sync: update 4 file(s) in core/

.gaai/core/adapters/claude-code/__tests__/nested-claude-spawn.test.js

@@ -1189,3 +1189,76 @@ describe('nested-claude-spawn — hard ceiling', () => {
 function _isAliveSelfCheck() {
   try { process.kill(process.pid, 0); return true; } catch { return false; }
 }
+
+// ---------------------------------------------------------------------------
+// Handle identity
+//
+// A handle that cannot name its own story forces consumers to infer identity
+// from the worktree path. These pin the recorded fields, and pin the one thing
+// that must never appear in them.
+// ---------------------------------------------------------------------------
+
+describe('nested-claude-spawn — handle identity', () => {
+
+  afterEach(() => {
+    _resetSpawnFn();
+    _resetLogPath();
+    clearEnv();
+    delete process.env.GAAI_STORY_ID;
+    delete process.env.GAAI_WORKSPACE_ID;
+  });
+
+  test('the in-flight handle names its story, workspace, phase and command', async () => {
+    setValidEnv();
+    const handleDir = useTmpHandleDir();
+    process.env.GAAI_STORY_ID     = 'EXAMPLE-STORY';
+    process.env.GAAI_WORKSPACE_ID = 'ws-example';
+
+    _setSpawnFn(() => createMockChild({ exitCode: 0, stdoutData: '## Implementation\n', delay: 400 }));
+
+    // Deliberately not awaited: the record is written synchronously at spawn, and
+    // it is removed once the run concludes — mid-flight is the only time to read it.
+    const pending = _spawnWithTimerOverride(
+      'a-prompt-that-must-not-be-recorded', '', [],
+      { globalTimeoutMs: 60_000, heartbeatTimeoutMs: 60_000 }
+    );
+
+    const files = readdirSync(handleDir).filter(f => f.endsWith('.json'));
+    assert.equal(files.length, 1, 'exactly one in-flight handle');
+    const raw = readFileSync(join(handleDir, files[0]), 'utf8');
+    const rec = JSON.parse(raw);
+
+    assert.equal(rec.story_id, 'EXAMPLE-STORY');
+    assert.equal(rec.workspace_id, 'ws-example');
+    assert.equal(rec.phase, 'impl');
+    assert.equal(typeof rec.command?.bin, 'string');
+    assert.ok(rec.started_at, 'started_at recorded');
+    assert.ok(rec.last_activity_at, 'last_activity_at recorded');
+
+    // The argv carries the prompt. A handle file is not a place for prompt content.
+    assert.equal(raw.includes('a-prompt-that-must-not-be-recorded'), false,
+      'the prompt must never be persisted into the handle');
+
+    await pending;
+    assert.deepEqual(readdirSync(handleDir).filter(f => f.endsWith('.json')), [],
+      'the handle is cleared once the run concludes');
+  });
+
+  test('reconcile reports the story so callers need not parse a path', () => {
+    const handleDir = useTmpHandleDir();
+    writeFileSync(join(handleDir, 'h.json'), JSON.stringify({
+      trace_id: 'h',
+      pid: process.pid,
+      story_id: 'EXAMPLE-STORY',
+      cwd: '/tmp/some-unrelated-path',
+      started_at: new Date().toISOString(),
+      state: 'running',
+    }), 'utf8');
+
+    const report = reconcileHandles();
+    assert.equal(report.live.length, 1);
+    assert.equal(report.live[0].story_id, 'EXAMPLE-STORY',
+      'story identity survives reconciliation without path inference');
+  });
+
+});

.gaai/core/adapters/claude-code/nested-claude-spawn.js

@@ -40,7 +40,7 @@ import {
   existsSync, appendFileSync, mkdirSync,
   writeFileSync, readFileSync, readdirSync, unlinkSync,
 } from 'node:fs';
-import { dirname, join } from 'node:path';
+import { basename, dirname, join } from 'node:path';
 import { fileURLToPath } from 'node:url';
 import { randomUUID } from 'node:crypto';
 import { logPhase, formatPhaseStdout } from './runtime-routing-logger.js';
@@ -78,6 +78,7 @@ const MAX_TURNS            =        150;
 // reported successful long runs as failures, discarded the session_id that would
 // have allowed correlation, and truncated the output to whatever had arrived.
 const POLL_INTERVAL_MS     =     30_000; // handle probe cadence while in poll mode
+const ACTIVITY_FLUSH_MS    =     10_000; // how often last_activity_at is persisted
 
 /**
  * Poll budget once an observation window expires. Defaults to one extra heartbeat
@@ -720,12 +721,22 @@ export function reconcileHandles({ reap = false } = {}) {
  */
 function spawnCore(prompt, implReportPath, extraArgs, globalTimeoutMs, heartbeatTimeoutMs, logFile,
                    envFn = buildChildEnv, model = 'opus', includeFallbackModel = true, collectTelemetry = false,
-                   cwd = '') {
+                   cwd = '', context = {}) {
   const traceId   = randomUUID();
   const startMs   = Date.now();
   const modelReq  = process.env.GAAI_IMPL_MODEL   || '';
   const baseUrl   = process.env.GAAI_IMPL_BASE_URL || '';
 
+  // What this run belongs to. A handle that cannot name its own story forces
+  // consumers to infer identity from the worktree path, which is a string-shape
+  // guess rather than a fact. Explicit context wins; the daemon exports the same
+  // values into the child env, which covers callers that pass nothing.
+  const identity = {
+    story_id:     context.storyId     || process.env.GAAI_STORY_ID     || null,
+    workspace_id: context.workspaceId || process.env.GAAI_WORKSPACE_ID || null,
+    phase:        context.phase       || 'impl',
+  };
+
   // Log spawn start — never log token values
   console.log(`[nested-claude-spawn] spawn trace_id=${traceId} model=${modelReq} url=${baseUrl}`);
 
@@ -785,21 +796,34 @@ function spawnCore(prompt, implReportPath, extraArgs, globalTimeoutMs, heartbeat
     const streamState  = { sessionId: null, terminalReceipt: null };
     const pollBudgetMs = _pollBudgetMs(heartbeatTimeoutMs);
     let scanBuffer     = '';
+    let lastActivityAt = startMs;
+    let activityFlushAt = 0;
+
+    // Command identity, deliberately narrow: the resolved binary and the model.
+    // The argv is never recorded — it carries the prompt, and a handle file is
+    // not a place to put prompt content.
+    const command = { bin: basename(claudePath), model };
 
     // Handle record — written before any window can expire, so a run is never
     // unrecoverable, and refreshed as soon as the session_id is announced.
     _writeHandle(traceId, {
+      ...identity,
+      command,
       pid:         child.pid ?? null,
       session_id:  null,
       state:       'running',
       started_at:  new Date(startMs).toISOString(),
+      last_activity_at: new Date(startMs).toISOString(),
       log_file:    logFile || null,
       report_path: implReportPath || null,
       cwd:         cwd || null,
     });
 
     function refreshHandle(state, extra = {}) {
       _writeHandle(traceId, {
+        ...identity,
+        command,
+        last_activity_at: new Date(lastActivityAt).toISOString(),
         pid:         child.pid ?? null,
         session_id:  streamState.sessionId,
         state,
@@ -926,6 +950,15 @@ function spawnCore(prompt, implReportPath, extraArgs, globalTimeoutMs, heartbeat
         scanBuffer = scanBuffer.slice(lastNewline + 1);
         if (!hadSession && streamState.sessionId) refreshHandle(inPollMode ? 'polling' : 'running');
       }
+      // Liveness as an observed fact, not an inference from the file's mtime.
+      // Flushed on an interval rather than per chunk: a chatty run emits far too
+      // often to justify a write each time, and consumers only need the field
+      // accurate to the order of the windows they compare it against.
+      lastActivityAt = Date.now();
+      if (lastActivityAt - activityFlushAt >= ACTIVITY_FLUSH_MS) {
+        activityFlushAt = lastActivityAt;
+        refreshHandle(inPollMode ? 'polling' : 'running');
+      }
       exitPollMode();
       resetHeartbeat();
     });
@@ -1183,7 +1216,7 @@ export async function runImpl({ implModelTag, prompt, reportPath, storyId, extra
       prompt, reportPath, extraArgs,
       GLOBAL_TIMEOUT_MS, HEARTBEAT_TIMEOUT_MS, logFile,
       buildChildEnv, 'opus', /* includeFallbackModel */ true, /* collectTelemetry */ true,
-      worktreePath
+      worktreePath, { storyId }
     );
 
     const logFailed = _emitLog({
@@ -1209,7 +1242,7 @@ export async function runImpl({ implModelTag, prompt, reportPath, storyId, extra
         prompt, reportPath, extraArgs,
         GLOBAL_TIMEOUT_MS, HEARTBEAT_TIMEOUT_MS, logFile,
         buildPrimaryChildEnv, 'sonnet', /* includeFallbackModel */ false,
-        /* collectTelemetry */ false, worktreePath
+        /* collectTelemetry */ false, worktreePath, { storyId }
       );
 
       const primaryLogFailed = _emitLog({
@@ -1242,7 +1275,7 @@ export async function runImpl({ implModelTag, prompt, reportPath, storyId, extra
     prompt, reportPath, extraArgs,
     GLOBAL_TIMEOUT_MS, HEARTBEAT_TIMEOUT_MS, logFile,
     buildPrimaryChildEnv, 'sonnet', /* includeFallbackModel */ false,
-    /* collectTelemetry */ false, worktreePath
+    /* collectTelemetry */ false, worktreePath, { storyId }
   );
 
   const logFailed = _emitLog({

.gaai/core/scripts/delivery-daemon.sh

@@ -2103,6 +2103,11 @@ _recovery_relaunch() {
   # running; a live handle rooted in this story's worktree means there is nothing
   # to relaunch yet. Best-effort: any failure here falls through to the relaunch,
   # preserving prior behaviour rather than stalling recovery.
+  #
+  # Matching is on the handle's own story_id. The worktree-path substring below is
+  # only a fallback for handles written before that field existed — inferring a
+  # story from a path shape is a guess, and it is wrong for any layout that does
+  # not embed the id.
   local _handles_json _live_here=""
   _handles_json=$(node "$PROJECT_DIR/.gaai/core/adapters/claude-code/nested-claude-spawn.js" \
     --reconcile-handles 2>/dev/null || true)
@@ -2113,8 +2118,13 @@ try:
     live = json.load(sys.stdin).get("live", [])
 except Exception:
     sys.exit(0)
-needle = sys.argv[1] + "-workspace"
-print("1" if any(needle in (h.get("cwd") or "") for h in live) else "")
+sid = sys.argv[1]
+legacy = sid + "-workspace"
+def matches(h):
+    if h.get("story_id"):
+        return h["story_id"] == sid
+    return legacy in (h.get("cwd") or "")
+print("1" if any(matches(h) for h in live) else "")
 ' "$sid" 2>/dev/null || true)
   fi
   if [[ "$_live_here" == "1" ]]; then

CHANGELOG.md

@@ -10,6 +10,7 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 ## [Unreleased]
 
 ### Changed
+- fix: record run identity on the handle instead of inferring it
 - fix: keep and poll the handle until a terminal receipt
 - fix: make generated execution plans carry artefact frontmatter
 -: Make hosted CI the sole daemon merge authority