Derive required Node version from package.json

As explained on https://crates.io/crates/regex-lite, `regex-lite` is a
lighter alternative to `regex` and seems to be sufficient for our
purposes.

.github/actions/codex/action.yml

@@ -15,14 +15,14 @@ inputs:
   codex_args:
     description: "A whitespace-delimited list of arguments to pass to Codex. Due to limitations in YAML, arguments with spaces are not supported. For more complex configurations, use the `codex_home` input."
     required: false
-    default: "--full-auto"
+    default: "--config hide_agent_reasoning=true --full-auto"
   codex_home:
     description: "Value to use as the CODEX_HOME environment variable when running Codex."
     required: false
   codex_release_tag:
     description: "The release tag of the Codex model to run."
     required: false
-    default: "codex-rs-d519bd8bbd1e1fd9efdc5d68cf7bebdec0dd0f28-1-rust-v0.0.2505270918"
+    default: "codex-rs-ca8e97fcbcb991e542b8689f2d4eab9d30c399d6-1-rust-v0.0.2505302325"
 
 runs:
   using: "composite"

codex-cli/bin/codex.js

@@ -81,6 +81,18 @@ if (wantsNative) {
   process.exit(exitCode);
 }
 
+// Ensure a compatible Node version before falling back to the JavaScript CLI.
+const pkgJsonPath = path.join(__dirname, "../package.json");
+const { engines } = JSON.parse(fs.readFileSync(pkgJsonPath, "utf8"));
+const requiredNodeMajor = Number.parseInt(engines.node.match(/\d+/)[0], 10);
+const currentMajor = Number.parseInt(process.versions.node.split(".")[0], 10);
+if (currentMajor < requiredNodeMajor) {
+  console.error(
+    `Codex CLI requires Node.js ${requiredNodeMajor} or newer. Detected ${process.version}. Please upgrade your Node.js installation.`
+  );
+  process.exit(1);
+}
+
 // Fallback: execute the original JavaScript CLI.
 
 // Resolve the path to the compiled CLI bundle

codex-cli/scripts/install_native_deps.sh

@@ -65,7 +65,7 @@ mkdir -p "$BIN_DIR"
 # Until we start publishing stable GitHub releases, we have to grab the binaries
 # from the GitHub Action that created them. Update the URL below to point to the
 # appropriate workflow run:
-WORKFLOW_URL="https://github.com/openai/codex/actions/runs/15334411824"
+WORKFLOW_URL="https://github.com/openai/codex/actions/runs/15361005231"
 WORKFLOW_ID="${WORKFLOW_URL##*/}"
 
 ARTIFACTS_DIR="$(mktemp -d)"

codex-rs/Cargo.lock

@@ -567,7 +567,6 @@ version = "0.0.0"
 dependencies = [
  "anyhow",
  "pretty_assertions",
- "regex",
  "serde_json",
  "similar",
  "tempfile",
@@ -635,6 +634,8 @@ dependencies = [
  "seccompiler",
  "serde",
  "serde_json",
+ "strum 0.27.1",
+ "strum_macros 0.27.1",
  "tempfile",
  "thiserror 2.0.12",
  "time",
@@ -680,7 +681,7 @@ dependencies = [
  "log",
  "multimap",
  "path-absolutize",
- "regex",
+ "regex-lite",
  "serde",
  "serde_json",
  "serde_with",
@@ -755,7 +756,7 @@ dependencies = [
  "pretty_assertions",
  "ratatui",
  "ratatui-image",
- "regex",
+ "regex-lite",
  "serde_json",
  "shlex",
  "strum 0.27.1",
@@ -3321,6 +3322,12 @@ dependencies = [
  "regex-syntax 0.8.5",
 ]
 
+[[package]]
+name = "regex-lite"
+version = "0.1.6"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "53a49587ad06b26609c52e423de037e7f57f20d53535d66e08c695f347df952a"
+
 [[package]]
 name = "regex-syntax"
 version = "0.6.29"

codex-rs/apply-patch/Cargo.toml

@@ -12,7 +12,6 @@ workspace = true
 
 [dependencies]
 anyhow = "1"
-regex = "1.11.1"
 serde_json = "1.0.110"
 similar = "2.7.0"
 thiserror = "2.0.12"

codex-rs/config.md

@@ -142,6 +142,34 @@ Users can specify config values at multiple levels. Order of precedence is as fo
 3. as an entry in `config.toml`, e.g., `model = "o3"`
 4. the default value that comes with Codex CLI (i.e., Codex CLI defaults to `codex-mini-latest`)
 
+## model_reasoning_effort
+
+If the model name starts with `"o"` (as in `"o3"` or `"o4-mini"`) or `"codex"`, reasoning is enabled by default when using the Responses API. As explained in the [OpenAI Platform documentation](https://platform.openai.com/docs/guides/reasoning?api-mode=responses#get-started-with-reasoning), this can be set to:
+
+- `"low"`
+- `"medium"` (default)
+- `"high"`
+
+To disable reasoning, set `model_reasoning_effort` to `"none"` in your config:
+
+```toml
+model_reasoning_effort = "none"  # disable reasoning
+```
+
+## model_reasoning_summary
+
+If the model name starts with `"o"` (as in `"o3"` or `"o4-mini"`) or `"codex"`, reasoning is enabled by default when using the Responses API. As explained in the [OpenAI Platform documentation](https://platform.openai.com/docs/guides/reasoning?api-mode=responses#reasoning-summaries), this can be set to:
+
+- `"auto"` (default)
+- `"concise"`
+- `"detailed"`
+
+To disable reasoning summaries, set `model_reasoning_summary` to `"none"` in your config:
+
+```toml
+model_reasoning_summary = "none"  # disable reasoning summaries
+```
+
 ## sandbox_permissions
 
 List of permissions to grant to the sandbox that Codex uses to execute untrusted commands:

codex-rs/core/Cargo.toml

@@ -31,6 +31,8 @@ rand = "0.9"
 reqwest = { version = "0.12", features = ["json", "stream"] }
 serde = { version = "1", features = ["derive"] }
 serde_json = "1"
+strum = "0.27.1"
+strum_macros = "0.27.1"
 thiserror = "2.0.12"
 time = { version = "0.3", features = ["formatting", "local-offset", "macros"] }
 tokio = { version = "1", features = [

codex-rs/core/src/chat_completions.rs

@@ -28,8 +28,7 @@ use crate::models::ResponseItem;
 use crate::openai_tools::create_tools_json_for_chat_completions_api;
 use crate::util::backoff;
 
-/// Implementation for the classic Chat Completions API. This is intentionally
-/// minimal: we only stream back plain assistant text.
+/// Implementation for the classic Chat Completions API.
 pub(crate) async fn stream_chat_completions(
     prompt: &Prompt,
     model: &str,
@@ -43,17 +42,67 @@ pub(crate) async fn stream_chat_completions(
     messages.push(json!({"role": "system", "content": full_instructions}));
 
     for item in &prompt.input {
-        if let ResponseItem::Message { role, content } = item {
-            let mut text = String::new();
-            for c in content {
-                match c {
-                    ContentItem::InputText { text: t } | ContentItem::OutputText { text: t } => {
-                        text.push_str(t);
+        match item {
+            ResponseItem::Message { role, content } => {
+                let mut text = String::new();
+                for c in content {
+                    match c {
+                        ContentItem::InputText { text: t }
+                        | ContentItem::OutputText { text: t } => {
+                            text.push_str(t);
+                        }
+                        _ => {}
                     }
-                    _ => {}
                 }
+                messages.push(json!({"role": role, "content": text}));
+            }
+            ResponseItem::FunctionCall {
+                name,
+                arguments,
+                call_id,
+            } => {
+                messages.push(json!({
+                    "role": "assistant",
+                    "content": null,
+                    "tool_calls": [{
+                        "id": call_id,
+                        "type": "function",
+                        "function": {
+                            "name": name,
+                            "arguments": arguments,
+                        }
+                    }]
+                }));
+            }
+            ResponseItem::LocalShellCall {
+                id,
+                call_id: _,
+                status,
+                action,
+            } => {
+                // Confirm with API team.
+                messages.push(json!({
+                    "role": "assistant",
+                    "content": null,
+                    "tool_calls": [{
+                        "id": id.clone().unwrap_or_else(|| "".to_string()),
+                        "type": "local_shell_call",
+                        "status": status,
+                        "action": action,
+                    }]
+                }));
+            }
+            ResponseItem::FunctionCallOutput { call_id, output } => {
+                messages.push(json!({
+                    "role": "tool",
+                    "tool_call_id": call_id,
+                    "content": output.content,
+                }));
+            }
+            ResponseItem::Reasoning { .. } | ResponseItem::Other => {
+                // Omit these items from the conversation history.
+                continue;
             }
-            messages.push(json!({"role": role, "content": text}));
         }
     }
 
@@ -68,9 +117,8 @@ pub(crate) async fn stream_chat_completions(
     let base_url = provider.base_url.trim_end_matches('/');
     let url = format!("{}/chat/completions", base_url);
 
-    debug!(url, "POST (chat)");
-    trace!(
-        "request payload: {}",
+    debug!(
+        "POST to {url}: {}",
         serde_json::to_string_pretty(&payload).unwrap_or_default()
     );
 
@@ -140,6 +188,21 @@ where
 
     let idle_timeout = *OPENAI_STREAM_IDLE_TIMEOUT_MS;
 
+    // State to accumulate a function call across streaming chunks.
+    // OpenAI may split the `arguments` string over multiple `delta` events
+    // until the chunk whose `finish_reason` is `tool_calls` is emitted. We
+    // keep collecting the pieces here and forward a single
+    // `ResponseItem::FunctionCall` once the call is complete.
+    #[derive(Default)]
+    struct FunctionCallState {
+        name: Option<String>,
+        arguments: String,
+        call_id: Option<String>,
+        active: bool,
+    }
+
+    let mut fn_call_state = FunctionCallState::default();
+
     loop {
         let sse = match timeout(idle_timeout, stream.next()).await {
             Ok(Some(Ok(ev))) => ev,
@@ -179,23 +242,89 @@ where
             Ok(v) => v,
             Err(_) => continue,
         };
+        trace!("chat_completions received SSE chunk: {chunk:?}");
 
-        let content_opt = chunk
-            .get("choices")
-            .and_then(|c| c.get(0))
-            .and_then(|c| c.get("delta"))
-            .and_then(|d| d.get("content"))
-            .and_then(|c| c.as_str());
+        let choice_opt = chunk.get("choices").and_then(|c| c.get(0));
 
-        if let Some(content) = content_opt {
-            let item = ResponseItem::Message {
-                role: "assistant".to_string(),
-                content: vec![ContentItem::OutputText {
-                    text: content.to_string(),
-                }],
-            };
+        if let Some(choice) = choice_opt {
+            // Handle assistant content tokens.
+            if let Some(content) = choice
+                .get("delta")
+                .and_then(|d| d.get("content"))
+                .and_then(|c| c.as_str())
+            {
+                let item = ResponseItem::Message {
+                    role: "assistant".to_string(),
+                    content: vec![ContentItem::OutputText {
+                        text: content.to_string(),
+                    }],
+                };
 
-            let _ = tx_event.send(Ok(ResponseEvent::OutputItemDone(item))).await;
+                let _ = tx_event.send(Ok(ResponseEvent::OutputItemDone(item))).await;
+            }
+
+            // Handle streaming function / tool calls.
+            if let Some(tool_calls) = choice
+                .get("delta")
+                .and_then(|d| d.get("tool_calls"))
+                .and_then(|tc| tc.as_array())
+            {
+                if let Some(tool_call) = tool_calls.first() {
+                    // Mark that we have an active function call in progress.
+                    fn_call_state.active = true;
+
+                    // 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());
+                    }
+
+                    // 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 let Some(args_fragment) =
+                            function.get("arguments").and_then(|a| a.as_str())
+                        {
+                            fn_call_state.arguments.push_str(args_fragment);
+                        }
+                    }
+                }
+            }
+
+            // Emit end-of-turn when finish_reason signals completion.
+            if let Some(finish_reason) = choice.get("finish_reason").and_then(|v| v.as_str()) {
+                match finish_reason {
+                    "tool_calls" if fn_call_state.active => {
+                        // Build the FunctionCall response item.
+                        let item = ResponseItem::FunctionCall {
+                            name: fn_call_state.name.clone().unwrap_or_else(|| "".to_string()),
+                            arguments: fn_call_state.arguments.clone(),
+                            call_id: fn_call_state.call_id.clone().unwrap_or_else(String::new),
+                        };
+
+                        // Emit it downstream.
+                        let _ = tx_event.send(Ok(ResponseEvent::OutputItemDone(item))).await;
+                    }
+                    "stop" => {
+                        // Regular turn without tool-call.
+                    }
+                    _ => {}
+                }
+
+                // Emit Completed regardless of reason so the agent can advance.
+                let _ = tx_event
+                    .send(Ok(ResponseEvent::Completed {
+                        response_id: String::new(),
+                    }))
+                    .await;
+
+                // Prepare for potential next turn (should not happen in same stream).
+                // fn_call_state = FunctionCallState::default();
+
+                return; // End processing for this SSE stream.
+            }
         }
     }
 }
@@ -242,9 +371,14 @@ where
                 Poll::Ready(None) => return Poll::Ready(None),
                 Poll::Ready(Some(Err(e))) => return Poll::Ready(Some(Err(e))),
                 Poll::Ready(Some(Ok(ResponseEvent::OutputItemDone(item)))) => {
-                    // Accumulate *assistant* text but do not emit yet.
-                    if let crate::models::ResponseItem::Message { role, content } = &item {
-                        if role == "assistant" {
+                    // If this is an incremental assistant message chunk, accumulate but
+                    // do NOT emit yet. Forward any other item (e.g. FunctionCall) right
+                    // away so downstream consumers see it.
+
+                    let is_assistant_delta = matches!(&item, crate::models::ResponseItem::Message { role, .. } if role == "assistant");
+
+                    if is_assistant_delta {
+                        if let crate::models::ResponseItem::Message { content, .. } = &item {
                             if let Some(text) = content.iter().find_map(|c| match c {
                                 crate::models::ContentItem::OutputText { text } => Some(text),
                                 _ => None,
@@ -252,10 +386,13 @@ where
                                 this.cumulative.push_str(text);
                             }
                         }
+
+                        // Swallow partial assistant chunk; keep polling.
+                        continue;
                     }
 
-                    // Swallow partial event; keep polling.
-                    continue;
+                    // Not an assistant message – forward immediately.
+                    return Poll::Ready(Some(Ok(ResponseEvent::OutputItemDone(item))));
                 }
                 Poll::Ready(Some(Ok(ResponseEvent::Completed { response_id }))) => {
                     if !this.cumulative.is_empty() {

codex-rs/core/src/client.rs

@@ -18,12 +18,13 @@ use tracing::warn;
 
 use crate::chat_completions::AggregateStreamExt;
 use crate::chat_completions::stream_chat_completions;
-use crate::client_common::Payload;
 use crate::client_common::Prompt;
-use crate::client_common::Reasoning;
 use crate::client_common::ResponseEvent;
 use crate::client_common::ResponseStream;
-use crate::client_common::Summary;
+use crate::client_common::ResponsesApiRequest;
+use crate::client_common::create_reasoning_param_for_request;
+use crate::config_types::ReasoningEffort as ReasoningEffortConfig;
+use crate::config_types::ReasoningSummary as ReasoningSummaryConfig;
 use crate::error::CodexErr;
 use crate::error::EnvVarError;
 use crate::error::Result;
@@ -41,14 +42,23 @@ pub struct ModelClient {
     model: String,
     client: reqwest::Client,
     provider: ModelProviderInfo,
+    effort: ReasoningEffortConfig,
+    summary: ReasoningSummaryConfig,
 }
 
 impl ModelClient {
-    pub fn new(model: impl ToString, provider: ModelProviderInfo) -> Self {
+    pub fn new(
+        model: impl ToString,
+        provider: ModelProviderInfo,
+        effort: ReasoningEffortConfig,
+        summary: ReasoningSummaryConfig,
+    ) -> Self {
         Self {
             model: model.to_string(),
             client: reqwest::Client::new(),
             provider,
+            effort,
+            summary,
         }
     }
 
@@ -98,17 +108,15 @@ impl ModelClient {
 
         let full_instructions = prompt.get_full_instructions();
         let tools_json = create_tools_json_for_responses_api(prompt, &self.model)?;
-        let payload = Payload {
+        let reasoning = create_reasoning_param_for_request(&self.model, self.effort, self.summary);
+        let payload = ResponsesApiRequest {
             model: &self.model,
             instructions: &full_instructions,
             input: &prompt.input,
             tools: &tools_json,
             tool_choice: "auto",
             parallel_tool_calls: false,
-            reasoning: Some(Reasoning {
-                effort: "high",
-                summary: Some(Summary::Auto),
-            }),
+            reasoning,
             previous_response_id: prompt.prev_id.clone(),
             store: prompt.store,
             stream: true,
@@ -117,8 +125,7 @@ impl ModelClient {
         let base_url = self.provider.base_url.clone();
         let base_url = base_url.trim_end_matches('/');
         let url = format!("{}/responses", base_url);
-        debug!(url, "POST");
-        trace!("request payload: {}", serde_json::to_string(&payload)?);
+        trace!("POST to {url}: {}", serde_json::to_string(&payload)?);
 
         let mut attempt = 0;
         loop {
@@ -303,6 +310,19 @@ where
                     };
                 };
             }
+            "response.content_part.done"
+            | "response.created"
+            | "response.function_call_arguments.delta"
+            | "response.in_progress"
+            | "response.output_item.added"
+            | "response.output_text.delta"
+            | "response.output_text.done"
+            | "response.reasoning_summary_part.added"
+            | "response.reasoning_summary_text.delta"
+            | "response.reasoning_summary_text.done" => {
+                // Currently, we ignore these events, but we handle them
+                // separately to skip the logging message in the `other` case.
+            }
             other => debug!(other, "sse event"),
         }
     }

codex-rs/core/src/client_common.rs

@@ -1,3 +1,5 @@
+use crate::config_types::ReasoningEffort as ReasoningEffortConfig;
+use crate::config_types::ReasoningSummary as ReasoningSummaryConfig;
 use crate::error::Result;
 use crate::models::ResponseItem;
 use futures::Stream;
@@ -52,25 +54,59 @@ pub enum ResponseEvent {
 
 #[derive(Debug, Serialize)]
 pub(crate) struct Reasoning {
-    pub(crate) effort: &'static str,
+    pub(crate) effort: OpenAiReasoningEffort,
     #[serde(skip_serializing_if = "Option::is_none")]
-    pub(crate) summary: Option<Summary>,
+    pub(crate) summary: Option<OpenAiReasoningSummary>,
+}
+
+/// See https://platform.openai.com/docs/guides/reasoning?api-mode=responses#get-started-with-reasoning
+#[derive(Debug, Serialize, Default, Clone, Copy)]
+#[serde(rename_all = "lowercase")]
+pub(crate) enum OpenAiReasoningEffort {
+    Low,
+    #[default]
+    Medium,
+    High,
+}
+
+impl From<ReasoningEffortConfig> for Option<OpenAiReasoningEffort> {
+    fn from(effort: ReasoningEffortConfig) -> Self {
+        match effort {
+            ReasoningEffortConfig::Low => Some(OpenAiReasoningEffort::Low),
+            ReasoningEffortConfig::Medium => Some(OpenAiReasoningEffort::Medium),
+            ReasoningEffortConfig::High => Some(OpenAiReasoningEffort::High),
+            ReasoningEffortConfig::None => None,
+        }
+    }
 }
 
 /// A summary of the reasoning performed by the model. This can be useful for
 /// debugging and understanding the model's reasoning process.
-#[derive(Debug, Serialize)]
+/// See https://platform.openai.com/docs/guides/reasoning?api-mode=responses#reasoning-summaries
+#[derive(Debug, Serialize, Default, Clone, Copy)]
 #[serde(rename_all = "lowercase")]
-pub(crate) enum Summary {
+pub(crate) enum OpenAiReasoningSummary {
+    #[default]
     Auto,
-    #[allow(dead_code)] // Will go away once this is configurable.
     Concise,
-    #[allow(dead_code)] // Will go away once this is configurable.
     Detailed,
 }
 
+impl From<ReasoningSummaryConfig> for Option<OpenAiReasoningSummary> {
+    fn from(summary: ReasoningSummaryConfig) -> Self {
+        match summary {
+            ReasoningSummaryConfig::Auto => Some(OpenAiReasoningSummary::Auto),
+            ReasoningSummaryConfig::Concise => Some(OpenAiReasoningSummary::Concise),
+            ReasoningSummaryConfig::Detailed => Some(OpenAiReasoningSummary::Detailed),
+            ReasoningSummaryConfig::None => None,
+        }
+    }
+}
+
+/// Request object that is serialized as JSON and POST'ed when using the
+/// Responses API.
 #[derive(Debug, Serialize)]
-pub(crate) struct Payload<'a> {
+pub(crate) struct ResponsesApiRequest<'a> {
     pub(crate) model: &'a str,
     pub(crate) instructions: &'a str,
     // TODO(mbolin): ResponseItem::Other should not be serialized. Currently,
@@ -88,6 +124,40 @@ pub(crate) struct Payload<'a> {
     pub(crate) stream: bool,
 }
 
+pub(crate) fn create_reasoning_param_for_request(
+    model: &str,
+    effort: ReasoningEffortConfig,
+    summary: ReasoningSummaryConfig,
+) -> Option<Reasoning> {
+    let effort: Option<OpenAiReasoningEffort> = effort.into();
+    let effort = effort?;
+
+    if model_supports_reasoning_summaries(model) {
+        Some(Reasoning {
+            effort,
+            summary: summary.into(),
+        })
+    } else {
+        None
+    }
+}
+
+pub fn model_supports_reasoning_summaries(model: &str) -> bool {
+    // Currently, we hardcode this rule to decide whether enable reasoning.
+    // We expect reasoning to apply only to OpenAI models, but we do not want
+    // users to have to mess with their config to disable reasoning for models
+    // that do not support it, such as `gpt-4.1`.
+    //
+    // Though if a user is using Codex with non-OpenAI models that, say, happen
+    // to start with "o", then they can set `model_reasoning_effort = "none` in
+    // config.toml to disable reasoning.
+    //
+    // Ultimately, this should also be configurable in config.toml, but we
+    // need to have defaults that "just work." Perhaps we could have a
+    // "reasoning models pattern" as part of ModelProviderInfo?
+    model.starts_with("o") || model.starts_with("codex")
+}
+
 pub(crate) struct ResponseStream {
     pub(crate) rx_event: mpsc::Receiver<Result<ResponseEvent>>,
 }

codex-rs/core/src/codex.rs

@@ -20,6 +20,7 @@ use codex_apply_patch::MaybeApplyPatchVerified;
 use codex_apply_patch::maybe_parse_apply_patch_verified;
 use codex_apply_patch::print_summary;
 use futures::prelude::*;
+use mcp_types::CallToolResult;
 use serde::Serialize;
 use serde_json;
 use tokio::sync::Notify;
@@ -107,6 +108,8 @@ impl Codex {
         let configure_session = Op::ConfigureSession {
             provider: config.model_provider.clone(),
             model: config.model.clone(),
+            model_reasoning_effort: config.model_reasoning_effort,
+            model_reasoning_summary: config.model_reasoning_summary,
             instructions,
             approval_policy: config.approval_policy,
             sandbox_policy: config.sandbox_policy.clone(),
@@ -295,6 +298,17 @@ impl Session {
         state.approved_commands.insert(cmd);
     }
 
+    /// Records items to both the rollout and the chat completions/ZDR
+    /// transcript, if enabled.
+    async fn record_conversation_items(&self, items: &[ResponseItem]) {
+        debug!("Recording items for conversation: {items:?}");
+        self.record_rollout_items(items).await;
+
+        if let Some(transcript) = self.state.lock().unwrap().zdr_transcript.as_mut() {
+            transcript.record_items(items);
+        }
+    }
+
     /// Append the given items to the session's rollout transcript (if enabled)
     /// and persist them to disk.
     async fn record_rollout_items(&self, items: &[ResponseItem]) {
@@ -388,7 +402,7 @@ impl Session {
         tool: &str,
         arguments: Option<serde_json::Value>,
         timeout: Option<Duration>,
-    ) -> anyhow::Result<mcp_types::CallToolResult> {
+    ) -> anyhow::Result<CallToolResult> {
         self.mcp_connection_manager
             .call_tool(server, tool, arguments, timeout)
             .await
@@ -542,6 +556,8 @@ async fn submission_loop(
             Op::ConfigureSession {
                 provider,
                 model,
+                model_reasoning_effort,
+                model_reasoning_summary,
                 instructions,
                 approval_policy,
                 sandbox_policy,
@@ -563,7 +579,12 @@ async fn submission_loop(
                     return;
                 }
 
-                let client = ModelClient::new(model.clone(), provider.clone());
+                let client = ModelClient::new(
+                    model.clone(),
+                    provider.clone(),
+                    model_reasoning_effort,
+                    model_reasoning_summary,
+                );
 
                 // abort any current running session and clone its state
                 let retain_zdr_transcript =
@@ -760,6 +781,19 @@ async fn submission_loop(
     debug!("Agent loop exited");
 }
 
+/// Takes a user message as input and runs a loop where, at each turn, the model
+/// replies with either:
+///
+/// - requested function calls
+/// - an assistant message
+///
+/// While it is possible for the model to return multiple of these items in a
+/// single turn, in practice, we generally one item per turn:
+///
+/// - If the model requests a function call, we execute it and send the output
+///   back to the model in the next turn.
+/// - If the model sends only an assistant message, we record it in the
+///   conversation history and consider the task complete.
 async fn run_task(sess: Arc<Session>, sub_id: String, input: Vec<InputItem>) {
     if input.is_empty() {
         return;
@@ -772,10 +806,14 @@ async fn run_task(sess: Arc<Session>, sub_id: String, input: Vec<InputItem>) {
         return;
     }
 
-    let mut pending_response_input: Vec<ResponseInputItem> = vec![ResponseInputItem::from(input)];
+    let initial_input_for_turn = ResponseInputItem::from(input);
+    sess.record_conversation_items(&[initial_input_for_turn.clone().into()])
+        .await;
+
+    let mut input_for_next_turn: Vec<ResponseInputItem> = vec![initial_input_for_turn];
     let last_agent_message: Option<String>;
     loop {
-        let mut net_new_turn_input = pending_response_input
+        let mut net_new_turn_input = input_for_next_turn
             .drain(..)
             .map(ResponseItem::from)
             .collect::<Vec<_>>();
@@ -783,11 +821,12 @@ async fn run_task(sess: Arc<Session>, sub_id: String, input: Vec<InputItem>) {
         // Note that pending_input would be something like a message the user
         // submitted through the UI while the model was running. Though the UI
         // may support this, the model might not.
-        let pending_input = sess.get_pending_input().into_iter().map(ResponseItem::from);
-        net_new_turn_input.extend(pending_input);
-
-        // Persist only the net-new items of this turn to the rollout.
-        sess.record_rollout_items(&net_new_turn_input).await;
+        let pending_input = sess
+            .get_pending_input()
+            .into_iter()
+            .map(ResponseItem::from)
+            .collect::<Vec<ResponseItem>>();
+        sess.record_conversation_items(&pending_input).await;
 
         // Construct the input that we will send to the model. When using the
         // Chat completions API (or ZDR clients), the model needs the full
@@ -796,20 +835,24 @@ async fn run_task(sess: Arc<Session>, sub_id: String, input: Vec<InputItem>) {
         // represents an append-only log without duplicates.
         let turn_input: Vec<ResponseItem> =
             if let Some(transcript) = sess.state.lock().unwrap().zdr_transcript.as_mut() {
-                // If we are using Chat/ZDR, we need to send the transcript with every turn.
-
-                // 1. Build up the conversation history for the next turn.
-                let full_transcript = [transcript.contents(), net_new_turn_input.clone()].concat();
-
-                // 2. Update the in-memory transcript so that future turns
-                // include these items as part of the history.
-                transcript.record_items(&net_new_turn_input);
-
-                // Note that `transcript.record_items()` does some filtering
-                // such that `full_transcript` may include items that were
-                // excluded from `transcript`.
-                full_transcript
+                // If we are using Chat/ZDR, we need to send the transcript with
+                // every turn. By induction, `transcript` already contains:
+                // - The `input` that kicked off this task.
+                // - Each `ResponseItem` that was recorded in the previous turn.
+                // - Each response to a `ResponseItem` (in practice, the only
+                //   response type we seem to have is `FunctionCallOutput`).
+                //
+                // The only thing the `transcript` does not contain is the
+                // `pending_input` that was injected while the model was
+                // running. We need to add that to the conversation history
+                // so that the model can see it in the next turn.
+                [transcript.contents(), pending_input].concat()
             } else {
+                // In practice, net_new_turn_input should contain only:
+                // - User messages
+                // - Outputs for function calls requested by the model
+                net_new_turn_input.extend(pending_input);
+
                 // Responses API path – we can just send the new items and
                 // record the same.
                 net_new_turn_input
@@ -830,29 +873,86 @@ async fn run_task(sess: Arc<Session>, sub_id: String, input: Vec<InputItem>) {
             .collect();
         match run_turn(&sess, sub_id.clone(), turn_input).await {
             Ok(turn_output) => {
-                let (items, responses): (Vec<_>, Vec<_>) = turn_output
-                    .into_iter()
-                    .map(|p| (p.item, p.response))
-                    .unzip();
-                let responses = responses
-                    .into_iter()
-                    .flatten()
-                    .collect::<Vec<ResponseInputItem>>();
+                let mut items_to_record_in_conversation_history = Vec::<ResponseItem>::new();
+                let mut responses = Vec::<ResponseInputItem>::new();
+                for processed_response_item in turn_output {
+                    let ProcessedResponseItem { item, response } = processed_response_item;
+                    match (&item, &response) {
+                        (ResponseItem::Message { role, .. }, None) if role == "assistant" => {
+                            // If the model returned a message, we need to record it.
+                            items_to_record_in_conversation_history.push(item);
+                        }
+                        (
+                            ResponseItem::LocalShellCall { .. },
+                            Some(ResponseInputItem::FunctionCallOutput { call_id, output }),
+                        ) => {
+                            items_to_record_in_conversation_history.push(item);
+                            items_to_record_in_conversation_history.push(
+                                ResponseItem::FunctionCallOutput {
+                                    call_id: call_id.clone(),
+                                    output: output.clone(),
+                                },
+                            );
+                        }
+                        (
+                            ResponseItem::FunctionCall { .. },
+                            Some(ResponseInputItem::FunctionCallOutput { call_id, output }),
+                        ) => {
+                            items_to_record_in_conversation_history.push(item);
+                            items_to_record_in_conversation_history.push(
+                                ResponseItem::FunctionCallOutput {
+                                    call_id: call_id.clone(),
+                                    output: output.clone(),
+                                },
+                            );
+                        }
+                        (
+                            ResponseItem::FunctionCall { .. },
+                            Some(ResponseInputItem::McpToolCallOutput { call_id, result }),
+                        ) => {
+                            items_to_record_in_conversation_history.push(item);
+                            let (content, success): (String, Option<bool>) = match result {
+                                Ok(CallToolResult { content, is_error }) => {
+                                    match serde_json::to_string(content) {
+                                        Ok(content) => (content, *is_error),
+                                        Err(e) => {
+                                            warn!("Failed to serialize MCP tool call output: {e}");
+                                            (e.to_string(), Some(true))
+                                        }
+                                    }
+                                }
+                                Err(e) => (e.clone(), Some(true)),
+                            };
+                            items_to_record_in_conversation_history.push(
+                                ResponseItem::FunctionCallOutput {
+                                    call_id: call_id.clone(),
+                                    output: FunctionCallOutputPayload { content, success },
+                                },
+                            );
+                        }
+                        (ResponseItem::Reasoning { .. }, None) => {
+                            // Omit from conversation history.
+                        }
+                        _ => {
+                            warn!("Unexpected response item: {item:?} with response: {response:?}");
+                        }
+                    };
+                    if let Some(response) = response {
+                        responses.push(response);
+                    }
+                }
 
                 // Only attempt to take the lock if there is something to record.
-                if !items.is_empty() {
-                    // First persist model-generated output to the rollout file – this only borrows.
-                    sess.record_rollout_items(&items).await;
-
-                    // For ZDR we also need to keep a transcript clone.
-                    if let Some(transcript) = sess.state.lock().unwrap().zdr_transcript.as_mut() {
-                        transcript.record_items(&items);
-                    }
+                if !items_to_record_in_conversation_history.is_empty() {
+                    sess.record_conversation_items(&items_to_record_in_conversation_history)
+                        .await;
                 }
 
                 if responses.is_empty() {
                     debug!("Turn completed");
-                    last_agent_message = get_last_assistant_message_from_turn(&items);
+                    last_agent_message = get_last_assistant_message_from_turn(
+                        &items_to_record_in_conversation_history,
+                    );
                     sess.maybe_notify(UserNotification::AgentTurnComplete {
                         turn_id: sub_id.clone(),
                         input_messages: turn_input_messages,
@@ -861,7 +961,7 @@ async fn run_task(sess: Arc<Session>, sub_id: String, input: Vec<InputItem>) {
                     break;
                 }
 
-                pending_response_input = responses;
+                input_for_next_turn = responses;
             }
             Err(e) => {
                 info!("Turn error: {e:#}");

codex-rs/core/src/config.rs

@@ -1,6 +1,8 @@
 use crate::config_profile::ConfigProfile;
 use crate::config_types::History;
 use crate::config_types::McpServerConfig;
+use crate::config_types::ReasoningEffort;
+use crate::config_types::ReasoningSummary;
 use crate::config_types::ShellEnvironmentPolicy;
 use crate::config_types::ShellEnvironmentPolicyToml;
 use crate::config_types::Tui;
@@ -112,6 +114,14 @@ pub struct Config {
     ///
     /// When this program is invoked, arg0 will be set to `codex-linux-sandbox`.
     pub codex_linux_sandbox_exe: Option<PathBuf>,
+
+    /// If not "none", the value to use for `reasoning.effort` when making a
+    /// request using the Responses API.
+    pub model_reasoning_effort: ReasoningEffort,
+
+    /// If not "none", the value to use for `reasoning.summary` when making a
+    /// request using the Responses API.
+    pub model_reasoning_summary: ReasoningSummary,
 }
 
 impl Config {
@@ -281,6 +291,9 @@ pub struct ConfigToml {
     /// When set to `true`, `AgentReasoning` events will be hidden from the
     /// UI/output. Defaults to `false`.
     pub hide_agent_reasoning: Option<bool>,
+
+    pub model_reasoning_effort: Option<ReasoningEffort>,
+    pub model_reasoning_summary: Option<ReasoningSummary>,
 }
 
 fn deserialize_sandbox_permissions<'de, D>(
@@ -444,6 +457,8 @@ impl Config {
             codex_linux_sandbox_exe,
 
             hide_agent_reasoning: cfg.hide_agent_reasoning.unwrap_or(false),
+            model_reasoning_effort: cfg.model_reasoning_effort.unwrap_or_default(),
+            model_reasoning_summary: cfg.model_reasoning_summary.unwrap_or_default(),
         };
         Ok(config)
     }
@@ -786,6 +801,8 @@ disable_response_storage = true
                 tui: Tui::default(),
                 codex_linux_sandbox_exe: None,
                 hide_agent_reasoning: false,
+                model_reasoning_effort: ReasoningEffort::default(),
+                model_reasoning_summary: ReasoningSummary::default(),
             },
             o3_profile_config
         );
@@ -826,6 +843,8 @@ disable_response_storage = true
             tui: Tui::default(),
             codex_linux_sandbox_exe: None,
             hide_agent_reasoning: false,
+            model_reasoning_effort: ReasoningEffort::default(),
+            model_reasoning_summary: ReasoningSummary::default(),
         };
 
         assert_eq!(expected_gpt3_profile_config, gpt3_profile_config);
@@ -881,6 +900,8 @@ disable_response_storage = true
             tui: Tui::default(),
             codex_linux_sandbox_exe: None,
             hide_agent_reasoning: false,
+            model_reasoning_effort: ReasoningEffort::default(),
+            model_reasoning_summary: ReasoningSummary::default(),
         };
 
         assert_eq!(expected_zdr_profile_config, zdr_profile_config);

codex-rs/core/src/config_types.rs

@@ -4,9 +4,11 @@
 // definitions that do not contain business logic.
 
 use std::collections::HashMap;
+use strum_macros::Display;
 use wildmatch::WildMatchPattern;
 
 use serde::Deserialize;
+use serde::Serialize;
 
 #[derive(Deserialize, Debug, Clone, PartialEq)]
 pub struct McpServerConfig {
@@ -175,3 +177,31 @@ impl From<ShellEnvironmentPolicyToml> for ShellEnvironmentPolicy {
         }
     }
 }
+
+/// See https://platform.openai.com/docs/guides/reasoning?api-mode=responses#get-started-with-reasoning
+#[derive(Debug, Serialize, Deserialize, Default, Clone, Copy, PartialEq, Eq, Display)]
+#[serde(rename_all = "lowercase")]
+#[strum(serialize_all = "lowercase")]
+pub enum ReasoningEffort {
+    Low,
+    #[default]
+    Medium,
+    High,
+    /// Option to disable reasoning.
+    None,
+}
+
+/// A summary of the reasoning performed by the model. This can be useful for
+/// debugging and understanding the model's reasoning process.
+/// See https://platform.openai.com/docs/guides/reasoning?api-mode=responses#reasoning-summaries
+#[derive(Debug, Serialize, Deserialize, Default, Clone, Copy, PartialEq, Eq, Display)]
+#[serde(rename_all = "lowercase")]
+#[strum(serialize_all = "lowercase")]
+pub enum ReasoningSummary {
+    #[default]
+    Auto,
+    Concise,
+    Detailed,
+    /// Option to disable reasoning summaries.
+    None,
+}

codex-rs/core/src/lib.rs

@@ -34,3 +34,5 @@ mod rollout;
 mod safety;
 mod user_notification;
 pub mod util;
+
+pub use client_common::model_supports_reasoning_summaries;

codex-rs/core/src/openai_tools.rs

@@ -93,7 +93,6 @@ pub(crate) fn create_tools_json_for_responses_api(
             .map(|(name, tool)| mcp_tool_to_openai_tool(name, tool)),
     );
 
-    tracing::debug!("tools_json: {}", serde_json::to_string_pretty(&tools_json)?);
     Ok(tools_json)
 }
 

codex-rs/core/src/protocol.rs

@@ -12,6 +12,8 @@ use serde::Deserialize;
 use serde::Serialize;
 use uuid::Uuid;
 
+use crate::config_types::ReasoningEffort as ReasoningEffortConfig;
+use crate::config_types::ReasoningSummary as ReasoningSummaryConfig;
 use crate::message_history::HistoryEntry;
 use crate::model_provider_info::ModelProviderInfo;
 
@@ -37,6 +39,10 @@ pub enum Op {
 
         /// If not specified, server will use its default model.
         model: String,
+
+        model_reasoning_effort: ReasoningEffortConfig,
+        model_reasoning_summary: ReasoningSummaryConfig,
+
         /// Model instructions
         instructions: Option<String>,
         /// When to escalate for approval for execution

codex-rs/exec/src/event_processor.rs

@@ -1,5 +1,7 @@
 use codex_common::elapsed::format_elapsed;
+use codex_core::WireApi;
 use codex_core::config::Config;
+use codex_core::model_supports_reasoning_summaries;
 use codex_core::protocol::AgentMessageEvent;
 use codex_core::protocol::BackgroundEventEvent;
 use codex_core::protocol::ErrorEvent;
@@ -127,16 +129,28 @@ impl EventProcessor {
             VERSION
         );
 
-        let entries = vec![
+        let mut entries = vec![
             ("workdir", config.cwd.display().to_string()),
             ("model", config.model.clone()),
             ("provider", config.model_provider_id.clone()),
             ("approval", format!("{:?}", config.approval_policy)),
             ("sandbox", format!("{:?}", config.sandbox_policy)),
         ];
+        if config.model_provider.wire_api == WireApi::Responses
+            && model_supports_reasoning_summaries(&config.model)
+        {
+            entries.push((
+                "reasoning effort",
+                config.model_reasoning_effort.to_string(),
+            ));
+            entries.push((
+                "reasoning summaries",
+                config.model_reasoning_summary.to_string(),
+            ));
+        }
 
         for (key, value) in entries {
-            println!("{} {}", format!("{key}: ").style(self.bold), value);
+            println!("{} {}", format!("{key}:").style(self.bold), value);
         }
 
         println!("--------");

codex-rs/execpolicy/Cargo.toml

@@ -24,7 +24,7 @@ env_logger = "0.11.5"
 log = "0.4"
 multimap = "0.10.0"
 path-absolutize = "3.1.1"
-regex = "1.11.1"
+regex-lite = "0.1"
 serde = { version = "1.0.194", features = ["derive"] }
 serde_json = "1.0.110"
 serde_with = { version = "3", features = ["macros"] }

codex-rs/execpolicy/src/policy.rs

@@ -1,6 +1,6 @@
 use multimap::MultiMap;
-use regex::Error as RegexError;
-use regex::Regex;
+use regex_lite::Error as RegexError;
+use regex_lite::Regex;
 
 use crate::ExecCall;
 use crate::Forbidden;
@@ -29,7 +29,7 @@ impl Policy {
         } else {
             let escaped_substrings = forbidden_substrings
                 .iter()
-                .map(|s| regex::escape(s))
+                .map(|s| regex_lite::escape(s))
                 .collect::<Vec<_>>()
                 .join("|");
             Some(Regex::new(&format!("({escaped_substrings})"))?)

codex-rs/execpolicy/src/policy_parser.rs

@@ -7,7 +7,7 @@ use crate::arg_matcher::ArgMatcher;
 use crate::opt::OptMeta;
 use log::info;
 use multimap::MultiMap;
-use regex::Regex;
+use regex_lite::Regex;
 use starlark::any::ProvidesStaticType;
 use starlark::environment::GlobalsBuilder;
 use starlark::environment::LibraryExtension;
@@ -73,7 +73,7 @@ impl PolicyParser {
 
 #[derive(Debug)]
 pub struct ForbiddenProgramRegex {
-    pub regex: regex::Regex,
+    pub regex: regex_lite::Regex,
     pub reason: String,
 }
 
@@ -93,7 +93,7 @@ impl PolicyBuilder {
         }
     }
 
-    fn build(self) -> Result<Policy, regex::Error> {
+    fn build(self) -> Result<Policy, regex_lite::Error> {
         let programs = self.programs.into_inner();
         let forbidden_program_regexes = self.forbidden_program_regexes.into_inner();
         let forbidden_substrings = self.forbidden_substrings.into_inner();
@@ -207,7 +207,7 @@ fn policy_builtins(builder: &mut GlobalsBuilder) {
             .unwrap()
             .downcast_ref::<PolicyBuilder>()
             .unwrap();
-        let compiled_regex = regex::Regex::new(&regex)?;
+        let compiled_regex = regex_lite::Regex::new(&regex)?;
         policy_builder.add_forbidden_program_regex(compiled_regex, reason);
         Ok(NoneType)
     }

codex-rs/tui/Cargo.toml

@@ -33,7 +33,7 @@ ratatui = { version = "0.29.0", features = [
     "unstable-rendered-line-info",
 ] }
 ratatui-image = "8.0.0"
-regex = "1"
+regex-lite = "0.1"
 serde_json = "1"
 shlex = "1.3.0"
 strum = "0.27.1"

codex-rs/tui/src/citation_regex.rs

@@ -1,6 +1,6 @@
 #![allow(clippy::expect_used)]
 
-use regex::Regex;
+use regex_lite::Regex;
 
 // This is defined in its own file so we can limit the scope of
 // `allow(clippy::expect_used)` because we cannot scope it to the `lazy_static!`

codex-rs/tui/src/history_cell.rs

@@ -5,7 +5,9 @@ use crate::text_block::TextBlock;
 use base64::Engine;
 use codex_ansi_escape::ansi_escape_line;
 use codex_common::elapsed::format_duration;
+use codex_core::WireApi;
 use codex_core::config::Config;
+use codex_core::model_supports_reasoning_summaries;
 use codex_core::protocol::FileChange;
 use codex_core::protocol::SessionConfiguredEvent;
 use image::DynamicImage;
@@ -147,13 +149,25 @@ impl HistoryCell {
                 ]),
             ];
 
-            let entries = vec![
+            let mut entries = vec![
                 ("workdir", config.cwd.display().to_string()),
                 ("model", config.model.clone()),
                 ("provider", config.model_provider_id.clone()),
                 ("approval", format!("{:?}", config.approval_policy)),
                 ("sandbox", format!("{:?}", config.sandbox_policy)),
             ];
+            if config.model_provider.wire_api == WireApi::Responses
+                && model_supports_reasoning_summaries(&config.model)
+            {
+                entries.push((
+                    "reasoning effort",
+                    config.model_reasoning_effort.to_string(),
+                ));
+                entries.push((
+                    "reasoning summaries",
+                    config.model_reasoning_summary.to_string(),
+                ));
+            }
             for (key, value) in entries {
                 lines.push(Line::from(vec![format!("{key}: ").bold(), value.into()]));
             }

codex-rs/tui/src/markdown.rs

@@ -71,7 +71,7 @@ fn rewrite_file_citations<'a>(
         None => return Cow::Borrowed(src),
     };
 
-    CITATION_REGEX.replace_all(src, |caps: &regex::Captures<'_>| {
+    CITATION_REGEX.replace_all(src, |caps: &regex_lite::Captures<'_>| {
         let file = &caps[1];
         let start_line = &caps[2];