400 error tool calling fix

Co-authored-by: Michael Bolin <mbolin@openai.com>

.vscode/launch.json

@@ -0,0 +1,18 @@
+{
+    "version": "0.2.0",
+    "configurations": [
+        {
+            "type": "lldb",
+            "request": "launch",
+            "name": "Cargo launch",
+            "cargo": {
+                "cwd": "${workspaceFolder}/codex-rs",
+                "args": [
+                    "build",
+                    "--bin=codex-tui"
+                ]
+            },
+            "args": []
+        }
+    ]
+}

.vscode/settings.json

@@ -0,0 +1,10 @@
+{
+    "rust-analyzer.checkOnSave": true,
+    "rust-analyzer.check.command": "clippy",
+    "rust-analyzer.check.extraArgs": ["--all-features", "--tests"],
+    "rust-analyzer.rustfmt.extraArgs": ["--config", "imports_granularity=Item"],
+    "[rust]": {
+        "editor.defaultFormatter": "rust-lang.rust-analyzer",
+        "editor.formatOnSave": true,
+    }
+}

AGENTS.md

@@ -3,3 +3,7 @@
 In the codex-rs folder where the rust code lives:
 
 - Never add or modify any code related to `CODEX_SANDBOX_NETWORK_DISABLED_ENV_VAR`. You operate in a sandbox where `CODEX_SANDBOX_NETWORK_DISABLED=1` will be set whenever you use the `shell` tool. Any existing code that uses `CODEX_SANDBOX_NETWORK_DISABLED_ENV_VAR` was authored with this fact in mind. It is often used to early exit out of tests that the author knew you would not be able to run given your sandbox limitations.
+
+Before creating a pull request with changes to `codex-rs`, run `just fmt` (in `codex-rs` directory) to format the code and `just fix` (in `codex-rs` directory) to fix any linter issues in the code, ensure the test suite passes by running `cargo test --all-features` in the `codex-rs` directory.
+
+When making individual changes prefer running tests on individual files or projects first.

codex-cli/bin/codex.js

@@ -15,7 +15,6 @@
  *      current platform / architecture, an error is thrown.
  */
 
-import { spawnSync } from "child_process";
 import fs from "fs";
 import path from "path";
 import { fileURLToPath, pathToFileURL } from "url";
@@ -35,7 +34,7 @@ const wantsNative = fs.existsSync(path.join(__dirname, "use-native")) ||
     : false);
 
 // Try native binary if requested.
-if (wantsNative) {
+if (wantsNative && process.platform !== 'win32') {
   const { platform, arch } = process;
 
   let targetTriple = null;
@@ -74,22 +73,76 @@ if (wantsNative) {
   }
 
   const binaryPath = path.join(__dirname, "..", "bin", `codex-${targetTriple}`);
-  const result = spawnSync(binaryPath, process.argv.slice(2), {
+
+  // Use an asynchronous spawn instead of spawnSync so that Node is able to
+  // respond to signals (e.g. Ctrl-C / SIGINT) while the native binary is
+  // executing. This allows us to forward those signals to the child process
+  // and guarantees that when either the child terminates or the parent
+  // receives a fatal signal, both processes exit in a predictable manner.
+  const { spawn } = await import("child_process");
+
+  const child = spawn(binaryPath, process.argv.slice(2), {
     stdio: "inherit",
   });
 
-  const exitCode = typeof result.status === "number" ? result.status : 1;
-  process.exit(exitCode);
-}
+  child.on("error", (err) => {
+    // Typically triggered when the binary is missing or not executable.
+    // Re-throwing here will terminate the parent with a non-zero exit code
+    // while still printing a helpful stack trace.
+    // eslint-disable-next-line no-console
+    console.error(err);
+    process.exit(1);
+  });
 
-// Fallback: execute the original JavaScript CLI.
+  // Forward common termination signals to the child so that it shuts down
+  // gracefully. In the handler we temporarily disable the default behavior of
+  // exiting immediately; once the child has been signaled we simply wait for
+  // its exit event which will in turn terminate the parent (see below).
+  const forwardSignal = (signal) => {
+    if (child.killed) {
+      return;
+    }
+    try {
+      child.kill(signal);
+    } catch {
+      /* ignore */
+    }
+  };
 
-// Resolve the path to the compiled CLI bundle
-const cliPath = path.resolve(__dirname, "../dist/cli.js");
-const cliUrl = pathToFileURL(cliPath).href;
+  ["SIGINT", "SIGTERM", "SIGHUP"].forEach((sig) => {
+    process.on(sig, () => forwardSignal(sig));
+  });
 
-// Load and execute the CLI
-(async () => {
+  // When the child exits, mirror its termination reason in the parent so that
+  // shell scripts and other tooling observe the correct exit status.
+  // Wrap the lifetime of the child process in a Promise so that we can await
+  // its termination in a structured way. The Promise resolves with an object
+  // describing how the child exited: either via exit code or due to a signal.
+  const childResult = await new Promise((resolve) => {
+    child.on("exit", (code, signal) => {
+      if (signal) {
+        resolve({ type: "signal", signal });
+      } else {
+        resolve({ type: "code", exitCode: code ?? 1 });
+      }
+    });
+  });
+
+  if (childResult.type === "signal") {
+    // Re-emit the same signal so that the parent terminates with the expected
+    // semantics (this also sets the correct exit code of 128 + n).
+    process.kill(process.pid, childResult.signal);
+  } else {
+    process.exit(childResult.exitCode);
+  }
+} else {
+  // Fallback: execute the original JavaScript CLI.
+
+  // Resolve the path to the compiled CLI bundle
+  const cliPath = path.resolve(__dirname, "../dist/cli.js");
+  const cliUrl = pathToFileURL(cliPath).href;
+
+  // Load and execute the CLI
   try {
     await import(cliUrl);
   } catch (err) {
@@ -97,4 +150,4 @@ const cliUrl = pathToFileURL(cliPath).href;
     console.error(err);
     process.exit(1);
   }
-})();
+}

codex-cli/src/components/chat/terminal-chat-input.tsx

@@ -10,25 +10,19 @@ import type {
 import MultilineTextEditor from "./multiline-editor";
 import { TerminalChatCommandReview } from "./terminal-chat-command-review.js";
 import TextCompletions from "./terminal-chat-completions.js";
-import { loadConfig, type AppConfig } from "../../utils/config.js";
+import { loadConfig } from "../../utils/config.js";
 import { getFileSystemSuggestions } from "../../utils/file-system-suggestions.js";
 import { expandFileTags } from "../../utils/file-tag-utils";
 import { createInputItem } from "../../utils/input-utils.js";
 import { log } from "../../utils/logger/log.js";
 import { setSessionId } from "../../utils/session.js";
 import { SLASH_COMMANDS, type SlashCommand } from "../../utils/slash-commands";
-import {
-  runSecurityReview,
-  SecurityReviewError,
-} from "../../utils/security-review.js";
-import type { SecurityReviewMode } from "../../utils/security-review.js";
 import {
   loadCommandHistory,
   addToHistory,
 } from "../../utils/storage/command-history.js";
 import { clearTerminal, onExit } from "../../utils/terminal.js";
 import { Box, Text, useApp, useInput, useStdin } from "ink";
-import path from "node:path";
 import { fileURLToPath } from "node:url";
 import React, {
   useCallback,
@@ -45,130 +39,6 @@ const suggestions = [
   "are there any bugs in my code?",
 ];
 
-const SEC_REVIEW_COMMAND = "/secreview";
-
-type SecReviewCommandOptions = {
-  mode: SecurityReviewMode;
-  includePaths: Array<string>;
-  outputPath?: string;
-  repoPath?: string;
-  modelName?: string;
-};
-
-function tokenizeCommand(input: string): Array<string> {
-  const tokens: Array<string> = [];
-  const regex = /"([^"]*)"|'([^']*)'|(\S+)/g;
-  let match: RegExpExecArray | null;
-  while ((match = regex.exec(input)) !== null) {
-    if (match[1] != null) {
-      tokens.push(match[1]);
-    } else if (match[2] != null) {
-      tokens.push(match[2]);
-    } else if (match[3] != null) {
-      tokens.push(match[3]);
-    }
-  }
-  return tokens;
-}
-
-function parseSecReviewCommand(input: string): SecReviewCommandOptions {
-  const tokens = tokenizeCommand(input).slice(1); // drop the command itself
-  let mode: SecurityReviewMode = "full";
-  const includePaths: Array<string> = [];
-  let outputPath: string | undefined;
-  let repoPath: string | undefined;
-  let modelName: string | undefined;
-
-  const parseMode = (value: string, option: string) => {
-    if (value === "bugs") {
-      mode = "bugs";
-    } else if (value === "full") {
-      mode = "full";
-    } else {
-      throw new Error(`Unknown ${option} value "${value}". Use "full" or "bugs".`);
-    }
-  };
-
-  for (let i = 0; i < tokens.length; i += 1) {
-    const token = tokens[i];
-
-    const expectValue = (label: string): string => {
-      if (i + 1 >= tokens.length) {
-        throw new Error(`Expected value after ${label}`);
-      }
-      i += 1;
-      return tokens[i];
-    };
-
-    if (token === "--") {
-      break;
-    } else if (token === "bugs" || token === "--bugs" || token === "--mode=bugs") {
-      mode = "bugs";
-    } else if (token === "full" || token === "--full" || token === "--mode=full") {
-      mode = "full";
-    } else if (token === "--mode") {
-      parseMode(expectValue("--mode"), "--mode");
-    } else if (token.startsWith("--mode=")) {
-      parseMode(token.slice("--mode=".length), "--mode");
-    } else if (token === "--path" || token === "-p") {
-      includePaths.push(expectValue(token));
-    } else if (token.startsWith("--path=")) {
-      includePaths.push(token.slice("--path=".length));
-    } else if (token.startsWith("-p=")) {
-      includePaths.push(token.slice("-p=".length));
-    } else if (
-      token === "--output" ||
-      token === "-o" ||
-      token === "--output-location"
-    ) {
-      outputPath = expectValue(token);
-    } else if (token.startsWith("--output=")) {
-      outputPath = token.slice("--output=".length);
-    } else if (token.startsWith("-o=")) {
-      outputPath = token.slice("-o=".length);
-    } else if (
-      token === "--repo" ||
-      token === "--repo-location" ||
-      token === "--repository"
-    ) {
-      repoPath = expectValue(token);
-    } else if (token.startsWith("--repo=")) {
-      repoPath = token.slice("--repo=".length);
-    } else if (token.startsWith("--repo-location=")) {
-      repoPath = token.slice("--repo-location=".length);
-    } else if (token === "--model" || token === "--model-name") {
-      modelName = expectValue(token);
-    } else if (token.startsWith("--model=")) {
-      modelName = token.slice("--model=".length);
-    } else if (token.startsWith("--model-name=")) {
-      modelName = token.slice("--model-name=".length);
-    } else if (token.length > 0) {
-      includePaths.push(token);
-    }
-  }
-
-  return {
-    mode,
-    includePaths: includePaths.filter((p) => p.length > 0),
-    outputPath,
-    repoPath,
-    modelName,
-  };
-}
-
-function trimLogOutput(logText: string, maxLines: number = 40): string {
-  const normalised = logText.replace(/\r\n/g, "\n").trimEnd();
-  if (normalised === "") {
-    return "(empty)";
-  }
-  const lines = normalised.split("\n");
-  if (lines.length <= maxLines) {
-    return normalised;
-  }
-  const tail = lines.slice(-maxLines);
-  return ["… (showing last " + maxLines + " lines)", ...tail].join("\n");
-}
-
 export default function TerminalChatInput({
   isNew,
   loading,
@@ -190,7 +60,6 @@ export default function TerminalChatInput({
   active,
   thinkingSeconds,
   items = [],
-  config,
 }: {
   isNew: boolean;
   loading: boolean;
@@ -216,7 +85,6 @@ export default function TerminalChatInput({
   thinkingSeconds: number;
   // New: current conversation items so we can include them in bug reports
   items?: Array<ResponseItem>;
-  config: AppConfig;
 }): React.ReactElement {
   // Slash command suggestion index
   const [selectedSlashSuggestion, setSelectedSlashSuggestion] =
@@ -644,230 +512,6 @@ export default function TerminalChatInput({
       } else if (inputValue.startsWith("/approval")) {
         setInput("");
         openApprovalOverlay();
-        return;
-      } else if (inputValue.startsWith(SEC_REVIEW_COMMAND)) {
-        setInput("");
-        const commandId = `secreview-${Date.now()}`;
-
-        let parsed: SecReviewCommandOptions;
-        try {
-          parsed = parseSecReviewCommand(inputValue);
-        } catch (error) {
-          const message =
-            error instanceof Error ? error.message : String(error);
-          setItems((prev) => [
-            ...prev,
-            {
-              id: `${commandId}-parse-error`,
-              type: "message",
-              role: "system",
-              content: [
-                {
-                  type: "input_text",
-                  text: `⚠️ Unable to parse ${SEC_REVIEW_COMMAND} arguments: ${message}`,
-                },
-              ],
-            },
-          ]);
-          return;
-        }
-
-        const repoPath = parsed.repoPath
-          ? path.isAbsolute(parsed.repoPath)
-            ? parsed.repoPath
-            : path.resolve(process.cwd(), parsed.repoPath)
-          : process.cwd();
-
-        const resolvedOutputPath =
-          parsed.outputPath != null
-            ? path.isAbsolute(parsed.outputPath)
-              ? parsed.outputPath
-              : path.resolve(repoPath, parsed.outputPath)
-            : undefined;
-
-        const scopeDescription =
-          parsed.includePaths.length > 0
-            ? parsed.includePaths.join(", ")
-            : "entire repository";
-
-        const introLines = [
-          `🔐 Running AppSec security review (mode: ${parsed.mode}).`,
-          `Repository: ${repoPath}`,
-          `Scope: ${scopeDescription}`,
-        ];
-
-        if (resolvedOutputPath) {
-          introLines.push(`Output: ${resolvedOutputPath}`);
-        }
-
-        if (parsed.modelName) {
-          introLines.push(`Model override: ${parsed.modelName}`);
-        }
-
-        setItems((prev) => [
-          ...prev,
-          {
-            id: `${commandId}-start`,
-            type: "message",
-            role: "system",
-            content: [
-              {
-                type: "input_text",
-                text: introLines.join("\n"),
-              },
-            ],
-          },
-        ]);
-
-        try {
-          const result = await runSecurityReview({
-            repoPath,
-            includePaths: parsed.includePaths,
-            outputPath: resolvedOutputPath,
-            modelName: parsed.modelName,
-            mode: parsed.mode,
-            config,
-          });
-
-          const summaryLines = [
-            "✅ AppSec review complete.",
-            `Artifacts: ${result.outputRoot}`,
-          ];
-          if (!result.reportContent) {
-            summaryLines.push("ℹ️ report.md not found in output.");
-          }
-          if (!result.bugsContent) {
-            summaryLines.push("ℹ️ context/bugs.md not found in output.");
-          }
-
-          setItems((prev) => [
-            ...prev,
-            {
-              id: `${commandId}-complete`,
-              type: "message",
-              role: "system",
-              content: [
-                {
-                  type: "input_text",
-                  text: summaryLines.join("\n"),
-                },
-              ],
-            },
-          ]);
-
-          if (parsed.mode === "full" && result.reportContent) {
-            setItems((prev) => [
-              ...prev,
-              {
-                id: `${commandId}-report`,
-                type: "message",
-                role: "assistant",
-                content: [
-                  {
-                    type: "output_text",
-                    text: `# AppSec Security Review Report\n\n${result.reportContent}`,
-                  },
-                ],
-              },
-            ]);
-          }
-
-          if (result.bugsContent) {
-            const heading =
-              parsed.mode === "full"
-                ? "## Bugs Summary"
-                : "# AppSec Bugs Summary";
-            setItems((prev) => [
-              ...prev,
-              {
-                id: `${commandId}-bugs`,
-                type: "message",
-                role: "assistant",
-                content: [
-                  {
-                    type: "output_text",
-                    text: `${heading}\n\n${result.bugsContent}`,
-                  },
-                ],
-              },
-            ]);
-          } else {
-            setItems((prev) => [
-              ...prev,
-              {
-                id: `${commandId}-no-bugs`,
-                type: "message",
-                role: "system",
-                content: [
-                  {
-                    type: "input_text",
-                    text:
-                      "No bug summary produced. Check the output directory for details.",
-                  },
-                ],
-              },
-            ]);
-          }
-
-          if (parsed.mode === "bugs" && result.reportContent) {
-            setItems((prev) => [
-              ...prev,
-              {
-                id: `${commandId}-report-location`,
-                type: "message",
-                role: "system",
-                content: [
-                  {
-                    type: "input_text",
-                    text: `Full report available at ${result.reportPath}`,
-                  },
-                ],
-              },
-            ]);
-          }
-          if (result.stdout.trim()) {
-            setItems((prev) => [
-              ...prev,
-              {
-                id: `${commandId}-logs`,
-                type: "message",
-                role: "system",
-                content: [
-                  {
-                    type: "input_text",
-                    text: `Logs:\n${trimLogOutput(result.stdout)}`,
-                  },
-                ],
-              },
-            ]);
-          }
-        } catch (error) {
-          const message =
-            error instanceof Error ? error.message : String(error);
-          const stderr =
-            error instanceof SecurityReviewError && error.stderr
-              ? `\n\nstderr last lines:\n${trimLogOutput(error.stderr)}`
-              : "";
-          const stdout =
-            error instanceof SecurityReviewError && error.stdout
-              ? `\n\nstdout last lines:\n${trimLogOutput(error.stdout)}`
-              : "";
-          setItems((prev) => [
-            ...prev,
-            {
-              id: `${commandId}-error`,
-              type: "message",
-              role: "system",
-              content: [
-                {
-                  type: "input_text",
-                  text: `❌ AppSec review failed: ${message}${stderr}${stdout}`,
-                },
-              ],
-            },
-          ]);
-        }
-
         return;
       } else if (["exit", "q", ":q"].includes(inputValue)) {
         setInput("");
@@ -1063,13 +707,13 @@ export default function TerminalChatInput({
       submitInput([inputItem]);
 
       // Get config for history persistence.
-      const historyConfig = loadConfig();
+      const config = loadConfig();
 
       // Add to history and update state.
       const updatedHistory = await addToHistory(value, history, {
-        maxSize: historyConfig.history?.maxSize ?? 1000,
-        saveHistory: historyConfig.history?.saveHistory ?? true,
-        sensitivePatterns: historyConfig.history?.sensitivePatterns ?? [],
+        maxSize: config.history?.maxSize ?? 1000,
+        saveHistory: config.history?.saveHistory ?? true,
+        sensitivePatterns: config.history?.sensitivePatterns ?? [],
       });
 
       setHistory(updatedHistory);
@@ -1098,7 +742,6 @@ export default function TerminalChatInput({
       onCompact,
       skipNextSubmit,
       items,
-      config,
     ],
   );
 

codex-cli/src/components/chat/terminal-chat.tsx

@@ -580,7 +580,6 @@ export default function TerminalChat({
             }}
             items={items}
             thinkingSeconds={thinkingSeconds}
-            config={config}
           />
         )}
         {overlayMode === "history" && (

codex-cli/src/components/help-overlay.tsx

@@ -56,10 +56,6 @@ export default function HelpOverlay({
           <Text color="cyan">/bug</Text> – generate a prefilled GitHub issue URL
           with session log
         </Text>
-        <Text>
-          <Text color="cyan">/secreview</Text> – run AppSec security review and
-          show the results
-        </Text>
         <Text>
           <Text color="cyan">/diff</Text> – view working tree git diff
         </Text>

codex-cli/src/utils/slash-commands.ts

@@ -24,10 +24,6 @@ export const SLASH_COMMANDS: Array<SlashCommand> = [
   { command: "/help", description: "Show list of commands" },
   { command: "/model", description: "Open model selection panel" },
   { command: "/approval", description: "Open approval mode selection panel" },
-  {
-    command: "/secreview",
-    description: "Run AppSec security review and display the generated reports",
-  },
   {
     command: "/bug",
     description: "Generate a prefilled GitHub issue URL with session log",

codex-cli/tests/clear-command.test.tsx

@@ -61,11 +61,6 @@ describe("/clear command", () => {
       active: true,
       thinkingSeconds: 0,
       items: existingItems,
-      config: {
-        model: "codex-mini-latest",
-        instructions: "",
-        provider: "openai",
-      },
     };
 
     const { stdin, flush, cleanup } = renderTui(

codex-cli/tests/multiline-history-behavior.test.tsx

@@ -66,27 +66,17 @@ function stubProps(): any {
     loading: false,
     submitInput: vi.fn(),
     confirmationPrompt: null,
-    explanation: undefined,
     submitConfirmation: vi.fn(),
     setLastResponseId: vi.fn(),
-    setItems: vi.fn(),
+    // Cast to any to satisfy the generic React.Dispatch signature without
+    // pulling the ResponseItem type into the test bundle.
+    setItems: (() => {}) as any,
     contextLeftPercent: 100,
     openOverlay: vi.fn(),
     openModelOverlay: vi.fn(),
     openHelpOverlay: vi.fn(),
-    openApprovalOverlay: vi.fn(),
-    openSessionsOverlay: vi.fn(),
-    openDiffOverlay: vi.fn(),
-    onCompact: vi.fn(),
     interruptAgent: vi.fn(),
     active: true,
-    thinkingSeconds: 0,
-    items: [],
-    config: {
-      model: "codex-mini-latest",
-      instructions: "",
-      provider: "openai",
-    },
   };
 }
 

codex-cli/tests/slash-commands.test.ts

@@ -10,7 +10,6 @@ test("SLASH_COMMANDS includes expected commands", () => {
   expect(commands).toContain("/help");
   expect(commands).toContain("/model");
   expect(commands).toContain("/approval");
-  expect(commands).toContain("/secreview");
   expect(commands).toContain("/clearhistory");
   expect(commands).toContain("/diff");
 });

codex-cli/tests/terminal-chat-input-compact.test.tsx

@@ -26,11 +26,6 @@ describe("TerminalChatInput compact command", () => {
       interruptAgent: () => {},
       active: true,
       thinkingSeconds: 0,
-      config: {
-        model: "codex-mini-latest",
-        instructions: "",
-        provider: "openai",
-      },
     };
     const { lastFrameStripped } = renderTui(<TerminalChatInput {...props} />);
     const frame = lastFrameStripped();

codex-cli/tests/terminal-chat-input-file-tag-suggestions.test.tsx

@@ -81,11 +81,6 @@ describe("TerminalChatInput file tag suggestions", () => {
     interruptAgent: vi.fn(),
     active: true,
     thinkingSeconds: 0,
-    config: {
-      model: "codex-mini-latest",
-      instructions: "",
-      provider: "openai",
-    },
   };
 
   beforeEach(() => {

codex-cli/tests/terminal-chat-input-multiline.test.tsx

@@ -47,11 +47,6 @@ describe("TerminalChatInput multiline functionality", () => {
       interruptAgent: () => {},
       active: true,
       thinkingSeconds: 0,
-      config: {
-        model: "codex-mini-latest",
-        instructions: "",
-        provider: "openai",
-      },
     };
 
     const { stdin, lastFrameStripped, flush, cleanup } = renderTui(
@@ -104,11 +99,6 @@ describe("TerminalChatInput multiline functionality", () => {
       interruptAgent: () => {},
       active: true,
       thinkingSeconds: 0,
-      config: {
-        model: "codex-mini-latest",
-        instructions: "",
-        provider: "openai",
-      },
     };
 
     const { stdin, lastFrameStripped, flush, cleanup } = renderTui(

codex-rs/Cargo.lock

@@ -399,6 +399,15 @@ version = "2.6.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "6099cdc01846bc367c4e7dd630dc5966dccf36b652fae7a74e17b640411a91b2"
 
+[[package]]
+name = "block-buffer"
+version = "0.10.4"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "3078c7629b62d3f0439517fa394996acacc5cbc91c5a20d8c658e77abd503a71"
+dependencies = [
+ "generic-array",
+]
+
 [[package]]
 name = "bstr"
 version = "1.12.0"
@@ -671,6 +680,7 @@ dependencies = [
  "seccompiler",
  "serde",
  "serde_json",
+ "sha1",
  "strum_macros 0.27.1",
  "tempfile",
  "thiserror 2.0.12",
@@ -683,6 +693,7 @@ dependencies = [
  "tree-sitter",
  "tree-sitter-bash",
  "uuid",
+ "walkdir",
  "wildmatch",
  "wiremock",
 ]
@@ -818,12 +829,10 @@ dependencies = [
  "ratatui",
  "ratatui-image",
  "regex-lite",
- "reqwest",
  "serde_json",
  "shlex",
  "strum 0.27.1",
  "strum_macros 0.27.1",
- "time",
  "tokio",
  "tracing",
  "tracing-appender",
@@ -934,6 +943,15 @@ version = "0.8.7"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "773648b94d0e5d620f64f280777445740e61fe701025087ec8b57f45c791888b"
 
+[[package]]
+name = "cpufeatures"
+version = "0.2.17"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "59ed5838eebb26a2bb2e58f6d5b5316989ae9d08bab10e0e6d103e656d1b0280"
+dependencies = [
+ "libc",
+]
+
 [[package]]
 name = "crc32fast"
 version = "1.4.2"
@@ -1008,6 +1026,16 @@ version = "0.2.4"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "460fbee9c2c2f33933d720630a6a0bac33ba7053db5344fac858d4b8952d77d5"
 
+[[package]]
+name = "crypto-common"
+version = "0.1.6"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "1bfb12502f3fc46cca1bb51ac28df9d618d813cdc3d2f25b9fe775a34af26bb3"
+dependencies = [
+ "generic-array",
+ "typenum",
+]
+
 [[package]]
 name = "ctor"
 version = "0.1.26"
@@ -1158,6 +1186,16 @@ version = "0.4.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "6184e33543162437515c2e2b48714794e37845ec9851711914eec9d308f6ebe8"
 
+[[package]]
+name = "digest"
+version = "0.10.7"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "9ed9a281f7bc9b7576e61468ba615a66a5c8cfdff42420a70aa82701a3b1e292"
+dependencies = [
+ "block-buffer",
+ "crypto-common",
+]
+
 [[package]]
 name = "dirs"
 version = "6.0.0"
@@ -1647,6 +1685,16 @@ dependencies = [
  "byteorder",
 ]
 
+[[package]]
+name = "generic-array"
+version = "0.14.7"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "85649ca51fd72272d7821adaf274ad91c288277713d9c18820d8499a7ff69e9a"
+dependencies = [
+ "typenum",
+ "version_check",
+]
+
 [[package]]
 name = "getopts"
 version = "0.2.23"
@@ -3946,6 +3994,17 @@ dependencies = [
  "syn 2.0.104",
 ]
 
+[[package]]
+name = "sha1"
+version = "0.10.6"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "e3bf829a2d51ab4a5ddf1352d8470c140cadc8301b2ae1789db023f01cedd6ba"
+dependencies = [
+ "cfg-if",
+ "cpufeatures",
+ "digest",
+]
+
 [[package]]
 name = "sharded-slab"
 version = "0.1.7"
@@ -4853,6 +4912,12 @@ dependencies = [
  "unicode-width 0.2.0",
 ]
 
+[[package]]
+name = "typenum"
+version = "1.18.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "1dccffe3ce07af9386bfd29e80c0ab1a8205a2fc34e4bcd40364df902cfa8f3f"
+
 [[package]]
 name = "unicase"
 version = "2.8.1"

codex-rs/config.md

@@ -92,6 +92,32 @@ http_headers = { "X-Example-Header" = "example-value" }
 env_http_headers = { "X-Example-Features": "EXAMPLE_FEATURES" }
 ```
 
+### Per-provider network tuning
+
+The following optional settings control retry behaviour and streaming idle timeouts **per model provider**. They must be specified inside the corresponding `[model_providers.<id>]` block in `config.toml`. (Older releases accepted top‑level keys; those are now ignored.)
+
+Example:
+
+```toml
+[model_providers.openai]
+name = "OpenAI"
+base_url = "https://api.openai.com/v1"
+env_key = "OPENAI_API_KEY"
+# network tuning overrides (all optional; falls back to built‑in defaults)
+request_max_retries = 4            # retry failed HTTP requests
+stream_max_retries = 10            # retry dropped SSE streams
+stream_idle_timeout_ms = 300000    # 5m idle timeout
+```
+
+#### request_max_retries
+How many times Codex will retry a failed HTTP request to the model provider. Defaults to `4`.
+
+#### stream_max_retries
+Number of times Codex will attempt to reconnect when a streaming response is interrupted. Defaults to `10`.
+
+#### stream_idle_timeout_ms
+How long Codex will wait for activity on a streaming response before treating the connection as lost. Defaults to `300_000` (5 minutes).
+
 ## model_provider
 
 Identifies which provider to use from the `model_providers` map. Defaults to `"openai"`. You can override the `base_url` for the built-in `openai` provider via the `OPENAI_BASE_URL` environment variable.
@@ -444,7 +470,7 @@ Currently, `"vscode"` is the default, though Codex does not verify VS Code is in
 
 ## hide_agent_reasoning
 
-Codex intermittently emits "reasoning" events that show the model’s internal "thinking" before it produces a final answer. Some users may find these events distracting, especially in CI logs or minimal terminal output.
+Codex intermittently emits "reasoning" events that show the model's internal "thinking" before it produces a final answer. Some users may find these events distracting, especially in CI logs or minimal terminal output.
 
 Setting `hide_agent_reasoning` to `true` suppresses these events in **both** the TUI as well as the headless `exec` sub-command:
 

codex-rs/core/Cargo.toml

@@ -28,6 +28,7 @@ rand = "0.9"
 reqwest = { version = "0.12", features = ["json", "stream"] }
 serde = { version = "1", features = ["derive"] }
 serde_json = "1"
+sha1 = "0.10.6"
 strum_macros = "0.27.1"
 thiserror = "2.0.12"
 time = { version = "0.3", features = ["formatting", "local-offset", "macros"] }
@@ -65,4 +66,5 @@ predicates = "3"
 pretty_assertions = "1.4.1"
 tempfile = "3"
 tokio-test = "0.4"
+walkdir = "2.5.0"
 wiremock = "0.6"

codex-rs/core/src/chat_completions.rs

@@ -21,8 +21,6 @@ use crate::client_common::ResponseEvent;
 use crate::client_common::ResponseStream;
 use crate::error::CodexErr;
 use crate::error::Result;
-use crate::flags::OPENAI_REQUEST_MAX_RETRIES;
-use crate::flags::OPENAI_STREAM_IDLE_TIMEOUT_MS;
 use crate::models::ContentItem;
 use crate::models::ResponseItem;
 use crate::openai_tools::create_tools_json_for_chat_completions_api;
@@ -121,6 +119,7 @@ pub(crate) async fn stream_chat_completions(
     );
 
     let mut attempt = 0;
+    let max_retries = provider.request_max_retries();
     loop {
         attempt += 1;
 
@@ -134,9 +133,13 @@ pub(crate) async fn stream_chat_completions(
 
         match res {
             Ok(resp) if resp.status().is_success() => {
-                let (tx_event, rx_event) = mpsc::channel::<Result<ResponseEvent>>(16);
+                let (tx_event, rx_event) = mpsc::channel::<Result<ResponseEvent>>(1600);
                 let stream = resp.bytes_stream().map_err(CodexErr::Reqwest);
-                tokio::spawn(process_chat_sse(stream, tx_event));
+                tokio::spawn(process_chat_sse(
+                    stream,
+                    tx_event,
+                    provider.stream_idle_timeout(),
+                ));
                 return Ok(ResponseStream { rx_event });
             }
             Ok(res) => {
@@ -146,7 +149,7 @@ pub(crate) async fn stream_chat_completions(
                     return Err(CodexErr::UnexpectedStatus(status, body));
                 }
 
-                if attempt > *OPENAI_REQUEST_MAX_RETRIES {
+                if attempt > max_retries {
                     return Err(CodexErr::RetryLimit(status));
                 }
 
@@ -162,7 +165,7 @@ pub(crate) async fn stream_chat_completions(
                 tokio::time::sleep(delay).await;
             }
             Err(e) => {
-                if attempt > *OPENAI_REQUEST_MAX_RETRIES {
+                if attempt > max_retries {
                     return Err(e.into());
                 }
                 let delay = backoff(attempt);
@@ -175,14 +178,15 @@ pub(crate) async fn stream_chat_completions(
 /// Lightweight SSE processor for the Chat Completions streaming format. The
 /// output is mapped onto Codex's internal [`ResponseEvent`] so that the rest
 /// of the pipeline can stay agnostic of the underlying wire format.
-async fn process_chat_sse<S>(stream: S, tx_event: mpsc::Sender<Result<ResponseEvent>>)
-where
+async fn process_chat_sse<S>(
+    stream: S,
+    tx_event: mpsc::Sender<Result<ResponseEvent>>,
+    idle_timeout: Duration,
+) where
     S: Stream<Item = Result<Bytes>> + Unpin,
 {
     let mut stream = stream.eventsource();
 
-    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
@@ -426,6 +430,12 @@ where
                     // will never appear in a Chat Completions stream.
                     continue;
                 }
+                Poll::Ready(Some(Ok(ResponseEvent::OutputTextDelta(_))))
+                | Poll::Ready(Some(Ok(ResponseEvent::ReasoningSummaryDelta(_)))) => {
+                    // Deltas are ignored here since aggregation waits for the
+                    // final OutputItemDone.
+                    continue;
+                }
             }
         }
     }

codex-rs/core/src/client.rs

@@ -29,8 +29,6 @@ use crate::config_types::ReasoningSummary as ReasoningSummaryConfig;
 use crate::error::CodexErr;
 use crate::error::Result;
 use crate::flags::CODEX_RS_SSE_FIXTURE;
-use crate::flags::OPENAI_REQUEST_MAX_RETRIES;
-use crate::flags::OPENAI_STREAM_IDLE_TIMEOUT_MS;
 use crate::model_provider_info::ModelProviderInfo;
 use crate::model_provider_info::WireApi;
 use crate::models::ResponseItem;
@@ -109,7 +107,7 @@ impl ModelClient {
         if let Some(path) = &*CODEX_RS_SSE_FIXTURE {
             // short circuit for tests
             warn!(path, "Streaming from fixture");
-            return stream_from_fixture(path).await;
+            return stream_from_fixture(path, self.provider.clone()).await;
         }
 
         let full_instructions = prompt.get_full_instructions(&self.config.model);
@@ -125,6 +123,7 @@ impl ModelClient {
             reasoning,
             previous_response_id: prompt.prev_id.clone(),
             store: prompt.store,
+            // TODO: make this configurable
             stream: true,
         };
 
@@ -135,6 +134,7 @@ impl ModelClient {
         );
 
         let mut attempt = 0;
+        let max_retries = self.provider.request_max_retries();
         loop {
             attempt += 1;
 
@@ -148,11 +148,15 @@ impl ModelClient {
             let res = req_builder.send().await;
             match res {
                 Ok(resp) if resp.status().is_success() => {
-                    let (tx_event, rx_event) = mpsc::channel::<Result<ResponseEvent>>(16);
+                    let (tx_event, rx_event) = mpsc::channel::<Result<ResponseEvent>>(1600);
 
                     // spawn task to process SSE
                     let stream = resp.bytes_stream().map_err(CodexErr::Reqwest);
-                    tokio::spawn(process_sse(stream, tx_event));
+                    tokio::spawn(process_sse(
+                        stream,
+                        tx_event,
+                        self.provider.stream_idle_timeout(),
+                    ));
 
                     return Ok(ResponseStream { rx_event });
                 }
@@ -171,7 +175,7 @@ impl ModelClient {
                         return Err(CodexErr::UnexpectedStatus(status, body));
                     }
 
-                    if attempt > *OPENAI_REQUEST_MAX_RETRIES {
+                    if attempt > max_retries {
                         return Err(CodexErr::RetryLimit(status));
                     }
 
@@ -188,7 +192,7 @@ impl ModelClient {
                     tokio::time::sleep(delay).await;
                 }
                 Err(e) => {
-                    if attempt > *OPENAI_REQUEST_MAX_RETRIES {
+                    if attempt > max_retries {
                         return Err(e.into());
                     }
                     let delay = backoff(attempt);
@@ -197,6 +201,10 @@ impl ModelClient {
             }
         }
     }
+
+    pub fn get_provider(&self) -> ModelProviderInfo {
+        self.provider.clone()
+    }
 }
 
 #[derive(Debug, Deserialize, Serialize)]
@@ -205,6 +213,7 @@ struct SseEvent {
     kind: String,
     response: Option<Value>,
     item: Option<Value>,
+    delta: Option<String>,
 }
 
 #[derive(Debug, Deserialize)]
@@ -247,14 +256,16 @@ struct ResponseCompletedOutputTokensDetails {
     reasoning_tokens: u64,
 }
 
-async fn process_sse<S>(stream: S, tx_event: mpsc::Sender<Result<ResponseEvent>>)
-where
+async fn process_sse<S>(
+    stream: S,
+    tx_event: mpsc::Sender<Result<ResponseEvent>>,
+    idle_timeout: Duration,
+) where
     S: Stream<Item = Result<Bytes>> + Unpin,
 {
     let mut stream = stream.eventsource();
 
     // If the stream stays completely silent for an extended period treat it as disconnected.
-    let idle_timeout = *OPENAI_STREAM_IDLE_TIMEOUT_MS;
     // The response id returned from the "complete" message.
     let mut response_completed: Option<ResponseCompleted> = None;
 
@@ -315,7 +326,7 @@ where
             // duplicated `output` array embedded in the `response.completed`
             // payload.  That produced two concrete issues:
             //   1. No real‑time streaming – the user only saw output after the
-            //      entire turn had finished, which broke the “typing” UX and
+            //      entire turn had finished, which broke the "typing" UX and
             //      made long‑running turns look stalled.
             //   2. Duplicate `function_call_output` items – both the
             //      individual *and* the completed array were forwarded, which
@@ -337,6 +348,22 @@ where
                     return;
                 }
             }
+            "response.output_text.delta" => {
+                if let Some(delta) = event.delta {
+                    let event = ResponseEvent::OutputTextDelta(delta);
+                    if tx_event.send(Ok(event)).await.is_err() {
+                        return;
+                    }
+                }
+            }
+            "response.reasoning_summary_text.delta" => {
+                if let Some(delta) = event.delta {
+                    let event = ResponseEvent::ReasoningSummaryDelta(delta);
+                    if tx_event.send(Ok(event)).await.is_err() {
+                        return;
+                    }
+                }
+            }
             "response.created" => {
                 if event.response.is_some() {
                     let _ = tx_event.send(Ok(ResponseEvent::Created {})).await;
@@ -360,10 +387,8 @@ where
             | "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.
@@ -374,8 +399,11 @@ where
 }
 
 /// used in tests to stream from a text SSE file
-async fn stream_from_fixture(path: impl AsRef<Path>) -> Result<ResponseStream> {
-    let (tx_event, rx_event) = mpsc::channel::<Result<ResponseEvent>>(16);
+async fn stream_from_fixture(
+    path: impl AsRef<Path>,
+    provider: ModelProviderInfo,
+) -> Result<ResponseStream> {
+    let (tx_event, rx_event) = mpsc::channel::<Result<ResponseEvent>>(1600);
     let f = std::fs::File::open(path.as_ref())?;
     let lines = std::io::BufReader::new(f).lines();
 
@@ -388,7 +416,11 @@ async fn stream_from_fixture(path: impl AsRef<Path>) -> Result<ResponseStream> {
 
     let rdr = std::io::Cursor::new(content);
     let stream = ReaderStream::new(rdr).map_err(CodexErr::Io);
-    tokio::spawn(process_sse(stream, tx_event));
+    tokio::spawn(process_sse(
+        stream,
+        tx_event,
+        provider.stream_idle_timeout(),
+    ));
     Ok(ResponseStream { rx_event })
 }
 
@@ -408,7 +440,10 @@ mod tests {
 
     /// Runs the SSE parser on pre-chunked byte slices and returns every event
     /// (including any final `Err` from a stream-closure check).
-    async fn collect_events(chunks: &[&[u8]]) -> Vec<Result<ResponseEvent>> {
+    async fn collect_events(
+        chunks: &[&[u8]],
+        provider: ModelProviderInfo,
+    ) -> Vec<Result<ResponseEvent>> {
         let mut builder = IoBuilder::new();
         for chunk in chunks {
             builder.read(chunk);
@@ -417,7 +452,7 @@ mod tests {
         let reader = builder.build();
         let stream = ReaderStream::new(reader).map_err(CodexErr::Io);
         let (tx, mut rx) = mpsc::channel::<Result<ResponseEvent>>(16);
-        tokio::spawn(process_sse(stream, tx));
+        tokio::spawn(process_sse(stream, tx, provider.stream_idle_timeout()));
 
         let mut events = Vec::new();
         while let Some(ev) = rx.recv().await {
@@ -428,7 +463,10 @@ mod tests {
 
     /// Builds an in-memory SSE stream from JSON fixtures and returns only the
     /// successfully parsed events (panics on internal channel errors).
-    async fn run_sse(events: Vec<serde_json::Value>) -> Vec<ResponseEvent> {
+    async fn run_sse(
+        events: Vec<serde_json::Value>,
+        provider: ModelProviderInfo,
+    ) -> Vec<ResponseEvent> {
         let mut body = String::new();
         for e in events {
             let kind = e
@@ -444,7 +482,7 @@ mod tests {
 
         let (tx, mut rx) = mpsc::channel::<Result<ResponseEvent>>(8);
         let stream = ReaderStream::new(std::io::Cursor::new(body)).map_err(CodexErr::Io);
-        tokio::spawn(process_sse(stream, tx));
+        tokio::spawn(process_sse(stream, tx, provider.stream_idle_timeout()));
 
         let mut out = Vec::new();
         while let Some(ev) = rx.recv().await {
@@ -489,7 +527,25 @@ mod tests {
         let sse2 = format!("event: response.output_item.done\ndata: {item2}\n\n");
         let sse3 = format!("event: response.completed\ndata: {completed}\n\n");
 
-        let events = collect_events(&[sse1.as_bytes(), sse2.as_bytes(), sse3.as_bytes()]).await;
+        let provider = ModelProviderInfo {
+            name: "test".to_string(),
+            base_url: "https://test.com".to_string(),
+            env_key: Some("TEST_API_KEY".to_string()),
+            env_key_instructions: None,
+            wire_api: WireApi::Responses,
+            query_params: None,
+            http_headers: None,
+            env_http_headers: None,
+            request_max_retries: Some(0),
+            stream_max_retries: Some(0),
+            stream_idle_timeout_ms: Some(1000),
+        };
+
+        let events = collect_events(
+            &[sse1.as_bytes(), sse2.as_bytes(), sse3.as_bytes()],
+            provider,
+        )
+        .await;
 
         assert_eq!(events.len(), 3);
 
@@ -530,8 +586,21 @@ mod tests {
         .to_string();
 
         let sse1 = format!("event: response.output_item.done\ndata: {item1}\n\n");
+        let provider = ModelProviderInfo {
+            name: "test".to_string(),
+            base_url: "https://test.com".to_string(),
+            env_key: Some("TEST_API_KEY".to_string()),
+            env_key_instructions: None,
+            wire_api: WireApi::Responses,
+            query_params: None,
+            http_headers: None,
+            env_http_headers: None,
+            request_max_retries: Some(0),
+            stream_max_retries: Some(0),
+            stream_idle_timeout_ms: Some(1000),
+        };
 
-        let events = collect_events(&[sse1.as_bytes()]).await;
+        let events = collect_events(&[sse1.as_bytes()], provider).await;
 
         assert_eq!(events.len(), 2);
 
@@ -619,7 +688,21 @@ mod tests {
             let mut evs = vec![case.event];
             evs.push(completed.clone());
 
-            let out = run_sse(evs).await;
+            let provider = ModelProviderInfo {
+                name: "test".to_string(),
+                base_url: "https://test.com".to_string(),
+                env_key: Some("TEST_API_KEY".to_string()),
+                env_key_instructions: None,
+                wire_api: WireApi::Responses,
+                query_params: None,
+                http_headers: None,
+                env_http_headers: None,
+                request_max_retries: Some(0),
+                stream_max_retries: Some(0),
+                stream_idle_timeout_ms: Some(1000),
+            };
+
+            let out = run_sse(evs, provider).await;
             assert_eq!(out.len(), case.expected_len, "case {}", case.name);
             assert!(
                 (case.expect_first)(&out[0]),

codex-rs/core/src/client_common.rs

@@ -57,6 +57,8 @@ pub enum ResponseEvent {
         response_id: String,
         token_usage: Option<TokenUsage>,
     },
+    OutputTextDelta(String),
+    ReasoningSummaryDelta(String),
 }
 
 #[derive(Debug, Serialize)]

codex-rs/core/src/codex.rs

@@ -49,9 +49,7 @@ use crate::exec::ExecToolCallOutput;
 use crate::exec::SandboxType;
 use crate::exec::process_exec_tool_call;
 use crate::exec_env::create_env;
-use crate::flags::OPENAI_STREAM_MAX_RETRIES;
 use crate::mcp_connection_manager::McpConnectionManager;
-use crate::mcp_connection_manager::try_parse_fully_qualified_tool_name;
 use crate::mcp_tool_call::handle_mcp_tool_call;
 use crate::models::ContentItem;
 use crate::models::FunctionCallOutputPayload;
@@ -61,7 +59,9 @@ use crate::models::ResponseInputItem;
 use crate::models::ResponseItem;
 use crate::models::ShellToolCallParams;
 use crate::project_doc::get_user_instructions;
+use crate::protocol::AgentMessageDeltaEvent;
 use crate::protocol::AgentMessageEvent;
+use crate::protocol::AgentReasoningDeltaEvent;
 use crate::protocol::AgentReasoningEvent;
 use crate::protocol::ApplyPatchApprovalRequestEvent;
 use crate::protocol::AskForApproval;
@@ -103,7 +103,7 @@ impl Codex {
     /// submitted to start the session.
     pub async fn spawn(config: Config, ctrl_c: Arc<Notify>) -> CodexResult<(Codex, String)> {
         let (tx_sub, rx_sub) = async_channel::bounded(64);
-        let (tx_event, rx_event) = async_channel::bounded(64);
+        let (tx_event, rx_event) = async_channel::bounded(1600);
 
         let instructions = get_user_instructions(&config).await;
         let configure_session = Op::ConfigureSession {
@@ -990,6 +990,52 @@ async fn run_task(sess: Arc<Session>, sub_id: String, input: Vec<InputItem>) {
     sess.tx_event.send(event).await.ok();
 }
 
+// ---
+// Helpers --------------------------------------------------------------------
+//
+// When a turn is interrupted before Codex can deliver tool output(s) back to
+// the model, the next request can fail with a 400 from the OpenAI API:
+//   {"error": {"message": "No tool output found for function call call_XXXXX", ...}}
+// Historically this manifested as a confusing retry loop ("stream error: 400 …")
+// because we never learned about the missing `call_id` (the stream was aborted
+// before we observed the `ResponseEvent::OutputItemDone` that would have let us
+// record it in `pending_call_ids`).
+//
+// To make interruption robust we parse the error body for the offending call id
+// and add it to `pending_call_ids` so the very next retry can inject a synthetic
+// `FunctionCallOutput { content: "aborted" }` and satisfy the API contract.
+// -----------------------------------------------------------------------------
+fn extract_missing_tool_call_id(body: &str) -> Option<String> {
+    // Try to parse the canonical JSON error shape first.
+    if let Ok(v) = serde_json::from_str::<serde_json::Value>(body) {
+        if let Some(msg) = v
+            .get("error")
+            .and_then(|e| e.get("message"))
+            .and_then(|m| m.as_str())
+        {
+            if let Some(id) = extract_missing_tool_call_id_from_msg(msg) {
+                return Some(id);
+            }
+        }
+    }
+    // Fallback: scan the raw body.
+    extract_missing_tool_call_id_from_msg(body)
+}
+
+fn extract_missing_tool_call_id_from_msg(msg: &str) -> Option<String> {
+    const NEEDLE: &str = "No tool output found for function call";
+    let idx = msg.find(NEEDLE)?;
+    let rest = &msg[idx + NEEDLE.len()..];
+    // Find the beginning of the call id (typically starts with "call_").
+    let start = rest.find("call_")?;
+    let rest = &rest[start..];
+    // Capture valid id chars [A-Za-z0-9_-/]. Hyphen shows up in some IDs; be permissive.
+    let end = rest
+        .find(|c: char| !(c.is_ascii_alphanumeric() || c == '_' || c == '-' || c == '/'))
+        .unwrap_or(rest.len());
+    Some(rest[..end].to_string())
+}
+
 async fn run_turn(
     sess: &Session,
     sub_id: String,
@@ -1024,13 +1070,58 @@ async fn run_turn(
             Ok(output) => return Ok(output),
             Err(CodexErr::Interrupted) => return Err(CodexErr::Interrupted),
             Err(CodexErr::EnvVar(var)) => return Err(CodexErr::EnvVar(var)),
-            Err(e) => {
-                if retries < *OPENAI_STREAM_MAX_RETRIES {
+            Err(CodexErr::UnexpectedStatus(status, body)) => {
+                // Detect the specific 400 "No tool output found for function call ..." error that
+                // occurs when a user interrupted before Codex could answer a tool call.
+                if status == reqwest::StatusCode::BAD_REQUEST {
+                    if let Some(call_id) = extract_missing_tool_call_id(&body) {
+                        {
+                            let mut state = sess.state.lock().unwrap();
+                            state.pending_call_ids.insert(call_id.clone());
+                        }
+                        // Surface a friendlier background event so users understand the recovery.
+                        sess
+                            .notify_background_event(
+                                &sub_id,
+                                format!(
+                                    "previous turn interrupted before responding to tool {call_id}; sending aborted output and retrying…",
+                                ),
+                            )
+                            .await;
+                        // Immediately retry the turn without consuming a provider stream retry budget.
+                        continue;
+                    }
+                }
+                // Fall through to generic retry path if we could not auto‑recover.
+                let e = CodexErr::UnexpectedStatus(status, body);
+                // Use the configured provider-specific stream retry budget.
+                let max_retries = sess.client.get_provider().stream_max_retries();
+                if retries < max_retries {
                     retries += 1;
                     let delay = backoff(retries);
                     warn!(
-                        "stream disconnected - retrying turn ({retries}/{} in {delay:?})...",
-                        *OPENAI_STREAM_MAX_RETRIES
+                        "stream disconnected - retrying turn ({retries}/{max_retries} in {delay:?})...",
+                    );
+                    sess.notify_background_event(
+                        &sub_id,
+                        format!(
+                            "stream error: {e}; retrying {retries}/{max_retries} in {delay:?}…",
+                        ),
+                    )
+                    .await;
+                    tokio::time::sleep(delay).await;
+                } else {
+                    return Err(e);
+                }
+            }
+            Err(e) => {
+                // Use the configured provider-specific stream retry budget.
+                let max_retries = sess.client.get_provider().stream_max_retries();
+                if retries < max_retries {
+                    retries += 1;
+                    let delay = backoff(retries);
+                    warn!(
+                        "stream disconnected - retrying turn ({retries}/{max_retries} in {delay:?})...",
                     );
 
                     // Surface retry information to any UI/front‑end so the
@@ -1039,8 +1130,7 @@ async fn run_turn(
                     sess.notify_background_event(
                         &sub_id,
                         format!(
-                            "stream error: {e}; retrying {retries}/{} in {:?}…",
-                            *OPENAI_STREAM_MAX_RETRIES, delay
+                            "stream error: {e}; retrying {retries}/{max_retries} in {delay:?}…",
                         ),
                     )
                     .await;
@@ -1121,15 +1211,29 @@ async fn try_run_turn(
 
     let mut stream = sess.client.clone().stream(&prompt).await?;
 
-    // Buffer all the incoming messages from the stream first, then execute them.
-    // If we execute a function call in the middle of handling the stream, it can time out.
-    let mut input = Vec::new();
-    while let Some(event) = stream.next().await {
-        input.push(event?);
-    }
-
     let mut output = Vec::new();
-    for event in input {
+    loop {
+        // Poll the next item from the model stream. We must inspect *both* Ok and Err
+        // cases so that transient stream failures (e.g., dropped SSE connection before
+        // `response.completed`) bubble up and trigger the caller's retry logic.
+        let event = stream.next().await;
+        let Some(event) = event else {
+            // Channel closed without yielding a final Completed event or explicit error.
+            // Treat as a disconnected stream so the caller can retry.
+            return Err(CodexErr::Stream(
+                "stream closed before response.completed".into(),
+            ));
+        };
+
+        let event = match event {
+            Ok(ev) => ev,
+            Err(e) => {
+                // Propagate the underlying stream error to the caller (run_turn), which
+                // will apply the configured `stream_max_retries` policy.
+                return Err(e);
+            }
+        };
+
         match event {
             ResponseEvent::Created => {
                 let mut state = sess.state.lock().unwrap();
@@ -1170,11 +1274,24 @@ async fn try_run_turn(
 
                 let mut state = sess.state.lock().unwrap();
                 state.previous_response_id = Some(response_id);
-                break;
+                return Ok(output);
+            }
+            ResponseEvent::OutputTextDelta(delta) => {
+                let event = Event {
+                    id: sub_id.to_string(),
+                    msg: EventMsg::AgentMessageDelta(AgentMessageDeltaEvent { delta }),
+                };
+                sess.tx_event.send(event).await.ok();
+            }
+            ResponseEvent::ReasoningSummaryDelta(delta) => {
+                let event = Event {
+                    id: sub_id.to_string(),
+                    msg: EventMsg::AgentReasoningDelta(AgentReasoningDeltaEvent { delta }),
+                };
+                sess.tx_event.send(event).await.ok();
             }
         }
     }
-    Ok(output)
 }
 
 async fn handle_response_item(
@@ -1283,7 +1400,7 @@ async fn handle_function_call(
             handle_container_exec_with_params(params, sess, sub_id, call_id).await
         }
         _ => {
-            match try_parse_fully_qualified_tool_name(&name) {
+            match sess.mcp_connection_manager.parse_tool_name(&name) {
                 Some((server, tool_name)) => {
                     // TODO(mbolin): Determine appropriate timeout for tool call.
                     let timeout = None;

codex-rs/core/src/config.rs

@@ -682,6 +682,9 @@ name = "OpenAI using Chat Completions"
 base_url = "https://api.openai.com/v1"
 env_key = "OPENAI_API_KEY"
 wire_api = "chat"
+request_max_retries = 4            # retry failed HTTP requests
+stream_max_retries = 10            # retry dropped SSE streams
+stream_idle_timeout_ms = 300000    # 5m idle timeout
 
 [profiles.o3]
 model = "o3"
@@ -722,7 +725,9 @@ disable_response_storage = true
             query_params: None,
             http_headers: None,
             env_http_headers: None,
-            supports_temperature: true,
+            request_max_retries: Some(4),
+            stream_max_retries: Some(10),
+            stream_idle_timeout_ms: Some(300_000),
         };
         let model_provider_map = {
             let mut model_provider_map = built_in_model_providers();

codex-rs/core/src/flags.rs

@@ -11,14 +11,6 @@ env_flags! {
     pub OPENAI_TIMEOUT_MS: Duration = Duration::from_millis(300_000), |value| {
         value.parse().map(Duration::from_millis)
     };
-    pub OPENAI_REQUEST_MAX_RETRIES: u64 = 4;
-    pub OPENAI_STREAM_MAX_RETRIES: u64 = 10;
-
-    // We generally don't want to disconnect; this updates the timeout to be five minutes
-    // which matches the upstream typescript codex impl.
-    pub OPENAI_STREAM_IDLE_TIMEOUT_MS: Duration = Duration::from_millis(300_000), |value| {
-        value.parse().map(Duration::from_millis)
-    };
 
     /// Fixture path for offline tests (see client.rs).
     pub CODEX_RS_SSE_FIXTURE: Option<&str> = None;

codex-rs/core/src/mcp_connection_manager.rs

@@ -7,6 +7,7 @@
 //! `"<server><MCP_TOOL_NAME_DELIMITER><tool>"` as the key.
 
 use std::collections::HashMap;
+use std::collections::HashSet;
 use std::time::Duration;
 
 use anyhow::Context;
@@ -16,8 +17,12 @@ use codex_mcp_client::McpClient;
 use mcp_types::ClientCapabilities;
 use mcp_types::Implementation;
 use mcp_types::Tool;
+
+use sha1::Digest;
+use sha1::Sha1;
 use tokio::task::JoinSet;
 use tracing::info;
+use tracing::warn;
 
 use crate::config_types::McpServerConfig;
 
@@ -26,7 +31,8 @@ use crate::config_types::McpServerConfig;
 ///
 /// OpenAI requires tool names to conform to `^[a-zA-Z0-9_-]+$`, so we must
 /// choose a delimiter from this character set.
-const MCP_TOOL_NAME_DELIMITER: &str = "__OAI_CODEX_MCP__";
+const MCP_TOOL_NAME_DELIMITER: &str = "__";
+const MAX_TOOL_NAME_LENGTH: usize = 64;
 
 /// Timeout for the `tools/list` request.
 const LIST_TOOLS_TIMEOUT: Duration = Duration::from_secs(10);
@@ -35,16 +41,42 @@ const LIST_TOOLS_TIMEOUT: Duration = Duration::from_secs(10);
 /// spawned successfully.
 pub type ClientStartErrors = HashMap<String, anyhow::Error>;
 
-fn fully_qualified_tool_name(server: &str, tool: &str) -> String {
-    format!("{server}{MCP_TOOL_NAME_DELIMITER}{tool}")
+fn qualify_tools(tools: Vec<ToolInfo>) -> HashMap<String, ToolInfo> {
+    let mut used_names = HashSet::new();
+    let mut qualified_tools = HashMap::new();
+    for tool in tools {
+        let mut qualified_name = format!(
+            "{}{}{}",
+            tool.server_name, MCP_TOOL_NAME_DELIMITER, tool.tool_name
+        );
+        if qualified_name.len() > MAX_TOOL_NAME_LENGTH {
+            let mut hasher = Sha1::new();
+            hasher.update(qualified_name.as_bytes());
+            let sha1 = hasher.finalize();
+            let sha1_str = format!("{sha1:x}");
+
+            // Truncate to make room for the hash suffix
+            let prefix_len = MAX_TOOL_NAME_LENGTH - sha1_str.len();
+
+            qualified_name = format!("{}{}", &qualified_name[..prefix_len], sha1_str);
+        }
+
+        if used_names.contains(&qualified_name) {
+            warn!("skipping duplicated tool {}", qualified_name);
+            continue;
+        }
+
+        used_names.insert(qualified_name.clone());
+        qualified_tools.insert(qualified_name, tool);
+    }
+
+    qualified_tools
 }
 
-pub(crate) fn try_parse_fully_qualified_tool_name(fq_name: &str) -> Option<(String, String)> {
-    let (server, tool) = fq_name.split_once(MCP_TOOL_NAME_DELIMITER)?;
-    if server.is_empty() || tool.is_empty() {
-        return None;
-    }
-    Some((server.to_string(), tool.to_string()))
+struct ToolInfo {
+    server_name: String,
+    tool_name: String,
+    tool: Tool,
 }
 
 /// A thin wrapper around a set of running [`McpClient`] instances.
@@ -57,7 +89,7 @@ pub(crate) struct McpConnectionManager {
     clients: HashMap<String, std::sync::Arc<McpClient>>,
 
     /// Fully qualified tool name -> tool instance.
-    tools: HashMap<String, Tool>,
+    tools: HashMap<String, ToolInfo>,
 }
 
 impl McpConnectionManager {
@@ -79,9 +111,19 @@ impl McpConnectionManager {
 
         // Launch all configured servers concurrently.
         let mut join_set = JoinSet::new();
+        let mut errors = ClientStartErrors::new();
 
         for (server_name, cfg) in mcp_servers {
-            // TODO: Verify server name: require `^[a-zA-Z0-9_-]+$`?
+            // Validate server name before spawning
+            if !is_valid_mcp_server_name(&server_name) {
+                let error = anyhow::anyhow!(
+                    "invalid server name '{}': must match pattern ^[a-zA-Z0-9_-]+$",
+                    server_name
+                );
+                errors.insert(server_name, error);
+                continue;
+            }
+
             join_set.spawn(async move {
                 let McpServerConfig { command, args, env } = cfg;
                 let client_res = McpClient::new_stdio_client(command, args, env).await;
@@ -117,7 +159,6 @@ impl McpConnectionManager {
 
         let mut clients: HashMap<String, std::sync::Arc<McpClient>> =
             HashMap::with_capacity(join_set.len());
-        let mut errors = ClientStartErrors::new();
 
         while let Some(res) = join_set.join_next().await {
             let (server_name, client_res) = res?; // JoinError propagation
@@ -132,7 +173,9 @@ impl McpConnectionManager {
             }
         }
 
-        let tools = list_all_tools(&clients).await?;
+        let all_tools = list_all_tools(&clients).await?;
+
+        let tools = qualify_tools(all_tools);
 
         Ok((Self { clients, tools }, errors))
     }
@@ -140,7 +183,10 @@ impl McpConnectionManager {
     /// Returns a single map that contains **all** tools. Each key is the
     /// fully-qualified name for the tool.
     pub fn list_all_tools(&self) -> HashMap<String, Tool> {
-        self.tools.clone()
+        self.tools
+            .iter()
+            .map(|(name, tool)| (name.clone(), tool.tool.clone()))
+            .collect()
     }
 
     /// Invoke the tool indicated by the (server, tool) pair.
@@ -162,13 +208,19 @@ impl McpConnectionManager {
             .await
             .with_context(|| format!("tool call failed for `{server}/{tool}`"))
     }
+
+    pub fn parse_tool_name(&self, tool_name: &str) -> Option<(String, String)> {
+        self.tools
+            .get(tool_name)
+            .map(|tool| (tool.server_name.clone(), tool.tool_name.clone()))
+    }
 }
 
 /// Query every server for its available tools and return a single map that
 /// contains **all** tools. Each key is the fully-qualified name for the tool.
-pub async fn list_all_tools(
+async fn list_all_tools(
     clients: &HashMap<String, std::sync::Arc<McpClient>>,
-) -> Result<HashMap<String, Tool>> {
+) -> Result<Vec<ToolInfo>> {
     let mut join_set = JoinSet::new();
 
     // Spawn one task per server so we can query them concurrently. This
@@ -185,18 +237,19 @@ pub async fn list_all_tools(
         });
     }
 
-    let mut aggregated: HashMap<String, Tool> = HashMap::with_capacity(join_set.len());
+    let mut aggregated: Vec<ToolInfo> = Vec::with_capacity(join_set.len());
 
     while let Some(join_res) = join_set.join_next().await {
         let (server_name, list_result) = join_res?;
         let list_result = list_result?;
 
         for tool in list_result.tools {
-            // TODO(mbolin): escape tool names that contain invalid characters.
-            let fq_name = fully_qualified_tool_name(&server_name, &tool.name);
-            if aggregated.insert(fq_name.clone(), tool).is_some() {
-                panic!("tool name collision for '{fq_name}': suspicious");
-            }
+            let tool_info = ToolInfo {
+                server_name: server_name.clone(),
+                tool_name: tool.name.clone(),
+                tool,
+            };
+            aggregated.push(tool_info);
         }
     }
 
@@ -208,3 +261,97 @@ pub async fn list_all_tools(
 
     Ok(aggregated)
 }
+
+fn is_valid_mcp_server_name(server_name: &str) -> bool {
+    !server_name.is_empty()
+        && server_name
+            .chars()
+            .all(|c| c.is_ascii_alphanumeric() || c == '_' || c == '-')
+}
+
+#[cfg(test)]
+#[allow(clippy::unwrap_used)]
+mod tests {
+    use super::*;
+    use mcp_types::ToolInputSchema;
+
+    fn create_test_tool(server_name: &str, tool_name: &str) -> ToolInfo {
+        ToolInfo {
+            server_name: server_name.to_string(),
+            tool_name: tool_name.to_string(),
+            tool: Tool {
+                annotations: None,
+                description: Some(format!("Test tool: {tool_name}")),
+                input_schema: ToolInputSchema {
+                    properties: None,
+                    required: None,
+                    r#type: "object".to_string(),
+                },
+                name: tool_name.to_string(),
+            },
+        }
+    }
+
+    #[test]
+    fn test_qualify_tools_short_non_duplicated_names() {
+        let tools = vec![
+            create_test_tool("server1", "tool1"),
+            create_test_tool("server1", "tool2"),
+        ];
+
+        let qualified_tools = qualify_tools(tools);
+
+        assert_eq!(qualified_tools.len(), 2);
+        assert!(qualified_tools.contains_key("server1__tool1"));
+        assert!(qualified_tools.contains_key("server1__tool2"));
+    }
+
+    #[test]
+    fn test_qualify_tools_duplicated_names_skipped() {
+        let tools = vec![
+            create_test_tool("server1", "duplicate_tool"),
+            create_test_tool("server1", "duplicate_tool"),
+        ];
+
+        let qualified_tools = qualify_tools(tools);
+
+        // Only the first tool should remain, the second is skipped
+        assert_eq!(qualified_tools.len(), 1);
+        assert!(qualified_tools.contains_key("server1__duplicate_tool"));
+    }
+
+    #[test]
+    fn test_qualify_tools_long_names_same_server() {
+        let server_name = "my_server";
+
+        let tools = vec![
+            create_test_tool(
+                server_name,
+                "extremely_lengthy_function_name_that_absolutely_surpasses_all_reasonable_limits",
+            ),
+            create_test_tool(
+                server_name,
+                "yet_another_extremely_lengthy_function_name_that_absolutely_surpasses_all_reasonable_limits",
+            ),
+        ];
+
+        let qualified_tools = qualify_tools(tools);
+
+        assert_eq!(qualified_tools.len(), 2);
+
+        let mut keys: Vec<_> = qualified_tools.keys().cloned().collect();
+        keys.sort();
+
+        assert_eq!(keys[0].len(), 64);
+        assert_eq!(
+            keys[0],
+            "my_server__extremely_lena02e507efc5a9de88637e436690364fd4219e4ef"
+        );
+
+        assert_eq!(keys[1].len(), 64);
+        assert_eq!(
+            keys[1],
+            "my_server__yet_another_e1c3987bd9c50b826cbe1687966f79f0c602d19ca"
+        );
+    }
+}

codex-rs/core/src/model_provider_info.rs

@@ -9,6 +9,7 @@ use serde::Deserialize;
 use serde::Serialize;
 use std::collections::HashMap;
 use std::env::VarError;
+use std::time::Duration;
 
 use crate::error::EnvVarError;
 use crate::openai_api_key::get_openai_api_key;
@@ -16,6 +17,9 @@ use crate::openai_api_key::get_openai_api_key;
 /// Value for the `OpenAI-Originator` header that is sent with requests to
 /// OpenAI.
 const OPENAI_ORIGINATOR_HEADER: &str = "codex_cli_rs";
+const DEFAULT_STREAM_IDLE_TIMEOUT_MS: u64 = 300_000;
+const DEFAULT_STREAM_MAX_RETRIES: u64 = 10;
+const DEFAULT_REQUEST_MAX_RETRIES: u64 = 4;
 
 /// Wire protocol that the provider speaks. Most third-party services only
 /// implement the classic OpenAI Chat Completions JSON schema, whereas OpenAI
@@ -26,7 +30,7 @@ const OPENAI_ORIGINATOR_HEADER: &str = "codex_cli_rs";
 #[derive(Debug, Clone, Copy, Default, PartialEq, Eq, Serialize, Deserialize)]
 #[serde(rename_all = "lowercase")]
 pub enum WireApi {
-    /// The experimental “Responses” API exposed by OpenAI at `/v1/responses`.
+    /// The experimental "Responses" API exposed by OpenAI at `/v1/responses`.
     Responses,
 
     /// Regular Chat Completions compatible with `/v1/chat/completions`.
@@ -65,13 +69,15 @@ pub struct ModelProviderInfo {
     /// value is empty, the header will not be included in the request.
     pub env_http_headers: Option<HashMap<String, String>>,
 
-    /// Whether the provider accepts an explicit `temperature` parameter.
-    #[serde(default = "default_supports_temperature")]
-    pub supports_temperature: bool,
-}
+    /// Maximum number of times to retry a failed HTTP request to this provider.
+    pub request_max_retries: Option<u64>,
 
-const fn default_supports_temperature() -> bool {
-    true
+    /// Number of times to retry reconnecting a dropped streaming response before failing.
+    pub stream_max_retries: Option<u64>,
+
+    /// Idle timeout (in milliseconds) to wait for activity on a streaming response before treating
+    /// the connection as lost.
+    pub stream_idle_timeout_ms: Option<u64>,
 }
 
 impl ModelProviderInfo {
@@ -169,6 +175,25 @@ impl ModelProviderInfo {
             None => Ok(None),
         }
     }
+
+    /// Effective maximum number of request retries for this provider.
+    pub fn request_max_retries(&self) -> u64 {
+        self.request_max_retries
+            .unwrap_or(DEFAULT_REQUEST_MAX_RETRIES)
+    }
+
+    /// Effective maximum number of stream reconnection attempts for this provider.
+    pub fn stream_max_retries(&self) -> u64 {
+        self.stream_max_retries
+            .unwrap_or(DEFAULT_STREAM_MAX_RETRIES)
+    }
+
+    /// Effective idle timeout for streaming responses.
+    pub fn stream_idle_timeout(&self) -> Duration {
+        self.stream_idle_timeout_ms
+            .map(Duration::from_millis)
+            .unwrap_or(Duration::from_millis(DEFAULT_STREAM_IDLE_TIMEOUT_MS))
+    }
 }
 
 /// Built-in default provider list.
@@ -213,7 +238,10 @@ pub fn built_in_model_providers() -> HashMap<String, ModelProviderInfo> {
                         .into_iter()
                         .collect(),
                 ),
-                supports_temperature: false,
+                // Use global defaults for retry/timeout unless overridden in config.toml.
+                request_max_retries: None,
+                stream_max_retries: None,
+                stream_idle_timeout_ms: None,
             },
         ),
     ]
@@ -243,7 +271,9 @@ base_url = "http://localhost:11434/v1"
             query_params: None,
             http_headers: None,
             env_http_headers: None,
-            supports_temperature: true,
+            request_max_retries: None,
+            stream_max_retries: None,
+            stream_idle_timeout_ms: None,
         };
 
         let provider: ModelProviderInfo = toml::from_str(azure_provider_toml).unwrap();
@@ -269,7 +299,9 @@ query_params = { api-version = "2025-04-01-preview" }
             }),
             http_headers: None,
             env_http_headers: None,
-            supports_temperature: true,
+            request_max_retries: None,
+            stream_max_retries: None,
+            stream_idle_timeout_ms: None,
         };
 
         let provider: ModelProviderInfo = toml::from_str(azure_provider_toml).unwrap();
@@ -298,7 +330,9 @@ env_http_headers = { "X-Example-Env-Header" = "EXAMPLE_ENV_VAR" }
             env_http_headers: Some(maplit::hashmap! {
                 "X-Example-Env-Header".to_string() => "EXAMPLE_ENV_VAR".to_string(),
             }),
-            supports_temperature: true,
+            request_max_retries: None,
+            stream_max_retries: None,
+            stream_idle_timeout_ms: None,
         };
 
         let provider: ModelProviderInfo = toml::from_str(azure_provider_toml).unwrap();

codex-rs/core/src/protocol.rs

@@ -282,9 +282,15 @@ pub enum EventMsg {
     /// Agent text output message
     AgentMessage(AgentMessageEvent),
 
+    /// Agent text output delta message
+    AgentMessageDelta(AgentMessageDeltaEvent),
+
     /// Reasoning event from agent.
     AgentReasoning(AgentReasoningEvent),
 
+    /// Agent reasoning delta event from agent.
+    AgentReasoningDelta(AgentReasoningDeltaEvent),
+
     /// Ack the client's configure message.
     SessionConfigured(SessionConfiguredEvent),
 
@@ -340,11 +346,21 @@ pub struct AgentMessageEvent {
     pub message: String,
 }
 
+#[derive(Debug, Clone, Deserialize, Serialize)]
+pub struct AgentMessageDeltaEvent {
+    pub delta: String,
+}
+
 #[derive(Debug, Clone, Deserialize, Serialize)]
 pub struct AgentReasoningEvent {
     pub text: String,
 }
 
+#[derive(Debug, Clone, Deserialize, Serialize)]
+pub struct AgentReasoningDeltaEvent {
+    pub delta: String,
+}
+
 #[derive(Debug, Clone, Deserialize, Serialize)]
 pub struct McpToolCallBeginEvent {
     /// Identifier so this can be paired with the McpToolCallEnd event.

codex-rs/core/src/rollout.rs

@@ -153,13 +153,15 @@ struct LogFileInfo {
 }
 
 fn create_log_file(config: &Config, session_id: Uuid) -> std::io::Result<LogFileInfo> {
-    // Resolve ~/.codex/sessions and create it if missing.
-    let mut dir = config.codex_home.clone();
-    dir.push(SESSIONS_SUBDIR);
-    fs::create_dir_all(&dir)?;
-
+    // Resolve ~/.codex/sessions/YYYY/MM/DD and create it if missing.
     let timestamp = OffsetDateTime::now_local()
         .map_err(|e| IoError::other(format!("failed to get local time: {e}")))?;
+    let mut dir = config.codex_home.clone();
+    dir.push(SESSIONS_SUBDIR);
+    dir.push(timestamp.year().to_string());
+    dir.push(format!("{:02}", u8::from(timestamp.month())));
+    dir.push(format!("{:02}", timestamp.day()));
+    fs::create_dir_all(&dir)?;
 
     // Custom format for YYYY-MM-DDThh-mm-ss. Use `-` instead of `:` for
     // compatibility with filesystems that do not allow colons in filenames.

codex-rs/core/tests/cli_stream.rs

@@ -2,7 +2,12 @@
 
 use assert_cmd::Command as AssertCommand;
 use codex_core::exec::CODEX_SANDBOX_NETWORK_DISABLED_ENV_VAR;
+use serde_json::Value;
+use std::time::Duration;
+use std::time::Instant;
 use tempfile::TempDir;
+use uuid::Uuid;
+use walkdir::WalkDir;
 use wiremock::Mock;
 use wiremock::MockServer;
 use wiremock::ResponseTemplate;
@@ -71,8 +76,8 @@ async fn chat_mode_stream_cli() {
     println!("Stderr:\n{}", String::from_utf8_lossy(&output.stderr));
     assert!(output.status.success());
     let stdout = String::from_utf8_lossy(&output.stdout);
-    assert!(stdout.contains("hi"));
-    assert_eq!(stdout.matches("hi").count(), 1);
+    let hi_lines = stdout.lines().filter(|line| line.trim() == "hi").count();
+    assert_eq!(hi_lines, 1, "Expected exactly one line with 'hi'");
 
     server.verify().await;
 }
@@ -117,3 +122,154 @@ async fn responses_api_stream_cli() {
     let stdout = String::from_utf8_lossy(&output.stdout);
     assert!(stdout.contains("fixture hello"));
 }
+
+#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
+async fn integration_creates_and_checks_session_file() {
+    // Honor sandbox network restrictions for CI parity with the other tests.
+    if std::env::var(CODEX_SANDBOX_NETWORK_DISABLED_ENV_VAR).is_ok() {
+        println!(
+            "Skipping test because it cannot execute when network is disabled in a Codex sandbox."
+        );
+        return;
+    }
+
+    // 1. Temp home so we read/write isolated session files.
+    let home = TempDir::new().unwrap();
+
+    // 2. Unique marker we'll look for in the session log.
+    let marker = format!("integration-test-{}", Uuid::new_v4());
+    let prompt = format!("echo {marker}");
+
+    // 3. Use the same offline SSE fixture as responses_api_stream_cli so the test is hermetic.
+    let fixture =
+        std::path::Path::new(env!("CARGO_MANIFEST_DIR")).join("tests/cli_responses_fixture.sse");
+
+    // 4. Run the codex CLI through cargo (ensures the right bin is built) and invoke `exec`,
+    //    which is what records a session.
+    let mut cmd = AssertCommand::new("cargo");
+    cmd.arg("run")
+        .arg("-p")
+        .arg("codex-cli")
+        .arg("--quiet")
+        .arg("--")
+        .arg("exec")
+        .arg("--skip-git-repo-check")
+        .arg("-C")
+        .arg(env!("CARGO_MANIFEST_DIR"))
+        .arg(&prompt);
+    cmd.env("CODEX_HOME", home.path())
+        .env("OPENAI_API_KEY", "dummy")
+        .env("CODEX_RS_SSE_FIXTURE", &fixture)
+        // Required for CLI arg parsing even though fixture short-circuits network usage.
+        .env("OPENAI_BASE_URL", "http://unused.local");
+
+    let output = cmd.output().unwrap();
+    assert!(
+        output.status.success(),
+        "codex-cli exec failed: {}",
+        String::from_utf8_lossy(&output.stderr)
+    );
+
+    // 5. Sessions are written asynchronously; wait briefly for the directory to appear.
+    let sessions_dir = home.path().join("sessions");
+    let start = Instant::now();
+    while !sessions_dir.exists() && start.elapsed() < Duration::from_secs(3) {
+        std::thread::sleep(Duration::from_millis(50));
+    }
+
+    // 6. Scan all session files and find the one that contains our marker.
+    let mut matching_files = vec![];
+    for entry in WalkDir::new(&sessions_dir) {
+        let entry = entry.unwrap();
+        if entry.file_type().is_file() && entry.file_name().to_string_lossy().ends_with(".jsonl") {
+            let path = entry.path();
+            let content = std::fs::read_to_string(path).unwrap();
+            let mut lines = content.lines();
+            // Skip SessionMeta (first line)
+            let _ = lines.next();
+            for line in lines {
+                let item: Value = serde_json::from_str(line).unwrap();
+                if let Some("message") = item.get("type").and_then(|t| t.as_str()) {
+                    if let Some(content) = item.get("content") {
+                        if content.to_string().contains(&marker) {
+                            matching_files.push(path.to_owned());
+                            break;
+                        }
+                    }
+                }
+            }
+        }
+    }
+    assert_eq!(
+        matching_files.len(),
+        1,
+        "Expected exactly one session file containing the marker, found {}",
+        matching_files.len()
+    );
+    let path = &matching_files[0];
+
+    // 7. Verify directory structure: sessions/YYYY/MM/DD/filename.jsonl
+    let rel = match path.strip_prefix(&sessions_dir) {
+        Ok(r) => r,
+        Err(_) => panic!("session file should live under sessions/"),
+    };
+    let comps: Vec<String> = rel
+        .components()
+        .map(|c| c.as_os_str().to_string_lossy().into_owned())
+        .collect();
+    assert_eq!(
+        comps.len(),
+        4,
+        "Expected sessions/YYYY/MM/DD/<file>, got {rel:?}"
+    );
+    let year = &comps[0];
+    let month = &comps[1];
+    let day = &comps[2];
+    assert!(
+        year.len() == 4 && year.chars().all(|c| c.is_ascii_digit()),
+        "Year dir not 4-digit numeric: {year}"
+    );
+    assert!(
+        month.len() == 2 && month.chars().all(|c| c.is_ascii_digit()),
+        "Month dir not zero-padded 2-digit numeric: {month}"
+    );
+    assert!(
+        day.len() == 2 && day.chars().all(|c| c.is_ascii_digit()),
+        "Day dir not zero-padded 2-digit numeric: {day}"
+    );
+    // Range checks (best-effort; won't fail on leading zeros)
+    if let Ok(m) = month.parse::<u8>() {
+        assert!((1..=12).contains(&m), "Month out of range: {m}");
+    }
+    if let Ok(d) = day.parse::<u8>() {
+        assert!((1..=31).contains(&d), "Day out of range: {d}");
+    }
+
+    // 8. Parse SessionMeta line and basic sanity checks.
+    let content = std::fs::read_to_string(path).unwrap();
+    let mut lines = content.lines();
+    let meta: Value = serde_json::from_str(lines.next().unwrap()).unwrap();
+    assert!(meta.get("id").is_some(), "SessionMeta missing id");
+    assert!(
+        meta.get("timestamp").is_some(),
+        "SessionMeta missing timestamp"
+    );
+
+    // 9. Confirm at least one message contains the marker.
+    let mut found_message = false;
+    for line in lines {
+        let item: Value = serde_json::from_str(line).unwrap();
+        if item.get("type").map(|t| t == "message").unwrap_or(false) {
+            if let Some(content) = item.get("content") {
+                if content.to_string().contains(&marker) {
+                    found_message = true;
+                    break;
+                }
+            }
+        }
+    }
+    assert!(
+        found_message,
+        "No message found in session file containing the marker"
+    );
+}

codex-rs/core/tests/live_agent.rs

@@ -45,22 +45,10 @@ async fn spawn_codex() -> Result<Codex, CodexErr> {
         "OPENAI_API_KEY must be set for live tests"
     );
 
-    // Environment tweaks to keep the tests snappy and inexpensive while still
-    // exercising retry/robustness logic.
-    //
-    // NOTE: Starting with the 2024 edition `std::env::set_var` is `unsafe`
-    // because changing the process environment races with any other threads
-    // that might be performing environment look-ups at the same time.
-    // Restrict the unsafety to this tiny block that happens at the very
-    // beginning of the test, before we spawn any background tasks that could
-    // observe the environment.
-    unsafe {
-        std::env::set_var("OPENAI_REQUEST_MAX_RETRIES", "2");
-        std::env::set_var("OPENAI_STREAM_MAX_RETRIES", "2");
-    }
-
     let codex_home = TempDir::new().unwrap();
-    let config = load_default_config_for_test(&codex_home);
+    let mut config = load_default_config_for_test(&codex_home);
+    config.model_provider.request_max_retries = Some(2);
+    config.model_provider.stream_max_retries = Some(2);
     let (agent, _init_id) = Codex::spawn(config, std::sync::Arc::new(Notify::new())).await?;
 
     Ok(agent)
@@ -79,7 +67,7 @@ async fn live_streaming_and_prev_id_reset() {
 
     let codex = spawn_codex().await.unwrap();
 
-    // ---------- Task 1 ----------
+    // ---------- Task 1 ----------
     codex
         .submit(Op::UserInput {
             items: vec![InputItem::Text {
@@ -113,7 +101,7 @@ async fn live_streaming_and_prev_id_reset() {
         "Agent did not stream any AgentMessage before TaskComplete"
     );
 
-    // ---------- Task 2 (same session) ----------
+    // ---------- Task 2 (same session) ----------
     codex
         .submit(Op::UserInput {
             items: vec![InputItem::Text {

codex-rs/core/tests/previous_response_id.rs

@@ -88,13 +88,8 @@ async fn keeps_previous_response_id_between_tasks() {
         .mount(&server)
         .await;
 
-    // Environment
-    // Update environment – `set_var` is `unsafe` starting with the 2024
-    // edition so we group the calls into a single `unsafe { … }` block.
-    unsafe {
-        std::env::set_var("OPENAI_REQUEST_MAX_RETRIES", "0");
-        std::env::set_var("OPENAI_STREAM_MAX_RETRIES", "0");
-    }
+    // Configure retry behavior explicitly to avoid mutating process-wide
+    // environment variables.
     let model_provider = ModelProviderInfo {
         name: "openai".into(),
         base_url: format!("{}/v1", server.uri()),
@@ -107,7 +102,10 @@ async fn keeps_previous_response_id_between_tasks() {
         query_params: None,
         http_headers: None,
         env_http_headers: None,
-        supports_temperature: true,
+        // disable retries so we don't get duplicate calls in this test
+        request_max_retries: Some(0),
+        stream_max_retries: Some(0),
+        stream_idle_timeout_ms: None,
     };
 
     // Init session

codex-rs/core/tests/stream_no_completed.rs

@@ -70,19 +70,8 @@ async fn retries_on_early_close() {
         .mount(&server)
         .await;
 
-    // Environment
-    //
-    // As of Rust 2024 `std::env::set_var` has been made `unsafe` because
-    // mutating the process environment is inherently racy when other threads
-    // are running.  We therefore have to wrap every call in an explicit
-    // `unsafe` block.  These are limited to the test-setup section so the
-    // scope is very small and clearly delineated.
-
-    unsafe {
-        std::env::set_var("OPENAI_REQUEST_MAX_RETRIES", "0");
-        std::env::set_var("OPENAI_STREAM_MAX_RETRIES", "1");
-        std::env::set_var("OPENAI_STREAM_IDLE_TIMEOUT_MS", "2000");
-    }
+    // Configure retry behavior explicitly to avoid mutating process-wide
+    // environment variables.
 
     let model_provider = ModelProviderInfo {
         name: "openai".into(),
@@ -96,7 +85,10 @@ async fn retries_on_early_close() {
         query_params: None,
         http_headers: None,
         env_http_headers: None,
-        supports_temperature: true,
+        // exercise retry path: first attempt yields incomplete stream, so allow 1 retry
+        request_max_retries: Some(0),
+        stream_max_retries: Some(1),
+        stream_idle_timeout_ms: Some(2000),
     };
 
     let ctrl_c = std::sync::Arc::new(tokio::sync::Notify::new());

codex-rs/exec/src/cli.rs

@@ -51,6 +51,10 @@ pub struct Cli {
     #[arg(long = "color", value_enum, default_value_t = Color::Auto)]
     pub color: Color,
 
+    /// Print events to stdout as JSONL.
+    #[arg(long = "json", default_value_t = false)]
+    pub json: bool,
+
     /// Specifies file where the last message from the agent should be written.
     #[arg(long = "output-last-message")]
     pub last_message_file: Option<PathBuf>,

codex-rs/exec/src/event_processor.rs

@@ -1,492 +1,37 @@
-use codex_common::elapsed::format_elapsed;
 use codex_common::summarize_sandbox_policy;
 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;
 use codex_core::protocol::Event;
-use codex_core::protocol::EventMsg;
-use codex_core::protocol::ExecCommandBeginEvent;
-use codex_core::protocol::ExecCommandEndEvent;
-use codex_core::protocol::FileChange;
-use codex_core::protocol::McpToolCallBeginEvent;
-use codex_core::protocol::McpToolCallEndEvent;
-use codex_core::protocol::PatchApplyBeginEvent;
-use codex_core::protocol::PatchApplyEndEvent;
-use codex_core::protocol::SessionConfiguredEvent;
-use codex_core::protocol::TokenUsage;
-use owo_colors::OwoColorize;
-use owo_colors::Style;
-use shlex::try_join;
-use std::collections::HashMap;
-use std::time::Instant;
 
-/// This should be configurable. When used in CI, users may not want to impose
-/// a limit so they can see the full transcript.
-const MAX_OUTPUT_LINES_FOR_EXEC_TOOL_CALL: usize = 20;
+pub(crate) trait EventProcessor {
+    /// Print summary of effective configuration and user prompt.
+    fn print_config_summary(&mut self, config: &Config, prompt: &str);
 
-pub(crate) struct EventProcessor {
-    call_id_to_command: HashMap<String, ExecCommandBegin>,
-    call_id_to_patch: HashMap<String, PatchApplyBegin>,
-
-    /// Tracks in-flight MCP tool calls so we can calculate duration and print
-    /// a concise summary when the corresponding `McpToolCallEnd` event is
-    /// received.
-    call_id_to_tool_call: HashMap<String, McpToolCallBegin>,
-
-    // To ensure that --color=never is respected, ANSI escapes _must_ be added
-    // using .style() with one of these fields. If you need a new style, add a
-    // new field here.
-    bold: Style,
-    italic: Style,
-    dimmed: Style,
-
-    magenta: Style,
-    red: Style,
-    green: Style,
-    cyan: Style,
-
-    /// Whether to include `AgentReasoning` events in the output.
-    show_agent_reasoning: bool,
+    /// Handle a single event emitted by the agent.
+    fn process_event(&mut self, event: Event);
 }
 
-impl EventProcessor {
-    pub(crate) fn create_with_ansi(with_ansi: bool, show_agent_reasoning: bool) -> Self {
-        let call_id_to_command = HashMap::new();
-        let call_id_to_patch = HashMap::new();
-        let call_id_to_tool_call = HashMap::new();
-
-        if with_ansi {
-            Self {
-                call_id_to_command,
-                call_id_to_patch,
-                bold: Style::new().bold(),
-                italic: Style::new().italic(),
-                dimmed: Style::new().dimmed(),
-                magenta: Style::new().magenta(),
-                red: Style::new().red(),
-                green: Style::new().green(),
-                cyan: Style::new().cyan(),
-                call_id_to_tool_call,
-                show_agent_reasoning,
-            }
-        } else {
-            Self {
-                call_id_to_command,
-                call_id_to_patch,
-                bold: Style::new(),
-                italic: Style::new(),
-                dimmed: Style::new(),
-                magenta: Style::new(),
-                red: Style::new(),
-                green: Style::new(),
-                cyan: Style::new(),
-                call_id_to_tool_call,
-                show_agent_reasoning,
-            }
-        }
-    }
-}
-
-struct ExecCommandBegin {
-    command: Vec<String>,
-    start_time: Instant,
-}
-
-/// Metadata captured when an `McpToolCallBegin` event is received.
-struct McpToolCallBegin {
-    /// Formatted invocation string, e.g. `server.tool({"city":"sf"})`.
-    invocation: String,
-    /// Timestamp when the call started so we can compute duration later.
-    start_time: Instant,
-}
-
-struct PatchApplyBegin {
-    start_time: Instant,
-    auto_approved: bool,
-}
-
-// Timestamped println helper. The timestamp is styled with self.dimmed.
-#[macro_export]
-macro_rules! ts_println {
-    ($self:ident, $($arg:tt)*) => {{
-        let now = chrono::Utc::now();
-        let formatted = now.format("[%Y-%m-%dT%H:%M:%S]");
-        print!("{} ", formatted.style($self.dimmed));
-        println!($($arg)*);
-    }};
-}
-
-impl EventProcessor {
-    /// Print a concise summary of the effective configuration that will be used
-    /// for the session. This mirrors the information shown in the TUI welcome
-    /// screen.
-    pub(crate) fn print_config_summary(&mut self, config: &Config, prompt: &str) {
-        const VERSION: &str = env!("CARGO_PKG_VERSION");
-        ts_println!(
-            self,
-            "OpenAI Codex v{} (research preview)\n--------",
-            VERSION
-        );
-
-        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", summarize_sandbox_policy(&config.sandbox_policy)),
-        ];
-        if config.model_provider.wire_api == WireApi::Responses
-            && model_supports_reasoning_summaries(config)
-        {
-            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!("--------");
-
-        // Echo the prompt that will be sent to the agent so it is visible in the
-        // transcript/logs before any events come in. Note the prompt may have been
-        // read from stdin, so it may not be visible in the terminal otherwise.
-        ts_println!(
-            self,
-            "{}\n{}",
-            "User instructions:".style(self.bold).style(self.cyan),
-            prompt
-        );
+pub(crate) fn create_config_summary_entries(config: &Config) -> Vec<(&'static str, String)> {
+    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", summarize_sandbox_policy(&config.sandbox_policy)),
+    ];
+    if config.model_provider.wire_api == WireApi::Responses
+        && model_supports_reasoning_summaries(config)
+    {
+        entries.push((
+            "reasoning effort",
+            config.model_reasoning_effort.to_string(),
+        ));
+        entries.push((
+            "reasoning summaries",
+            config.model_reasoning_summary.to_string(),
+        ));
     }
 
-    pub(crate) fn process_event(&mut self, event: Event) {
-        let Event { id: _, msg } = event;
-        match msg {
-            EventMsg::Error(ErrorEvent { message }) => {
-                let prefix = "ERROR:".style(self.red);
-                ts_println!(self, "{prefix} {message}");
-            }
-            EventMsg::BackgroundEvent(BackgroundEventEvent { message }) => {
-                ts_println!(self, "{}", message.style(self.dimmed));
-            }
-            EventMsg::TaskStarted | EventMsg::TaskComplete(_) => {
-                // Ignore.
-            }
-            EventMsg::TokenCount(TokenUsage { total_tokens, .. }) => {
-                ts_println!(self, "tokens used: {total_tokens}");
-            }
-            EventMsg::AgentMessage(AgentMessageEvent { message }) => {
-                ts_println!(
-                    self,
-                    "{}\n{message}",
-                    "codex".style(self.bold).style(self.magenta)
-                );
-            }
-            EventMsg::ExecCommandBegin(ExecCommandBeginEvent {
-                call_id,
-                command,
-                cwd,
-            }) => {
-                self.call_id_to_command.insert(
-                    call_id.clone(),
-                    ExecCommandBegin {
-                        command: command.clone(),
-                        start_time: Instant::now(),
-                    },
-                );
-                ts_println!(
-                    self,
-                    "{} {} in {}",
-                    "exec".style(self.magenta),
-                    escape_command(&command).style(self.bold),
-                    cwd.to_string_lossy(),
-                );
-            }
-            EventMsg::ExecCommandEnd(ExecCommandEndEvent {
-                call_id,
-                stdout,
-                stderr,
-                exit_code,
-            }) => {
-                let exec_command = self.call_id_to_command.remove(&call_id);
-                let (duration, call) = if let Some(ExecCommandBegin {
-                    command,
-                    start_time,
-                }) = exec_command
-                {
-                    (
-                        format!(" in {}", format_elapsed(start_time)),
-                        format!("{}", escape_command(&command).style(self.bold)),
-                    )
-                } else {
-                    ("".to_string(), format!("exec('{call_id}')"))
-                };
-
-                let output = if exit_code == 0 { stdout } else { stderr };
-                let truncated_output = output
-                    .lines()
-                    .take(MAX_OUTPUT_LINES_FOR_EXEC_TOOL_CALL)
-                    .collect::<Vec<_>>()
-                    .join("\n");
-                match exit_code {
-                    0 => {
-                        let title = format!("{call} succeeded{duration}:");
-                        ts_println!(self, "{}", title.style(self.green));
-                    }
-                    _ => {
-                        let title = format!("{call} exited {exit_code}{duration}:");
-                        ts_println!(self, "{}", title.style(self.red));
-                    }
-                }
-                println!("{}", truncated_output.style(self.dimmed));
-            }
-            EventMsg::McpToolCallBegin(McpToolCallBeginEvent {
-                call_id,
-                server,
-                tool,
-                arguments,
-            }) => {
-                // Build fully-qualified tool name: server.tool
-                let fq_tool_name = format!("{server}.{tool}");
-
-                // Format arguments as compact JSON so they fit on one line.
-                let args_str = arguments
-                    .as_ref()
-                    .map(|v: &serde_json::Value| {
-                        serde_json::to_string(v).unwrap_or_else(|_| v.to_string())
-                    })
-                    .unwrap_or_default();
-
-                let invocation = if args_str.is_empty() {
-                    format!("{fq_tool_name}()")
-                } else {
-                    format!("{fq_tool_name}({args_str})")
-                };
-
-                self.call_id_to_tool_call.insert(
-                    call_id.clone(),
-                    McpToolCallBegin {
-                        invocation: invocation.clone(),
-                        start_time: Instant::now(),
-                    },
-                );
-
-                ts_println!(
-                    self,
-                    "{} {}",
-                    "tool".style(self.magenta),
-                    invocation.style(self.bold),
-                );
-            }
-            EventMsg::McpToolCallEnd(tool_call_end_event) => {
-                let is_success = tool_call_end_event.is_success();
-                let McpToolCallEndEvent { call_id, result } = tool_call_end_event;
-                // Retrieve start time and invocation for duration calculation and labeling.
-                let info = self.call_id_to_tool_call.remove(&call_id);
-
-                let (duration, invocation) = if let Some(McpToolCallBegin {
-                    invocation,
-                    start_time,
-                    ..
-                }) = info
-                {
-                    (format!(" in {}", format_elapsed(start_time)), invocation)
-                } else {
-                    (String::new(), format!("tool('{call_id}')"))
-                };
-
-                let status_str = if is_success { "success" } else { "failed" };
-                let title_style = if is_success { self.green } else { self.red };
-                let title = format!("{invocation} {status_str}{duration}:");
-
-                ts_println!(self, "{}", title.style(title_style));
-
-                if let Ok(res) = result {
-                    let val: serde_json::Value = res.into();
-                    let pretty =
-                        serde_json::to_string_pretty(&val).unwrap_or_else(|_| val.to_string());
-
-                    for line in pretty.lines().take(MAX_OUTPUT_LINES_FOR_EXEC_TOOL_CALL) {
-                        println!("{}", line.style(self.dimmed));
-                    }
-                }
-            }
-            EventMsg::PatchApplyBegin(PatchApplyBeginEvent {
-                call_id,
-                auto_approved,
-                changes,
-            }) => {
-                // Store metadata so we can calculate duration later when we
-                // receive the corresponding PatchApplyEnd event.
-                self.call_id_to_patch.insert(
-                    call_id.clone(),
-                    PatchApplyBegin {
-                        start_time: Instant::now(),
-                        auto_approved,
-                    },
-                );
-
-                ts_println!(
-                    self,
-                    "{} auto_approved={}:",
-                    "apply_patch".style(self.magenta),
-                    auto_approved,
-                );
-
-                // Pretty-print the patch summary with colored diff markers so
-                // it’s easy to scan in the terminal output.
-                for (path, change) in changes.iter() {
-                    match change {
-                        FileChange::Add { content } => {
-                            let header = format!(
-                                "{} {}",
-                                format_file_change(change),
-                                path.to_string_lossy()
-                            );
-                            println!("{}", header.style(self.magenta));
-                            for line in content.lines() {
-                                println!("{}", line.style(self.green));
-                            }
-                        }
-                        FileChange::Delete => {
-                            let header = format!(
-                                "{} {}",
-                                format_file_change(change),
-                                path.to_string_lossy()
-                            );
-                            println!("{}", header.style(self.magenta));
-                        }
-                        FileChange::Update {
-                            unified_diff,
-                            move_path,
-                        } => {
-                            let header = if let Some(dest) = move_path {
-                                format!(
-                                    "{} {} -> {}",
-                                    format_file_change(change),
-                                    path.to_string_lossy(),
-                                    dest.to_string_lossy()
-                                )
-                            } else {
-                                format!("{} {}", format_file_change(change), path.to_string_lossy())
-                            };
-                            println!("{}", header.style(self.magenta));
-
-                            // Colorize diff lines. We keep file header lines
-                            // (--- / +++) without extra coloring so they are
-                            // still readable.
-                            for diff_line in unified_diff.lines() {
-                                if diff_line.starts_with('+') && !diff_line.starts_with("+++") {
-                                    println!("{}", diff_line.style(self.green));
-                                } else if diff_line.starts_with('-')
-                                    && !diff_line.starts_with("---")
-                                {
-                                    println!("{}", diff_line.style(self.red));
-                                } else {
-                                    println!("{diff_line}");
-                                }
-                            }
-                        }
-                    }
-                }
-            }
-            EventMsg::PatchApplyEnd(PatchApplyEndEvent {
-                call_id,
-                stdout,
-                stderr,
-                success,
-            }) => {
-                let patch_begin = self.call_id_to_patch.remove(&call_id);
-
-                // Compute duration and summary label similar to exec commands.
-                let (duration, label) = if let Some(PatchApplyBegin {
-                    start_time,
-                    auto_approved,
-                }) = patch_begin
-                {
-                    (
-                        format!(" in {}", format_elapsed(start_time)),
-                        format!("apply_patch(auto_approved={auto_approved})"),
-                    )
-                } else {
-                    (String::new(), format!("apply_patch('{call_id}')"))
-                };
-
-                let (exit_code, output, title_style) = if success {
-                    (0, stdout, self.green)
-                } else {
-                    (1, stderr, self.red)
-                };
-
-                let title = format!("{label} exited {exit_code}{duration}:");
-                ts_println!(self, "{}", title.style(title_style));
-                for line in output.lines() {
-                    println!("{}", line.style(self.dimmed));
-                }
-            }
-            EventMsg::ExecApprovalRequest(_) => {
-                // Should we exit?
-            }
-            EventMsg::ApplyPatchApprovalRequest(_) => {
-                // Should we exit?
-            }
-            EventMsg::AgentReasoning(agent_reasoning_event) => {
-                if self.show_agent_reasoning {
-                    ts_println!(
-                        self,
-                        "{}\n{}",
-                        "thinking".style(self.italic).style(self.magenta),
-                        agent_reasoning_event.text
-                    );
-                }
-            }
-            EventMsg::SessionConfigured(session_configured_event) => {
-                let SessionConfiguredEvent {
-                    session_id,
-                    model,
-                    history_log_id: _,
-                    history_entry_count: _,
-                } = session_configured_event;
-
-                ts_println!(
-                    self,
-                    "{} {}",
-                    "codex session".style(self.magenta).style(self.bold),
-                    session_id.to_string().style(self.dimmed)
-                );
-
-                ts_println!(self, "model: {}", model);
-                println!();
-            }
-            EventMsg::GetHistoryEntryResponse(_) => {
-                // Currently ignored in exec output.
-            }
-        }
-    }
-}
-
-fn escape_command(command: &[String]) -> String {
-    try_join(command.iter().map(|s| s.as_str())).unwrap_or_else(|_| command.join(" "))
-}
-
-fn format_file_change(change: &FileChange) -> &'static str {
-    match change {
-        FileChange::Add { .. } => "A",
-        FileChange::Delete => "D",
-        FileChange::Update {
-            move_path: Some(_), ..
-        } => "R",
-        FileChange::Update {
-            move_path: None, ..
-        } => "M",
-    }
+    entries
 }

codex-rs/exec/src/event_processor_with_human_output.rs

@@ -0,0 +1,520 @@
+use codex_common::elapsed::format_elapsed;
+use codex_core::config::Config;
+use codex_core::protocol::AgentMessageDeltaEvent;
+use codex_core::protocol::AgentMessageEvent;
+use codex_core::protocol::AgentReasoningDeltaEvent;
+use codex_core::protocol::BackgroundEventEvent;
+use codex_core::protocol::ErrorEvent;
+use codex_core::protocol::Event;
+use codex_core::protocol::EventMsg;
+use codex_core::protocol::ExecCommandBeginEvent;
+use codex_core::protocol::ExecCommandEndEvent;
+use codex_core::protocol::FileChange;
+use codex_core::protocol::McpToolCallBeginEvent;
+use codex_core::protocol::McpToolCallEndEvent;
+use codex_core::protocol::PatchApplyBeginEvent;
+use codex_core::protocol::PatchApplyEndEvent;
+use codex_core::protocol::SessionConfiguredEvent;
+use codex_core::protocol::TokenUsage;
+use owo_colors::OwoColorize;
+use owo_colors::Style;
+use shlex::try_join;
+use std::collections::HashMap;
+use std::io::Write;
+use std::time::Instant;
+
+use crate::event_processor::EventProcessor;
+use crate::event_processor::create_config_summary_entries;
+
+/// This should be configurable. When used in CI, users may not want to impose
+/// a limit so they can see the full transcript.
+const MAX_OUTPUT_LINES_FOR_EXEC_TOOL_CALL: usize = 20;
+pub(crate) struct EventProcessorWithHumanOutput {
+    call_id_to_command: HashMap<String, ExecCommandBegin>,
+    call_id_to_patch: HashMap<String, PatchApplyBegin>,
+
+    /// Tracks in-flight MCP tool calls so we can calculate duration and print
+    /// a concise summary when the corresponding `McpToolCallEnd` event is
+    /// received.
+    call_id_to_tool_call: HashMap<String, McpToolCallBegin>,
+
+    // To ensure that --color=never is respected, ANSI escapes _must_ be added
+    // using .style() with one of these fields. If you need a new style, add a
+    // new field here.
+    bold: Style,
+    italic: Style,
+    dimmed: Style,
+
+    magenta: Style,
+    red: Style,
+    green: Style,
+    cyan: Style,
+
+    /// Whether to include `AgentReasoning` events in the output.
+    show_agent_reasoning: bool,
+    answer_started: bool,
+    reasoning_started: bool,
+}
+
+impl EventProcessorWithHumanOutput {
+    pub(crate) fn create_with_ansi(with_ansi: bool, config: &Config) -> Self {
+        let call_id_to_command = HashMap::new();
+        let call_id_to_patch = HashMap::new();
+        let call_id_to_tool_call = HashMap::new();
+
+        if with_ansi {
+            Self {
+                call_id_to_command,
+                call_id_to_patch,
+                bold: Style::new().bold(),
+                italic: Style::new().italic(),
+                dimmed: Style::new().dimmed(),
+                magenta: Style::new().magenta(),
+                red: Style::new().red(),
+                green: Style::new().green(),
+                cyan: Style::new().cyan(),
+                call_id_to_tool_call,
+                show_agent_reasoning: !config.hide_agent_reasoning,
+                answer_started: false,
+                reasoning_started: false,
+            }
+        } else {
+            Self {
+                call_id_to_command,
+                call_id_to_patch,
+                bold: Style::new(),
+                italic: Style::new(),
+                dimmed: Style::new(),
+                magenta: Style::new(),
+                red: Style::new(),
+                green: Style::new(),
+                cyan: Style::new(),
+                call_id_to_tool_call,
+                show_agent_reasoning: !config.hide_agent_reasoning,
+                answer_started: false,
+                reasoning_started: false,
+            }
+        }
+    }
+}
+
+struct ExecCommandBegin {
+    command: Vec<String>,
+    start_time: Instant,
+}
+
+/// Metadata captured when an `McpToolCallBegin` event is received.
+struct McpToolCallBegin {
+    /// Formatted invocation string, e.g. `server.tool({"city":"sf"})`.
+    invocation: String,
+    /// Timestamp when the call started so we can compute duration later.
+    start_time: Instant,
+}
+
+struct PatchApplyBegin {
+    start_time: Instant,
+    auto_approved: bool,
+}
+
+// Timestamped println helper. The timestamp is styled with self.dimmed.
+#[macro_export]
+macro_rules! ts_println {
+    ($self:ident, $($arg:tt)*) => {{
+        let now = chrono::Utc::now();
+        let formatted = now.format("[%Y-%m-%dT%H:%M:%S]");
+        print!("{} ", formatted.style($self.dimmed));
+        println!($($arg)*);
+    }};
+}
+
+impl EventProcessor for EventProcessorWithHumanOutput {
+    /// Print a concise summary of the effective configuration that will be used
+    /// for the session. This mirrors the information shown in the TUI welcome
+    /// screen.
+    fn print_config_summary(&mut self, config: &Config, prompt: &str) {
+        const VERSION: &str = env!("CARGO_PKG_VERSION");
+        ts_println!(
+            self,
+            "OpenAI Codex v{} (research preview)\n--------",
+            VERSION
+        );
+
+        let entries = create_config_summary_entries(config);
+
+        for (key, value) in entries {
+            println!("{} {}", format!("{key}:").style(self.bold), value);
+        }
+
+        println!("--------");
+
+        // Echo the prompt that will be sent to the agent so it is visible in the
+        // transcript/logs before any events come in. Note the prompt may have been
+        // read from stdin, so it may not be visible in the terminal otherwise.
+        ts_println!(
+            self,
+            "{}\n{}",
+            "User instructions:".style(self.bold).style(self.cyan),
+            prompt
+        );
+    }
+
+    fn process_event(&mut self, event: Event) {
+        let Event { id: _, msg } = event;
+        match msg {
+            EventMsg::Error(ErrorEvent { message }) => {
+                let prefix = "ERROR:".style(self.red);
+                ts_println!(self, "{prefix} {message}");
+            }
+            EventMsg::BackgroundEvent(BackgroundEventEvent { message }) => {
+                ts_println!(self, "{}", message.style(self.dimmed));
+            }
+            EventMsg::TaskStarted | EventMsg::TaskComplete(_) => {
+                // Ignore.
+            }
+            EventMsg::TokenCount(TokenUsage { total_tokens, .. }) => {
+                ts_println!(self, "tokens used: {total_tokens}");
+            }
+            EventMsg::AgentMessageDelta(AgentMessageDeltaEvent { delta }) => {
+                if !self.answer_started {
+                    ts_println!(self, "{}\n", "codex".style(self.italic).style(self.magenta));
+                    self.answer_started = true;
+                }
+                print!("{delta}");
+                #[allow(clippy::expect_used)]
+                std::io::stdout().flush().expect("could not flush stdout");
+            }
+            EventMsg::AgentReasoningDelta(AgentReasoningDeltaEvent { delta }) => {
+                if !self.show_agent_reasoning {
+                    return;
+                }
+                if !self.reasoning_started {
+                    ts_println!(
+                        self,
+                        "{}\n",
+                        "thinking".style(self.italic).style(self.magenta),
+                    );
+                    self.reasoning_started = true;
+                }
+                print!("{delta}");
+                #[allow(clippy::expect_used)]
+                std::io::stdout().flush().expect("could not flush stdout");
+            }
+            EventMsg::AgentMessage(AgentMessageEvent { message }) => {
+                // if answer_started is false, this means we haven't received any
+                // delta. Thus, we need to print the message as a new answer.
+                if !self.answer_started {
+                    ts_println!(
+                        self,
+                        "{}\n{}",
+                        "codex".style(self.italic).style(self.magenta),
+                        message,
+                    );
+                } else {
+                    println!();
+                    self.answer_started = false;
+                }
+            }
+            EventMsg::ExecCommandBegin(ExecCommandBeginEvent {
+                call_id,
+                command,
+                cwd,
+            }) => {
+                self.call_id_to_command.insert(
+                    call_id.clone(),
+                    ExecCommandBegin {
+                        command: command.clone(),
+                        start_time: Instant::now(),
+                    },
+                );
+                ts_println!(
+                    self,
+                    "{} {} in {}",
+                    "exec".style(self.magenta),
+                    escape_command(&command).style(self.bold),
+                    cwd.to_string_lossy(),
+                );
+            }
+            EventMsg::ExecCommandEnd(ExecCommandEndEvent {
+                call_id,
+                stdout,
+                stderr,
+                exit_code,
+            }) => {
+                let exec_command = self.call_id_to_command.remove(&call_id);
+                let (duration, call) = if let Some(ExecCommandBegin {
+                    command,
+                    start_time,
+                }) = exec_command
+                {
+                    (
+                        format!(" in {}", format_elapsed(start_time)),
+                        format!("{}", escape_command(&command).style(self.bold)),
+                    )
+                } else {
+                    ("".to_string(), format!("exec('{call_id}')"))
+                };
+
+                let output = if exit_code == 0 { stdout } else { stderr };
+                let truncated_output = output
+                    .lines()
+                    .take(MAX_OUTPUT_LINES_FOR_EXEC_TOOL_CALL)
+                    .collect::<Vec<_>>()
+                    .join("\n");
+                match exit_code {
+                    0 => {
+                        let title = format!("{call} succeeded{duration}:");
+                        ts_println!(self, "{}", title.style(self.green));
+                    }
+                    _ => {
+                        let title = format!("{call} exited {exit_code}{duration}:");
+                        ts_println!(self, "{}", title.style(self.red));
+                    }
+                }
+                println!("{}", truncated_output.style(self.dimmed));
+            }
+            EventMsg::McpToolCallBegin(McpToolCallBeginEvent {
+                call_id,
+                server,
+                tool,
+                arguments,
+            }) => {
+                // Build fully-qualified tool name: server.tool
+                let fq_tool_name = format!("{server}.{tool}");
+
+                // Format arguments as compact JSON so they fit on one line.
+                let args_str = arguments
+                    .as_ref()
+                    .map(|v: &serde_json::Value| {
+                        serde_json::to_string(v).unwrap_or_else(|_| v.to_string())
+                    })
+                    .unwrap_or_default();
+
+                let invocation = if args_str.is_empty() {
+                    format!("{fq_tool_name}()")
+                } else {
+                    format!("{fq_tool_name}({args_str})")
+                };
+
+                self.call_id_to_tool_call.insert(
+                    call_id.clone(),
+                    McpToolCallBegin {
+                        invocation: invocation.clone(),
+                        start_time: Instant::now(),
+                    },
+                );
+
+                ts_println!(
+                    self,
+                    "{} {}",
+                    "tool".style(self.magenta),
+                    invocation.style(self.bold),
+                );
+            }
+            EventMsg::McpToolCallEnd(tool_call_end_event) => {
+                let is_success = tool_call_end_event.is_success();
+                let McpToolCallEndEvent { call_id, result } = tool_call_end_event;
+                // Retrieve start time and invocation for duration calculation and labeling.
+                let info = self.call_id_to_tool_call.remove(&call_id);
+
+                let (duration, invocation) = if let Some(McpToolCallBegin {
+                    invocation,
+                    start_time,
+                    ..
+                }) = info
+                {
+                    (format!(" in {}", format_elapsed(start_time)), invocation)
+                } else {
+                    (String::new(), format!("tool('{call_id}')"))
+                };
+
+                let status_str = if is_success { "success" } else { "failed" };
+                let title_style = if is_success { self.green } else { self.red };
+                let title = format!("{invocation} {status_str}{duration}:");
+
+                ts_println!(self, "{}", title.style(title_style));
+
+                if let Ok(res) = result {
+                    let val: serde_json::Value = res.into();
+                    let pretty =
+                        serde_json::to_string_pretty(&val).unwrap_or_else(|_| val.to_string());
+
+                    for line in pretty.lines().take(MAX_OUTPUT_LINES_FOR_EXEC_TOOL_CALL) {
+                        println!("{}", line.style(self.dimmed));
+                    }
+                }
+            }
+            EventMsg::PatchApplyBegin(PatchApplyBeginEvent {
+                call_id,
+                auto_approved,
+                changes,
+            }) => {
+                // Store metadata so we can calculate duration later when we
+                // receive the corresponding PatchApplyEnd event.
+                self.call_id_to_patch.insert(
+                    call_id.clone(),
+                    PatchApplyBegin {
+                        start_time: Instant::now(),
+                        auto_approved,
+                    },
+                );
+
+                ts_println!(
+                    self,
+                    "{} auto_approved={}:",
+                    "apply_patch".style(self.magenta),
+                    auto_approved,
+                );
+
+                // Pretty-print the patch summary with colored diff markers so
+                // it's easy to scan in the terminal output.
+                for (path, change) in changes.iter() {
+                    match change {
+                        FileChange::Add { content } => {
+                            let header = format!(
+                                "{} {}",
+                                format_file_change(change),
+                                path.to_string_lossy()
+                            );
+                            println!("{}", header.style(self.magenta));
+                            for line in content.lines() {
+                                println!("{}", line.style(self.green));
+                            }
+                        }
+                        FileChange::Delete => {
+                            let header = format!(
+                                "{} {}",
+                                format_file_change(change),
+                                path.to_string_lossy()
+                            );
+                            println!("{}", header.style(self.magenta));
+                        }
+                        FileChange::Update {
+                            unified_diff,
+                            move_path,
+                        } => {
+                            let header = if let Some(dest) = move_path {
+                                format!(
+                                    "{} {} -> {}",
+                                    format_file_change(change),
+                                    path.to_string_lossy(),
+                                    dest.to_string_lossy()
+                                )
+                            } else {
+                                format!("{} {}", format_file_change(change), path.to_string_lossy())
+                            };
+                            println!("{}", header.style(self.magenta));
+
+                            // Colorize diff lines. We keep file header lines
+                            // (--- / +++) without extra coloring so they are
+                            // still readable.
+                            for diff_line in unified_diff.lines() {
+                                if diff_line.starts_with('+') && !diff_line.starts_with("+++") {
+                                    println!("{}", diff_line.style(self.green));
+                                } else if diff_line.starts_with('-')
+                                    && !diff_line.starts_with("---")
+                                {
+                                    println!("{}", diff_line.style(self.red));
+                                } else {
+                                    println!("{diff_line}");
+                                }
+                            }
+                        }
+                    }
+                }
+            }
+            EventMsg::PatchApplyEnd(PatchApplyEndEvent {
+                call_id,
+                stdout,
+                stderr,
+                success,
+            }) => {
+                let patch_begin = self.call_id_to_patch.remove(&call_id);
+
+                // Compute duration and summary label similar to exec commands.
+                let (duration, label) = if let Some(PatchApplyBegin {
+                    start_time,
+                    auto_approved,
+                }) = patch_begin
+                {
+                    (
+                        format!(" in {}", format_elapsed(start_time)),
+                        format!("apply_patch(auto_approved={auto_approved})"),
+                    )
+                } else {
+                    (String::new(), format!("apply_patch('{call_id}')"))
+                };
+
+                let (exit_code, output, title_style) = if success {
+                    (0, stdout, self.green)
+                } else {
+                    (1, stderr, self.red)
+                };
+
+                let title = format!("{label} exited {exit_code}{duration}:");
+                ts_println!(self, "{}", title.style(title_style));
+                for line in output.lines() {
+                    println!("{}", line.style(self.dimmed));
+                }
+            }
+            EventMsg::ExecApprovalRequest(_) => {
+                // Should we exit?
+            }
+            EventMsg::ApplyPatchApprovalRequest(_) => {
+                // Should we exit?
+            }
+            EventMsg::AgentReasoning(agent_reasoning_event) => {
+                if self.show_agent_reasoning {
+                    if !self.reasoning_started {
+                        ts_println!(
+                            self,
+                            "{}\n{}",
+                            "codex".style(self.italic).style(self.magenta),
+                            agent_reasoning_event.text,
+                        );
+                    } else {
+                        println!();
+                        self.reasoning_started = false;
+                    }
+                }
+            }
+            EventMsg::SessionConfigured(session_configured_event) => {
+                let SessionConfiguredEvent {
+                    session_id,
+                    model,
+                    history_log_id: _,
+                    history_entry_count: _,
+                } = session_configured_event;
+
+                ts_println!(
+                    self,
+                    "{} {}",
+                    "codex session".style(self.magenta).style(self.bold),
+                    session_id.to_string().style(self.dimmed)
+                );
+
+                ts_println!(self, "model: {}", model);
+                println!();
+            }
+            EventMsg::GetHistoryEntryResponse(_) => {
+                // Currently ignored in exec output.
+            }
+        }
+    }
+}
+
+fn escape_command(command: &[String]) -> String {
+    try_join(command.iter().map(|s| s.as_str())).unwrap_or_else(|_| command.join(" "))
+}
+
+fn format_file_change(change: &FileChange) -> &'static str {
+    match change {
+        FileChange::Add { .. } => "A",
+        FileChange::Delete => "D",
+        FileChange::Update {
+            move_path: Some(_), ..
+        } => "R",
+        FileChange::Update {
+            move_path: None, ..
+        } => "M",
+    }
+}

codex-rs/exec/src/event_processor_with_json_output.rs

@@ -0,0 +1,48 @@
+use std::collections::HashMap;
+
+use codex_core::config::Config;
+use codex_core::protocol::Event;
+use codex_core::protocol::EventMsg;
+use serde_json::json;
+
+use crate::event_processor::EventProcessor;
+use crate::event_processor::create_config_summary_entries;
+
+pub(crate) struct EventProcessorWithJsonOutput;
+
+impl EventProcessorWithJsonOutput {
+    pub fn new() -> Self {
+        Self {}
+    }
+}
+
+impl EventProcessor for EventProcessorWithJsonOutput {
+    fn print_config_summary(&mut self, config: &Config, prompt: &str) {
+        let entries = create_config_summary_entries(config)
+            .into_iter()
+            .map(|(key, value)| (key.to_string(), value))
+            .collect::<HashMap<String, String>>();
+        #[allow(clippy::expect_used)]
+        let config_json =
+            serde_json::to_string(&entries).expect("Failed to serialize config summary to JSON");
+        println!("{config_json}");
+
+        let prompt_json = json!({
+            "prompt": prompt,
+        });
+        println!("{prompt_json}");
+    }
+
+    fn process_event(&mut self, event: Event) {
+        match event.msg {
+            EventMsg::AgentMessageDelta(_) | EventMsg::AgentReasoningDelta(_) => {
+                // Suppress streaming events in JSON mode.
+            }
+            _ => {
+                if let Ok(line) = serde_json::to_string(&event) {
+                    println!("{line}");
+                }
+            }
+        }
+    }
+}

codex-rs/exec/src/lib.rs

@@ -1,5 +1,7 @@
 mod cli;
 mod event_processor;
+mod event_processor_with_human_output;
+mod event_processor_with_json_output;
 
 use std::io::IsTerminal;
 use std::io::Read;
@@ -19,12 +21,15 @@ use codex_core::protocol::InputItem;
 use codex_core::protocol::Op;
 use codex_core::protocol::TaskCompleteEvent;
 use codex_core::util::is_inside_git_repo;
-use event_processor::EventProcessor;
+use event_processor_with_human_output::EventProcessorWithHumanOutput;
+use event_processor_with_json_output::EventProcessorWithJsonOutput;
 use tracing::debug;
 use tracing::error;
 use tracing::info;
 use tracing_subscriber::EnvFilter;
 
+use crate::event_processor::EventProcessor;
+
 pub async fn run_main(cli: Cli, codex_linux_sandbox_exe: Option<PathBuf>) -> anyhow::Result<()> {
     let Cli {
         images,
@@ -36,6 +41,7 @@ pub async fn run_main(cli: Cli, codex_linux_sandbox_exe: Option<PathBuf>) -> any
         skip_git_repo_check,
         color,
         last_message_file,
+        json: json_mode,
         sandbox_mode: sandbox_mode_cli_arg,
         prompt,
         config_overrides,
@@ -115,8 +121,15 @@ pub async fn run_main(cli: Cli, codex_linux_sandbox_exe: Option<PathBuf>) -> any
     };
 
     let config = Config::load_with_cli_overrides(cli_kv_overrides, overrides)?;
-    let mut event_processor =
-        EventProcessor::create_with_ansi(stdout_with_ansi, !config.hide_agent_reasoning);
+    let mut event_processor: Box<dyn EventProcessor> = if json_mode {
+        Box::new(EventProcessorWithJsonOutput::new())
+    } else {
+        Box::new(EventProcessorWithHumanOutput::create_with_ansi(
+            stdout_with_ansi,
+            &config,
+        ))
+    };
+
     // Print the effective configuration and prompt so users can see what Codex
     // is using.
     event_processor.print_config_summary(&config, &prompt);

codex-rs/justfile

@@ -23,3 +23,10 @@ file-search *args:
 # format code
 fmt:
     cargo fmt -- --config imports_granularity=Item
+
+fix:
+    cargo clippy --fix --all-features --tests --allow-dirty
+
+install:
+    rustup show active-toolchain
+    cargo fetch

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

@@ -171,6 +171,12 @@ pub async fn run_codex_tool_session(
                     EventMsg::SessionConfigured(_) => {
                         tracing::error!("unexpected SessionConfigured event");
                     }
+                    EventMsg::AgentMessageDelta(_) => {
+                        // TODO: think how we want to support this in the MCP
+                    }
+                    EventMsg::AgentReasoningDelta(_) => {
+                        // TODO: think how we want to support this in the MCP
+                    }
                     EventMsg::Error(_)
                     | EventMsg::TaskStarted
                     | EventMsg::TokenCount(_)

codex-rs/rust-toolchain.toml

@@ -0,0 +1,3 @@
+[toolchain]
+channel = "1.88.0"
+components = [ "clippy", "rustfmt", "rust-src"]

codex-rs/tui/Cargo.toml

@@ -40,7 +40,6 @@ ratatui = { version = "0.29.0", features = [
 ] }
 ratatui-image = "8.0.0"
 regex-lite = "0.1"
-reqwest = { version = "0.12", features = ["json"] }
 serde_json = { version = "1", features = ["preserve_order"] }
 shlex = "1.3.0"
 strum = "0.27.1"
@@ -60,7 +59,6 @@ tui-markdown = "0.3.3"
 tui-textarea = "0.7.0"
 unicode-segmentation = "1.12.0"
 uuid = "1"
-time = { version = "0.3", features = ["formatting"] }
 
 [dev-dependencies]
 insta = "1.43.1"

codex-rs/tui/src/app.rs

@@ -18,8 +18,15 @@ use crossterm::event::KeyEvent;
 use crossterm::event::MouseEvent;
 use crossterm::event::MouseEventKind;
 use std::path::PathBuf;
+use std::sync::Arc;
+use std::sync::Mutex;
 use std::sync::mpsc::Receiver;
 use std::sync::mpsc::channel;
+use std::thread;
+use std::time::Duration;
+
+/// Time window for debouncing redraw requests.
+const REDRAW_DEBOUNCE: Duration = Duration::from_millis(10);
 
 /// Top-level application state: which full-screen view is currently active.
 #[allow(clippy::large_enum_variant)]
@@ -46,6 +53,9 @@ pub(crate) struct App<'a> {
 
     file_search: FileSearchManager,
 
+    /// True when a redraw has been scheduled but not yet executed.
+    pending_redraw: Arc<Mutex<bool>>,
+
     /// Stored parameters needed to instantiate the ChatWidget later, e.g.,
     /// after dismissing the Git-repo warning.
     chat_args: Option<ChatWidgetArgs>,
@@ -60,7 +70,7 @@ struct ChatWidgetArgs {
     initial_images: Vec<PathBuf>,
 }
 
-impl<'a> App<'a> {
+impl App<'_> {
     pub(crate) fn new(
         config: Config,
         initial_prompt: Option<String>,
@@ -70,6 +80,7 @@ impl<'a> App<'a> {
     ) -> Self {
         let (app_event_tx, app_event_rx) = channel();
         let app_event_tx = AppEventSender::new(app_event_tx);
+        let pending_redraw = Arc::new(Mutex::new(false));
         let scroll_event_helper = ScrollEventHelper::new(app_event_tx.clone());
 
         // Spawn a dedicated thread for reading the crossterm event loop and
@@ -83,7 +94,7 @@ impl<'a> App<'a> {
                             app_event_tx.send(AppEvent::KeyEvent(key_event));
                         }
                         crossterm::event::Event::Resize(_, _) => {
-                            app_event_tx.send(AppEvent::Redraw);
+                            app_event_tx.send(AppEvent::RequestRedraw);
                         }
                         crossterm::event::Event::Mouse(MouseEvent {
                             kind: MouseEventKind::ScrollUp,
@@ -152,6 +163,7 @@ impl<'a> App<'a> {
             app_state,
             config,
             file_search,
+            pending_redraw,
             chat_args,
         }
     }
@@ -162,6 +174,29 @@ impl<'a> App<'a> {
         self.app_event_tx.clone()
     }
 
+    /// Schedule a redraw if one is not already pending.
+    #[allow(clippy::unwrap_used)]
+    fn schedule_redraw(&self) {
+        {
+            #[allow(clippy::unwrap_used)]
+            let mut flag = self.pending_redraw.lock().unwrap();
+            if *flag {
+                return;
+            }
+            *flag = true;
+        }
+
+        let tx = self.app_event_tx.clone();
+        let pending_redraw = self.pending_redraw.clone();
+        thread::spawn(move || {
+            thread::sleep(REDRAW_DEBOUNCE);
+            tx.send(AppEvent::Redraw);
+            #[allow(clippy::unwrap_used)]
+            let mut f = pending_redraw.lock().unwrap();
+            *f = false;
+        });
+    }
+
     pub(crate) fn run(
         &mut self,
         terminal: &mut tui::Tui,
@@ -169,10 +204,13 @@ impl<'a> App<'a> {
     ) -> Result<()> {
         // Insert an event to trigger the first render.
         let app_event_tx = self.app_event_tx.clone();
-        app_event_tx.send(AppEvent::Redraw);
+        app_event_tx.send(AppEvent::RequestRedraw);
 
         while let Ok(event) = self.app_event_rx.recv() {
             match event {
+                AppEvent::RequestRedraw => {
+                    self.schedule_redraw();
+                }
                 AppEvent::Redraw => {
                     self.draw_next_frame(terminal)?;
                 }
@@ -199,7 +237,21 @@ impl<'a> App<'a> {
                             modifiers: crossterm::event::KeyModifiers::CONTROL,
                             ..
                         } => {
-                            self.app_event_tx.send(AppEvent::ExitRequest);
+                            match &mut self.app_state {
+                                AppState::Chat { widget } => {
+                                    if widget.composer_is_empty() {
+                                        self.app_event_tx.send(AppEvent::ExitRequest);
+                                    } else {
+                                        // Treat Ctrl+D as a normal key event when the composer
+                                        // is not empty so that it doesn't quit the application
+                                        // prematurely.
+                                        self.dispatch_key_event(key_event);
+                                    }
+                                }
+                                AppState::Login { .. } | AppState::GitWarning { .. } => {
+                                    self.app_event_tx.send(AppEvent::ExitRequest);
+                                }
+                            }
                         }
                         _ => {
                             self.dispatch_key_event(key_event);
@@ -235,7 +287,7 @@ impl<'a> App<'a> {
                             Vec::new(),
                         ));
                         self.app_state = AppState::Chat { widget: new_widget };
-                        self.app_event_tx.send(AppEvent::Redraw);
+                        self.app_event_tx.send(AppEvent::RequestRedraw);
                     }
                     SlashCommand::ToggleMouseMode => {
                         if let Err(e) = mouse_capture.toggle() {
@@ -266,11 +318,6 @@ impl<'a> App<'a> {
                             widget.add_diff_output(text);
                         }
                     }
-                    SlashCommand::SecurityReview => {
-                        if let AppState::Chat { widget } = &mut self.app_state {
-                            widget.start_security_review_with_defaults();
-                        }
-                    }
                 },
                 AppEvent::StartFileSearch(query) => {
                     self.file_search.on_user_query(query);
@@ -280,11 +327,6 @@ impl<'a> App<'a> {
                         widget.apply_file_search_result(query, matches);
                     }
                 }
-                AppEvent::SecurityReviewFinished { mode, outcome } => {
-                    if let AppState::Chat { widget } = &mut self.app_state {
-                        widget.handle_security_review_finished(mode, outcome);
-                    }
-                }
             }
         }
         terminal.clear()?;
@@ -293,6 +335,8 @@ impl<'a> App<'a> {
     }
 
     fn draw_next_frame(&mut self, terminal: &mut tui::Tui) -> Result<()> {
+        // TODO: add a throttle to avoid redrawing too often
+
         match &mut self.app_state {
             AppState::Chat { widget } => {
                 terminal.draw(|frame| frame.render_widget_ref(&**widget, frame.area()))?;
@@ -330,7 +374,7 @@ impl<'a> App<'a> {
                         args.initial_images,
                     ));
                     self.app_state = AppState::Chat { widget };
-                    self.app_event_tx.send(AppEvent::Redraw);
+                    self.app_event_tx.send(AppEvent::RequestRedraw);
                 }
                 GitWarningOutcome::Quit => {
                     self.app_event_tx.send(AppEvent::ExitRequest);

codex-rs/tui/src/app_event.rs

@@ -2,13 +2,16 @@ use codex_core::protocol::Event;
 use codex_file_search::FileMatch;
 use crossterm::event::KeyEvent;
 
-use crate::security_review::{SecurityReviewFailure, SecurityReviewMode, SecurityReviewResult};
 use crate::slash_command::SlashCommand;
 
 #[allow(clippy::large_enum_variant)]
 pub(crate) enum AppEvent {
     CodexEvent(Event),
 
+    /// Request a redraw which will be debounced by the [`App`].
+    RequestRedraw,
+
+    /// Actually draw the next frame.
     Redraw,
 
     KeyEvent(KeyEvent),
@@ -46,10 +49,4 @@ pub(crate) enum AppEvent {
         query: String,
         matches: Vec<FileMatch>,
     },
-
-    /// Completion event for a `/secreview` invocation.
-    SecurityReviewFinished {
-        mode: SecurityReviewMode,
-        outcome: Result<SecurityReviewResult, SecurityReviewFailure>,
-    },
 }

codex-rs/tui/src/bottom_pane/chat_composer.rs

@@ -76,6 +76,11 @@ impl ChatComposer<'_> {
         this
     }
 
+    /// Returns true if the composer currently contains no user input.
+    pub(crate) fn is_empty(&self) -> bool {
+        self.textarea.is_empty()
+    }
+
     /// Update the cached *context-left* percentage and refresh the placeholder
     /// text. The UI relies on the placeholder to convey the remaining
     /// context when the composer is empty.

codex-rs/tui/src/bottom_pane/chat_composer_history.rs

@@ -72,8 +72,7 @@ impl ChatComposerHistory {
             return false;
         }
 
-        let lines = textarea.lines();
-        if lines.len() == 1 && lines[0].is_empty() {
+        if textarea.is_empty() {
             return true;
         }
 
@@ -85,6 +84,7 @@ impl ChatComposerHistory {
             return false;
         }
 
+        let lines = textarea.lines();
         matches!(&self.last_history_text, Some(prev) if prev == &lines.join("\n"))
     }
 

codex-rs/tui/src/bottom_pane/mod.rs

@@ -162,6 +162,10 @@ impl BottomPane<'_> {
         }
     }
 
+    pub(crate) fn composer_is_empty(&self) -> bool {
+        self.composer.is_empty()
+    }
+
     pub(crate) fn is_task_running(&self) -> bool {
         self.is_task_running
     }
@@ -208,7 +212,7 @@ impl BottomPane<'_> {
     }
 
     pub(crate) fn request_redraw(&self) {
-        self.app_event_tx.send(AppEvent::Redraw)
+        self.app_event_tx.send(AppEvent::RequestRedraw)
     }
 
     /// Returns true when a popup inside the composer is visible.

codex-rs/tui/src/bottom_pane/status_indicator_view.rs

@@ -24,7 +24,7 @@ impl StatusIndicatorView {
     }
 }
 
-impl<'a> BottomPaneView<'a> for StatusIndicatorView {
+impl BottomPaneView<'_> for StatusIndicatorView {
     fn update_status_text(&mut self, text: String) -> ConditionalUpdate {
         self.update_text(text);
         ConditionalUpdate::NeedsRedraw

codex-rs/tui/src/chatwidget.rs

@@ -1,9 +1,11 @@
-use std::path::{Path, PathBuf};
+use std::path::PathBuf;
 use std::sync::Arc;
 
 use codex_core::codex_wrapper::init_codex;
 use codex_core::config::Config;
+use codex_core::protocol::AgentMessageDeltaEvent;
 use codex_core::protocol::AgentMessageEvent;
+use codex_core::protocol::AgentReasoningDeltaEvent;
 use codex_core::protocol::AgentReasoningEvent;
 use codex_core::protocol::ApplyPatchApprovalRequestEvent;
 use codex_core::protocol::ErrorEvent;
@@ -29,7 +31,6 @@ use ratatui::widgets::Widget;
 use ratatui::widgets::WidgetRef;
 use tokio::sync::mpsc::UnboundedSender;
 use tokio::sync::mpsc::unbounded_channel;
-use tokio::task::JoinHandle;
 
 use crate::app_event::AppEvent;
 use crate::app_event_sender::AppEventSender;
@@ -39,10 +40,7 @@ use crate::bottom_pane::InputResult;
 use crate::conversation_history_widget::ConversationHistoryWidget;
 use crate::history_cell::PatchEventType;
 use crate::user_approval_widget::ApprovalRequest;
-use crate::security_review::{run_security_review, SecurityReviewFailure, SecurityReviewMode, SecurityReviewRequest, SecurityReviewResult};
 use codex_file_search::FileMatch;
-use path_clean::PathClean;
-use shlex;
 
 pub(crate) struct ChatWidget<'a> {
     app_event_tx: AppEventSender,
@@ -53,8 +51,9 @@ pub(crate) struct ChatWidget<'a> {
     config: Config,
     initial_user_message: Option<UserMessage>,
     token_usage: TokenUsage,
-    security_review_handle: Option<JoinHandle<()>>,
-    active_security_review_mode: Option<SecurityReviewMode>,
+    reasoning_buffer: String,
+    answer_buffer: String,
+    active_task_id: Option<String>,
 }
 
 #[derive(Clone, Copy, Eq, PartialEq)]
@@ -85,15 +84,6 @@ fn create_initial_user_message(text: String, image_paths: Vec<PathBuf>) -> Optio
     }
 }
 
-#[derive(Debug, Default)]
-struct ParsedSecReviewCommand {
-    mode: SecurityReviewMode,
-    include_paths: Vec<String>,
-    output_path: Option<String>,
-    repo_path: Option<String>,
-    model_name: Option<String>,
-}
-
 impl ChatWidget<'_> {
     pub(crate) fn new(
         config: Config,
@@ -150,8 +140,9 @@ impl ChatWidget<'_> {
                 initial_images,
             ),
             token_usage: TokenUsage::default(),
-            security_review_handle: None,
-            active_security_review_mode: None,
+            reasoning_buffer: String::new(),
+            answer_buffer: String::new(),
+            active_task_id: None,
         }
     }
 
@@ -199,11 +190,6 @@ impl ChatWidget<'_> {
 
     fn submit_user_message(&mut self, user_message: UserMessage) {
         let UserMessage { text, image_paths } = user_message;
-
-        if self.try_handle_slash_command(&text) {
-            return;
-        }
-
         let mut items: Vec<InputItem> = Vec::new();
 
         if !text.is_empty() {
@@ -238,244 +224,30 @@ impl ChatWidget<'_> {
             self.conversation_history.add_user_message(text);
         }
         self.conversation_history.scroll_to_bottom();
-    }
 
-    fn try_handle_slash_command(&mut self, text: &str) -> bool {
-        let trimmed = text.trim();
-        if trimmed.starts_with("/secreview") {
-            match parse_security_review_command(trimmed) {
-                Ok(command) => {
-                    if let Err(err) = self.launch_security_review(command) {
-                        self.report_security_review_error(err);
-                    }
-                }
-                Err(err) => self.report_security_review_error(err),
-            }
-            return true;
-        }
-        false
-    }
-
-    pub(crate) fn start_security_review_with_defaults(&mut self) {
-        let command = ParsedSecReviewCommand::default();
-        if let Err(err) = self.launch_security_review(command) {
-            self.report_security_review_error(err);
-        }
-    }
-
-    fn launch_security_review(&mut self, command: ParsedSecReviewCommand) -> Result<(), String> {
-        let repo_candidate = if let Some(repo_override) = command.repo_path.as_ref() {
-            let candidate = Path::new(repo_override);
-            if candidate.is_absolute() {
-                candidate.to_path_buf()
-            } else {
-                self.config.cwd.join(candidate)
-            }
-        } else {
-            self.config.cwd.clone()
-        }
-        .clean();
-
-        let repo_path = match repo_candidate.canonicalize() {
-            Ok(path) => path,
-            Err(_) => repo_candidate.clone(),
-        };
-
-        if !repo_path.exists() {
-            return Err(format!(
-                "Repository path '{}' does not exist.",
-                repo_path.display()
-            ));
-        }
-        if !repo_path.is_dir() {
-            return Err(format!(
-                "Repository path '{}' is not a directory.",
-                repo_path.display()
-            ));
-        }
-
-        let mut resolved_paths: Vec<PathBuf> = Vec::new();
-        let mut display_paths: Vec<String> = Vec::new();
-
-        for include in &command.include_paths {
-            let candidate = resolve_path(&repo_path, include);
-            let canonical = match candidate.canonicalize() {
-                Ok(path) => path,
-                Err(_) => candidate.clone(),
-            };
-
-            if !canonical.exists() {
-                return Err(format!("Path '{}' does not exist.", canonical.display()));
-            }
-            if !canonical.starts_with(&repo_path) {
-                return Err(format!(
-                    "Path '{}' is outside the repository root '{}'.",
-                    canonical.display(),
-                    repo_path.display()
-                ));
-            }
-
-            let relative = canonical
-                .strip_prefix(&repo_path)
-                .unwrap_or(&canonical)
-                .display()
-                .to_string();
-            display_paths.push(relative);
-            resolved_paths.push(canonical);
-        }
-
-        let output_root = if let Some(output_override) = command.output_path.as_ref() {
-            let candidate = Path::new(output_override);
-            if candidate.is_absolute() {
-                candidate.to_path_buf()
-            } else {
-                repo_path.join(candidate)
-            }
-        } else {
-            repo_path.join("appsec_review")
-        }
-        .clean();
-
-        let model_name = command
-            .model_name
-            .clone()
-            .unwrap_or_else(|| self.config.model.clone());
-
-        if self.security_review_handle.is_some() {
-            return Err("A security review is already running. Please wait for it to finish or abort it before starting another.".to_string());
-        }
-
-        let scope_description = if resolved_paths.is_empty() {
-            "entire repository".to_string()
-        } else {
-            display_paths.join(", ")
-        };
-
-        let summary = format!(
-            "🔐 Running AppSec security review (mode: {}).\nRepository: {}\nScope: {}\nOutput: {}\nModel: {}",
-            command.mode.as_str(),
-            repo_path.display(),
-            scope_description,
-            output_root.display(),
-            model_name
-        );
-        self.conversation_history.add_background_event(summary);
-        self.conversation_history.scroll_to_bottom();
-        self.bottom_pane.set_task_running(true);
-        self.request_redraw();
-
-        let provider = self.config.model_provider.clone();
-        let request = SecurityReviewRequest {
-            repo_path: repo_path.clone(),
-            include_paths: resolved_paths,
-            output_root: output_root.clone(),
-            mode: command.mode,
-            model: model_name,
-            provider,
-            progress_sender: Some(self.app_event_tx.clone()),
-        };
-
-        let app_event_tx = self.app_event_tx.clone();
-        let mode = command.mode;
-        let handle = tokio::spawn(async move {
-            let outcome = run_security_review(request).await;
-            app_event_tx.send(AppEvent::SecurityReviewFinished { mode, outcome });
-        });
-        self.security_review_handle = Some(handle);
-        self.active_security_review_mode = Some(mode);
-
-        Ok(())
-    }
-
-    fn report_security_review_error(&mut self, message: String) {
-        self.security_review_handle = None;
-        self.active_security_review_mode = None;
-        self.bottom_pane.set_task_running(false);
-        self.conversation_history
-            .add_background_event(format!("❌ {message}"));
-        self.conversation_history.scroll_to_bottom();
-        self.request_redraw();
-    }
-
-    pub(crate) fn handle_security_review_finished(
-        &mut self,
-        mode: SecurityReviewMode,
-        outcome: Result<SecurityReviewResult, SecurityReviewFailure>,
-    ) {
-        self.security_review_handle = None;
-        self.active_security_review_mode = None;
-        self.bottom_pane.set_task_running(false);
-        match outcome {
-            Ok(result) => {
-                let SecurityReviewResult {
-                    bugs_markdown,
-                    report_markdown,
-                    bugs_path,
-                    report_path,
-                    logs,
-                } = result;
-
-                let mut summary = format!(
-                    "✅ AppSec security review complete (mode: {}).\nBugs saved to {}.",
-                    mode.as_str(),
-                    bugs_path.display()
-                );
-                if let Some(report_path) = report_path.as_ref() {
-                    summary.push_str(&format!(
-                        "\nReport saved to {}.",
-                        report_path.display()
-                    ));
-                }
-                self.conversation_history.add_background_event(summary);
-
-                if matches!(mode, SecurityReviewMode::Full) {
-                    if let Some(markdown) = report_markdown.and_then(|m| {
-                        if m.trim().is_empty() {
-                            None
-                        } else {
-                            Some(m)
-                        }
-                    }) {
-                        self.conversation_history.add_agent_message(
-                            &self.config,
-                            format!("# AppSec Security Review Report\n\n{markdown}"),
-                        );
-                    }
-                }
-
-                if !bugs_markdown.trim().is_empty() {
-                    let heading = if matches!(mode, SecurityReviewMode::Full) {
-                        "## Bugs Summary"
-                    } else {
-                        "# AppSec Bugs Summary"
-                    };
-                    self.conversation_history.add_agent_message(
-                        &self.config,
-                        format!("{heading}\n\n{bugs_markdown}"),
-                    );
-                }
-
-                if let Some(log_text) = format_security_review_logs(&logs) {
-                    self.conversation_history
-                        .add_background_event(format!("Logs:\n{log_text}"));
-                }
-            }
-            Err(error) => {
-                let SecurityReviewFailure { message, logs } = error;
-                let mut summary = format!("❌ AppSec security review failed: {message}");
-                if let Some(log_text) = format_security_review_logs(&logs) {
-                    summary.push_str(&format!("\n\nLogs:\n{log_text}"));
-                }
-                self.conversation_history.add_background_event(summary);
-            }
-        }
-
-        self.conversation_history.scroll_to_bottom();
-        self.request_redraw();
+        // IMPORTANT: Starting a *new* user turn. Clear any partially streamed
+        // answer from a previous turn (e.g., one that was interrupted) so that
+        // the next AgentMessageDelta spawns a fresh agent message cell instead
+        // of overwriting the last one.
+        self.answer_buffer.clear();
+        self.reasoning_buffer.clear();
     }
 
     pub(crate) fn handle_codex_event(&mut self, event: Event) {
-        let Event { id, msg } = event;
+        // Retain the event ID so we can refer to it after destructuring.
+        let event_id = event.id.clone();
+        let Event { id: _, msg } = event;
+
+        // When we are in the middle of a task (active_task_id is Some) we drop
+        // streaming text/reasoning events for *other* task IDs. This prevents
+        // late tokens from an interrupted run from bleeding into the current
+        // answer.
+        let should_drop_streaming = self
+            .active_task_id
+            .as_ref()
+            .map(|active| active != &event_id)
+            .unwrap_or(false);
+
         match msg {
             EventMsg::SessionConfigured(event) => {
                 // Record session information at the top of the conversation.
@@ -496,18 +268,69 @@ impl ChatWidget<'_> {
                 self.request_redraw();
             }
             EventMsg::AgentMessage(AgentMessageEvent { message }) => {
+                if should_drop_streaming {
+                    return;
+                }
+                // if the answer buffer is empty, this means we haven't received any
+                // delta. Thus, we need to print the message as a new answer.
+                if self.answer_buffer.is_empty() {
+                    self.conversation_history
+                        .add_agent_message(&self.config, message);
+                } else {
+                    self.conversation_history
+                        .replace_prev_agent_message(&self.config, message);
+                }
+                self.answer_buffer.clear();
+                self.request_redraw();
+            }
+            EventMsg::AgentMessageDelta(AgentMessageDeltaEvent { delta }) => {
+                if should_drop_streaming {
+                    return;
+                }
+                if self.answer_buffer.is_empty() {
+                    self.conversation_history
+                        .add_agent_message(&self.config, "".to_string());
+                }
+                self.answer_buffer.push_str(&delta.clone());
                 self.conversation_history
-                    .add_agent_message(&self.config, message);
+                    .replace_prev_agent_message(&self.config, self.answer_buffer.clone());
+                self.request_redraw();
+            }
+            EventMsg::AgentReasoningDelta(AgentReasoningDeltaEvent { delta }) => {
+                if should_drop_streaming {
+                    return;
+                }
+                if self.reasoning_buffer.is_empty() {
+                    self.conversation_history
+                        .add_agent_reasoning(&self.config, "".to_string());
+                }
+                self.reasoning_buffer.push_str(&delta.clone());
+                self.conversation_history
+                    .replace_prev_agent_reasoning(&self.config, self.reasoning_buffer.clone());
                 self.request_redraw();
             }
             EventMsg::AgentReasoning(AgentReasoningEvent { text }) => {
-                if !self.config.hide_agent_reasoning {
-                    self.conversation_history
-                        .add_agent_reasoning(&self.config, text);
-                    self.request_redraw();
+                if should_drop_streaming {
+                    return;
                 }
+                // if the reasoning buffer is empty, this means we haven't received any
+                // delta. Thus, we need to print the message as a new reasoning.
+                if self.reasoning_buffer.is_empty() {
+                    self.conversation_history
+                        .add_agent_reasoning(&self.config, "".to_string());
+                } else {
+                    // else, we rerender one last time.
+                    self.conversation_history
+                        .replace_prev_agent_reasoning(&self.config, text);
+                }
+                self.reasoning_buffer.clear();
+                self.request_redraw();
             }
             EventMsg::TaskStarted => {
+                // New task has begun – update state and clear any stale buffers.
+                self.active_task_id = Some(event_id);
+                self.answer_buffer.clear();
+                self.reasoning_buffer.clear();
                 self.bottom_pane.clear_ctrl_c_quit_hint();
                 self.bottom_pane.set_task_running(true);
                 self.request_redraw();
@@ -515,6 +338,10 @@ impl ChatWidget<'_> {
             EventMsg::TaskComplete(TaskCompleteEvent {
                 last_agent_message: _,
             }) => {
+                // Task finished; clear active_task_id so that subsequent events are processed.
+                if self.active_task_id.as_ref() == Some(&event_id) {
+                    self.active_task_id = None;
+                }
                 self.bottom_pane.set_task_running(false);
                 self.request_redraw();
             }
@@ -524,16 +351,25 @@ impl ChatWidget<'_> {
                     .set_token_usage(self.token_usage.clone(), self.config.model_context_window);
             }
             EventMsg::Error(ErrorEvent { message }) => {
+                // Error events always get surfaced (even for stale task IDs) so that the user sees
+                // why a run stopped. However, only clear the running indicator if this is the
+                // active task.
+                if self.active_task_id.as_ref() == Some(&event_id) {
+                    self.bottom_pane.set_task_running(false);
+                    self.active_task_id = None;
+                }
                 self.conversation_history.add_error(message);
-                self.bottom_pane.set_task_running(false);
             }
             EventMsg::ExecApprovalRequest(ExecApprovalRequestEvent {
                 command,
                 cwd,
                 reason,
             }) => {
+                if should_drop_streaming {
+                    return;
+                }
                 let request = ApprovalRequest::Exec {
-                    id,
+                    id: event_id,
                     command,
                     cwd,
                     reason,
@@ -545,6 +381,9 @@ impl ChatWidget<'_> {
                 reason,
                 grant_root,
             }) => {
+                if should_drop_streaming {
+                    return;
+                }
                 // ------------------------------------------------------------------
                 // Before we even prompt the user for approval we surface the patch
                 // summary in the main conversation so that the dialog appears in a
@@ -563,7 +402,7 @@ impl ChatWidget<'_> {
 
                 // Now surface the approval request in the BottomPane as before.
                 let request = ApprovalRequest::ApplyPatch {
-                    id,
+                    id: event_id,
                     reason,
                     grant_root,
                 };
@@ -575,6 +414,9 @@ impl ChatWidget<'_> {
                 command,
                 cwd: _,
             }) => {
+                if should_drop_streaming {
+                    return;
+                }
                 self.conversation_history
                     .add_active_exec_command(call_id, command);
                 self.request_redraw();
@@ -584,6 +426,9 @@ impl ChatWidget<'_> {
                 auto_approved,
                 changes,
             }) => {
+                if should_drop_streaming {
+                    return;
+                }
                 // Even when a patch is auto‑approved we still display the
                 // summary so the user can follow along.
                 self.conversation_history
@@ -599,6 +444,9 @@ impl ChatWidget<'_> {
                 stdout,
                 stderr,
             }) => {
+                if should_drop_streaming {
+                    return;
+                }
                 self.conversation_history
                     .record_completed_exec_command(call_id, stdout, stderr, exit_code);
                 self.request_redraw();
@@ -609,11 +457,17 @@ impl ChatWidget<'_> {
                 tool,
                 arguments,
             }) => {
+                if should_drop_streaming {
+                    return;
+                }
                 self.conversation_history
                     .add_active_mcp_tool_call(call_id, server, tool, arguments);
                 self.request_redraw();
             }
             EventMsg::McpToolCallEnd(mcp_tool_call_end_event) => {
+                if should_drop_streaming {
+                    return;
+                }
                 let success = mcp_tool_call_end_event.is_success();
                 let McpToolCallEndEvent { call_id, result } = mcp_tool_call_end_event;
                 self.conversation_history
@@ -646,7 +500,7 @@ impl ChatWidget<'_> {
     }
 
     fn request_redraw(&mut self) {
-        self.app_event_tx.send(AppEvent::Redraw);
+        self.app_event_tx.send(AppEvent::RequestRedraw);
     }
 
     pub(crate) fn add_diff_output(&mut self, diff_output: String) {
@@ -676,20 +530,6 @@ impl ChatWidget<'_> {
     /// Returns true if the key press was handled, false if it was not.
     /// If the key press was not handled, the caller should handle it (likely by exiting the process).
     pub(crate) fn on_ctrl_c(&mut self) -> bool {
-        if let Some(handle) = self.security_review_handle.take() {
-            handle.abort();
-            let mode = self
-                .active_security_review_mode
-                .take()
-                .unwrap_or_else(SecurityReviewMode::default);
-            let failure = SecurityReviewFailure {
-                message: "AppSec security review aborted by user.".to_string(),
-                logs: vec!["AppSec security review aborted by user.".to_string()],
-            };
-            self.handle_security_review_finished(mode, Err(failure));
-            return true;
-        }
-
         if self.bottom_pane.is_task_running() {
             self.bottom_pane.clear_ctrl_c_quit_hint();
             self.submit_op(Op::Interrupt);
@@ -702,6 +542,10 @@ impl ChatWidget<'_> {
         }
     }
 
+    pub(crate) fn composer_is_empty(&self) -> bool {
+        self.bottom_pane.composer_is_empty()
+    }
+
     /// Forward an `Op` directly to codex.
     pub(crate) fn submit_op(&self, op: Op) {
         if let Err(e) = self.codex_op_tx.send(op) {
@@ -751,137 +595,3 @@ fn add_token_usage(current_usage: &TokenUsage, new_usage: &TokenUsage) -> TokenU
         total_tokens: current_usage.total_tokens + new_usage.total_tokens,
     }
 }
-
-fn parse_security_review_command(input: &str) -> Result<ParsedSecReviewCommand, String> {
-    let tokens = shlex::split(input).ok_or_else(|| "Unable to parse command arguments.".to_string())?;
-    if tokens.is_empty() {
-        return Err("Empty command.".to_string());
-    }
-
-    if tokens[0] != "/secreview" {
-        return Err("Unrecognized command.".to_string());
-    }
-
-    let mut command = ParsedSecReviewCommand::default();
-    let mut idx = 1;
-    while idx < tokens.len() {
-        let token = &tokens[idx];
-
-        if token == "--" {
-            for extra in tokens.iter().skip(idx + 1) {
-                if !extra.is_empty() {
-                    command.include_paths.push(extra.to_string());
-                }
-            }
-            break;
-        } else if matches!(
-            token.as_str(),
-            "bugs" | "--bugs" | "--mode=bugs"
-        ) {
-            command.mode = SecurityReviewMode::Bugs;
-        } else if matches!(
-            token.as_str(),
-            "full" | "--full" | "--mode=full"
-        ) {
-            command.mode = SecurityReviewMode::Full;
-        } else if token == "--mode" {
-            idx += 1;
-            if idx >= tokens.len() {
-                return Err("Expected value after --mode.".to_string());
-            }
-            command.mode = parse_mode(&tokens[idx])?;
-        } else if let Some(value) = token.strip_prefix("--mode=") {
-            command.mode = parse_mode(value)?;
-        } else if token == "--path" || token == "-p" {
-            idx += 1;
-            if idx >= tokens.len() {
-                return Err(format!("Expected value after {token}."));
-            }
-            command.include_paths.push(tokens[idx].clone());
-        } else if let Some(value) = token.strip_prefix("--path=") {
-            command.include_paths.push(value.to_string());
-        } else if let Some(value) = token.strip_prefix("-p=") {
-            command.include_paths.push(value.to_string());
-        } else if matches!(
-            token.as_str(),
-            "--output" | "-o" | "--output-location"
-        ) {
-            idx += 1;
-            if idx >= tokens.len() {
-                return Err(format!("Expected value after {token}."));
-            }
-            command.output_path = Some(tokens[idx].clone());
-        } else if let Some(value) = token.strip_prefix("--output=") {
-            command.output_path = Some(value.to_string());
-        } else if let Some(value) = token.strip_prefix("-o=") {
-            command.output_path = Some(value.to_string());
-        } else if matches!(
-            token.as_str(),
-            "--repo" | "--repo-location" | "--repository"
-        ) {
-            idx += 1;
-            if idx >= tokens.len() {
-                return Err(format!("Expected value after {token}."));
-            }
-            command.repo_path = Some(tokens[idx].clone());
-        } else if let Some(value) = token.strip_prefix("--repo=") {
-            command.repo_path = Some(value.to_string());
-        } else if let Some(value) = token.strip_prefix("--repo-location=") {
-            command.repo_path = Some(value.to_string());
-        } else if token == "--model" || token == "--model-name" {
-            idx += 1;
-            if idx >= tokens.len() {
-                return Err(format!("Expected value after {token}."));
-            }
-            command.model_name = Some(tokens[idx].clone());
-        } else if let Some(value) = token.strip_prefix("--model=") {
-            command.model_name = Some(value.to_string());
-        } else if let Some(value) = token.strip_prefix("--model-name=") {
-            command.model_name = Some(value.to_string());
-        } else if !token.is_empty() {
-            command.include_paths.push(token.clone());
-        }
-
-        idx += 1;
-    }
-
-    Ok(command)
-}
-
-fn parse_mode(value: &str) -> Result<SecurityReviewMode, String> {
-    match value.to_ascii_lowercase().as_str() {
-        "full" => Ok(SecurityReviewMode::Full),
-        "bugs" | "bugs-only" | "bugsonly" => Ok(SecurityReviewMode::Bugs),
-        other => Err(format!("Unknown mode '{other}'. Use 'full' or 'bugs'.")),
-    }
-}
-
-fn resolve_path(base: &Path, candidate: &str) -> PathBuf {
-    let path = Path::new(candidate);
-    if path.is_absolute() {
-        path.to_path_buf()
-    } else {
-        base.join(path)
-    }
-    .clean()
-}
-
-fn format_security_review_logs(logs: &[String]) -> Option<String> {
-    if logs.is_empty() {
-        return None;
-    }
-
-    let joined = logs.join("\n");
-    if joined.trim().is_empty() {
-        return None;
-    }
-
-    let lines: Vec<&str> = joined.lines().collect();
-    const MAX_LINES: usize = 40;
-    if lines.len() <= MAX_LINES {
-        Some(joined)
-    } else {
-        let tail = lines[lines.len().saturating_sub(MAX_LINES)..].join("\n");
-        Some(format!("… (showing last {MAX_LINES} lines)\n{tail}"))
-    }
-}

codex-rs/tui/src/conversation_history_widget.rs

@@ -202,6 +202,14 @@ impl ConversationHistoryWidget {
         self.add_to_history(HistoryCell::new_agent_reasoning(config, text));
     }
 
+    pub fn replace_prev_agent_reasoning(&mut self, config: &Config, text: String) {
+        self.replace_last_agent_reasoning(config, text);
+    }
+
+    pub fn replace_prev_agent_message(&mut self, config: &Config, text: String) {
+        self.replace_last_agent_message(config, text);
+    }
+
     pub fn add_background_event(&mut self, message: String) {
         self.add_to_history(HistoryCell::new_background_event(message));
     }
@@ -249,6 +257,42 @@ impl ConversationHistoryWidget {
         });
     }
 
+    pub fn replace_last_agent_reasoning(&mut self, config: &Config, text: String) {
+        if let Some(idx) = self
+            .entries
+            .iter()
+            .rposition(|entry| matches!(entry.cell, HistoryCell::AgentReasoning { .. }))
+        {
+            let width = self.cached_width.get();
+            let entry = &mut self.entries[idx];
+            entry.cell = HistoryCell::new_agent_reasoning(config, text);
+            let height = if width > 0 {
+                entry.cell.height(width)
+            } else {
+                0
+            };
+            entry.line_count.set(height);
+        }
+    }
+
+    pub fn replace_last_agent_message(&mut self, config: &Config, text: String) {
+        if let Some(idx) = self
+            .entries
+            .iter()
+            .rposition(|entry| matches!(entry.cell, HistoryCell::AgentMessage { .. }))
+        {
+            let width = self.cached_width.get();
+            let entry = &mut self.entries[idx];
+            entry.cell = HistoryCell::new_agent_message(config, text);
+            let height = if width > 0 {
+                entry.cell.height(width)
+            } else {
+                0
+            };
+            entry.line_count.set(height);
+        }
+    }
+
     pub fn record_completed_exec_command(
         &mut self,
         call_id: String,
@@ -454,7 +498,7 @@ impl WidgetRef for ConversationHistoryWidget {
 
         {
             // Choose a thumb color that stands out only when this pane has focus so that the
-            // user’s attention is naturally drawn to the active viewport. When unfocused we show
+            // user's attention is naturally drawn to the active viewport. When unfocused we show
             // a low-contrast thumb so the scrollbar fades into the background without becoming
             // invisible.
             let thumb_style = if self.has_input_focus {

codex-rs/tui/src/lib.rs

@@ -20,7 +20,6 @@ use tracing_subscriber::EnvFilter;
 use tracing_subscriber::prelude::*;
 
 mod app;
-mod security_review;
 mod app_event;
 mod app_event_sender;
 mod bottom_pane;

codex-rs/tui/src/slash_command.rs

@@ -14,8 +14,6 @@ pub enum SlashCommand {
     // more frequently used commands should be listed first.
     New,
     Diff,
-    #[strum(serialize = "secreview")]
-    SecurityReview,
     Quit,
     ToggleMouseMode,
 }
@@ -32,9 +30,6 @@ impl SlashCommand {
             SlashCommand::Diff => {
                 "Show git diff of the working directory (including untracked files)"
             }
-            SlashCommand::SecurityReview => {
-                "Run AppSec security review and display the generated reports"
-            }
         }
     }
 

codex-rs/tui/src/status_indicator_widget.rs

@@ -65,7 +65,7 @@ impl StatusIndicatorWidget {
                     std::thread::sleep(Duration::from_millis(200));
                     counter = counter.wrapping_add(1);
                     frame_idx_clone.store(counter, Ordering::Relaxed);
-                    app_event_tx_clone.send(AppEvent::Redraw);
+                    app_event_tx_clone.send(AppEvent::RequestRedraw);
                 }
             });
         }