@claude-flow/cli 3.7.0-alpha.8 → 3.7.0-alpha.80

package/dist/src/services/worker-daemon.js

@@ -34,6 +34,9 @@ export class WorkerDaemon extends EventEmitter {
     config;
     workers = new Map();
     timers = new Map();
+    // #1845: separate timer for the MCP-dispatch queue poller. Kept off
+    // the per-worker map so stop() clears both kinds without confusion.
+    queuePollTimer;
     running = false;
     startedAt;
     projectRoot;
@@ -76,6 +79,9 @@ export class WorkerDaemon extends EventEmitter {
         };
         // Setup graceful shutdown handlers
         this.setupShutdownHandlers();
+        // #1855: install crash handlers so uncaught exceptions and unhandled
+        // rejections don't leak the PID file or orphan child processes.
+        this.installCrashHandlers();
         // Ensure directories exist
         if (!existsSync(claudeFlowDir)) {
             mkdirSync(claudeFlowDir, { recursive: true });
@@ -101,14 +107,19 @@ export class WorkerDaemon extends EventEmitter {
             this.headlessAvailable = await this.headlessExecutor.isAvailable();
             if (this.headlessAvailable) {
                 this.log('info', 'Claude Code headless mode available - AI workers enabled');
-                // Forward headless executor events
+                // Forward headless executor events. #1855: also snapshot the
+                // active child PIDs to disk on every transition so the next
+                // lifetime can reap orphans after a hard crash.
                 this.headlessExecutor.on('execution:start', (data) => {
+                    this.writeChildrenSnapshot();
                     this.emit('headless:start', data);
                 });
                 this.headlessExecutor.on('execution:complete', (data) => {
+                    this.writeChildrenSnapshot();
                     this.emit('headless:complete', data);
                 });
                 this.headlessExecutor.on('execution:error', (data) => {
+                    this.writeChildrenSnapshot();
                     this.emit('headless:error', data);
                 });
                 this.headlessExecutor.on('output', (data) => {
@@ -169,22 +180,52 @@ export class WorkerDaemon extends EventEmitter {
         return cpus().length || 1;
     }
     /**
-     * Read daemon-specific config from .claude-flow/config.json
-     * Supports dot-notation keys like 'daemon.resourceThresholds.maxCpuLoad'
+     * Read daemon-specific config from .claude-flow/config.{json,yaml,yml}.
+     * Supports dot-notation keys like 'daemon.resourceThresholds.maxCpuLoad'.
+     * #1844: prefer JSON when both exist (existing behavior) but fall back
+     * to YAML so operators using the v3 canonical YAML format aren't silently
+     * ignored. The chosen path is logged at info level.
      */
     readDaemonConfigFromFile(claudeFlowDir) {
-        const configPath = join(claudeFlowDir, 'config.json');
-        if (!existsSync(configPath)) {
-            // Warn if config.yaml exists but config.json does not (#1395 Bug 4)
-            const yamlPath = join(claudeFlowDir, 'config.yaml');
-            const ymlPath = join(claudeFlowDir, 'config.yml');
-            if (existsSync(yamlPath) || existsSync(ymlPath)) {
-                this.log('warn', `Found ${existsSync(yamlPath) ? 'config.yaml' : 'config.yml'} but daemon reads only config.json — YAML config is being ignored. Convert to JSON or create config.json.`);
+        const jsonPath = join(claudeFlowDir, 'config.json');
+        const yamlPath = join(claudeFlowDir, 'config.yaml');
+        const ymlPath = join(claudeFlowDir, 'config.yml');
+        // eslint-disable-next-line @typescript-eslint/no-explicit-any
+        let raw;
+        let chosenPath;
+        if (existsSync(jsonPath)) {
+            try {
+                raw = JSON.parse(readFileSync(jsonPath, 'utf-8'));
+                chosenPath = jsonPath;
+            }
+            catch {
+                return {};
+            }
+        }
+        else if (existsSync(yamlPath) || existsSync(ymlPath)) {
+            const yPath = existsSync(yamlPath) ? yamlPath : ymlPath;
+            try {
+                // Lazy-load yaml so the daemon doesn't hard-require it; if the
+                // dep isn't installed, fall back to the previous warn-only path.
+                // eslint-disable-next-line @typescript-eslint/no-var-requires
+                const yamlMod = require('yaml');
+                const parsed = yamlMod.parse(readFileSync(yPath, 'utf-8'));
+                if (parsed && typeof parsed === 'object') {
+                    // eslint-disable-next-line @typescript-eslint/no-explicit-any
+                    raw = parsed;
+                    chosenPath = yPath;
+                }
             }
+            catch {
+                this.log('warn', `Found ${yPath} but yaml parser unavailable. Install \`yaml\` or convert to JSON. Falling back to defaults.`);
+                return {};
+            }
+        }
+        if (!raw || !chosenPath) {
             return {};
         }
+        this.log('info', `Daemon config loaded from ${chosenPath}`);
         try {
-            const raw = JSON.parse(readFileSync(configPath, 'utf-8'));
             // Support both flat keys at root and nested under scopes.project
             const cfg = raw?.scopes?.project ?? raw;
             const rawCpuLoad = cfg['daemon.resourceThresholds.maxCpuLoad'] ?? raw['daemon.resourceThresholds.maxCpuLoad'];
@@ -216,6 +257,153 @@ export class WorkerDaemon extends EventEmitter {
         process.on('SIGINT', shutdown);
         process.on('SIGHUP', shutdown);
     }
+    /**
+     * #1855: install crash handlers for uncaught exceptions and unhandled
+     * rejections. Without these, a thrown error from any timer callback,
+     * worker logic path, or transitive import crashes the daemon process
+     * silently — the PID file leaks and any in-flight child processes
+     * orphan. With these, we log a structured crash record, run stop()
+     * to clean up, then exit 1 so the process actually dies (otherwise
+     * Node would crash anyway after the handler returns).
+     */
+    installCrashHandlers() {
+        const onCrash = (kind, err) => {
+            // Best-effort logging; never throw from inside the crash handler.
+            try {
+                this.writeCrashRecord(kind, err);
+            }
+            catch { /* nothing more we can do */ }
+            try {
+                // Synchronous stop — don't await; the process is dying. Just
+                // remove the PID file and snapshot state so the next start
+                // sees a clean slate.
+                this.removePidFile();
+                this.saveState();
+                // Snapshot any in-flight child PIDs one last time so the next
+                // lifetime can reap them.
+                this.writeChildrenSnapshot();
+            }
+            catch { /* ignore */ }
+            // Exit non-zero so supervisors / shells see the failure.
+            process.exit(1);
+        };
+        process.on('uncaughtException', (err) => onCrash('uncaughtException', err));
+        process.on('unhandledRejection', (err) => onCrash('unhandledRejection', err));
+    }
+    /**
+     * Append a structured crash record to .claude-flow/logs/crash.log.
+     * Inspectable by hand or via `ruflo daemon status` follow-ups.
+     */
+    writeCrashRecord(kind, err) {
+        const logDir = this.config.logDir;
+        if (!existsSync(logDir))
+            mkdirSync(logDir, { recursive: true });
+        const crashLog = join(logDir, 'crash.log');
+        const ts = new Date().toISOString();
+        const message = err instanceof Error ? err.message : String(err);
+        const stack = err instanceof Error && err.stack ? err.stack : '<no stack>';
+        const record = `[${ts}] [${kind}] pid=${process.pid} ${message}\n${stack}\n---\n`;
+        appendFileSync(crashLog, record, 'utf-8');
+        this.log('warn', `Daemon crashed (${kind}): ${message} — see ${crashLog}`);
+    }
+    /**
+     * Path to the on-disk children registry — list of headless worker
+     * child PIDs the daemon currently owns. #1855: written on every
+     * execution:start / :complete / :error transition; read by the next
+     * lifetime to reap orphans after a hard crash.
+     */
+    get childrenFile() {
+        return join(this.projectRoot, '.claude-flow', 'daemon-children.json');
+    }
+    /**
+     * #1856: detect workers that were mid-flight when the previous daemon
+     * lifetime ended. A mid-flight worker has `lastStartedAt > lastRun`
+     * (started after the last successful completion). On crash recovery
+     * we count these as failures so the run-counter math stays consistent
+     * (`runCount === successCount + failureCount`). Workers naturally
+     * retry at their next scheduled interval; we deliberately don't
+     * immediately re-run because the failure may have been deterministic.
+     */
+    detectMidFlightFailures() {
+        let detected = 0;
+        for (const [type, state] of this.workers.entries()) {
+            const startedAt = state.lastStartedAt?.getTime() ?? 0;
+            const lastRunAt = state.lastRun?.getTime() ?? 0;
+            // started after the last successful completion → was mid-flight
+            if (startedAt > 0 && startedAt > lastRunAt) {
+                state.failureCount++;
+                state.isRunning = false;
+                // Don't bump runCount — it was already incremented at start
+                this.log('info', `Worker ${type} was mid-flight at last crash (started ${state.lastStartedAt?.toISOString()}); counted as failure, will retry at next scheduled interval`);
+                detected++;
+            }
+        }
+        if (detected > 0) {
+            this.saveState();
+        }
+    }
+    /**
+     * Snapshot the currently-active headless worker child PIDs to disk.
+     * Best-effort; failures don't propagate.
+     */
+    writeChildrenSnapshot() {
+        if (!this.headlessExecutor)
+            return;
+        try {
+            const pids = this.headlessExecutor.getActiveChildPids();
+            const dir = join(this.projectRoot, '.claude-flow');
+            if (!existsSync(dir))
+                mkdirSync(dir, { recursive: true });
+            writeFileSync(this.childrenFile, JSON.stringify({ pids, daemonPid: process.pid, timestamp: new Date().toISOString() }, null, 2), 'utf-8');
+        }
+        catch { /* best-effort */ }
+    }
+    /**
+     * #1855: reap orphan headless worker children left behind by a
+     * previous crashed lifetime. Reads `.claude-flow/daemon-children.json`,
+     * SIGTERMs any PID still alive that doesn't belong to the current
+     * daemon, then truncates the file. Called at the top of `start()`
+     * so the next lifetime starts with a clean process tree.
+     */
+    reapOrphanedChildren() {
+        const file = this.childrenFile;
+        if (!existsSync(file))
+            return;
+        let snapshot;
+        try {
+            snapshot = JSON.parse(readFileSync(file, 'utf-8'));
+        }
+        catch {
+            try {
+                unlinkSync(file);
+            }
+            catch { /* ignore */ }
+            return;
+        }
+        const pids = Array.isArray(snapshot.pids) ? snapshot.pids : [];
+        let reaped = 0;
+        for (const pid of pids) {
+            if (typeof pid !== 'number' || pid <= 0)
+                continue;
+            if (pid === process.pid)
+                continue; // never our own PID
+            try {
+                process.kill(pid, 0); // is alive?
+                process.kill(pid, 'SIGTERM');
+                reaped++;
+            }
+            catch {
+                // already dead — fine
+            }
+        }
+        if (reaped > 0) {
+            this.log('info', `Reaped ${reaped} orphan headless worker child(ren) from previous lifetime`);
+        }
+        try {
+            unlinkSync(file);
+        }
+        catch { /* ignore */ }
+    }
     /**
      * Check if system resources allow worker execution
      */
@@ -299,12 +487,14 @@ export class WorkerDaemon extends EventEmitter {
                     for (const [type, state] of Object.entries(saved.workers)) {
                         const savedState = state;
                         const lastRunValue = savedState.lastRun;
+                        const lastStartedAtValue = savedState.lastStartedAt;
                         this.workers.set(type, {
                             runCount: savedState.runCount || 0,
                             successCount: savedState.successCount || 0,
                             failureCount: savedState.failureCount || 0,
                             averageDurationMs: savedState.averageDurationMs || 0,
                             lastRun: lastRunValue ? new Date(lastRunValue) : undefined,
+                            lastStartedAt: lastStartedAtValue ? new Date(lastStartedAtValue) : undefined,
                             nextRun: undefined,
                             isRunning: false,
                         });
@@ -337,6 +527,15 @@ export class WorkerDaemon extends EventEmitter {
     /**
      * Check if another daemon instance is already running.
      * Returns the existing PID if alive, or null if no daemon is running.
+     *
+     * #1853: ignore self-PID matches. The detached-spawn path in
+     * `commands/daemon.ts` writes the child's PID into the file as a
+     * fallback after a 500ms wait. If the child reaches `start()` slower
+     * than the parent's 500ms wait (observed on Node 25 / macOS 26), the
+     * child reads its own PID back from the file and concludes "another
+     * daemon is already running" — so it exits before scheduling workers
+     * and `daemon status` reports STOPPED forever. A daemon process is
+     * never "another instance" of itself; treat self-match as absence.
      */
     checkExistingDaemon() {
         if (!existsSync(this.pidFile))
@@ -345,6 +544,10 @@ export class WorkerDaemon extends EventEmitter {
             const pid = parseInt(readFileSync(this.pidFile, 'utf-8').trim(), 10);
             if (isNaN(pid))
                 return null;
+            // #1853: a PID file containing our own PID is not "another daemon".
+            // Treat as absent so the start() path proceeds normally.
+            if (pid === process.pid)
+                return null;
             // Check if process is alive (signal 0 = existence check)
             process.kill(pid, 0);
             return pid; // Process is alive
@@ -391,6 +594,18 @@ export class WorkerDaemon extends EventEmitter {
             this.emit('warning', `Daemon already running (PID: ${existingPid})`);
             return;
         }
+        // #1855: reap orphan headless worker children left by a previous
+        // crashed lifetime, BEFORE we mark ourselves running and start
+        // accepting new work. The children file from the prior daemon's
+        // last-snapshot is the authoritative list.
+        this.reapOrphanedChildren();
+        // #1856: detect workers that were mid-flight at the previous crash
+        // and count them as failures so runCount/successCount/failureCount
+        // stay consistent. Workers retry naturally at their next scheduled
+        // interval — we don't immediately re-run them, which avoids a
+        // freshly-recovered daemon hammering the same code path that just
+        // killed it.
+        this.detectMidFlightFailures();
         this.running = true;
         this.startedAt = new Date();
         this.writePidFile();
@@ -401,10 +616,90 @@ export class WorkerDaemon extends EventEmitter {
                 this.scheduleWorker(workerConfig);
             }
         }
+        // #1845: poll the MCP-dispatch queue directory so workers requested
+        // via mcp__hooks_worker-dispatch (in a separate process) actually
+        // execute here. Previously the dispatch wrote to a process-local Map
+        // that the daemon could never see.
+        this.queuePollTimer = setInterval(() => {
+            void this.processDispatchQueue();
+        }, 5_000);
+        if (typeof this.queuePollTimer.unref === 'function') {
+            this.queuePollTimer.unref();
+        }
         // Save state
         this.saveState();
         this.log('info', `Daemon started (PID: ${process.pid}, CPUs: ${cpus().length}, workers: ${this.config.workers.filter(w => w.enabled).length}, maxCpuLoad: ${this.config.resourceThresholds.maxCpuLoad}, minFreeMemoryPercent: ${this.config.resourceThresholds.minFreeMemoryPercent}%)`);
     }
+    /**
+     * #1845: ingest queue entries written by mcp__hooks_worker-dispatch.
+     * Each entry is a JSON file at `.claude-flow/daemon-queue/<id>.json`
+     * with `{ workerId, trigger, context, enqueuedAt }`. We move processed
+     * files to `.claude-flow/daemon-queue/.processed/` so the daemon never
+     * re-runs the same dispatch and operators can inspect history.
+     */
+    async processDispatchQueue() {
+        if (!this.running)
+            return;
+        const queueDir = join(this.projectRoot, '.claude-flow', 'daemon-queue');
+        if (!existsSync(queueDir))
+            return;
+        let entries;
+        try {
+            const fs = await import('fs');
+            entries = fs.readdirSync(queueDir).filter((n) => n.endsWith('.json'));
+        }
+        catch {
+            return;
+        }
+        if (entries.length === 0)
+            return;
+        const fs = await import('fs');
+        const processedDir = join(queueDir, '.processed');
+        if (!existsSync(processedDir)) {
+            try {
+                fs.mkdirSync(processedDir, { recursive: true });
+            }
+            catch { /* race ok */ }
+        }
+        for (const entry of entries) {
+            const src = join(queueDir, entry);
+            // eslint-disable-next-line @typescript-eslint/no-explicit-any
+            let payload;
+            try {
+                payload = JSON.parse(fs.readFileSync(src, 'utf-8'));
+            }
+            catch {
+                // Malformed entry — quarantine so we don't loop on it
+                try {
+                    fs.renameSync(src, join(processedDir, `bad-${entry}`));
+                }
+                catch { /* nothing more we can do */ }
+                continue;
+            }
+            const trigger = payload?.trigger;
+            const workerId = payload?.workerId;
+            if (!trigger || !this.config.workers.some((w) => w.type === trigger)) {
+                try {
+                    fs.renameSync(src, join(processedDir, `unknown-${entry}`));
+                }
+                catch { /* ok */ }
+                continue;
+            }
+            try {
+                this.log('info', `Dequeued ${trigger}${workerId ? ` (id=${workerId})` : ''} from MCP dispatch queue`);
+                await this.triggerWorker(trigger);
+            }
+            catch (err) {
+                this.log('warn', `Queued worker ${trigger} failed: ${err.message}`);
+            }
+            finally {
+                try {
+                    fs.renameSync(src, join(processedDir, entry));
+                }
+                catch { /* ignore */ }
+            }
+        }
+    }
     /**
      * Stop the daemon and all workers
      */
@@ -420,6 +715,11 @@ export class WorkerDaemon extends EventEmitter {
             this.log('info', `Stopped worker: ${type}`);
         }
         this.timers.clear();
+        // #1845: stop the MCP-dispatch queue poller too.
+        if (this.queuePollTimer) {
+            clearInterval(this.queuePollTimer);
+            this.queuePollTimer = undefined;
+        }
         this.running = false;
         this.removePidFile();
         this.saveState();
@@ -500,6 +800,8 @@ export class WorkerDaemon extends EventEmitter {
         // Track running worker
         this.runningWorkers.add(workerConfig.type);
         state.isRunning = true;
+        state.lastStartedAt = new Date(); // #1856: timestamp the start
+        this.saveState(); // persist before we run anything
         this.emit('worker:start', { workerId, type: workerConfig.type });
         this.log('info', `Starting worker: ${workerConfig.type} (${this.runningWorkers.size}/${this.config.maxConcurrent} concurrent)`);
         try {
@@ -607,6 +909,17 @@ export class WorkerDaemon extends EventEmitter {
             try {
                 this.log('info', `Running ${workerConfig.type} in headless mode (Claude Code AI)`);
                 const result = await this.headlessExecutor.execute(workerConfig.type);
+                // #1793: persist the headless result to the same metrics files the
+                // local workers write to. Without this, AI-mode runs produced rich
+                // parsedOutput that lived only in `.claude-flow/logs/headless/*` and
+                // never reached `.claude-flow/metrics/<name>.json` — `memory stats`
+                // and downstream consumers saw nothing despite successful runs.
+                try {
+                    this.persistHeadlessResult(workerConfig.type, result);
+                }
+                catch (persistError) {
+                    this.log('warn', `Failed to persist headless result for ${workerConfig.type}: ${persistError.message}`);
+                }
                 return {
                     mode: 'headless',
                     ...result,
@@ -651,6 +964,53 @@ export class WorkerDaemon extends EventEmitter {
                 return { status: 'unknown worker type', mode: 'local' };
         }
     }
+    /**
+     * #1793: persist a headless worker result to the same metrics file the
+     * local fallback writes to. Without this, AI-mode workers produced rich
+     * structured output (audit findings, perf signals, test-gap analysis)
+     * that lived only in `.claude-flow/logs/headless/*_result.log` and was
+     * invisible to `npx ruflo memory stats` or the metrics consumers.
+     *
+     * The mapping mirrors the `*Local` worker implementations below so a
+     * single consumer path works regardless of execution mode.
+     */
+    persistHeadlessResult(workerType, result) {
+        const metricsDir = join(this.projectRoot, '.claude-flow', 'metrics');
+        if (!existsSync(metricsDir))
+            mkdirSync(metricsDir, { recursive: true });
+        // Filename mirrors the local-mode worker writes (security-audit.json,
+        // performance.json, test-gaps.json) so a downstream reader doesn't
+        // care which mode produced the data.
+        const filenameMap = {
+            audit: 'security-audit.json',
+            optimize: 'performance.json',
+            testgaps: 'test-gaps.json',
+            document: 'documentation.json',
+            refactor: 'refactor.json',
+            deepdive: 'deepdive.json',
+            ultralearn: 'ultralearn.json',
+            predict: 'predictions.json',
+        };
+        const filename = filenameMap[workerType] ?? `${workerType}.json`;
+        const metricsFile = join(metricsDir, filename);
+        const persisted = {
+            timestamp: result.timestamp instanceof Date ? result.timestamp.toISOString() : new Date().toISOString(),
+            mode: 'headless',
+            workerType,
+            model: result.model,
+            durationMs: result.durationMs,
+            tokensUsed: result.tokensUsed,
+            executionId: result.executionId,
+            success: result.success,
+            // Structured findings live here when the worker emits JSON (e.g. the
+            // audit worker's vulnerability list). Fall back to a raw-output
+            // pointer so consumers can still locate the full log.
+            findings: result.parsedOutput ?? null,
+            rawOutputPreview: typeof result.output === 'string' ? result.output.slice(0, 2000) : undefined,
+            rawOutputLength: typeof result.output === 'string' ? result.output.length : 0,
+        };
+        writeFileSync(metricsFile, JSON.stringify(persisted, null, 2));
+    }
     // Worker implementations
     async runMapWorker() {
         // Scan project structure and update metrics
@@ -893,6 +1253,7 @@ export class WorkerDaemon extends EventEmitter {
                 {
                     ...state,
                     lastRun: state.lastRun?.toISOString(),
+                    lastStartedAt: state.lastStartedAt?.toISOString(),
                     nextRun: state.nextRun?.toISOString(),
                 }
             ])),