commandmate 0.32.0 → 0.33.0

package/dist/server/src/lib/detection/cli-patterns.js

@@ -4,13 +4,15 @@
  * Shared between response-poller.ts and API routes
  */
 Object.defineProperty(exports, "__esModule", { value: true });
-exports.COPILOT_TOOL_ROW_PATTERN = exports.COPILOT_TOOL_VERBS = exports.COPILOT_TRANSCRIPT_DIVIDER_PATTERN = exports.COPILOT_USER_ECHO_PATTERN = exports.COPILOT_REASONING_HEADER_PATTERN = exports.COPILOT_BOX_ROW_PATTERN = exports.COPILOT_SEPARATOR_PATTERN = exports.COPILOT_IDLE_STATUS_PATTERN = exports.COPILOT_WORKING_STATUS_PATTERN = exports.COPILOT_THINKING_PATTERN = exports.COPILOT_PROMPT_PATTERN = exports.OPENCODE_SKIP_PATTERNS = exports.OPENCODE_SEPARATOR_PATTERN = exports.CLAUDE_SELECTION_LIST_FOOTER = exports.CLAUDE_MODEL_OVERLAY_FOOTER_PATTERN = exports.OPENCODE_SELECTION_LIST_PATTERN = exports.OPENCODE_FOOTER_STATUS_PATTERN = exports.OPENCODE_USER_ECHO_PATTERN = exports.OPENCODE_GUTTER_ROW_PATTERN = exports.OPENCODE_COMPOSER_BOTTOM_BORDER = exports.OPENCODE_PROCESSING_INDICATOR = exports.OPENCODE_PERMISSION_PATTERN = exports.OPENCODE_TURN_COMPLETE_PATTERN = exports.OPENCODE_RESPONSE_COMPLETE = exports.OPENCODE_LOADING_PATTERN = exports.OPENCODE_THINKING_PATTERN = exports.OPENCODE_PROMPT_AFTER_RESPONSE = exports.OPENCODE_IDLE_COMPOSER_PATTERN = exports.OPENCODE_PROMPT_PATTERN = exports.GEMINI_THINKING_PATTERN = exports.GEMINI_PROMPT_PATTERN = exports.MAX_PASTED_TEXT_RETRIES = exports.PASTED_TEXT_DETECT_DELAY = exports.PASTED_TEXT_PATTERN = exports.CODEX_STATUS_BAR_PATTERN = exports.CODEX_PAGER_FOOTER_PATTERN = exports.CODEX_APPROVAL_FOOTER_PATTERN = exports.CODEX_SELECTION_LIST_PATTERN = exports.CODEX_SEPARATOR_PATTERN = exports.CODEX_HOOKS_REVIEW_ANCHORS = exports.CODEX_DIALOG_PATTERN = exports.CODEX_PROMPT_PATTERN = exports.CLAUDE_TRUST_DIALOG_PATTERN = exports.CLAUDE_SEPARATOR_PATTERN = exports.CLAUDE_PROMPT_PATTERN = exports.CODEX_INTERRUPT_HINT_PATTERN = exports.CODEX_THINKING_PATTERN = exports.CLAUDE_INTERRUPT_HINT_PATTERN = exports.CLAUDE_THINKING_PATTERN = exports.CLAUDE_SPINNER_CHARS = void 0;
-exports.CLAUDE_SESSION_ERROR_REGEX_PATTERNS = exports.CLAUDE_SESSION_ERROR_PATTERNS = exports.extractAnsiSequences = exports.stripAnsi = exports.COMMAND_CODE_SKIP_PATTERNS = exports.COMMAND_CODE_HOOK_NOTICE_PATTERN = exports.COMMAND_CODE_BANNER_PATTERNS = exports.COMMAND_CODE_MODE_INDICATOR_PATTERN = exports.COMMAND_CODE_COMPLETION_PATTERN = exports.COMMAND_CODE_RESPONSE_MARKER_PATTERN = exports.COMMAND_CODE_THINKING_PATTERN = exports.COMMAND_CODE_INTERRUPT_HINT_PATTERN = exports.COMMAND_CODE_SEPARATOR_PATTERN = exports.COMMAND_CODE_PROMPT_PATTERN = exports.COMMAND_CODE_SPINNER_CHARS = exports.ANTIGRAVITY_SKIP_PATTERNS = exports.ANTIGRAVITY_SURVEY_PATTERN = exports.ANTIGRAVITY_DIALOG_MAX_ROWS = exports.ANTIGRAVITY_DIALOG_BOUNDARY_PATTERN = exports.ANTIGRAVITY_SWITCH_MODEL_HEADER_PATTERN = exports.ANTIGRAVITY_NUMBERED_OPTION_PATTERN = exports.ANTIGRAVITY_NAVIGATE_FOOTER_PATTERN = exports.ANTIGRAVITY_SELECTION_LIST_PATTERN = exports.ANTIGRAVITY_SEPARATOR_PATTERN = exports.ANTIGRAVITY_THINKING_PATTERN = exports.ANTIGRAVITY_PROMPT_PATTERN = exports.VIBE_LOCAL_THINKING_PATTERN = exports.VIBE_LOCAL_PROMPT_PATTERN = exports.COPILOT_SKIP_PATTERNS = exports.COPILOT_FOLDER_TRUST_ANSWER_KEY = exports.COPILOT_FOLDER_TRUST_SESSION_OPTION_PATTERN = exports.COPILOT_FOLDER_TRUST_ANCHORS = exports.COPILOT_SELECTION_FOOTER_PATTERN = exports.COPILOT_BOOT_BANNER_ANCHORS = exports.COPILOT_TRANSCRIPT_CONTINUATION_PATTERN = void 0;
+exports.COPILOT_TOOL_VERBS = exports.COPILOT_TRANSCRIPT_DIVIDER_PATTERN = exports.COPILOT_USER_ECHO_PATTERN = exports.COPILOT_REASONING_HEADER_PATTERN = exports.COPILOT_BOX_ROW_PATTERN = exports.COPILOT_SEPARATOR_PATTERN = exports.COPILOT_IDLE_STATUS_PATTERN = exports.COPILOT_WORKING_STATUS_PATTERN = exports.COPILOT_THINKING_PATTERN = exports.COPILOT_PROMPT_PATTERN = exports.OPENCODE_SKIP_PATTERNS = exports.OPENCODE_SEPARATOR_PATTERN = exports.CLAUDE_SELECTION_LIST_FOOTER = exports.CLAUDE_MODEL_OVERLAY_FOOTER_PATTERN = exports.OPENCODE_SELECTION_LIST_PATTERN = exports.OPENCODE_FOOTER_STATUS_PATTERN = exports.OPENCODE_USER_ECHO_PATTERN = exports.OPENCODE_GUTTER_ROW_PATTERN = exports.OPENCODE_COMPOSER_BOTTOM_BORDER = exports.OPENCODE_PROCESSING_INDICATOR = exports.OPENCODE_PERMISSION_PATTERN = exports.OPENCODE_TURN_COMPLETE_PATTERN = exports.OPENCODE_RESPONSE_COMPLETE = exports.OPENCODE_LOADING_PATTERN = exports.OPENCODE_THINKING_PATTERN = exports.OPENCODE_PROMPT_AFTER_RESPONSE = exports.OPENCODE_IDLE_COMPOSER_PATTERN = exports.OPENCODE_PROMPT_PATTERN = exports.GEMINI_THINKING_PATTERN = exports.GEMINI_PROMPT_PATTERN = exports.MAX_PASTED_TEXT_RETRIES = exports.PASTED_TEXT_DETECT_DELAY = exports.PASTED_TEXT_PATTERN = exports.CODEX_USER_ECHO_PATTERN = exports.CODEX_STATUS_BAR_PATTERN = exports.CODEX_PAGER_FOOTER_PATTERN = exports.CODEX_APPROVAL_FOOTER_PATTERN = exports.CODEX_SELECTION_LIST_PATTERN = exports.CODEX_SEPARATOR_PATTERN = exports.CODEX_HOOKS_REVIEW_ANCHORS = exports.CODEX_DIALOG_PATTERN = exports.CODEX_PROMPT_PATTERN = exports.CLAUDE_TRUST_DIALOG_PATTERN = exports.CLAUDE_SEPARATOR_PATTERN = exports.CLAUDE_PROMPT_PATTERN = exports.CODEX_INTERRUPT_HINT_PATTERN = exports.CODEX_THINKING_PATTERN = exports.CLAUDE_INTERRUPT_HINT_PATTERN = exports.CLAUDE_THINKING_PATTERN = exports.CLAUDE_SPINNER_CHARS = void 0;
+exports.CLAUDE_SESSION_ERROR_REGEX_PATTERNS = exports.CLAUDE_SESSION_ERROR_PATTERNS = exports.extractAnsiSequences = exports.stripAnsi = exports.COMMAND_CODE_SKIP_PATTERNS = exports.COMMAND_CODE_HOOK_NOTICE_PATTERN = exports.COMMAND_CODE_BANNER_PATTERNS = exports.COMMAND_CODE_MODE_INDICATOR_PATTERN = exports.COMMAND_CODE_COMPLETION_PATTERN = exports.COMMAND_CODE_RESPONSE_MARKER_PATTERN = exports.COMMAND_CODE_THINKING_PATTERN = exports.COMMAND_CODE_INTERRUPT_HINT_PATTERN = exports.COMMAND_CODE_SEPARATOR_PATTERN = exports.COMMAND_CODE_PROMPT_PATTERN = exports.COMMAND_CODE_SPINNER_CHARS = exports.ANTIGRAVITY_SKIP_PATTERNS = exports.ANTIGRAVITY_SURVEY_PATTERN = exports.ANTIGRAVITY_DIALOG_MAX_ROWS = exports.ANTIGRAVITY_DIALOG_BOUNDARY_PATTERN = exports.ANTIGRAVITY_SWITCH_MODEL_HEADER_PATTERN = exports.ANTIGRAVITY_NUMBERED_OPTION_PATTERN = exports.ANTIGRAVITY_NAVIGATE_FOOTER_PATTERN = exports.ANTIGRAVITY_SELECTION_LIST_PATTERN = exports.ANTIGRAVITY_SEPARATOR_PATTERN = exports.ANTIGRAVITY_THINKING_PATTERN = exports.ANTIGRAVITY_PROMPT_PATTERN = exports.VIBE_LOCAL_THINKING_PATTERN = exports.VIBE_LOCAL_PROMPT_PATTERN = exports.COPILOT_SKIP_PATTERNS = exports.COPILOT_FOLDER_TRUST_ANSWER_KEY = exports.COPILOT_FOLDER_TRUST_SESSION_OPTION_PATTERN = exports.COPILOT_FOLDER_TRUST_ANCHORS = exports.COPILOT_SELECTION_FOOTER_PATTERN = exports.COPILOT_BOOT_BANNER_ANCHORS = exports.COPILOT_TRANSCRIPT_CONTINUATION_PATTERN = exports.COPILOT_TOOL_ROW_PATTERN = void 0;
 exports.isCodexTurnActive = isCodexTurnActive;
 exports.findClaudeChromeStart = findClaudeChromeStart;
 exports.isCodexPromptReady = isCodexPromptReady;
 exports.getCodexActiveDialog = getCodexActiveDialog;
 exports.getCodexLifecycleDialog = getCodexLifecycleDialog;
+exports.findCodexChromeStart = findCodexChromeStart;
+exports.findCodexUserEchoIndex = findCodexUserEchoIndex;
 exports.findOpenCodeChromeStart = findOpenCodeChromeStart;
 exports.findOpenCodeUserEchoEnd = findOpenCodeUserEchoEnd;
 exports.findOpenCodeComposerRows = findOpenCodeComposerRows;
@@ -29,6 +31,7 @@ exports.buildDetectPromptOptions = buildDetectPromptOptions;
 const logger_1 = require("../../lib/logger");
 const ansi_1 = require("./ansi");
 const composer_text_1 = require("./composer-text");
+const cli_patterns_1 = require("./tools/codex/cli-patterns");
 const thinking_constants_1 = require("../../config/thinking-constants");
 const logger = (0, logger_1.createLogger)('cli-patterns');
 /**
@@ -589,6 +592,207 @@ exports.CODEX_PAGER_FOOTER_PATTERN = /(?:↑\/↓|pgup\/pgdn|home\/end)\s+to\s+(
  * line. No nested quantifiers (ReDoS-safe; adjacent greedy quantifiers only).
  */
 exports.CODEX_STATUS_BAR_PATTERN = /^\s*\S.*·\s*~?\/\S*\s*$/;
+/**
+ * How far above the last non-blank row {@link findCodexChromeStart} looks for the
+ * composer.
+ *
+ * Same allowance and the same reason as {@link CODEX_COMPOSER_SEARCH_ROWS}: codex
+ * pins the composer two to three rows above the bottom in every measured frame
+ * (`tests/fixtures/codex-live-2310/`), and a wider band would let the search walk
+ * into the transcript and mistake the echoed user message — drawn with the same
+ * `›` — for the input box.
+ *
+ * Wider than the 8 rows of the liveness search because 0.15x can draw notices
+ * BELOW the composer (`N background terminal running · /ps to view · /stop to
+ * close`) that the liveness search never had to step over.
+ */
+const CODEX_CHROME_SEARCH_ROWS = 12;
+/**
+ * A row whose first character is codex's `›` glyph, with something after it.
+ *
+ * Deliberately not {@link CODEX_PROMPT_PATTERN}: that one is multiline and
+ * matches a bare `›`, which is right for "is a prompt on screen anywhere?" and
+ * wrong for classifying ONE row. Anchored at column 0 because all three of
+ * codex's `›` uses are, and an indented `›` in a reply is quoted text.
+ */
+const CODEX_CHROME_GLYPH_ROW_PATTERN = /^›(\s|$)/;
+/**
+ * Locate the start of codex's bottom-pinned chrome within a captured pane.
+ *
+ * The fifth reader of this shape, after {@link findClaudeChromeStart} (#1289),
+ * {@link findCopilotChromeStart} (#1897), {@link findOpenCodeChromeStart}
+ * (#1911) and {@link findCommandCodeChromeStart} (#2250). codex is the tool that
+ * never got one, and Issue #2400 is the bill for that.
+ *
+ * codex renders inline and pins two rows to the bottom of a settled pane:
+ *
+ * ```text
+ * › Ask Codex to do anything                      ← composer (placeholder or typed text)
+ *
+ *   gpt-6-astra xhigh · ~/share/work/…/CommandMate ← status bar (model · cwd)
+ * ```
+ *
+ * Below the composer codex may also draw its own notices (`N background terminal
+ * running · /ps to view`), so the boundary is "the composer row" rather than a
+ * list of footer shapes: everything from the composer down is chrome by
+ * construction, whatever codex adds there next.
+ *
+ * ## What went wrong without it (#2400)
+ *
+ * While the capture window is NOT saturated codex's extraction starts at
+ * `lastCapturedLine`, so the composer only ever mattered as the `endIndex` break
+ * — which the extraction loop already had. Once the pane outgrows
+ * `CACHE_MAX_CAPTURE_LINES` (#1670) the cursor stops being a position in the
+ * capture and `resolveExtractionStartIndex` switches to the newest echoed user
+ * prompt. With no `contentEnd`, that backwards search started at the very bottom
+ * of the pane and the first `›` it met was the COMPOSER. Extraction then began
+ * on the row after it, i.e. on the status bar, and the saved "reply" for every
+ * turn on a saturated pane was one row:
+ *
+ * ```text
+ * gpt-6-astra xhigh · ~/share/work/github_kewton/CommandAgent-develop
+ * ```
+ *
+ * Identical on every turn, so `isDuplicateResponse` then locked on it and the
+ * pane could not record another reply at all. That is #1289's defect verbatim,
+ * one tool later — the same reason `findCommandCodeChromeStart` exists.
+ *
+ * ## Why the attributes and not the placeholder text
+ *
+ * The codex branch of `findRecentUserPromptIndex` used to exclude the composer by
+ * naming its placeholders (`Implement`, `Find and fix`, `Type`, `Summarize`).
+ * Those are codex 0.1x wording, and 0.15x draws `Ask Codex to do anything`, so
+ * the list silently stopped matching the thing it was written for. Issue #2310
+ * measured what actually separates codex's three uses of `›` (U+203A), and it is
+ * the SGR attributes, not the text: the composer glyph is bold (`ESC[1m›`), a
+ * transcript echo is dim (`ESC[1;2m›`), a dialog option carries a coloured glyph
+ * or a bold label. {@link readCodexGlyphRowKind} is that measurement, and this
+ * reader is one of its callers.
+ *
+ * `-1` is returned for a frame whose bottom-most `›` is an option row: codex
+ * replaces the composer with the dialog, so there is no chrome to trim and the
+ * caller resolves the frame on the prompt path instead.
+ *
+ * ## The stripped-capture fallback
+ *
+ * Auto-Yes hands the detection layer a capture that has already been through
+ * `stripAnsi`, and there every `›` is the same byte — {@link
+ * readCodexGlyphRowKind} answers `null` on purpose rather than guessing. This
+ * reader still has to answer for those frames, so it falls back to the one
+ * structural landmark codex pins BELOW the composer and nowhere else: the status
+ * bar ({@link CODEX_STATUS_BAR_PATTERN}, `model · /path`). Requiring it means a
+ * frame with no bar — a pane mid-redraw, an overlay — yields `-1` and the
+ * pre-#2400 reading, which is the direction that costs nothing.
+ *
+ * @param lines - Captured pane lines, ANSI-bearing or not; trailing blanks tolerated
+ * @returns Index of the composer row, or -1 when no composer chrome is present
+ */
+function findCodexChromeStart(lines) {
+    let lastRow = lines.length - 1;
+    while (lastRow >= 0 && (0, ansi_1.stripAnsi)(lines[lastRow]).trim() === '')
+        lastRow--;
+    if (lastRow < 0)
+        return -1;
+    // Bottom-most `›` row within the band codex reserves for its chrome. Anything
+    // further up is transcript, and latching onto an echo there would cut the
+    // reply this whole reader exists to keep.
+    let glyphRow = -1;
+    for (let i = lastRow; i >= Math.max(0, lastRow - CODEX_CHROME_SEARCH_ROWS); i--) {
+        if (CODEX_CHROME_GLYPH_ROW_PATTERN.test((0, ansi_1.stripAnsi)(lines[i]))) {
+            glyphRow = i;
+            break;
+        }
+    }
+    if (glyphRow < 0)
+        return -1;
+    const kind = (0, cli_patterns_1.readCodexGlyphRowKind)(lines[glyphRow]);
+    if (kind === 'composer')
+        return glyphRow;
+    // A dialog is up (no composer drawn) or the bottom-most `›` is a transcript
+    // echo mid-redraw. Neither is chrome to trim.
+    if (kind !== null)
+        return -1;
+    // Stripped capture: no attributes to read. Accept the row as the composer only
+    // when codex's status bar is drawn below it, which is where it always sits and
+    // where a transcript echo can never be.
+    for (let i = glyphRow + 1; i <= lastRow; i++) {
+        if (exports.CODEX_STATUS_BAR_PATTERN.test((0, ansi_1.stripAnsi)(lines[i])))
+            return glyphRow;
+    }
+    return -1;
+}
+/**
+ * The shape of a codex row that could be the echo of a message the user sent.
+ *
+ * Shape only — `›` at column 0 with text after it — which all three of codex's
+ * `›` uses share. {@link findCodexUserEchoIndex} is what tells them apart.
+ */
+exports.CODEX_USER_ECHO_PATTERN = /^›\s+\S/;
+/**
+ * Find the newest transcript echo of an operator message in a codex capture.
+ *
+ * The reader `findRecentUserPromptIndex` anchors codex turns on (Issue #2400).
+ * It replaces a negative lookahead over composer placeholder strings
+ * (`(?!Implement|Find and fix|Type|Summarize)`) written against codex 0.1x:
+ * 0.15x draws `Ask Codex to do anything`, so the guard matched nothing it was
+ * written for. The composer became the newest "echo", and on a saturated pane —
+ * the one path where this anchor decides where extraction STARTS (#1670) — the
+ * reply saved for every turn was the single status-bar row below it.
+ *
+ * ## What the attributes can and cannot separate
+ *
+ * #2310 measured the three uses of `›` and this reader adds the fourth reading
+ * they left open, captured for #2400 on codex-cli 0.153.4
+ * (`tests/fixtures/codex-live-2310/turn-submitted-no-status.txt`): the echo of
+ * a message the operator has JUST submitted is drawn `ESC[1m› ESC[0m<text>` —
+ * bold glyph, plain label — and only becomes the dim `ESC[1;2m› ` of the
+ * measured frames once the turn settles. That shape is indistinguishable from a
+ * composer holding typed text, so no per-row attribute rule can separate them.
+ *
+ * What separates them is position, which codex's layout fixes: the composer is
+ * the BOTTOM-MOST `›` row of a frame. So the reader takes the boundary from its
+ * caller:
+ *
+ * - `composerTrimmed` — `findCodexChromeStart` located the chrome and `lines`
+ *   has already been cut above it, so every `›` row left is transcript and the
+ *   newest one wins outright.
+ * - otherwise — the composer may still be the bottom-most `›` row, so the first
+ *   candidate is stepped over. This is the structural spelling of the guard the
+ *   placeholder list used to be, and unlike that list it cannot go stale.
+ *
+ * Dialog option rows are refused wherever they appear: codex renders inline, so
+ * a dialog answered minutes ago is still in the scrollback with its options
+ * intact (#1160), and anchoring on one would cut the reply mid-way.
+ *
+ * @param lines - Captured pane rows, ANSI intact where the caller has it
+ * @param contentEnd - Exclusive end of the conversation region
+ * @param windowSize - How many rows above `contentEnd` to search
+ * @param composerTrimmed - Whether `contentEnd` already excludes the composer
+ * @returns Index of the newest echo row, or -1 when none is in the window
+ */
+function findCodexUserEchoIndex(lines, contentEnd, windowSize, composerTrimmed) {
+    let composerHandled = composerTrimmed;
+    for (let i = Math.min(contentEnd, lines.length) - 1; i >= Math.max(0, contentEnd - windowSize); i--) {
+        if (!exports.CODEX_USER_ECHO_PATTERN.test((0, ansi_1.stripAnsi)(lines[i])))
+            continue;
+        const kind = (0, cli_patterns_1.readCodexGlyphRowKind)(lines[i]);
+        // An option row of a dialog still sitting in the scrollback.
+        if (kind === 'option')
+            continue;
+        // Positively an echo: codex has settled the row and drawn its glyph dim.
+        if (kind === 'transcript-echo')
+            return i;
+        // `composer` (bold glyph) or `null` (an ANSI-stripped capture, where all
+        // three uses are one byte). Either could be the input box, and the input box
+        // is always the bottom-most `›` row.
+        if (!composerHandled) {
+            composerHandled = true;
+            continue;
+        }
+        return i;
+    }
+    return -1;
+}
 /**
  * Pasted text pattern
  *

package/dist/server/src/lib/hooks/agent-event-service.js

@@ -161,16 +161,29 @@ function safeRealpath(target) {
  * skip it, and after `recordAgentStopEvent` because the reader resolves its
  * transcript through the session pointer that call refreshes.
  *
- * It is awaited rather than detached. The turn has already ended, so the cost is
- * paid by nobody who is waiting for an answer, and a detached promise here would
- * be one whose failures nothing observes.
+ * ## Exactly one read is awaited (Issue #2398)
+ *
+ * The comment that stood here said the read could be awaited freely, because
+ * "the turn has already ended, so the cost is paid by nobody who is waiting for
+ * an answer". That is true of Claude Code, which writes its transcript and then
+ * fires the hook, and false of codex, which does not append the record closing
+ * the turn until this hook's command exits — so for codex the agent *is* waiting
+ * for this answer, and the retries that were waiting for the file were waiting
+ * for themselves. Measured 2026-09-07: 105 codex stop events, 0 captured, 1 s of
+ * the agent's stop path spent on each.
+ *
+ * So the reader is asked once here, and anything further happens after this
+ * function has returned — detached, and observed through the log rather than
+ * through the return value, which is what
+ * {@link AgentStopOutcome.structuredHistoryOutcome} exists to say.
  *
  * Never throws. A hook is fire-and-forget from a CLI's stop handler, and a
  * failure here must not become the agent's problem.
  */
 async function applyAgentStopEvent(db, worktree, cliToolId, instanceId) {
     (0, agent_event_state_1.recordAgentStopEvent)(worktree.id, cliToolId, instanceId);
-    const structuredHistoryCaptured = await (0, stop_history_capture_1.captureTranscriptTurnOnStop)(worktree, cliToolId, instanceId);
+    const { status: structuredHistoryOutcome } = await (0, stop_history_capture_1.resolveStopTranscriptCapture)(worktree, cliToolId, instanceId);
+    const structuredHistoryCaptured = structuredHistoryOutcome === 'captured';
     let task;
     try {
         task = (0, tasks_db_1.getActiveTaskForInstance)(db, worktree.id, cliToolId, instanceId);
@@ -185,6 +198,7 @@ async function applyAgentStopEvent(db, worktree, cliToolId, instanceId) {
             taskEventApplied: false,
             verificationRunId: null,
             structuredHistoryCaptured,
+            structuredHistoryOutcome,
         };
     }
     if (!task) {
@@ -193,6 +207,7 @@ async function applyAgentStopEvent(db, worktree, cliToolId, instanceId) {
             taskEventApplied: false,
             verificationRunId: null,
             structuredHistoryCaptured,
+            structuredHistoryOutcome,
         };
     }
     let taskEventApplied = false;
@@ -212,6 +227,7 @@ async function applyAgentStopEvent(db, worktree, cliToolId, instanceId) {
             taskEventApplied,
             verificationRunId: null,
             structuredHistoryCaptured,
+            structuredHistoryOutcome,
         };
     }
     let verificationRunId = null;
@@ -235,5 +251,11 @@ async function applyAgentStopEvent(db, worktree, cliToolId, instanceId) {
             });
         }
     }
-    return { taskId: task.id, taskEventApplied, verificationRunId, structuredHistoryCaptured };
+    return {
+        taskId: task.id,
+        taskEventApplied,
+        verificationRunId,
+        structuredHistoryCaptured,
+        structuredHistoryOutcome,
+    };
 }

package/dist/server/src/lib/hooks/sources/codex/history.js

@@ -323,6 +323,13 @@ async function scanForRollout(dir, sessionId, depth) {
  * attach` is the thing #2196 exists to record, and it is worth recording next to
  * a *scraped* reply just as much as next to a Markdown one.
  *
+ * A turn whose `task_started` fell outside the window is written too, behind
+ * `CODEX_TURN_HEAD_TRUNCATION_MARKER` and beside a
+ * `codex-transcript-turn-headless` report (Issue #2402). See `./transcript`'s
+ * module docblock for why codex may do that where `../claude/history` may not,
+ * and `docs/design/codex-transcript-reader.md` §4.5 for why the prompts that
+ * went out of the window with it are left unadopted rather than guessed at.
+ *
  * Never throws.
  *
  * @param target - The instance whose turn just ended
@@ -411,11 +418,17 @@ async function captureCodexTranscriptTurn(target, capture = {}) {
             const previousStartedAt = index === 0 ? pending.previousStartedAt : pending.turns[index - 1].startedAt;
             userRows.push(await recordCodexUserTurns(target, pending.turns[index], previousStartedAt));
         }
+        const windowFacts = {
+            path,
+            firstRecordAt: firstCodexRecordAt(parsed.records),
+            turnlessRecords: built.turnlessRecords,
+            malformedLines: parsed.malformedLines,
+        };
         const lastRecordAt = lastCodexRecordAt(parsed.records);
         let captured = false;
         for (let index = 0; index < pending.turns.length; index += 1) {
             const turn = pending.turns[index];
-            captured = await writeCodexTurn(target, turn, (0, transcript_1.renderCodexTurn)(turn), resolveAssistantTimestampMs(turn, userRows[index], lastRecordAt.get(turn.turnId) ?? 0, nextTurnOpensAt(pending.turns, userRows, index)), path);
+            captured = await writeCodexTurn(target, turn, (0, transcript_1.renderCodexTurn)(turn), resolveAssistantTimestampMs(turn, userRows[index], lastRecordAt.get(turn.turnId) ?? 0, nextTurnOpensAt(pending.turns, userRows, index)), windowFacts);
         }
         return captured;
     }
@@ -448,6 +461,21 @@ async function resolveCodexTranscriptPath(target, capture = {}) {
         return null;
     }
 }
+/**
+ * When the window's first timestamped record was written (Issue #2402).
+ *
+ * The floor on "everything this read could possibly have seen". For a headless
+ * turn it is also the only honest thing to say about where the missing head
+ * ends: the turn began before this instant, and by how much is not in the file
+ * this read opened.
+ */
+function firstCodexRecordAt(records) {
+    for (const record of records) {
+        if (record.timestampMs !== null)
+            return record.timestampMs;
+    }
+    return null;
+}
 /**
  * The turns {@link captureCodexTranscriptTurn} still has to write (Issue #2246).
  *
@@ -642,7 +670,7 @@ function nextTurnOpensAt(turns, userRows, index) {
  *
  * @returns Whether History holds this turn as the agent's own Markdown
  */
-async function writeCodexTurn(target, turn, rendered, timestampMs, path) {
+async function writeCodexTurn(target, turn, rendered, timestampMs, windowFacts) {
     const instanceId = target.instanceId ?? target.cliToolId;
     if (!turn.closed) {
         // codex has not written the `task_complete` for this `turn_id` yet, so what
@@ -706,6 +734,38 @@ async function writeCodexTurn(target, turn, rendered, timestampMs, path) {
         });
         return true;
     }
+    if (rendered.headless) {
+        // Issue #2402. The turn's `task_started` is outside the window, so this row
+        // carries `CODEX_TURN_HEAD_TRUNCATION_MARKER` and the operator can see that
+        // — but the operator cannot see WHY, and neither can anyone reading the logs
+        // afterwards. Everything below is measured: where the window opens, what it
+        // held that no turn claimed, and how many of the turn's prompts were inside
+        // it. How much fell out is deliberately NOT here. It is not knowable from a
+        // read that never saw those bytes, and a guess would be worse than silence.
+        //
+        // `promptsInWindow` is the one to read when a prompt is missing from the
+        // conversation: codex folds a prompt submitted mid-turn into that same turn
+        // (23 of 326 archived turns carry more than one), so a headless turn's
+        // earlier prompts can be outside the window too — and unlike the body, those
+        // are NOT recoverable. `recordCodexUserTurns` only ever sees `turn.prompts`,
+        // which is what the window held; a `/send` row for a prompt written before
+        // the window stays unkeyed. Adopting it would mean guessing which unkeyed
+        // row belongs to this turn with no text to match on — see
+        // `docs/design/codex-transcript-reader.md` §4.5.
+        logger.info('codex-transcript-turn-headless', {
+            worktreeId: target.worktreeId,
+            instanceId,
+            sessionId: rendered.sessionId,
+            turnId: rendered.turnId,
+            path: windowFacts.path,
+            windowBytes: exports.CODEX_TRANSCRIPT_TAIL_BYTES,
+            windowFirstRecordAt: windowFacts.firstRecordAt,
+            turnlessRecords: windowFacts.turnlessRecords,
+            malformedLines: windowFacts.malformedLines,
+            promptsInWindow: turn.prompts.length,
+            itemsInWindow: turn.items.length,
+        });
+    }
     const message = createMessage(db, {
         worktreeId: target.worktreeId,
         role: 'assistant',
@@ -722,7 +782,7 @@ async function writeCodexTurn(target, turn, rendered, timestampMs, path) {
         instanceId,
         sessionId: rendered.sessionId,
         requestId,
-        path,
+        path: windowFacts.path,
         bodyLength: rendered.body.length,
         textBlocks: rendered.textBlocks,
         toolBlocks: rendered.toolBlocks,

package/dist/server/src/lib/hooks/sources/codex/transcript.js

@@ -53,6 +53,34 @@
  * boundary is a field, and a turn is closed by the `task_complete` that carries
  * its id — measured on 326 of 326 archived turns.
  *
+ * ## A turn can begin outside the window (Issue #2402)
+ *
+ * `./history` reads the rollout's last `CODEX_TRANSCRIPT_TAIL_BYTES` and nothing
+ * more, because the largest rollout on this machine was 273 MB. A long turn
+ * outgrows that window: measured on 2026-09-07 on a 30 MB / 7,408-line rollout
+ * whose newest turn ran seven hours over 387 items and four compactions, so its
+ * `task_started` and its first two prompts sat outside the 4 MiB tail while its
+ * later items sat inside it.
+ *
+ * The `turn_id` stamp is what makes that recoverable. Because it is on **every**
+ * `item_completed` and not only on `task_started`, a window that opens in the
+ * middle of a turn still reads the turn's real id — so the turn can be rendered
+ * AND written under the key a later read will recognise. What is lost is the
+ * head of the body, and {@link CodexTurnAccumulator.started} is how that becomes
+ * a fact the writer can state: it is the opening mark, exactly as `closed` is
+ * the closing one, and a turn with `started: false` is rendered behind
+ * {@link CODEX_TURN_HEAD_TRUNCATION_MARKER}.
+ *
+ * **This is where codex and claude diverge, and the divergence is deliberate.**
+ * `../claude/history`'s `collectHeadlessClaudeTurn` renders the same situation
+ * and refuses to write it, because claude's transcript carries no link from a
+ * reply to the prompt it answers: a turn whose prompt record is outside the
+ * window has no key at all, so writing it would mean inventing one, and an
+ * invented key is a row no later run can recognise as already written. codex is
+ * not under that constraint — the key is a field on the records that ARE in the
+ * window — so the argument that stops claude does not reach here. Copying
+ * claude's refusal would drop a body codex handed over for free.
+ *
  * ## Pure on purpose
  *
  * No filesystem, no database, no `globalThis` — `./history` owns all three, the
@@ -63,7 +91,7 @@
  * @module lib/hooks/sources/codex/transcript
  */
 Object.defineProperty(exports, "__esModule", { value: true });
-exports.CODEX_THINKING_LABEL = exports.MAX_CODEX_TOOL_DETAIL_LENGTH = exports.MAX_CODEX_TURN_ITEMS = exports.CODEX_TURN_TRUNCATION_MARKER = exports.MAX_CODEX_TURN_BODY_LENGTH = exports.CODEX_ROLLOUT_EXTENSION = exports.CODEX_SESSIONS_DIR_SEGMENTS = void 0;
+exports.CODEX_COMMENTARY_PHASE = exports.CODEX_FINAL_ANSWER_PHASE = exports.CODEX_THINKING_LABEL = exports.MAX_CODEX_TOOL_DETAIL_LENGTH = exports.MAX_CODEX_TURN_ITEMS = exports.CODEX_TURN_HEAD_TRUNCATION_MARKER = exports.CODEX_TURN_TRUNCATION_MARKER = exports.MAX_CODEX_TURN_BODY_LENGTH = exports.CODEX_ROLLOUT_EXTENSION = exports.CODEX_SESSIONS_DIR_SEGMENTS = void 0;
 exports.readCodexRolloutItem = readCodexRolloutItem;
 exports.readCodexRolloutRecord = readCodexRolloutRecord;
 exports.parseCodexRollout = parseCodexRollout;
@@ -79,6 +107,17 @@ exports.CODEX_ROLLOUT_EXTENSION = '.jsonl';
 exports.MAX_CODEX_TURN_BODY_LENGTH = 200_000;
 /** Appended when {@link MAX_CODEX_TURN_BODY_LENGTH} truncates a turn. */
 exports.CODEX_TURN_TRUNCATION_MARKER = '\n\n_(truncated)_';
+/**
+ * Prepended when a turn's opening records fell outside the window (#2402).
+ *
+ * The mirror of {@link CODEX_TURN_TRUNCATION_MARKER} — same italic parenthetical
+ * in the same plain Markdown, on the other end of the body — because the two say
+ * the same thing about the same row and a reader who has learned to read one
+ * should not have to learn a second notation for the other. Written into
+ * `chat_messages.content` at read time like every other string these readers
+ * emit, so it is English and outlives any later locale change.
+ */
+exports.CODEX_TURN_HEAD_TRUNCATION_MARKER = '_(head truncated)_\n\n';
 /**
  * Cap on items kept for one turn.
  *
@@ -91,6 +130,21 @@ exports.MAX_CODEX_TURN_ITEMS = 2048;
 exports.MAX_CODEX_TOOL_DETAIL_LENGTH = 200;
 /** The label a reasoning summary is folded behind. Same word claude uses. */
 exports.CODEX_THINKING_LABEL = 'Thinking';
+/**
+ * `AgentMessage.phase` for the reply codex means the operator to read.
+ *
+ * Measured 268 of 2,211 archived `AgentMessage` items
+ * (`docs/design/codex-transcript-reader.md` §2.3).
+ */
+exports.CODEX_FINAL_ANSWER_PHASE = 'final_answer';
+/**
+ * `AgentMessage.phase` for the progress narration codex writes while it works.
+ *
+ * Measured 1,943 of 2,211 — the majority of what codex says in a turn is this.
+ * The one phase {@link renderCodexTurn} folds; see its docblock for why the
+ * classification is an allow-list of exactly this value and nothing else.
+ */
+exports.CODEX_COMMENTARY_PHASE = 'commentary';
 /**
  * Read one `item_completed` item.
  *
@@ -322,11 +376,19 @@ function buildCodexTurns(records, sessionId) {
                 startedAt: record.timestampMs ?? 0,
                 prompts: [],
                 items: [],
+                started: false,
                 closed: false,
                 overflowed: false,
             };
             turns.set(turnId, turn);
         }
+        if (record.payloadType === 'task_started') {
+            // The opening mark (Issue #2402). Recorded rather than assumed from "this
+            // is the first record we saw for the id", which is true of every turn in
+            // the window — including the one whose real first record is outside it.
+            turn.started = true;
+            continue;
+        }
         if (record.payloadType === 'task_complete') {
             turn.closed = true;
             continue;
@@ -390,14 +452,79 @@ function renderReasoningItem(text) {
         .join('\n');
     return `> **${exports.CODEX_THINKING_LABEL}**\n>\n${quoted}`;
 }
+/**
+ * The prose one `AgentMessage` contributes, or `''` when it contributes none.
+ *
+ * The single definition of "non-empty body" for this reader, shared by
+ * {@link hasFinalAnswerProse} and by {@link renderCodexTurn}'s loop on purpose:
+ * the fallback asks whether the turn renders an answer, and if the two spelled
+ * that test out separately they could disagree about a message that is nothing
+ * but whitespace — the exact turn the fallback exists for.
+ */
+function agentMessageText(item) {
+    return item.text?.trim() ?? '';
+}
+/**
+ * Whether this turn renders at least one non-empty `final_answer`.
+ *
+ * The test is on the **rendered body**, not on the presence of an item: item
+ * count and drawn paragraphs are different quantities, because a message whose
+ * text is empty or whitespace is dropped before it reaches the body. A turn
+ * whose only `final_answer` is blank would, on an item-count test, fold its
+ * commentary away and leave an empty bubble.
+ */
+function hasFinalAnswerProse(turn) {
+    return turn.items.some((item) => item.type === 'AgentMessage' &&
+        item.phase === exports.CODEX_FINAL_ANSWER_PHASE &&
+        agentMessageText(item).length > 0);
+}
 /**
  * Render one turn to Markdown.
  *
  * Transcript order within each kind — the same decision #2041 and #2121 took,
  * and for the same reason: the order is the only record of what happened when,
- * and this row is the record. Both `commentary` and `final_answer` messages are
- * kept; codex's TUI shows both, and dropping the commentary would remove the
- * sentence that explains what the tool line underneath it is for.
+ * and this row is the record.
+ *
+ * ## Why `commentary` is folded now (Issue #2420)
+ *
+ * #2197 kept `commentary` in the body on the argument that "the sentence that
+ * explains what the tool line underneath it is for" would otherwise be lost.
+ * **#2234 retired that argument by moving every tool line to a folded section
+ * at the end of the body** — there is no line underneath the commentary any
+ * more, so what leads the bubble is narration with nothing left to narrate.
+ * Measured on one live session (`codex-turn:01a07e37-0658`): 4 commentary
+ * blocks / 754 characters in front of a single 495-character answer, so the
+ * operator read four paragraphs of progress notes before reaching the reply.
+ *
+ * ## The classification is an allow-list
+ *
+ * Exactly `phase === 'commentary'` folds. `final_answer`, a missing `phase` and
+ * any value codex has not shipped yet all stay in the body. The deny-list
+ * spelling — "fold everything that is not `final_answer`" — is rejected on
+ * #2196's discipline: it hands codex the power to hide the body of a bubble by
+ * adding a phase name, and the failure would be silent. Every `final_answer` is
+ * kept, not just the last: the measured corpus has turns carrying more than one.
+ *
+ * ## The fallback
+ *
+ * A turn with no non-empty `final_answer` keeps its commentary in the body —
+ * see {@link hasFinalAnswerProse}. Folding is only ever an improvement when
+ * there is something better to lead with.
+ *
+ * ## Why the `Thinking` chip and not a new label
+ *
+ * The folded commentary rides `../turn-body`'s existing `reasoning` kind, so it
+ * lands under the `Thinking (N)` heading the other readers already write, and
+ * `splitChatThinking` in `components/worktree/ChatMessageBubble` draws it in
+ * the chip that is already there. A second label (`Progress`) would read more
+ * precisely, and it was rejected on cost: these labels are baked into
+ * `chat_messages.content` **at read time**, so a new one is a string every
+ * future reader of the table has to keep folding forever, for rows nothing can
+ * re-label. Against that, the chip already means the right thing — "subordinate
+ * narration the reader may want and does not want first" — and the operator who
+ * reported #2420 called this content "thinking" in their own words. Same
+ * reasoning as {@link CODEX_THINKING_LABEL}, which is `Thinking` for the same
+ * reason.
  *
  * The layout — prose first, the calls folded into one labelled section — is
  * `../turn-body`'s and is shared with the other three readers (#2234).
@@ -407,12 +534,14 @@ function renderCodexTurn(turn) {
     const unknown = new Set();
     let textBlocks = 0;
     let toolBlocks = 0;
+    const foldsCommentary = hasFinalAnswerProse(turn);
     for (const item of turn.items) {
         if (item.type === 'AgentMessage') {
-            const text = item.text?.trim() ?? '';
+            const text = agentMessageText(item);
             if (text.length === 0)
                 continue;
-            rendered.push({ kind: 'prose', text });
+            const folds = foldsCommentary && item.phase === exports.CODEX_COMMENTARY_PHASE;
+            rendered.push({ kind: folds ? 'reasoning' : 'prose', text });
             textBlocks += 1;
             continue;
         }
@@ -432,6 +561,10 @@ function renderCodexTurn(turn) {
             unknown.add(item.type);
     }
     let body = (0, turn_body_1.separateTurnBody)(rendered).body;
+    // The head mark goes on before the length bound, so a headless turn long
+    // enough to be truncated at both ends still says so at both ends.
+    if (!turn.started && body.length > 0)
+        body = exports.CODEX_TURN_HEAD_TRUNCATION_MARKER + body;
     if (body.length > exports.MAX_CODEX_TURN_BODY_LENGTH) {
         body =
             body.slice(0, exports.MAX_CODEX_TURN_BODY_LENGTH - exports.CODEX_TURN_TRUNCATION_MARKER.length) +
@@ -441,6 +574,7 @@ function renderCodexTurn(turn) {
         sessionId: turn.sessionId,
         turnId: turn.turnId,
         body,
+        headless: !turn.started,
         textBlocks,
         toolBlocks,
         unknownBlockTypes: [...unknown],