commandmate 0.29.2 → 0.30.0

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

@@ -0,0 +1,654 @@
+"use strict";
+/**
+ * Writing Claude Code's own words into conversation history (Issue #2121).
+ *
+ * The second writer of `chat_messages` for a Claude turn, on the same terms
+ * `../opencode/history` established for opencode (#2041): the first writer is
+ * `lib/polling/response-checker`, which captures the pane and cleans it, and the
+ * two are mutually exclusive.
+ *
+ * ## Pull, where opencode is push
+ *
+ * That is the one structural difference, and it decides the shape of this file.
+ * opencode has a server CommandMate holds an SSE connection to, so its history
+ * writer is driven by frames and flushed by `session.idle`. Claude has no
+ * connection at all — its hooks are one-way HTTP posts that this process answers
+ * and forgets. What it has instead is a **file**: every record of a session is
+ * appended to `~/.claude/projects/<slug>/<session-id>.jsonl` as it happens.
+ *
+ * So there is nothing to subscribe to and nothing to flush, and the trigger has
+ * to come from whatever already knows a turn just ended. That is the poller —
+ * which is also the writer being replaced, so the handover happens in one place
+ * and cannot half-happen: `lib/polling/structured-history-gate` calls
+ * {@link captureClaudeTranscriptTurn} at the moment the scraped reply would be
+ * saved, and saves it only if this returns false.
+ *
+ * ## Which file is this instance's
+ *
+ * The Issue's own framing, kept: **the session id is not a key, it is a mutable
+ * pointer.** Identity stays the (worktree, tool, instance) triple that
+ * `buildCompositeKey` has always spelled; the session id is only the answer to
+ * "which transcript does that triple point at *right now*", and `/clear`
+ * changing it is correct rather than a problem, because the conversation really
+ * did change.
+ *
+ * The pointer is read from the structured events the agent already sends —
+ * `getLastAgentEvent(...).sessionId`, which the `/api/hooks/agent-event`
+ * receiver records for every hook — and latched here, because most events carry
+ * a session id but the record is replaced by whichever event was newest. Two
+ * consequences worth stating:
+ *
+ *  - **Two Claude instances in one worktree do not collide.** They share the
+ *    project directory (the slug is a function of `cwd`), but the hook URL
+ *    CommandMate injects carries `instanceId`, so each triple latches its own
+ *    session id and therefore its own file.
+ *  - **A session with no hooks has no pointer**, and this module returns false
+ *    for it. That is the fail-open the acceptance criteria ask for: no
+ *    transcript we can name means the scraper is still the only record there is.
+ *
+ * ## Nothing here throws
+ *
+ * Same contract as `../opencode/history`, for the same reason, and one more: a
+ * throw here would propagate into the poller's save path and could cost the
+ * scraped reply *as well as* the structured one. The database imports are
+ * dynamic so that `better-sqlite3` does not enter the module graph of everything
+ * that imports `@/lib/hooks/sources`.
+ *
+ * @module lib/hooks/sources/claude/history
+ */
+var __createBinding = (this && this.__createBinding) || (Object.create ? (function(o, m, k, k2) {
+    if (k2 === undefined) k2 = k;
+    var desc = Object.getOwnPropertyDescriptor(m, k);
+    if (!desc || ("get" in desc ? !m.__esModule : desc.writable || desc.configurable)) {
+      desc = { enumerable: true, get: function() { return m[k]; } };
+    }
+    Object.defineProperty(o, k2, desc);
+}) : (function(o, m, k, k2) {
+    if (k2 === undefined) k2 = k;
+    o[k2] = m[k];
+}));
+var __setModuleDefault = (this && this.__setModuleDefault) || (Object.create ? (function(o, v) {
+    Object.defineProperty(o, "default", { enumerable: true, value: v });
+}) : function(o, v) {
+    o["default"] = v;
+});
+var __importStar = (this && this.__importStar) || (function () {
+    var ownKeys = function(o) {
+        ownKeys = Object.getOwnPropertyNames || function (o) {
+            var ar = [];
+            for (var k in o) if (Object.prototype.hasOwnProperty.call(o, k)) ar[ar.length] = k;
+            return ar;
+        };
+        return ownKeys(o);
+    };
+    return function (mod) {
+        if (mod && mod.__esModule) return mod;
+        var result = {};
+        if (mod != null) for (var k = ownKeys(mod), i = 0; i < k.length; i++) if (k[i] !== "default") __createBinding(result, mod, k[i]);
+        __setModuleDefault(result, mod);
+        return result;
+    };
+})();
+Object.defineProperty(exports, "__esModule", { value: true });
+exports.CLAUDE_TRANSCRIPT_TAIL_BYTES = void 0;
+exports.resetClaudeTranscriptSessions = resetClaudeTranscriptSessions;
+exports.resolveClaudeSessionId = resolveClaudeSessionId;
+exports.claudeProjectsRoot = claudeProjectsRoot;
+exports.claudeTranscriptPath = claudeTranscriptPath;
+exports.acceptClaudeTranscriptHint = acceptClaudeTranscriptHint;
+exports.captureClaudeTranscriptTurn = captureClaudeTranscriptTurn;
+exports.readClaudeTurnProgress = readClaudeTurnProgress;
+const promises_1 = require("fs/promises");
+const os_1 = require("os");
+const path_1 = require("path");
+const auto_yes_state_1 = require("../../../../lib/auto-yes-state");
+const user_turn_recorder_1 = require("../../../../lib/history/user-turn-recorder");
+const logger_1 = require("../../../../lib/logger");
+const agent_transcript_1 = require("../../../../types/agent-transcript");
+const transcript_1 = require("./transcript");
+const logger = (0, logger_1.createLogger)('lib/hooks/sources/claude/history');
+/**
+ * How much of the transcript's tail is read.
+ *
+ * The file grows for the life of the session and a long one is tens of
+ * megabytes — the largest on this machine on 2026-08-31 was 23 MB — so reading
+ * it whole on every finished turn would be the most expensive thing the poller
+ * does. Only the newest turn is ever written (see {@link captureClaudeTranscriptTurn}),
+ * so only the newest turn has to be in the window.
+ *
+ * 4 MiB is roughly two orders of magnitude above the measured size of a single
+ * turn and still small enough to read and parse in one tick. A turn that
+ * genuinely does not fit produces orphaned assistant records, which is detected
+ * and reported rather than written as a headless reply.
+ */
+exports.CLAUDE_TRANSCRIPT_TAIL_BYTES = 4 * 1024 * 1024;
+/** `.jsonl`; the only extension this reader will open. */
+const CLAUDE_TRANSCRIPT_EXTENSION = '.jsonl';
+/**
+ * The last session id seen for each instance.
+ *
+ * On `globalThis` for the reason every shared map in this subsystem is (#1736):
+ * under `next dev` the poller's bundle and the hook receiver's bundle would each
+ * get a private copy of a module-scoped map.
+ *
+ * A latch and not a cache: `getLastAgentEvent` holds only the newest event, and
+ * an event that carried no `session_id` would otherwise blank the pointer
+ * mid-session. Same reasoning as `getLastKnownAgentModel`, which latches for the
+ * same reason.
+ */
+const sessionPointers = (globalThis.__claudeTranscriptSessions ??= new Map());
+function keyOf(target) {
+    return (0, auto_yes_state_1.buildCompositeKey)(target.worktreeId, target.cliToolId, target.instanceId);
+}
+/** Forget every instance's pointer. Test seam. */
+function resetClaudeTranscriptSessions() {
+    sessionPointers.clear();
+}
+/**
+ * The session id this instance's transcript is under, or null.
+ *
+ * Reads the structured event state first and falls back to the latched value.
+ * The import is dynamic so that `agent-event-state`'s module graph does not
+ * become a static dependency of the poller.
+ */
+async function resolveClaudeSessionId(target) {
+    const key = keyOf(target);
+    try {
+        const { getLastAgentEvent } = await Promise.resolve().then(() => __importStar(require('../../../../lib/session/agent-event-state')));
+        const sessionId = getLastAgentEvent(target.worktreeId, target.cliToolId, target.instanceId)?.sessionId;
+        if (typeof sessionId === 'string' && sessionId.length > 0) {
+            sessionPointers.set(key, sessionId);
+            return sessionId;
+        }
+    }
+    catch (error) {
+        // A state module that cannot be reached is one that knows no session id.
+        logger.debug('claude-transcript-session-lookup-failed', {
+            worktreeId: target.worktreeId,
+            instanceId: target.instanceId ?? target.cliToolId,
+            error: error instanceof Error ? error.message : String(error),
+        });
+    }
+    return sessionPointers.get(key) ?? null;
+}
+/** `<home>/.claude/projects`. */
+function claudeProjectsRoot(homeDir) {
+    return (0, path_1.join)(homeDir, ...transcript_1.CLAUDE_PROJECTS_DIR_SEGMENTS);
+}
+/**
+ * Where this session's transcript is, from the worktree's own path.
+ *
+ * `cwd` is the agent's working directory, and CommandMate starts every pane at
+ * the worktree root, so the worktree path is the slug input. An agent the
+ * operator `cd`-ed somewhere else writes to a different directory and this
+ * answers a path that does not exist — read as "no transcript", which is the
+ * fail-open direction.
+ */
+function claudeTranscriptPath(homeDir, worktreePath, sessionId) {
+    return (0, path_1.join)(claudeProjectsRoot(homeDir), (0, transcript_1.claudeProjectSlug)(worktreePath), `${sessionId}${CLAUDE_TRANSCRIPT_EXTENSION}`);
+}
+/**
+ * A path named by something other than this module, accepted only if it is
+ * really a transcript.
+ *
+ * The one caller is the `📄 Session log: …jsonl` line `lib/claude-output` reads
+ * off the pane, which is text the agent printed and therefore text an agent
+ * could print. Two conditions, both necessary: it must be under
+ * `~/.claude/projects`, so a crafted line cannot make this open
+ * `/etc/passwd`; and it must end in `.jsonl`.
+ *
+ * Containment is checked on the resolved path so that `..` cannot climb out,
+ * and `resolve` is safe to use here — unlike in `validateHookCwd`, where the
+ * value is echoed back — because the resolved string is the only thing that is
+ * ever used.
+ *
+ * @returns The resolved path, or null when it is not acceptable
+ */
+function acceptClaudeTranscriptHint(homeDir, hint) {
+    if (!hint.endsWith(CLAUDE_TRANSCRIPT_EXTENSION))
+        return null;
+    if (hint.includes('\0'))
+        return null;
+    const root = (0, path_1.resolve)(claudeProjectsRoot(homeDir));
+    const resolved = (0, path_1.resolve)(hint);
+    if (resolved !== root && !resolved.startsWith(root + path_1.sep))
+        return null;
+    return resolved;
+}
+/**
+ * Read the newest turn out of this instance's transcript and write it.
+ *
+ * **Only the newest turn**, and that is the deliberate difference from
+ * opencode's backfill. Every earlier turn of this session already has a
+ * `chat_messages` row that the *scraper* wrote, and writing a Markdown row for
+ * it as well would put the same reply in History twice — once as prose and once
+ * as the pane drew it. The scraper wrote exactly one row per turn and this
+ * writes exactly one row per turn, so the swap is one-for-one and history is
+ * neither duplicated nor rewritten.
+ *
+ * The return value is the poller's instruction, so the two failure directions
+ * are worth stating plainly. **True** means this path has recorded the turn and
+ * the scrape must be dropped. **False** means it has not, for any reason at all
+ * — no session pointer, no file, an unreadable file, a turn with no assistant
+ * text yet — and the scrape must be saved. Everything that can go wrong answers
+ * false, which is the fail-open the acceptance criteria require: two writers
+ * duplicate a reply, no writer loses one.
+ *
+ * Never throws.
+ *
+ * @param target - The instance whose turn just ended
+ * @returns Whether History now holds this turn as the agent's own Markdown
+ */
+async function captureClaudeTranscriptTurn(target, capture) {
+    const instanceId = target.instanceId ?? target.cliToolId;
+    try {
+        const homeDir = capture.homeDir ?? (0, os_1.homedir)();
+        if (typeof capture.worktreePath !== 'string' || capture.worktreePath.length === 0) {
+            return false;
+        }
+        const sessionId = await resolveClaudeSessionId(target);
+        const path = await locateClaudeTranscript(homeDir, capture, sessionId);
+        if (!path) {
+            logger.debug('claude-transcript-unavailable', {
+                worktreeId: target.worktreeId,
+                instanceId,
+                sessionId,
+                reason: sessionId ? 'no-file' : 'no-session-pointer',
+            });
+            return false;
+        }
+        const text = await readClaudeTranscriptTail(path);
+        if (text === null)
+            return false;
+        const parsed = (0, transcript_1.parseClaudeTranscript)(text);
+        const built = (0, transcript_1.buildClaudeTurns)(parsed.records, sessionId ?? '');
+        const turn = built.turns.at(-1);
+        if (!turn) {
+            logger.info('claude-transcript-no-turn', {
+                worktreeId: target.worktreeId,
+                instanceId,
+                path,
+                records: parsed.records.length,
+                malformedLines: parsed.malformedLines,
+                orphanedAssistantRecords: built.orphanedAssistantRecords,
+            });
+            return false;
+        }
+        if (parsed.malformedLines > 0 || built.orphanedAssistantRecords > 0) {
+            // Both are expected in small numbers — a fragment at the tail of a file
+            // being appended to, and the head of the window landing mid-turn — and
+            // both are the kind of thing that must be visible when it stops being
+            // small.
+            //
+            // `orphanedAssistantRecords` is also the shape the Issue #2196 tail-window
+            // trap takes: a turn whose prompt record fell outside
+            // CLAUDE_TRANSCRIPT_TAIL_BYTES has assistant records and no prompt to
+            // record a user row from. It is reported here rather than dropped in
+            // silence, exactly as #2121 left it.
+            logger.info('claude-transcript-partial-read', {
+                worktreeId: target.worktreeId,
+                instanceId,
+                path,
+                malformedLines: parsed.malformedLines,
+                orphanedAssistantRecords: built.orphanedAssistantRecords,
+                sidechainRecords: built.sidechainRecords,
+            });
+        }
+        const userRow = await recordClaudeUserTurn(target, turn);
+        return await writeClaudeTurn(target, (0, transcript_1.renderClaudeTurn)(turn), resolveAssistantTimestampMs(turn, userRow), path);
+    }
+    catch (error) {
+        logger.error('claude-transcript-capture-failed', {
+            worktreeId: target.worktreeId,
+            instanceId,
+            error: error instanceof Error ? error.message : String(error),
+        });
+        return false;
+    }
+}
+/**
+ * The key for a body read with no prompt record in the window.
+ *
+ * `partial:` cannot be the prefix of a UUID, so this can never collide with
+ * `claudeTurnRequestId(promptUuid)` — which matters because the two mean
+ * different things to the client. A prompt-derived key is a promise that a row
+ * with the same id is coming; this one is a promise that no such row exists, so
+ * the bubble it draws is cleared by the session going idle rather than by a swap.
+ */
+function headlessClaudeTurnKey(sessionId) {
+    return (0, agent_transcript_1.claudeTurnRequestId)(`partial:${sessionId}`);
+}
+/**
+ * Everything the window holds for a turn whose prompt record is outside it.
+ *
+ * Built here rather than in `buildClaudeTurns` on purpose: that function drops
+ * orphaned records for a stated reason — their text must never be attached to an
+ * invented turn key, because an invented key is a row no later run can recognise
+ * as already written. That argument is about **writing**, and this path writes
+ * nothing. So the records are gathered separately, into an accumulator that is
+ * rendered by the same `renderClaudeTurn` and never handed to a writer.
+ *
+ * Only the records before the first prompt record are taken, which is exactly
+ * the set `buildClaudeTurns` counted as orphaned.
+ */
+function collectHeadlessClaudeTurn(records, sessionId) {
+    const blocks = [];
+    let assistantRecords = 0;
+    let overflowed = false;
+    for (const record of records) {
+        if ((0, transcript_1.isClaudePromptRecord)(record))
+            break;
+        if (record.isSidechain || record.type !== 'assistant')
+            continue;
+        assistantRecords += 1;
+        for (const block of record.blocks) {
+            if (blocks.length >= transcript_1.MAX_CLAUDE_TURN_BLOCKS) {
+                overflowed = true;
+                break;
+            }
+            blocks.push(block);
+        }
+    }
+    return {
+        sessionId,
+        promptUuid: `partial:${sessionId}`,
+        startedAt: 0,
+        promptText: '',
+        promptIsOperatorInput: false,
+        blocks,
+        assistantRecords,
+        overflowed,
+    };
+}
+/**
+ * Read the turn that is open right now, without writing anything (Issue #2199).
+ *
+ * The read half of {@link captureClaudeTranscriptTurn}, and *only* the read
+ * half: no `chat_messages` row, no user row, no `broadcastMessage`. That
+ * separation is the whole safety argument for this function. The write path is
+ * idempotent because `findMessageByRequestId` answers for a `requestId` derived
+ * from the prompt record's `uuid`; running it on every poll tick of an
+ * unfinished turn would mean writing the reply as it grows and then finding the
+ * row already there, so the row would freeze at whatever the first tick saw.
+ * This path never reaches that code at all.
+ *
+ * The two also cannot fight over the file: both open it read-only, and the
+ * writer is Claude.
+ *
+ * Deliberately **not** gated on whether the turn has ended. It cannot be — the
+ * transcript has no end-of-turn record this reader could trust, which is why
+ * #2121 put the trigger in the poller. The caller supplies that judgement (it
+ * asks only while the session is generating), and a body read from a turn that
+ * has just finished is harmless: the settled row carries the same
+ * {@link ClaudeTurnProgress.turnKey}, so the client replaces one with the other.
+ *
+ * Never throws.
+ *
+ * @param target - The instance whose turn is in flight
+ * @param capture - Where to look; the same {@link ClaudeTranscriptCapture} the writer takes
+ * @returns The open turn's body, or null when there is nothing to show yet
+ */
+async function readClaudeTurnProgress(target, capture) {
+    const instanceId = target.instanceId ?? target.cliToolId;
+    try {
+        const homeDir = capture.homeDir ?? (0, os_1.homedir)();
+        if (typeof capture.worktreePath !== 'string' || capture.worktreePath.length === 0) {
+            return null;
+        }
+        const sessionId = await resolveClaudeSessionId(target);
+        const path = await locateClaudeTranscript(homeDir, capture, sessionId);
+        if (!path)
+            return null;
+        const text = await readClaudeTranscriptTail(path);
+        if (text === null)
+            return null;
+        // A fragment at the tail of a file being appended to is the normal case here
+        // — this reader runs *while* Claude is writing — and `parseClaudeTranscript`
+        // already counts it rather than throwing. Nothing extra is needed, and that
+        // is the property `claude-transcript-progress-2199` pins.
+        const parsed = (0, transcript_1.parseClaudeTranscript)(text);
+        const built = (0, transcript_1.buildClaudeTurns)(parsed.records, sessionId ?? '');
+        const turn = built.turns.at(-1);
+        if (turn) {
+            const rendered = (0, transcript_1.renderClaudeTurn)(turn);
+            if (rendered.body.length === 0)
+                return null;
+            return {
+                turnKey: (0, agent_transcript_1.claudeTurnRequestId)(rendered.promptUuid),
+                body: rendered.body,
+                partial: false,
+            };
+        }
+        // No prompt record anywhere in the window. See {@link ClaudeTurnProgress.partial}.
+        if (built.orphanedAssistantRecords === 0)
+            return null;
+        const headless = collectHeadlessClaudeTurn(parsed.records, sessionId ?? '');
+        const rendered = (0, transcript_1.renderClaudeTurn)(headless);
+        if (rendered.body.length === 0)
+            return null;
+        logger.info('claude-transcript-progress-headless', {
+            worktreeId: target.worktreeId,
+            instanceId,
+            path,
+            orphanedAssistantRecords: built.orphanedAssistantRecords,
+            malformedLines: parsed.malformedLines,
+        });
+        return {
+            turnKey: headlessClaudeTurnKey(sessionId ?? ''),
+            body: rendered.body,
+            partial: true,
+        };
+    }
+    catch (error) {
+        logger.debug('claude-transcript-progress-failed', {
+            worktreeId: target.worktreeId,
+            instanceId,
+            error: error instanceof Error ? error.message : String(error),
+        });
+        return null;
+    }
+}
+/**
+ * Record the prompt this turn answers, before the reply is written (Issue #2196).
+ *
+ * Ordering is the reason this is a separate call and not a branch inside
+ * {@link writeClaudeTurn}: the user row has to exist *before* the assistant row
+ * so that {@link resolveAssistantTimestampMs} can put the reply after it, and so
+ * that a browser watching the broadcast sees a prompt appear and then an answer
+ * rather than the other way round.
+ *
+ * A turn whose prompt was not the operator's — a `<task-notification>`, a
+ * compaction summary, a headless `sdk` run; see `isClaudeOperatorPromptRecord` —
+ * is skipped and said so in the log. The assistant row is still written: the
+ * agent really did reply, and #2121's behaviour for those turns is unchanged.
+ *
+ * Never throws; `recordUserTurn` reports its failures in the return value.
+ */
+async function recordClaudeUserTurn(target, turn) {
+    const instanceId = target.instanceId ?? target.cliToolId;
+    if (!turn.promptIsOperatorInput) {
+        logger.debug('claude-transcript-user-turn-skipped', {
+            worktreeId: target.worktreeId,
+            instanceId,
+            promptUuid: turn.promptUuid,
+            reason: 'not-operator-input',
+        });
+        return { outcome: 'skipped', messageId: null, timestampMs: null };
+    }
+    const recorded = await (0, user_turn_recorder_1.recordUserTurn)(target, (0, agent_transcript_1.claudePromptRequestId)(turn.promptUuid), turn.promptText, turn.startedAt);
+    if (recorded.outcome === 'failed') {
+        logger.warn('claude-transcript-user-turn-failed', {
+            worktreeId: target.worktreeId,
+            instanceId,
+            sessionId: turn.sessionId,
+            promptUuid: turn.promptUuid,
+        });
+    }
+    return recorded;
+}
+/**
+ * When the assistant row for this turn is dated.
+ *
+ * #2121 dated it by the prompt record's clock so that a row written a poll late
+ * still sorts where the conversation put it, and that is still what happens for
+ * every turn that produced no user row.
+ *
+ * When there *is* a user row, it sits on that same instant — both come from the
+ * one prompt record — and `groupMessagesIntoPairs` orders by timestamp and
+ * nothing else. A tie there is decided by whichever row the database happened to
+ * return first, and the losing arrangement (assistant, then user) is exactly the
+ * `orphan` pair this Issue exists to remove. So the reply is moved to the first
+ * millisecond after the prompt: the smallest change that makes the order a
+ * property of the data rather than of the query plan.
+ */
+function resolveAssistantTimestampMs(turn, userRow) {
+    if (userRow.timestampMs === null)
+        return turn.startedAt;
+    return Math.max(turn.startedAt, userRow.timestampMs + 1);
+}
+/**
+ * The transcript file for this instance, or null.
+ *
+ * The session pointer first, the pane's own claim second. Both are checked
+ * against the filesystem rather than trusted, because "the path we would use"
+ * and "the path that exists" differ for every session the operator started
+ * outside CommandMate.
+ */
+async function locateClaudeTranscript(homeDir, capture, sessionId) {
+    if (sessionId) {
+        const path = claudeTranscriptPath(homeDir, capture.worktreePath, sessionId);
+        if (await isReadableFile(path))
+            return path;
+    }
+    const hint = capture.transcriptPathHint;
+    if (typeof hint === 'string' && hint.length > 0) {
+        const accepted = acceptClaudeTranscriptHint(homeDir, hint);
+        if (accepted && (await isReadableFile(accepted)))
+            return accepted;
+    }
+    return null;
+}
+async function isReadableFile(path) {
+    try {
+        return (await (0, promises_1.stat)(path)).isFile();
+    }
+    catch {
+        return false;
+    }
+}
+/**
+ * The last {@link CLAUDE_TRANSCRIPT_TAIL_BYTES} of the file, as UTF-8.
+ *
+ * Read at an offset rather than whole, and the first line of a windowed read is
+ * dropped: starting mid-line would hand `parseClaudeTranscript` a fragment that
+ * it would count as malformed anyway, and dropping it deliberately keeps that
+ * counter meaning "the writer was mid-append", which is the thing worth seeing.
+ *
+ * @returns The text, or null when the file could not be read
+ */
+async function readClaudeTranscriptTail(path) {
+    let handle;
+    try {
+        handle = await (0, promises_1.open)(path, 'r');
+        const { size } = await handle.stat();
+        const offset = Math.max(0, size - exports.CLAUDE_TRANSCRIPT_TAIL_BYTES);
+        const length = size - offset;
+        if (length <= 0)
+            return '';
+        const buffer = Buffer.alloc(length);
+        const { bytesRead } = await handle.read(buffer, 0, length, offset);
+        const text = buffer.subarray(0, bytesRead).toString('utf8');
+        if (offset === 0)
+            return text;
+        const firstBreak = text.indexOf('\n');
+        return firstBreak === -1 ? '' : text.slice(firstBreak + 1);
+    }
+    catch (error) {
+        logger.warn('claude-transcript-read-failed', {
+            path,
+            error: error instanceof Error ? error.message : String(error),
+        });
+        return null;
+    }
+    finally {
+        await handle?.close().catch(() => { });
+    }
+}
+/**
+ * Write one rendered turn, unless it is already there.
+ *
+ * `findMessageByRequestId` is both the idempotency check and the reason a
+ * repeat poll does not duplicate the row: the id is derived from the prompt
+ * record's `uuid`, which does not change between reads of the same file.
+ *
+ * Answering **true** for a turn that was already saved is deliberate. It means
+ * "History holds this turn as Markdown", which is exactly what the poller needs
+ * to know — a second poll of the same finished turn must not save the pane's
+ * copy on top of the row this path wrote for it.
+ *
+ * @returns Whether History holds this turn as the agent's own Markdown
+ */
+async function writeClaudeTurn(target, rendered, timestampMs, path) {
+    const instanceId = target.instanceId ?? target.cliToolId;
+    if (rendered.body.length === 0) {
+        // The turn is open but the agent has not written anything to the file yet —
+        // the prompt record is there and the assistant records are not. Answering
+        // false hands the turn back to the scraper, which is the only correct
+        // answer: an empty row would show as a blank reply forever, and suppressing
+        // the scrape would lose the reply outright.
+        logger.info('claude-transcript-turn-empty', {
+            worktreeId: target.worktreeId,
+            instanceId,
+            sessionId: rendered.sessionId,
+            promptUuid: rendered.promptUuid,
+        });
+        return false;
+    }
+    if (rendered.unknownBlockTypes.length > 0) {
+        logger.info('claude-transcript-unknown-blocks', {
+            worktreeId: target.worktreeId,
+            instanceId,
+            sessionId: rendered.sessionId,
+            blockTypes: rendered.unknownBlockTypes,
+        });
+    }
+    const requestId = (0, agent_transcript_1.claudeTurnRequestId)(rendered.promptUuid);
+    const [{ getDbInstance }, { createMessage, findMessageByRequestId }, { broadcastMessage }] = await Promise.all([
+        Promise.resolve().then(() => __importStar(require('../../../../lib/db/db-instance'))),
+        Promise.resolve().then(() => __importStar(require('../../../../lib/db'))),
+        Promise.resolve().then(() => __importStar(require('../../../../lib/ws-server'))),
+    ]);
+    const db = getDbInstance();
+    if (findMessageByRequestId(db, target.worktreeId, requestId)) {
+        logger.debug('claude-transcript-turn-already-saved', {
+            worktreeId: target.worktreeId,
+            instanceId,
+            requestId,
+        });
+        return true;
+    }
+    const message = createMessage(db, {
+        worktreeId: target.worktreeId,
+        role: 'assistant',
+        content: rendered.body,
+        messageType: 'normal',
+        // The agent's own clock, from the prompt record's `timestamp`, so a row
+        // written a poll late still sorts where the conversation put it.
+        timestamp: new Date(timestampMs > 0 ? timestampMs : Date.now()),
+        cliToolId: target.cliToolId,
+        instanceId,
+        requestId,
+    });
+    broadcastMessage('message', { worktreeId: target.worktreeId, message });
+    logger.info('claude-transcript-turn-saved', {
+        worktreeId: target.worktreeId,
+        instanceId,
+        sessionId: rendered.sessionId,
+        requestId,
+        path,
+        bodyLength: rendered.body.length,
+        textBlocks: rendered.textBlocks,
+        toolBlocks: rendered.toolBlocks,
+    });
+    return true;
+}

package/dist/server/src/lib/hooks/sources/claude/source.js

@@ -120,6 +120,10 @@ exports.claudeAgentEventSource = (0, define_source_1.definePushHookSource)({
         eventIdentity: null,
         // push. There is no connection to lose and nothing to re-read.
         resync: 'none',
+        // Issue #2197. `lib/hooks/sources/claude/history` reads the transcript
+        // JSONL when the poller asks (#2121); the gate dispatches on this word
+        // rather than on the tool id.
+        transcriptHistory: 'pull',
     },
     // S1. A plain name table is enough *for this tool* — one native name, one
     // word, no predicates — so it is expressed as rules built from the table