@cat-factory/executor-harness 1.64.0 → 1.64.4

package/dist/agent-runner.js

@@ -2,13 +2,14 @@ import { spawn } from 'node:child_process';
 import { mkdir, mkdtemp, rm, writeFile } from 'node:fs/promises';
 import { tmpdir } from 'node:os';
 import { dirname, join } from 'node:path';
-import { claudeAssistantContent, claudeCallUsage, isObject, numberOf, redactBody, } from './claude-stream.js';
+import { claudeAssistantContent, isObject, numberOf, redactBody } from './claude-stream.js';
+import { createClaudeRunTelemetry, subagentDispatchId } from './claude-call-aggregator.js';
 import { createCallMetricPublisher, publishCallMetric, } from './pi.js';
 import { ProgressGuard } from './progress-guard.js';
 import { killChildProcess, spawnDetached } from './process.js';
 import { redact, secretsToRedact } from './redact.js';
 import { createSliceTracker, startSubagentWatcher } from './subagents.js';
-import { createTaskPlanTracker, normalizeStatus, pickProgress, toProgress, todosToProgress, } from './progress.js';
+import { createTaskPlanTracker, mergeProgress, normalizeStatus, pickProgress, toProgress, todosToProgress, } from './progress.js';
 import { assertOnboardingKeysCurrent, writeOnboardingPreseed } from './onboarding-preseed.js';
 import { retainSessionTranscripts } from './transcript-retention.js';
 /**
@@ -214,44 +215,54 @@ export async function runClaudeCode(opts) {
         });
     }
     // Reconstruct the full per-call request/response bodies for telemetry from the
-    // stream. `--output-format stream-json --verbose` emits each turn as a near-verbatim
-    // Anthropic Messages envelope, so `assistant` events carry the complete response
-    // (text + tool_use blocks + usage), and `user` events carry the tool_result blocks
-    // fed back — together the growing prompt transcript. We seed it with the inputs the
-    // harness supplies (they never appear in the stream): the system + first user message
-    // when the prompt rides argv, or a single folded user turn when it doesn't — so the
-    // reconstruction never shows a system turn that was never sent. Bodies are
-    // credential-scrubbed (they can echo the leased token).
+    // stream. `--output-format stream-json --verbose` emits a near-verbatim Anthropic
+    // Messages envelope per response CONTENT BLOCK (not per call), so the aggregator below
+    // folds the envelopes sharing a `message.id` back into one call and buffers that call's
+    // `user` tool_result turns — together the growing prompt transcript, in the shape the
+    // model was actually sent. We seed it with the inputs the harness supplies (they never
+    // appear in the stream): the system + first user message when the prompt rides argv, or
+    // a single folded user turn when it doesn't — so the reconstruction never shows a system
+    // turn that was never sent. Bodies are credential-scrubbed (they can echo the leased token).
     const secrets = opts.subscriptionToken ? secretsToRedact(opts.subscriptionToken) : [];
-    const messages = folded
-        ? [{ role: 'user', content: prompt }]
-        : [
-            { role: 'system', content: opts.systemPrompt },
-            { role: 'user', content: opts.userPrompt },
-        ];
     const calls = [];
     // Streams each call as the CLI yields it, EXCEPT one whose tokens `attributeCumulativeUsage`
     // may still rewrite below (a published call must be final — see the publisher).
     const publisher = createCallMetricPublisher(calls, opts.onCallMetric);
-    // ADR 0026 D2.1 + ADR 0027 Defect B: surface live slice progress from TWO reconciled
-    // sources. The parent's subagent dispatches + their terminal tool_results DO appear on this
-    // stream (only a subagent's intermediate turns don't), so `sliceTracker` derives per-slice
-    // progress for the parallel shape; the parent's own plan (the sequential shape) is tracked
-    // by `planTracker` + `lastTodo`. `pickProgress` picks whichever is further along on each
-    // update, so neither masks the other — the pr-reviewer prompt writes its plan ONCE and never
-    // marks it done, which used to gate the slice signal off and pin progress at 0%.
+    // `watcherOwnsSubagents` tracks the `startSubagentWatcher` wiring below: it is started only when
+    // the CLI has an isolated config home to watch, which an `ambientAuth` run does not have. The
+    // telemetry routes the CLI's tagged subagent turns accordingly — see `createClaudeRunTelemetry`.
+    const telemetry = createClaudeRunTelemetry({
+        seed: folded
+            ? [{ role: 'user', content: prompt }]
+            : [
+                { role: 'system', content: opts.systemPrompt },
+                { role: 'user', content: opts.userPrompt },
+            ],
+        secrets,
+        watcherOwnsSubagents: !opts.ambientAuth,
+        publish: (metric) => publisher.publish(metric),
+    });
+    // ADR 0026 D2.1 + ADR 0027 Defect B: surface live slice progress from the two views the run
+    // produces of the SAME slicing. The parent's subagent dispatches + their terminal tool_results
+    // appear on this stream (as do the subagents' own intermediate turns, tagged with the dispatch
+    // that spawned them — see `isSubagentEvent`), so `sliceTracker` knows which slices are in flight
+    // and which have returned; the parent's own plan (tracked by `planTracker` + `lastTodo`) is the
+    // only place a not-yet-dispatched slice is named at all.
     //
     // The plan arrives in one of two tool vocabularies depending on the bundled CLI build:
     // `TodoWrite` (whole-list snapshots, tracked in `lastTodo`) or the incremental
     // `TaskCreate`/`TaskUpdate` pair (tracked by `planTracker`, which needs the tool RESULTS too
-    // because the task id is minted there). Both are read — see ./progress.ts.
+    // because the task id is minted there). Both are read, and `pickProgress` resolves that
+    // either/or; the plan then MERGES with the dispatch view (`mergeProgress`) rather than
+    // competing with it — picking the further-along view collapsed the list to the dispatched
+    // slices alone the moment the first subagent returned. See ./progress.ts.
     const sliceTracker = createSliceTracker();
     const planTracker = createTaskPlanTracker();
     let lastTodo;
     const emitProgress = () => {
         if (!opts.onProgress)
             return;
-        const progress = pickProgress(pickProgress(lastTodo, planTracker.progress()), sliceTracker.progress());
+        const progress = mergeProgress(pickProgress(lastTodo, planTracker.progress()), sliceTracker.progress());
         if (progress)
             opts.onProgress(progress);
     };
@@ -294,12 +305,19 @@ export async function runClaudeCode(opts) {
     };
     const onEvent = (event, meta) => {
         const type = event.type;
+        // A subagent's turns ride the parent's stdout tagged with the dispatch that spawned them;
+        // `telemetry` routes them off the parent's chain (and decides who bills them). Progress, slice
+        // tracking, the guard and `stats` below deliberately see EVERY event: a subagent grinding on
+        // errors should trip the guard exactly as the parent would, and whether the agent acted at all
+        // does not depend on which channel billed it.
+        const dispatchId = subagentDispatchId(event);
         if (type === 'assistant' && isObject(event.message)) {
             const message = event.message;
             const content = Array.isArray(message.content) ? message.content : [];
-            const { text, reasoning, toolUses } = claudeAssistantContent(content);
+            const { text, toolUses } = claudeAssistantContent(content);
             stats.assistantChars += text.length;
             stats.toolCalls += toolUses;
+            telemetry.onAssistant(dispatchId, message);
             for (const block of content) {
                 if (!isObject(block) || block.type !== 'tool_use')
                     continue;
@@ -317,22 +335,6 @@ export async function runClaudeCode(opts) {
             sliceTracker.onAssistant(content);
             planTracker.onAssistant(content);
             emitProgress();
-            // Record this call BEFORE appending its turn: the prompt is the history that
-            // produced this response. The append-only array keeps each call's prompt a strict
-            // prefix of the next, so the backend's telemetry chain delta-compresses cleanly.
-            const u = claudeCallUsage(message.usage);
-            publisher.publish({
-                ...(typeof message.model === 'string' ? { model: message.model } : {}),
-                promptText: redactBody(JSON.stringify(messages), secrets),
-                messageCount: messages.length,
-                responseText: redactBody(text, secrets),
-                reasoningText: redactBody(reasoning, secrets),
-                inputTokens: u.inputTokens,
-                cachedInputTokens: u.cachedInputTokens,
-                outputTokens: u.outputTokens,
-                finishReason: typeof message.stop_reason === 'string' ? message.stop_reason : null,
-            });
-            messages.push({ role: 'assistant', content });
         }
         else if (type === 'user' && isObject(event.message)) {
             // tool_result blocks the harness fed back to the model — part of the next prompt.
@@ -345,7 +347,7 @@ export async function runClaudeCode(opts) {
                 // would kill nothing and only convert a clean exit into a spurious failure.
                 if (!meta?.final)
                     feedGuard(content);
-                messages.push({ role: 'tool', content });
+                telemetry.onToolResult(dispatchId, content);
             }
         }
         else if (type === 'result') {
@@ -431,6 +433,8 @@ export async function runClaudeCode(opts) {
                 ...appendArgs,
             ],
         }, prompt, { ...opts, signal: runSignal }, env, opts.subscriptionToken ? secretsToRedact(opts.subscriptionToken) : [], onEvent);
+        // The stream has ended, so the last call has no successor envelope to complete it.
+        telemetry.flush();
         return await assembleClaudeOutcome({
             summary,
             stats,
@@ -439,9 +443,17 @@ export async function runClaudeCode(opts) {
             publisher,
             usage,
             subagents,
+            expectSubagentCalls: telemetry.expectsWatcherCalls(),
+            ...(opts.log ? { log: opts.log } : {}),
         });
     }
     catch (err) {
+        // The stream ended abnormally (guard trip, watchdog kill, CLI crash). Complete the call in
+        // flight anyway, and release whatever the publisher was withholding: a killed run never
+        // returns an outcome, so the live channel is the ONLY record of what it spent, and dropping
+        // its last turn is what the streaming exists to avoid.
+        telemetry.flush();
+        publisher.flush();
         // A tripped no-progress guard aborted the CLI; streamCli rejects with its generic abort
         // message, so replace it with the guard's actionable diagnostic — carrying the stderr tail
         // it attached, since that is usually the only evidence of what the CLI was doing when it was
@@ -497,6 +509,10 @@ function buildClaudeEnv(opts, configHome) {
  * terminal `result` event's cumulative) covers ONLY the parent loop, and the subagent tokens live
  * exclusively in the per-session `subagents/*.jsonl` transcripts the watcher reads. Extracted from
  * {@link runClaudeCode} verbatim to keep its cyclomatic complexity down.
+ *
+ * The invariant is about the aggregate `usage` only. `calls` was NEVER disjoint from the watcher's
+ * on its own: the CLI streams a subagent's turns onto the parent's stdout as well, so the parent
+ * loop's telemetry must filter them (`subagentDispatchId`) for this concatenation to hold.
  */
 async function assembleClaudeOutcome(args) {
     const { summary, stats, stderrTail, calls, publisher, usage, subagents } = args;
@@ -512,6 +528,10 @@ async function assembleClaudeOutcome(args) {
     await subagents?.stop();
     const subUsage = subagents?.usage() ?? { inputTokens: 0, outputTokens: 0 };
     const subCalls = subagents?.calls() ?? [];
+    if (args.expectSubagentCalls && !subCalls.length) {
+        args.log?.warn('subagent turns were streamed but the transcript watcher captured no calls; their token ' +
+            'spend is missing from this run’s telemetry (check the CLI’s subagents/*.jsonl layout)');
+    }
     const mergedCalls = [...calls, ...subCalls];
     const mergedUsage = usage || subUsage.inputTokens || subUsage.outputTokens
         ? {

package/dist/claude-call-aggregator.js

@@ -0,0 +1,229 @@
+import { claudeAssistantContent, claudeCallUsage, isObject, redactBody } from './claude-stream.js';
+/**
+ * Assemble per-call telemetry out of Claude Code's per-block stream envelopes.
+ *
+ * `onCallStart` fires when a call's FIRST envelope arrives, which is the moment the caller must
+ * snapshot the prompt: the history at that point is what produced the response. `onCall` fires
+ * once the call is complete (a different `message.id` began, or the stream ended).
+ *
+ * Usage is merged as the MAXIMUM of each bucket across the call's envelopes rather than the last
+ * one seen. The envelopes carry a snapshot of the same call's usage, and which of them holds the
+ * final output count is a CLI detail we should not depend on; a max is right whether the value is
+ * repeated verbatim or grows.
+ *
+ * An envelope with no `message.id` cannot be attributed, so it is treated as a call of its own —
+ * the pre-aggregation behaviour, kept so a CLI build (or a transcript) that omits the id degrades
+ * to over-counting rather than to silently merging unrelated calls.
+ */
+export function createClaudeCallAggregator(handlers) {
+    let pending;
+    let anonymous = 0;
+    const complete = () => {
+        if (!pending)
+            return;
+        const { id: _id, ...call } = pending;
+        pending = undefined;
+        handlers.onCall(call);
+    };
+    return {
+        onAssistant(message) {
+            // `#anon-<n>` cannot collide with a real id (the API mints `msg_…`), so an envelope
+            // with no id keeps its own call rather than merging into whatever came before it.
+            const id = typeof message.id === 'string' && message.id ? message.id : `#anon-${anonymous++}`;
+            if (pending && pending.id !== id)
+                complete();
+            const content = Array.isArray(message.content) ? message.content : [];
+            const { text, reasoning, toolUses } = claudeAssistantContent(content);
+            const usage = claudeCallUsage(message.usage);
+            const stopReason = typeof message.stop_reason === 'string' ? message.stop_reason : null;
+            const model = typeof message.model === 'string' ? message.model : undefined;
+            if (!pending) {
+                pending = {
+                    id,
+                    content: [],
+                    text: '',
+                    reasoning: '',
+                    stopReason: null,
+                    inputTokens: 0,
+                    cachedInputTokens: 0,
+                    outputTokens: 0,
+                    toolResults: [],
+                    toolUses: 0,
+                };
+                handlers.onCallStart?.();
+            }
+            pending.content.push(...content);
+            pending.text += text;
+            pending.reasoning += reasoning;
+            pending.toolUses += toolUses;
+            pending.inputTokens = Math.max(pending.inputTokens, usage.inputTokens);
+            pending.cachedInputTokens = Math.max(pending.cachedInputTokens, usage.cachedInputTokens);
+            pending.outputTokens = Math.max(pending.outputTokens, usage.outputTokens);
+            // A block-split response reports its stop reason on the envelope that carries the end of the
+            // message; earlier ones report none. Keep the first non-null rather than the last seen.
+            if (stopReason && !pending.stopReason)
+                pending.stopReason = stopReason;
+            if (model && !pending.model)
+                pending.model = model;
+        },
+        onToolResult(content) {
+            // Results can only belong to the tool_use blocks of the call in flight. Before the first
+            // assistant envelope there is nothing they could attach to.
+            if (pending)
+                pending.toolResults.push(content);
+        },
+        flush: complete,
+    };
+}
+/**
+ * Assemble ONE conversation's per-call telemetry from the CLI stream: the growing request
+ * transcript and the per-call token/body metrics.
+ *
+ * Owns the transcript because the two are one concern — a call's `promptText` is the transcript as
+ * of that call, and its turns may only be appended once the call that produced them is complete.
+ * `seed` is what the harness supplied and the stream therefore never shows (the system + first user
+ * message, or the single folded user turn), so the reconstruction never claims a system turn that
+ * was not sent. A subagent's conversation seeds EMPTY — its prompt was minted by the CLI and never
+ * crosses this stream — which is also why its first call carries `messageCount: 0` (the backend's
+ * `latestChainTip` skips those on purpose: there is no re-sendable chain to delta against).
+ * Bodies are credential-scrubbed; they can echo the leased token.
+ *
+ * Deliberately does NOT touch {@link PiRunStats}: the run's tool/output counters describe whether
+ * the agent ACTED at all (`agentNeverActed`), which is true of a subagent's turns whichever channel
+ * ends up owning their telemetry rows. The caller accumulates them off the raw stream instead.
+ */
+export function createClaudeStreamTelemetry(opts) {
+    const messages = [...opts.seed];
+    let callPrompt = '';
+    let callMessageCount = 0;
+    // The aggregator IS the surface: the transcript and metric work happens in its callbacks, so
+    // there is nothing to wrap it in.
+    return createClaudeCallAggregator({
+        // Snapshotted when a call's FIRST envelope arrives: the history at that moment is what
+        // produced the response, and later envelopes of the same call must not see the turns it
+        // went on to add.
+        onCallStart: () => {
+            callPrompt = redactBody(JSON.stringify(messages), opts.secrets);
+            callMessageCount = messages.length;
+        },
+        onCall: (call) => {
+            opts.publish({
+                ...(call.model ? { model: call.model } : {}),
+                promptText: callPrompt,
+                messageCount: callMessageCount,
+                responseText: redactBody(call.text, opts.secrets),
+                reasoningText: redactBody(call.reasoning, opts.secrets),
+                inputTokens: call.inputTokens,
+                cachedInputTokens: call.cachedInputTokens,
+                outputTokens: call.outputTokens,
+                finishReason: call.stopReason,
+            });
+            // Appended only now, so each call's prompt stays a strict prefix of the next and the
+            // backend's telemetry chain delta-compresses cleanly.
+            messages.push({ role: 'assistant', content: call.content });
+            for (const result of call.toolResults)
+                messages.push({ role: 'tool', content: result });
+        },
+    });
+}
+/**
+ * The dispatch (`Agent`/`Task` tool_use) id a stream envelope is tagged with, or `undefined` for a
+ * parent-loop turn.
+ *
+ * Claude Code streams the turns of the subagents it dispatches onto the parent's stdout, tagged
+ * with the tool_use id that spawned them. Those same turns are also written to the per-session
+ * `subagents/*.jsonl` transcripts the watcher reads, so recording both channels counted every
+ * subagent call twice — and splicing them into the parent's message reconstruction produced a
+ * `promptText` chain that interleaves several conversations and therefore matches no real request.
+ *
+ * The id is what makes the fallback below possible: concurrent subagents interleave on one stdout,
+ * so it is the ONLY thing separating their conversations.
+ */
+export function subagentDispatchId(event) {
+    if (!isObject(event))
+        return undefined;
+    const id = event.parent_tool_use_id;
+    return typeof id === 'string' && id ? id : undefined;
+}
+/** Whether a stream envelope describes a SUBAGENT's turn rather than the parent loop's. */
+export function isSubagentEvent(event) {
+    return subagentDispatchId(event) !== undefined;
+}
+/**
+ * Per-call telemetry for the subagents whose turns ride the parent's stdout — the FALLBACK channel,
+ * used only when no `subagents/*.jsonl` watcher will run (see `startSubagentWatcher`, which is
+ * wired only when the CLI has an isolated config home; an `ambientAuth` run has none).
+ *
+ * Without this, filtering tagged events out of the parent's telemetry leaves a subagent-heavy run
+ * with its spend recorded by NEITHER channel — an under-count, which reads as a cheap run and is
+ * the worse failure direction than the double-count the filter exists to fix.
+ *
+ * Each dispatch id gets its OWN transcript, because concurrent subagents interleave arbitrarily on
+ * one stream: folding them into a single chain is exactly the defect this whole module removes,
+ * one level down.
+ */
+function createSubagentStreamTelemetry(opts) {
+    const perDispatch = new Map();
+    const forDispatch = (dispatchId) => {
+        let telemetry = perDispatch.get(dispatchId);
+        if (!telemetry) {
+            // Seeded EMPTY: the CLI minted this subagent's prompt and it never crossed the stream.
+            telemetry = createClaudeStreamTelemetry({
+                seed: [],
+                secrets: opts.secrets,
+                publish: opts.publish,
+            });
+            perDispatch.set(dispatchId, telemetry);
+        }
+        return telemetry;
+    };
+    return {
+        onAssistant: (dispatchId, message) => forDispatch(dispatchId).onAssistant(message),
+        // Only against a dispatch already seen: a result for a subagent whose assistant turns never
+        // reached us has no conversation to attach to, and minting one would publish a call that is
+        // all tool output and no request.
+        onToolResult: (dispatchId, content) => perDispatch.get(dispatchId)?.onToolResult(content),
+        flush: () => {
+            for (const telemetry of perDispatch.values())
+                telemetry.flush();
+        },
+    };
+}
+/**
+ * Assemble a run's per-call telemetry, routing each envelope to the conversation it belongs to.
+ *
+ * The routing is the whole point. A subagent's turns ride the parent's stdout tagged with the
+ * dispatch that spawned them, and they must never join the PARENT's chain — that splice produced a
+ * `promptText` interleaving several conversations, matching no request that was ever sent.
+ *
+ * Who RECORDS them is a separate question, decided once per run rather than per event:
+ * `watcherOwnsSubagents` says a `subagents/*.jsonl` watcher will run, and it is the better source
+ * (it reads the settled transcript, so its usage and stop reason are final). With no watcher — an
+ * `ambientAuth` run has no isolated config home to watch — the tagged turns are recorded here
+ * instead, on per-dispatch transcripts of their own. Dropping them in that case would leave the run
+ * billed by neither channel, and an under-count reads as a cheap run rather than as an error.
+ */
+export function createClaudeRunTelemetry(opts) {
+    const parent = createClaudeStreamTelemetry(opts);
+    const subagents = opts.watcherOwnsSubagents ? undefined : createSubagentStreamTelemetry(opts);
+    let sawSubagentTurn = false;
+    return {
+        onAssistant(dispatchId, message) {
+            if (!dispatchId)
+                return parent.onAssistant(message);
+            sawSubagentTurn = true;
+            subagents?.onAssistant(dispatchId, message);
+        },
+        onToolResult(dispatchId, content) {
+            if (!dispatchId)
+                return parent.onToolResult(content);
+            sawSubagentTurn = true;
+            subagents?.onToolResult(dispatchId, content);
+        },
+        flush() {
+            parent.flush();
+            subagents?.flush();
+        },
+        expectsWatcherCalls: () => opts.watcherOwnsSubagents && sawSubagentTurn,
+    };
+}

package/dist/progress.js

@@ -1,7 +1,8 @@
 import { isObject } from './claude-stream.js';
-// The parent agent's own PLAN, as progress counts. This is one of the two redundant views a
-// pr-reviewer run produces (the other is the parallel-subagent dispatch view in
-// `subagents.ts`); {@link pickProgress} reconciles them.
+// The parent agent's own PLAN, as progress counts. This is one of the two views a pr-reviewer
+// run produces of the same slicing (the other is the parallel-subagent dispatch view in
+// `subagents.ts`). The plan is the INVENTORY, the dispatches are the live STATUS, and
+// {@link mergeProgress} folds them into the one list the board renders.
 //
 // The Claude Code CLI exposes the plan through TWO different tool vocabularies, and which one
 // a run uses depends on the CLI build, not on anything the harness controls:
@@ -190,17 +191,14 @@ export function createTaskPlanTracker() {
     };
 }
 /**
- * Reconcile the redundant views of the same work into the one to surface (ADR 0027 Defect B).
- * A pr-reviewer run has BOTH a parent plan (`TodoWrite` or `TaskCreate`/`TaskUpdate`) and the
- * `SliceTracker`'s subagent-dispatch view. The sequential shape advances the plan; the parallel
- * shape advances ONLY the slice tracker (the reviewer writes its plan once and the parallel
- * subagents report in-flight/complete). Neither alone covers both shapes, and gating the slice
- * tracker off whenever a plan exists (the original behaviour) pinned parallel runs at 0%.
- *
- * So prefer whichever view is further along: more `completed`, then more `inProgress` (an
- * all-pending plan must not beat live in-flight slices), then more `total` (the richer view — a
- * plan can carry an extra "aggregate" entry), else the plan. Pure + total; returns whichever
+ * Reconcile the parent's TWO plan vocabularies (`TodoWrite` snapshots vs the incremental
+ * `TaskCreate`/`TaskUpdate` pair) into one plan. A run uses one or the other, so this is a
+ * genuine either/or: prefer whichever is further along — more `completed`, then more
+ * `inProgress`, then more `total` — else the `TodoWrite` view. Pure + total; returns whichever
  * single input is present when only one is.
+ *
+ * This is NOT how the plan reconciles with the parallel-subagent view — those describe the same
+ * slices from two angles and are MERGED, see {@link mergeProgress}.
  */
 export function pickProgress(todo, slice) {
     if (!todo)
@@ -215,3 +213,114 @@ export function pickProgress(todo, slice) {
         return slice.total > todo.total ? slice : todo;
     return todo;
 }
+/** Status ordering, so a merge can only ever ADVANCE an entry, never walk it back. */
+const STATUS_RANK = {
+    pending: 0,
+    in_progress: 1,
+    completed: 2,
+};
+/**
+ * Words that carry no identity in a slice label, so `Review identity/auth slice` (the subagent
+ * description) and `identity/auth` (the plan entry's subject) compare equal.
+ */
+const LABEL_FILLER = new Set([
+    'a',
+    'agent',
+    'an',
+    'and',
+    'chunk',
+    'chunks',
+    'for',
+    'of',
+    'pass',
+    'review',
+    'reviewing',
+    'slice',
+    'slices',
+    'subagent',
+    'the',
+]);
+/**
+ * A slice label reduced to its identifying words, for pairing a plan entry with the subagent
+ * dispatched to review it. Case, punctuation and the boilerplate around the slice name all
+ * differ between the two vocabularies; the slice NAME does not.
+ */
+export function sliceLabelKey(label) {
+    return label
+        .toLowerCase()
+        .replace(/[^a-z0-9]+/g, ' ')
+        .split(' ')
+        .filter((w) => w.length > 0 && !LABEL_FILLER.has(w))
+        .join(' ');
+}
+/** Advance an entry to the stronger of its current status and the dispatch's. */
+function advance(entry, status) {
+    if (STATUS_RANK[status] > STATUS_RANK[entry.status])
+        entry.status = status;
+    entry.paired = true;
+}
+/**
+ * MERGE the parent's plan with the `SliceTracker`'s subagent-dispatch view into the single list
+ * the board renders (ADR 0027 Defect B, corrected).
+ *
+ * The two are not competing answers, they are two halves of one: the plan is the INVENTORY (it
+ * names every slice, including the ones not dispatched yet, which is the only place a `pending`
+ * slice exists at all), and the dispatch view is the live STATUS (the plan advances only when
+ * the agent remembers to update it, which it does unreliably). Picking whichever looked "further
+ * along" — the previous behaviour — made the rendered list SHRINK the moment the first subagent
+ * returned: the dispatch view won on `completed`, and it only knows the slices dispatched so far,
+ * so every queued slice vanished from the window and reappeared one at a time as it was dispatched.
+ *
+ * Pairing is by normalised label ({@link sliceLabelKey}) — exact first, then containment — and
+ * finally positionally into the leftover pending entries, in dispatch order (the agent dispatches
+ * in plan order). A dispatch that pairs with nothing is APPENDED rather than dropped, so the list
+ * is at worst a union and can never lose a slice. Statuses only ever advance, so a plan entry the
+ * agent already marked done is not walked back by a re-dispatch.
+ *
+ * Pure + total. Falls back to {@link pickProgress} when either side carries counts but no items
+ * (nothing to merge onto).
+ */
+export function mergeProgress(plan, slice) {
+    if (!plan)
+        return slice;
+    if (!slice)
+        return plan;
+    const planItems = plan.items ?? [];
+    const sliceItems = slice.items ?? [];
+    if (planItems.length === 0 || sliceItems.length === 0)
+        return pickProgress(plan, slice);
+    const entries = planItems.map((i) => ({
+        label: i.label,
+        status: normalizeStatus(i.status),
+        key: sliceLabelKey(i.label),
+        paired: false,
+    }));
+    const take = (match) => entries.find((e) => !e.paired && match(e));
+    // Pass 1 — the same slice named the same way.
+    // Pass 2 — one label contains the other (a dispatch description often expands the plan's short
+    //          name). Length-guarded so a one-word residue can't match everything.
+    // Pass 3 — no words in common at all (renamed between planning and dispatch): absorb into the
+    //          still-untouched pending entries in dispatch order.
+    // Anything still unpaired is a slice the plan never mentioned, so it JOINS the list.
+    const matchers = [
+        (key) => (e) => key.length > 0 && e.key === key,
+        (key) => (e) => key.length >= 3 && e.key.length >= 3 && (e.key.includes(key) || key.includes(e.key)),
+        () => (e) => e.status === 'pending',
+    ];
+    let unpaired = sliceItems;
+    for (const matcher of matchers) {
+        const rest = [];
+        for (const item of unpaired) {
+            const hit = take(matcher(sliceLabelKey(item.label)));
+            if (hit)
+                advance(hit, normalizeStatus(item.status));
+            else
+                rest.push(item);
+        }
+        unpaired = rest;
+    }
+    return toProgress([
+        ...entries.map((e) => ({ label: e.label, status: e.status })),
+        ...unpaired.map((i) => ({ label: i.label, status: normalizeStatus(i.status) })),
+    ]);
+}

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@cat-factory/executor-harness",
-  "version": "1.64.0",
+  "version": "1.64.4",
   "description": "Container payload: a thin TypeScript wrapper that runs the Pi coding agent against a cloned repo and opens a PR. Runs in the Cloudflare Container (and, in local native mode, as a host process); carries no secrets.",
   "repository": {
     "type": "git",
@@ -26,9 +26,9 @@
     "hono": "^4.12.32",
     "typescript": "7.0.2",
     "vitest": "^4.1.10",
-    "@cat-factory/kernel": "0.170.0",
-    "@cat-factory/server": "0.160.0",
-    "@cat-factory/spend": "0.12.99"
+    "@cat-factory/kernel": "0.171.0",
+    "@cat-factory/server": "0.162.1",
+    "@cat-factory/spend": "0.12.100"
   },
   "scripts": {
     "build": "tsc -p tsconfig.json",