Merge branch 'main' into token-usage-heuristic

This PR truncates the text output of MCP tool

codex-rs/Cargo.lock

@@ -843,6 +843,7 @@ dependencies = [
  "codex-backend-client",
  "codex-common",
  "codex-core",
+ "codex-feedback",
  "codex-file-search",
  "codex-login",
  "codex-protocol",
@@ -853,6 +854,7 @@ dependencies = [
  "pretty_assertions",
  "serde",
  "serde_json",
+ "serial_test",
  "tempfile",
  "tokio",
  "toml",
@@ -1061,10 +1063,13 @@ dependencies = [
  "codex-apply-patch",
  "codex-async-utils",
  "codex-file-search",
+ "codex-git-tooling",
+ "codex-keyring-store",
  "codex-otel",
  "codex-protocol",
  "codex-rmcp-client",
  "codex-utils-pty",
+ "codex-utils-readiness",
  "codex-utils-string",
  "codex-utils-tokenizer",
  "core-foundation 0.9.4",
@@ -1076,7 +1081,9 @@ dependencies = [
  "eventsource-stream",
  "futures",
  "http",
+ "image",
  "indexmap 2.10.0",
+ "keyring",
  "landlock",
  "libc",
  "maplit",
@@ -1093,6 +1100,7 @@ dependencies = [
  "serde_json",
  "serial_test",
  "sha1",
+ "sha2",
  "shlex",
  "similar",
  "strum_macros 0.27.2",
@@ -1208,11 +1216,22 @@ version = "0.0.0"
 dependencies = [
  "assert_matches",
  "pretty_assertions",
+ "schemars 0.8.22",
+ "serde",
  "tempfile",
  "thiserror 2.0.16",
+ "ts-rs",
  "walkdir",
 ]
 
+[[package]]
+name = "codex-keyring-store"
+version = "0.0.0"
+dependencies = [
+ "keyring",
+ "tracing",
+]
+
 [[package]]
 name = "codex-linux-sandbox"
 version = "0.0.0"
@@ -1327,6 +1346,8 @@ version = "0.0.0"
 dependencies = [
  "anyhow",
  "base64",
+ "codex-git-tooling",
+ "codex-utils-image",
  "icu_decimal",
  "icu_locale_core",
  "mcp-types",
@@ -1376,6 +1397,7 @@ version = "0.0.0"
 dependencies = [
  "anyhow",
  "axum",
+ "codex-keyring-store",
  "codex-protocol",
  "dirs",
  "escargot",
@@ -1427,7 +1449,6 @@ dependencies = [
  "codex-core",
  "codex-feedback",
  "codex-file-search",
- "codex-git-tooling",
  "codex-login",
  "codex-ollama",
  "codex-protocol",
@@ -1472,6 +1493,27 @@ dependencies = [
  "vt100",
 ]
 
+[[package]]
+name = "codex-utils-cache"
+version = "0.0.0"
+dependencies = [
+ "lru",
+ "sha1",
+ "tokio",
+]
+
+[[package]]
+name = "codex-utils-image"
+version = "0.0.0"
+dependencies = [
+ "base64",
+ "codex-utils-cache",
+ "image",
+ "tempfile",
+ "thiserror 2.0.16",
+ "tokio",
+]
+
 [[package]]
 name = "codex-utils-json-to-toml"
 version = "0.0.0"
@@ -5457,9 +5499,9 @@ dependencies = [
 
 [[package]]
 name = "serde"
-version = "1.0.226"
+version = "1.0.228"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "0dca6411025b24b60bfa7ec1fe1f8e710ac09782dca409ee8237ba74b51295fd"
+checksum = "9a8e94ea7f378bd32cbbd37198a4a91436180c5bb472411e48b5ec2e2124ae9e"
 dependencies = [
  "serde_core",
  "serde_derive",
@@ -5467,18 +5509,18 @@ dependencies = [
 
 [[package]]
 name = "serde_core"
-version = "1.0.226"
+version = "1.0.228"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "ba2ba63999edb9dac981fb34b3e5c0d111a69b0924e253ed29d83f7c99e966a4"
+checksum = "41d385c7d4ca58e59fc732af25c3983b67ac852c1a25000afe1175de458b67ad"
 dependencies = [
  "serde_derive",
 ]
 
 [[package]]
 name = "serde_derive"
-version = "1.0.226"
+version = "1.0.228"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "8db53ae22f34573731bafa1db20f04027b2d25e02d8205921b569171699cdb33"
+checksum = "d540f220d3187173da220f885ab66608367b6574e925011a9353e4badda91d79"
 dependencies = [
  "proc-macro2",
  "quote",

codex-rs/Cargo.toml

@@ -16,6 +16,7 @@ members = [
     "core",
     "exec",
     "execpolicy",
+    "keyring-store",
     "file-search",
     "git-tooling",
     "linux-sandbox",
@@ -32,9 +33,11 @@ members = [
     "otel",
     "tui",
     "git-apply",
+    "utils/cache",
+    "utils/image",
     "utils/json-to-toml",
-    "utils/readiness",
     "utils/pty",
+    "utils/readiness",
     "utils/string",
     "utils/tokenizer",
 ]
@@ -65,6 +68,7 @@ codex-exec = { path = "exec" }
 codex-feedback = { path = "feedback" }
 codex-file-search = { path = "file-search" }
 codex-git-tooling = { path = "git-tooling" }
+codex-keyring-store = { path = "keyring-store" }
 codex-linux-sandbox = { path = "linux-sandbox" }
 codex-login = { path = "login" }
 codex-mcp-server = { path = "mcp-server" }
@@ -77,6 +81,8 @@ codex-responses-api-proxy = { path = "responses-api-proxy" }
 codex-rmcp-client = { path = "rmcp-client" }
 codex-stdio-to-uds = { path = "stdio-to-uds" }
 codex-tui = { path = "tui" }
+codex-utils-cache = { path = "utils/cache" }
+codex-utils-image = { path = "utils/image" }
 codex-utils-json-to-toml = { path = "utils/json-to-toml" }
 codex-utils-pty = { path = "utils/pty" }
 codex-utils-readiness = { path = "utils/readiness" }
@@ -129,6 +135,7 @@ landlock = "0.4.1"
 lazy_static = "1"
 libc = "0.2.175"
 log = "0.4"
+lru = "0.12.5"
 maplit = "1.0.2"
 mime_guess = "2.0.5"
 multimap = "0.10.0"

codex-rs/app-server-protocol/src/export.rs

@@ -23,7 +23,6 @@ use std::io::Write;
 use std::path::Path;
 use std::path::PathBuf;
 use std::process::Command;
-use ts_rs::ExportError;
 use ts_rs::TS;
 
 const HEADER: &str = "// GENERATED CODE! DO NOT MODIFY BY HAND!\n\n";
@@ -105,19 +104,6 @@ macro_rules! for_each_schema_type {
     };
 }
 
-fn export_ts_with_context<F>(label: &str, export: F) -> Result<()>
-where
-    F: FnOnce() -> std::result::Result<(), ExportError>,
-{
-    match export() {
-        Ok(()) => Ok(()),
-        Err(ExportError::CannotBeExported(ty)) => Err(anyhow!(
-            "failed to export {label}: dependency {ty} cannot be exported"
-        )),
-        Err(err) => Err(err.into()),
-    }
-}
-
 pub fn generate_types(out_dir: &Path, prettier: Option<&Path>) -> Result<()> {
     generate_ts(out_dir, prettier)?;
     generate_json(out_dir)?;
@@ -127,17 +113,13 @@ pub fn generate_types(out_dir: &Path, prettier: Option<&Path>) -> Result<()> {
 pub fn generate_ts(out_dir: &Path, prettier: Option<&Path>) -> Result<()> {
     ensure_dir(out_dir)?;
 
-    export_ts_with_context("ClientRequest", || ClientRequest::export_all_to(out_dir))?;
-    export_ts_with_context("client responses", || export_client_responses(out_dir))?;
-    export_ts_with_context("ClientNotification", || {
-        ClientNotification::export_all_to(out_dir)
-    })?;
+    ClientRequest::export_all_to(out_dir)?;
+    export_client_responses(out_dir)?;
+    ClientNotification::export_all_to(out_dir)?;
 
-    export_ts_with_context("ServerRequest", || ServerRequest::export_all_to(out_dir))?;
-    export_ts_with_context("server responses", || export_server_responses(out_dir))?;
-    export_ts_with_context("ServerNotification", || {
-        ServerNotification::export_all_to(out_dir)
-    })?;
+    ServerRequest::export_all_to(out_dir)?;
+    export_server_responses(out_dir)?;
+    ServerNotification::export_all_to(out_dir)?;
 
     generate_index_ts(out_dir)?;
 

codex-rs/app-server-protocol/src/protocol.rs

@@ -5,7 +5,7 @@ use crate::JSONRPCNotification;
 use crate::JSONRPCRequest;
 use crate::RequestId;
 use codex_protocol::ConversationId;
-use codex_protocol::account::Account;
+use codex_protocol::account::PlanType;
 use codex_protocol::config_types::ForcedLoginMethod;
 use codex_protocol::config_types::ReasoningEffort;
 use codex_protocol::config_types::ReasoningSummary;
@@ -17,6 +17,7 @@ use codex_protocol::protocol::EventMsg;
 use codex_protocol::protocol::FileChange;
 use codex_protocol::protocol::RateLimitSnapshot;
 use codex_protocol::protocol::ReviewDecision;
+use codex_protocol::protocol::SandboxCommandAssessment;
 use codex_protocol::protocol::SandboxPolicy;
 use codex_protocol::protocol::TurnAbortReason;
 use paste::paste;
@@ -123,6 +124,13 @@ client_request_definitions! {
         response: GetAccountRateLimitsResponse,
     },
 
+    #[serde(rename = "feedback/upload")]
+    #[ts(rename = "feedback/upload")]
+    UploadFeedback {
+        params: UploadFeedbackParams,
+        response: UploadFeedbackResponse,
+    },
+
     #[serde(rename = "account/read")]
     #[ts(rename = "account/read")]
     GetAccount {
@@ -139,6 +147,10 @@ client_request_definitions! {
         params: NewConversationParams,
         response: NewConversationResponse,
     },
+    GetConversationSummary {
+        params: GetConversationSummaryParams,
+        response: GetConversationSummaryResponse,
+    },
     /// List recorded Codex conversations (rollouts) with optional pagination and search.
     ListConversations {
         params: ListConversationsParams,
@@ -224,6 +236,28 @@ client_request_definitions! {
     },
 }
 
+#[derive(Debug, Clone, PartialEq, Deserialize, Serialize, JsonSchema, TS)]
+#[serde(tag = "type", rename_all = "camelCase")]
+#[ts(tag = "type")]
+pub enum Account {
+    #[serde(rename = "apiKey", rename_all = "camelCase")]
+    #[ts(rename = "apiKey", rename_all = "camelCase")]
+    ApiKey { api_key: String },
+
+    #[serde(rename = "chatgpt", rename_all = "camelCase")]
+    #[ts(rename = "chatgpt", rename_all = "camelCase")]
+    ChatGpt {
+        email: Option<String>,
+        plan_type: PlanType,
+    },
+}
+
+#[derive(Serialize, Deserialize, Debug, Clone, PartialEq, JsonSchema, TS)]
+#[serde(rename_all = "camelCase")]
+pub struct GetAccountResponse {
+    pub account: Account,
+}
+
 #[derive(Serialize, Deserialize, Debug, Clone, PartialEq, Default, JsonSchema, TS)]
 #[serde(rename_all = "camelCase")]
 pub struct InitializeParams {
@@ -252,6 +286,10 @@ pub struct NewConversationParams {
     #[serde(skip_serializing_if = "Option::is_none")]
     pub model: Option<String>,
 
+    /// Override the model provider to use for this session.
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub model_provider: Option<String>,
+
     /// Configuration profile from config.toml to specify default options.
     #[serde(skip_serializing_if = "Option::is_none")]
     pub profile: Option<String>,
@@ -304,6 +342,18 @@ pub struct ResumeConversationResponse {
     pub initial_messages: Option<Vec<EventMsg>>,
 }
 
+#[derive(Serialize, Deserialize, Debug, Clone, PartialEq, JsonSchema, TS)]
+#[serde(rename_all = "camelCase")]
+pub struct GetConversationSummaryParams {
+    pub rollout_path: PathBuf,
+}
+
+#[derive(Serialize, Deserialize, Debug, Clone, JsonSchema, TS)]
+#[serde(rename_all = "camelCase")]
+pub struct GetConversationSummaryResponse {
+    pub summary: ConversationSummary,
+}
+
 #[derive(Serialize, Deserialize, Debug, Clone, PartialEq, Default, JsonSchema, TS)]
 #[serde(rename_all = "camelCase")]
 pub struct ListConversationsParams {
@@ -313,6 +363,12 @@ pub struct ListConversationsParams {
     /// Opaque pagination cursor returned by a previous call.
     #[serde(skip_serializing_if = "Option::is_none")]
     pub cursor: Option<String>,
+    /// Optional model provider filter (matches against session metadata).
+    /// - None => filter by the server's default model provider
+    /// - Some([]) => no filtering, include all providers
+    /// - Some([...]) => only include sessions with one of the specified providers
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub model_providers: Option<Vec<String>>,
 }
 
 #[derive(Serialize, Deserialize, Debug, Clone, PartialEq, JsonSchema, TS)]
@@ -324,6 +380,8 @@ pub struct ConversationSummary {
     /// RFC3339 timestamp string for the session start, if available.
     #[serde(skip_serializing_if = "Option::is_none")]
     pub timestamp: Option<String>,
+    /// Model provider recorded for the session (resolved when absent in metadata).
+    pub model_provider: String,
 }
 
 #[derive(Serialize, Deserialize, Debug, Clone, PartialEq, JsonSchema, TS)]
@@ -377,6 +435,23 @@ pub struct ListModelsResponse {
     pub next_cursor: Option<String>,
 }
 
+#[derive(Serialize, Deserialize, Debug, Clone, PartialEq, JsonSchema, TS)]
+#[serde(rename_all = "camelCase")]
+pub struct UploadFeedbackParams {
+    pub classification: String,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub reason: Option<String>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub conversation_id: Option<ConversationId>,
+    pub include_logs: bool,
+}
+
+#[derive(Serialize, Deserialize, Debug, Clone, PartialEq, JsonSchema, TS)]
+#[serde(rename_all = "camelCase")]
+pub struct UploadFeedbackResponse {
+    pub thread_id: String,
+}
+
 #[derive(Serialize, Deserialize, Debug, Clone, PartialEq, JsonSchema, TS)]
 #[serde(tag = "type")]
 #[ts(tag = "type")]
@@ -534,12 +609,6 @@ pub struct GetAccountRateLimitsResponse {
     pub rate_limits: RateLimitSnapshot,
 }
 
-#[derive(Serialize, Deserialize, Debug, Clone, PartialEq, JsonSchema, TS)]
-#[serde(transparent)]
-#[ts(export)]
-#[ts(type = "Account | null")]
-pub struct GetAccountResponse(#[ts(type = "Account | null")] pub Option<Account>);
-
 #[derive(Serialize, Deserialize, Debug, Clone, PartialEq, JsonSchema, TS)]
 #[serde(rename_all = "camelCase")]
 pub struct GetAuthStatusResponse {
@@ -716,6 +785,8 @@ pub struct SendUserMessageResponse {}
 #[serde(rename_all = "camelCase")]
 pub struct AddConversationListenerParams {
     pub conversation_id: ConversationId,
+    #[serde(default)]
+    pub experimental_raw_events: bool,
 }
 
 #[derive(Serialize, Deserialize, Debug, Clone, PartialEq, JsonSchema, TS)]
@@ -847,6 +918,8 @@ pub struct ExecCommandApprovalParams {
     pub cwd: PathBuf,
     #[serde(skip_serializing_if = "Option::is_none")]
     pub reason: Option<String>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub risk: Option<SandboxCommandAssessment>,
     pub parsed_cmd: Vec<ParsedCommand>,
 }
 
@@ -995,6 +1068,7 @@ mod tests {
             request_id: RequestId::Integer(42),
             params: NewConversationParams {
                 model: Some("gpt-5-codex".to_string()),
+                model_provider: None,
                 profile: None,
                 cwd: None,
                 approval_policy: Some(AskForApproval::OnRequest),
@@ -1063,6 +1137,7 @@ mod tests {
             command: vec!["echo".to_string(), "hello".to_string()],
             cwd: PathBuf::from("/tmp"),
             reason: Some("because tests".to_string()),
+            risk: None,
             parsed_cmd: vec![ParsedCommand::Unknown {
                 cmd: "echo hello".to_string(),
             }],
@@ -1187,6 +1262,35 @@ mod tests {
         Ok(())
     }
 
+    #[test]
+    fn account_serializes_fields_in_camel_case() -> Result<()> {
+        let api_key = Account::ApiKey {
+            api_key: "secret".to_string(),
+        };
+        assert_eq!(
+            json!({
+                "type": "apiKey",
+                "apiKey": "secret",
+            }),
+            serde_json::to_value(&api_key)?,
+        );
+
+        let chatgpt = Account::ChatGpt {
+            email: Some("user@example.com".to_string()),
+            plan_type: PlanType::Plus,
+        };
+        assert_eq!(
+            json!({
+                "type": "chatgpt",
+                "email": "user@example.com",
+                "planType": "plus",
+            }),
+            serde_json::to_value(&chatgpt)?,
+        );
+
+        Ok(())
+    }
+
     #[test]
     fn serialize_list_models() -> Result<()> {
         let request = ClientRequest::ListModels {

codex-rs/app-server/Cargo.toml

@@ -24,6 +24,7 @@ codex-file-search = { workspace = true }
 codex-login = { workspace = true }
 codex-protocol = { workspace = true }
 codex-app-server-protocol = { workspace = true }
+codex-feedback = { workspace = true }
 codex-utils-json-to-toml = { workspace = true }
 chrono = { workspace = true }
 serde = { workspace = true, features = ["derive"] }
@@ -47,6 +48,7 @@ base64 = { workspace = true }
 core_test_support = { workspace = true }
 os_info = { workspace = true }
 pretty_assertions = { workspace = true }
+serial_test = { workspace = true }
 tempfile = { workspace = true }
 toml = { workspace = true }
 wiremock = { workspace = true }

codex-rs/app-server/src/codex_message_processor.rs

@@ -21,6 +21,8 @@ use codex_app_server_protocol::ExecOneOffCommandResponse;
 use codex_app_server_protocol::FuzzyFileSearchParams;
 use codex_app_server_protocol::FuzzyFileSearchResponse;
 use codex_app_server_protocol::GetAccountRateLimitsResponse;
+use codex_app_server_protocol::GetConversationSummaryParams;
+use codex_app_server_protocol::GetConversationSummaryResponse;
 use codex_app_server_protocol::GetUserAgentResponse;
 use codex_app_server_protocol::GetUserSavedConfigResponse;
 use codex_app_server_protocol::GitDiffToRemoteResponse;
@@ -52,6 +54,8 @@ use codex_app_server_protocol::ServerRequestPayload;
 use codex_app_server_protocol::SessionConfiguredNotification;
 use codex_app_server_protocol::SetDefaultModelParams;
 use codex_app_server_protocol::SetDefaultModelResponse;
+use codex_app_server_protocol::UploadFeedbackParams;
+use codex_app_server_protocol::UploadFeedbackResponse;
 use codex_app_server_protocol::UserInfoResponse;
 use codex_app_server_protocol::UserSavedConfig;
 use codex_backend_client::Client as BackendClient;
@@ -64,9 +68,7 @@ use codex_core::NewConversation;
 use codex_core::RolloutRecorder;
 use codex_core::SessionMeta;
 use codex_core::auth::CLIENT_ID;
-use codex_core::auth::get_auth_file;
 use codex_core::auth::login_with_api_key;
-use codex_core::auth::try_read_auth_json;
 use codex_core::config::Config;
 use codex_core::config::ConfigOverrides;
 use codex_core::config::ConfigToml;
@@ -85,6 +87,8 @@ use codex_core::protocol::EventMsg;
 use codex_core::protocol::ExecApprovalRequestEvent;
 use codex_core::protocol::Op;
 use codex_core::protocol::ReviewDecision;
+use codex_core::read_head_for_summary;
+use codex_feedback::CodexFeedback;
 use codex_login::ServerOptions as LoginServerOptions;
 use codex_login::ShutdownHandle;
 use codex_login::run_login_server;
@@ -98,6 +102,8 @@ use codex_protocol::user_input::UserInput as CoreInputItem;
 use codex_utils_json_to_toml::json_to_toml;
 use std::collections::HashMap;
 use std::ffi::OsStr;
+use std::io::Error as IoError;
+use std::path::Path;
 use std::path::PathBuf;
 use std::sync::Arc;
 use std::sync::atomic::AtomicBool;
@@ -136,6 +142,7 @@ pub(crate) struct CodexMessageProcessor {
     // Queue of pending interrupt requests per conversation. We reply when TurnAborted arrives.
     pending_interrupts: Arc<Mutex<HashMap<ConversationId, Vec<RequestId>>>>,
     pending_fuzzy_searches: Arc<Mutex<HashMap<String, Arc<AtomicBool>>>>,
+    feedback: CodexFeedback,
 }
 
 impl CodexMessageProcessor {
@@ -145,6 +152,7 @@ impl CodexMessageProcessor {
         outgoing: Arc<OutgoingMessageSender>,
         codex_linux_sandbox_exe: Option<PathBuf>,
         config: Arc<Config>,
+        feedback: CodexFeedback,
     ) -> Self {
         Self {
             auth_manager,
@@ -156,6 +164,7 @@ impl CodexMessageProcessor {
             active_login: Arc::new(Mutex::new(None)),
             pending_interrupts: Arc::new(Mutex::new(HashMap::new())),
             pending_fuzzy_searches: Arc::new(Mutex::new(HashMap::new())),
+            feedback,
         }
     }
 
@@ -170,6 +179,9 @@ impl CodexMessageProcessor {
                 // created before processing any subsequent messages.
                 self.process_new_conversation(request_id, params).await;
             }
+            ClientRequest::GetConversationSummary { request_id, params } => {
+                self.get_conversation_summary(request_id, params).await;
+            }
             ClientRequest::ListConversations { request_id, params } => {
                 self.handle_list_conversations(request_id, params).await;
             }
@@ -275,6 +287,9 @@ impl CodexMessageProcessor {
             } => {
                 self.get_account_rate_limits(request_id).await;
             }
+            ClientRequest::UploadFeedback { request_id, params } => {
+                self.upload_feedback(request_id, params).await;
+            }
         }
     }
 
@@ -654,12 +669,8 @@ impl CodexMessageProcessor {
     }
 
     async fn get_user_info(&self, request_id: RequestId) {
-        // Read alleged user email from auth.json (best-effort; not verified).
-        let auth_path = get_auth_file(&self.config.codex_home);
-        let alleged_user_email = match try_read_auth_json(&auth_path) {
-            Ok(auth) => auth.tokens.and_then(|t| t.id_token.email),
-            Err(_) => None,
-        };
+        // Read alleged user email from cached auth (best-effort; not verified).
+        let alleged_user_email = self.auth_manager.auth().and_then(|a| a.get_account_email());
 
         let response = UserInfoResponse { alleged_user_email };
         self.outgoing.send_response(request_id, response).await;
@@ -813,24 +824,76 @@ impl CodexMessageProcessor {
         }
     }
 
+    async fn get_conversation_summary(
+        &self,
+        request_id: RequestId,
+        params: GetConversationSummaryParams,
+    ) {
+        let GetConversationSummaryParams { rollout_path } = params;
+        let path = if rollout_path.is_relative() {
+            self.config.codex_home.join(&rollout_path)
+        } else {
+            rollout_path.clone()
+        };
+        let fallback_provider = self.config.model_provider_id.as_str();
+
+        match read_summary_from_rollout(&path, fallback_provider).await {
+            Ok(summary) => {
+                let response = GetConversationSummaryResponse { summary };
+                self.outgoing.send_response(request_id, response).await;
+            }
+            Err(err) => {
+                let error = JSONRPCErrorError {
+                    code: INTERNAL_ERROR_CODE,
+                    message: format!(
+                        "failed to load conversation summary from {}: {}",
+                        path.display(),
+                        err
+                    ),
+                    data: None,
+                };
+                self.outgoing.send_error(request_id, error).await;
+            }
+        }
+    }
+
     async fn handle_list_conversations(
         &self,
         request_id: RequestId,
         params: ListConversationsParams,
     ) {
-        let page_size = params.page_size.unwrap_or(25);
+        let ListConversationsParams {
+            page_size,
+            cursor,
+            model_providers: model_provider,
+        } = params;
+        let page_size = page_size.unwrap_or(25);
         // Decode the optional cursor string to a Cursor via serde (Cursor implements Deserialize from string)
-        let cursor_obj: Option<RolloutCursor> = match params.cursor {
+        let cursor_obj: Option<RolloutCursor> = match cursor {
             Some(s) => serde_json::from_str::<RolloutCursor>(&format!("\"{s}\"")).ok(),
             None => None,
         };
         let cursor_ref = cursor_obj.as_ref();
+        let model_provider_filter = match model_provider {
+            Some(providers) => {
+                if providers.is_empty() {
+                    None
+                } else {
+                    Some(providers)
+                }
+            }
+            None => Some(vec![self.config.model_provider_id.clone()]),
+        };
+        let model_provider_slice = model_provider_filter.as_deref();
+        let fallback_provider = self.config.model_provider_id.clone();
 
         let page = match RolloutRecorder::list_conversations(
             &self.config.codex_home,
             page_size,
             cursor_ref,
             INTERACTIVE_SESSION_SOURCES,
+            model_provider_slice,
+            fallback_provider.as_str(),
         )
         .await
         {
@@ -849,7 +912,7 @@ impl CodexMessageProcessor {
         let items = page
             .items
             .into_iter()
-            .filter_map(|it| extract_conversation_summary(it.path, &it.head))
+            .filter_map(|it| extract_conversation_summary(it.path, &it.head, &fallback_provider))
             .collect();
 
         // Encode next_cursor as a plain string
@@ -1256,7 +1319,10 @@ impl CodexMessageProcessor {
         request_id: RequestId,
         params: AddConversationListenerParams,
     ) {
-        let AddConversationListenerParams { conversation_id } = params;
+        let AddConversationListenerParams {
+            conversation_id,
+            experimental_raw_events,
+        } = params;
         let Ok(conversation) = self
             .conversation_manager
             .get_conversation(conversation_id)
@@ -1293,6 +1359,11 @@ impl CodexMessageProcessor {
                             }
                         };
 
+                        if let EventMsg::RawResponseItem(_) = &event.msg
+                            && !experimental_raw_events {
+                                continue;
+                            }
+
                         // For now, we send a notification for every event,
                         // JSON-serializing the `Event` as-is, but these should
                         // be migrated to be variants of `ServerNotification`
@@ -1410,6 +1481,77 @@ impl CodexMessageProcessor {
         let response = FuzzyFileSearchResponse { files: results };
         self.outgoing.send_response(request_id, response).await;
     }
+
+    async fn upload_feedback(&self, request_id: RequestId, params: UploadFeedbackParams) {
+        let UploadFeedbackParams {
+            classification,
+            reason,
+            conversation_id,
+            include_logs,
+        } = params;
+
+        let snapshot = self.feedback.snapshot(conversation_id);
+        let thread_id = snapshot.thread_id.clone();
+
+        let validated_rollout_path = if include_logs {
+            match conversation_id {
+                Some(conv_id) => self.resolve_rollout_path(conv_id).await,
+                None => None,
+            }
+        } else {
+            None
+        };
+
+        let upload_result = tokio::task::spawn_blocking(move || {
+            let rollout_path_ref = validated_rollout_path.as_deref();
+            snapshot.upload_feedback(
+                &classification,
+                reason.as_deref(),
+                include_logs,
+                rollout_path_ref,
+            )
+        })
+        .await;
+
+        let upload_result = match upload_result {
+            Ok(result) => result,
+            Err(join_err) => {
+                let error = JSONRPCErrorError {
+                    code: INTERNAL_ERROR_CODE,
+                    message: format!("failed to upload feedback: {join_err}"),
+                    data: None,
+                };
+                self.outgoing.send_error(request_id, error).await;
+                return;
+            }
+        };
+
+        match upload_result {
+            Ok(()) => {
+                let response = UploadFeedbackResponse { thread_id };
+                self.outgoing.send_response(request_id, response).await;
+            }
+            Err(err) => {
+                let error = JSONRPCErrorError {
+                    code: INTERNAL_ERROR_CODE,
+                    message: format!("failed to upload feedback: {err}"),
+                    data: None,
+                };
+                self.outgoing.send_error(request_id, error).await;
+            }
+        }
+    }
+
+    async fn resolve_rollout_path(&self, conversation_id: ConversationId) -> Option<PathBuf> {
+        match self
+            .conversation_manager
+            .get_conversation(conversation_id)
+            .await
+        {
+            Ok(conv) => Some(conv.rollout_path()),
+            Err(_) => None,
+        }
+    }
 }
 
 async fn apply_bespoke_event_handling(
@@ -1447,6 +1589,7 @@ async fn apply_bespoke_event_handling(
             command,
             cwd,
             reason,
+            risk,
             parsed_cmd,
         }) => {
             let params = ExecCommandApprovalParams {
@@ -1455,6 +1598,7 @@ async fn apply_bespoke_event_handling(
                 command,
                 cwd,
                 reason,
+                risk,
                 parsed_cmd,
             };
             let rx = outgoing
@@ -1501,6 +1645,7 @@ async fn derive_config_from_params(
 ) -> std::io::Result<Config> {
     let NewConversationParams {
         model,
+        model_provider,
         profile,
         cwd,
         approval_policy,
@@ -1516,13 +1661,14 @@ async fn derive_config_from_params(
         cwd: cwd.map(PathBuf::from),
         approval_policy,
         sandbox_mode,
-        model_provider: None,
+        model_provider,
         codex_linux_sandbox_exe,
         base_instructions,
         include_apply_patch_tool,
         include_view_image_tool: None,
         show_raw_agent_reasoning: None,
         tools_web_search_request: None,
+        experimental_sandbox_command_assessment: None,
         additional_writable_roots: Vec::new(),
     };
 
@@ -1613,9 +1759,54 @@ async fn on_exec_approval_response(
     }
 }
 
+async fn read_summary_from_rollout(
+    path: &Path,
+    fallback_provider: &str,
+) -> std::io::Result<ConversationSummary> {
+    let head = read_head_for_summary(path).await?;
+
+    let Some(first) = head.first() else {
+        return Err(IoError::other(format!(
+            "rollout at {} is empty",
+            path.display()
+        )));
+    };
+
+    let session_meta = serde_json::from_value::<SessionMeta>(first.clone()).map_err(|_| {
+        IoError::other(format!(
+            "rollout at {} does not start with session metadata",
+            path.display()
+        ))
+    })?;
+
+    if let Some(summary) =
+        extract_conversation_summary(path.to_path_buf(), &head, fallback_provider)
+    {
+        return Ok(summary);
+    }
+
+    let timestamp = if session_meta.timestamp.is_empty() {
+        None
+    } else {
+        Some(session_meta.timestamp.clone())
+    };
+    let model_provider = session_meta
+        .model_provider
+        .unwrap_or_else(|| fallback_provider.to_string());
+
+    Ok(ConversationSummary {
+        conversation_id: session_meta.id,
+        timestamp,
+        path: path.to_path_buf(),
+        preview: String::new(),
+        model_provider,
+    })
+}
+
 fn extract_conversation_summary(
     path: PathBuf,
     head: &[serde_json::Value],
+    fallback_provider: &str,
 ) -> Option<ConversationSummary> {
     let session_meta = match head.first() {
         Some(first_line) => serde_json::from_value::<SessionMeta>(first_line.clone()).ok()?,
@@ -1640,12 +1831,17 @@ fn extract_conversation_summary(
     } else {
         Some(session_meta.timestamp.clone())
     };
+    let conversation_id = session_meta.id;
+    let model_provider = session_meta
+        .model_provider
+        .unwrap_or_else(|| fallback_provider.to_string());
 
     Some(ConversationSummary {
-        conversation_id: session_meta.id,
+        conversation_id,
         timestamp,
         path,
         preview: preview.to_string(),
+        model_provider,
     })
 }
 
@@ -1655,6 +1851,7 @@ mod tests {
     use anyhow::Result;
     use pretty_assertions::assert_eq;
     use serde_json::json;
+    use tempfile::TempDir;
 
     #[test]
     fn extract_conversation_summary_prefers_plain_user_messages() -> Result<()> {
@@ -1669,7 +1866,8 @@ mod tests {
                 "cwd": "/",
                 "originator": "codex",
                 "cli_version": "0.0.0",
-                "instructions": null
+                "instructions": null,
+                "model_provider": "test-provider"
             }),
             json!({
                 "type": "message",
@@ -1689,15 +1887,62 @@ mod tests {
             }),
         ];
 
-        let summary = extract_conversation_summary(path.clone(), &head).expect("summary");
+        let summary =
+            extract_conversation_summary(path.clone(), &head, "test-provider").expect("summary");
 
-        assert_eq!(summary.conversation_id, conversation_id);
-        assert_eq!(
-            summary.timestamp,
-            Some("2025-09-05T16:53:11.850Z".to_string())
-        );
-        assert_eq!(summary.path, path);
-        assert_eq!(summary.preview, "Count to 5");
+        let expected = ConversationSummary {
+            conversation_id,
+            timestamp,
+            path,
+            preview: "Count to 5".to_string(),
+            model_provider: "test-provider".to_string(),
+        };
+
+        assert_eq!(summary, expected);
+        Ok(())
+    }
+
+    #[tokio::test]
+    async fn read_summary_from_rollout_returns_empty_preview_when_no_user_message() -> Result<()> {
+        use codex_protocol::protocol::RolloutItem;
+        use codex_protocol::protocol::RolloutLine;
+        use codex_protocol::protocol::SessionMetaLine;
+        use std::fs;
+
+        let temp_dir = TempDir::new()?;
+        let path = temp_dir.path().join("rollout.jsonl");
+
+        let conversation_id = ConversationId::from_string("bfd12a78-5900-467b-9bc5-d3d35df08191")?;
+        let timestamp = "2025-09-05T16:53:11.850Z".to_string();
+
+        let session_meta = SessionMeta {
+            id: conversation_id,
+            timestamp: timestamp.clone(),
+            model_provider: None,
+            ..SessionMeta::default()
+        };
+
+        let line = RolloutLine {
+            timestamp: timestamp.clone(),
+            item: RolloutItem::SessionMeta(SessionMetaLine {
+                meta: session_meta.clone(),
+                git: None,
+            }),
+        };
+
+        fs::write(&path, format!("{}\n", serde_json::to_string(&line)?))?;
+
+        let summary = read_summary_from_rollout(path.as_path(), "fallback").await?;
+
+        let expected = ConversationSummary {
+            conversation_id,
+            timestamp: Some(timestamp),
+            path: path.clone(),
+            preview: String::new(),
+            model_provider: "fallback".to_string(),
+        };
+
+        assert_eq!(summary, expected);
         Ok(())
     }
 }

codex-rs/app-server/src/lib.rs

@@ -12,16 +12,19 @@ use crate::message_processor::MessageProcessor;
 use crate::outgoing_message::OutgoingMessage;
 use crate::outgoing_message::OutgoingMessageSender;
 use codex_app_server_protocol::JSONRPCMessage;
+use codex_feedback::CodexFeedback;
 use tokio::io::AsyncBufReadExt;
 use tokio::io::AsyncWriteExt;
 use tokio::io::BufReader;
 use tokio::io::{self};
 use tokio::sync::mpsc;
+use tracing::Level;
 use tracing::debug;
 use tracing::error;
 use tracing::info;
 use tracing_subscriber::EnvFilter;
 use tracing_subscriber::Layer;
+use tracing_subscriber::filter::Targets;
 use tracing_subscriber::layer::SubscriberExt;
 use tracing_subscriber::util::SubscriberInitExt;
 
@@ -82,6 +85,8 @@ pub async fn run_main(
             std::io::Error::new(ErrorKind::InvalidData, format!("error loading config: {e}"))
         })?;
 
+    let feedback = CodexFeedback::new();
+
     let otel =
         codex_core::otel_init::build_provider(&config, env!("CARGO_PKG_VERSION")).map_err(|e| {
             std::io::Error::new(
@@ -96,8 +101,15 @@ pub async fn run_main(
         .with_writer(std::io::stderr)
         .with_filter(EnvFilter::from_default_env());
 
+    let feedback_layer = tracing_subscriber::fmt::layer()
+        .with_writer(feedback.make_writer())
+        .with_ansi(false)
+        .with_target(false)
+        .with_filter(Targets::new().with_default(Level::TRACE));
+
     let _ = tracing_subscriber::registry()
         .with(stderr_fmt)
+        .with(feedback_layer)
         .with(otel.as_ref().map(|provider| {
             OpenTelemetryTracingBridge::new(&provider.logger).with_filter(
                 tracing_subscriber::filter::filter_fn(codex_core::otel_init::codex_export_filter),
@@ -112,6 +124,7 @@ pub async fn run_main(
             outgoing_message_sender,
             codex_linux_sandbox_exe,
             std::sync::Arc::new(config),
+            feedback.clone(),
         );
         async move {
             while let Some(msg) = incoming_rx.recv().await {

codex-rs/app-server/src/message_processor.rs

@@ -17,6 +17,7 @@ use codex_core::ConversationManager;
 use codex_core::config::Config;
 use codex_core::default_client::USER_AGENT_SUFFIX;
 use codex_core::default_client::get_codex_user_agent;
+use codex_feedback::CodexFeedback;
 use codex_protocol::protocol::SessionSource;
 use std::sync::Arc;
 
@@ -33,6 +34,7 @@ impl MessageProcessor {
         outgoing: OutgoingMessageSender,
         codex_linux_sandbox_exe: Option<PathBuf>,
         config: Arc<Config>,
+        feedback: CodexFeedback,
     ) -> Self {
         let outgoing = Arc::new(outgoing);
         let auth_manager = AuthManager::shared(config.codex_home.clone(), false);
@@ -46,6 +48,7 @@ impl MessageProcessor {
             outgoing.clone(),
             codex_linux_sandbox_exe,
             config,
+            feedback,
         );
 
         Self {

codex-rs/app-server/tests/common/auth_fixtures.rs

@@ -7,8 +7,7 @@ use base64::engine::general_purpose::URL_SAFE_NO_PAD;
 use chrono::DateTime;
 use chrono::Utc;
 use codex_core::auth::AuthDotJson;
-use codex_core::auth::get_auth_file;
-use codex_core::auth::write_auth_json;
+use codex_core::auth::save_auth;
 use codex_core::token_data::TokenData;
 use codex_core::token_data::parse_id_token;
 use serde_json::json;
@@ -127,5 +126,5 @@ pub fn write_chatgpt_auth(codex_home: &Path, fixture: ChatGptAuthFixture) -> Res
         last_refresh,
     };
 
-    write_auth_json(&get_auth_file(codex_home), &auth).context("write auth.json")
+    save_auth(codex_home, &auth).context("write auth.json")
 }

codex-rs/app-server/tests/common/mcp_process.rs

@@ -30,6 +30,7 @@ use codex_app_server_protocol::SendUserMessageParams;
 use codex_app_server_protocol::SendUserTurnParams;
 use codex_app_server_protocol::ServerRequest;
 use codex_app_server_protocol::SetDefaultModelParams;
+use codex_app_server_protocol::UploadFeedbackParams;
 
 use codex_app_server_protocol::JSONRPCError;
 use codex_app_server_protocol::JSONRPCMessage;
@@ -242,6 +243,15 @@ impl McpProcess {
         self.send_request("account/rateLimits/read", None).await
     }
 
+    /// Send a `feedback/upload` JSON-RPC request.
+    pub async fn send_upload_feedback_request(
+        &mut self,
+        params: UploadFeedbackParams,
+    ) -> anyhow::Result<i64> {
+        let params = Some(serde_json::to_value(params)?);
+        self.send_request("feedback/upload", params).await
+    }
+
     /// Send a `userInfo` JSON-RPC request.
     pub async fn send_user_info_request(&mut self) -> anyhow::Result<i64> {
         self.send_request("userInfo", None).await

codex-rs/app-server/tests/suite/codex_message_processor_flow.rs

@@ -103,7 +103,10 @@ async fn test_codex_jsonrpc_conversation_flow() {
 
     // 2) addConversationListener
     let add_listener_id = mcp
-        .send_add_conversation_listener_request(AddConversationListenerParams { conversation_id })
+        .send_add_conversation_listener_request(AddConversationListenerParams {
+            conversation_id,
+            experimental_raw_events: false,
+        })
         .await
         .expect("send addConversationListener");
     let add_listener_resp: JSONRPCResponse = timeout(
@@ -252,7 +255,10 @@ async fn test_send_user_turn_changes_approval_policy_behavior() {
 
     // 2) addConversationListener
     let add_listener_id = mcp
-        .send_add_conversation_listener_request(AddConversationListenerParams { conversation_id })
+        .send_add_conversation_listener_request(AddConversationListenerParams {
+            conversation_id,
+            experimental_raw_events: false,
+        })
         .await
         .expect("send addConversationListener");
     let _: AddConversationSubscriptionResponse =
@@ -311,6 +317,7 @@ async fn test_send_user_turn_changes_approval_policy_behavior() {
             ],
             cwd: working_directory.clone(),
             reason: None,
+            risk: None,
             parsed_cmd: vec![ParsedCommand::Unknown {
                 cmd: "python3 -c 'print(42)'".to_string()
             }],
@@ -458,7 +465,10 @@ async fn test_send_user_turn_updates_sandbox_and_cwd_between_turns() {
         .expect("deserialize newConversation response");
 
     let add_listener_id = mcp
-        .send_add_conversation_listener_request(AddConversationListenerParams { conversation_id })
+        .send_add_conversation_listener_request(AddConversationListenerParams {
+            conversation_id,
+            experimental_raw_events: false,
+        })
         .await
         .expect("send addConversationListener");
     timeout(

codex-rs/app-server/tests/suite/create_conversation.rs

@@ -67,7 +67,10 @@ async fn test_conversation_create_and_send_message_ok() {
 
     // Add a listener so we receive notifications for this conversation (not strictly required for this test).
     let add_listener_id = mcp
-        .send_add_conversation_listener_request(AddConversationListenerParams { conversation_id })
+        .send_add_conversation_listener_request(AddConversationListenerParams {
+            conversation_id,
+            experimental_raw_events: false,
+        })
         .await
         .expect("send addConversationListener");
     let _sub: AddConversationSubscriptionResponse =

codex-rs/app-server/tests/suite/interrupt.rs

@@ -88,7 +88,10 @@ async fn shell_command_interruption() -> anyhow::Result<()> {
 
     // 2) addConversationListener
     let add_listener_id = mcp
-        .send_add_conversation_listener_request(AddConversationListenerParams { conversation_id })
+        .send_add_conversation_listener_request(AddConversationListenerParams {
+            conversation_id,
+            experimental_raw_events: false,
+        })
         .await?;
     let _add_listener_resp: JSONRPCResponse = timeout(
         DEFAULT_READ_TIMEOUT,

codex-rs/app-server/tests/suite/list_resume.rs

@@ -30,18 +30,21 @@ async fn test_list_and_resume_conversations() {
         "2025-01-02T12-00-00",
         "2025-01-02T12:00:00Z",
         "Hello A",
+        Some("openai"),
     );
     create_fake_rollout(
         codex_home.path(),
         "2025-01-01T13-00-00",
         "2025-01-01T13:00:00Z",
         "Hello B",
+        Some("openai"),
     );
     create_fake_rollout(
         codex_home.path(),
         "2025-01-01T12-00-00",
         "2025-01-01T12:00:00Z",
         "Hello C",
+        None,
     );
 
     let mut mcp = McpProcess::new(codex_home.path())
@@ -57,6 +60,7 @@ async fn test_list_and_resume_conversations() {
         .send_list_conversations_request(ListConversationsParams {
             page_size: Some(2),
             cursor: None,
+            model_providers: None,
         })
         .await
         .expect("send listConversations");
@@ -74,6 +78,8 @@ async fn test_list_and_resume_conversations() {
     // Newest first; preview text should match
     assert_eq!(items[0].preview, "Hello A");
     assert_eq!(items[1].preview, "Hello B");
+    assert_eq!(items[0].model_provider, "openai");
+    assert_eq!(items[1].model_provider, "openai");
     assert!(items[0].path.is_absolute());
     assert!(next_cursor.is_some());
 
@@ -82,6 +88,7 @@ async fn test_list_and_resume_conversations() {
         .send_list_conversations_request(ListConversationsParams {
             page_size: Some(2),
             cursor: next_cursor,
+            model_providers: None,
         })
         .await
         .expect("send listConversations page 2");
@@ -99,7 +106,88 @@ async fn test_list_and_resume_conversations() {
     } = to_response::<ListConversationsResponse>(resp2).expect("deserialize response");
     assert_eq!(items2.len(), 1);
     assert_eq!(items2[0].preview, "Hello C");
-    assert!(next2.is_some());
+    assert_eq!(items2[0].model_provider, "openai");
+    assert_eq!(next2, None);
+
+    // Add a conversation with an explicit non-OpenAI provider for filter tests.
+    create_fake_rollout(
+        codex_home.path(),
+        "2025-01-01T11-30-00",
+        "2025-01-01T11:30:00Z",
+        "Hello TP",
+        Some("test-provider"),
+    );
+
+    // Filtering by model provider should return only matching sessions.
+    let filter_req_id = mcp
+        .send_list_conversations_request(ListConversationsParams {
+            page_size: Some(10),
+            cursor: None,
+            model_providers: Some(vec!["test-provider".to_string()]),
+        })
+        .await
+        .expect("send listConversations filtered");
+    let filter_resp: JSONRPCResponse = timeout(
+        DEFAULT_READ_TIMEOUT,
+        mcp.read_stream_until_response_message(RequestId::Integer(filter_req_id)),
+    )
+    .await
+    .expect("listConversations filtered timeout")
+    .expect("listConversations filtered resp");
+    let ListConversationsResponse {
+        items: filtered_items,
+        next_cursor: filtered_next,
+    } = to_response::<ListConversationsResponse>(filter_resp).expect("deserialize filtered");
+    assert_eq!(filtered_items.len(), 1);
+    assert_eq!(filtered_next, None);
+    assert_eq!(filtered_items[0].preview, "Hello TP");
+    assert_eq!(filtered_items[0].model_provider, "test-provider");
+
+    // Empty filter should include every session regardless of provider metadata.
+    let unfiltered_req_id = mcp
+        .send_list_conversations_request(ListConversationsParams {
+            page_size: Some(10),
+            cursor: None,
+            model_providers: Some(Vec::new()),
+        })
+        .await
+        .expect("send listConversations unfiltered");
+    let unfiltered_resp: JSONRPCResponse = timeout(
+        DEFAULT_READ_TIMEOUT,
+        mcp.read_stream_until_response_message(RequestId::Integer(unfiltered_req_id)),
+    )
+    .await
+    .expect("listConversations unfiltered timeout")
+    .expect("listConversations unfiltered resp");
+    let ListConversationsResponse {
+        items: unfiltered_items,
+        next_cursor: unfiltered_next,
+    } = to_response::<ListConversationsResponse>(unfiltered_resp)
+        .expect("deserialize unfiltered response");
+    assert_eq!(unfiltered_items.len(), 4);
+    assert!(unfiltered_next.is_none());
+
+    let empty_req_id = mcp
+        .send_list_conversations_request(ListConversationsParams {
+            page_size: Some(10),
+            cursor: None,
+            model_providers: Some(vec!["other".to_string()]),
+        })
+        .await
+        .expect("send listConversations filtered empty");
+    let empty_resp: JSONRPCResponse = timeout(
+        DEFAULT_READ_TIMEOUT,
+        mcp.read_stream_until_response_message(RequestId::Integer(empty_req_id)),
+    )
+    .await
+    .expect("listConversations filtered empty timeout")
+    .expect("listConversations filtered empty resp");
+    let ListConversationsResponse {
+        items: empty_items,
+        next_cursor: empty_next,
+    } = to_response::<ListConversationsResponse>(empty_resp).expect("deserialize filtered empty");
+    assert!(empty_items.is_empty());
+    assert!(empty_next.is_none());
 
     // Now resume one of the sessions and expect a SessionConfigured notification and response.
     let resume_req_id = mcp
@@ -152,7 +240,13 @@ async fn test_list_and_resume_conversations() {
     assert!(!conversation_id.to_string().is_empty());
 }
 
-fn create_fake_rollout(codex_home: &Path, filename_ts: &str, meta_rfc3339: &str, preview: &str) {
+fn create_fake_rollout(
+    codex_home: &Path,
+    filename_ts: &str,
+    meta_rfc3339: &str,
+    preview: &str,
+    model_provider: Option<&str>,
+) {
     let uuid = Uuid::new_v4();
     // sessions/YYYY/MM/DD/ derived from filename_ts (YYYY-MM-DDThh-mm-ss)
     let year = &filename_ts[0..4];
@@ -164,18 +258,22 @@ fn create_fake_rollout(codex_home: &Path, filename_ts: &str, meta_rfc3339: &str,
     let file_path = dir.join(format!("rollout-{filename_ts}-{uuid}.jsonl"));
     let mut lines = Vec::new();
     // Meta line with timestamp (flattened meta in payload for new schema)
+    let mut payload = json!({
+        "id": uuid,
+        "timestamp": meta_rfc3339,
+        "cwd": "/",
+        "originator": "codex",
+        "cli_version": "0.0.0",
+        "instructions": null,
+    });
+    if let Some(provider) = model_provider {
+        payload["model_provider"] = json!(provider);
+    }
     lines.push(
         json!({
             "timestamp": meta_rfc3339,
             "type": "session_meta",
-            "payload": {
-                "id": uuid,
-                "timestamp": meta_rfc3339,
-                "cwd": "/",
-                "originator": "codex",
-                "cli_version": "0.0.0",
-                "instructions": null
-            }
+            "payload": payload
         })
         .to_string(),
     );

codex-rs/app-server/tests/suite/login.rs

@@ -13,6 +13,7 @@ use codex_app_server_protocol::LoginChatGptResponse;
 use codex_app_server_protocol::LogoutChatGptResponse;
 use codex_app_server_protocol::RequestId;
 use codex_login::login_with_api_key;
+use serial_test::serial;
 use tempfile::TempDir;
 use tokio::time::timeout;
 
@@ -94,6 +95,8 @@ async fn logout_chatgpt_removes_auth() {
 }
 
 #[tokio::test(flavor = "multi_thread", worker_threads = 2)]
+// Serialize tests that launch the login server since it binds to a fixed port.
+#[serial(login_port)]
 async fn login_and_cancel_chatgpt() {
     let codex_home = TempDir::new().unwrap_or_else(|e| panic!("create tempdir: {e}"));
     create_config_toml(codex_home.path()).unwrap_or_else(|err| panic!("write config.toml: {err}"));
@@ -208,6 +211,8 @@ async fn login_chatgpt_rejected_when_forced_api() {
 }
 
 #[tokio::test(flavor = "multi_thread", worker_threads = 2)]
+// Serialize tests that launch the login server since it binds to a fixed port.
+#[serial(login_port)]
 async fn login_chatgpt_includes_forced_workspace_query_param() {
     let codex_home = TempDir::new().unwrap_or_else(|e| panic!("create tempdir: {e}"));
     create_config_toml_forced_workspace(codex_home.path(), "ws-forced")

codex-rs/app-server/tests/suite/send_message.rs

@@ -15,6 +15,8 @@ use codex_app_server_protocol::RequestId;
 use codex_app_server_protocol::SendUserMessageParams;
 use codex_app_server_protocol::SendUserMessageResponse;
 use codex_protocol::ConversationId;
+use codex_protocol::models::ContentItem;
+use codex_protocol::models::ResponseItem;
 use pretty_assertions::assert_eq;
 use tempfile::TempDir;
 use tokio::time::timeout;
@@ -62,7 +64,10 @@ async fn test_send_message_success() {
 
     // 2) addConversationListener
     let add_listener_id = mcp
-        .send_add_conversation_listener_request(AddConversationListenerParams { conversation_id })
+        .send_add_conversation_listener_request(AddConversationListenerParams {
+            conversation_id,
+            experimental_raw_events: false,
+        })
         .await
         .expect("send addConversationListener");
     let add_listener_resp: JSONRPCResponse = timeout(
@@ -124,6 +129,105 @@ async fn send_message(message: &str, conversation_id: ConversationId, mcp: &mut
             .expect("should have conversationId"),
         &serde_json::Value::String(conversation_id.to_string())
     );
+
+    let raw_attempt = tokio::time::timeout(
+        std::time::Duration::from_millis(200),
+        mcp.read_stream_until_notification_message("codex/event/raw_response_item"),
+    )
+    .await;
+    assert!(
+        raw_attempt.is_err(),
+        "unexpected raw item notification when not opted in"
+    );
+}
+
+#[tokio::test]
+async fn test_send_message_raw_notifications_opt_in() {
+    let responses = vec![
+        create_final_assistant_message_sse_response("Done").expect("build mock assistant message"),
+    ];
+    let server = create_mock_chat_completions_server(responses).await;
+
+    let codex_home = TempDir::new().expect("create temp dir");
+    create_config_toml(codex_home.path(), &server.uri()).expect("write config.toml");
+
+    let mut mcp = McpProcess::new(codex_home.path())
+        .await
+        .expect("spawn mcp process");
+    timeout(DEFAULT_READ_TIMEOUT, mcp.initialize())
+        .await
+        .expect("init timed out")
+        .expect("init failed");
+
+    let new_conv_id = mcp
+        .send_new_conversation_request(NewConversationParams::default())
+        .await
+        .expect("send newConversation");
+    let new_conv_resp: JSONRPCResponse = timeout(
+        DEFAULT_READ_TIMEOUT,
+        mcp.read_stream_until_response_message(RequestId::Integer(new_conv_id)),
+    )
+    .await
+    .expect("newConversation timeout")
+    .expect("newConversation resp");
+    let NewConversationResponse {
+        conversation_id, ..
+    } = to_response::<_>(new_conv_resp).expect("deserialize newConversation response");
+
+    let add_listener_id = mcp
+        .send_add_conversation_listener_request(AddConversationListenerParams {
+            conversation_id,
+            experimental_raw_events: true,
+        })
+        .await
+        .expect("send addConversationListener");
+    let add_listener_resp: JSONRPCResponse = timeout(
+        DEFAULT_READ_TIMEOUT,
+        mcp.read_stream_until_response_message(RequestId::Integer(add_listener_id)),
+    )
+    .await
+    .expect("addConversationListener timeout")
+    .expect("addConversationListener resp");
+    let AddConversationSubscriptionResponse { subscription_id: _ } =
+        to_response::<_>(add_listener_resp).expect("deserialize addConversationListener response");
+
+    let send_id = mcp
+        .send_send_user_message_request(SendUserMessageParams {
+            conversation_id,
+            items: vec![InputItem::Text {
+                text: "Hello".to_string(),
+            }],
+        })
+        .await
+        .expect("send sendUserMessage");
+
+    let instructions = read_raw_response_item(&mut mcp, conversation_id).await;
+    assert_instructions_message(&instructions);
+
+    let environment = read_raw_response_item(&mut mcp, conversation_id).await;
+    assert_environment_message(&environment);
+
+    let response: JSONRPCResponse = timeout(
+        DEFAULT_READ_TIMEOUT,
+        mcp.read_stream_until_response_message(RequestId::Integer(send_id)),
+    )
+    .await
+    .expect("sendUserMessage response timeout")
+    .expect("sendUserMessage response error");
+    let _ok: SendUserMessageResponse = to_response::<SendUserMessageResponse>(response)
+        .expect("deserialize sendUserMessage response");
+
+    let user_message = read_raw_response_item(&mut mcp, conversation_id).await;
+    assert_user_message(&user_message, "Hello");
+
+    let assistant_message = read_raw_response_item(&mut mcp, conversation_id).await;
+    assert_assistant_message(&assistant_message, "Done");
+
+    let _ = tokio::time::timeout(
+        std::time::Duration::from_millis(250),
+        mcp.read_stream_until_notification_message("codex/event/task_complete"),
+    )
+    .await;
 }
 
 #[tokio::test]
@@ -184,3 +288,108 @@ stream_max_retries = 0
         ),
     )
 }
+
+#[expect(clippy::expect_used)]
+async fn read_raw_response_item(
+    mcp: &mut McpProcess,
+    conversation_id: ConversationId,
+) -> ResponseItem {
+    let raw_notification: JSONRPCNotification = timeout(
+        DEFAULT_READ_TIMEOUT,
+        mcp.read_stream_until_notification_message("codex/event/raw_response_item"),
+    )
+    .await
+    .expect("codex/event/raw_response_item notification timeout")
+    .expect("codex/event/raw_response_item notification resp");
+
+    let serde_json::Value::Object(params) = raw_notification
+        .params
+        .expect("codex/event/raw_response_item should have params")
+    else {
+        panic!("codex/event/raw_response_item should have params");
+    };
+
+    let conversation_id_value = params
+        .get("conversationId")
+        .and_then(|value| value.as_str())
+        .expect("raw response item should include conversationId");
+
+    assert_eq!(
+        conversation_id_value,
+        conversation_id.to_string(),
+        "raw response item conversation mismatch"
+    );
+
+    let msg_value = params
+        .get("msg")
+        .cloned()
+        .expect("raw response item should include msg payload");
+
+    serde_json::from_value(msg_value).expect("deserialize raw response item")
+}
+
+fn assert_instructions_message(item: &ResponseItem) {
+    match item {
+        ResponseItem::Message { role, content, .. } => {
+            assert_eq!(role, "user");
+            let texts = content_texts(content);
+            assert!(
+                texts
+                    .iter()
+                    .any(|text| text.contains("<user_instructions>")),
+                "expected instructions message, got {texts:?}"
+            );
+        }
+        other => panic!("expected instructions message, got {other:?}"),
+    }
+}
+
+fn assert_environment_message(item: &ResponseItem) {
+    match item {
+        ResponseItem::Message { role, content, .. } => {
+            assert_eq!(role, "user");
+            let texts = content_texts(content);
+            assert!(
+                texts
+                    .iter()
+                    .any(|text| text.contains("<environment_context>")),
+                "expected environment context message, got {texts:?}"
+            );
+        }
+        other => panic!("expected environment message, got {other:?}"),
+    }
+}
+
+fn assert_user_message(item: &ResponseItem, expected_text: &str) {
+    match item {
+        ResponseItem::Message { role, content, .. } => {
+            assert_eq!(role, "user");
+            let texts = content_texts(content);
+            assert_eq!(texts, vec![expected_text]);
+        }
+        other => panic!("expected user message, got {other:?}"),
+    }
+}
+
+fn assert_assistant_message(item: &ResponseItem, expected_text: &str) {
+    match item {
+        ResponseItem::Message { role, content, .. } => {
+            assert_eq!(role, "assistant");
+            let texts = content_texts(content);
+            assert_eq!(texts, vec![expected_text]);
+        }
+        other => panic!("expected assistant message, got {other:?}"),
+    }
+}
+
+fn content_texts(content: &[ContentItem]) -> Vec<&str> {
+    content
+        .iter()
+        .filter_map(|item| match item {
+            ContentItem::InputText { text } | ContentItem::OutputText { text } => {
+                Some(text.as_str())
+            }
+            _ => None,
+        })
+        .collect()
+}

codex-rs/chatgpt/src/chatgpt_token.rs

@@ -19,7 +19,7 @@ pub fn set_chatgpt_token_data(value: TokenData) {
 
 /// Initialize the ChatGPT token from auth.json file
 pub async fn init_chatgpt_token_from_auth(codex_home: &Path) -> std::io::Result<()> {
-    let auth = CodexAuth::from_codex_home(codex_home)?;
+    let auth = CodexAuth::from_auth_storage(codex_home)?;
     if let Some(auth) = auth {
         let token_data = auth.get_token_data().await?;
         set_chatgpt_token_data(token_data);

codex-rs/cli/src/login.rs

@@ -140,7 +140,7 @@ pub async fn run_login_with_device_code(
 pub async fn run_login_status(cli_config_overrides: CliConfigOverrides) -> ! {
     let config = load_config_or_exit(cli_config_overrides).await;
 
-    match CodexAuth::from_codex_home(&config.codex_home) {
+    match CodexAuth::from_auth_storage(&config.codex_home) {
         Ok(Some(auth)) => match auth.mode {
             AuthMode::ApiKey => match auth.get_token().await {
                 Ok(api_key) => {

codex-rs/cli/src/main.rs

@@ -29,6 +29,7 @@ mod mcp_cmd;
 use crate::mcp_cmd::McpCli;
 use codex_core::config::Config;
 use codex_core::config::ConfigOverrides;
+use codex_core::features::is_known_feature_key;
 
 /// Codex CLI
 ///
@@ -286,15 +287,25 @@ struct FeatureToggles {
 }
 
 impl FeatureToggles {
-    fn to_overrides(&self) -> Vec<String> {
+    fn to_overrides(&self) -> anyhow::Result<Vec<String>> {
         let mut v = Vec::new();
-        for k in &self.enable {
-            v.push(format!("features.{k}=true"));
+        for feature in &self.enable {
+            Self::validate_feature(feature)?;
+            v.push(format!("features.{feature}=true"));
         }
-        for k in &self.disable {
-            v.push(format!("features.{k}=false"));
+        for feature in &self.disable {
+            Self::validate_feature(feature)?;
+            v.push(format!("features.{feature}=false"));
+        }
+        Ok(v)
+    }
+
+    fn validate_feature(feature: &str) -> anyhow::Result<()> {
+        if is_known_feature_key(feature) {
+            Ok(())
+        } else {
+            anyhow::bail!("Unknown feature flag: {feature}")
         }
-        v
     }
 }
 
@@ -345,9 +356,8 @@ async fn cli_main(codex_linux_sandbox_exe: Option<PathBuf>) -> anyhow::Result<()
     } = MultitoolCli::parse();
 
     // Fold --enable/--disable into config overrides so they flow to all subcommands.
-    root_config_overrides
-        .raw_overrides
-        .extend(feature_toggles.to_overrides());
+    let toggle_overrides = feature_toggles.to_overrides()?;
+    root_config_overrides.raw_overrides.extend(toggle_overrides);
 
     match subcommand {
         None => {
@@ -605,6 +615,7 @@ mod tests {
     use assert_matches::assert_matches;
     use codex_core::protocol::TokenUsage;
     use codex_protocol::ConversationId;
+    use pretty_assertions::assert_eq;
 
     fn finalize_from_args(args: &[&str]) -> TuiCli {
         let cli = MultitoolCli::try_parse_from(args).expect("parse");
@@ -781,4 +792,32 @@ mod tests {
         assert!(!interactive.resume_last);
         assert_eq!(interactive.resume_session_id, None);
     }
+
+    #[test]
+    fn feature_toggles_known_features_generate_overrides() {
+        let toggles = FeatureToggles {
+            enable: vec!["web_search_request".to_string()],
+            disable: vec!["unified_exec".to_string()],
+        };
+        let overrides = toggles.to_overrides().expect("valid features");
+        assert_eq!(
+            overrides,
+            vec![
+                "features.web_search_request=true".to_string(),
+                "features.unified_exec=false".to_string(),
+            ]
+        );
+    }
+
+    #[test]
+    fn feature_toggles_unknown_feature_errors() {
+        let toggles = FeatureToggles {
+            enable: vec!["does_not_exist".to_string()],
+            disable: Vec::new(),
+        };
+        let err = toggles
+            .to_overrides()
+            .expect_err("feature should be rejected");
+        assert_eq!(err.to_string(), "Unknown feature flag: does_not_exist");
+    }
 }

codex-rs/cli/src/mcp_cmd.rs

@@ -274,19 +274,33 @@ async fn run_add(config_overrides: &CliConfigOverrides, add_args: AddArgs) -> Re
         http_headers,
         env_http_headers,
     } = transport
-        && matches!(supports_oauth_login(&url).await, Ok(true))
     {
-        println!("Detected OAuth support. Starting OAuth flow…");
-        perform_oauth_login(
-            &name,
-            &url,
-            config.mcp_oauth_credentials_store_mode,
-            http_headers.clone(),
-            env_http_headers.clone(),
-            &Vec::new(),
-        )
-        .await?;
-        println!("Successfully logged in.");
+        match supports_oauth_login(&url).await {
+            Ok(true) => {
+                if !config.features.enabled(Feature::RmcpClient) {
+                    println!(
+                        "MCP server supports login. Add `experimental_use_rmcp_client = true` \
+                         to your config.toml and run `codex mcp login {name}` to login."
+                    );
+                } else {
+                    println!("Detected OAuth support. Starting OAuth flow…");
+                    perform_oauth_login(
+                        &name,
+                        &url,
+                        config.mcp_oauth_credentials_store_mode,
+                        http_headers.clone(),
+                        env_http_headers.clone(),
+                        &Vec::new(),
+                    )
+                    .await?;
+                    println!("Successfully logged in.");
+                }
+            }
+            Ok(false) => {}
+            Err(_) => println!(
+                "MCP server may or may not require login. Run `codex mcp login {name}` to login."
+            ),
+        }
     }
 
     Ok(())
@@ -523,10 +537,12 @@ async fn run_list(config_overrides: &CliConfigOverrides, list_args: ListArgs) ->
                     .map(|entry| entry.auth_status)
                     .unwrap_or(McpAuthStatus::Unsupported)
                     .to_string();
+                let bearer_token_display =
+                    bearer_token_env_var.as_deref().unwrap_or("-").to_string();
                 http_rows.push([
                     name.clone(),
                     url.clone(),
-                    bearer_token_env_var.clone().unwrap_or("-".to_string()),
+                    bearer_token_display,
                     status,
                     auth_status,
                 ]);
@@ -752,15 +768,15 @@ async fn run_get(config_overrides: &CliConfigOverrides, get_args: GetArgs) -> Re
         } => {
             println!("  transport: streamable_http");
             println!("  url: {url}");
-            let env_var = bearer_token_env_var.as_deref().unwrap_or("-");
-            println!("  bearer_token_env_var: {env_var}");
+            let bearer_token_display = bearer_token_env_var.as_deref().unwrap_or("-");
+            println!("  bearer_token_env_var: {bearer_token_display}");
             let headers_display = match http_headers {
                 Some(map) if !map.is_empty() => {
                     let mut pairs: Vec<_> = map.iter().collect();
                     pairs.sort_by(|(a, _), (b, _)| a.cmp(b));
                     pairs
                         .into_iter()
-                        .map(|(k, v)| format!("{k}={v}"))
+                        .map(|(k, _)| format!("{k}=*****"))
                         .collect::<Vec<_>>()
                         .join(", ")
                 }
@@ -773,7 +789,7 @@ async fn run_get(config_overrides: &CliConfigOverrides, get_args: GetArgs) -> Re
                     pairs.sort_by(|(a, _), (b, _)| a.cmp(b));
                     pairs
                         .into_iter()
-                        .map(|(k, v)| format!("{k}={v}"))
+                        .map(|(k, var)| format!("{k}={var}"))
                         .collect::<Vec<_>>()
                         .join(", ")
                 }

codex-rs/cli/tests/mcp_list.rs

@@ -68,9 +68,9 @@ async fn list_and_get_render_expected_output() -> Result<()> {
     assert!(stdout.contains("Name"));
     assert!(stdout.contains("docs"));
     assert!(stdout.contains("docs-server"));
-    assert!(stdout.contains("TOKEN=secret"));
-    assert!(stdout.contains("APP_TOKEN=$APP_TOKEN"));
-    assert!(stdout.contains("WORKSPACE_ID=$WORKSPACE_ID"));
+    assert!(stdout.contains("TOKEN=*****"));
+    assert!(stdout.contains("APP_TOKEN=*****"));
+    assert!(stdout.contains("WORKSPACE_ID=*****"));
     assert!(stdout.contains("Status"));
     assert!(stdout.contains("Auth"));
     assert!(stdout.contains("enabled"));
@@ -119,9 +119,9 @@ async fn list_and_get_render_expected_output() -> Result<()> {
     assert!(stdout.contains("transport: stdio"));
     assert!(stdout.contains("command: docs-server"));
     assert!(stdout.contains("args: --port 4000"));
-    assert!(stdout.contains("env: TOKEN=secret"));
-    assert!(stdout.contains("APP_TOKEN=$APP_TOKEN"));
-    assert!(stdout.contains("WORKSPACE_ID=$WORKSPACE_ID"));
+    assert!(stdout.contains("env: TOKEN=*****"));
+    assert!(stdout.contains("APP_TOKEN=*****"));
+    assert!(stdout.contains("WORKSPACE_ID=*****"));
     assert!(stdout.contains("enabled: true"));
     assert!(stdout.contains("remove: codex mcp remove docs"));
 

codex-rs/common/src/format_env_display.rs

@@ -6,15 +6,11 @@ pub fn format_env_display(env: Option<&HashMap<String, String>>, env_vars: &[Str
     if let Some(map) = env {
         let mut pairs: Vec<_> = map.iter().collect();
         pairs.sort_by(|(a, _), (b, _)| a.cmp(b));
-        parts.extend(
-            pairs
-                .into_iter()
-                .map(|(key, value)| format!("{key}={value}")),
-        );
+        parts.extend(pairs.into_iter().map(|(key, _)| format!("{key}=*****")));
     }
 
     if !env_vars.is_empty() {
-        parts.extend(env_vars.iter().map(|var| format!("{var}=${var}")));
+        parts.extend(env_vars.iter().map(|var| format!("{var}=*****")));
     }
 
     if parts.is_empty() {
@@ -42,14 +38,14 @@ mod tests {
         env.insert("B".to_string(), "two".to_string());
         env.insert("A".to_string(), "one".to_string());
 
-        assert_eq!(format_env_display(Some(&env), &[]), "A=one, B=two");
+        assert_eq!(format_env_display(Some(&env), &[]), "A=*****, B=*****");
     }
 
     #[test]
     fn formats_env_vars_with_dollar_prefix() {
         let vars = vec!["TOKEN".to_string(), "PATH".to_string()];
 
-        assert_eq!(format_env_display(None, &vars), "TOKEN=$TOKEN, PATH=$PATH");
+        assert_eq!(format_env_display(None, &vars), "TOKEN=*****, PATH=*****");
     }
 
     #[test]
@@ -60,7 +56,7 @@ mod tests {
 
         assert_eq!(
             format_env_display(Some(&env), &vars),
-            "HOME=/tmp, TOKEN=$TOKEN"
+            "HOME=*****, TOKEN=*****"
         );
     }
 }

codex-rs/core/Cargo.toml

@@ -21,13 +21,16 @@ bytes = { workspace = true }
 chrono = { workspace = true, features = ["serde"] }
 codex-app-server-protocol = { workspace = true }
 codex-apply-patch = { workspace = true }
+codex-async-utils = { workspace = true }
 codex-file-search = { workspace = true }
+codex-git-tooling = { workspace = true }
+codex-keyring-store = { workspace = true }
 codex-otel = { workspace = true, features = ["otel"] }
 codex-protocol = { workspace = true }
 codex-rmcp-client = { workspace = true }
-codex-async-utils = { workspace = true }
-codex-utils-string = { workspace = true }
 codex-utils-pty = { workspace = true }
+codex-utils-readiness = { workspace = true }
+codex-utils-string = { workspace = true }
 codex-utils-tokenizer = { workspace = true }
 dirs = { workspace = true }
 dunce = { workspace = true }
@@ -36,6 +39,7 @@ eventsource-stream = { workspace = true }
 futures = { workspace = true }
 http = { workspace = true }
 indexmap = { workspace = true }
+keyring = { workspace = true }
 libc = { workspace = true }
 mcp-types = { workspace = true }
 os_info = { workspace = true }
@@ -45,6 +49,7 @@ reqwest = { workspace = true, features = ["json", "stream"] }
 serde = { workspace = true, features = ["derive"] }
 serde_json = { workspace = true }
 sha1 = { workspace = true }
+sha2 = { workspace = true }
 shlex = { workspace = true }
 similar = { workspace = true }
 strum_macros = { workspace = true }
@@ -95,6 +100,7 @@ assert_cmd = { workspace = true }
 assert_matches = { workspace = true }
 core_test_support = { workspace = true }
 escargot = { workspace = true }
+image = { workspace = true, features = ["jpeg", "png"] }
 maplit = { workspace = true }
 predicates = { workspace = true }
 pretty_assertions = { workspace = true }

codex-rs/core/src/auth.rs

@@ -1,16 +1,12 @@
-use chrono::DateTime;
+mod storage;
+
 use chrono::Utc;
 use serde::Deserialize;
 use serde::Serialize;
 #[cfg(test)]
 use serial_test::serial;
 use std::env;
-use std::fs::File;
-use std::fs::OpenOptions;
-use std::io::Read;
-use std::io::Write;
-#[cfg(unix)]
-use std::os::unix::fs::OpenOptionsExt;
+use std::fmt::Debug;
 use std::path::Path;
 use std::path::PathBuf;
 use std::sync::Arc;
@@ -20,6 +16,10 @@ use std::time::Duration;
 use codex_app_server_protocol::AuthMode;
 use codex_protocol::config_types::ForcedLoginMethod;
 
+pub use crate::auth::storage::AuthCredentialsStoreMode;
+pub use crate::auth::storage::AuthDotJson;
+use crate::auth::storage::AuthStorageBackend;
+use crate::auth::storage::create_auth_storage;
 use crate::config::Config;
 use crate::default_client::CodexHttpClient;
 use crate::token_data::PlanType;
@@ -32,7 +32,7 @@ pub struct CodexAuth {
 
     pub(crate) api_key: Option<String>,
     pub(crate) auth_dot_json: Arc<Mutex<Option<AuthDotJson>>>,
-    pub(crate) auth_file: PathBuf,
+    storage: Arc<dyn AuthStorageBackend>,
     pub(crate) client: CodexHttpClient,
 }
 
@@ -56,7 +56,7 @@ impl CodexAuth {
             .map_err(std::io::Error::other)?;
 
         let updated = update_tokens(
-            &self.auth_file,
+            &self.storage,
             refresh_response.id_token,
             refresh_response.access_token,
             refresh_response.refresh_token,
@@ -78,8 +78,8 @@ impl CodexAuth {
         Ok(access)
     }
 
-    /// Loads the available auth information from the auth.json.
-    pub fn from_codex_home(codex_home: &Path) -> std::io::Result<Option<CodexAuth>> {
+    /// Loads the available auth information from auth storage.
+    pub fn from_auth_storage(codex_home: &Path) -> std::io::Result<Option<CodexAuth>> {
         load_auth(codex_home, false)
     }
 
@@ -103,7 +103,7 @@ impl CodexAuth {
                     .map_err(std::io::Error::other)?;
 
                     let updated_auth_dot_json = update_tokens(
-                        &self.auth_file,
+                        &self.storage,
                         refresh_response.id_token,
                         refresh_response.access_token,
                         refresh_response.refresh_token,
@@ -177,7 +177,7 @@ impl CodexAuth {
         Self {
             api_key: None,
             mode: AuthMode::ChatGPT,
-            auth_file: PathBuf::new(),
+            storage: create_auth_storage(PathBuf::new(), AuthCredentialsStoreMode::File),
             auth_dot_json,
             client: crate::default_client::create_client(),
         }
@@ -187,7 +187,7 @@ impl CodexAuth {
         Self {
             api_key: Some(api_key.to_owned()),
             mode: AuthMode::ApiKey,
-            auth_file: PathBuf::new(),
+            storage: create_auth_storage(PathBuf::new(), AuthCredentialsStoreMode::File),
             auth_dot_json: Arc::new(Mutex::new(None)),
             client,
         }
@@ -215,19 +215,11 @@ pub fn read_codex_api_key_from_env() -> Option<String> {
         .filter(|value| !value.is_empty())
 }
 
-pub fn get_auth_file(codex_home: &Path) -> PathBuf {
-    codex_home.join("auth.json")
-}
-
 /// Delete the auth.json file inside `codex_home` if it exists. Returns `Ok(true)`
 /// if a file was removed, `Ok(false)` if no auth file was present.
 pub fn logout(codex_home: &Path) -> std::io::Result<bool> {
-    let auth_file = get_auth_file(codex_home);
-    match std::fs::remove_file(&auth_file) {
-        Ok(_) => Ok(true),
-        Err(err) if err.kind() == std::io::ErrorKind::NotFound => Ok(false),
-        Err(err) => Err(err),
-    }
+    let storage = create_auth_storage(codex_home.to_path_buf(), AuthCredentialsStoreMode::File);
+    storage.delete()
 }
 
 /// Writes an `auth.json` that contains only the API key.
@@ -237,7 +229,20 @@ pub fn login_with_api_key(codex_home: &Path, api_key: &str) -> std::io::Result<(
         tokens: None,
         last_refresh: None,
     };
-    write_auth_json(&get_auth_file(codex_home), &auth_dot_json)
+    save_auth(codex_home, &auth_dot_json)
+}
+
+/// Persist the provided auth payload using the specified backend.
+pub fn save_auth(codex_home: &Path, auth: &AuthDotJson) -> std::io::Result<()> {
+    let storage = create_auth_storage(codex_home.to_path_buf(), AuthCredentialsStoreMode::File);
+    storage.save(auth)
+}
+
+/// Load CLI auth data using the configured credential store backend.
+/// Returns `None` when no credentials are stored.
+pub fn load_auth_dot_json(codex_home: &Path) -> std::io::Result<Option<AuthDotJson>> {
+    let storage = create_auth_storage(codex_home.to_path_buf(), AuthCredentialsStoreMode::File);
+    storage.load()
 }
 
 pub async fn enforce_login_restrictions(config: &Config) -> std::io::Result<()> {
@@ -320,12 +325,12 @@ fn load_auth(
         )));
     }
 
-    let auth_file = get_auth_file(codex_home);
+    let storage = create_auth_storage(codex_home.to_path_buf(), AuthCredentialsStoreMode::File);
+
     let client = crate::default_client::create_client();
-    let auth_dot_json = match try_read_auth_json(&auth_file) {
-        Ok(auth) => auth,
-        Err(err) if err.kind() == std::io::ErrorKind::NotFound => return Ok(None),
-        Err(err) => return Err(err),
+    let auth_dot_json = match storage.load()? {
+        Some(auth) => auth,
+        None => return Ok(None),
     };
 
     let AuthDotJson {
@@ -342,7 +347,7 @@ fn load_auth(
     Ok(Some(CodexAuth {
         api_key: None,
         mode: AuthMode::ChatGPT,
-        auth_file,
+        storage: storage.clone(),
         auth_dot_json: Arc::new(Mutex::new(Some(AuthDotJson {
             openai_api_key: None,
             tokens,
@@ -352,44 +357,20 @@ fn load_auth(
     }))
 }
 
-/// Attempt to read and refresh the `auth.json` file in the given `CODEX_HOME` directory.
-/// Returns the full AuthDotJson structure after refreshing if necessary.
-pub fn try_read_auth_json(auth_file: &Path) -> std::io::Result<AuthDotJson> {
-    let mut file = File::open(auth_file)?;
-    let mut contents = String::new();
-    file.read_to_string(&mut contents)?;
-    let auth_dot_json: AuthDotJson = serde_json::from_str(&contents)?;
-
-    Ok(auth_dot_json)
-}
-
-pub fn write_auth_json(auth_file: &Path, auth_dot_json: &AuthDotJson) -> std::io::Result<()> {
-    if let Some(parent) = auth_file.parent() {
-        std::fs::create_dir_all(parent)?;
-    }
-    let json_data = serde_json::to_string_pretty(auth_dot_json)?;
-    let mut options = OpenOptions::new();
-    options.truncate(true).write(true).create(true);
-    #[cfg(unix)]
-    {
-        options.mode(0o600);
-    }
-    let mut file = options.open(auth_file)?;
-    file.write_all(json_data.as_bytes())?;
-    file.flush()?;
-    Ok(())
-}
-
 async fn update_tokens(
-    auth_file: &Path,
-    id_token: String,
+    storage: &Arc<dyn AuthStorageBackend>,
+    id_token: Option<String>,
     access_token: Option<String>,
     refresh_token: Option<String>,
 ) -> std::io::Result<AuthDotJson> {
-    let mut auth_dot_json = try_read_auth_json(auth_file)?;
+    let mut auth_dot_json = storage
+        .load()?
+        .ok_or(std::io::Error::other("Token data is not available."))?;
 
     let tokens = auth_dot_json.tokens.get_or_insert_with(TokenData::default);
-    tokens.id_token = parse_id_token(&id_token).map_err(std::io::Error::other)?;
+    if let Some(id_token) = id_token {
+        tokens.id_token = parse_id_token(&id_token).map_err(std::io::Error::other)?;
+    }
     if let Some(access_token) = access_token {
         tokens.access_token = access_token;
     }
@@ -397,7 +378,7 @@ async fn update_tokens(
         tokens.refresh_token = refresh_token;
     }
     auth_dot_json.last_refresh = Some(Utc::now());
-    write_auth_json(auth_file, &auth_dot_json)?;
+    storage.save(&auth_dot_json)?;
     Ok(auth_dot_json)
 }
 
@@ -445,24 +426,11 @@ struct RefreshRequest {
 
 #[derive(Deserialize, Clone)]
 struct RefreshResponse {
-    id_token: String,
+    id_token: Option<String>,
     access_token: Option<String>,
     refresh_token: Option<String>,
 }
 
-/// Expected structure for $CODEX_HOME/auth.json.
-#[derive(Deserialize, Serialize, Clone, Debug, PartialEq)]
-pub struct AuthDotJson {
-    #[serde(rename = "OPENAI_API_KEY")]
-    pub openai_api_key: Option<String>,
-
-    #[serde(default, skip_serializing_if = "Option::is_none")]
-    pub tokens: Option<TokenData>,
-
-    #[serde(default, skip_serializing_if = "Option::is_none")]
-    pub last_refresh: Option<DateTime<Utc>>,
-}
-
 // Shared constant for token refresh (client id used for oauth token refresh flow)
 pub const CLIENT_ID: &str = "app_EMoamEEZ73f0CkXaXp7hrann";
 
@@ -477,12 +445,15 @@ struct CachedAuth {
 #[cfg(test)]
 mod tests {
     use super::*;
+    use crate::auth::storage::FileAuthStorage;
+    use crate::auth::storage::get_auth_file;
     use crate::config::Config;
     use crate::config::ConfigOverrides;
     use crate::config::ConfigToml;
     use crate::token_data::IdTokenInfo;
     use crate::token_data::KnownPlan;
     use crate::token_data::PlanType;
+
     use base64::Engine;
     use codex_protocol::config_types::ForcedLoginMethod;
     use pretty_assertions::assert_eq;
@@ -491,9 +462,9 @@ mod tests {
     use tempfile::tempdir;
 
     #[tokio::test]
-    async fn roundtrip_auth_dot_json() {
+    async fn refresh_without_id_token() {
         let codex_home = tempdir().unwrap();
-        let _ = write_auth_file(
+        let fake_jwt = write_auth_file(
             AuthFileParams {
                 openai_api_key: None,
                 chatgpt_plan_type: "pro".to_string(),
@@ -503,12 +474,23 @@ mod tests {
         )
         .expect("failed to write auth file");
 
-        let file = get_auth_file(codex_home.path());
-        let auth_dot_json = try_read_auth_json(&file).unwrap();
-        write_auth_json(&file, &auth_dot_json).unwrap();
+        let storage = create_auth_storage(
+            codex_home.path().to_path_buf(),
+            AuthCredentialsStoreMode::File,
+        );
+        let updated = super::update_tokens(
+            &storage,
+            None,
+            Some("new-access-token".to_string()),
+            Some("new-refresh-token".to_string()),
+        )
+        .await
+        .expect("update_tokens should succeed");
 
-        let same_auth_dot_json = try_read_auth_json(&file).unwrap();
-        assert_eq!(auth_dot_json, same_auth_dot_json);
+        let tokens = updated.tokens.expect("tokens should exist");
+        assert_eq!(tokens.id_token.raw_jwt, fake_jwt);
+        assert_eq!(tokens.access_token, "new-access-token");
+        assert_eq!(tokens.refresh_token, "new-refresh-token");
     }
 
     #[test]
@@ -532,7 +514,10 @@ mod tests {
 
         super::login_with_api_key(dir.path(), "sk-new").expect("login_with_api_key should succeed");
 
-        let auth = super::try_read_auth_json(&auth_path).expect("auth.json should parse");
+        let storage = FileAuthStorage::new(dir.path().to_path_buf());
+        let auth = storage
+            .try_read_auth_json(&auth_path)
+            .expect("auth.json should parse");
         assert_eq!(auth.openai_api_key.as_deref(), Some("sk-new"));
         assert!(auth.tokens.is_none(), "tokens should be cleared");
     }
@@ -540,7 +525,7 @@ mod tests {
     #[test]
     fn missing_auth_json_returns_none() {
         let dir = tempdir().unwrap();
-        let auth = CodexAuth::from_codex_home(dir.path()).expect("call should succeed");
+        let auth = CodexAuth::from_auth_storage(dir.path()).expect("call should succeed");
         assert_eq!(auth, None);
     }
 
@@ -562,7 +547,7 @@ mod tests {
             api_key,
             mode,
             auth_dot_json,
-            auth_file: _,
+            storage: _,
             ..
         } = super::load_auth(codex_home.path(), false).unwrap().unwrap();
         assert_eq!(None, api_key);
@@ -620,11 +605,11 @@ mod tests {
             tokens: None,
             last_refresh: None,
         };
-        write_auth_json(&get_auth_file(dir.path()), &auth_dot_json)?;
-        assert!(dir.path().join("auth.json").exists());
-        let removed = logout(dir.path())?;
-        assert!(removed);
-        assert!(!dir.path().join("auth.json").exists());
+        super::save_auth(dir.path(), &auth_dot_json)?;
+        let auth_file = get_auth_file(dir.path());
+        assert!(auth_file.exists());
+        assert!(logout(dir.path())?);
+        assert!(!auth_file.exists());
         Ok(())
     }
 

codex-rs/core/src/auth/storage.rs

@@ -0,0 +1,672 @@
+use chrono::DateTime;
+use chrono::Utc;
+use serde::Deserialize;
+use serde::Serialize;
+use sha2::Digest;
+use sha2::Sha256;
+use std::fmt::Debug;
+use std::fs::File;
+use std::fs::OpenOptions;
+use std::io::Read;
+use std::io::Write;
+#[cfg(unix)]
+use std::os::unix::fs::OpenOptionsExt;
+use std::path::Path;
+use std::path::PathBuf;
+use std::sync::Arc;
+use tracing::warn;
+
+use crate::token_data::TokenData;
+use codex_keyring_store::DefaultKeyringStore;
+use codex_keyring_store::KeyringStore;
+
+/// Determine where Codex should store CLI auth credentials.
+#[derive(Debug, Default, Copy, Clone, PartialEq, Eq, Serialize, Deserialize)]
+#[serde(rename_all = "lowercase")]
+pub enum AuthCredentialsStoreMode {
+    #[default]
+    /// Persist credentials in CODEX_HOME/auth.json.
+    File,
+    /// Persist credentials in the keyring. Fail if unavailable.
+    Keyring,
+    /// Use keyring when available; otherwise, fall back to a file in CODEX_HOME.
+    Auto,
+}
+
+/// Expected structure for $CODEX_HOME/auth.json.
+#[derive(Deserialize, Serialize, Clone, Debug, PartialEq)]
+pub struct AuthDotJson {
+    #[serde(rename = "OPENAI_API_KEY")]
+    pub openai_api_key: Option<String>,
+
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub tokens: Option<TokenData>,
+
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub last_refresh: Option<DateTime<Utc>>,
+}
+
+pub(super) fn get_auth_file(codex_home: &Path) -> PathBuf {
+    codex_home.join("auth.json")
+}
+
+pub(super) fn delete_file_if_exists(codex_home: &Path) -> std::io::Result<bool> {
+    let auth_file = get_auth_file(codex_home);
+    match std::fs::remove_file(&auth_file) {
+        Ok(()) => Ok(true),
+        Err(err) if err.kind() == std::io::ErrorKind::NotFound => Ok(false),
+        Err(err) => Err(err),
+    }
+}
+
+pub(super) trait AuthStorageBackend: Debug + Send + Sync {
+    fn load(&self) -> std::io::Result<Option<AuthDotJson>>;
+    fn save(&self, auth: &AuthDotJson) -> std::io::Result<()>;
+    fn delete(&self) -> std::io::Result<bool>;
+}
+
+#[derive(Clone, Debug)]
+pub(super) struct FileAuthStorage {
+    codex_home: PathBuf,
+}
+
+impl FileAuthStorage {
+    pub(super) fn new(codex_home: PathBuf) -> Self {
+        Self { codex_home }
+    }
+
+    /// Attempt to read and refresh the `auth.json` file in the given `CODEX_HOME` directory.
+    /// Returns the full AuthDotJson structure after refreshing if necessary.
+    pub(super) fn try_read_auth_json(&self, auth_file: &Path) -> std::io::Result<AuthDotJson> {
+        let mut file = File::open(auth_file)?;
+        let mut contents = String::new();
+        file.read_to_string(&mut contents)?;
+        let auth_dot_json: AuthDotJson = serde_json::from_str(&contents)?;
+
+        Ok(auth_dot_json)
+    }
+}
+
+impl AuthStorageBackend for FileAuthStorage {
+    fn load(&self) -> std::io::Result<Option<AuthDotJson>> {
+        let auth_file = get_auth_file(&self.codex_home);
+        let auth_dot_json = match self.try_read_auth_json(&auth_file) {
+            Ok(auth) => auth,
+            Err(err) if err.kind() == std::io::ErrorKind::NotFound => return Ok(None),
+            Err(err) => return Err(err),
+        };
+        Ok(Some(auth_dot_json))
+    }
+
+    fn save(&self, auth_dot_json: &AuthDotJson) -> std::io::Result<()> {
+        let auth_file = get_auth_file(&self.codex_home);
+
+        if let Some(parent) = auth_file.parent() {
+            std::fs::create_dir_all(parent)?;
+        }
+        let json_data = serde_json::to_string_pretty(auth_dot_json)?;
+        let mut options = OpenOptions::new();
+        options.truncate(true).write(true).create(true);
+        #[cfg(unix)]
+        {
+            options.mode(0o600);
+        }
+        let mut file = options.open(auth_file)?;
+        file.write_all(json_data.as_bytes())?;
+        file.flush()?;
+        Ok(())
+    }
+
+    fn delete(&self) -> std::io::Result<bool> {
+        delete_file_if_exists(&self.codex_home)
+    }
+}
+
+const KEYRING_SERVICE: &str = "Codex Auth";
+
+// turns codex_home path into a stable, short key string
+fn compute_store_key(codex_home: &Path) -> std::io::Result<String> {
+    let canonical = codex_home
+        .canonicalize()
+        .unwrap_or_else(|_| codex_home.to_path_buf());
+    let path_str = canonical.to_string_lossy();
+    let mut hasher = Sha256::new();
+    hasher.update(path_str.as_bytes());
+    let digest = hasher.finalize();
+    let hex = format!("{digest:x}");
+    let truncated = hex.get(..16).unwrap_or(&hex);
+    Ok(format!("cli|{truncated}"))
+}
+
+#[derive(Clone, Debug)]
+struct KeyringAuthStorage {
+    codex_home: PathBuf,
+    keyring_store: Arc<dyn KeyringStore>,
+}
+
+impl KeyringAuthStorage {
+    fn new(codex_home: PathBuf, keyring_store: Arc<dyn KeyringStore>) -> Self {
+        Self {
+            codex_home,
+            keyring_store,
+        }
+    }
+
+    fn load_from_keyring(&self, key: &str) -> std::io::Result<Option<AuthDotJson>> {
+        match self.keyring_store.load(KEYRING_SERVICE, key) {
+            Ok(Some(serialized)) => serde_json::from_str(&serialized).map(Some).map_err(|err| {
+                std::io::Error::other(format!(
+                    "failed to deserialize CLI auth from keyring: {err}"
+                ))
+            }),
+            Ok(None) => Ok(None),
+            Err(error) => Err(std::io::Error::other(format!(
+                "failed to load CLI auth from keyring: {}",
+                error.message()
+            ))),
+        }
+    }
+
+    fn save_to_keyring(&self, key: &str, value: &str) -> std::io::Result<()> {
+        match self.keyring_store.save(KEYRING_SERVICE, key, value) {
+            Ok(()) => Ok(()),
+            Err(error) => {
+                let message = format!(
+                    "failed to write OAuth tokens to keyring: {}",
+                    error.message()
+                );
+                warn!("{message}");
+                Err(std::io::Error::other(message))
+            }
+        }
+    }
+}
+
+impl AuthStorageBackend for KeyringAuthStorage {
+    fn load(&self) -> std::io::Result<Option<AuthDotJson>> {
+        let key = compute_store_key(&self.codex_home)?;
+        self.load_from_keyring(&key)
+    }
+
+    fn save(&self, auth: &AuthDotJson) -> std::io::Result<()> {
+        let key = compute_store_key(&self.codex_home)?;
+        // Simpler error mapping per style: prefer method reference over closure
+        let serialized = serde_json::to_string(auth).map_err(std::io::Error::other)?;
+        self.save_to_keyring(&key, &serialized)?;
+        if let Err(err) = delete_file_if_exists(&self.codex_home) {
+            warn!("failed to remove CLI auth fallback file: {err}");
+        }
+        Ok(())
+    }
+
+    fn delete(&self) -> std::io::Result<bool> {
+        let key = compute_store_key(&self.codex_home)?;
+        let keyring_removed = self
+            .keyring_store
+            .delete(KEYRING_SERVICE, &key)
+            .map_err(|err| {
+                std::io::Error::other(format!("failed to delete auth from keyring: {err}"))
+            })?;
+        let file_removed = delete_file_if_exists(&self.codex_home)?;
+        Ok(keyring_removed || file_removed)
+    }
+}
+
+#[derive(Clone, Debug)]
+struct AutoAuthStorage {
+    keyring_storage: Arc<KeyringAuthStorage>,
+    file_storage: Arc<FileAuthStorage>,
+}
+
+impl AutoAuthStorage {
+    fn new(codex_home: PathBuf, keyring_store: Arc<dyn KeyringStore>) -> Self {
+        Self {
+            keyring_storage: Arc::new(KeyringAuthStorage::new(codex_home.clone(), keyring_store)),
+            file_storage: Arc::new(FileAuthStorage::new(codex_home)),
+        }
+    }
+}
+
+impl AuthStorageBackend for AutoAuthStorage {
+    fn load(&self) -> std::io::Result<Option<AuthDotJson>> {
+        match self.keyring_storage.load() {
+            Ok(Some(auth)) => Ok(Some(auth)),
+            Ok(None) => self.file_storage.load(),
+            Err(err) => {
+                warn!("failed to load CLI auth from keyring, falling back to file storage: {err}");
+                self.file_storage.load()
+            }
+        }
+    }
+
+    fn save(&self, auth: &AuthDotJson) -> std::io::Result<()> {
+        match self.keyring_storage.save(auth) {
+            Ok(()) => Ok(()),
+            Err(err) => {
+                warn!("failed to save auth to keyring, falling back to file storage: {err}");
+                self.file_storage.save(auth)
+            }
+        }
+    }
+
+    fn delete(&self) -> std::io::Result<bool> {
+        // Keyring storage will delete from disk as well
+        self.keyring_storage.delete()
+    }
+}
+
+pub(super) fn create_auth_storage(
+    codex_home: PathBuf,
+    mode: AuthCredentialsStoreMode,
+) -> Arc<dyn AuthStorageBackend> {
+    let keyring_store: Arc<dyn KeyringStore> = Arc::new(DefaultKeyringStore);
+    create_auth_storage_with_keyring_store(codex_home, mode, keyring_store)
+}
+
+fn create_auth_storage_with_keyring_store(
+    codex_home: PathBuf,
+    mode: AuthCredentialsStoreMode,
+    keyring_store: Arc<dyn KeyringStore>,
+) -> Arc<dyn AuthStorageBackend> {
+    match mode {
+        AuthCredentialsStoreMode::File => Arc::new(FileAuthStorage::new(codex_home)),
+        AuthCredentialsStoreMode::Keyring => {
+            Arc::new(KeyringAuthStorage::new(codex_home, keyring_store))
+        }
+        AuthCredentialsStoreMode::Auto => Arc::new(AutoAuthStorage::new(codex_home, keyring_store)),
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use crate::token_data::IdTokenInfo;
+    use anyhow::Context;
+    use base64::Engine;
+    use pretty_assertions::assert_eq;
+    use serde_json::json;
+    use tempfile::tempdir;
+
+    use codex_keyring_store::tests::MockKeyringStore;
+    use keyring::Error as KeyringError;
+
+    #[tokio::test]
+    async fn file_storage_load_returns_auth_dot_json() -> anyhow::Result<()> {
+        let codex_home = tempdir()?;
+        let storage = FileAuthStorage::new(codex_home.path().to_path_buf());
+        let auth_dot_json = AuthDotJson {
+            openai_api_key: Some("test-key".to_string()),
+            tokens: None,
+            last_refresh: Some(Utc::now()),
+        };
+
+        storage
+            .save(&auth_dot_json)
+            .context("failed to save auth file")?;
+
+        let loaded = storage.load().context("failed to load auth file")?;
+        assert_eq!(Some(auth_dot_json), loaded);
+        Ok(())
+    }
+
+    #[tokio::test]
+    async fn file_storage_save_persists_auth_dot_json() -> anyhow::Result<()> {
+        let codex_home = tempdir()?;
+        let storage = FileAuthStorage::new(codex_home.path().to_path_buf());
+        let auth_dot_json = AuthDotJson {
+            openai_api_key: Some("test-key".to_string()),
+            tokens: None,
+            last_refresh: Some(Utc::now()),
+        };
+
+        let file = get_auth_file(codex_home.path());
+        storage
+            .save(&auth_dot_json)
+            .context("failed to save auth file")?;
+
+        let same_auth_dot_json = storage
+            .try_read_auth_json(&file)
+            .context("failed to read auth file after save")?;
+        assert_eq!(auth_dot_json, same_auth_dot_json);
+        Ok(())
+    }
+
+    #[test]
+    fn file_storage_delete_removes_auth_file() -> anyhow::Result<()> {
+        let dir = tempdir()?;
+        let auth_dot_json = AuthDotJson {
+            openai_api_key: Some("sk-test-key".to_string()),
+            tokens: None,
+            last_refresh: None,
+        };
+        let storage = create_auth_storage(dir.path().to_path_buf(), AuthCredentialsStoreMode::File);
+        storage.save(&auth_dot_json)?;
+        assert!(dir.path().join("auth.json").exists());
+        let storage = FileAuthStorage::new(dir.path().to_path_buf());
+        let removed = storage.delete()?;
+        assert!(removed);
+        assert!(!dir.path().join("auth.json").exists());
+        Ok(())
+    }
+
+    fn seed_keyring_and_fallback_auth_file_for_delete<F>(
+        mock_keyring: &MockKeyringStore,
+        codex_home: &Path,
+        compute_key: F,
+    ) -> anyhow::Result<(String, PathBuf)>
+    where
+        F: FnOnce() -> std::io::Result<String>,
+    {
+        let key = compute_key()?;
+        mock_keyring.save(KEYRING_SERVICE, &key, "{}")?;
+        let auth_file = get_auth_file(codex_home);
+        std::fs::write(&auth_file, "stale")?;
+        Ok((key, auth_file))
+    }
+
+    fn seed_keyring_with_auth<F>(
+        mock_keyring: &MockKeyringStore,
+        compute_key: F,
+        auth: &AuthDotJson,
+    ) -> anyhow::Result<()>
+    where
+        F: FnOnce() -> std::io::Result<String>,
+    {
+        let key = compute_key()?;
+        let serialized = serde_json::to_string(auth)?;
+        mock_keyring.save(KEYRING_SERVICE, &key, &serialized)?;
+        Ok(())
+    }
+
+    fn assert_keyring_saved_auth_and_removed_fallback(
+        mock_keyring: &MockKeyringStore,
+        key: &str,
+        codex_home: &Path,
+        expected: &AuthDotJson,
+    ) {
+        let saved_value = mock_keyring
+            .saved_value(key)
+            .expect("keyring entry should exist");
+        let expected_serialized = serde_json::to_string(expected).expect("serialize expected auth");
+        assert_eq!(saved_value, expected_serialized);
+        let auth_file = get_auth_file(codex_home);
+        assert!(
+            !auth_file.exists(),
+            "fallback auth.json should be removed after keyring save"
+        );
+    }
+
+    fn id_token_with_prefix(prefix: &str) -> IdTokenInfo {
+        #[derive(Serialize)]
+        struct Header {
+            alg: &'static str,
+            typ: &'static str,
+        }
+
+        let header = Header {
+            alg: "none",
+            typ: "JWT",
+        };
+        let payload = json!({
+            "email": format!("{prefix}@example.com"),
+            "https://api.openai.com/auth": {
+                "chatgpt_account_id": format!("{prefix}-account"),
+            },
+        });
+        let encode = |bytes: &[u8]| base64::engine::general_purpose::URL_SAFE_NO_PAD.encode(bytes);
+        let header_b64 = encode(&serde_json::to_vec(&header).expect("serialize header"));
+        let payload_b64 = encode(&serde_json::to_vec(&payload).expect("serialize payload"));
+        let signature_b64 = encode(b"sig");
+        let fake_jwt = format!("{header_b64}.{payload_b64}.{signature_b64}");
+
+        crate::token_data::parse_id_token(&fake_jwt).expect("fake JWT should parse")
+    }
+
+    fn auth_with_prefix(prefix: &str) -> AuthDotJson {
+        AuthDotJson {
+            openai_api_key: Some(format!("{prefix}-api-key")),
+            tokens: Some(TokenData {
+                id_token: id_token_with_prefix(prefix),
+                access_token: format!("{prefix}-access"),
+                refresh_token: format!("{prefix}-refresh"),
+                account_id: Some(format!("{prefix}-account-id")),
+            }),
+            last_refresh: None,
+        }
+    }
+
+    #[test]
+    fn keyring_auth_storage_load_returns_deserialized_auth() -> anyhow::Result<()> {
+        let codex_home = tempdir()?;
+        let mock_keyring = MockKeyringStore::default();
+        let storage = KeyringAuthStorage::new(
+            codex_home.path().to_path_buf(),
+            Arc::new(mock_keyring.clone()),
+        );
+        let expected = AuthDotJson {
+            openai_api_key: Some("sk-test".to_string()),
+            tokens: None,
+            last_refresh: None,
+        };
+        seed_keyring_with_auth(
+            &mock_keyring,
+            || compute_store_key(codex_home.path()),
+            &expected,
+        )?;
+
+        let loaded = storage.load()?;
+        assert_eq!(Some(expected), loaded);
+        Ok(())
+    }
+
+    #[test]
+    fn keyring_auth_storage_compute_store_key_for_home_directory() -> anyhow::Result<()> {
+        let codex_home = PathBuf::from("~/.codex");
+
+        let key = compute_store_key(codex_home.as_path())?;
+
+        assert_eq!(key, "cli|940db7b1d0e4eb40");
+        Ok(())
+    }
+
+    #[test]
+    fn keyring_auth_storage_save_persists_and_removes_fallback_file() -> anyhow::Result<()> {
+        let codex_home = tempdir()?;
+        let mock_keyring = MockKeyringStore::default();
+        let storage = KeyringAuthStorage::new(
+            codex_home.path().to_path_buf(),
+            Arc::new(mock_keyring.clone()),
+        );
+        let auth_file = get_auth_file(codex_home.path());
+        std::fs::write(&auth_file, "stale")?;
+        let auth = AuthDotJson {
+            openai_api_key: None,
+            tokens: Some(TokenData {
+                id_token: Default::default(),
+                access_token: "access".to_string(),
+                refresh_token: "refresh".to_string(),
+                account_id: Some("account".to_string()),
+            }),
+            last_refresh: Some(Utc::now()),
+        };
+
+        storage.save(&auth)?;
+
+        let key = compute_store_key(codex_home.path())?;
+        assert_keyring_saved_auth_and_removed_fallback(
+            &mock_keyring,
+            &key,
+            codex_home.path(),
+            &auth,
+        );
+        Ok(())
+    }
+
+    #[test]
+    fn keyring_auth_storage_delete_removes_keyring_and_file() -> anyhow::Result<()> {
+        let codex_home = tempdir()?;
+        let mock_keyring = MockKeyringStore::default();
+        let storage = KeyringAuthStorage::new(
+            codex_home.path().to_path_buf(),
+            Arc::new(mock_keyring.clone()),
+        );
+        let (key, auth_file) = seed_keyring_and_fallback_auth_file_for_delete(
+            &mock_keyring,
+            codex_home.path(),
+            || compute_store_key(codex_home.path()),
+        )?;
+
+        let removed = storage.delete()?;
+
+        assert!(removed, "delete should report removal");
+        assert!(
+            !mock_keyring.contains(&key),
+            "keyring entry should be removed"
+        );
+        assert!(
+            !auth_file.exists(),
+            "fallback auth.json should be removed after keyring delete"
+        );
+        Ok(())
+    }
+
+    #[test]
+    fn auto_auth_storage_load_prefers_keyring_value() -> anyhow::Result<()> {
+        let codex_home = tempdir()?;
+        let mock_keyring = MockKeyringStore::default();
+        let storage = AutoAuthStorage::new(
+            codex_home.path().to_path_buf(),
+            Arc::new(mock_keyring.clone()),
+        );
+        let keyring_auth = auth_with_prefix("keyring");
+        seed_keyring_with_auth(
+            &mock_keyring,
+            || compute_store_key(codex_home.path()),
+            &keyring_auth,
+        )?;
+
+        let file_auth = auth_with_prefix("file");
+        storage.file_storage.save(&file_auth)?;
+
+        let loaded = storage.load()?;
+        assert_eq!(loaded, Some(keyring_auth));
+        Ok(())
+    }
+
+    #[test]
+    fn auto_auth_storage_load_uses_file_when_keyring_empty() -> anyhow::Result<()> {
+        let codex_home = tempdir()?;
+        let mock_keyring = MockKeyringStore::default();
+        let storage = AutoAuthStorage::new(codex_home.path().to_path_buf(), Arc::new(mock_keyring));
+
+        let expected = auth_with_prefix("file-only");
+        storage.file_storage.save(&expected)?;
+
+        let loaded = storage.load()?;
+        assert_eq!(loaded, Some(expected));
+        Ok(())
+    }
+
+    #[test]
+    fn auto_auth_storage_load_falls_back_when_keyring_errors() -> anyhow::Result<()> {
+        let codex_home = tempdir()?;
+        let mock_keyring = MockKeyringStore::default();
+        let storage = AutoAuthStorage::new(
+            codex_home.path().to_path_buf(),
+            Arc::new(mock_keyring.clone()),
+        );
+        let key = compute_store_key(codex_home.path())?;
+        mock_keyring.set_error(&key, KeyringError::Invalid("error".into(), "load".into()));
+
+        let expected = auth_with_prefix("fallback");
+        storage.file_storage.save(&expected)?;
+
+        let loaded = storage.load()?;
+        assert_eq!(loaded, Some(expected));
+        Ok(())
+    }
+
+    #[test]
+    fn auto_auth_storage_save_prefers_keyring() -> anyhow::Result<()> {
+        let codex_home = tempdir()?;
+        let mock_keyring = MockKeyringStore::default();
+        let storage = AutoAuthStorage::new(
+            codex_home.path().to_path_buf(),
+            Arc::new(mock_keyring.clone()),
+        );
+        let key = compute_store_key(codex_home.path())?;
+
+        let stale = auth_with_prefix("stale");
+        storage.file_storage.save(&stale)?;
+
+        let expected = auth_with_prefix("to-save");
+        storage.save(&expected)?;
+
+        assert_keyring_saved_auth_and_removed_fallback(
+            &mock_keyring,
+            &key,
+            codex_home.path(),
+            &expected,
+        );
+        Ok(())
+    }
+
+    #[test]
+    fn auto_auth_storage_save_falls_back_when_keyring_errors() -> anyhow::Result<()> {
+        let codex_home = tempdir()?;
+        let mock_keyring = MockKeyringStore::default();
+        let storage = AutoAuthStorage::new(
+            codex_home.path().to_path_buf(),
+            Arc::new(mock_keyring.clone()),
+        );
+        let key = compute_store_key(codex_home.path())?;
+        mock_keyring.set_error(&key, KeyringError::Invalid("error".into(), "save".into()));
+
+        let auth = auth_with_prefix("fallback");
+        storage.save(&auth)?;
+
+        let auth_file = get_auth_file(codex_home.path());
+        assert!(
+            auth_file.exists(),
+            "fallback auth.json should be created when keyring save fails"
+        );
+        let saved = storage
+            .file_storage
+            .load()?
+            .context("fallback auth should exist")?;
+        assert_eq!(saved, auth);
+        assert!(
+            mock_keyring.saved_value(&key).is_none(),
+            "keyring should not contain value when save fails"
+        );
+        Ok(())
+    }
+
+    #[test]
+    fn auto_auth_storage_delete_removes_keyring_and_file() -> anyhow::Result<()> {
+        let codex_home = tempdir()?;
+        let mock_keyring = MockKeyringStore::default();
+        let storage = AutoAuthStorage::new(
+            codex_home.path().to_path_buf(),
+            Arc::new(mock_keyring.clone()),
+        );
+        let (key, auth_file) = seed_keyring_and_fallback_auth_file_for_delete(
+            &mock_keyring,
+            codex_home.path(),
+            || compute_store_key(codex_home.path()),
+        )?;
+
+        let removed = storage.delete()?;
+
+        assert!(removed, "delete should report removal");
+        assert!(
+            !mock_keyring.contains(&key),
+            "keyring entry should be removed"
+        );
+        assert!(
+            !auth_file.exists(),
+            "fallback auth.json should be removed after delete"
+        );
+        Ok(())
+    }
+}

codex-rs/core/src/chat_completions.rs

@@ -1,3 +1,4 @@
+use std::sync::Arc;
 use std::time::Duration;
 
 use crate::ModelProviderInfo;
@@ -12,13 +13,16 @@ use crate::error::Result;
 use crate::error::RetryLimitReachedError;
 use crate::error::UnexpectedResponseError;
 use crate::model_family::ModelFamily;
+use crate::protocol::TokenUsage;
 use crate::tools::spec::create_tools_json_for_chat_completions_api;
 use crate::util::backoff;
 use bytes::Bytes;
 use codex_otel::otel_event_manager::OtelEventManager;
 use codex_protocol::models::ContentItem;
+use codex_protocol::models::FunctionCallOutputContentItem;
 use codex_protocol::models::ReasoningItemContent;
 use codex_protocol::models::ResponseItem;
+use codex_utils_tokenizer::Tokenizer;
 use eventsource_stream::Eventsource;
 use futures::Stream;
 use futures::StreamExt;
@@ -33,6 +37,102 @@ use tokio::time::timeout;
 use tracing::debug;
 use tracing::trace;
 
+struct ChatUsageHeuristic {
+    tokenizer: Arc<Tokenizer>,
+    input_tokens: i64,
+    output_tokens: i64,
+    reasoning_tokens: i64,
+}
+
+impl ChatUsageHeuristic {
+    fn new(model: &str, messages: &[serde_json::Value]) -> Option<Self> {
+        let tokenizer = match Tokenizer::for_model(model) {
+            Ok(tok) => tok,
+            Err(err) => {
+                debug!(
+                    "failed to build tokenizer for model {model}; falling back to default: {err:?}"
+                );
+                match Tokenizer::try_default() {
+                    Ok(tok) => tok,
+                    Err(fallback_err) => {
+                        debug!(
+                            "failed to fall back to default tokenizer for model {model}: {fallback_err:?}"
+                        );
+                        return None;
+                    }
+                }
+            }
+        };
+
+        let tokenizer = Arc::new(tokenizer);
+        let mut input_tokens =
+            4_i64.saturating_mul(i64::try_from(messages.len()).unwrap_or(i64::MAX));
+
+        for message in messages {
+            input_tokens =
+                input_tokens.saturating_add(Self::count_value_tokens(tokenizer.as_ref(), message));
+
+            if let Some(tool_calls) = message.get("tool_calls").and_then(|v| v.as_array()) {
+                input_tokens = input_tokens.saturating_add(
+                    8_i64.saturating_mul(i64::try_from(tool_calls.len()).unwrap_or(i64::MAX)),
+                );
+            }
+        }
+
+        Some(Self {
+            tokenizer,
+            input_tokens,
+            output_tokens: 0,
+            reasoning_tokens: 0,
+        })
+    }
+
+    fn record_output(&mut self, text: &str) {
+        if text.is_empty() {
+            return;
+        }
+        self.output_tokens = self
+            .output_tokens
+            .saturating_add(self.tokenizer.count(text));
+    }
+
+    fn record_reasoning(&mut self, text: &str) {
+        if text.is_empty() {
+            return;
+        }
+        self.reasoning_tokens = self
+            .reasoning_tokens
+            .saturating_add(self.tokenizer.count(text));
+    }
+
+    fn to_usage(&self) -> TokenUsage {
+        let total = self
+            .input_tokens
+            .saturating_add(self.output_tokens)
+            .saturating_add(self.reasoning_tokens);
+        TokenUsage {
+            input_tokens: self.input_tokens,
+            cached_input_tokens: 0,
+            output_tokens: self.output_tokens,
+            reasoning_output_tokens: self.reasoning_tokens,
+            total_tokens: total,
+        }
+    }
+
+    fn count_value_tokens(tokenizer: &Tokenizer, value: &serde_json::Value) -> i64 {
+        match value {
+            serde_json::Value::String(s) => tokenizer.count(s),
+            serde_json::Value::Array(items) => items.iter().fold(0_i64, |acc, item| {
+                acc.saturating_add(Self::count_value_tokens(tokenizer, item))
+            }),
+            serde_json::Value::Object(map) => map.values().fold(0_i64, |acc, item| {
+                acc.saturating_add(Self::count_value_tokens(tokenizer, item))
+            }),
+            _ => 0,
+        }
+    }
+}
+
 /// Implementation for the classic Chat Completions API.
 pub(crate) async fn stream_chat_completions(
     prompt: &Prompt,
@@ -76,6 +176,7 @@ pub(crate) async fn stream_chat_completions(
             ResponseItem::CustomToolCall { .. } => {}
             ResponseItem::CustomToolCallOutput { .. } => {}
             ResponseItem::WebSearchCall { .. } => {}
+            ResponseItem::GhostSnapshot { .. } => {}
         }
     }
 
@@ -158,16 +259,26 @@ pub(crate) async fn stream_chat_completions(
     for (idx, item) in input.iter().enumerate() {
         match item {
             ResponseItem::Message { role, content, .. } => {
+                // Build content either as a plain string (typical for assistant text)
+                // or as an array of content items when images are present (user/tool multimodal).
                 let mut text = String::new();
+                let mut items: Vec<serde_json::Value> = Vec::new();
+                let mut saw_image = false;
+
                 for c in content {
                     match c {
                         ContentItem::InputText { text: t }
                         | ContentItem::OutputText { text: t } => {
                             text.push_str(t);
+                            items.push(json!({"type":"text","text": t}));
+                        }
+                        ContentItem::InputImage { image_url } => {
+                            saw_image = true;
+                            items.push(json!({"type":"image_url","image_url": {"url": image_url}}));
                         }
-                        _ => {}
                     }
                 }
+
                 // Skip exact-duplicate assistant messages.
                 if role == "assistant" {
                     if let Some(prev) = &last_assistant_text
@@ -178,7 +289,17 @@ pub(crate) async fn stream_chat_completions(
                     last_assistant_text = Some(text.clone());
                 }
 
-                let mut msg = json!({"role": role, "content": text});
+                // For assistant messages, always send a plain string for compatibility.
+                // For user messages, if an image is present, send an array of content items.
+                let content_value = if role == "assistant" {
+                    json!(text)
+                } else if saw_image {
+                    json!(items)
+                } else {
+                    json!(text)
+                };
+
+                let mut msg = json!({"role": role, "content": content_value});
                 if role == "assistant"
                     && let Some(reasoning) = reasoning_by_anchor_index.get(&idx)
                     && let Some(obj) = msg.as_object_mut()
@@ -237,10 +358,29 @@ pub(crate) async fn stream_chat_completions(
                 messages.push(msg);
             }
             ResponseItem::FunctionCallOutput { call_id, output } => {
+                // Prefer structured content items when available (e.g., images)
+                // otherwise fall back to the legacy plain-string content.
+                let content_value = if let Some(items) = &output.content_items {
+                    let mapped: Vec<serde_json::Value> = items
+                        .iter()
+                        .map(|it| match it {
+                            FunctionCallOutputContentItem::InputText { text } => {
+                                json!({"type":"text","text": text})
+                            }
+                            FunctionCallOutputContentItem::InputImage { image_url } => {
+                                json!({"type":"image_url","image_url": {"url": image_url}})
+                            }
+                        })
+                        .collect();
+                    json!(mapped)
+                } else {
+                    json!(output.content)
+                };
+
                 messages.push(json!({
                     "role": "tool",
                     "tool_call_id": call_id,
-                    "content": output.content,
+                    "content": content_value,
                 }));
             }
             ResponseItem::CustomToolCall {
@@ -270,6 +410,10 @@ pub(crate) async fn stream_chat_completions(
                     "content": output,
                 }));
             }
+            ResponseItem::GhostSnapshot { .. } => {
+                // Ghost snapshots annotate history but are not sent to the model.
+                continue;
+            }
             ResponseItem::Reasoning { .. }
             | ResponseItem::WebSearchCall { .. }
             | ResponseItem::Other => {
@@ -280,6 +424,8 @@ pub(crate) async fn stream_chat_completions(
     }
 
     let tools_json = create_tools_json_for_chat_completions_api(&prompt.tools)?;
+    let usage_heuristic = ChatUsageHeuristic::new(model_family.slug.as_str(), &messages);
+
     let payload = json!({
         "model": model_family.slug,
         "messages": messages,
@@ -323,6 +469,7 @@ pub(crate) async fn stream_chat_completions(
                     tx_event,
                     provider.stream_idle_timeout(),
                     otel_event_manager.clone(),
+                    usage_heuristic,
                 ));
                 return Ok(ResponseStream { rx_event });
             }
@@ -376,6 +523,7 @@ async fn process_chat_sse<S>(
     tx_event: mpsc::Sender<Result<ResponseEvent>>,
     idle_timeout: Duration,
     otel_event_manager: OtelEventManager,
+    mut usage_heuristic: Option<ChatUsageHeuristic>,
 ) where
     S: Stream<Item = Result<Bytes>> + Unpin,
 {
@@ -414,10 +562,11 @@ async fn process_chat_sse<S>(
             }
             Ok(None) => {
                 // Stream closed gracefully – emit Completed with dummy id.
+                let token_usage = usage_heuristic.as_ref().map(ChatUsageHeuristic::to_usage);
                 let _ = tx_event
                     .send(Ok(ResponseEvent::Completed {
                         response_id: String::new(),
-                        token_usage: None,
+                        token_usage,
                     }))
                     .await;
                 return;
@@ -460,10 +609,11 @@ async fn process_chat_sse<S>(
                 let _ = tx_event.send(Ok(ResponseEvent::OutputItemDone(item))).await;
             }
 
+            let token_usage = usage_heuristic.as_ref().map(ChatUsageHeuristic::to_usage);
             let _ = tx_event
                 .send(Ok(ResponseEvent::Completed {
                     response_id: String::new(),
-                    token_usage: None,
+                    token_usage,
                 }))
                 .await;
             return;
@@ -487,6 +637,9 @@ async fn process_chat_sse<S>(
                 && !content.is_empty()
             {
                 assistant_text.push_str(content);
+                if let Some(usage) = usage_heuristic.as_mut() {
+                    usage.record_output(content);
+                }
                 let _ = tx_event
                     .send(Ok(ResponseEvent::OutputTextDelta(content.to_string())))
                     .await;
@@ -520,6 +673,9 @@ async fn process_chat_sse<S>(
                 if let Some(reasoning) = maybe_text {
                     // Accumulate so we can emit a terminal Reasoning item at the end.
                     reasoning_text.push_str(&reasoning);
+                    if let Some(usage) = usage_heuristic.as_mut() {
+                        usage.record_reasoning(&reasoning);
+                    }
                     let _ = tx_event
                         .send(Ok(ResponseEvent::ReasoningContentDelta(reasoning)))
                         .await;
@@ -533,6 +689,9 @@ async fn process_chat_sse<S>(
                 if let Some(s) = message_reasoning.as_str() {
                     if !s.is_empty() {
                         reasoning_text.push_str(s);
+                        if let Some(usage) = usage_heuristic.as_mut() {
+                            usage.record_reasoning(s);
+                        }
                         let _ = tx_event
                             .send(Ok(ResponseEvent::ReasoningContentDelta(s.to_string())))
                             .await;
@@ -545,6 +704,9 @@ async fn process_chat_sse<S>(
                     && !s.is_empty()
                 {
                     reasoning_text.push_str(s);
+                    if let Some(usage) = usage_heuristic.as_mut() {
+                        usage.record_reasoning(s);
+                    }
                     let _ = tx_event
                         .send(Ok(ResponseEvent::ReasoningContentDelta(s.to_string())))
                         .await;
@@ -563,18 +725,31 @@ async fn process_chat_sse<S>(
 
                 // Extract call_id if present.
                 if let Some(id) = tool_call.get("id").and_then(|v| v.as_str()) {
-                    fn_call_state.call_id.get_or_insert_with(|| id.to_string());
+                    if fn_call_state.call_id.is_none() {
+                        if let Some(usage) = usage_heuristic.as_mut() {
+                            usage.record_output(id);
+                        }
+                        fn_call_state.call_id = Some(id.to_string());
+                    }
                 }
 
                 // Extract function details if present.
                 if let Some(function) = tool_call.get("function") {
                     if let Some(name) = function.get("name").and_then(|n| n.as_str()) {
-                        fn_call_state.name.get_or_insert_with(|| name.to_string());
+                        if fn_call_state.name.is_none() {
+                            if let Some(usage) = usage_heuristic.as_mut() {
+                                usage.record_output(name);
+                            }
+                            fn_call_state.name = Some(name.to_string());
+                        }
                     }
 
                     if let Some(args_fragment) = function.get("arguments").and_then(|a| a.as_str())
                     {
                         fn_call_state.arguments.push_str(args_fragment);
+                        if let Some(usage) = usage_heuristic.as_mut() {
+                            usage.record_output(args_fragment);
+                        }
                     }
                 }
             }
@@ -637,10 +812,11 @@ async fn process_chat_sse<S>(
                 }
 
                 // Emit Completed regardless of reason so the agent can advance.
+                let token_usage = usage_heuristic.as_ref().map(ChatUsageHeuristic::to_usage);
                 let _ = tx_event
                     .send(Ok(ResponseEvent::Completed {
                         response_id: String::new(),
-                        token_usage: None,
+                        token_usage,
                     }))
                     .await;
 

codex-rs/core/src/client.rs

@@ -134,6 +134,14 @@ impl ModelClient {
         self.stream_with_task_kind(prompt, TaskKind::Regular).await
     }
 
+    pub fn config(&self) -> Arc<Config> {
+        Arc::clone(&self.config)
+    }
+
+    pub fn provider(&self) -> &ModelProviderInfo {
+        &self.provider
+    }
+
     pub(crate) async fn stream_with_task_kind(
         &self,
         prompt: &Prompt,
@@ -215,18 +223,14 @@ impl ModelClient {
 
         let input_with_instructions = prompt.get_formatted_input();
 
-        let verbosity = match &self.config.model_family.family {
-            family if family == "gpt-5" => self.config.model_verbosity,
-            _ => {
-                if self.config.model_verbosity.is_some() {
-                    warn!(
-                        "model_verbosity is set but ignored for non-gpt-5 model family: {}",
-                        self.config.model_family.family
-                    );
-                }
-
-                None
-            }
+        let verbosity = if self.config.model_family.support_verbosity {
+            self.config.model_verbosity
+        } else {
+            warn!(
+                "model_verbosity is set but ignored as the model does not support verbosity: {}",
+                self.config.model_family.family
+            );
+            None
         };
 
         // Only include `text.verbosity` for GPT-5 family models
@@ -381,9 +385,14 @@ impl ModelClient {
 
                 if status == StatusCode::UNAUTHORIZED
                     && let Some(manager) = auth_manager.as_ref()
-                    && manager.auth().is_some()
+                    && let Some(auth) = auth.as_ref()
+                    && auth.mode == AuthMode::ChatGPT
                 {
-                    let _ = manager.refresh_token().await;
+                    manager.refresh_token().await.map_err(|err| {
+                        StreamAttemptError::Fatal(CodexErr::Fatal(format!(
+                            "Failed to refresh ChatGPT credentials: {err}"
+                        )))
+                    })?;
                 }
 
                 // The OpenAI Responses endpoint returns structured JSON bodies even for 4xx/5xx

codex-rs/core/src/codex.rs

@@ -20,7 +20,6 @@ use async_channel::Sender;
 use codex_apply_patch::ApplyPatchAction;
 use codex_protocol::ConversationId;
 use codex_protocol::items::TurnItem;
-use codex_protocol::protocol::ConversationPathResponseEvent;
 use codex_protocol::protocol::ExitedReviewModeEvent;
 use codex_protocol::protocol::ItemCompletedEvent;
 use codex_protocol::protocol::ItemStartedEvent;
@@ -88,6 +87,7 @@ use crate::protocol::Op;
 use crate::protocol::RateLimitSnapshot;
 use crate::protocol::ReviewDecision;
 use crate::protocol::ReviewOutputEvent;
+use crate::protocol::SandboxCommandAssessment;
 use crate::protocol::SandboxPolicy;
 use crate::protocol::SessionConfiguredEvent;
 use crate::protocol::StreamErrorEvent;
@@ -104,8 +104,12 @@ use crate::state::SessionServices;
 use crate::state::SessionState;
 use crate::state::TaskKind;
 use crate::tasks::CompactTask;
+use crate::tasks::GhostSnapshotTask;
 use crate::tasks::RegularTask;
 use crate::tasks::ReviewTask;
+use crate::tasks::SessionTask;
+use crate::tasks::SessionTaskContext;
+use crate::tasks::UndoTask;
 use crate::tools::ToolRouter;
 use crate::tools::context::SharedTurnDiffTracker;
 use crate::tools::parallel::ToolCallRuntime;
@@ -128,6 +132,8 @@ use codex_protocol::models::ResponseInputItem;
 use codex_protocol::models::ResponseItem;
 use codex_protocol::protocol::InitialHistory;
 use codex_protocol::user_input::UserInput;
+use codex_utils_readiness::Readiness;
+use codex_utils_readiness::ReadinessFlag;
 
 pub mod compact;
 use self::compact::build_compacted_history;
@@ -178,6 +184,7 @@ impl Codex {
             sandbox_policy: config.sandbox_policy.clone(),
             cwd: config.cwd.clone(),
             original_config_do_not_use: Arc::clone(&config),
+            features: config.features.clone(),
         };
 
         // Generate a unique ID for the lifetime of this Codex session.
@@ -271,6 +278,7 @@ pub(crate) struct TurnContext {
     pub(crate) is_review_mode: bool,
     pub(crate) final_output_json_schema: Option<Value>,
     pub(crate) codex_linux_sandbox_exe: Option<PathBuf>,
+    pub(crate) tool_call_gate: Arc<ReadinessFlag>,
 }
 
 impl TurnContext {
@@ -312,6 +320,9 @@ pub(crate) struct SessionConfiguration {
     /// operate deterministically.
     cwd: PathBuf,
 
+    /// Set of feature flags for this session
+    features: Features,
+
     // TODO(pakrym): Remove config from here
     original_config_do_not_use: Arc<Config>,
 }
@@ -406,6 +417,7 @@ impl Session {
             is_review_mode: false,
             final_output_json_schema: None,
             codex_linux_sandbox_exe: config.codex_linux_sandbox_exe.clone(),
+            tool_call_gate: Arc::new(ReadinessFlag::new()),
         }
     }
 
@@ -569,7 +581,6 @@ impl Session {
         // Dispatch the SessionConfiguredEvent first and then report any errors.
         // If resuming, include converted initial messages in the payload so UIs can render them immediately.
         let initial_messages = initial_history.get_event_msgs();
-        sess.record_initial_history(initial_history).await;
 
         let events = std::iter::once(Event {
             id: INITIAL_SUBMIT_ID.to_owned(),
@@ -588,6 +599,9 @@ impl Session {
             sess.send_event_raw(event).await;
         }
 
+        // record_initial_history can emit events. We record only after the SessionConfiguredEvent is emitted.
+        sess.record_initial_history(initial_history).await;
+
         Ok(sess)
     }
 
@@ -595,6 +609,19 @@ impl Session {
         self.tx_event.clone()
     }
 
+    /// Ensure all rollout writes are durably flushed.
+    pub(crate) async fn flush_rollout(&self) {
+        let recorder = {
+            let guard = self.services.rollout.lock().await;
+            guard.clone()
+        };
+        if let Some(rec) = recorder
+            && let Err(e) = rec.flush().await
+        {
+            warn!("failed to flush rollout recorder: {e}");
+        }
+    }
+
     fn next_internal_sub_id(&self) -> String {
         let id = self
             .next_internal_sub_id
@@ -608,7 +635,9 @@ impl Session {
             InitialHistory::New => {
                 // Build and record initial items (user instructions + environment context)
                 let items = self.build_initial_context(&turn_context);
-                self.record_conversation_items(&items).await;
+                self.record_conversation_items(&turn_context, &items).await;
+                // Ensure initial items are visible to immediate readers (e.g., tests, forks).
+                self.flush_rollout().await;
             }
             InitialHistory::Resumed(_) | InitialHistory::Forked(_) => {
                 let rollout_items = conversation_history.get_rollout_items();
@@ -625,6 +654,8 @@ impl Session {
                 if persist && !rollout_items.is_empty() {
                     self.persist_rollout_items(&rollout_items).await;
                 }
+                // Flush after seeding history and any persisted rollout copy.
+                self.flush_rollout().await;
             }
         }
     }
@@ -755,6 +786,32 @@ impl Session {
         }
     }
 
+    pub(crate) async fn assess_sandbox_command(
+        &self,
+        turn_context: &TurnContext,
+        call_id: &str,
+        command: &[String],
+        failure_message: Option<&str>,
+    ) -> Option<SandboxCommandAssessment> {
+        let config = turn_context.client.config();
+        let provider = turn_context.client.provider().clone();
+        let auth_manager = Arc::clone(&self.services.auth_manager);
+        let otel = self.services.otel_event_manager.clone();
+        crate::sandboxing::assessment::assess_command(
+            config,
+            provider,
+            auth_manager,
+            &otel,
+            self.conversation_id,
+            call_id,
+            command,
+            &turn_context.sandbox_policy,
+            &turn_context.cwd,
+            failure_message,
+        )
+        .await
+    }
+
     /// Emit an exec approval request event and await the user's decision.
     ///
     /// The request is keyed by `sub_id`/`call_id` so matching responses are delivered
@@ -767,6 +824,7 @@ impl Session {
         command: Vec<String>,
         cwd: PathBuf,
         reason: Option<String>,
+        risk: Option<SandboxCommandAssessment>,
     ) -> ReviewDecision {
         let sub_id = turn_context.sub_id.clone();
         // Add the tx_approve callback to the map before sending the request.
@@ -792,6 +850,7 @@ impl Session {
             command,
             cwd,
             reason,
+            risk,
             parsed_cmd,
         });
         self.send_event(turn_context, event).await;
@@ -857,9 +916,14 @@ impl Session {
 
     /// Records input items: always append to conversation history and
     /// persist these response items to rollout.
-    pub(crate) async fn record_conversation_items(&self, items: &[ResponseItem]) {
+    pub(crate) async fn record_conversation_items(
+        &self,
+        turn_context: &TurnContext,
+        items: &[ResponseItem],
+    ) {
         self.record_into_history(items).await;
         self.persist_rollout_response_items(items).await;
+        self.send_raw_response_items(turn_context, items).await;
     }
 
     fn reconstruct_history_from_rollout(
@@ -895,7 +959,7 @@ impl Session {
         state.record_items(items.iter());
     }
 
-    async fn replace_history(&self, items: Vec<ResponseItem>) {
+    pub(crate) async fn replace_history(&self, items: Vec<ResponseItem>) {
         let mut state = self.state.lock().await;
         state.replace_history(items);
     }
@@ -909,6 +973,13 @@ impl Session {
         self.persist_rollout_items(&rollout_items).await;
     }
 
+    async fn send_raw_response_items(&self, turn_context: &TurnContext, items: &[ResponseItem]) {
+        for item in items {
+            self.send_event(turn_context, EventMsg::RawResponseItem(item.clone()))
+                .await;
+        }
+    }
+
     pub(crate) fn build_initial_context(&self, turn_context: &TurnContext) -> Vec<ResponseItem> {
         let mut items = Vec::<ResponseItem>::with_capacity(2);
         if let Some(user_instructions) = turn_context.user_instructions.as_deref() {
@@ -1004,7 +1075,7 @@ impl Session {
     ) {
         let response_item: ResponseItem = response_input.clone().into();
         // Add to conversation history and persist response item to rollout
-        self.record_conversation_items(std::slice::from_ref(&response_item))
+        self.record_conversation_items(turn_context, std::slice::from_ref(&response_item))
             .await;
 
         // Derive user message events and persist only UserMessage to rollout
@@ -1037,6 +1108,43 @@ impl Session {
         self.send_event(turn_context, event).await;
     }
 
+    async fn maybe_start_ghost_snapshot(
+        self: &Arc<Self>,
+        turn_context: Arc<TurnContext>,
+        cancellation_token: CancellationToken,
+    ) {
+        if turn_context.is_review_mode
+            || !self
+                .state
+                .lock()
+                .await
+                .session_configuration
+                .features
+                .enabled(Feature::GhostCommit)
+        {
+            return;
+        }
+
+        let token = match turn_context.tool_call_gate.subscribe().await {
+            Ok(token) => token,
+            Err(err) => {
+                warn!("failed to subscribe to ghost snapshot readiness: {err}");
+                return;
+            }
+        };
+
+        info!("spawning ghost snapshot task");
+        let task = GhostSnapshotTask::new(token);
+        Arc::new(task)
+            .run(
+                Arc::new(SessionTaskContext::new(self.clone())),
+                turn_context.clone(),
+                Vec::new(),
+                cancellation_token,
+            )
+            .await;
+    }
+
     /// Returns the input if there was no task running to inject into
     pub async fn inject_input(&self, input: Vec<UserInput>) -> Result<(), Vec<UserInput>> {
         let mut active = self.active_turn.lock().await;
@@ -1195,8 +1303,11 @@ async fn submission_loop(sess: Arc<Session>, config: Arc<Config>, rx_sub: Receiv
                     if let Some(env_item) = sess
                         .build_environment_update_item(previous_context.as_ref(), &current_context)
                     {
-                        sess.record_conversation_items(std::slice::from_ref(&env_item))
-                            .await;
+                        sess.record_conversation_items(
+                            &current_context,
+                            std::slice::from_ref(&env_item),
+                        )
+                        .await;
                     }
 
                     sess.spawn_task(Arc::clone(&current_context), items, RegularTask)
@@ -1310,6 +1421,13 @@ async fn submission_loop(sess: Arc<Session>, config: Arc<Config>, rx_sub: Receiv
                 };
                 sess.send_event_raw(event).await;
             }
+            Op::Undo => {
+                let turn_context = sess
+                    .new_turn_with_sub_id(sub.id.clone(), SessionSettingsUpdate::default())
+                    .await;
+                sess.spawn_task(turn_context, Vec::new(), UndoTask::new())
+                    .await;
+            }
             Op::Compact => {
                 let turn_context = sess
                     .new_turn_with_sub_id(sub.id.clone(), SessionSettingsUpdate::default())
@@ -1355,33 +1473,7 @@ async fn submission_loop(sess: Arc<Session>, config: Arc<Config>, rx_sub: Receiv
                 sess.send_event_raw(event).await;
                 break;
             }
-            Op::GetPath => {
-                let sub_id = sub.id.clone();
-                // Flush rollout writes before returning the path so readers observe a consistent file.
-                let (path, rec_opt) = {
-                    let guard = sess.services.rollout.lock().await;
-                    match guard.as_ref() {
-                        Some(rec) => (rec.get_rollout_path(), Some(rec.clone())),
-                        None => {
-                            error!("rollout recorder not found");
-                            continue;
-                        }
-                    }
-                };
-                if let Some(rec) = rec_opt
-                    && let Err(e) = rec.flush().await
-                {
-                    warn!("failed to flush rollout recorder before GetHistory: {e}");
-                }
-                let event = Event {
-                    id: sub_id.clone(),
-                    msg: EventMsg::ConversationPath(ConversationPathResponseEvent {
-                        conversation_id: sess.conversation_id,
-                        path,
-                    }),
-                };
-                sess.send_event_raw(event).await;
-            }
+
             Op::Review { review_request } => {
                 let turn_context = sess
                     .new_turn_with_sub_id(sub.id.clone(), SessionSettingsUpdate::default())
@@ -1472,6 +1564,7 @@ async fn spawn_review_thread(
         is_review_mode: true,
         final_output_json_schema: None,
         codex_linux_sandbox_exe: parent_turn_context.codex_linux_sandbox_exe.clone(),
+        tool_call_gate: Arc::new(ReadinessFlag::new()),
     };
 
     // Seed the child task with the review prompt as the initial user message.
@@ -1535,6 +1628,8 @@ pub(crate) async fn run_task(
             .await;
     }
 
+    sess.maybe_start_ghost_snapshot(Arc::clone(&turn_context), cancellation_token.child_token())
+        .await;
     let mut last_agent_message: Option<String> = None;
     // Although from the perspective of codex.rs, TurnDiffTracker has the lifecycle of a Task which contains
     // many turns, from the perspective of the user, it is a single turn.
@@ -1568,7 +1663,8 @@ pub(crate) async fn run_task(
             }
             review_thread_history.get_history()
         } else {
-            sess.record_conversation_items(&pending_input).await;
+            sess.record_conversation_items(&turn_context, &pending_input)
+                .await;
             sess.history_snapshot().await
         };
 
@@ -1615,6 +1711,7 @@ pub(crate) async fn run_task(
                     is_review_mode,
                     &mut review_thread_history,
                     &sess,
+                    &turn_context,
                 )
                 .await;
 
@@ -1663,6 +1760,7 @@ pub(crate) async fn run_task(
                     is_review_mode,
                     &mut review_thread_history,
                     &sess,
+                    &turn_context,
                 )
                 .await;
                 // Aborted turn is reported via a different event.
@@ -1724,6 +1822,13 @@ fn parse_review_output_event(text: &str) -> ReviewOutputEvent {
     }
 }
 
+fn filter_model_visible_history(input: Vec<ResponseItem>) -> Vec<ResponseItem> {
+    input
+        .into_iter()
+        .filter(|item| !matches!(item, ResponseItem::GhostSnapshot { .. }))
+        .collect()
+}
+
 async fn run_turn(
     sess: Arc<Session>,
     turn_context: Arc<TurnContext>,
@@ -1744,7 +1849,7 @@ async fn run_turn(
         .supports_parallel_tool_calls;
     let parallel_tool_calls = model_supports_parallel;
     let prompt = Prompt {
-        input,
+        input: filter_model_visible_history(input),
         tools: router.specs(),
         parallel_tool_calls,
         base_instructions_override: turn_context.base_instructions.clone(),
@@ -1806,7 +1911,7 @@ async fn run_turn(
                     // at a seemingly frozen screen.
                     sess.notify_stream_error(
                         turn_context.as_ref(),
-                        format!("Re-connecting... {retries}/{max_retries}"),
+                        format!("Reconnecting... {retries}/{max_retries}"),
                     )
                     .await;
 
@@ -1942,7 +2047,7 @@ async fn try_run_turn(
                             call_id: String::new(),
                             output: FunctionCallOutputPayload {
                                 content: msg.to_string(),
-                                success: None,
+                                ..Default::default()
                             },
                         };
                         add_completed(ProcessedResponseItem {
@@ -1956,7 +2061,7 @@ async fn try_run_turn(
                             call_id: String::new(),
                             output: FunctionCallOutputPayload {
                                 content: message,
-                                success: None,
+                                ..Default::default()
                             },
                         };
                         add_completed(ProcessedResponseItem {
@@ -2094,41 +2199,6 @@ pub(super) fn get_last_assistant_message_from_turn(responses: &[ResponseItem]) -
         }
     })
 }
-pub(crate) fn convert_call_tool_result_to_function_call_output_payload(
-    call_tool_result: &CallToolResult,
-) -> FunctionCallOutputPayload {
-    let CallToolResult {
-        content,
-        is_error,
-        structured_content,
-    } = call_tool_result;
-
-    // In terms of what to send back to the model, we prefer structured_content,
-    // if available, and fallback to content, otherwise.
-    let mut is_success = is_error != &Some(true);
-    let content = if let Some(structured_content) = structured_content
-        && structured_content != &serde_json::Value::Null
-        && let Ok(serialized_structured_content) = serde_json::to_string(&structured_content)
-    {
-        serialized_structured_content
-    } else {
-        match serde_json::to_string(&content) {
-            Ok(serialized_content) => serialized_content,
-            Err(err) => {
-                // If we could not serialize either content or structured_content to
-                // JSON, flag this as an error.
-                is_success = false;
-                err.to_string()
-            }
-        }
-    };
-
-    FunctionCallOutputPayload {
-        content,
-        success: Some(is_success),
-    }
-}
-
 /// Emits an ExitedReviewMode Event with optional ReviewOutput,
 /// and records a developer message with the review output.
 pub(crate) async fn exit_review_mode(
@@ -2173,12 +2243,17 @@ pub(crate) async fn exit_review_mode(
     }
 
     session
-        .record_conversation_items(&[ResponseItem::Message {
-            id: None,
-            role: "user".to_string(),
-            content: vec![ContentItem::InputText { text: user_message }],
-        }])
+        .record_conversation_items(
+            &turn_context,
+            &[ResponseItem::Message {
+                id: None,
+                role: "user".to_string(),
+                content: vec![ContentItem::InputText { text: user_message }],
+            }],
+        )
         .await;
+    // Make the recorded review note visible immediately for readers.
+    session.flush_rollout().await;
 }
 
 fn mcp_init_error_display(
@@ -2234,6 +2309,8 @@ fn is_mcp_client_startup_timeout_error(error: &anyhow::Error) -> bool {
         || error_message.contains("timed out handshaking with MCP server")
 }
 
+use crate::features::Feature;
+use crate::features::Features;
 #[cfg(test)]
 pub(crate) use tests::make_session_and_context;
 
@@ -2254,10 +2331,6 @@ mod tests {
     use crate::state::TaskKind;
     use crate::tasks::SessionTask;
     use crate::tasks::SessionTaskContext;
-    use crate::tools::MODEL_FORMAT_HEAD_LINES;
-    use crate::tools::MODEL_FORMAT_MAX_BYTES;
-    use crate::tools::MODEL_FORMAT_MAX_LINES;
-    use crate::tools::MODEL_FORMAT_TAIL_LINES;
     use crate::tools::ToolRouter;
     use crate::tools::context::ToolInvocation;
     use crate::tools::context::ToolOutput;
@@ -2331,7 +2404,7 @@ mod tests {
             })),
         };
 
-        let got = convert_call_tool_result_to_function_call_output_payload(&ctr);
+        let got = FunctionCallOutputPayload::from(&ctr);
         let expected = FunctionCallOutputPayload {
             content: serde_json::to_string(&json!({
                 "ok": true,
@@ -2339,100 +2412,12 @@ mod tests {
             }))
             .unwrap(),
             success: Some(true),
+            ..Default::default()
         };
 
         assert_eq!(expected, got);
     }
 
-    #[test]
-    fn model_truncation_head_tail_by_lines() {
-        // Build 400 short lines so line-count limit, not byte budget, triggers truncation
-        let lines: Vec<String> = (1..=400).map(|i| format!("line{i}")).collect();
-        let full = lines.join("\n");
-
-        let exec = ExecToolCallOutput {
-            exit_code: 0,
-            stdout: StreamOutput::new(String::new()),
-            stderr: StreamOutput::new(String::new()),
-            aggregated_output: StreamOutput::new(full),
-            duration: StdDuration::from_secs(1),
-            timed_out: false,
-        };
-
-        let out = format_exec_output_str(&exec);
-
-        // Strip truncation header if present for subsequent assertions
-        let body = out
-            .strip_prefix("Total output lines: ")
-            .and_then(|rest| rest.split_once("\n\n").map(|x| x.1))
-            .unwrap_or(out.as_str());
-
-        // Expect elision marker with correct counts
-        let omitted = 400 - MODEL_FORMAT_MAX_LINES; // 144
-        let marker = format!("\n[... omitted {omitted} of 400 lines ...]\n\n");
-        assert!(out.contains(&marker), "missing marker: {out}");
-
-        // Validate head and tail
-        let parts: Vec<&str> = body.split(&marker).collect();
-        assert_eq!(parts.len(), 2, "expected one marker split");
-        let head = parts[0];
-        let tail = parts[1];
-
-        let expected_head: String = (1..=MODEL_FORMAT_HEAD_LINES)
-            .map(|i| format!("line{i}"))
-            .collect::<Vec<_>>()
-            .join("\n");
-        assert!(head.starts_with(&expected_head), "head mismatch");
-
-        let expected_tail: String = ((400 - MODEL_FORMAT_TAIL_LINES + 1)..=400)
-            .map(|i| format!("line{i}"))
-            .collect::<Vec<_>>()
-            .join("\n");
-        assert!(tail.ends_with(&expected_tail), "tail mismatch");
-    }
-
-    #[test]
-    fn model_truncation_respects_byte_budget() {
-        // Construct a large output (about 100kB) so byte budget dominates
-        let big_line = "x".repeat(100);
-        let full = std::iter::repeat_n(big_line, 1000)
-            .collect::<Vec<_>>()
-            .join("\n");
-
-        let exec = ExecToolCallOutput {
-            exit_code: 0,
-            stdout: StreamOutput::new(String::new()),
-            stderr: StreamOutput::new(String::new()),
-            aggregated_output: StreamOutput::new(full.clone()),
-            duration: StdDuration::from_secs(1),
-            timed_out: false,
-        };
-
-        let out = format_exec_output_str(&exec);
-        // Keep strict budget on the truncated body (excluding header)
-        let body = out
-            .strip_prefix("Total output lines: ")
-            .and_then(|rest| rest.split_once("\n\n").map(|x| x.1))
-            .unwrap_or(out.as_str());
-        assert!(body.len() <= MODEL_FORMAT_MAX_BYTES, "exceeds byte budget");
-        assert!(out.contains("omitted"), "should contain elision marker");
-
-        // Ensure head and tail are drawn from the original
-        assert!(full.starts_with(body.chars().take(8).collect::<String>().as_str()));
-        assert!(
-            full.ends_with(
-                body.chars()
-                    .rev()
-                    .take(8)
-                    .collect::<String>()
-                    .chars()
-                    .rev()
-                    .collect::<String>()
-                    .as_str()
-            )
-        );
-    }
-
     #[test]
     fn includes_timed_out_message() {
         let exec = ExecToolCallOutput {
@@ -2460,11 +2445,12 @@ mod tests {
             structured_content: Some(serde_json::Value::Null),
         };
 
-        let got = convert_call_tool_result_to_function_call_output_payload(&ctr);
+        let got = FunctionCallOutputPayload::from(&ctr);
         let expected = FunctionCallOutputPayload {
             content: serde_json::to_string(&vec![text_block("hello"), text_block("world")])
                 .unwrap(),
             success: Some(true),
+            ..Default::default()
         };
 
         assert_eq!(expected, got);
@@ -2478,10 +2464,11 @@ mod tests {
             structured_content: Some(json!({ "message": "bad" })),
         };
 
-        let got = convert_call_tool_result_to_function_call_output_payload(&ctr);
+        let got = FunctionCallOutputPayload::from(&ctr);
         let expected = FunctionCallOutputPayload {
             content: serde_json::to_string(&json!({ "message": "bad" })).unwrap(),
             success: Some(false),
+            ..Default::default()
         };
 
         assert_eq!(expected, got);
@@ -2495,10 +2482,11 @@ mod tests {
             structured_content: None,
         };
 
-        let got = convert_call_tool_result_to_function_call_output_payload(&ctr);
+        let got = FunctionCallOutputPayload::from(&ctr);
         let expected = FunctionCallOutputPayload {
             content: serde_json::to_string(&vec![text_block("alpha")]).unwrap(),
             success: Some(true),
+            ..Default::default()
         };
 
         assert_eq!(expected, got);
@@ -2550,6 +2538,7 @@ mod tests {
             sandbox_policy: config.sandbox_policy.clone(),
             cwd: config.cwd.clone(),
             original_config_do_not_use: Arc::clone(&config),
+            features: Features::default(),
         };
 
         let state = SessionState::new(session_configuration.clone());
@@ -2618,6 +2607,7 @@ mod tests {
             sandbox_policy: config.sandbox_policy.clone(),
             cwd: config.cwd.clone(),
             original_config_do_not_use: Arc::clone(&config),
+            features: Features::default(),
         };
 
         let state = SessionState::new(session_configuration.clone());
@@ -2772,13 +2762,19 @@ mod tests {
             EventMsg::ExitedReviewMode(ev) => assert!(ev.review_output.is_none()),
             other => panic!("unexpected first event: {other:?}"),
         }
-        let second = tokio::time::timeout(std::time::Duration::from_secs(2), rx.recv())
-            .await
-            .expect("timeout waiting for second event")
-            .expect("second event");
-        match second.msg {
-            EventMsg::TurnAborted(e) => assert_eq!(TurnAbortReason::Interrupted, e.reason),
-            other => panic!("unexpected second event: {other:?}"),
+        loop {
+            let evt = tokio::time::timeout(std::time::Duration::from_secs(2), rx.recv())
+                .await
+                .expect("timeout waiting for next event")
+                .expect("event");
+            match evt.msg {
+                EventMsg::RawResponseItem(_) => continue,
+                EventMsg::TurnAborted(e) => {
+                    assert_eq!(TurnAbortReason::Interrupted, e.reason);
+                    break;
+                }
+                other => panic!("unexpected second event: {other:?}"),
+            }
         }
 
         let history = sess.history_snapshot().await;

codex-rs/core/src/codex/compact.rs

@@ -2,6 +2,7 @@ use std::sync::Arc;
 
 use super::Session;
 use super::TurnContext;
+use super::filter_model_visible_history;
 use super::get_last_assistant_message_from_turn;
 use crate::Prompt;
 use crate::client_common::ResponseEvent;
@@ -86,8 +87,9 @@ async fn run_compact_task_inner(
 
     loop {
         let turn_input = history.get_history();
+        let prompt_input = filter_model_visible_history(turn_input.clone());
         let prompt = Prompt {
-            input: turn_input.clone(),
+            input: prompt_input.clone(),
             ..Default::default()
         };
         let attempt_result = drain_to_completed(&sess, turn_context.as_ref(), &prompt).await;
@@ -109,7 +111,7 @@ async fn run_compact_task_inner(
                 return;
             }
             Err(e @ CodexErr::ContextWindowExceeded) => {
-                if turn_input.len() > 1 {
+                if prompt_input.len() > 1 {
                     // Trim from the beginning to preserve cache (prefix-based) and keep recent messages intact.
                     error!(
                         "Context window exceeded while compacting; removing oldest history item. Error: {e}"
@@ -132,7 +134,7 @@ async fn run_compact_task_inner(
                     let delay = backoff(retries);
                     sess.notify_stream_error(
                         turn_context.as_ref(),
-                        format!("Re-connecting... {retries}/{max_retries}"),
+                        format!("Reconnecting... {retries}/{max_retries}"),
                     )
                     .await;
                     tokio::time::sleep(delay).await;
@@ -152,7 +154,13 @@ async fn run_compact_task_inner(
     let summary_text = get_last_assistant_message_from_turn(&history_snapshot).unwrap_or_default();
     let user_messages = collect_user_messages(&history_snapshot);
     let initial_context = sess.build_initial_context(turn_context.as_ref());
-    let new_history = build_compacted_history(initial_context, &user_messages, &summary_text);
+    let mut new_history = build_compacted_history(initial_context, &user_messages, &summary_text);
+    let ghost_snapshots: Vec<ResponseItem> = history_snapshot
+        .iter()
+        .filter(|item| matches!(item, ResponseItem::GhostSnapshot { .. }))
+        .cloned()
+        .collect();
+    new_history.extend(ghost_snapshots);
     sess.replace_history(new_history).await;
 
     let rollout_item = RolloutItem::Compacted(CompactedItem {
@@ -200,7 +208,20 @@ pub(crate) fn build_compacted_history(
     user_messages: &[String],
     summary_text: &str,
 ) -> Vec<ResponseItem> {
-    let mut history = initial_context;
+    build_compacted_history_with_limit(
+        initial_context,
+        user_messages,
+        summary_text,
+        COMPACT_USER_MESSAGE_MAX_TOKENS * 4,
+    )
+}
+
+fn build_compacted_history_with_limit(
+    mut history: Vec<ResponseItem>,
+    user_messages: &[String],
+    summary_text: &str,
+    max_bytes: usize,
+) -> Vec<ResponseItem> {
     let mut user_messages_text = if user_messages.is_empty() {
         "(none)".to_string()
     } else {
@@ -208,7 +229,6 @@ pub(crate) fn build_compacted_history(
     };
     // Truncate the concatenated prior user messages so the bridge message
     // stays well under the context window (approx. 4 bytes/token).
-    let max_bytes = COMPACT_USER_MESSAGE_MAX_TOKENS * 4;
     if user_messages_text.len() > max_bytes {
         user_messages_text = truncate_middle(&user_messages_text, max_bytes).0;
     }
@@ -361,11 +381,16 @@ mod tests {
 
     #[test]
     fn build_compacted_history_truncates_overlong_user_messages() {
-        // Prepare a very large prior user message so the aggregated
-        // `user_messages_text` exceeds the truncation threshold used by
-        // `build_compacted_history` (80k bytes).
-        let big = "X".repeat(200_000);
-        let history = build_compacted_history(Vec::new(), std::slice::from_ref(&big), "SUMMARY");
+        // Use a small truncation limit so the test remains fast while still validating
+        // that oversized user content is truncated.
+        let max_bytes = 128;
+        let big = "X".repeat(max_bytes + 50);
+        let history = super::build_compacted_history_with_limit(
+            Vec::new(),
+            std::slice::from_ref(&big),
+            "SUMMARY",
+            max_bytes,
+        );
 
         // Expect exactly one bridge message added to history (plus any initial context we provided, which is none).
         assert_eq!(history.len(), 1);

codex-rs/core/src/codex_conversation.rs

@@ -3,16 +3,21 @@ use crate::error::Result as CodexResult;
 use crate::protocol::Event;
 use crate::protocol::Op;
 use crate::protocol::Submission;
+use std::path::PathBuf;
 
 pub struct CodexConversation {
     codex: Codex,
+    rollout_path: PathBuf,
 }
 
 /// Conduit for the bidirectional stream of messages that compose a conversation
 /// in Codex.
 impl CodexConversation {
-    pub(crate) fn new(codex: Codex) -> Self {
-        Self { codex }
+    pub(crate) fn new(codex: Codex, rollout_path: PathBuf) -> Self {
+        Self {
+            codex,
+            rollout_path,
+        }
     }
 
     pub async fn submit(&self, op: Op) -> CodexResult<String> {
@@ -27,4 +32,8 @@ impl CodexConversation {
     pub async fn next_event(&self) -> CodexResult<Event> {
         self.codex.next_event().await
     }
+
+    pub fn rollout_path(&self) -> PathBuf {
+        self.rollout_path.clone()
+    }
 }

codex-rs/core/src/config.rs

@@ -223,6 +223,9 @@ pub struct Config {
 
     pub tools_web_search_request: bool,
 
+    /// When `true`, run a model-based assessment for commands denied by the sandbox.
+    pub experimental_sandbox_command_assessment: bool,
+
     pub use_experimental_streamable_shell_tool: bool,
 
     /// If set to `true`, used only the experimental unified exec tool.
@@ -958,6 +961,7 @@ pub struct ConfigToml {
     pub experimental_use_unified_exec_tool: Option<bool>,
     pub experimental_use_rmcp_client: Option<bool>,
     pub experimental_use_freeform_apply_patch: Option<bool>,
+    pub experimental_sandbox_command_assessment: Option<bool>,
 }
 
 impl From<ConfigToml> for UserSavedConfig {
@@ -1023,9 +1027,11 @@ impl ConfigToml {
     fn derive_sandbox_policy(
         &self,
         sandbox_mode_override: Option<SandboxMode>,
+        profile_sandbox_mode: Option<SandboxMode>,
         resolved_cwd: &Path,
     ) -> SandboxPolicy {
         let resolved_sandbox_mode = sandbox_mode_override
+            .or(profile_sandbox_mode)
             .or(self.sandbox_mode)
             .or_else(|| {
                 // if no sandbox_mode is set, but user has marked directory as trusted, use WorkspaceWrite
@@ -1118,6 +1124,7 @@ pub struct ConfigOverrides {
     pub include_view_image_tool: Option<bool>,
     pub show_raw_agent_reasoning: Option<bool>,
     pub tools_web_search_request: Option<bool>,
+    pub experimental_sandbox_command_assessment: Option<bool>,
     /// Additional directories that should be treated as writable roots for this session.
     pub additional_writable_roots: Vec<PathBuf>,
 }
@@ -1147,6 +1154,7 @@ impl Config {
             include_view_image_tool: include_view_image_tool_override,
             show_raw_agent_reasoning,
             tools_web_search_request: override_tools_web_search_request,
+            experimental_sandbox_command_assessment: sandbox_command_assessment_override,
             additional_writable_roots,
         } = overrides;
 
@@ -1172,6 +1180,7 @@ impl Config {
             include_apply_patch_tool: include_apply_patch_tool_override,
             include_view_image_tool: include_view_image_tool_override,
             web_search_request: override_tools_web_search_request,
+            experimental_sandbox_command_assessment: sandbox_command_assessment_override,
         };
 
         let features = Features::from_config(&cfg, &config_profile, feature_overrides);
@@ -1212,7 +1221,8 @@ impl Config {
             .get_active_project(&resolved_cwd)
             .unwrap_or(ProjectConfig { trust_level: None });
 
-        let mut sandbox_policy = cfg.derive_sandbox_policy(sandbox_mode, &resolved_cwd);
+        let mut sandbox_policy =
+            cfg.derive_sandbox_policy(sandbox_mode, config_profile.sandbox_mode, &resolved_cwd);
         if let SandboxPolicy::WorkspaceWrite { writable_roots, .. } = &mut sandbox_policy {
             for path in additional_writable_roots {
                 if !writable_roots.iter().any(|existing| existing == &path) {
@@ -1235,8 +1245,8 @@ impl Config {
             .is_some()
             || config_profile.approval_policy.is_some()
             || cfg.approval_policy.is_some()
-            // TODO(#3034): profile.sandbox_mode is not implemented
             || sandbox_mode.is_some()
+            || config_profile.sandbox_mode.is_some()
             || cfg.sandbox_mode.is_some();
 
         let mut model_providers = built_in_model_providers();
@@ -1269,6 +1279,8 @@ impl Config {
         let use_experimental_streamable_shell_tool = features.enabled(Feature::StreamableShell);
         let use_experimental_unified_exec_tool = features.enabled(Feature::UnifiedExec);
         let use_experimental_use_rmcp_client = features.enabled(Feature::RmcpClient);
+        let experimental_sandbox_command_assessment =
+            features.enabled(Feature::SandboxCommandAssessment);
 
         let forced_chatgpt_workspace_id =
             cfg.forced_chatgpt_workspace_id.as_ref().and_then(|value| {
@@ -1390,6 +1402,7 @@ impl Config {
             forced_login_method,
             include_apply_patch_tool: include_apply_patch_tool_flag,
             tools_web_search_request,
+            experimental_sandbox_command_assessment,
             use_experimental_streamable_shell_tool,
             use_experimental_unified_exec_tool,
             use_experimental_use_rmcp_client,
@@ -1593,8 +1606,11 @@ network_access = false  # This should be ignored.
         let sandbox_mode_override = None;
         assert_eq!(
             SandboxPolicy::DangerFullAccess,
-            sandbox_full_access_cfg
-                .derive_sandbox_policy(sandbox_mode_override, &PathBuf::from("/tmp/test"))
+            sandbox_full_access_cfg.derive_sandbox_policy(
+                sandbox_mode_override,
+                None,
+                &PathBuf::from("/tmp/test")
+            )
         );
 
         let sandbox_read_only = r#"
@@ -1609,8 +1625,11 @@ network_access = true  # This should be ignored.
         let sandbox_mode_override = None;
         assert_eq!(
             SandboxPolicy::ReadOnly,
-            sandbox_read_only_cfg
-                .derive_sandbox_policy(sandbox_mode_override, &PathBuf::from("/tmp/test"))
+            sandbox_read_only_cfg.derive_sandbox_policy(
+                sandbox_mode_override,
+                None,
+                &PathBuf::from("/tmp/test")
+            )
         );
 
         let sandbox_workspace_write = r#"
@@ -1634,8 +1653,11 @@ exclude_slash_tmp = true
                 exclude_tmpdir_env_var: true,
                 exclude_slash_tmp: true,
             },
-            sandbox_workspace_write_cfg
-                .derive_sandbox_policy(sandbox_mode_override, &PathBuf::from("/tmp/test"))
+            sandbox_workspace_write_cfg.derive_sandbox_policy(
+                sandbox_mode_override,
+                None,
+                &PathBuf::from("/tmp/test")
+            )
         );
 
         let sandbox_workspace_write = r#"
@@ -1662,8 +1684,11 @@ trust_level = "trusted"
                 exclude_tmpdir_env_var: true,
                 exclude_slash_tmp: true,
             },
-            sandbox_workspace_write_cfg
-                .derive_sandbox_policy(sandbox_mode_override, &PathBuf::from("/tmp/test"))
+            sandbox_workspace_write_cfg.derive_sandbox_policy(
+                sandbox_mode_override,
+                None,
+                &PathBuf::from("/tmp/test")
+            )
         );
     }
 
@@ -1755,6 +1780,75 @@ trust_level = "trusted"
         Ok(())
     }
 
+    #[test]
+    fn profile_sandbox_mode_overrides_base() -> std::io::Result<()> {
+        let codex_home = TempDir::new()?;
+        let mut profiles = HashMap::new();
+        profiles.insert(
+            "work".to_string(),
+            ConfigProfile {
+                sandbox_mode: Some(SandboxMode::DangerFullAccess),
+                ..Default::default()
+            },
+        );
+        let cfg = ConfigToml {
+            profiles,
+            profile: Some("work".to_string()),
+            sandbox_mode: Some(SandboxMode::ReadOnly),
+            ..Default::default()
+        };
+
+        let config = Config::load_from_base_config_with_overrides(
+            cfg,
+            ConfigOverrides::default(),
+            codex_home.path().to_path_buf(),
+        )?;
+
+        assert!(matches!(
+            config.sandbox_policy,
+            SandboxPolicy::DangerFullAccess
+        ));
+        assert!(config.did_user_set_custom_approval_policy_or_sandbox_mode);
+
+        Ok(())
+    }
+
+    #[test]
+    fn cli_override_takes_precedence_over_profile_sandbox_mode() -> std::io::Result<()> {
+        let codex_home = TempDir::new()?;
+        let mut profiles = HashMap::new();
+        profiles.insert(
+            "work".to_string(),
+            ConfigProfile {
+                sandbox_mode: Some(SandboxMode::DangerFullAccess),
+                ..Default::default()
+            },
+        );
+        let cfg = ConfigToml {
+            profiles,
+            profile: Some("work".to_string()),
+            ..Default::default()
+        };
+
+        let overrides = ConfigOverrides {
+            sandbox_mode: Some(SandboxMode::WorkspaceWrite),
+            ..Default::default()
+        };
+
+        let config = Config::load_from_base_config_with_overrides(
+            cfg,
+            overrides,
+            codex_home.path().to_path_buf(),
+        )?;
+
+        assert!(matches!(
+            config.sandbox_policy,
+            SandboxPolicy::WorkspaceWrite { .. }
+        ));
+
+        Ok(())
+    }
+
     #[test]
     fn feature_table_overrides_legacy_flags() -> std::io::Result<()> {
         let codex_home = TempDir::new()?;
@@ -2873,6 +2967,7 @@ model_verbosity = "high"
                 forced_login_method: None,
                 include_apply_patch_tool: false,
                 tools_web_search_request: false,
+                experimental_sandbox_command_assessment: false,
                 use_experimental_streamable_shell_tool: false,
                 use_experimental_unified_exec_tool: false,
                 use_experimental_use_rmcp_client: false,
@@ -2941,6 +3036,7 @@ model_verbosity = "high"
             forced_login_method: None,
             include_apply_patch_tool: false,
             tools_web_search_request: false,
+            experimental_sandbox_command_assessment: false,
             use_experimental_streamable_shell_tool: false,
             use_experimental_unified_exec_tool: false,
             use_experimental_use_rmcp_client: false,
@@ -3024,6 +3120,7 @@ model_verbosity = "high"
             forced_login_method: None,
             include_apply_patch_tool: false,
             tools_web_search_request: false,
+            experimental_sandbox_command_assessment: false,
             use_experimental_streamable_shell_tool: false,
             use_experimental_unified_exec_tool: false,
             use_experimental_use_rmcp_client: false,
@@ -3093,6 +3190,7 @@ model_verbosity = "high"
             forced_login_method: None,
             include_apply_patch_tool: false,
             tools_web_search_request: false,
+            experimental_sandbox_command_assessment: false,
             use_experimental_streamable_shell_tool: false,
             use_experimental_unified_exec_tool: false,
             use_experimental_use_rmcp_client: false,

codex-rs/core/src/config_profile.rs

@@ -4,6 +4,7 @@ use std::path::PathBuf;
 use crate::protocol::AskForApproval;
 use codex_protocol::config_types::ReasoningEffort;
 use codex_protocol::config_types::ReasoningSummary;
+use codex_protocol::config_types::SandboxMode;
 use codex_protocol::config_types::Verbosity;
 
 /// Collection of common configuration options that a user can define as a unit
@@ -15,6 +16,7 @@ pub struct ConfigProfile {
     /// [`ModelProviderInfo`] to use.
     pub model_provider: Option<String>,
     pub approval_policy: Option<AskForApproval>,
+    pub sandbox_mode: Option<SandboxMode>,
     pub model_reasoning_effort: Option<ReasoningEffort>,
     pub model_reasoning_summary: Option<ReasoningSummary>,
     pub model_verbosity: Option<Verbosity>,
@@ -26,6 +28,7 @@ pub struct ConfigProfile {
     pub experimental_use_exec_command_tool: Option<bool>,
     pub experimental_use_rmcp_client: Option<bool>,
     pub experimental_use_freeform_apply_patch: Option<bool>,
+    pub experimental_sandbox_command_assessment: Option<bool>,
     pub tools_web_search: Option<bool>,
     pub tools_view_image: Option<bool>,
     /// Optional feature toggles scoped to this profile.

codex-rs/core/src/conversation_history.rs

@@ -1,9 +1,20 @@
+use codex_protocol::models::FunctionCallOutputContentItem;
 use codex_protocol::models::FunctionCallOutputPayload;
 use codex_protocol::models::ResponseItem;
 use codex_protocol::protocol::TokenUsage;
 use codex_protocol::protocol::TokenUsageInfo;
+use codex_utils_string::take_bytes_at_char_boundary;
+use codex_utils_string::take_last_bytes_at_char_boundary;
+use std::ops::Deref;
 use tracing::error;
 
+// Model-formatting limits: clients get full streams; only content sent to the model is truncated.
+pub(crate) const MODEL_FORMAT_MAX_BYTES: usize = 10 * 1024; // 10 KiB
+pub(crate) const MODEL_FORMAT_MAX_LINES: usize = 256; // lines
+pub(crate) const MODEL_FORMAT_HEAD_LINES: usize = MODEL_FORMAT_MAX_LINES / 2;
+pub(crate) const MODEL_FORMAT_TAIL_LINES: usize = MODEL_FORMAT_MAX_LINES - MODEL_FORMAT_HEAD_LINES; // 128
+pub(crate) const MODEL_FORMAT_HEAD_BYTES: usize = MODEL_FORMAT_MAX_BYTES / 2;
+
 /// Transcript of conversation history
 #[derive(Debug, Clone, Default)]
 pub(crate) struct ConversationHistory {
@@ -40,11 +51,14 @@ impl ConversationHistory {
         I::Item: std::ops::Deref<Target = ResponseItem>,
     {
         for item in items {
-            if !is_api_message(&item) {
+            let item_ref = item.deref();
+            let is_ghost_snapshot = matches!(item_ref, ResponseItem::GhostSnapshot { .. });
+            if !is_api_message(item_ref) && !is_ghost_snapshot {
                 continue;
             }
 
-            self.items.push(item.clone());
+            let processed = Self::process_item(&item);
+            self.items.push(processed);
         }
     }
 
@@ -65,6 +79,22 @@ impl ConversationHistory {
         }
     }
 
+    pub(crate) fn replace(&mut self, items: Vec<ResponseItem>) {
+        self.items = items;
+    }
+
+    pub(crate) fn update_token_info(
+        &mut self,
+        usage: &TokenUsage,
+        model_context_window: Option<i64>,
+    ) {
+        self.token_info = TokenUsageInfo::new_or_append(
+            &self.token_info,
+            &Some(usage.clone()),
+            model_context_window,
+        );
+    }
+
     /// This function enforces a couple of invariants on the in-memory history:
     /// 1. every call (function/custom) has a corresponding output entry
     /// 2. every output has a corresponding call entry
@@ -107,7 +137,7 @@ impl ConversationHistory {
                                 call_id: call_id.clone(),
                                 output: FunctionCallOutputPayload {
                                     content: "aborted".to_string(),
-                                    success: None,
+                                    ..Default::default()
                                 },
                             },
                         ));
@@ -154,7 +184,7 @@ impl ConversationHistory {
                                     call_id: call_id.clone(),
                                     output: FunctionCallOutputPayload {
                                         content: "aborted".to_string(),
-                                        success: None,
+                                        ..Default::default()
                                     },
                                 },
                             ));
@@ -165,6 +195,7 @@ impl ConversationHistory {
                 | ResponseItem::WebSearchCall { .. }
                 | ResponseItem::FunctionCallOutput { .. }
                 | ResponseItem::CustomToolCallOutput { .. }
+                | ResponseItem::GhostSnapshot { .. }
                 | ResponseItem::Other
                 | ResponseItem::Message { .. } => {
                     // nothing to do for these variants
@@ -231,6 +262,7 @@ impl ConversationHistory {
                 | ResponseItem::LocalShellCall { .. }
                 | ResponseItem::Reasoning { .. }
                 | ResponseItem::WebSearchCall { .. }
+                | ResponseItem::GhostSnapshot { .. }
                 | ResponseItem::Other
                 | ResponseItem::Message { .. } => {
                     // nothing to do for these variants
@@ -248,10 +280,6 @@ impl ConversationHistory {
         }
     }
 
-    pub(crate) fn replace(&mut self, items: Vec<ResponseItem>) {
-        self.items = items;
-    }
-
     /// Removes the corresponding paired item for the provided `item`, if any.
     ///
     /// Pairs:
@@ -321,19 +349,126 @@ impl ConversationHistory {
         }
     }
 
-    pub(crate) fn update_token_info(
-        &mut self,
-        usage: &TokenUsage,
-        model_context_window: Option<i64>,
-    ) {
-        self.token_info = TokenUsageInfo::new_or_append(
-            &self.token_info,
-            &Some(usage.clone()),
-            model_context_window,
-        );
+    fn process_item(item: &ResponseItem) -> ResponseItem {
+        match item {
+            ResponseItem::FunctionCallOutput { call_id, output } => {
+                let truncated = format_output_for_model_body(output.content.as_str());
+                let truncated_items = output.content_items.as_ref().map(|items| {
+                    items
+                        .iter()
+                        .map(|it| match it {
+                            FunctionCallOutputContentItem::InputText { text } => {
+                                FunctionCallOutputContentItem::InputText {
+                                    text: format_output_for_model_body(text),
+                                }
+                            }
+                            FunctionCallOutputContentItem::InputImage { image_url } => {
+                                FunctionCallOutputContentItem::InputImage {
+                                    image_url: image_url.clone(),
+                                }
+                            }
+                        })
+                        .collect()
+                });
+                ResponseItem::FunctionCallOutput {
+                    call_id: call_id.clone(),
+                    output: FunctionCallOutputPayload {
+                        content: truncated,
+                        content_items: truncated_items,
+                        success: output.success,
+                    },
+                }
+            }
+            ResponseItem::CustomToolCallOutput { call_id, output } => {
+                let truncated = format_output_for_model_body(output);
+                ResponseItem::CustomToolCallOutput {
+                    call_id: call_id.clone(),
+                    output: truncated,
+                }
+            }
+            ResponseItem::Message { .. }
+            | ResponseItem::Reasoning { .. }
+            | ResponseItem::LocalShellCall { .. }
+            | ResponseItem::FunctionCall { .. }
+            | ResponseItem::WebSearchCall { .. }
+            | ResponseItem::CustomToolCall { .. }
+            | ResponseItem::GhostSnapshot { .. }
+            | ResponseItem::Other => item.clone(),
+        }
     }
 }
 
+pub(crate) fn format_output_for_model_body(content: &str) -> String {
+    // Head+tail truncation for the model: show the beginning and end with an elision.
+    // Clients still receive full streams; only this formatted summary is capped.
+    let total_lines = content.lines().count();
+    if content.len() <= MODEL_FORMAT_MAX_BYTES && total_lines <= MODEL_FORMAT_MAX_LINES {
+        return content.to_string();
+    }
+    let output = truncate_formatted_exec_output(content, total_lines);
+    format!("Total output lines: {total_lines}\n\n{output}")
+}
+
+fn truncate_formatted_exec_output(content: &str, total_lines: usize) -> String {
+    let segments: Vec<&str> = content.split_inclusive('\n').collect();
+    let head_take = MODEL_FORMAT_HEAD_LINES.min(segments.len());
+    let tail_take = MODEL_FORMAT_TAIL_LINES.min(segments.len().saturating_sub(head_take));
+    let omitted = segments.len().saturating_sub(head_take + tail_take);
+
+    let head_slice_end: usize = segments
+        .iter()
+        .take(head_take)
+        .map(|segment| segment.len())
+        .sum();
+    let tail_slice_start: usize = if tail_take == 0 {
+        content.len()
+    } else {
+        content.len()
+            - segments
+                .iter()
+                .rev()
+                .take(tail_take)
+                .map(|segment| segment.len())
+                .sum::<usize>()
+    };
+    let head_slice = &content[..head_slice_end];
+    let tail_slice = &content[tail_slice_start..];
+    let truncated_by_bytes = content.len() > MODEL_FORMAT_MAX_BYTES;
+    // this is a bit wrong. We are counting metadata lines and not just shell output lines.
+    let marker = if omitted > 0 {
+        Some(format!(
+            "\n[... omitted {omitted} of {total_lines} lines ...]\n\n"
+        ))
+    } else if truncated_by_bytes {
+        Some(format!(
+            "\n[... output truncated to fit {MODEL_FORMAT_MAX_BYTES} bytes ...]\n\n"
+        ))
+    } else {
+        None
+    };
+
+    let marker_len = marker.as_ref().map_or(0, String::len);
+    let base_head_budget = MODEL_FORMAT_HEAD_BYTES.min(MODEL_FORMAT_MAX_BYTES);
+    let head_budget = base_head_budget.min(MODEL_FORMAT_MAX_BYTES.saturating_sub(marker_len));
+    let head_part = take_bytes_at_char_boundary(head_slice, head_budget);
+    let mut result = String::with_capacity(MODEL_FORMAT_MAX_BYTES.min(content.len()));
+
+    result.push_str(head_part);
+    if let Some(marker_text) = marker.as_ref() {
+        result.push_str(marker_text);
+    }
+
+    let remaining = MODEL_FORMAT_MAX_BYTES.saturating_sub(result.len());
+    if remaining == 0 {
+        return result;
+    }
+
+    let tail_part = take_last_bytes_at_char_boundary(tail_slice, remaining);
+    result.push_str(tail_part);
+
+    result
+}
+
 #[inline]
 fn error_or_panic(message: String) {
     if cfg!(debug_assertions) || env!("CARGO_PKG_VERSION").contains("alpha") {
@@ -355,6 +490,7 @@ fn is_api_message(message: &ResponseItem) -> bool {
         | ResponseItem::LocalShellCall { .. }
         | ResponseItem::Reasoning { .. }
         | ResponseItem::WebSearchCall { .. } => true,
+        ResponseItem::GhostSnapshot { .. } => false,
         ResponseItem::Other => false,
     }
 }
@@ -448,7 +584,7 @@ mod tests {
                 call_id: "call-1".to_string(),
                 output: FunctionCallOutputPayload {
                     content: "ok".to_string(),
-                    success: None,
+                    ..Default::default()
                 },
             },
         ];
@@ -464,7 +600,7 @@ mod tests {
                 call_id: "call-2".to_string(),
                 output: FunctionCallOutputPayload {
                     content: "ok".to_string(),
-                    success: None,
+                    ..Default::default()
                 },
             },
             ResponseItem::FunctionCall {
@@ -498,7 +634,7 @@ mod tests {
                 call_id: "call-3".to_string(),
                 output: FunctionCallOutputPayload {
                     content: "ok".to_string(),
-                    success: None,
+                    ..Default::default()
                 },
             },
         ];
@@ -527,6 +663,184 @@ mod tests {
         assert_eq!(h.contents(), vec![]);
     }
 
+    #[test]
+    fn record_items_truncates_function_call_output_content() {
+        let mut history = ConversationHistory::new();
+        let long_line = "a very long line to trigger truncation\n";
+        let long_output = long_line.repeat(2_500);
+        let item = ResponseItem::FunctionCallOutput {
+            call_id: "call-100".to_string(),
+            output: FunctionCallOutputPayload {
+                content: long_output.clone(),
+                success: Some(true),
+                ..Default::default()
+            },
+        };
+
+        history.record_items([&item]);
+
+        assert_eq!(history.items.len(), 1);
+        match &history.items[0] {
+            ResponseItem::FunctionCallOutput { output, .. } => {
+                assert_ne!(output.content, long_output);
+                assert!(
+                    output.content.starts_with("Total output lines:"),
+                    "expected truncated summary, got {}",
+                    output.content
+                );
+            }
+            other => panic!("unexpected history item: {other:?}"),
+        }
+    }
+
+    #[test]
+    fn record_items_truncates_custom_tool_call_output_content() {
+        let mut history = ConversationHistory::new();
+        let line = "custom output that is very long\n";
+        let long_output = line.repeat(2_500);
+        let item = ResponseItem::CustomToolCallOutput {
+            call_id: "tool-200".to_string(),
+            output: long_output.clone(),
+        };
+
+        history.record_items([&item]);
+
+        assert_eq!(history.items.len(), 1);
+        match &history.items[0] {
+            ResponseItem::CustomToolCallOutput { output, .. } => {
+                assert_ne!(output, &long_output);
+                assert!(
+                    output.starts_with("Total output lines:"),
+                    "expected truncated summary, got {output}"
+                );
+            }
+            other => panic!("unexpected history item: {other:?}"),
+        }
+    }
+
+    // The following tests were adapted from tools::mod truncation tests to
+    // target the new truncation functions in conversation_history.
+
+    use regex_lite::Regex;
+
+    fn assert_truncated_message_matches(message: &str, line: &str, total_lines: usize) {
+        let pattern = truncated_message_pattern(line, total_lines);
+        let regex = Regex::new(&pattern).unwrap_or_else(|err| {
+            panic!("failed to compile regex {pattern}: {err}");
+        });
+        let captures = regex
+            .captures(message)
+            .unwrap_or_else(|| panic!("message failed to match pattern {pattern}: {message}"));
+        let body = captures
+            .name("body")
+            .expect("missing body capture")
+            .as_str();
+        assert!(
+            body.len() <= MODEL_FORMAT_MAX_BYTES,
+            "body exceeds byte limit: {} bytes",
+            body.len()
+        );
+    }
+
+    fn truncated_message_pattern(line: &str, total_lines: usize) -> String {
+        let head_take = MODEL_FORMAT_HEAD_LINES.min(total_lines);
+        let tail_take = MODEL_FORMAT_TAIL_LINES.min(total_lines.saturating_sub(head_take));
+        let omitted = total_lines.saturating_sub(head_take + tail_take);
+        let escaped_line = regex_lite::escape(line);
+        if omitted == 0 {
+            return format!(
+                r"(?s)^Total output lines: {total_lines}\n\n(?P<body>{escaped_line}.*\n\[\.{{3}} output truncated to fit {MODEL_FORMAT_MAX_BYTES} bytes \.{{3}}]\n\n.*)$",
+            );
+        }
+        format!(
+            r"(?s)^Total output lines: {total_lines}\n\n(?P<body>{escaped_line}.*\n\[\.{{3}} omitted {omitted} of {total_lines} lines \.{{3}}]\n\n.*)$",
+        )
+    }
+
+    #[test]
+    fn format_exec_output_truncates_large_error() {
+        let line = "very long execution error line that should trigger truncation\n";
+        let large_error = line.repeat(2_500); // way beyond both byte and line limits
+
+        let truncated = format_output_for_model_body(&large_error);
+
+        let total_lines = large_error.lines().count();
+        assert_truncated_message_matches(&truncated, line, total_lines);
+        assert_ne!(truncated, large_error);
+    }
+
+    #[test]
+    fn format_exec_output_marks_byte_truncation_without_omitted_lines() {
+        let long_line = "a".repeat(MODEL_FORMAT_MAX_BYTES + 50);
+        let truncated = format_output_for_model_body(&long_line);
+
+        assert_ne!(truncated, long_line);
+        let marker_line =
+            format!("[... output truncated to fit {MODEL_FORMAT_MAX_BYTES} bytes ...]");
+        assert!(
+            truncated.contains(&marker_line),
+            "missing byte truncation marker: {truncated}"
+        );
+        assert!(
+            !truncated.contains("omitted"),
+            "line omission marker should not appear when no lines were dropped: {truncated}"
+        );
+    }
+
+    #[test]
+    fn format_exec_output_returns_original_when_within_limits() {
+        let content = "example output\n".repeat(10);
+
+        assert_eq!(format_output_for_model_body(&content), content);
+    }
+
+    #[test]
+    fn format_exec_output_reports_omitted_lines_and_keeps_head_and_tail() {
+        let total_lines = MODEL_FORMAT_MAX_LINES + 100;
+        let content: String = (0..total_lines)
+            .map(|idx| format!("line-{idx}\n"))
+            .collect();
+
+        let truncated = format_output_for_model_body(&content);
+        let omitted = total_lines - MODEL_FORMAT_MAX_LINES;
+        let expected_marker = format!("[... omitted {omitted} of {total_lines} lines ...]");
+
+        assert!(
+            truncated.contains(&expected_marker),
+            "missing omitted marker: {truncated}"
+        );
+        assert!(
+            truncated.contains("line-0\n"),
+            "expected head line to remain: {truncated}"
+        );
+
+        let last_line = format!("line-{}\n", total_lines - 1);
+        assert!(
+            truncated.contains(&last_line),
+            "expected tail line to remain: {truncated}"
+        );
+    }
+
+    #[test]
+    fn format_exec_output_prefers_line_marker_when_both_limits_exceeded() {
+        let total_lines = MODEL_FORMAT_MAX_LINES + 42;
+        let long_line = "x".repeat(256);
+        let content: String = (0..total_lines)
+            .map(|idx| format!("line-{idx}-{long_line}\n"))
+            .collect();
+
+        let truncated = format_output_for_model_body(&content);
+
+        assert!(
+            truncated.contains("[... omitted 42 of 298 lines ...]"),
+            "expected omitted marker when line count exceeds limit: {truncated}"
+        );
+        assert!(
+            !truncated.contains("output truncated to fit"),
+            "line omission marker should take precedence over byte marker: {truncated}"
+        );
+    }
+
     //TODO(aibrahim): run CI in release mode.
     #[cfg(not(debug_assertions))]
     #[test]
@@ -554,7 +868,7 @@ mod tests {
                     call_id: "call-x".to_string(),
                     output: FunctionCallOutputPayload {
                         content: "aborted".to_string(),
-                        success: None,
+                        ..Default::default()
                     },
                 },
             ]
@@ -631,7 +945,7 @@ mod tests {
                     call_id: "shell-1".to_string(),
                     output: FunctionCallOutputPayload {
                         content: "aborted".to_string(),
-                        success: None,
+                        ..Default::default()
                     },
                 },
             ]
@@ -645,7 +959,7 @@ mod tests {
             call_id: "orphan-1".to_string(),
             output: FunctionCallOutputPayload {
                 content: "ok".to_string(),
-                success: None,
+                ..Default::default()
             },
         }];
         let mut h = create_history_with_items(items);
@@ -685,7 +999,7 @@ mod tests {
                 call_id: "c2".to_string(),
                 output: FunctionCallOutputPayload {
                     content: "ok".to_string(),
-                    success: None,
+                    ..Default::default()
                 },
             },
             // Will get an inserted custom tool output
@@ -727,7 +1041,7 @@ mod tests {
                     call_id: "c1".to_string(),
                     output: FunctionCallOutputPayload {
                         content: "aborted".to_string(),
-                        success: None,
+                        ..Default::default()
                     },
                 },
                 ResponseItem::CustomToolCall {
@@ -757,7 +1071,7 @@ mod tests {
                     call_id: "s1".to_string(),
                     output: FunctionCallOutputPayload {
                         content: "aborted".to_string(),
-                        success: None,
+                        ..Default::default()
                     },
                 },
             ]
@@ -822,7 +1136,7 @@ mod tests {
             call_id: "orphan-1".to_string(),
             output: FunctionCallOutputPayload {
                 content: "ok".to_string(),
-                success: None,
+                ..Default::default()
             },
         }];
         let mut h = create_history_with_items(items);
@@ -856,7 +1170,7 @@ mod tests {
                 call_id: "c2".to_string(),
                 output: FunctionCallOutputPayload {
                     content: "ok".to_string(),
-                    success: None,
+                    ..Default::default()
                 },
             },
             ResponseItem::CustomToolCall {

codex-rs/core/src/conversation_manager.rs

@@ -98,7 +98,10 @@ impl ConversationManager {
             }
         };
 
-        let conversation = Arc::new(CodexConversation::new(codex));
+        let conversation = Arc::new(CodexConversation::new(
+            codex,
+            session_configured.rollout_path.clone(),
+        ));
         self.conversations
             .write()
             .await

codex-rs/core/src/error.rs

@@ -55,7 +55,7 @@ pub enum SandboxErr {
 #[derive(Error, Debug)]
 pub enum CodexErr {
     // todo(aibrahim): git rid of this error carrying the dangling artifacts
-    #[error("turn aborted")]
+    #[error("turn aborted. Something went wrong? Hit `/feedback` to report the issue.")]
     TurnAborted {
         dangling_artifacts: Vec<ProcessedResponseItem>,
     },
@@ -91,7 +91,7 @@ pub enum CodexErr {
 
     /// Returned by run_command_stream when the user pressed Ctrl‑C (SIGINT). Session uses this to
     /// surface a polite FunctionCallOutput back to the model instead of crashing the CLI.
-    #[error("interrupted (Ctrl-C)")]
+    #[error("interrupted (Ctrl-C). Something went wrong? Hit `/feedback` to report the issue.")]
     Interrupted,
 
     /// Unexpected HTTP status code.

codex-rs/core/src/features.rs

@@ -39,6 +39,10 @@ pub enum Feature {
     ViewImageTool,
     /// Allow the model to request web searches.
     WebSearchRequest,
+    /// Enable the model-based risk assessments for sandboxed commands.
+    SandboxCommandAssessment,
+    /// Create a ghost commit at each turn.
+    GhostCommit,
 }
 
 impl Feature {
@@ -73,6 +77,7 @@ pub struct FeatureOverrides {
     pub include_apply_patch_tool: Option<bool>,
     pub include_view_image_tool: Option<bool>,
     pub web_search_request: Option<bool>,
+    pub experimental_sandbox_command_assessment: Option<bool>,
 }
 
 impl FeatureOverrides {
@@ -137,6 +142,7 @@ impl Features {
         let mut features = Features::with_defaults();
 
         let base_legacy = LegacyFeatureToggles {
+            experimental_sandbox_command_assessment: cfg.experimental_sandbox_command_assessment,
             experimental_use_freeform_apply_patch: cfg.experimental_use_freeform_apply_patch,
             experimental_use_exec_command_tool: cfg.experimental_use_exec_command_tool,
             experimental_use_unified_exec_tool: cfg.experimental_use_unified_exec_tool,
@@ -154,6 +160,8 @@ impl Features {
         let profile_legacy = LegacyFeatureToggles {
             include_apply_patch_tool: config_profile.include_apply_patch_tool,
             include_view_image_tool: config_profile.include_view_image_tool,
+            experimental_sandbox_command_assessment: config_profile
+                .experimental_sandbox_command_assessment,
             experimental_use_freeform_apply_patch: config_profile
                 .experimental_use_freeform_apply_patch,
             experimental_use_exec_command_tool: config_profile.experimental_use_exec_command_tool,
@@ -183,6 +191,11 @@ fn feature_for_key(key: &str) -> Option<Feature> {
     legacy::feature_for_key(key)
 }
 
+/// Returns `true` if the provided string matches a known feature toggle key.
+pub fn is_known_feature_key(key: &str) -> bool {
+    feature_for_key(key).is_some()
+}
+
 /// Deserializable features table for TOML.
 #[derive(Deserialize, Debug, Clone, Default, PartialEq)]
 pub struct FeaturesToml {
@@ -236,4 +249,16 @@ pub const FEATURES: &[FeatureSpec] = &[
         stage: Stage::Stable,
         default_enabled: false,
     },
+    FeatureSpec {
+        id: Feature::SandboxCommandAssessment,
+        key: "experimental_sandbox_command_assessment",
+        stage: Stage::Experimental,
+        default_enabled: false,
+    },
+    FeatureSpec {
+        id: Feature::GhostCommit,
+        key: "ghost_commit",
+        stage: Stage::Experimental,
+        default_enabled: false,
+    },
 ];

codex-rs/core/src/features/legacy.rs

@@ -9,6 +9,10 @@ struct Alias {
 }
 
 const ALIASES: &[Alias] = &[
+    Alias {
+        legacy_key: "experimental_sandbox_command_assessment",
+        feature: Feature::SandboxCommandAssessment,
+    },
     Alias {
         legacy_key: "experimental_use_unified_exec_tool",
         feature: Feature::UnifiedExec,
@@ -53,6 +57,7 @@ pub(crate) fn feature_for_key(key: &str) -> Option<Feature> {
 pub struct LegacyFeatureToggles {
     pub include_apply_patch_tool: Option<bool>,
     pub include_view_image_tool: Option<bool>,
+    pub experimental_sandbox_command_assessment: Option<bool>,
     pub experimental_use_freeform_apply_patch: Option<bool>,
     pub experimental_use_exec_command_tool: Option<bool>,
     pub experimental_use_unified_exec_tool: Option<bool>,
@@ -69,6 +74,12 @@ impl LegacyFeatureToggles {
             self.include_apply_patch_tool,
             "include_apply_patch_tool",
         );
+        set_if_some(
+            features,
+            Feature::SandboxCommandAssessment,
+            self.experimental_sandbox_command_assessment,
+            "experimental_sandbox_command_assessment",
+        );
         set_if_some(
             features,
             Feature::ApplyPatchFreeform,

codex-rs/core/src/lib.rs

@@ -77,6 +77,7 @@ pub use rollout::find_conversation_path_by_id_str;
 pub use rollout::list::ConversationItem;
 pub use rollout::list::ConversationsPage;
 pub use rollout::list::Cursor;
+pub use rollout::list::read_head_for_summary;
 mod function_tool;
 mod state;
 mod tasks;

codex-rs/core/src/mcp_tool_call.rs

@@ -35,6 +35,7 @@ pub(crate) async fn handle_mcp_tool_call(
                     output: FunctionCallOutputPayload {
                         content: format!("err: {e}"),
                         success: Some(false),
+                        ..Default::default()
                     },
                 };
             }

codex-rs/core/src/model_family.rs

@@ -54,6 +54,9 @@ pub struct ModelFamily {
     /// This is applied when computing the effective context window seen by
     /// consumers.
     pub effective_context_window_percent: i64,
+
+    /// If the model family supports setting the verbosity level when using Responses API.
+    pub support_verbosity: bool,
 }
 
 macro_rules! model_family {
@@ -73,6 +76,7 @@ macro_rules! model_family {
             base_instructions: BASE_INSTRUCTIONS.to_string(),
             experimental_supported_tools: Vec::new(),
             effective_context_window_percent: 95,
+            support_verbosity: false,
         };
         // apply overrides
         $(
@@ -128,10 +132,11 @@ pub fn find_family_for_model(slug: &str) -> Option<ModelFamily> {
                 "test_sync_tool".to_string(),
             ],
             supports_parallel_tool_calls: true,
+            support_verbosity: true,
         )
 
     // Internal models.
-    } else if slug.starts_with("codex-") {
+    } else if slug.starts_with("codex-exp-") {
         model_family!(
             slug, slug,
             supports_reasoning_summaries: true,
@@ -144,22 +149,25 @@ pub fn find_family_for_model(slug: &str) -> Option<ModelFamily> {
                 "read_file".to_string(),
             ],
             supports_parallel_tool_calls: true,
+            support_verbosity: true,
         )
 
     // Production models.
-    } else if slug.starts_with("gpt-5-codex") {
+    } else if slug.starts_with("gpt-5-codex") || slug.starts_with("codex-") {
         model_family!(
             slug, slug,
             supports_reasoning_summaries: true,
             reasoning_summary_format: ReasoningSummaryFormat::Experimental,
             base_instructions: GPT_5_CODEX_INSTRUCTIONS.to_string(),
             apply_patch_tool_type: Some(ApplyPatchToolType::Freeform),
+            support_verbosity: true,
         )
     } else if slug.starts_with("gpt-5") {
         model_family!(
             slug, "gpt-5",
             supports_reasoning_summaries: true,
             needs_special_apply_patch_instructions: true,
+            support_verbosity: true,
         )
     } else {
         None
@@ -179,5 +187,6 @@ pub fn derive_default_model_family(model: &str) -> ModelFamily {
         base_instructions: BASE_INSTRUCTIONS.to_string(),
         experimental_supported_tools: Vec::new(),
         effective_context_window_percent: 95,
+        support_verbosity: false,
     }
 }

codex-rs/core/src/openai_model_info.rs

@@ -37,8 +37,10 @@ impl ModelInfo {
 }
 
 pub(crate) fn get_model_info(model_family: &ModelFamily) -> Option<ModelInfo> {
-    let slug = model_family.slug.as_str();
-    match slug {
+    let raw_slug = model_family.slug.as_str();
+    let slug = raw_slug.strip_prefix("openai/").unwrap_or(raw_slug);
+    let normalized_slug = slug.replace(':', "-");
+    match normalized_slug.as_str() {
         // OSS models have a 128k shared token pool.
         // Arbitrarily splitting it: 3/4 input context, 1/4 output.
         // https://openai.com/index/gpt-oss-model-card/

codex-rs/core/src/response_processing.rs

@@ -1,4 +1,5 @@
 use crate::codex::Session;
+use crate::codex::TurnContext;
 use crate::conversation_history::ConversationHistory;
 use codex_protocol::models::FunctionCallOutputPayload;
 use codex_protocol::models::ResponseInputItem;
@@ -13,6 +14,7 @@ pub(crate) async fn process_items(
     is_review_mode: bool,
     review_thread_history: &mut ConversationHistory,
     sess: &Session,
+    turn_context: &TurnContext,
 ) -> (Vec<ResponseInputItem>, Vec<ResponseItem>) {
     let mut items_to_record_in_conversation_history = Vec::<ResponseItem>::new();
     let mut responses = Vec::<ResponseInputItem>::new();
@@ -59,14 +61,11 @@ pub(crate) async fn process_items(
             ) => {
                 items_to_record_in_conversation_history.push(item);
                 let output = match result {
-                    Ok(call_tool_result) => {
-                        crate::codex::convert_call_tool_result_to_function_call_output_payload(
-                            call_tool_result,
-                        )
-                    }
+                    Ok(call_tool_result) => FunctionCallOutputPayload::from(call_tool_result),
                     Err(err) => FunctionCallOutputPayload {
                         content: err.clone(),
                         success: Some(false),
+                        ..Default::default()
                     },
                 };
                 items_to_record_in_conversation_history.push(ResponseItem::FunctionCallOutput {
@@ -104,7 +103,7 @@ pub(crate) async fn process_items(
         if is_review_mode {
             review_thread_history.record_items(items_to_record_in_conversation_history.iter());
         } else {
-            sess.record_conversation_items(&items_to_record_in_conversation_history)
+            sess.record_conversation_items(turn_context, &items_to_record_in_conversation_history)
                 .await;
         }
     }

codex-rs/core/src/rollout/list.rs

@@ -54,6 +54,7 @@ struct HeadTailSummary {
     saw_session_meta: bool,
     saw_user_event: bool,
     source: Option<SessionSource>,
+    model_provider: Option<String>,
     created_at: Option<String>,
     updated_at: Option<String>,
 }
@@ -109,6 +110,8 @@ pub(crate) async fn get_conversations(
     page_size: usize,
     cursor: Option<&Cursor>,
     allowed_sources: &[SessionSource],
+    model_providers: Option<&[String]>,
+    default_provider: &str,
 ) -> io::Result<ConversationsPage> {
     let mut root = codex_home.to_path_buf();
     root.push(SESSIONS_SUBDIR);
@@ -124,8 +127,17 @@ pub(crate) async fn get_conversations(
 
     let anchor = cursor.cloned();
 
-    let result =
-        traverse_directories_for_paths(root.clone(), page_size, anchor, allowed_sources).await?;
+    let provider_matcher =
+        model_providers.and_then(|filters| ProviderMatcher::new(filters, default_provider));
+
+    let result = traverse_directories_for_paths(
+        root.clone(),
+        page_size,
+        anchor,
+        allowed_sources,
+        provider_matcher.as_ref(),
+    )
+    .await?;
     Ok(result)
 }
 
@@ -145,6 +157,7 @@ async fn traverse_directories_for_paths(
     page_size: usize,
     anchor: Option<Cursor>,
     allowed_sources: &[SessionSource],
+    provider_matcher: Option<&ProviderMatcher<'_>>,
 ) -> io::Result<ConversationsPage> {
     let mut items: Vec<ConversationItem> = Vec::with_capacity(page_size);
     let mut scanned_files = 0usize;
@@ -153,6 +166,7 @@ async fn traverse_directories_for_paths(
         Some(c) => (c.ts, c.id),
         None => (OffsetDateTime::UNIX_EPOCH, Uuid::nil()),
     };
+    let mut more_matches_available = false;
 
     let year_dirs = collect_dirs_desc(&root, |s| s.parse::<u16>().ok()).await?;
 
@@ -184,6 +198,7 @@ async fn traverse_directories_for_paths(
                 for (ts, sid, _name_str, path) in day_files.into_iter() {
                     scanned_files += 1;
                     if scanned_files >= MAX_SCAN_FILES && items.len() >= page_size {
+                        more_matches_available = true;
                         break 'outer;
                     }
                     if !anchor_passed {
@@ -194,6 +209,7 @@ async fn traverse_directories_for_paths(
                         }
                     }
                     if items.len() == page_size {
+                        more_matches_available = true;
                         break 'outer;
                     }
                     // Read head and simultaneously detect message events within the same
@@ -208,6 +224,11 @@ async fn traverse_directories_for_paths(
                     {
                         continue;
                     }
+                    if let Some(matcher) = provider_matcher
+                        && !matcher.matches(summary.model_provider.as_deref())
+                    {
+                        continue;
+                    }
                     // Apply filters: must have session meta and at least one user message event
                     if summary.saw_session_meta && summary.saw_user_event {
                         let HeadTailSummary {
@@ -231,12 +252,21 @@ async fn traverse_directories_for_paths(
         }
     }
 
-    let next = build_next_cursor(&items);
+    let reached_scan_cap = scanned_files >= MAX_SCAN_FILES;
+    if reached_scan_cap && !items.is_empty() {
+        more_matches_available = true;
+    }
+
+    let next = if more_matches_available {
+        build_next_cursor(&items)
+    } else {
+        None
+    };
     Ok(ConversationsPage {
         items,
         next_cursor: next,
         num_scanned_files: scanned_files,
-        reached_scan_cap: scanned_files >= MAX_SCAN_FILES,
+        reached_scan_cap,
     })
 }
 
@@ -328,6 +358,32 @@ fn parse_timestamp_uuid_from_filename(name: &str) -> Option<(OffsetDateTime, Uui
     Some((ts, uuid))
 }
 
+struct ProviderMatcher<'a> {
+    filters: &'a [String],
+    matches_default_provider: bool,
+}
+
+impl<'a> ProviderMatcher<'a> {
+    fn new(filters: &'a [String], default_provider: &'a str) -> Option<Self> {
+        if filters.is_empty() {
+            return None;
+        }
+
+        let matches_default_provider = filters.iter().any(|provider| provider == default_provider);
+        Some(Self {
+            filters,
+            matches_default_provider,
+        })
+    }
+
+    fn matches(&self, session_provider: Option<&str>) -> bool {
+        match session_provider {
+            Some(provider) => self.filters.iter().any(|candidate| candidate == provider),
+            None => self.matches_default_provider,
+        }
+    }
+}
+
 async fn read_head_and_tail(
     path: &Path,
     head_limit: usize,
@@ -354,6 +410,7 @@ async fn read_head_and_tail(
         match rollout_line.item {
             RolloutItem::SessionMeta(session_meta_line) => {
                 summary.source = Some(session_meta_line.meta.source);
+                summary.model_provider = session_meta_line.meta.model_provider.clone();
                 summary.created_at = summary
                     .created_at
                     .clone()
@@ -394,6 +451,13 @@ async fn read_head_and_tail(
     Ok(summary)
 }
 
+/// Read up to `HEAD_RECORD_LIMIT` records from the start of the rollout file at `path`.
+/// This should be enough to produce a summary including the session meta line.
+pub async fn read_head_for_summary(path: &Path) -> io::Result<Vec<serde_json::Value>> {
+    let summary = read_head_and_tail(path, HEAD_RECORD_LIMIT, 0).await?;
+    Ok(summary.head)
+}
+
 async fn read_tail_records(
     path: &Path,
     max_records: usize,

codex-rs/core/src/rollout/policy.rs

@@ -26,7 +26,8 @@ pub(crate) fn should_persist_response_item(item: &ResponseItem) -> bool {
         | ResponseItem::FunctionCallOutput { .. }
         | ResponseItem::CustomToolCall { .. }
         | ResponseItem::CustomToolCallOutput { .. }
-        | ResponseItem::WebSearchCall { .. } => true,
+        | ResponseItem::WebSearchCall { .. }
+        | ResponseItem::GhostSnapshot { .. } => true,
         ResponseItem::Other => false,
     }
 }
@@ -42,6 +43,7 @@ pub(crate) fn should_persist_event_msg(ev: &EventMsg) -> bool {
         | EventMsg::TokenCount(_)
         | EventMsg::EnteredReviewMode(_)
         | EventMsg::ExitedReviewMode(_)
+        | EventMsg::UndoCompleted(_)
         | EventMsg::TurnAborted(_) => true,
         EventMsg::Error(_)
         | EventMsg::TaskStarted(_)
@@ -50,6 +52,7 @@ pub(crate) fn should_persist_event_msg(ev: &EventMsg) -> bool {
         | EventMsg::AgentReasoningDelta(_)
         | EventMsg::AgentReasoningRawContentDelta(_)
         | EventMsg::AgentReasoningSectionBreak(_)
+        | EventMsg::RawResponseItem(_)
         | EventMsg::SessionConfigured(_)
         | EventMsg::McpToolCallBegin(_)
         | EventMsg::McpToolCallEnd(_)
@@ -66,12 +69,12 @@ pub(crate) fn should_persist_event_msg(ev: &EventMsg) -> bool {
         | EventMsg::PatchApplyEnd(_)
         | EventMsg::TurnDiff(_)
         | EventMsg::GetHistoryEntryResponse(_)
+        | EventMsg::UndoStarted(_)
         | EventMsg::McpListToolsResponse(_)
         | EventMsg::ListCustomPromptsResponse(_)
         | EventMsg::PlanUpdate(_)
         | EventMsg::ShutdownComplete
         | EventMsg::ViewImageToolCall(_)
-        | EventMsg::ConversationPath(_)
         | EventMsg::ItemStarted(_)
         | EventMsg::ItemCompleted(_) => false,
     }

codex-rs/core/src/rollout/recorder.rs

@@ -97,8 +97,18 @@ impl RolloutRecorder {
         page_size: usize,
         cursor: Option<&Cursor>,
         allowed_sources: &[SessionSource],
+        model_providers: Option<&[String]>,
+        default_provider: &str,
     ) -> std::io::Result<ConversationsPage> {
-        get_conversations(codex_home, page_size, cursor, allowed_sources).await
+        get_conversations(
+            codex_home,
+            page_size,
+            cursor,
+            allowed_sources,
+            model_providers,
+            default_provider,
+        )
+        .await
     }
 
     /// Attempt to create a new [`RolloutRecorder`]. If the sessions directory
@@ -137,6 +147,7 @@ impl RolloutRecorder {
                         cli_version: env!("CARGO_PKG_VERSION").to_string(),
                         instructions,
                         source,
+                        model_provider: Some(config.model_provider_id.clone()),
                     }),
                 )
             }
@@ -267,10 +278,6 @@ impl RolloutRecorder {
         }))
     }
 
-    pub(crate) fn get_rollout_path(&self) -> PathBuf {
-        self.rollout_path.clone()
-    }
-
     pub async fn shutdown(&self) -> std::io::Result<()> {
         let (tx_done, rx_done) = oneshot::channel();
         match self.tx.send(RolloutCmd::Shutdown { ack: tx_done }).await {

codex-rs/core/src/rollout/tests.rs

@@ -32,6 +32,14 @@ use codex_protocol::protocol::SessionSource;
 use codex_protocol::protocol::UserMessageEvent;
 
 const NO_SOURCE_FILTER: &[SessionSource] = &[];
+const TEST_PROVIDER: &str = "test-provider";
+
+fn provider_vec(providers: &[&str]) -> Vec<String> {
+    providers
+        .iter()
+        .map(std::string::ToString::to_string)
+        .collect()
+}
 
 fn write_session_file(
     root: &Path,
@@ -39,6 +47,24 @@ fn write_session_file(
     uuid: Uuid,
     num_records: usize,
     source: Option<SessionSource>,
+) -> std::io::Result<(OffsetDateTime, Uuid)> {
+    write_session_file_with_provider(
+        root,
+        ts_str,
+        uuid,
+        num_records,
+        source,
+        Some("test-provider"),
+    )
+}
+
+fn write_session_file_with_provider(
+    root: &Path,
+    ts_str: &str,
+    uuid: Uuid,
+    num_records: usize,
+    source: Option<SessionSource>,
+    model_provider: Option<&str>,
 ) -> std::io::Result<(OffsetDateTime, Uuid)> {
     let format: &[FormatItem] =
         format_description!("[year]-[month]-[day]T[hour]-[minute]-[second]");
@@ -68,6 +94,9 @@ fn write_session_file(
     if let Some(source) = source {
         payload["source"] = serde_json::to_value(source).unwrap();
     }
+    if let Some(provider) = model_provider {
+        payload["model_provider"] = serde_json::Value::String(provider.to_string());
+    }
 
     let meta = serde_json::json!({
         "timestamp": ts_str,
@@ -134,9 +163,17 @@ async fn test_list_conversations_latest_first() {
     )
     .unwrap();
 
-    let page = get_conversations(home, 10, None, INTERACTIVE_SESSION_SOURCES)
-        .await
-        .unwrap();
+    let provider_filter = provider_vec(&[TEST_PROVIDER]);
+    let page = get_conversations(
+        home,
+        10,
+        None,
+        INTERACTIVE_SESSION_SOURCES,
+        Some(provider_filter.as_slice()),
+        TEST_PROVIDER,
+    )
+    .await
+    .unwrap();
 
     // Build expected objects
     let p1 = home
@@ -166,6 +203,7 @@ async fn test_list_conversations_latest_first() {
         "originator": "test_originator",
         "cli_version": "test_version",
         "source": "vscode",
+        "model_provider": "test-provider",
     })];
     let head_2 = vec![serde_json::json!({
         "id": u2,
@@ -175,6 +213,7 @@ async fn test_list_conversations_latest_first() {
         "originator": "test_originator",
         "cli_version": "test_version",
         "source": "vscode",
+        "model_provider": "test-provider",
     })];
     let head_1 = vec![serde_json::json!({
         "id": u1,
@@ -184,11 +223,9 @@ async fn test_list_conversations_latest_first() {
         "originator": "test_originator",
         "cli_version": "test_version",
         "source": "vscode",
+        "model_provider": "test-provider",
     })];
 
-    let expected_cursor: Cursor =
-        serde_json::from_str(&format!("\"2025-01-01T12-00-00|{u1}\"")).unwrap();
-
     let expected = ConversationsPage {
         items: vec![
             ConversationItem {
@@ -213,7 +250,7 @@ async fn test_list_conversations_latest_first() {
                 updated_at: Some("2025-01-01T12-00-00".into()),
             },
         ],
-        next_cursor: Some(expected_cursor),
+        next_cursor: None,
         num_scanned_files: 3,
         reached_scan_cap: false,
     };
@@ -275,9 +312,17 @@ async fn test_pagination_cursor() {
     )
     .unwrap();
 
-    let page1 = get_conversations(home, 2, None, INTERACTIVE_SESSION_SOURCES)
-        .await
-        .unwrap();
+    let provider_filter = provider_vec(&[TEST_PROVIDER]);
+    let page1 = get_conversations(
+        home,
+        2,
+        None,
+        INTERACTIVE_SESSION_SOURCES,
+        Some(provider_filter.as_slice()),
+        TEST_PROVIDER,
+    )
+    .await
+    .unwrap();
     let p5 = home
         .join("sessions")
         .join("2025")
@@ -298,6 +343,7 @@ async fn test_pagination_cursor() {
         "originator": "test_originator",
         "cli_version": "test_version",
         "source": "vscode",
+        "model_provider": "test-provider",
     })];
     let head_4 = vec![serde_json::json!({
         "id": u4,
@@ -307,6 +353,7 @@ async fn test_pagination_cursor() {
         "originator": "test_originator",
         "cli_version": "test_version",
         "source": "vscode",
+        "model_provider": "test-provider",
     })];
     let expected_cursor1: Cursor =
         serde_json::from_str(&format!("\"2025-03-04T09-00-00|{u4}\"")).unwrap();
@@ -338,6 +385,8 @@ async fn test_pagination_cursor() {
         2,
         page1.next_cursor.as_ref(),
         INTERACTIVE_SESSION_SOURCES,
+        Some(provider_filter.as_slice()),
+        TEST_PROVIDER,
     )
     .await
     .unwrap();
@@ -361,6 +410,7 @@ async fn test_pagination_cursor() {
         "originator": "test_originator",
         "cli_version": "test_version",
         "source": "vscode",
+        "model_provider": "test-provider",
     })];
     let head_2 = vec![serde_json::json!({
         "id": u2,
@@ -370,6 +420,7 @@ async fn test_pagination_cursor() {
         "originator": "test_originator",
         "cli_version": "test_version",
         "source": "vscode",
+        "model_provider": "test-provider",
     })];
     let expected_cursor2: Cursor =
         serde_json::from_str(&format!("\"2025-03-02T09-00-00|{u2}\"")).unwrap();
@@ -401,6 +452,8 @@ async fn test_pagination_cursor() {
         2,
         page2.next_cursor.as_ref(),
         INTERACTIVE_SESSION_SOURCES,
+        Some(provider_filter.as_slice()),
+        TEST_PROVIDER,
     )
     .await
     .unwrap();
@@ -418,9 +471,8 @@ async fn test_pagination_cursor() {
         "originator": "test_originator",
         "cli_version": "test_version",
         "source": "vscode",
+        "model_provider": "test-provider",
     })];
-    let expected_cursor3: Cursor =
-        serde_json::from_str(&format!("\"2025-03-01T09-00-00|{u1}\"")).unwrap();
     let expected_page3 = ConversationsPage {
         items: vec![ConversationItem {
             path: p1,
@@ -429,7 +481,7 @@ async fn test_pagination_cursor() {
             created_at: Some("2025-03-01T09-00-00".into()),
             updated_at: Some("2025-03-01T09-00-00".into()),
         }],
-        next_cursor: Some(expected_cursor3),
+        next_cursor: None,
         num_scanned_files: 5, // scanned 05, 04 (anchor), 03, 02 (anchor), 01
         reached_scan_cap: false,
     };
@@ -445,9 +497,17 @@ async fn test_get_conversation_contents() {
     let ts = "2025-04-01T10-30-00";
     write_session_file(home, ts, uuid, 2, Some(SessionSource::VSCode)).unwrap();
 
-    let page = get_conversations(home, 1, None, INTERACTIVE_SESSION_SOURCES)
-        .await
-        .unwrap();
+    let provider_filter = provider_vec(&[TEST_PROVIDER]);
+    let page = get_conversations(
+        home,
+        1,
+        None,
+        INTERACTIVE_SESSION_SOURCES,
+        Some(provider_filter.as_slice()),
+        TEST_PROVIDER,
+    )
+    .await
+    .unwrap();
     let path = &page.items[0].path;
 
     let content = get_conversation(path).await.unwrap();
@@ -467,8 +527,8 @@ async fn test_get_conversation_contents() {
         "originator": "test_originator",
         "cli_version": "test_version",
         "source": "vscode",
+        "model_provider": "test-provider",
     })];
-    let expected_cursor: Cursor = serde_json::from_str(&format!("\"{ts}|{uuid}\"")).unwrap();
     let expected_page = ConversationsPage {
         items: vec![ConversationItem {
             path: expected_path,
@@ -477,7 +537,7 @@ async fn test_get_conversation_contents() {
             created_at: Some(ts.into()),
             updated_at: Some(ts.into()),
         }],
-        next_cursor: Some(expected_cursor),
+        next_cursor: None,
         num_scanned_files: 1,
         reached_scan_cap: false,
     };
@@ -495,6 +555,7 @@ async fn test_get_conversation_contents() {
             "originator": "test_originator",
             "cli_version": "test_version",
             "source": "vscode",
+            "model_provider": "test-provider",
         }
     });
     let user_event = serde_json::json!({
@@ -532,6 +593,7 @@ async fn test_tail_includes_last_response_items() -> Result<()> {
                 originator: "test_originator".into(),
                 cli_version: "test_version".into(),
                 source: SessionSource::VSCode,
+                model_provider: Some("test-provider".into()),
             },
             git: None,
         }),
@@ -563,7 +625,16 @@ async fn test_tail_includes_last_response_items() -> Result<()> {
     }
     drop(file);
 
-    let page = get_conversations(home, 1, None, INTERACTIVE_SESSION_SOURCES).await?;
+    let provider_filter = provider_vec(&[TEST_PROVIDER]);
+    let page = get_conversations(
+        home,
+        1,
+        None,
+        INTERACTIVE_SESSION_SOURCES,
+        Some(provider_filter.as_slice()),
+        TEST_PROVIDER,
+    )
+    .await?;
     let item = page.items.first().expect("conversation item");
     let tail_len = item.tail.len();
     assert_eq!(tail_len, 10usize.min(total_messages));
@@ -615,6 +686,7 @@ async fn test_tail_handles_short_sessions() -> Result<()> {
                 originator: "test_originator".into(),
                 cli_version: "test_version".into(),
                 source: SessionSource::VSCode,
+                model_provider: Some("test-provider".into()),
             },
             git: None,
         }),
@@ -645,7 +717,16 @@ async fn test_tail_handles_short_sessions() -> Result<()> {
     }
     drop(file);
 
-    let page = get_conversations(home, 1, None, INTERACTIVE_SESSION_SOURCES).await?;
+    let provider_filter = provider_vec(&[TEST_PROVIDER]);
+    let page = get_conversations(
+        home,
+        1,
+        None,
+        INTERACTIVE_SESSION_SOURCES,
+        Some(provider_filter.as_slice()),
+        TEST_PROVIDER,
+    )
+    .await?;
     let tail = &page.items.first().expect("conversation item").tail;
 
     assert_eq!(tail.len(), 3);
@@ -699,6 +780,7 @@ async fn test_tail_skips_trailing_non_responses() -> Result<()> {
                 originator: "test_originator".into(),
                 cli_version: "test_version".into(),
                 source: SessionSource::VSCode,
+                model_provider: Some("test-provider".into()),
             },
             git: None,
         }),
@@ -743,7 +825,16 @@ async fn test_tail_skips_trailing_non_responses() -> Result<()> {
     writeln!(file, "{}", serde_json::to_string(&shutdown_event)?)?;
     drop(file);
 
-    let page = get_conversations(home, 1, None, INTERACTIVE_SESSION_SOURCES).await?;
+    let provider_filter = provider_vec(&[TEST_PROVIDER]);
+    let page = get_conversations(
+        home,
+        1,
+        None,
+        INTERACTIVE_SESSION_SOURCES,
+        Some(provider_filter.as_slice()),
+        TEST_PROVIDER,
+    )
+    .await?;
     let tail = &page.items.first().expect("conversation item").tail;
 
     let expected: Vec<serde_json::Value> = (0..4)
@@ -785,9 +876,17 @@ async fn test_stable_ordering_same_second_pagination() {
     write_session_file(home, ts, u2, 0, Some(SessionSource::VSCode)).unwrap();
     write_session_file(home, ts, u3, 0, Some(SessionSource::VSCode)).unwrap();
 
-    let page1 = get_conversations(home, 2, None, INTERACTIVE_SESSION_SOURCES)
-        .await
-        .unwrap();
+    let provider_filter = provider_vec(&[TEST_PROVIDER]);
+    let page1 = get_conversations(
+        home,
+        2,
+        None,
+        INTERACTIVE_SESSION_SOURCES,
+        Some(provider_filter.as_slice()),
+        TEST_PROVIDER,
+    )
+    .await
+    .unwrap();
 
     let p3 = home
         .join("sessions")
@@ -810,6 +909,7 @@ async fn test_stable_ordering_same_second_pagination() {
             "originator": "test_originator",
             "cli_version": "test_version",
             "source": "vscode",
+            "model_provider": "test-provider",
         })]
     };
     let expected_cursor1: Cursor = serde_json::from_str(&format!("\"{ts}|{u2}\"")).unwrap();
@@ -841,6 +941,8 @@ async fn test_stable_ordering_same_second_pagination() {
         2,
         page1.next_cursor.as_ref(),
         INTERACTIVE_SESSION_SOURCES,
+        Some(provider_filter.as_slice()),
+        TEST_PROVIDER,
     )
     .await
     .unwrap();
@@ -850,7 +952,6 @@ async fn test_stable_ordering_same_second_pagination() {
         .join("07")
         .join("01")
         .join(format!("rollout-2025-07-01T00-00-00-{u1}.jsonl"));
-    let expected_cursor2: Cursor = serde_json::from_str(&format!("\"{ts}|{u1}\"")).unwrap();
     let expected_page2 = ConversationsPage {
         items: vec![ConversationItem {
             path: p1,
@@ -859,7 +960,7 @@ async fn test_stable_ordering_same_second_pagination() {
             created_at: Some(ts.to_string()),
             updated_at: Some(ts.to_string()),
         }],
-        next_cursor: Some(expected_cursor2),
+        next_cursor: None,
         num_scanned_files: 3, // scanned u3, u2 (anchor), u1
         reached_scan_cap: false,
     };
@@ -891,9 +992,17 @@ async fn test_source_filter_excludes_non_matching_sessions() {
     )
     .unwrap();
 
-    let interactive_only = get_conversations(home, 10, None, INTERACTIVE_SESSION_SOURCES)
-        .await
-        .unwrap();
+    let provider_filter = provider_vec(&[TEST_PROVIDER]);
+    let interactive_only = get_conversations(
+        home,
+        10,
+        None,
+        INTERACTIVE_SESSION_SOURCES,
+        Some(provider_filter.as_slice()),
+        TEST_PROVIDER,
+    )
+    .await
+    .unwrap();
     let paths: Vec<_> = interactive_only
         .items
         .iter()
@@ -905,7 +1014,7 @@ async fn test_source_filter_excludes_non_matching_sessions() {
         path.ends_with("rollout-2025-08-02T10-00-00-00000000-0000-0000-0000-00000000002a.jsonl")
     }));
 
-    let all_sessions = get_conversations(home, 10, None, NO_SOURCE_FILTER)
+    let all_sessions = get_conversations(home, 10, None, NO_SOURCE_FILTER, None, TEST_PROVIDER)
         .await
         .unwrap();
     let all_paths: Vec<_> = all_sessions
@@ -921,3 +1030,102 @@ async fn test_source_filter_excludes_non_matching_sessions() {
         path.ends_with("rollout-2025-08-01T10-00-00-00000000-0000-0000-0000-00000000004d.jsonl")
     }));
 }
+
+#[tokio::test]
+async fn test_model_provider_filter_selects_only_matching_sessions() -> Result<()> {
+    let temp = TempDir::new().unwrap();
+    let home = temp.path();
+
+    let openai_id = Uuid::from_u128(1);
+    let beta_id = Uuid::from_u128(2);
+    let none_id = Uuid::from_u128(3);
+
+    write_session_file_with_provider(
+        home,
+        "2025-09-01T12-00-00",
+        openai_id,
+        1,
+        Some(SessionSource::VSCode),
+        Some("openai"),
+    )?;
+    write_session_file_with_provider(
+        home,
+        "2025-09-01T11-00-00",
+        beta_id,
+        1,
+        Some(SessionSource::VSCode),
+        Some("beta"),
+    )?;
+    write_session_file_with_provider(
+        home,
+        "2025-09-01T10-00-00",
+        none_id,
+        1,
+        Some(SessionSource::VSCode),
+        None,
+    )?;
+
+    let openai_id_str = openai_id.to_string();
+    let none_id_str = none_id.to_string();
+    let openai_filter = provider_vec(&["openai"]);
+    let openai_sessions = get_conversations(
+        home,
+        10,
+        None,
+        NO_SOURCE_FILTER,
+        Some(openai_filter.as_slice()),
+        "openai",
+    )
+    .await?;
+    assert_eq!(openai_sessions.items.len(), 2);
+    let openai_ids: Vec<_> = openai_sessions
+        .items
+        .iter()
+        .filter_map(|item| {
+            item.head
+                .first()
+                .and_then(|value| value.get("id"))
+                .and_then(serde_json::Value::as_str)
+                .map(str::to_string)
+        })
+        .collect();
+    assert!(openai_ids.contains(&openai_id_str));
+    assert!(openai_ids.contains(&none_id_str));
+
+    let beta_filter = provider_vec(&["beta"]);
+    let beta_sessions = get_conversations(
+        home,
+        10,
+        None,
+        NO_SOURCE_FILTER,
+        Some(beta_filter.as_slice()),
+        "openai",
+    )
+    .await?;
+    assert_eq!(beta_sessions.items.len(), 1);
+    let beta_id_str = beta_id.to_string();
+    let beta_head = beta_sessions
+        .items
+        .first()
+        .and_then(|item| item.head.first())
+        .and_then(|value| value.get("id"))
+        .and_then(serde_json::Value::as_str);
+    assert_eq!(beta_head, Some(beta_id_str.as_str()));
+
+    let unknown_filter = provider_vec(&["unknown"]);
+    let unknown_sessions = get_conversations(
+        home,
+        10,
+        None,
+        NO_SOURCE_FILTER,
+        Some(unknown_filter.as_slice()),
+        "openai",
+    )
+    .await?;
+    assert!(unknown_sessions.items.is_empty());
+
+    let all_sessions = get_conversations(home, 10, None, NO_SOURCE_FILTER, None, "openai").await?;
+    assert_eq!(all_sessions.items.len(), 3);
+
+    Ok(())
+}

codex-rs/core/src/sandboxing/assessment.rs

@@ -0,0 +1,275 @@
+use std::path::Path;
+use std::path::PathBuf;
+use std::sync::Arc;
+use std::time::Duration;
+use std::time::Instant;
+
+use crate::AuthManager;
+use crate::ModelProviderInfo;
+use crate::client::ModelClient;
+use crate::client_common::Prompt;
+use crate::client_common::ResponseEvent;
+use crate::config::Config;
+use crate::protocol::SandboxPolicy;
+use askama::Template;
+use codex_otel::otel_event_manager::OtelEventManager;
+use codex_protocol::ConversationId;
+use codex_protocol::models::ContentItem;
+use codex_protocol::models::ResponseItem;
+use codex_protocol::protocol::SandboxCommandAssessment;
+use futures::StreamExt;
+use serde_json::json;
+use tokio::time::timeout;
+use tracing::warn;
+
+const SANDBOX_ASSESSMENT_TIMEOUT: Duration = Duration::from_secs(5);
+
+const SANDBOX_RISK_CATEGORY_VALUES: &[&str] = &[
+    "data_deletion",
+    "data_exfiltration",
+    "privilege_escalation",
+    "system_modification",
+    "network_access",
+    "resource_exhaustion",
+    "compliance",
+];
+
+#[derive(Template)]
+#[template(path = "sandboxing/assessment_prompt.md", escape = "none")]
+struct SandboxAssessmentPromptTemplate<'a> {
+    platform: &'a str,
+    sandbox_policy: &'a str,
+    filesystem_roots: Option<&'a str>,
+    working_directory: &'a str,
+    command_argv: &'a str,
+    command_joined: &'a str,
+    sandbox_failure_message: Option<&'a str>,
+}
+
+#[allow(clippy::too_many_arguments)]
+pub(crate) async fn assess_command(
+    config: Arc<Config>,
+    provider: ModelProviderInfo,
+    auth_manager: Arc<AuthManager>,
+    parent_otel: &OtelEventManager,
+    conversation_id: ConversationId,
+    call_id: &str,
+    command: &[String],
+    sandbox_policy: &SandboxPolicy,
+    cwd: &Path,
+    failure_message: Option<&str>,
+) -> Option<SandboxCommandAssessment> {
+    if !config.experimental_sandbox_command_assessment || command.is_empty() {
+        return None;
+    }
+
+    let command_json = serde_json::to_string(command).unwrap_or_else(|_| "[]".to_string());
+    let command_joined =
+        shlex::try_join(command.iter().map(String::as_str)).unwrap_or_else(|_| command.join(" "));
+    let failure = failure_message
+        .map(str::trim)
+        .filter(|msg| !msg.is_empty())
+        .map(str::to_string);
+
+    let cwd_str = cwd.to_string_lossy().to_string();
+    let sandbox_summary = summarize_sandbox_policy(sandbox_policy);
+    let mut roots = sandbox_roots_for_prompt(sandbox_policy, cwd);
+    roots.sort();
+    roots.dedup();
+
+    let platform = std::env::consts::OS;
+    let roots_formatted = roots.iter().map(|root| root.to_string_lossy().to_string());
+    let filesystem_roots = match roots_formatted.collect::<Vec<_>>() {
+        collected if collected.is_empty() => None,
+        collected => Some(collected.join(", ")),
+    };
+
+    let prompt_template = SandboxAssessmentPromptTemplate {
+        platform,
+        sandbox_policy: sandbox_summary.as_str(),
+        filesystem_roots: filesystem_roots.as_deref(),
+        working_directory: cwd_str.as_str(),
+        command_argv: command_json.as_str(),
+        command_joined: command_joined.as_str(),
+        sandbox_failure_message: failure.as_deref(),
+    };
+    let rendered_prompt = match prompt_template.render() {
+        Ok(rendered) => rendered,
+        Err(err) => {
+            warn!("failed to render sandbox assessment prompt: {err}");
+            return None;
+        }
+    };
+    let (system_prompt_section, user_prompt_section) = match rendered_prompt.split_once("\n---\n") {
+        Some(split) => split,
+        None => {
+            warn!("rendered sandbox assessment prompt missing separator");
+            return None;
+        }
+    };
+    let system_prompt = system_prompt_section
+        .strip_prefix("System Prompt:\n")
+        .unwrap_or(system_prompt_section)
+        .trim()
+        .to_string();
+    let user_prompt = user_prompt_section
+        .strip_prefix("User Prompt:\n")
+        .unwrap_or(user_prompt_section)
+        .trim()
+        .to_string();
+
+    let prompt = Prompt {
+        input: vec![ResponseItem::Message {
+            id: None,
+            role: "user".to_string(),
+            content: vec![ContentItem::InputText { text: user_prompt }],
+        }],
+        tools: Vec::new(),
+        parallel_tool_calls: false,
+        base_instructions_override: Some(system_prompt),
+        output_schema: Some(sandbox_assessment_schema()),
+    };
+
+    let child_otel =
+        parent_otel.with_model(config.model.as_str(), config.model_family.slug.as_str());
+
+    let client = ModelClient::new(
+        Arc::clone(&config),
+        Some(auth_manager),
+        child_otel,
+        provider,
+        config.model_reasoning_effort,
+        config.model_reasoning_summary,
+        conversation_id,
+    );
+
+    let start = Instant::now();
+    let assessment_result = timeout(SANDBOX_ASSESSMENT_TIMEOUT, async move {
+        let mut stream = client.stream(&prompt).await?;
+        let mut last_json: Option<String> = None;
+        while let Some(event) = stream.next().await {
+            match event {
+                Ok(ResponseEvent::OutputItemDone(item)) => {
+                    if let Some(text) = response_item_text(&item) {
+                        last_json = Some(text);
+                    }
+                }
+                Ok(ResponseEvent::RateLimits(_)) => {}
+                Ok(ResponseEvent::Completed { .. }) => break,
+                Ok(_) => continue,
+                Err(err) => return Err(err),
+            }
+        }
+        Ok(last_json)
+    })
+    .await;
+    let duration = start.elapsed();
+    parent_otel.sandbox_assessment_latency(call_id, duration);
+
+    match assessment_result {
+        Ok(Ok(Some(raw))) => match serde_json::from_str::<SandboxCommandAssessment>(raw.trim()) {
+            Ok(assessment) => {
+                parent_otel.sandbox_assessment(
+                    call_id,
+                    "success",
+                    Some(assessment.risk_level),
+                    &assessment.risk_categories,
+                    duration,
+                );
+                return Some(assessment);
+            }
+            Err(err) => {
+                warn!("failed to parse sandbox assessment JSON: {err}");
+                parent_otel.sandbox_assessment(call_id, "parse_error", None, &[], duration);
+            }
+        },
+        Ok(Ok(None)) => {
+            warn!("sandbox assessment response did not include any message");
+            parent_otel.sandbox_assessment(call_id, "no_output", None, &[], duration);
+        }
+        Ok(Err(err)) => {
+            warn!("sandbox assessment failed: {err}");
+            parent_otel.sandbox_assessment(call_id, "model_error", None, &[], duration);
+        }
+        Err(_) => {
+            warn!("sandbox assessment timed out");
+            parent_otel.sandbox_assessment(call_id, "timeout", None, &[], duration);
+        }
+    }
+
+    None
+}
+
+fn summarize_sandbox_policy(policy: &SandboxPolicy) -> String {
+    match policy {
+        SandboxPolicy::DangerFullAccess => "danger-full-access".to_string(),
+        SandboxPolicy::ReadOnly => "read-only".to_string(),
+        SandboxPolicy::WorkspaceWrite { network_access, .. } => {
+            let network = if *network_access {
+                "network"
+            } else {
+                "no-network"
+            };
+            format!("workspace-write (network_access={network})")
+        }
+    }
+}
+
+fn sandbox_roots_for_prompt(policy: &SandboxPolicy, cwd: &Path) -> Vec<PathBuf> {
+    let mut roots = vec![cwd.to_path_buf()];
+    if let SandboxPolicy::WorkspaceWrite { writable_roots, .. } = policy {
+        roots.extend(writable_roots.iter().cloned());
+    }
+    roots
+}
+
+fn sandbox_assessment_schema() -> serde_json::Value {
+    json!({
+        "type": "object",
+        "required": ["description", "risk_level", "risk_categories"],
+        "properties": {
+            "description": {
+                "type": "string",
+                "minLength": 1,
+                "maxLength": 500
+            },
+            "risk_level": {
+                "type": "string",
+                "enum": ["low", "medium", "high"]
+            },
+            "risk_categories": {
+                "type": "array",
+                "items": {
+                    "type": "string",
+                    "enum": SANDBOX_RISK_CATEGORY_VALUES
+                }
+            }
+        },
+        "additionalProperties": false
+    })
+}
+
+fn response_item_text(item: &ResponseItem) -> Option<String> {
+    match item {
+        ResponseItem::Message { content, .. } => {
+            let mut buffers: Vec<&str> = Vec::new();
+            for segment in content {
+                match segment {
+                    ContentItem::InputText { text } | ContentItem::OutputText { text } => {
+                        if !text.is_empty() {
+                            buffers.push(text);
+                        }
+                    }
+                    ContentItem::InputImage { .. } => {}
+                }
+            }
+            if buffers.is_empty() {
+                None
+            } else {
+                Some(buffers.join("\n"))
+            }
+        }
+        ResponseItem::FunctionCallOutput { output, .. } => Some(output.content.clone()),
+        _ => None,
+    }
+}

codex-rs/core/src/sandboxing/mod.rs

@@ -5,6 +5,9 @@ Build platform wrappers and produce ExecEnv for execution. Owns low‑level
 sandbox placement and transformation of portable CommandSpec into a
 ready‑to‑spawn environment.
 */
+
+pub mod assessment;
+
 use crate::exec::ExecToolCallOutput;
 use crate::exec::SandboxType;
 use crate::exec::StdoutStream;

codex-rs/core/src/tasks/ghost_snapshot.rs

@@ -0,0 +1,110 @@
+use crate::codex::TurnContext;
+use crate::state::TaskKind;
+use crate::tasks::SessionTask;
+use crate::tasks::SessionTaskContext;
+use async_trait::async_trait;
+use codex_git_tooling::CreateGhostCommitOptions;
+use codex_git_tooling::GitToolingError;
+use codex_git_tooling::create_ghost_commit;
+use codex_protocol::models::ResponseItem;
+use codex_protocol::user_input::UserInput;
+use codex_utils_readiness::Readiness;
+use codex_utils_readiness::Token;
+use std::sync::Arc;
+use tokio_util::sync::CancellationToken;
+use tracing::info;
+use tracing::warn;
+
+pub(crate) struct GhostSnapshotTask {
+    token: Token,
+}
+
+#[async_trait]
+impl SessionTask for GhostSnapshotTask {
+    fn kind(&self) -> TaskKind {
+        TaskKind::Regular
+    }
+
+    async fn run(
+        self: Arc<Self>,
+        session: Arc<SessionTaskContext>,
+        ctx: Arc<TurnContext>,
+        _input: Vec<UserInput>,
+        cancellation_token: CancellationToken,
+    ) -> Option<String> {
+        tokio::task::spawn(async move {
+            let token = self.token;
+            let ctx_for_task = Arc::clone(&ctx);
+            let cancelled = tokio::select! {
+                _ = cancellation_token.cancelled() => true,
+                _ = async {
+                    let repo_path = ctx_for_task.cwd.clone();
+                    // Required to run in a dedicated blocking pool.
+                    match tokio::task::spawn_blocking(move || {
+                        let options = CreateGhostCommitOptions::new(&repo_path);
+                        create_ghost_commit(&options)
+                    })
+                    .await
+                    {
+                        Ok(Ok(ghost_commit)) => {
+                            info!("ghost snapshot blocking task finished");
+                            session
+                                .session
+                                .record_conversation_items(&ctx, &[ResponseItem::GhostSnapshot {
+                                    ghost_commit: ghost_commit.clone(),
+                                }])
+                                .await;
+                            info!("ghost commit captured: {}", ghost_commit.id());
+                        }
+                        Ok(Err(err)) => {
+                            warn!(
+                                sub_id = ctx_for_task.sub_id.as_str(),
+                                "failed to capture ghost snapshot: {err}"
+                            );
+                            let message = match err {
+                                GitToolingError::NotAGitRepository { .. } => {
+                                    "Snapshots disabled: current directory is not a Git repository."
+                                        .to_string()
+                                }
+                                _ => format!("Snapshots disabled after ghost snapshot error: {err}."),
+                            };
+                            session
+                                .session
+                                .notify_background_event(&ctx_for_task, message)
+                                .await;
+                        }
+                        Err(err) => {
+                            warn!(
+                                sub_id = ctx_for_task.sub_id.as_str(),
+                                "ghost snapshot task panicked: {err}"
+                            );
+                            let message =
+                                format!("Snapshots disabled after ghost snapshot panic: {err}.");
+                            session
+                                .session
+                                .notify_background_event(&ctx_for_task, message)
+                                .await;
+                        }
+                    }
+                } => false,
+            };
+
+            if cancelled {
+                info!("ghost snapshot task cancelled");
+            }
+
+            match ctx.tool_call_gate.mark_ready(token).await {
+                Ok(true) => info!("ghost snapshot gate marked ready"),
+                Ok(false) => warn!("ghost snapshot gate already ready"),
+                Err(err) => warn!("failed to mark ghost snapshot ready: {err}"),
+            }
+        });
+        None
+    }
+}
+
+impl GhostSnapshotTask {
+    pub(crate) fn new(token: Token) -> Self {
+        Self { token }
+    }
+}

codex-rs/core/src/tasks/mod.rs

@@ -1,6 +1,8 @@
 mod compact;
+mod ghost_snapshot;
 mod regular;
 mod review;
+mod undo;
 
 use std::sync::Arc;
 use std::time::Duration;
@@ -25,8 +27,10 @@ use crate::state::TaskKind;
 use codex_protocol::user_input::UserInput;
 
 pub(crate) use compact::CompactTask;
+pub(crate) use ghost_snapshot::GhostSnapshotTask;
 pub(crate) use regular::RegularTask;
 pub(crate) use review::ReviewTask;
+pub(crate) use undo::UndoTask;
 
 const GRACEFULL_INTERRUPTION_TIMEOUT_MS: u64 = 100;
 
@@ -46,10 +50,28 @@ impl SessionTaskContext {
     }
 }
 
+/// Async task that drives a [`Session`] turn.
+///
+/// Implementations encapsulate a specific Codex workflow (regular chat,
+/// reviews, ghost snapshots, etc.). Each task instance is owned by a
+/// [`Session`] and executed on a background Tokio task. The trait is
+/// intentionally small: implementers identify themselves via
+/// [`SessionTask::kind`], perform their work in [`SessionTask::run`], and may
+/// release resources in [`SessionTask::abort`].
 #[async_trait]
 pub(crate) trait SessionTask: Send + Sync + 'static {
+    /// Describes the type of work the task performs so the session can
+    /// surface it in telemetry and UI.
     fn kind(&self) -> TaskKind;
 
+    /// Executes the task until completion or cancellation.
+    ///
+    /// Implementations typically stream protocol events using `session` and
+    /// `ctx`, returning an optional final agent message when finished. The
+    /// provided `cancellation_token` is cancelled when the session requests an
+    /// abort; implementers should watch for it and terminate quickly once it
+    /// fires. Returning [`Some`] yields a final message that
+    /// [`Session::on_task_finished`] will emit to the client.
     async fn run(
         self: Arc<Self>,
         session: Arc<SessionTaskContext>,
@@ -58,6 +80,11 @@ pub(crate) trait SessionTask: Send + Sync + 'static {
         cancellation_token: CancellationToken,
     ) -> Option<String>;
 
+    /// Gives the task a chance to perform cleanup after an abort.
+    ///
+    /// The default implementation is a no-op; override this if additional
+    /// teardown or notifications are required once
+    /// [`Session::abort_all_tasks`] cancels the task.
     async fn abort(&self, session: Arc<SessionTaskContext>, ctx: Arc<TurnContext>) {
         let _ = (session, ctx);
     }

codex-rs/core/src/tasks/undo.rs

@@ -0,0 +1,117 @@
+use std::sync::Arc;
+
+use crate::codex::TurnContext;
+use crate::protocol::EventMsg;
+use crate::protocol::UndoCompletedEvent;
+use crate::protocol::UndoStartedEvent;
+use crate::state::TaskKind;
+use crate::tasks::SessionTask;
+use crate::tasks::SessionTaskContext;
+use async_trait::async_trait;
+use codex_git_tooling::restore_ghost_commit;
+use codex_protocol::models::ResponseItem;
+use codex_protocol::user_input::UserInput;
+use tokio_util::sync::CancellationToken;
+use tracing::error;
+use tracing::info;
+use tracing::warn;
+
+pub(crate) struct UndoTask;
+
+impl UndoTask {
+    pub(crate) fn new() -> Self {
+        Self
+    }
+}
+
+#[async_trait]
+impl SessionTask for UndoTask {
+    fn kind(&self) -> TaskKind {
+        TaskKind::Regular
+    }
+
+    async fn run(
+        self: Arc<Self>,
+        session: Arc<SessionTaskContext>,
+        ctx: Arc<TurnContext>,
+        _input: Vec<UserInput>,
+        cancellation_token: CancellationToken,
+    ) -> Option<String> {
+        let sess = session.clone_session();
+        sess.send_event(
+            ctx.as_ref(),
+            EventMsg::UndoStarted(UndoStartedEvent {
+                message: Some("Undo in progress...".to_string()),
+            }),
+        )
+        .await;
+
+        if cancellation_token.is_cancelled() {
+            sess.send_event(
+                ctx.as_ref(),
+                EventMsg::UndoCompleted(UndoCompletedEvent {
+                    success: false,
+                    message: Some("Undo cancelled.".to_string()),
+                }),
+            )
+            .await;
+            return None;
+        }
+
+        let mut history = sess.clone_history().await;
+        let mut items = history.get_history();
+        let mut completed = UndoCompletedEvent {
+            success: false,
+            message: None,
+        };
+
+        let Some((idx, ghost_commit)) =
+            items
+                .iter()
+                .enumerate()
+                .rev()
+                .find_map(|(idx, item)| match item {
+                    ResponseItem::GhostSnapshot { ghost_commit } => {
+                        Some((idx, ghost_commit.clone()))
+                    }
+                    _ => None,
+                })
+        else {
+            completed.message = Some("No ghost snapshot available to undo.".to_string());
+            sess.send_event(ctx.as_ref(), EventMsg::UndoCompleted(completed))
+                .await;
+            return None;
+        };
+
+        let commit_id = ghost_commit.id().to_string();
+        let repo_path = ctx.cwd.clone();
+        let restore_result =
+            tokio::task::spawn_blocking(move || restore_ghost_commit(&repo_path, &ghost_commit))
+                .await;
+
+        match restore_result {
+            Ok(Ok(())) => {
+                items.remove(idx);
+                sess.replace_history(items).await;
+                let short_id: String = commit_id.chars().take(7).collect();
+                info!(commit_id = commit_id, "Undo restored ghost snapshot");
+                completed.success = true;
+                completed.message = Some(format!("Undo restored snapshot {short_id}."));
+            }
+            Ok(Err(err)) => {
+                let message = format!("Failed to restore snapshot {commit_id}: {err}");
+                warn!("{message}");
+                completed.message = Some(message);
+            }
+            Err(err) => {
+                let message = format!("Failed to restore snapshot {commit_id}: {err}");
+                error!("{message}");
+                completed.message = Some(message);
+            }
+        }
+
+        sess.send_event(ctx.as_ref(), EventMsg::UndoCompleted(completed))
+            .await;
+        None
+    }
+}

codex-rs/core/src/tools/context.rs

@@ -5,6 +5,7 @@ use crate::tools::TELEMETRY_PREVIEW_MAX_LINES;
 use crate::tools::TELEMETRY_PREVIEW_TRUNCATION_NOTICE;
 use crate::turn_diff_tracker::TurnDiffTracker;
 use codex_otel::otel_event_manager::OtelEventManager;
+use codex_protocol::models::FunctionCallOutputContentItem;
 use codex_protocol::models::FunctionCallOutputPayload;
 use codex_protocol::models::ResponseInputItem;
 use codex_protocol::models::ShellToolCallParams;
@@ -65,7 +66,10 @@ impl ToolPayload {
 #[derive(Clone)]
 pub enum ToolOutput {
     Function {
+        // Plain text representation of the tool output.
         content: String,
+        // Some tool calls such as MCP calls may return structured content that can get parsed into an array of polymorphic content items.
+        content_items: Option<Vec<FunctionCallOutputContentItem>>,
         success: Option<bool>,
     },
     Mcp {
@@ -90,7 +94,11 @@ impl ToolOutput {
 
     pub fn into_response(self, call_id: &str, payload: &ToolPayload) -> ResponseInputItem {
         match self {
-            ToolOutput::Function { content, success } => {
+            ToolOutput::Function {
+                content,
+                content_items,
+                success,
+            } => {
                 if matches!(payload, ToolPayload::Custom { .. }) {
                     ResponseInputItem::CustomToolCallOutput {
                         call_id: call_id.to_string(),
@@ -99,7 +107,11 @@ impl ToolOutput {
                 } else {
                     ResponseInputItem::FunctionCallOutput {
                         call_id: call_id.to_string(),
-                        output: FunctionCallOutputPayload { content, success },
+                        output: FunctionCallOutputPayload {
+                            content,
+                            content_items,
+                            success,
+                        },
                     }
                 }
             }
@@ -163,6 +175,7 @@ mod tests {
         };
         let response = ToolOutput::Function {
             content: "patched".to_string(),
+            content_items: None,
             success: Some(true),
         }
         .into_response("call-42", &payload);
@@ -183,6 +196,7 @@ mod tests {
         };
         let response = ToolOutput::Function {
             content: "ok".to_string(),
+            content_items: None,
             success: Some(true),
         }
         .into_response("fn-1", &payload);
@@ -191,6 +205,7 @@ mod tests {
             ResponseInputItem::FunctionCallOutput { call_id, output } => {
                 assert_eq!(call_id, "fn-1");
                 assert_eq!(output.content, "ok");
+                assert!(output.content_items.is_none());
                 assert_eq!(output.success, Some(true));
             }
             other => panic!("expected FunctionCallOutput, got {other:?}"),

codex-rs/core/src/tools/events.rs

@@ -19,7 +19,6 @@ use std::path::Path;
 use std::path::PathBuf;
 use std::time::Duration;
 
-use super::format_exec_output;
 use super::format_exec_output_str;
 
 #[derive(Clone, Copy)]
@@ -146,7 +145,7 @@ impl ToolEmitter {
                     (*message).to_string(),
                     -1,
                     Duration::ZERO,
-                    format_exec_output(&message),
+                    message.clone(),
                 )
                 .await;
             }
@@ -241,7 +240,7 @@ impl ToolEmitter {
                     (*message).to_string(),
                     -1,
                     Duration::ZERO,
-                    format_exec_output(&message),
+                    message.clone(),
                 )
                 .await;
             }
@@ -277,7 +276,7 @@ impl ToolEmitter {
             }
             Err(ToolError::Codex(err)) => {
                 let message = format!("execution error: {err:?}");
-                let response = super::format_exec_output(&message);
+                let response = message.clone();
                 event = ToolEventStage::Failure(ToolEventFailure::Message(message));
                 Err(FunctionCallError::RespondToModel(response))
             }
@@ -289,9 +288,9 @@ impl ToolEmitter {
                 } else {
                     msg
                 };
-                let response = super::format_exec_output(&normalized);
-                event = ToolEventStage::Failure(ToolEventFailure::Message(normalized));
-                Err(FunctionCallError::RespondToModel(response))
+                let response = &normalized;
+                event = ToolEventStage::Failure(ToolEventFailure::Message(normalized.clone()));
+                Err(FunctionCallError::RespondToModel(response.clone()))
             }
         };
         self.emit(ctx, event).await;

codex-rs/core/src/tools/handlers/apply_patch.rs

@@ -82,6 +82,7 @@ impl ToolHandler for ApplyPatchHandler {
                         let content = item?;
                         Ok(ToolOutput::Function {
                             content,
+                            content_items: None,
                             success: Some(true),
                         })
                     }
@@ -126,6 +127,7 @@ impl ToolHandler for ApplyPatchHandler {
                         let content = emitter.finish(event_ctx, out).await?;
                         Ok(ToolOutput::Function {
                             content,
+                            content_items: None,
                             success: Some(true),
                         })
                     }

codex-rs/core/src/tools/handlers/grep_files.rs

@@ -90,11 +90,13 @@ impl ToolHandler for GrepFilesHandler {
         if search_results.is_empty() {
             Ok(ToolOutput::Function {
                 content: "No matches found.".to_string(),
+                content_items: None,
                 success: Some(false),
             })
         } else {
             Ok(ToolOutput::Function {
                 content: search_results.join("\n"),
+                content_items: None,
                 success: Some(true),
             })
         }

codex-rs/core/src/tools/handlers/list_dir.rs

@@ -106,6 +106,7 @@ impl ToolHandler for ListDirHandler {
         output.extend(entries);
         Ok(ToolOutput::Function {
             content: output.join("\n"),
+            content_items: None,
             success: Some(true),
         })
     }

codex-rs/core/src/tools/handlers/mcp.rs

@@ -56,8 +56,16 @@ impl ToolHandler for McpHandler {
                 Ok(ToolOutput::Mcp { result })
             }
             codex_protocol::models::ResponseInputItem::FunctionCallOutput { output, .. } => {
-                let codex_protocol::models::FunctionCallOutputPayload { content, success } = output;
-                Ok(ToolOutput::Function { content, success })
+                let codex_protocol::models::FunctionCallOutputPayload {
+                    content,
+                    content_items,
+                    success,
+                } = output;
+                Ok(ToolOutput::Function {
+                    content,
+                    content_items,
+                    success,
+                })
             }
             _ => Err(FunctionCallError::RespondToModel(
                 "mcp handler received unexpected response variant".to_string(),

codex-rs/core/src/tools/handlers/mcp_resource.rs

@@ -297,7 +297,10 @@ async fn handle_list_resources(
     match payload_result {
         Ok(payload) => match serialize_function_output(payload) {
             Ok(output) => {
-                let ToolOutput::Function { content, success } = &output else {
+                let ToolOutput::Function {
+                    content, success, ..
+                } = &output
+                else {
                     unreachable!("MCP resource handler should return function output");
                 };
                 let duration = start.elapsed();
@@ -403,7 +406,10 @@ async fn handle_list_resource_templates(
     match payload_result {
         Ok(payload) => match serialize_function_output(payload) {
             Ok(output) => {
-                let ToolOutput::Function { content, success } = &output else {
+                let ToolOutput::Function {
+                    content, success, ..
+                } = &output
+                else {
                     unreachable!("MCP resource handler should return function output");
                 };
                 let duration = start.elapsed();
@@ -489,7 +495,10 @@ async fn handle_read_resource(
     match payload_result {
         Ok(payload) => match serialize_function_output(payload) {
             Ok(output) => {
-                let ToolOutput::Function { content, success } = &output else {
+                let ToolOutput::Function {
+                    content, success, ..
+                } = &output
+                else {
                     unreachable!("MCP resource handler should return function output");
                 };
                 let duration = start.elapsed();
@@ -618,6 +627,7 @@ where
 
     Ok(ToolOutput::Function {
         content,
+        content_items: None,
         success: Some(true),
     })
 }

codex-rs/core/src/tools/handlers/plan.rs

@@ -88,6 +88,7 @@ impl ToolHandler for PlanHandler {
 
         Ok(ToolOutput::Function {
             content,
+            content_items: None,
             success: Some(true),
         })
     }

codex-rs/core/src/tools/handlers/read_file.rs

@@ -149,6 +149,7 @@ impl ToolHandler for ReadFileHandler {
         };
         Ok(ToolOutput::Function {
             content: collected.join("\n"),
+            content_items: None,
             success: Some(true),
         })
     }

codex-rs/core/src/tools/handlers/shell.rs

@@ -136,6 +136,7 @@ impl ShellHandler {
                         let content = item?;
                         return Ok(ToolOutput::Function {
                             content,
+                            content_items: None,
                             success: Some(true),
                         });
                     }
@@ -179,6 +180,7 @@ impl ShellHandler {
                         let content = emitter.finish(event_ctx, out).await?;
                         return Ok(ToolOutput::Function {
                             content,
+                            content_items: None,
                             success: Some(true),
                         });
                     }
@@ -226,6 +228,7 @@ impl ShellHandler {
         let content = emitter.finish(event_ctx, out).await?;
         Ok(ToolOutput::Function {
             content,
+            content_items: None,
             success: Some(true),
         })
     }

codex-rs/core/src/tools/handlers/test_sync.rs

@@ -95,6 +95,7 @@ impl ToolHandler for TestSyncHandler {
 
         Ok(ToolOutput::Function {
             content: "ok".to_string(),
+            content_items: None,
             success: Some(true),
         })
     }

codex-rs/core/src/tools/handlers/unified_exec.rs

@@ -171,6 +171,7 @@ impl ToolHandler for UnifiedExecHandler {
 
         Ok(ToolOutput::Function {
             content,
+            content_items: None,
             success: Some(true),
         })
     }

codex-rs/core/src/tools/handlers/view_image.rs

@@ -85,6 +85,7 @@ impl ToolHandler for ViewImageHandler {
 
         Ok(ToolOutput::Function {
             content: "attached local image path".to_string(),
+            content_items: None,
             success: Some(true),
         })
     }

codex-rs/core/src/tools/mod.rs

@@ -9,19 +9,11 @@ pub mod runtimes;
 pub mod sandboxing;
 pub mod spec;
 
+use crate::conversation_history::format_output_for_model_body;
 use crate::exec::ExecToolCallOutput;
-use codex_utils_string::take_bytes_at_char_boundary;
-use codex_utils_string::take_last_bytes_at_char_boundary;
 pub use router::ToolRouter;
 use serde::Serialize;
 
-// Model-formatting limits: clients get full streams; only content sent to the model is truncated.
-pub(crate) const MODEL_FORMAT_MAX_BYTES: usize = 10 * 1024; // 10 KiB
-pub(crate) const MODEL_FORMAT_MAX_LINES: usize = 256; // lines
-pub(crate) const MODEL_FORMAT_HEAD_LINES: usize = MODEL_FORMAT_MAX_LINES / 2;
-pub(crate) const MODEL_FORMAT_TAIL_LINES: usize = MODEL_FORMAT_MAX_LINES - MODEL_FORMAT_HEAD_LINES; // 128
-pub(crate) const MODEL_FORMAT_HEAD_BYTES: usize = MODEL_FORMAT_MAX_BYTES / 2;
-
 // Telemetry preview limits: keep log events smaller than model budgets.
 pub(crate) const TELEMETRY_PREVIEW_MAX_BYTES: usize = 2 * 1024; // 2 KiB
 pub(crate) const TELEMETRY_PREVIEW_MAX_LINES: usize = 64; // lines
@@ -73,249 +65,15 @@ pub fn format_exec_output_str(exec_output: &ExecToolCallOutput) -> String {
 
     let content = aggregated_output.text.as_str();
 
-    if exec_output.timed_out {
-        let prefixed = format!(
+    let body = if exec_output.timed_out {
+        format!(
             "command timed out after {} milliseconds\n{content}",
             exec_output.duration.as_millis()
-        );
-        return format_exec_output(&prefixed);
-    }
-
-    format_exec_output(content)
-}
-
-pub(super) fn format_exec_output(content: &str) -> String {
-    // Head+tail truncation for the model: show the beginning and end with an elision.
-    // Clients still receive full streams; only this formatted summary is capped.
-    let total_lines = content.lines().count();
-    if content.len() <= MODEL_FORMAT_MAX_BYTES && total_lines <= MODEL_FORMAT_MAX_LINES {
-        return content.to_string();
-    }
-    let output = truncate_formatted_exec_output(content, total_lines);
-    format!("Total output lines: {total_lines}\n\n{output}")
-}
-
-fn truncate_formatted_exec_output(content: &str, total_lines: usize) -> String {
-    let segments: Vec<&str> = content.split_inclusive('\n').collect();
-    let head_take = MODEL_FORMAT_HEAD_LINES.min(segments.len());
-    let tail_take = MODEL_FORMAT_TAIL_LINES.min(segments.len().saturating_sub(head_take));
-    let omitted = segments.len().saturating_sub(head_take + tail_take);
-
-    let head_slice_end: usize = segments
-        .iter()
-        .take(head_take)
-        .map(|segment| segment.len())
-        .sum();
-    let tail_slice_start: usize = if tail_take == 0 {
-        content.len()
-    } else {
-        content.len()
-            - segments
-                .iter()
-                .rev()
-                .take(tail_take)
-                .map(|segment| segment.len())
-                .sum::<usize>()
-    };
-    let head_slice = &content[..head_slice_end];
-    let tail_slice = &content[tail_slice_start..];
-    let truncated_by_bytes = content.len() > MODEL_FORMAT_MAX_BYTES;
-    let marker = if omitted > 0 {
-        Some(format!(
-            "\n[... omitted {omitted} of {total_lines} lines ...]\n\n"
-        ))
-    } else if truncated_by_bytes {
-        Some(format!(
-            "\n[... output truncated to fit {MODEL_FORMAT_MAX_BYTES} bytes ...]\n\n"
-        ))
-    } else {
-        None
-    };
-
-    let marker_len = marker.as_ref().map_or(0, String::len);
-    let base_head_budget = MODEL_FORMAT_HEAD_BYTES.min(MODEL_FORMAT_MAX_BYTES);
-    let head_budget = base_head_budget.min(MODEL_FORMAT_MAX_BYTES.saturating_sub(marker_len));
-    let head_part = take_bytes_at_char_boundary(head_slice, head_budget);
-    let mut result = String::with_capacity(MODEL_FORMAT_MAX_BYTES.min(content.len()));
-
-    result.push_str(head_part);
-    if let Some(marker_text) = marker.as_ref() {
-        result.push_str(marker_text);
-    }
-
-    let remaining = MODEL_FORMAT_MAX_BYTES.saturating_sub(result.len());
-    if remaining == 0 {
-        return result;
-    }
-
-    let tail_part = take_last_bytes_at_char_boundary(tail_slice, remaining);
-    result.push_str(tail_part);
-
-    result
-}
-
-#[cfg(test)]
-mod tests {
-    use super::*;
-    use crate::function_tool::FunctionCallError;
-    use regex_lite::Regex;
-
-    fn truncate_function_error(err: FunctionCallError) -> FunctionCallError {
-        match err {
-            FunctionCallError::RespondToModel(msg) => {
-                FunctionCallError::RespondToModel(format_exec_output(&msg))
-            }
-            FunctionCallError::Denied(msg) => FunctionCallError::Denied(format_exec_output(&msg)),
-            FunctionCallError::Fatal(msg) => FunctionCallError::Fatal(format_exec_output(&msg)),
-            other => other,
-        }
-    }
-
-    fn assert_truncated_message_matches(message: &str, line: &str, total_lines: usize) {
-        let pattern = truncated_message_pattern(line, total_lines);
-        let regex = Regex::new(&pattern).unwrap_or_else(|err| {
-            panic!("failed to compile regex {pattern}: {err}");
-        });
-        let captures = regex
-            .captures(message)
-            .unwrap_or_else(|| panic!("message failed to match pattern {pattern}: {message}"));
-        let body = captures
-            .name("body")
-            .expect("missing body capture")
-            .as_str();
-        assert!(
-            body.len() <= MODEL_FORMAT_MAX_BYTES,
-            "body exceeds byte limit: {} bytes",
-            body.len()
-        );
-    }
-
-    fn truncated_message_pattern(line: &str, total_lines: usize) -> String {
-        let head_take = MODEL_FORMAT_HEAD_LINES.min(total_lines);
-        let tail_take = MODEL_FORMAT_TAIL_LINES.min(total_lines.saturating_sub(head_take));
-        let omitted = total_lines.saturating_sub(head_take + tail_take);
-        let escaped_line = regex_lite::escape(line);
-        if omitted == 0 {
-            return format!(
-                r"(?s)^Total output lines: {total_lines}\n\n(?P<body>{escaped_line}.*\n\[\.{{3}} output truncated to fit {MODEL_FORMAT_MAX_BYTES} bytes \.{{3}}]\n\n.*)$",
-            );
-        }
-        format!(
-            r"(?s)^Total output lines: {total_lines}\n\n(?P<body>{escaped_line}.*\n\[\.{{3}} omitted {omitted} of {total_lines} lines \.{{3}}]\n\n.*)$",
         )
-    }
+    } else {
+        content.to_string()
+    };
 
-    #[test]
-    fn truncate_formatted_exec_output_truncates_large_error() {
-        let line = "very long execution error line that should trigger truncation\n";
-        let large_error = line.repeat(2_500); // way beyond both byte and line limits
-
-        let truncated = format_exec_output(&large_error);
-
-        let total_lines = large_error.lines().count();
-        assert_truncated_message_matches(&truncated, line, total_lines);
-        assert_ne!(truncated, large_error);
-    }
-
-    #[test]
-    fn truncate_function_error_trims_respond_to_model() {
-        let line = "respond-to-model error that should be truncated\n";
-        let huge = line.repeat(3_000);
-        let total_lines = huge.lines().count();
-
-        let err = truncate_function_error(FunctionCallError::RespondToModel(huge));
-        match err {
-            FunctionCallError::RespondToModel(message) => {
-                assert_truncated_message_matches(&message, line, total_lines);
-            }
-            other => panic!("unexpected error variant: {other:?}"),
-        }
-    }
-
-    #[test]
-    fn truncate_function_error_trims_fatal() {
-        let line = "fatal error output that should be truncated\n";
-        let huge = line.repeat(3_000);
-        let total_lines = huge.lines().count();
-
-        let err = truncate_function_error(FunctionCallError::Fatal(huge));
-        match err {
-            FunctionCallError::Fatal(message) => {
-                assert_truncated_message_matches(&message, line, total_lines);
-            }
-            other => panic!("unexpected error variant: {other:?}"),
-        }
-    }
-
-    #[test]
-    fn truncate_formatted_exec_output_marks_byte_truncation_without_omitted_lines() {
-        let long_line = "a".repeat(MODEL_FORMAT_MAX_BYTES + 50);
-        let truncated = format_exec_output(&long_line);
-
-        assert_ne!(truncated, long_line);
-        let marker_line =
-            format!("[... output truncated to fit {MODEL_FORMAT_MAX_BYTES} bytes ...]");
-        assert!(
-            truncated.contains(&marker_line),
-            "missing byte truncation marker: {truncated}"
-        );
-        assert!(
-            !truncated.contains("omitted"),
-            "line omission marker should not appear when no lines were dropped: {truncated}"
-        );
-    }
-
-    #[test]
-    fn truncate_formatted_exec_output_returns_original_when_within_limits() {
-        let content = "example output\n".repeat(10);
-
-        assert_eq!(format_exec_output(&content), content);
-    }
-
-    #[test]
-    fn truncate_formatted_exec_output_reports_omitted_lines_and_keeps_head_and_tail() {
-        let total_lines = MODEL_FORMAT_MAX_LINES + 100;
-        let content: String = (0..total_lines)
-            .map(|idx| format!("line-{idx}\n"))
-            .collect();
-
-        let truncated = format_exec_output(&content);
-        let omitted = total_lines - MODEL_FORMAT_MAX_LINES;
-        let expected_marker = format!("[... omitted {omitted} of {total_lines} lines ...]");
-
-        assert!(
-            truncated.contains(&expected_marker),
-            "missing omitted marker: {truncated}"
-        );
-        assert!(
-            truncated.contains("line-0\n"),
-            "expected head line to remain: {truncated}"
-        );
-
-        let last_line = format!("line-{}\n", total_lines - 1);
-        assert!(
-            truncated.contains(&last_line),
-            "expected tail line to remain: {truncated}"
-        );
-    }
-
-    #[test]
-    fn truncate_formatted_exec_output_prefers_line_marker_when_both_limits_exceeded() {
-        let total_lines = MODEL_FORMAT_MAX_LINES + 42;
-        let long_line = "x".repeat(256);
-        let content: String = (0..total_lines)
-            .map(|idx| format!("line-{idx}-{long_line}\n"))
-            .collect();
-
-        let truncated = format_exec_output(&content);
-
-        assert!(
-            truncated.contains("[... omitted 42 of 298 lines ...]"),
-            "expected omitted marker when line count exceeds limit: {truncated}"
-        );
-        assert!(
-            !truncated.contains("output truncated to fit"),
-            "line omission marker should take precedence over byte marker: {truncated}"
-        );
-    }
+    // Truncate for model consumption before serialization.
+    format_output_for_model_body(&body)
 }

codex-rs/core/src/tools/orchestrator.rs

@@ -7,9 +7,11 @@ retry without sandbox on denial (no re‑approval thanks to caching).
 */
 use crate::error::CodexErr;
 use crate::error::SandboxErr;
+use crate::error::get_error_message_ui;
 use crate::exec::ExecToolCallOutput;
 use crate::sandboxing::SandboxManager;
 use crate::tools::sandboxing::ApprovalCtx;
+use crate::tools::sandboxing::ProvidesSandboxRetryData;
 use crate::tools::sandboxing::SandboxAttempt;
 use crate::tools::sandboxing::ToolCtx;
 use crate::tools::sandboxing::ToolError;
@@ -38,6 +40,7 @@ impl ToolOrchestrator {
     ) -> Result<Out, ToolError>
     where
         T: ToolRuntime<Rq, Out>,
+        Rq: ProvidesSandboxRetryData,
     {
         let otel = turn_ctx.client.get_otel_event_manager();
         let otel_tn = &tool_ctx.tool_name;
@@ -56,6 +59,7 @@ impl ToolOrchestrator {
                 turn: turn_ctx,
                 call_id: &tool_ctx.call_id,
                 retry_reason: None,
+                risk: None,
             };
             let decision = tool.start_approval_async(req, approval_ctx).await;
 
@@ -107,12 +111,33 @@ impl ToolOrchestrator {
 
                 // Ask for approval before retrying without sandbox.
                 if !tool.should_bypass_approval(approval_policy, already_approved) {
+                    let mut risk = None;
+
+                    if let Some(metadata) = req.sandbox_retry_data() {
+                        let err = SandboxErr::Denied {
+                            output: output.clone(),
+                        };
+                        let friendly = get_error_message_ui(&CodexErr::Sandbox(err));
+                        let failure_summary = format!("failed in sandbox: {friendly}");
+
+                        risk = tool_ctx
+                            .session
+                            .assess_sandbox_command(
+                                turn_ctx,
+                                &tool_ctx.call_id,
+                                &metadata.command,
+                                Some(failure_summary.as_str()),
+                            )
+                            .await;
+                    }
+
                     let reason_msg = build_denial_reason_from_output(output.as_ref());
                     let approval_ctx = ApprovalCtx {
                         session: tool_ctx.session,
                         turn: turn_ctx,
                         call_id: &tool_ctx.call_id,
                         retry_reason: Some(reason_msg),
+                        risk,
                     };
 
                     let decision = tool.start_approval_async(req, approval_ctx).await;

codex-rs/core/src/tools/parallel.rs

@@ -15,6 +15,7 @@ use crate::tools::router::ToolCall;
 use crate::tools::router::ToolRouter;
 use codex_protocol::models::FunctionCallOutputPayload;
 use codex_protocol::models::ResponseInputItem;
+use codex_utils_readiness::Readiness;
 
 pub(crate) struct ToolCallRuntime {
     router: Arc<ToolRouter>,
@@ -53,12 +54,16 @@ impl ToolCallRuntime {
         let tracker = Arc::clone(&self.tracker);
         let lock = Arc::clone(&self.parallel_execution);
         let aborted_response = Self::aborted_response(&call);
+        let readiness = self.turn_context.tool_call_gate.clone();
 
         let handle: AbortOnDropHandle<Result<ResponseInputItem, FunctionCallError>> =
             AbortOnDropHandle::new(tokio::spawn(async move {
                 tokio::select! {
                     _ = cancellation_token.cancelled() => Ok(aborted_response),
                     res = async {
+                        tracing::info!("waiting for tool gate");
+                        readiness.wait_ready().await;
+                        tracing::info!("tool gate released");
                         let _guard = if supports_parallel {
                             Either::Left(lock.read().await)
                         } else {
@@ -100,7 +105,7 @@ impl ToolCallRuntime {
                 call_id: call.call_id.clone(),
                 output: FunctionCallOutputPayload {
                     content: "aborted".to_string(),
-                    success: None,
+                    ..Default::default()
                 },
             },
         }

codex-rs/core/src/tools/router.rs

@@ -181,6 +181,7 @@ impl ToolRouter {
                 output: codex_protocol::models::FunctionCallOutputPayload {
                     content: message,
                     success: Some(false),
+                    ..Default::default()
                 },
             }
         }

codex-rs/core/src/tools/runtimes/apply_patch.rs

@@ -10,7 +10,9 @@ use crate::sandboxing::CommandSpec;
 use crate::sandboxing::execute_env;
 use crate::tools::sandboxing::Approvable;
 use crate::tools::sandboxing::ApprovalCtx;
+use crate::tools::sandboxing::ProvidesSandboxRetryData;
 use crate::tools::sandboxing::SandboxAttempt;
+use crate::tools::sandboxing::SandboxRetryData;
 use crate::tools::sandboxing::Sandboxable;
 use crate::tools::sandboxing::SandboxablePreference;
 use crate::tools::sandboxing::ToolCtx;
@@ -32,6 +34,12 @@ pub struct ApplyPatchRequest {
     pub codex_exe: Option<PathBuf>,
 }
 
+impl ProvidesSandboxRetryData for ApplyPatchRequest {
+    fn sandbox_retry_data(&self) -> Option<SandboxRetryData> {
+        None
+    }
+}
+
 #[derive(Default)]
 pub struct ApplyPatchRuntime;
 
@@ -106,9 +114,10 @@ impl Approvable<ApplyPatchRequest> for ApplyPatchRuntime {
         let call_id = ctx.call_id.to_string();
         let cwd = req.cwd.clone();
         let retry_reason = ctx.retry_reason.clone();
+        let risk = ctx.risk.clone();
         let user_explicitly_approved = req.user_explicitly_approved;
         Box::pin(async move {
-            with_cached_approval(&session.services, key, || async move {
+            with_cached_approval(&session.services, key, move || async move {
                 if let Some(reason) = retry_reason {
                     session
                         .request_command_approval(
@@ -117,6 +126,7 @@ impl Approvable<ApplyPatchRequest> for ApplyPatchRuntime {
                             vec!["apply_patch".to_string()],
                             cwd,
                             Some(reason),
+                            risk,
                         )
                         .await
                 } else if user_explicitly_approved {

codex-rs/core/src/tools/runtimes/shell.rs

@@ -12,7 +12,9 @@ use crate::sandboxing::execute_env;
 use crate::tools::runtimes::build_command_spec;
 use crate::tools::sandboxing::Approvable;
 use crate::tools::sandboxing::ApprovalCtx;
+use crate::tools::sandboxing::ProvidesSandboxRetryData;
 use crate::tools::sandboxing::SandboxAttempt;
+use crate::tools::sandboxing::SandboxRetryData;
 use crate::tools::sandboxing::Sandboxable;
 use crate::tools::sandboxing::SandboxablePreference;
 use crate::tools::sandboxing::ToolCtx;
@@ -34,6 +36,15 @@ pub struct ShellRequest {
     pub justification: Option<String>,
 }
 
+impl ProvidesSandboxRetryData for ShellRequest {
+    fn sandbox_retry_data(&self) -> Option<SandboxRetryData> {
+        Some(SandboxRetryData {
+            command: self.command.clone(),
+            cwd: self.cwd.clone(),
+        })
+    }
+}
+
 #[derive(Default)]
 pub struct ShellRuntime;
 
@@ -90,13 +101,14 @@ impl Approvable<ShellRequest> for ShellRuntime {
             .retry_reason
             .clone()
             .or_else(|| req.justification.clone());
+        let risk = ctx.risk.clone();
         let session = ctx.session;
         let turn = ctx.turn;
         let call_id = ctx.call_id.to_string();
         Box::pin(async move {
-            with_cached_approval(&session.services, key, || async move {
+            with_cached_approval(&session.services, key, move || async move {
                 session
-                    .request_command_approval(turn, call_id, command, cwd, reason)
+                    .request_command_approval(turn, call_id, command, cwd, reason, risk)
                     .await
             })
             .await

codex-rs/core/src/tools/runtimes/unified_exec.rs

@@ -9,7 +9,9 @@ use crate::error::SandboxErr;
 use crate::tools::runtimes::build_command_spec;
 use crate::tools::sandboxing::Approvable;
 use crate::tools::sandboxing::ApprovalCtx;
+use crate::tools::sandboxing::ProvidesSandboxRetryData;
 use crate::tools::sandboxing::SandboxAttempt;
+use crate::tools::sandboxing::SandboxRetryData;
 use crate::tools::sandboxing::Sandboxable;
 use crate::tools::sandboxing::SandboxablePreference;
 use crate::tools::sandboxing::ToolCtx;
@@ -31,6 +33,15 @@ pub struct UnifiedExecRequest {
     pub env: HashMap<String, String>,
 }
 
+impl ProvidesSandboxRetryData for UnifiedExecRequest {
+    fn sandbox_retry_data(&self) -> Option<SandboxRetryData> {
+        Some(SandboxRetryData {
+            command: self.command.clone(),
+            cwd: self.cwd.clone(),
+        })
+    }
+}
+
 #[derive(serde::Serialize, Clone, Debug, Eq, PartialEq, Hash)]
 pub struct UnifiedExecApprovalKey {
     pub command: Vec<String>,
@@ -85,10 +96,11 @@ impl Approvable<UnifiedExecRequest> for UnifiedExecRuntime<'_> {
         let command = req.command.clone();
         let cwd = req.cwd.clone();
         let reason = ctx.retry_reason.clone();
+        let risk = ctx.risk.clone();
         Box::pin(async move {
             with_cached_approval(&session.services, key, || async move {
                 session
-                    .request_command_approval(turn, call_id, command, cwd, reason)
+                    .request_command_approval(turn, call_id, command, cwd, reason, risk)
                     .await
             })
             .await

codex-rs/core/src/tools/sandboxing.rs

@@ -7,6 +7,7 @@
 use crate::codex::Session;
 use crate::codex::TurnContext;
 use crate::error::CodexErr;
+use crate::protocol::SandboxCommandAssessment;
 use crate::protocol::SandboxPolicy;
 use crate::sandboxing::CommandSpec;
 use crate::sandboxing::SandboxManager;
@@ -18,6 +19,7 @@ use std::collections::HashMap;
 use std::fmt::Debug;
 use std::hash::Hash;
 use std::path::Path;
+use std::path::PathBuf;
 
 use futures::Future;
 use futures::future::BoxFuture;
@@ -81,6 +83,7 @@ pub(crate) struct ApprovalCtx<'a> {
     pub turn: &'a TurnContext,
     pub call_id: &'a str,
     pub retry_reason: Option<String>,
+    pub risk: Option<SandboxCommandAssessment>,
 }
 
 pub(crate) trait Approvable<Req> {
@@ -156,6 +159,17 @@ pub(crate) struct ToolCtx<'a> {
     pub tool_name: String,
 }
 
+/// Captures the command metadata needed to re-run a tool request without sandboxing.
+#[derive(Clone, Debug, PartialEq, Eq)]
+pub(crate) struct SandboxRetryData {
+    pub command: Vec<String>,
+    pub cwd: PathBuf,
+}
+
+pub(crate) trait ProvidesSandboxRetryData {
+    fn sandbox_retry_data(&self) -> Option<SandboxRetryData>;
+}
+
 #[derive(Debug)]
 pub(crate) enum ToolError {
     Rejected(String),

codex-rs/core/templates/sandboxing/assessment_prompt.md

@@ -0,0 +1,27 @@
+You are a security analyst evaluating shell commands that were blocked by a sandbox. Given the provided metadata, summarize the command's likely intent and assess the risk. Return strictly valid JSON with the keys:
+- description (concise summary, at most two sentences)
+- risk_level ("low", "medium", or "high")
+- risk_categories (optional array of zero or more category strings)
+Risk level examples:
+- low: read-only inspections, listing files, printing configuration
+- medium: modifying project files, installing dependencies, fetching artifacts from trusted sources
+- high: deleting or overwriting data, exfiltrating secrets, escalating privileges, or disabling security controls
+Recognized risk_categories: data_deletion, data_exfiltration, privilege_escalation, system_modification, network_access, resource_exhaustion, compliance.
+Use multiple categories when appropriate.
+If information is insufficient, choose the most cautious risk level supported by the evidence.
+Respond with JSON only, without markdown code fences or extra commentary.
+
+---
+
+Command metadata:
+Platform: {{ platform }}
+Sandbox policy: {{ sandbox_policy }}
+{% if let Some(roots) = filesystem_roots %}
+Filesystem roots: {{ roots }}
+{% endif %}
+Working directory: {{ working_directory }}
+Command argv: {{ command_argv }}
+Command (joined): {{ command_joined }}
+{% if let Some(message) = sandbox_failure_message %}
+Sandbox failure message: {{ message }}
+{% endif %}

codex-rs/core/tests/chat_completions_sse.rs

@@ -185,6 +185,49 @@ async fn streams_text_without_reasoning() {
     assert_matches!(events[2], ResponseEvent::Completed { .. });
 }
 
+#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
+async fn completed_event_includes_usage_estimate() {
+    if network_disabled() {
+        println!(
+            "Skipping test because it cannot execute when network is disabled in a Codex sandbox."
+        );
+        return;
+    }
+
+    let sse = concat!(
+        "data: {\"choices\":[{\"delta\":{\"content\":\"hi\"}}]}\n\n",
+        "data: {\"choices\":[{\"delta\":{}}]}\n\n",
+        "data: [DONE]\n\n",
+    );
+
+    let events = run_stream(sse).await;
+    assert_eq!(events.len(), 3, "unexpected events: {events:?}");
+
+    let usage = events
+        .iter()
+        .find_map(|event| match event {
+            ResponseEvent::Completed {
+                token_usage: Some(usage),
+                ..
+            } => Some(usage.clone()),
+            _ => None,
+        })
+        .expect("missing usage estimate on Completed event");
+
+    assert!(
+        usage.input_tokens > 0,
+        "expected input tokens > 0, got {usage:?}"
+    );
+    assert!(
+        usage.output_tokens > 0,
+        "expected output tokens > 0, got {usage:?}"
+    );
+    assert!(
+        usage.total_tokens >= usage.input_tokens + usage.output_tokens,
+        "expected total tokens to cover input + output, got {usage:?}"
+    );
+}
+
 #[tokio::test(flavor = "multi_thread", worker_threads = 2)]
 async fn streams_reasoning_from_string_delta() {
     if network_disabled() {

codex-rs/core/tests/suite/cli_stream.rs

@@ -75,9 +75,17 @@ async fn chat_mode_stream_cli() {
     server.verify().await;
 
     // Verify a new session rollout was created and is discoverable via list_conversations
-    let page = RolloutRecorder::list_conversations(home.path(), 10, None, &[])
-        .await
-        .expect("list conversations");
+    let provider_filter = vec!["mock".to_string()];
+    let page = RolloutRecorder::list_conversations(
+        home.path(),
+        10,
+        None,
+        &[],
+        Some(provider_filter.as_slice()),
+        "mock",
+    )
+    .await
+    .expect("list conversations");
     assert!(
         !page.items.is_empty(),
         "expected at least one session to be listed"

codex-rs/core/tests/suite/client.rs

@@ -154,7 +154,8 @@ async fn resume_includes_initial_messages_and_sends_prior_items() {
                 "instructions": "be nice",
                 "cwd": ".",
                 "originator": "test_originator",
-                "cli_version": "test_version"
+                "cli_version": "test_version",
+                "model_provider": "test-provider"
             }
         })
     )
@@ -524,7 +525,7 @@ async fn prefers_apikey_when_config_prefers_apikey_even_with_chatgpt_tokens() {
     let mut config = load_default_config_for_test(&codex_home);
     config.model_provider = model_provider;
 
-    let auth_manager = match CodexAuth::from_codex_home(codex_home.path()) {
+    let auth_manager = match CodexAuth::from_auth_storage(codex_home.path()) {
         Ok(Some(auth)) => codex_core::AuthManager::from_auth_for_testing(auth),
         Ok(None) => panic!("No CodexAuth found in codex_home"),
         Err(e) => panic!("Failed to load CodexAuth: {e}"),

codex-rs/core/tests/suite/compact_resume_fork.rs

@@ -18,7 +18,6 @@ use codex_core::built_in_model_providers;
 use codex_core::codex::compact::SUMMARIZATION_PROMPT;
 use codex_core::config::Config;
 use codex_core::config::OPENAI_DEFAULT_MODEL;
-use codex_core::protocol::ConversationPathResponseEvent;
 use codex_core::protocol::EventMsg;
 use codex_core::protocol::Op;
 use codex_core::spawn::CODEX_SANDBOX_NETWORK_DISABLED_ENV_VAR;
@@ -42,6 +41,29 @@ fn network_disabled() -> bool {
     std::env::var(CODEX_SANDBOX_NETWORK_DISABLED_ENV_VAR).is_ok()
 }
 
+fn filter_out_ghost_snapshot_entries(items: &[Value]) -> Vec<Value> {
+    items
+        .iter()
+        .filter(|item| !is_ghost_snapshot_message(item))
+        .cloned()
+        .collect()
+}
+
+fn is_ghost_snapshot_message(item: &Value) -> bool {
+    if item.get("type").and_then(Value::as_str) != Some("message") {
+        return false;
+    }
+    if item.get("role").and_then(Value::as_str) != Some("user") {
+        return false;
+    }
+    item.get("content")
+        .and_then(Value::as_array)
+        .and_then(|content| content.first())
+        .and_then(|entry| entry.get("text"))
+        .and_then(Value::as_str)
+        .is_some_and(|text| text.trim_start().starts_with("<ghost_snapshot>"))
+}
+
 #[tokio::test(flavor = "multi_thread", worker_threads = 2)]
 /// Scenario: compact an initial conversation, resume it, fork one turn back, and
 /// ensure the model-visible history matches expectations at each request.
@@ -61,7 +83,7 @@ async fn compact_resume_and_fork_preserve_model_history_view() {
     user_turn(&base, "hello world").await;
     compact_conversation(&base).await;
     user_turn(&base, "AFTER_COMPACT").await;
-    let base_path = fetch_conversation_path(&base, "base conversation").await;
+    let base_path = fetch_conversation_path(&base).await;
     assert!(
         base_path.exists(),
         "compact+resume test expects base path {base_path:?} to exist",
@@ -69,7 +91,7 @@ async fn compact_resume_and_fork_preserve_model_history_view() {
 
     let resumed = resume_conversation(&manager, &config, base_path).await;
     user_turn(&resumed, "AFTER_RESUME").await;
-    let resumed_path = fetch_conversation_path(&resumed, "resumed conversation").await;
+    let resumed_path = fetch_conversation_path(&resumed).await;
     assert!(
         resumed_path.exists(),
         "compact+resume test expects resumed path {resumed_path:?} to exist",
@@ -518,7 +540,7 @@ async fn compact_resume_after_second_compaction_preserves_history() {
     user_turn(&base, "hello world").await;
     compact_conversation(&base).await;
     user_turn(&base, "AFTER_COMPACT").await;
-    let base_path = fetch_conversation_path(&base, "base conversation").await;
+    let base_path = fetch_conversation_path(&base).await;
     assert!(
         base_path.exists(),
         "second compact test expects base path {base_path:?} to exist",
@@ -526,7 +548,7 @@ async fn compact_resume_after_second_compaction_preserves_history() {
 
     let resumed = resume_conversation(&manager, &config, base_path).await;
     user_turn(&resumed, "AFTER_RESUME").await;
-    let resumed_path = fetch_conversation_path(&resumed, "resumed conversation").await;
+    let resumed_path = fetch_conversation_path(&resumed).await;
     assert!(
         resumed_path.exists(),
         "second compact test expects resumed path {resumed_path:?} to exist",
@@ -537,7 +559,7 @@ async fn compact_resume_after_second_compaction_preserves_history() {
 
     compact_conversation(&forked).await;
     user_turn(&forked, "AFTER_COMPACT_2").await;
-    let forked_path = fetch_conversation_path(&forked, "forked conversation").await;
+    let forked_path = fetch_conversation_path(&forked).await;
     assert!(
         forked_path.exists(),
         "second compact test expects forked path {forked_path:?} to exist",
@@ -557,13 +579,15 @@ async fn compact_resume_after_second_compaction_preserves_history() {
     let resume_input_array = input_after_resume
         .as_array()
         .expect("input after resume should be an array");
+    let compact_filtered = filter_out_ghost_snapshot_entries(compact_input_array);
+    let resume_filtered = filter_out_ghost_snapshot_entries(resume_input_array);
     assert!(
-        compact_input_array.len() <= resume_input_array.len(),
+        compact_filtered.len() <= resume_filtered.len(),
         "after-resume input should have at least as many items as after-compact"
     );
     assert_eq!(
-        compact_input_array.as_slice(),
-        &resume_input_array[..compact_input_array.len()]
+        compact_filtered.as_slice(),
+        &resume_filtered[..compact_filtered.len()]
     );
     // hard coded test
     let prompt = requests[0]["instructions"]
@@ -792,22 +816,8 @@ async fn compact_conversation(conversation: &Arc<CodexConversation>) {
     wait_for_event(conversation, |ev| matches!(ev, EventMsg::TaskComplete(_))).await;
 }
 
-async fn fetch_conversation_path(
-    conversation: &Arc<CodexConversation>,
-    context: &str,
-) -> std::path::PathBuf {
-    conversation
-        .submit(Op::GetPath)
-        .await
-        .expect("request conversation path");
-    match wait_for_event(conversation, |ev| {
-        matches!(ev, EventMsg::ConversationPath(_))
-    })
-    .await
-    {
-        EventMsg::ConversationPath(ConversationPathResponseEvent { path, .. }) => path,
-        _ => panic!("expected ConversationPath event for {context}"),
-    }
+async fn fetch_conversation_path(conversation: &Arc<CodexConversation>) -> std::path::PathBuf {
+    conversation.rollout_path()
 }
 
 async fn resume_conversation(

codex-rs/core/tests/suite/fork_conversation.rs

@@ -4,7 +4,6 @@ use codex_core::ModelProviderInfo;
 use codex_core::NewConversation;
 use codex_core::built_in_model_providers;
 use codex_core::parse_turn_item;
-use codex_core::protocol::ConversationPathResponseEvent;
 use codex_core::protocol::EventMsg;
 use codex_core::protocol::Op;
 use codex_core::protocol::RolloutItem;
@@ -79,13 +78,7 @@ async fn fork_conversation_twice_drops_to_first_message() {
     }
 
     // Request history from the base conversation to obtain rollout path.
-    codex.submit(Op::GetPath).await.unwrap();
-    let base_history =
-        wait_for_event(&codex, |ev| matches!(ev, EventMsg::ConversationPath(_))).await;
-    let base_path = match &base_history {
-        EventMsg::ConversationPath(ConversationPathResponseEvent { path, .. }) => path.clone(),
-        _ => panic!("expected ConversationHistory event"),
-    };
+    let base_path = codex.rollout_path();
 
     // GetHistory flushes before returning the path; no wait needed.
 
@@ -140,15 +133,7 @@ async fn fork_conversation_twice_drops_to_first_message() {
         .await
         .expect("fork 1");
 
-    codex_fork1.submit(Op::GetPath).await.unwrap();
-    let fork1_history = wait_for_event(&codex_fork1, |ev| {
-        matches!(ev, EventMsg::ConversationPath(_))
-    })
-    .await;
-    let fork1_path = match &fork1_history {
-        EventMsg::ConversationPath(ConversationPathResponseEvent { path, .. }) => path.clone(),
-        _ => panic!("expected ConversationHistory event after first fork"),
-    };
+    let fork1_path = codex_fork1.rollout_path();
 
     // GetHistory on fork1 flushed; the file is ready.
     let fork1_items = read_items(&fork1_path);
@@ -166,15 +151,7 @@ async fn fork_conversation_twice_drops_to_first_message() {
         .await
         .expect("fork 2");
 
-    codex_fork2.submit(Op::GetPath).await.unwrap();
-    let fork2_history = wait_for_event(&codex_fork2, |ev| {
-        matches!(ev, EventMsg::ConversationPath(_))
-    })
-    .await;
-    let fork2_path = match &fork2_history {
-        EventMsg::ConversationPath(ConversationPathResponseEvent { path, .. }) => path.clone(),
-        _ => panic!("expected ConversationHistory event after second fork"),
-    };
+    let fork2_path = codex_fork2.rollout_path();
     // GetHistory on fork2 flushed; the file is ready.
     let fork1_items = read_items(&fork1_path);
     let fork1_user_inputs = find_user_input_positions(&fork1_items);

codex-rs/core/tests/suite/mod.rs

@@ -33,6 +33,7 @@ mod stream_no_completed;
 mod tool_harness;
 mod tool_parallelism;
 mod tools;
+mod truncation;
 mod unified_exec;
 mod user_notification;
 mod view_image;

codex-rs/core/tests/suite/review.rs

@@ -7,7 +7,6 @@ use codex_core::REVIEW_PROMPT;
 use codex_core::ResponseItem;
 use codex_core::built_in_model_providers;
 use codex_core::config::Config;
-use codex_core::protocol::ConversationPathResponseEvent;
 use codex_core::protocol::ENVIRONMENT_CONTEXT_OPEN_TAG;
 use codex_core::protocol::EventMsg;
 use codex_core::protocol::ExitedReviewModeEvent;
@@ -120,13 +119,7 @@ async fn review_op_emits_lifecycle_and_review_output() {
 
     // Also verify that a user message with the header and a formatted finding
     // was recorded back in the parent session's rollout.
-    codex.submit(Op::GetPath).await.unwrap();
-    let history_event =
-        wait_for_event(&codex, |ev| matches!(ev, EventMsg::ConversationPath(_))).await;
-    let path = match history_event {
-        EventMsg::ConversationPath(ConversationPathResponseEvent { path, .. }) => path,
-        other => panic!("expected ConversationPath event, got {other:?}"),
-    };
+    let path = codex.rollout_path();
     let text = std::fs::read_to_string(&path).expect("read rollout file");
 
     let mut saw_header = false;
@@ -375,7 +368,8 @@ async fn review_input_isolated_from_parent_history() {
                 "instructions": null,
                 "cwd": ".",
                 "originator": "test_originator",
-                "cli_version": "test_version"
+                "cli_version": "test_version",
+                "model_provider": "test-provider"
             }
         });
         f.write_all(format!("{meta_line}\n").as_bytes())
@@ -482,13 +476,7 @@ async fn review_input_isolated_from_parent_history() {
     assert_eq!(instructions, REVIEW_PROMPT);
 
     // Also verify that a user interruption note was recorded in the rollout.
-    codex.submit(Op::GetPath).await.unwrap();
-    let history_event =
-        wait_for_event(&codex, |ev| matches!(ev, EventMsg::ConversationPath(_))).await;
-    let path = match history_event {
-        EventMsg::ConversationPath(ConversationPathResponseEvent { path, .. }) => path,
-        other => panic!("expected ConversationPath event, got {other:?}"),
-    };
+    let path = codex.rollout_path();
     let text = std::fs::read_to_string(&path).expect("read rollout file");
     let mut saw_interruption_message = false;
     for line in text.lines() {

codex-rs/core/tests/suite/rmcp_client.rs

@@ -14,6 +14,8 @@ use codex_core::features::Feature;
 
 use codex_core::protocol::AskForApproval;
 use codex_core::protocol::EventMsg;
+use codex_core::protocol::McpInvocation;
+use codex_core::protocol::McpToolCallBeginEvent;
 use codex_core::protocol::Op;
 use codex_core::protocol::SandboxPolicy;
 use codex_protocol::config_types::ReasoningSummary;
@@ -25,7 +27,9 @@ use core_test_support::test_codex::test_codex;
 use core_test_support::wait_for_event;
 use core_test_support::wait_for_event_with_timeout;
 use escargot::CargoBuild;
+use mcp_types::ContentBlock;
 use serde_json::Value;
+use serde_json::json;
 use serial_test::serial;
 use tempfile::tempdir;
 use tokio::net::TcpStream;
@@ -35,6 +39,8 @@ use tokio::time::Instant;
 use tokio::time::sleep;
 use wiremock::matchers::any;
 
+static OPENAI_PNG: &str = "data:image/png;base64,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";
+
 #[tokio::test(flavor = "multi_thread", worker_threads = 1)]
 #[serial(mcp_test_value)]
 async fn stdio_server_round_trip() -> anyhow::Result<()> {
@@ -175,6 +181,352 @@ async fn stdio_server_round_trip() -> anyhow::Result<()> {
     Ok(())
 }
 
+#[tokio::test(flavor = "multi_thread", worker_threads = 1)]
+#[serial(mcp_test_value)]
+async fn stdio_image_responses_round_trip() -> anyhow::Result<()> {
+    skip_if_no_network!(Ok(()));
+
+    let server = responses::start_mock_server().await;
+
+    let call_id = "img-1";
+    let server_name = "rmcp";
+    let tool_name = format!("mcp__{server_name}__image");
+
+    // First stream: model decides to call the image tool.
+    mount_sse_once_match(
+        &server,
+        any(),
+        responses::sse(vec![
+            responses::ev_response_created("resp-1"),
+            responses::ev_function_call(call_id, &tool_name, "{}"),
+            responses::ev_completed("resp-1"),
+        ]),
+    )
+    .await;
+    // Second stream: after tool execution, assistant emits a message and completes.
+    let final_mock = mount_sse_once_match(
+        &server,
+        any(),
+        responses::sse(vec![
+            responses::ev_assistant_message("msg-1", "rmcp image tool completed successfully."),
+            responses::ev_completed("resp-2"),
+        ]),
+    )
+    .await;
+
+    // Build the stdio rmcp server and pass the image as data URL so it can construct ImageContent.
+    let rmcp_test_server_bin = CargoBuild::new()
+        .package("codex-rmcp-client")
+        .bin("test_stdio_server")
+        .run()?
+        .path()
+        .to_string_lossy()
+        .into_owned();
+
+    let fixture = test_codex()
+        .with_config(move |config| {
+            config.features.enable(Feature::RmcpClient);
+            config.mcp_servers.insert(
+                server_name.to_string(),
+                McpServerConfig {
+                    transport: McpServerTransportConfig::Stdio {
+                        command: rmcp_test_server_bin,
+                        args: Vec::new(),
+                        env: Some(HashMap::from([(
+                            "MCP_TEST_IMAGE_DATA_URL".to_string(),
+                            OPENAI_PNG.to_string(),
+                        )])),
+                        env_vars: Vec::new(),
+                        cwd: None,
+                    },
+                    enabled: true,
+                    startup_timeout_sec: Some(Duration::from_secs(10)),
+                    tool_timeout_sec: None,
+                    enabled_tools: None,
+                    disabled_tools: None,
+                },
+            );
+        })
+        .build(&server)
+        .await?;
+    let session_model = fixture.session_configured.model.clone();
+
+    fixture
+        .codex
+        .submit(Op::UserTurn {
+            items: vec![UserInput::Text {
+                text: "call the rmcp image tool".into(),
+            }],
+            final_output_json_schema: None,
+            cwd: fixture.cwd.path().to_path_buf(),
+            approval_policy: AskForApproval::Never,
+            sandbox_policy: SandboxPolicy::ReadOnly,
+            model: session_model,
+            effort: None,
+            summary: ReasoningSummary::Auto,
+        })
+        .await?;
+
+    // Wait for tool begin/end and final completion.
+    let begin_event = wait_for_event_with_timeout(
+        &fixture.codex,
+        |ev| matches!(ev, EventMsg::McpToolCallBegin(_)),
+        Duration::from_secs(10),
+    )
+    .await;
+    let EventMsg::McpToolCallBegin(begin) = begin_event else {
+        unreachable!("begin");
+    };
+    assert_eq!(
+        begin,
+        McpToolCallBeginEvent {
+            call_id: call_id.to_string(),
+            invocation: McpInvocation {
+                server: server_name.to_string(),
+                tool: "image".to_string(),
+                arguments: Some(json!({})),
+            },
+        },
+    );
+
+    let end_event = wait_for_event(&fixture.codex, |ev| {
+        matches!(ev, EventMsg::McpToolCallEnd(_))
+    })
+    .await;
+    let EventMsg::McpToolCallEnd(end) = end_event else {
+        unreachable!("end");
+    };
+    assert_eq!(end.call_id, call_id);
+    assert_eq!(
+        end.invocation,
+        McpInvocation {
+            server: server_name.to_string(),
+            tool: "image".to_string(),
+            arguments: Some(json!({})),
+        }
+    );
+    let result = end.result.expect("rmcp image tool should return success");
+    assert_eq!(result.is_error, Some(false));
+    assert_eq!(result.content.len(), 1);
+    let base64_only = OPENAI_PNG
+        .strip_prefix("data:image/png;base64,")
+        .expect("data url prefix");
+    match &result.content[0] {
+        ContentBlock::ImageContent(img) => {
+            assert_eq!(img.mime_type, "image/png");
+            assert_eq!(img.r#type, "image");
+            assert_eq!(img.data, base64_only);
+        }
+        other => panic!("expected image content, got {other:?}"),
+    }
+
+    wait_for_event(&fixture.codex, |ev| matches!(ev, EventMsg::TaskComplete(_))).await;
+
+    let output_item = final_mock.single_request().function_call_output(call_id);
+    assert_eq!(
+        output_item,
+        json!({
+            "type": "function_call_output",
+            "call_id": call_id,
+            "output": [{
+                "type": "input_image",
+                "image_url": OPENAI_PNG
+            }]
+        })
+    );
+    server.verify().await;
+    Ok(())
+}
+
+#[tokio::test(flavor = "multi_thread", worker_threads = 1)]
+#[serial(mcp_test_value)]
+async fn stdio_image_completions_round_trip() -> anyhow::Result<()> {
+    skip_if_no_network!(Ok(()));
+
+    let server = responses::start_mock_server().await;
+
+    let call_id = "img-cc-1";
+    let server_name = "rmcp";
+    let tool_name = format!("mcp__{server_name}__image");
+
+    let tool_call = json!({
+        "choices": [
+            {
+                "delta": {
+                    "tool_calls": [
+                        {
+                            "id": call_id,
+                            "type": "function",
+                            "function": {"name": tool_name, "arguments": "{}"}
+                        }
+                    ]
+                },
+                "finish_reason": "tool_calls"
+            }
+        ]
+    });
+    let sse_tool_call = format!(
+        "data: {}\n\ndata: [DONE]\n\n",
+        serde_json::to_string(&tool_call)?
+    );
+
+    let final_assistant = json!({
+        "choices": [
+            {
+                "delta": {"content": "rmcp image tool completed successfully."},
+                "finish_reason": "stop"
+            }
+        ]
+    });
+    let sse_final = format!(
+        "data: {}\n\ndata: [DONE]\n\n",
+        serde_json::to_string(&final_assistant)?
+    );
+
+    use std::sync::atomic::AtomicUsize;
+    use std::sync::atomic::Ordering;
+    struct ChatSeqResponder {
+        num_calls: AtomicUsize,
+        bodies: Vec<String>,
+    }
+    impl wiremock::Respond for ChatSeqResponder {
+        fn respond(&self, _: &wiremock::Request) -> wiremock::ResponseTemplate {
+            let idx = self.num_calls.fetch_add(1, Ordering::SeqCst);
+            match self.bodies.get(idx) {
+                Some(body) => wiremock::ResponseTemplate::new(200)
+                    .insert_header("content-type", "text/event-stream")
+                    .set_body_string(body.clone()),
+                None => panic!("no chat completion response for index {idx}"),
+            }
+        }
+    }
+
+    let chat_seq = ChatSeqResponder {
+        num_calls: AtomicUsize::new(0),
+        bodies: vec![sse_tool_call, sse_final],
+    };
+    wiremock::Mock::given(wiremock::matchers::method("POST"))
+        .and(wiremock::matchers::path("/v1/chat/completions"))
+        .respond_with(chat_seq)
+        .expect(2)
+        .mount(&server)
+        .await;
+
+    let rmcp_test_server_bin = CargoBuild::new()
+        .package("codex-rmcp-client")
+        .bin("test_stdio_server")
+        .run()?
+        .path()
+        .to_string_lossy()
+        .into_owned();
+
+    let fixture = test_codex()
+        .with_config(move |config| {
+            config.model_provider.wire_api = codex_core::WireApi::Chat;
+            config.features.enable(Feature::RmcpClient);
+            config.mcp_servers.insert(
+                server_name.to_string(),
+                McpServerConfig {
+                    transport: McpServerTransportConfig::Stdio {
+                        command: rmcp_test_server_bin,
+                        args: Vec::new(),
+                        env: Some(HashMap::from([(
+                            "MCP_TEST_IMAGE_DATA_URL".to_string(),
+                            OPENAI_PNG.to_string(),
+                        )])),
+                        env_vars: Vec::new(),
+                        cwd: None,
+                    },
+                    enabled: true,
+                    startup_timeout_sec: Some(Duration::from_secs(10)),
+                    tool_timeout_sec: None,
+                    enabled_tools: None,
+                    disabled_tools: None,
+                },
+            );
+        })
+        .build(&server)
+        .await?;
+    let session_model = fixture.session_configured.model.clone();
+
+    fixture
+        .codex
+        .submit(Op::UserTurn {
+            items: vec![UserInput::Text {
+                text: "call the rmcp image tool".into(),
+            }],
+            final_output_json_schema: None,
+            cwd: fixture.cwd.path().to_path_buf(),
+            approval_policy: AskForApproval::Never,
+            sandbox_policy: SandboxPolicy::ReadOnly,
+            model: session_model,
+            effort: None,
+            summary: ReasoningSummary::Auto,
+        })
+        .await?;
+
+    let begin_event = wait_for_event_with_timeout(
+        &fixture.codex,
+        |ev| matches!(ev, EventMsg::McpToolCallBegin(_)),
+        Duration::from_secs(10),
+    )
+    .await;
+    let EventMsg::McpToolCallBegin(begin) = begin_event else {
+        unreachable!("begin");
+    };
+    assert_eq!(
+        begin,
+        McpToolCallBeginEvent {
+            call_id: call_id.to_string(),
+            invocation: McpInvocation {
+                server: server_name.to_string(),
+                tool: "image".to_string(),
+                arguments: Some(json!({})),
+            },
+        },
+    );
+
+    let end_event = wait_for_event(&fixture.codex, |ev| {
+        matches!(ev, EventMsg::McpToolCallEnd(_))
+    })
+    .await;
+    let EventMsg::McpToolCallEnd(end) = end_event else {
+        unreachable!("end");
+    };
+    assert!(end.result.as_ref().is_ok(), "tool call should succeed");
+
+    wait_for_event(&fixture.codex, |ev| matches!(ev, EventMsg::TaskComplete(_))).await;
+
+    // Chat Completions assertion: the second POST should include a tool role message
+    // with an array `content` containing an item with the expected data URL.
+    let requests = server.received_requests().await.expect("requests captured");
+    assert!(requests.len() >= 2, "expected two chat completion calls");
+    let second = &requests[1];
+    let body: Value = serde_json::from_slice(&second.body)?;
+    let messages = body
+        .get("messages")
+        .and_then(Value::as_array)
+        .cloned()
+        .expect("messages array");
+    let tool_msg = messages
+        .iter()
+        .find(|m| {
+            m.get("role") == Some(&json!("tool")) && m.get("tool_call_id") == Some(&json!(call_id))
+        })
+        .cloned()
+        .expect("tool message present");
+    assert_eq!(
+        tool_msg,
+        json!({
+            "role": "tool",
+            "tool_call_id": call_id,
+            "content": [{"type": "image_url", "image_url": {"url": OPENAI_PNG}}]
+        })
+    );
+
+    Ok(())
+}
+
 #[tokio::test(flavor = "multi_thread", worker_threads = 1)]
 #[serial(mcp_test_value)]
 async fn stdio_server_propagates_whitelisted_env_vars() -> anyhow::Result<()> {

codex-rs/core/tests/suite/rollout_list_find.rs

@@ -28,7 +28,8 @@ fn write_minimal_rollout_with_id(codex_home: &Path, id: Uuid) -> PathBuf {
                 "instructions": null,
                 "cwd": ".",
                 "originator": "test",
-                "cli_version": "test"
+                "cli_version": "test",
+                "model_provider": "test-provider"
             }
         })
     )

codex-rs/core/tests/suite/truncation.rs

@@ -0,0 +1,270 @@
+#![cfg(not(target_os = "windows"))]
+#![allow(clippy::unwrap_used, clippy::expect_used)]
+
+use anyhow::Context;
+use anyhow::Result;
+use codex_core::features::Feature;
+use codex_core::model_family::find_family_for_model;
+use codex_core::protocol::SandboxPolicy;
+use core_test_support::assert_regex_match;
+use core_test_support::responses;
+use core_test_support::responses::ev_assistant_message;
+use core_test_support::responses::ev_completed;
+use core_test_support::responses::ev_function_call;
+use core_test_support::responses::ev_response_created;
+use core_test_support::responses::mount_sse_once_match;
+use core_test_support::responses::mount_sse_sequence;
+use core_test_support::responses::sse;
+use core_test_support::responses::start_mock_server;
+use core_test_support::skip_if_no_network;
+use core_test_support::test_codex::test_codex;
+use escargot::CargoBuild;
+use regex_lite::Regex;
+use serde_json::Value;
+use serde_json::json;
+use wiremock::matchers::any;
+
+// Verifies byte-truncation formatting for function error output (RespondToModel errors)
+#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
+async fn truncate_function_error_trims_respond_to_model() -> Result<()> {
+    skip_if_no_network!(Ok(()));
+
+    let server = start_mock_server().await;
+    let mut builder = test_codex().with_config(|config| {
+        // Use the test model that wires function tools like grep_files
+        config.model = "test-gpt-5-codex".to_string();
+        config.model_family =
+            find_family_for_model("test-gpt-5-codex").expect("model family for test model");
+    });
+    let test = builder.build(&server).await?;
+
+    // Construct a very long, non-existent path to force a RespondToModel error with a large message
+    let long_path = "a".repeat(20_000);
+    let call_id = "grep-huge-error";
+    let args = json!({
+        "pattern": "alpha",
+        "path": long_path,
+        "limit": 10
+    });
+    let responses = vec![
+        sse(vec![
+            ev_response_created("resp-1"),
+            ev_function_call(call_id, "grep_files", &serde_json::to_string(&args)?),
+            ev_completed("resp-1"),
+        ]),
+        sse(vec![
+            ev_assistant_message("msg-1", "done"),
+            ev_completed("resp-2"),
+        ]),
+    ];
+    let mock = mount_sse_sequence(&server, responses).await;
+
+    test.submit_turn_with_policy(
+        "trigger grep_files with long path to test truncation",
+        SandboxPolicy::DangerFullAccess,
+    )
+    .await?;
+
+    let output = mock
+        .function_call_output_text(call_id)
+        .context("function error output present")?;
+
+    tracing::debug!(output = %output, "truncated function error output");
+
+    // Expect plaintext with byte-truncation marker and no omitted-lines marker
+    assert!(
+        serde_json::from_str::<serde_json::Value>(&output).is_err(),
+        "expected error output to be plain text",
+    );
+    let truncated_pattern = r#"(?s)^Total output lines: 1\s+.*\[\.\.\. output truncated to fit 10240 bytes \.\.\.\]\s*$"#;
+    assert_regex_match(truncated_pattern, &output);
+    assert!(
+        !output.contains("omitted"),
+        "line omission marker should not appear when no lines were dropped: {output}"
+    );
+
+    Ok(())
+}
+
+// Verifies that a standard tool call (shell) exceeding the model formatting
+// limits is truncated before being sent back to the model.
+#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
+async fn tool_call_output_exceeds_limit_truncated_for_model() -> Result<()> {
+    skip_if_no_network!(Ok(()));
+
+    let server = start_mock_server().await;
+
+    // Use a model that exposes the generic shell tool.
+    let mut builder = test_codex().with_config(|config| {
+        config.model = "gpt-5-codex".to_string();
+        config.model_family =
+            find_family_for_model("gpt-5-codex").expect("gpt-5-codex is a model family");
+    });
+    let fixture = builder.build(&server).await?;
+
+    let call_id = "shell-too-large";
+    let args = serde_json::json!({
+        "command": ["/bin/sh", "-c", "seq 1 400"],
+        "timeout_ms": 5_000,
+    });
+
+    // First response: model tells us to run the tool; second: complete the turn.
+    mount_sse_once_match(
+        &server,
+        any(),
+        sse(vec![
+            responses::ev_response_created("resp-1"),
+            responses::ev_function_call(call_id, "shell", &serde_json::to_string(&args)?),
+            responses::ev_completed("resp-1"),
+        ]),
+    )
+    .await;
+    let mock2 = mount_sse_once_match(
+        &server,
+        any(),
+        sse(vec![
+            responses::ev_assistant_message("msg-1", "done"),
+            responses::ev_completed("resp-2"),
+        ]),
+    )
+    .await;
+
+    fixture
+        .submit_turn_with_policy("trigger big shell output", SandboxPolicy::DangerFullAccess)
+        .await?;
+
+    // Inspect what we sent back to the model; it should contain a truncated
+    // function_call_output for the shell call.
+    let output = mock2
+        .single_request()
+        .function_call_output_text(call_id)
+        .context("function_call_output present for shell call")?;
+
+    // Expect plain text (not JSON) with truncation markers and line elision.
+    assert!(
+        serde_json::from_str::<Value>(&output).is_err(),
+        "expected truncated shell output to be plain text"
+    );
+    let truncated_pattern = r#"(?s)^Exit code: 0
+Wall time: .* seconds
+Total output lines: 400
+Output:
+1
+2
+3
+4
+5
+6
+.*
+\[\.{3} omitted 144 of 400 lines \.{3}\]
+
+.*
+396
+397
+398
+399
+400
+$"#;
+    assert_regex_match(truncated_pattern, &output);
+
+    Ok(())
+}
+
+// Verifies that an MCP tool call result exceeding the model formatting limits
+// is truncated before being sent back to the model.
+#[tokio::test(flavor = "multi_thread", worker_threads = 1)]
+async fn mcp_tool_call_output_exceeds_limit_truncated_for_model() -> Result<()> {
+    skip_if_no_network!(Ok(()));
+
+    let server = start_mock_server().await;
+
+    let call_id = "rmcp-truncated";
+    let server_name = "rmcp";
+    let tool_name = format!("mcp__{server_name}__echo");
+
+    // Build a very large message to exceed 10KiB once serialized.
+    let large_msg = "long-message-with-newlines-".repeat(600);
+    let args_json = serde_json::json!({ "message": large_msg });
+
+    mount_sse_once_match(
+        &server,
+        any(),
+        sse(vec![
+            responses::ev_response_created("resp-1"),
+            responses::ev_function_call(call_id, &tool_name, &args_json.to_string()),
+            responses::ev_completed("resp-1"),
+        ]),
+    )
+    .await;
+    let mock2 = mount_sse_once_match(
+        &server,
+        any(),
+        sse(vec![
+            responses::ev_assistant_message("msg-1", "rmcp echo tool completed."),
+            responses::ev_completed("resp-2"),
+        ]),
+    )
+    .await;
+
+    // Compile the rmcp stdio test server and configure it.
+    let rmcp_test_server_bin = CargoBuild::new()
+        .package("codex-rmcp-client")
+        .bin("test_stdio_server")
+        .run()?
+        .path()
+        .to_string_lossy()
+        .into_owned();
+
+    let mut builder = test_codex().with_config(move |config| {
+        config.features.enable(Feature::RmcpClient);
+        config.mcp_servers.insert(
+            server_name.to_string(),
+            codex_core::config_types::McpServerConfig {
+                transport: codex_core::config_types::McpServerTransportConfig::Stdio {
+                    command: rmcp_test_server_bin,
+                    args: Vec::new(),
+                    env: None,
+                    env_vars: Vec::new(),
+                    cwd: None,
+                },
+                enabled: true,
+                startup_timeout_sec: Some(std::time::Duration::from_secs(10)),
+                tool_timeout_sec: None,
+                enabled_tools: None,
+                disabled_tools: None,
+            },
+        );
+    });
+    let fixture = builder.build(&server).await?;
+
+    fixture
+        .submit_turn_with_policy(
+            "call the rmcp echo tool with a very large message",
+            SandboxPolicy::ReadOnly,
+        )
+        .await?;
+
+    // The MCP tool call output is converted to a function_call_output for the model.
+    let output = mock2
+        .single_request()
+        .function_call_output_text(call_id)
+        .context("function_call_output present for rmcp call")?;
+
+    // Expect plain text with byte-based truncation marker.
+    assert!(
+        serde_json::from_str::<Value>(&output).is_err(),
+        "expected truncated MCP output to be plain text"
+    );
+    assert!(
+        output.starts_with("Total output lines: 1\n\n{"),
+        "expected total line header and JSON head, got: {output}"
+    );
+    let byte_marker = Regex::new(r"\[\.\.\. output truncated to fit 10240 bytes \.\.\.\]")
+        .expect("compile regex");
+    assert!(
+        byte_marker.is_match(&output),
+        "expected byte truncation marker, got: {output}"
+    );
+
+    Ok(())
+}

codex-rs/core/tests/suite/unified_exec.rs

@@ -240,7 +240,7 @@ async fn unified_exec_emits_output_delta_for_exec_command() -> Result<()> {
     let call_id = "uexec-delta-1";
     let args = json!({
         "cmd": "printf 'HELLO-UEXEC'",
-        "yield_time_ms": 250,
+        "yield_time_ms": 1000,
     });
 
     let responses = vec![

codex-rs/core/tests/suite/view_image.rs

@@ -19,6 +19,10 @@ use core_test_support::skip_if_no_network;
 use core_test_support::test_codex::TestCodex;
 use core_test_support::test_codex::test_codex;
 use core_test_support::wait_for_event;
+use image::GenericImageView;
+use image::ImageBuffer;
+use image::Rgba;
+use image::load_from_memory;
 use serde_json::Value;
 use wiremock::matchers::any;
 
@@ -49,6 +53,88 @@ fn extract_output_text(item: &Value) -> Option<&str> {
     })
 }
 
+#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
+async fn user_turn_with_local_image_attaches_image() -> anyhow::Result<()> {
+    skip_if_no_network!(Ok(()));
+
+    let server = start_mock_server().await;
+
+    let TestCodex {
+        codex,
+        cwd,
+        session_configured,
+        ..
+    } = test_codex().build(&server).await?;
+
+    let rel_path = "user-turn/example.png";
+    let abs_path = cwd.path().join(rel_path);
+    if let Some(parent) = abs_path.parent() {
+        std::fs::create_dir_all(parent)?;
+    }
+    let image = ImageBuffer::from_pixel(4096, 1024, Rgba([20u8, 40, 60, 255]));
+    image.save(&abs_path)?;
+
+    let response = sse(vec![
+        ev_response_created("resp-1"),
+        ev_assistant_message("msg-1", "done"),
+        ev_completed("resp-1"),
+    ]);
+    let mock = responses::mount_sse_once_match(&server, any(), response).await;
+
+    let session_model = session_configured.model.clone();
+
+    codex
+        .submit(Op::UserTurn {
+            items: vec![UserInput::LocalImage {
+                path: abs_path.clone(),
+            }],
+            final_output_json_schema: None,
+            cwd: cwd.path().to_path_buf(),
+            approval_policy: AskForApproval::Never,
+            sandbox_policy: SandboxPolicy::DangerFullAccess,
+            model: session_model,
+            effort: None,
+            summary: ReasoningSummary::Auto,
+        })
+        .await?;
+
+    wait_for_event(&codex, |event| matches!(event, EventMsg::TaskComplete(_))).await;
+
+    let body = mock.single_request().body_json();
+    let image_message =
+        find_image_message(&body).expect("pending input image message not included in request");
+    let image_url = image_message
+        .get("content")
+        .and_then(Value::as_array)
+        .and_then(|content| {
+            content.iter().find_map(|span| {
+                if span.get("type").and_then(Value::as_str) == Some("input_image") {
+                    span.get("image_url").and_then(Value::as_str)
+                } else {
+                    None
+                }
+            })
+        })
+        .expect("image_url present");
+
+    let (prefix, encoded) = image_url
+        .split_once(',')
+        .expect("image url contains data prefix");
+    assert_eq!(prefix, "data:image/png;base64");
+
+    let decoded = BASE64_STANDARD
+        .decode(encoded)
+        .expect("image data decodes from base64 for request");
+    let resized = load_from_memory(&decoded).expect("load resized image");
+    let (width, height) = resized.dimensions();
+    assert!(width <= 2048);
+    assert!(height <= 768);
+    assert!(width < 4096);
+    assert!(height < 1024);
+
+    Ok(())
+}
+
 #[tokio::test(flavor = "multi_thread", worker_threads = 2)]
 async fn view_image_tool_attaches_local_image() -> anyhow::Result<()> {
     skip_if_no_network!(Ok(()));
@@ -67,8 +153,8 @@ async fn view_image_tool_attaches_local_image() -> anyhow::Result<()> {
     if let Some(parent) = abs_path.parent() {
         std::fs::create_dir_all(parent)?;
     }
-    let image_bytes = b"fake_png_bytes".to_vec();
-    std::fs::write(&abs_path, &image_bytes)?;
+    let image = ImageBuffer::from_pixel(4096, 1024, Rgba([255u8, 0, 0, 255]));
+    image.save(&abs_path)?;
 
     let call_id = "view-image-call";
     let arguments = serde_json::json!({ "path": rel_path }).to_string();
@@ -143,11 +229,20 @@ async fn view_image_tool_attaches_local_image() -> anyhow::Result<()> {
         })
         .expect("image_url present");
 
-    let expected_image_url = format!(
-        "data:image/png;base64,{}",
-        BASE64_STANDARD.encode(&image_bytes)
-    );
-    assert_eq!(image_url, expected_image_url);
+    let (prefix, encoded) = image_url
+        .split_once(',')
+        .expect("image url contains data prefix");
+    assert_eq!(prefix, "data:image/png;base64");
+
+    let decoded = BASE64_STANDARD
+        .decode(encoded)
+        .expect("image data decodes from base64 for request");
+    let resized = load_from_memory(&decoded).expect("load resized image");
+    let (resized_width, resized_height) = resized.dimensions();
+    assert!(resized_width <= 2048);
+    assert!(resized_height <= 768);
+    assert!(resized_width < 4096);
+    assert!(resized_height < 1024);
 
     Ok(())
 }

codex-rs/exec/src/event_processor_with_human_output.rs

@@ -20,7 +20,6 @@ use codex_core::protocol::StreamErrorEvent;
 use codex_core::protocol::TaskCompleteEvent;
 use codex_core::protocol::TurnAbortReason;
 use codex_core::protocol::TurnDiffEvent;
-use codex_core::protocol::WebSearchBeginEvent;
 use codex_core::protocol::WebSearchEndEvent;
 use codex_protocol::num_format::format_with_separators;
 use owo_colors::OwoColorize;
@@ -216,7 +215,6 @@ impl EventProcessor for EventProcessorWithHumanOutput {
                     cwd.to_string_lossy(),
                 );
             }
-            EventMsg::ExecCommandOutputDelta(_) => {}
             EventMsg::ExecCommandEnd(ExecCommandEndEvent {
                 aggregated_output,
                 duration,
@@ -283,7 +281,6 @@ impl EventProcessor for EventProcessorWithHumanOutput {
                     }
                 }
             }
-            EventMsg::WebSearchBegin(WebSearchBeginEvent { call_id: _ }) => {}
             EventMsg::WebSearchEnd(WebSearchEndEvent { call_id: _, query }) => {
                 ts_msg!(self, "🌐 Searched: {query}");
             }
@@ -411,12 +408,6 @@ impl EventProcessor for EventProcessorWithHumanOutput {
                 );
                 eprintln!("{unified_diff}");
             }
-            EventMsg::ExecApprovalRequest(_) => {
-                // Should we exit?
-            }
-            EventMsg::ApplyPatchApprovalRequest(_) => {
-                // Should we exit?
-            }
             EventMsg::AgentReasoning(agent_reasoning_event) => {
                 if self.show_agent_reasoning {
                     ts_msg!(
@@ -481,15 +472,6 @@ impl EventProcessor for EventProcessorWithHumanOutput {
                     }
                 }
             }
-            EventMsg::GetHistoryEntryResponse(_) => {
-                // Currently ignored in exec output.
-            }
-            EventMsg::McpListToolsResponse(_) => {
-                // Currently ignored in exec output.
-            }
-            EventMsg::ListCustomPromptsResponse(_) => {
-                // Currently ignored in exec output.
-            }
             EventMsg::ViewImageToolCall(view) => {
                 ts_msg!(
                     self,
@@ -510,15 +492,24 @@ impl EventProcessor for EventProcessorWithHumanOutput {
                 }
             },
             EventMsg::ShutdownComplete => return CodexStatus::Shutdown,
-            EventMsg::ConversationPath(_) => {}
-            EventMsg::UserMessage(_) => {}
-            EventMsg::EnteredReviewMode(_) => {}
-            EventMsg::ExitedReviewMode(_) => {}
-            EventMsg::AgentMessageDelta(_) => {}
-            EventMsg::AgentReasoningDelta(_) => {}
-            EventMsg::AgentReasoningRawContentDelta(_) => {}
-            EventMsg::ItemStarted(_) => {}
-            EventMsg::ItemCompleted(_) => {}
+            EventMsg::WebSearchBegin(_)
+            | EventMsg::ExecApprovalRequest(_)
+            | EventMsg::ApplyPatchApprovalRequest(_)
+            | EventMsg::ExecCommandOutputDelta(_)
+            | EventMsg::GetHistoryEntryResponse(_)
+            | EventMsg::McpListToolsResponse(_)
+            | EventMsg::ListCustomPromptsResponse(_)
+            | EventMsg::RawResponseItem(_)
+            | EventMsg::UserMessage(_)
+            | EventMsg::EnteredReviewMode(_)
+            | EventMsg::ExitedReviewMode(_)
+            | EventMsg::AgentMessageDelta(_)
+            | EventMsg::AgentReasoningDelta(_)
+            | EventMsg::AgentReasoningRawContentDelta(_)
+            | EventMsg::ItemStarted(_)
+            | EventMsg::ItemCompleted(_)
+            | EventMsg::UndoCompleted(_)
+            | EventMsg::UndoStarted(_) => {}
         }
         CodexStatus::Running
     }

codex-rs/exec/src/lib.rs

@@ -179,6 +179,7 @@ pub async fn run_main(cli: Cli, codex_linux_sandbox_exe: Option<PathBuf>) -> any
         include_view_image_tool: None,
         show_raw_agent_reasoning: oss.then_some(true),
         tools_web_search_request: None,
+        experimental_sandbox_command_assessment: None,
         additional_writable_roots: Vec::new(),
     };
     // Parse `-c` overrides.
@@ -388,8 +389,16 @@ async fn resolve_resume_path(
     args: &crate::cli::ResumeArgs,
 ) -> anyhow::Result<Option<PathBuf>> {
     if args.last {
-        match codex_core::RolloutRecorder::list_conversations(&config.codex_home, 1, None, &[])
-            .await
+        let default_provider_filter = vec![config.model_provider_id.clone()];
+        match codex_core::RolloutRecorder::list_conversations(
+            &config.codex_home,
+            1,
+            None,
+            &[],
+            Some(default_provider_filter.as_slice()),
+            &config.model_provider_id,
+        )
+        .await
         {
             Ok(page) => Ok(page.items.first().map(|it| it.path.clone())),
             Err(e) => {

codex-rs/feedback/src/lib.rs

@@ -167,8 +167,16 @@ impl CodexLogSnapshot {
         Ok(path)
     }
 
-    pub fn upload_to_sentry(&self) -> Result<()> {
+    /// Upload feedback to Sentry with optional attachments.
+    pub fn upload_feedback(
+        &self,
+        classification: &str,
+        reason: Option<&str>,
+        include_logs: bool,
+        rollout_path: Option<&std::path::Path>,
+    ) -> Result<()> {
         use std::collections::BTreeMap;
+        use std::fs;
         use std::str::FromStr;
         use std::sync::Arc;
 
@@ -182,36 +190,91 @@ impl CodexLogSnapshot {
         use sentry::transports::DefaultTransportFactory;
         use sentry::types::Dsn;
 
+        // Build Sentry client
         let client = Client::from_config(ClientOptions {
-            dsn: Some(Dsn::from_str(SENTRY_DSN).map_err(|e| anyhow!("invalid DSN: {}", e))?),
+            dsn: Some(Dsn::from_str(SENTRY_DSN).map_err(|e| anyhow!("invalid DSN: {e}"))?),
             transport: Some(Arc::new(DefaultTransportFactory {})),
             ..Default::default()
         });
 
-        let tags = BTreeMap::from([(String::from("thread_id"), self.thread_id.to_string())]);
+        let cli_version = env!("CARGO_PKG_VERSION");
+        let mut tags = BTreeMap::from([
+            (String::from("thread_id"), self.thread_id.to_string()),
+            (String::from("classification"), classification.to_string()),
+            (String::from("cli_version"), cli_version.to_string()),
+        ]);
+        if let Some(r) = reason {
+            tags.insert(String::from("reason"), r.to_string());
+        }
 
-        let event = Event {
-            level: Level::Error,
-            message: Some("Codex Log Upload ".to_string() + &self.thread_id),
+        let level = match classification {
+            "bug" | "bad_result" => Level::Error,
+            _ => Level::Info,
+        };
+
+        let mut envelope = Envelope::new();
+        let title = format!(
+            "[{}]: Codex session {}",
+            display_classification(classification),
+            self.thread_id
+        );
+
+        let mut event = Event {
+            level,
+            message: Some(title.clone()),
             tags,
             ..Default::default()
         };
-        let mut envelope = Envelope::new();
+        if let Some(r) = reason {
+            use sentry::protocol::Exception;
+            use sentry::protocol::Values;
+
+            event.exception = Values::from(vec![Exception {
+                ty: title.clone(),
+                value: Some(r.to_string()),
+                ..Default::default()
+            }]);
+        }
         envelope.add_item(EnvelopeItem::Event(event));
-        envelope.add_item(EnvelopeItem::Attachment(Attachment {
-            buffer: self.bytes.clone(),
-            filename: String::from("codex-logs.log"),
-            content_type: Some("text/plain".to_string()),
-            ty: None,
-        }));
+
+        if include_logs {
+            envelope.add_item(EnvelopeItem::Attachment(Attachment {
+                buffer: self.bytes.clone(),
+                filename: String::from("codex-logs.log"),
+                content_type: Some("text/plain".to_string()),
+                ty: None,
+            }));
+        }
+
+        if let Some((path, data)) = rollout_path.and_then(|p| fs::read(p).ok().map(|d| (p, d))) {
+            let fname = path
+                .file_name()
+                .map(|s| s.to_string_lossy().to_string())
+                .unwrap_or_else(|| "rollout.jsonl".to_string());
+            let content_type = "text/plain".to_string();
+            envelope.add_item(EnvelopeItem::Attachment(Attachment {
+                buffer: data,
+                filename: fname,
+                content_type: Some(content_type),
+                ty: None,
+            }));
+        }
 
         client.send_envelope(envelope);
         client.flush(Some(Duration::from_secs(UPLOAD_TIMEOUT_SECS)));
-
         Ok(())
     }
 }
 
+fn display_classification(classification: &str) -> String {
+    match classification {
+        "bug" => "Bug".to_string(),
+        "bad_result" => "Bad result".to_string(),
+        "good_result" => "Good result".to_string(),
+        _ => "Other".to_string(),
+    }
+}
+
 #[cfg(test)]
 mod tests {
     use super::*;

codex-rs/git-tooling/Cargo.toml

@@ -9,13 +9,20 @@ name = "codex_git_tooling"
 path = "src/lib.rs"
 
 [dependencies]
-tempfile = "3"
-thiserror = "2"
-walkdir = "2"
+tempfile = { workspace = true }
+thiserror = { workspace = true }
+walkdir = { workspace = true }
+schemars = { workspace = true }
+serde = { workspace = true, features = ["derive"] }
+ts-rs = { workspace = true, features = [
+    "uuid-impl",
+    "serde-json-impl",
+    "no-serde-warnings",
+] }
 
 [lints]
 workspace = true
 
 [dev-dependencies]
 assert_matches = { workspace = true }
-pretty_assertions = "1.4.1"
+pretty_assertions = { workspace = true }

codex-rs/git-tooling/src/ghost_commits.rs

@@ -1,4 +1,7 @@
+use std::collections::HashSet;
 use std::ffi::OsString;
+use std::fs;
+use std::io;
 use std::path::Path;
 use std::path::PathBuf;
 
@@ -14,6 +17,7 @@ use crate::operations::resolve_head;
 use crate::operations::resolve_repository_root;
 use crate::operations::run_git_for_status;
 use crate::operations::run_git_for_stdout;
+use crate::operations::run_git_for_stdout_all;
 
 /// Default commit message used for ghost commits when none is provided.
 const DEFAULT_COMMIT_MESSAGE: &str = "codex snapshot";
@@ -69,6 +73,8 @@ pub fn create_ghost_commit(
     let repo_root = resolve_repository_root(options.repo_path)?;
     let repo_prefix = repo_subdir(repo_root.as_path(), options.repo_path);
     let parent = resolve_head(repo_root.as_path())?;
+    let existing_untracked =
+        capture_existing_untracked(repo_root.as_path(), repo_prefix.as_deref())?;
 
     let normalized_force = options
         .force_include
@@ -84,6 +90,16 @@ pub fn create_ghost_commit(
         OsString::from(index_path.as_os_str()),
     )];
 
+    // Pre-populate the temporary index with HEAD so unchanged tracked files
+    // are included in the snapshot tree.
+    if let Some(parent_sha) = parent.as_deref() {
+        run_git_for_status(
+            repo_root.as_path(),
+            vec![OsString::from("read-tree"), OsString::from(parent_sha)],
+            Some(base_env.as_slice()),
+        )?;
+    }
+
     let mut add_args = vec![OsString::from("add"), OsString::from("--all")];
     if let Some(prefix) = repo_prefix.as_deref() {
         add_args.extend([OsString::from("--"), prefix.as_os_str().to_os_string()]);
@@ -127,12 +143,29 @@ pub fn create_ghost_commit(
         Some(commit_env.as_slice()),
     )?;
 
-    Ok(GhostCommit::new(commit_id, parent))
+    Ok(GhostCommit::new(
+        commit_id,
+        parent,
+        existing_untracked.files,
+        existing_untracked.dirs,
+    ))
 }
 
 /// Restore the working tree to match the provided ghost commit.
 pub fn restore_ghost_commit(repo_path: &Path, commit: &GhostCommit) -> Result<(), GitToolingError> {
-    restore_to_commit(repo_path, commit.id())
+    ensure_git_repository(repo_path)?;
+
+    let repo_root = resolve_repository_root(repo_path)?;
+    let repo_prefix = repo_subdir(repo_root.as_path(), repo_path);
+    let current_untracked =
+        capture_existing_untracked(repo_root.as_path(), repo_prefix.as_deref())?;
+    restore_to_commit_inner(repo_root.as_path(), repo_prefix.as_deref(), commit.id())?;
+    remove_new_untracked(
+        repo_root.as_path(),
+        commit.preexisting_untracked_files(),
+        commit.preexisting_untracked_dirs(),
+        current_untracked,
+    )
 }
 
 /// Restore the working tree to match the given commit ID.
@@ -141,7 +174,16 @@ pub fn restore_to_commit(repo_path: &Path, commit_id: &str) -> Result<(), GitToo
 
     let repo_root = resolve_repository_root(repo_path)?;
     let repo_prefix = repo_subdir(repo_root.as_path(), repo_path);
+    restore_to_commit_inner(repo_root.as_path(), repo_prefix.as_deref(), commit_id)
+}
 
+/// Restores the working tree and index to the given commit using `git restore`.
+/// The repository root and optional repository-relative prefix limit the restore scope.
+fn restore_to_commit_inner(
+    repo_root: &Path,
+    repo_prefix: Option<&Path>,
+    commit_id: &str,
+) -> Result<(), GitToolingError> {
     let mut restore_args = vec![
         OsString::from("restore"),
         OsString::from("--source"),
@@ -150,13 +192,143 @@ pub fn restore_to_commit(repo_path: &Path, commit_id: &str) -> Result<(), GitToo
         OsString::from("--staged"),
         OsString::from("--"),
     ];
-    if let Some(prefix) = repo_prefix.as_deref() {
+    if let Some(prefix) = repo_prefix {
         restore_args.push(prefix.as_os_str().to_os_string());
     } else {
         restore_args.push(OsString::from("."));
     }
 
-    run_git_for_status(repo_root.as_path(), restore_args, None)?;
+    run_git_for_status(repo_root, restore_args, None)?;
+    Ok(())
+}
+
+#[derive(Default)]
+struct UntrackedSnapshot {
+    files: Vec<PathBuf>,
+    dirs: Vec<PathBuf>,
+}
+
+/// Captures the untracked and ignored entries under `repo_root`, optionally limited by `repo_prefix`.
+/// Returns the result as an `UntrackedSnapshot`.
+fn capture_existing_untracked(
+    repo_root: &Path,
+    repo_prefix: Option<&Path>,
+) -> Result<UntrackedSnapshot, GitToolingError> {
+    // Ask git for the zero-delimited porcelain status so we can enumerate
+    // every untracked or ignored path (including ones filtered by prefix).
+    let mut args = vec![
+        OsString::from("status"),
+        OsString::from("--porcelain=2"),
+        OsString::from("-z"),
+        OsString::from("--ignored=matching"),
+        OsString::from("--untracked-files=all"),
+    ];
+    if let Some(prefix) = repo_prefix {
+        args.push(OsString::from("--"));
+        args.push(prefix.as_os_str().to_os_string());
+    }
+
+    let output = run_git_for_stdout_all(repo_root, args, None)?;
+    if output.is_empty() {
+        return Ok(UntrackedSnapshot::default());
+    }
+
+    let mut snapshot = UntrackedSnapshot::default();
+    // Each entry is of the form "<code> <path>" where code is '?' (untracked)
+    // or '!' (ignored); everything else is irrelevant to this snapshot.
+    for entry in output.split('\0') {
+        if entry.is_empty() {
+            continue;
+        }
+        let mut parts = entry.splitn(2, ' ');
+        let code = parts.next();
+        let path_part = parts.next();
+        let (Some(code), Some(path_part)) = (code, path_part) else {
+            continue;
+        };
+        if code != "?" && code != "!" {
+            continue;
+        }
+        if path_part.is_empty() {
+            continue;
+        }
+
+        let normalized = normalize_relative_path(Path::new(path_part))?;
+        let absolute = repo_root.join(&normalized);
+        let is_dir = absolute.is_dir();
+        if is_dir {
+            snapshot.dirs.push(normalized);
+        } else {
+            snapshot.files.push(normalized);
+        }
+    }
+
+    Ok(snapshot)
+}
+
+/// Removes untracked files and directories that were not present when the snapshot was captured.
+fn remove_new_untracked(
+    repo_root: &Path,
+    preserved_files: &[PathBuf],
+    preserved_dirs: &[PathBuf],
+    current: UntrackedSnapshot,
+) -> Result<(), GitToolingError> {
+    if current.files.is_empty() && current.dirs.is_empty() {
+        return Ok(());
+    }
+
+    let preserved_file_set: HashSet<PathBuf> = preserved_files.iter().cloned().collect();
+    let preserved_dirs_vec: Vec<PathBuf> = preserved_dirs.to_vec();
+
+    for path in current.files {
+        if should_preserve(&path, &preserved_file_set, &preserved_dirs_vec) {
+            continue;
+        }
+        remove_path(&repo_root.join(&path))?;
+    }
+
+    for dir in current.dirs {
+        if should_preserve(&dir, &preserved_file_set, &preserved_dirs_vec) {
+            continue;
+        }
+        remove_path(&repo_root.join(&dir))?;
+    }
+
+    Ok(())
+}
+
+/// Determines whether an untracked path should be kept because it existed in the snapshot.
+fn should_preserve(
+    path: &Path,
+    preserved_files: &HashSet<PathBuf>,
+    preserved_dirs: &[PathBuf],
+) -> bool {
+    if preserved_files.contains(path) {
+        return true;
+    }
+
+    preserved_dirs
+        .iter()
+        .any(|dir| path.starts_with(dir.as_path()))
+}
+
+/// Deletes the file or directory at the provided path, ignoring if it is already absent.
+fn remove_path(path: &Path) -> Result<(), GitToolingError> {
+    match fs::symlink_metadata(path) {
+        Ok(metadata) => {
+            if metadata.is_dir() {
+                fs::remove_dir_all(path)?;
+            } else {
+                fs::remove_file(path)?;
+            }
+        }
+        Err(err) => {
+            if err.kind() == io::ErrorKind::NotFound {
+                return Ok(());
+            }
+            return Err(err.into());
+        }
+    }
     Ok(())
 }
 
@@ -239,6 +411,9 @@ mod tests {
             ],
         );
 
+        let preexisting_untracked = repo.join("notes.txt");
+        std::fs::write(&preexisting_untracked, "notes before\n")?;
+
         let tracked_contents = "modified contents\n";
         std::fs::write(repo.join("tracked.txt"), tracked_contents)?;
         std::fs::remove_file(repo.join("delete-me.txt"))?;
@@ -267,6 +442,7 @@ mod tests {
         std::fs::write(repo.join("ignored.txt"), "changed\n")?;
         std::fs::remove_file(repo.join("new-file.txt"))?;
         std::fs::write(repo.join("ephemeral.txt"), "temp data\n")?;
+        std::fs::write(&preexisting_untracked, "notes after\n")?;
 
         restore_ghost_commit(repo, &ghost)?;
 
@@ -277,7 +453,9 @@ mod tests {
         let new_file_after = std::fs::read_to_string(repo.join("new-file.txt"))?;
         assert_eq!(new_file_after, new_file_contents);
         assert_eq!(repo.join("delete-me.txt").exists(), false);
-        assert!(repo.join("ephemeral.txt").exists());
+        assert!(!repo.join("ephemeral.txt").exists());
+        let notes_after = std::fs::read_to_string(&preexisting_untracked)?;
+        assert_eq!(notes_after, "notes before\n");
 
         Ok(())
     }
@@ -488,7 +666,43 @@ mod tests {
         assert!(vscode.join("settings.json").exists());
         let settings_after = std::fs::read_to_string(vscode.join("settings.json"))?;
         assert_eq!(settings_after, "{\n  \"after\": true\n}\n");
-        assert!(repo.join("temp.txt").exists());
+        assert!(!repo.join("temp.txt").exists());
+
+        Ok(())
+    }
+
+    #[test]
+    /// Restoring removes ignored directories created after the snapshot.
+    fn restore_removes_new_ignored_directory() -> Result<(), GitToolingError> {
+        let temp = tempfile::tempdir()?;
+        let repo = temp.path();
+        init_test_repo(repo);
+
+        std::fs::write(repo.join(".gitignore"), ".vscode/\n")?;
+        std::fs::write(repo.join("tracked.txt"), "snapshot version\n")?;
+        run_git_in(repo, &["add", ".gitignore", "tracked.txt"]);
+        run_git_in(
+            repo,
+            &[
+                "-c",
+                "user.name=Tester",
+                "-c",
+                "user.email=test@example.com",
+                "commit",
+                "-m",
+                "initial",
+            ],
+        );
+
+        let ghost = create_ghost_commit(&CreateGhostCommitOptions::new(repo))?;
+
+        let vscode = repo.join(".vscode");
+        std::fs::create_dir_all(&vscode)?;
+        std::fs::write(vscode.join("settings.json"), "{\n  \"after\": true\n}\n")?;
+
+        restore_ghost_commit(repo, &ghost)?;
+
+        assert!(!vscode.exists());
 
         Ok(())
     }

codex-rs/git-tooling/src/lib.rs

@@ -1,4 +1,5 @@
 use std::fmt;
+use std::path::PathBuf;
 
 mod errors;
 mod ghost_commits;
@@ -11,18 +12,36 @@ pub use ghost_commits::create_ghost_commit;
 pub use ghost_commits::restore_ghost_commit;
 pub use ghost_commits::restore_to_commit;
 pub use platform::create_symlink;
+use schemars::JsonSchema;
+use serde::Deserialize;
+use serde::Serialize;
+use ts_rs::TS;
+
+type CommitID = String;
 
 /// Details of a ghost commit created from a repository state.
-#[derive(Debug, Clone, PartialEq, Eq)]
+#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize, JsonSchema, TS)]
 pub struct GhostCommit {
-    id: String,
-    parent: Option<String>,
+    id: CommitID,
+    parent: Option<CommitID>,
+    preexisting_untracked_files: Vec<PathBuf>,
+    preexisting_untracked_dirs: Vec<PathBuf>,
 }
 
 impl GhostCommit {
     /// Create a new ghost commit wrapper from a raw commit ID and optional parent.
-    pub fn new(id: String, parent: Option<String>) -> Self {
-        Self { id, parent }
+    pub fn new(
+        id: CommitID,
+        parent: Option<CommitID>,
+        preexisting_untracked_files: Vec<PathBuf>,
+        preexisting_untracked_dirs: Vec<PathBuf>,
+    ) -> Self {
+        Self {
+            id,
+            parent,
+            preexisting_untracked_files,
+            preexisting_untracked_dirs,
+        }
     }
 
     /// Commit ID for the snapshot.
@@ -34,6 +53,16 @@ impl GhostCommit {
     pub fn parent(&self) -> Option<&str> {
         self.parent.as_deref()
     }
+
+    /// Untracked or ignored files that already existed when the snapshot was captured.
+    pub fn preexisting_untracked_files(&self) -> &[PathBuf] {
+        &self.preexisting_untracked_files
+    }
+
+    /// Untracked or ignored directories that already existed when the snapshot was captured.
+    pub fn preexisting_untracked_dirs(&self) -> &[PathBuf] {
+        &self.preexisting_untracked_dirs
+    }
 }
 
 impl fmt::Display for GhostCommit {

codex-rs/git-tooling/src/operations.rs

@@ -161,6 +161,27 @@ where
         })
 }
 
+/// Executes `git` and returns the full stdout without trimming so callers
+/// can parse delimiter-sensitive output, propagating UTF-8 errors with context.
+pub(crate) fn run_git_for_stdout_all<I, S>(
+    dir: &Path,
+    args: I,
+    env: Option<&[(OsString, OsString)]>,
+) -> Result<String, GitToolingError>
+where
+    I: IntoIterator<Item = S>,
+    S: AsRef<OsStr>,
+{
+    // Keep the raw stdout untouched so callers can parse delimiter-sensitive
+    // output (e.g. NUL-separated paths) without trimming artefacts.
+    let run = run_git(dir, args, env)?;
+    // Propagate UTF-8 conversion failures with the command context for debugging.
+    String::from_utf8(run.output.stdout).map_err(|source| GitToolingError::GitOutputUtf8 {
+        command: run.command,
+        source,
+    })
+}
+
 fn run_git<I, S>(
     dir: &Path,
     args: I,

codex-rs/keyring-store/Cargo.toml

@@ -0,0 +1,11 @@
+[package]
+edition = "2024"
+name = "codex-keyring-store"
+version = { workspace = true }
+
+[lints]
+workspace = true
+
+[dependencies]
+keyring = { workspace = true }
+tracing = { workspace = true }

codex-rs/keyring-store/src/lib.rs

@@ -0,0 +1,226 @@
+use keyring::Entry;
+use keyring::Error as KeyringError;
+use std::error::Error;
+use std::fmt;
+use std::fmt::Debug;
+use tracing::trace;
+
+#[derive(Debug)]
+pub enum CredentialStoreError {
+    Other(KeyringError),
+}
+
+impl CredentialStoreError {
+    pub fn new(error: KeyringError) -> Self {
+        Self::Other(error)
+    }
+
+    pub fn message(&self) -> String {
+        match self {
+            Self::Other(error) => error.to_string(),
+        }
+    }
+
+    pub fn into_error(self) -> KeyringError {
+        match self {
+            Self::Other(error) => error,
+        }
+    }
+}
+
+impl fmt::Display for CredentialStoreError {
+    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
+        match self {
+            Self::Other(error) => write!(f, "{error}"),
+        }
+    }
+}
+
+impl Error for CredentialStoreError {}
+
+/// Shared credential store abstraction for keyring-backed implementations.
+pub trait KeyringStore: Debug + Send + Sync {
+    fn load(&self, service: &str, account: &str) -> Result<Option<String>, CredentialStoreError>;
+    fn save(&self, service: &str, account: &str, value: &str) -> Result<(), CredentialStoreError>;
+    fn delete(&self, service: &str, account: &str) -> Result<bool, CredentialStoreError>;
+}
+
+#[derive(Debug)]
+pub struct DefaultKeyringStore;
+
+impl KeyringStore for DefaultKeyringStore {
+    fn load(&self, service: &str, account: &str) -> Result<Option<String>, CredentialStoreError> {
+        trace!("keyring.load start, service={service}, account={account}");
+        let entry = Entry::new(service, account).map_err(CredentialStoreError::new)?;
+        match entry.get_password() {
+            Ok(password) => {
+                trace!("keyring.load success, service={service}, account={account}");
+                Ok(Some(password))
+            }
+            Err(keyring::Error::NoEntry) => {
+                trace!("keyring.load no entry, service={service}, account={account}");
+                Ok(None)
+            }
+            Err(error) => {
+                trace!("keyring.load error, service={service}, account={account}, error={error}");
+                Err(CredentialStoreError::new(error))
+            }
+        }
+    }
+
+    fn save(&self, service: &str, account: &str, value: &str) -> Result<(), CredentialStoreError> {
+        trace!(
+            "keyring.save start, service={service}, account={account}, value_len={}",
+            value.len()
+        );
+        let entry = Entry::new(service, account).map_err(CredentialStoreError::new)?;
+        match entry.set_password(value) {
+            Ok(()) => {
+                trace!("keyring.save success, service={service}, account={account}");
+                Ok(())
+            }
+            Err(error) => {
+                trace!("keyring.save error, service={service}, account={account}, error={error}");
+                Err(CredentialStoreError::new(error))
+            }
+        }
+    }
+
+    fn delete(&self, service: &str, account: &str) -> Result<bool, CredentialStoreError> {
+        trace!("keyring.delete start, service={service}, account={account}");
+        let entry = Entry::new(service, account).map_err(CredentialStoreError::new)?;
+        match entry.delete_credential() {
+            Ok(()) => {
+                trace!("keyring.delete success, service={service}, account={account}");
+                Ok(true)
+            }
+            Err(keyring::Error::NoEntry) => {
+                trace!("keyring.delete no entry, service={service}, account={account}");
+                Ok(false)
+            }
+            Err(error) => {
+                trace!("keyring.delete error, service={service}, account={account}, error={error}");
+                Err(CredentialStoreError::new(error))
+            }
+        }
+    }
+}
+
+pub mod tests {
+    use super::CredentialStoreError;
+    use super::KeyringStore;
+    use keyring::Error as KeyringError;
+    use keyring::credential::CredentialApi as _;
+    use keyring::mock::MockCredential;
+    use std::collections::HashMap;
+    use std::sync::Arc;
+    use std::sync::Mutex;
+    use std::sync::PoisonError;
+
+    #[derive(Default, Clone, Debug)]
+    pub struct MockKeyringStore {
+        credentials: Arc<Mutex<HashMap<String, Arc<MockCredential>>>>,
+    }
+
+    impl MockKeyringStore {
+        pub fn credential(&self, account: &str) -> Arc<MockCredential> {
+            let mut guard = self
+                .credentials
+                .lock()
+                .unwrap_or_else(PoisonError::into_inner);
+            guard
+                .entry(account.to_string())
+                .or_insert_with(|| Arc::new(MockCredential::default()))
+                .clone()
+        }
+
+        pub fn saved_value(&self, account: &str) -> Option<String> {
+            let credential = {
+                let guard = self
+                    .credentials
+                    .lock()
+                    .unwrap_or_else(PoisonError::into_inner);
+                guard.get(account).cloned()
+            }?;
+            credential.get_password().ok()
+        }
+
+        pub fn set_error(&self, account: &str, error: KeyringError) {
+            let credential = self.credential(account);
+            credential.set_error(error);
+        }
+
+        pub fn contains(&self, account: &str) -> bool {
+            let guard = self
+                .credentials
+                .lock()
+                .unwrap_or_else(PoisonError::into_inner);
+            guard.contains_key(account)
+        }
+    }
+
+    impl KeyringStore for MockKeyringStore {
+        fn load(
+            &self,
+            _service: &str,
+            account: &str,
+        ) -> Result<Option<String>, CredentialStoreError> {
+            let credential = {
+                let guard = self
+                    .credentials
+                    .lock()
+                    .unwrap_or_else(PoisonError::into_inner);
+                guard.get(account).cloned()
+            };
+
+            let Some(credential) = credential else {
+                return Ok(None);
+            };
+
+            match credential.get_password() {
+                Ok(password) => Ok(Some(password)),
+                Err(KeyringError::NoEntry) => Ok(None),
+                Err(error) => Err(CredentialStoreError::new(error)),
+            }
+        }
+
+        fn save(
+            &self,
+            _service: &str,
+            account: &str,
+            value: &str,
+        ) -> Result<(), CredentialStoreError> {
+            let credential = self.credential(account);
+            credential
+                .set_password(value)
+                .map_err(CredentialStoreError::new)
+        }
+
+        fn delete(&self, _service: &str, account: &str) -> Result<bool, CredentialStoreError> {
+            let credential = {
+                let guard = self
+                    .credentials
+                    .lock()
+                    .unwrap_or_else(PoisonError::into_inner);
+                guard.get(account).cloned()
+            };
+
+            let Some(credential) = credential else {
+                return Ok(false);
+            };
+
+            let removed = match credential.delete_credential() {
+                Ok(()) => Ok(true),
+                Err(KeyringError::NoEntry) => Ok(false),
+                Err(error) => Err(CredentialStoreError::new(error)),
+            }?;
+
+            let mut guard = self
+                .credentials
+                .lock()
+                .unwrap_or_else(PoisonError::into_inner);
+            guard.remove(account);
+            Ok(removed)
+        }
+    }
+}

codex-rs/login/src/lib.rs

@@ -16,9 +16,7 @@ pub use codex_core::auth::AuthDotJson;
 pub use codex_core::auth::CLIENT_ID;
 pub use codex_core::auth::CODEX_API_KEY_ENV_VAR;
 pub use codex_core::auth::OPENAI_API_KEY_ENV_VAR;
-pub use codex_core::auth::get_auth_file;
 pub use codex_core::auth::login_with_api_key;
 pub use codex_core::auth::logout;
-pub use codex_core::auth::try_read_auth_json;
-pub use codex_core::auth::write_auth_json;
+pub use codex_core::auth::save_auth;
 pub use codex_core::token_data::TokenData;

codex-rs/login/src/server.rs

@@ -15,7 +15,7 @@ use crate::pkce::generate_pkce;
 use base64::Engine;
 use chrono::Utc;
 use codex_core::auth::AuthDotJson;
-use codex_core::auth::get_auth_file;
+use codex_core::auth::save_auth;
 use codex_core::default_client::originator;
 use codex_core::token_data::TokenData;
 use codex_core::token_data::parse_id_token;
@@ -540,13 +540,6 @@ pub(crate) async fn persist_tokens_async(
     // Reuse existing synchronous logic but run it off the async runtime.
     let codex_home = codex_home.to_path_buf();
     tokio::task::spawn_blocking(move || {
-        let auth_file = get_auth_file(&codex_home);
-        if let Some(parent) = auth_file.parent()
-            && !parent.exists()
-        {
-            std::fs::create_dir_all(parent).map_err(io::Error::other)?;
-        }
-
         let mut tokens = TokenData {
             id_token: parse_id_token(&id_token).map_err(io::Error::other)?,
             access_token,
@@ -564,7 +557,7 @@ pub(crate) async fn persist_tokens_async(
             tokens: Some(tokens),
             last_refresh: Some(Utc::now()),
         };
-        codex_core::auth::write_auth_json(&auth_file, &auth)
+        save_auth(&codex_home, &auth)
     })
     .await
     .map_err(|e| io::Error::other(format!("persist task failed: {e}")))?

codex-rs/login/tests/suite/device_code_login.rs

@@ -1,9 +1,9 @@
 #![allow(clippy::unwrap_used)]
 
+use anyhow::Context;
 use base64::Engine;
 use base64::engine::general_purpose::URL_SAFE_NO_PAD;
-use codex_core::auth::get_auth_file;
-use codex_core::auth::try_read_auth_json;
+use codex_core::auth::load_auth_dot_json;
 use codex_login::ServerOptions;
 use codex_login::run_device_code_login;
 use serde_json::json;
@@ -108,8 +108,8 @@ fn server_opts(codex_home: &tempfile::TempDir, issuer: String) -> ServerOptions
 }
 
 #[tokio::test]
-async fn device_code_login_integration_succeeds() {
-    skip_if_no_network!();
+async fn device_code_login_integration_succeeds() -> anyhow::Result<()> {
+    skip_if_no_network!(Ok(()));
 
     let codex_home = tempdir().unwrap();
     let mock_server = MockServer::start().await;
@@ -133,19 +133,21 @@ async fn device_code_login_integration_succeeds() {
         .await
         .expect("device code login integration should succeed");
 
-    let auth_path = get_auth_file(codex_home.path());
-    let auth = try_read_auth_json(&auth_path).expect("auth.json written");
+    let auth = load_auth_dot_json(codex_home.path())
+        .context("auth.json should load after login succeeds")?
+        .context("auth.json written")?;
     // assert_eq!(auth.openai_api_key.as_deref(), Some("api-key-321"));
     let tokens = auth.tokens.expect("tokens persisted");
     assert_eq!(tokens.access_token, "access-token-123");
     assert_eq!(tokens.refresh_token, "refresh-token-123");
     assert_eq!(tokens.id_token.raw_jwt, jwt);
     assert_eq!(tokens.account_id.as_deref(), Some("acct_321"));
+    Ok(())
 }
 
 #[tokio::test]
-async fn device_code_login_rejects_workspace_mismatch() {
-    skip_if_no_network!();
+async fn device_code_login_rejects_workspace_mismatch() -> anyhow::Result<()> {
+    skip_if_no_network!(Ok(()));
 
     let codex_home = tempdir().unwrap();
     let mock_server = MockServer::start().await;
@@ -172,16 +174,18 @@ async fn device_code_login_rejects_workspace_mismatch() {
         .expect_err("device code login should fail when workspace mismatches");
     assert_eq!(err.kind(), std::io::ErrorKind::PermissionDenied);
 
-    let auth_path = get_auth_file(codex_home.path());
+    let auth =
+        load_auth_dot_json(codex_home.path()).context("auth.json should load after login fails")?;
     assert!(
-        !auth_path.exists(),
+        auth.is_none(),
         "auth.json should not be created when workspace validation fails"
     );
+    Ok(())
 }
 
 #[tokio::test]
-async fn device_code_login_integration_handles_usercode_http_failure() {
-    skip_if_no_network!();
+async fn device_code_login_integration_handles_usercode_http_failure() -> anyhow::Result<()> {
+    skip_if_no_network!(Ok(()));
 
     let codex_home = tempdir().unwrap();
     let mock_server = MockServer::start().await;
@@ -201,13 +205,19 @@ async fn device_code_login_integration_handles_usercode_http_failure() {
         "unexpected error: {err:?}"
     );
 
-    let auth_path = get_auth_file(codex_home.path());
-    assert!(!auth_path.exists());
+    let auth =
+        load_auth_dot_json(codex_home.path()).context("auth.json should load after login fails")?;
+    assert!(
+        auth.is_none(),
+        "auth.json should not be created when login fails"
+    );
+    Ok(())
 }
 
 #[tokio::test]
-async fn device_code_login_integration_persists_without_api_key_on_exchange_failure() {
-    skip_if_no_network!();
+async fn device_code_login_integration_persists_without_api_key_on_exchange_failure()
+-> anyhow::Result<()> {
+    skip_if_no_network!(Ok(()));
 
     let codex_home = tempdir().unwrap();
 
@@ -235,18 +245,20 @@ async fn device_code_login_integration_persists_without_api_key_on_exchange_fail
         .await
         .expect("device login should succeed without API key exchange");
 
-    let auth_path = get_auth_file(codex_home.path());
-    let auth = try_read_auth_json(&auth_path).expect("auth.json written");
+    let auth = load_auth_dot_json(codex_home.path())
+        .context("auth.json should load after login succeeds")?
+        .context("auth.json written")?;
     assert!(auth.openai_api_key.is_none());
     let tokens = auth.tokens.expect("tokens persisted");
     assert_eq!(tokens.access_token, "access-token-123");
     assert_eq!(tokens.refresh_token, "refresh-token-123");
     assert_eq!(tokens.id_token.raw_jwt, jwt);
+    Ok(())
 }
 
 #[tokio::test]
-async fn device_code_login_integration_handles_error_payload() {
-    skip_if_no_network!();
+async fn device_code_login_integration_handles_error_payload() -> anyhow::Result<()> {
+    skip_if_no_network!(Ok(()));
 
     let codex_home = tempdir().unwrap();
 
@@ -288,9 +300,11 @@ async fn device_code_login_integration_handles_error_payload() {
         "Expected an authorization_declined / 400 / 404 error, got {err:?}"
     );
 
-    let auth_path = get_auth_file(codex_home.path());
+    let auth =
+        load_auth_dot_json(codex_home.path()).context("auth.json should load after login fails")?;
     assert!(
-        !auth_path.exists(),
+        auth.is_none(),
         "auth.json should not be created when device auth fails"
     );
+    Ok(())
 }

codex-rs/mcp-server/src/codex_tool_config.rs

@@ -158,6 +158,7 @@ impl CodexToolCallParam {
             include_view_image_tool: None,
             show_raw_agent_reasoning: None,
             tools_web_search_request: None,
+            experimental_sandbox_command_assessment: None,
             additional_writable_roots: Vec::new(),
         };
 

codex-rs/mcp-server/src/codex_tool_runner.rs

@@ -178,6 +178,7 @@ async fn run_codex_tool_session_inner(
                         cwd,
                         call_id,
                         reason: _,
+                        risk,
                         parsed_cmd,
                     }) => {
                         handle_exec_approval_request(
@@ -190,6 +191,7 @@ async fn run_codex_tool_session_inner(
                             event.id.clone(),
                             call_id,
                             parsed_cmd,
+                            risk,
                         )
                         .await;
                         continue;
@@ -279,13 +281,15 @@ async fn run_codex_tool_session_inner(
                     | EventMsg::GetHistoryEntryResponse(_)
                     | EventMsg::PlanUpdate(_)
                     | EventMsg::TurnAborted(_)
-                    | EventMsg::ConversationPath(_)
                     | EventMsg::UserMessage(_)
                     | EventMsg::ShutdownComplete
                     | EventMsg::ViewImageToolCall(_)
+                    | EventMsg::RawResponseItem(_)
                     | EventMsg::EnteredReviewMode(_)
                     | EventMsg::ItemStarted(_)
                     | EventMsg::ItemCompleted(_)
+                    | EventMsg::UndoStarted(_)
+                    | EventMsg::UndoCompleted(_)
                     | EventMsg::ExitedReviewMode(_) => {
                         // For now, we do not do anything extra for these
                         // events. Note that

codex-rs/mcp-server/src/exec_approval.rs

@@ -4,6 +4,7 @@ use std::sync::Arc;
 use codex_core::CodexConversation;
 use codex_core::protocol::Op;
 use codex_core::protocol::ReviewDecision;
+use codex_core::protocol::SandboxCommandAssessment;
 use codex_protocol::parse_command::ParsedCommand;
 use mcp_types::ElicitRequest;
 use mcp_types::ElicitRequestParamsRequestedSchema;
@@ -37,6 +38,8 @@ pub struct ExecApprovalElicitRequestParams {
     pub codex_command: Vec<String>,
     pub codex_cwd: PathBuf,
     pub codex_parsed_cmd: Vec<ParsedCommand>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub codex_risk: Option<SandboxCommandAssessment>,
 }
 
 // TODO(mbolin): ExecApprovalResponse does not conform to ElicitResult. See:
@@ -59,6 +62,7 @@ pub(crate) async fn handle_exec_approval_request(
     event_id: String,
     call_id: String,
     codex_parsed_cmd: Vec<ParsedCommand>,
+    codex_risk: Option<SandboxCommandAssessment>,
 ) {
     let escaped_command =
         shlex::try_join(command.iter().map(String::as_str)).unwrap_or_else(|_| command.join(" "));
@@ -81,6 +85,7 @@ pub(crate) async fn handle_exec_approval_request(
         codex_command: command,
         codex_cwd: cwd,
         codex_parsed_cmd,
+        codex_risk,
     };
     let params_json = match serde_json::to_value(&params) {
         Ok(value) => value,