@ouro.bot/cli 0.1.0-alpha.3 → 0.1.0-alpha.300

package/dist/heart/daemon/process-manager.js

@@ -51,15 +51,49 @@ class DaemonProcessManager {
     now;
     setTimeoutFn;
     clearTimeoutFn;
+    cooldownRecoveryMs;
+    maxCooldownRetries;
+    existsSyncFn;
+    configCheckFn;
+    onSnapshotChangeFn;
+    /**
+     * Notify the snapshot-change observer (if registered). Swallows any
+     * errors from the observer so process lifecycle code never fails
+     * because the observer threw.
+     */
+    notifySnapshotChange(snapshot) {
+        if (!this.onSnapshotChangeFn)
+            return;
+        try {
+            this.onSnapshotChangeFn(snapshot);
+        }
+        catch (error) {
+            (0, runtime_1.emitNervesEvent)({
+                level: "warn",
+                component: "daemon",
+                event: "daemon.snapshot_change_observer_error",
+                message: "snapshot-change observer threw",
+                meta: {
+                    agent: snapshot.name,
+                    error: error instanceof Error ? error.message : /* v8 ignore next -- defensive: non-Error catch branch @preserve */ String(error),
+                },
+            });
+        }
+    }
     constructor(options) {
         this.maxRestartsPerHour = options.maxRestartsPerHour ?? 10;
         this.stabilityThresholdMs = options.stabilityThresholdMs ?? 60_000;
         this.initialBackoffMs = options.initialBackoffMs ?? 1_000;
         this.maxBackoffMs = options.maxBackoffMs ?? 60_000;
+        this.cooldownRecoveryMs = options.cooldownRecoveryMs ?? 5 * 60 * 1_000;
+        this.maxCooldownRetries = options.maxCooldownRetries ?? 3;
         this.spawnFn = options.spawn ?? ((command, args, spawnOptions) => (0, child_process_1.spawn)(command, args, spawnOptions));
         this.now = options.now ?? (() => Date.now());
         this.setTimeoutFn = options.setTimeoutFn ?? ((cb, delay) => setTimeout(cb, delay));
         this.clearTimeoutFn = options.clearTimeoutFn ?? ((timer) => clearTimeout(timer));
+        this.existsSyncFn = options.existsSync ?? null;
+        this.configCheckFn = options.configCheck ?? null;
+        this.onSnapshotChangeFn = options.onSnapshotChange ?? null;
         for (const agent of options.agents) {
             this.agents.set(agent.name, {
                 config: agent,
@@ -67,6 +101,9 @@ class DaemonProcessManager {
                 restartTimer: null,
                 crashTimestamps: [],
                 stopRequested: false,
+                cooldownTimer: null,
+                cooldownRetryCount: 0,
+                fastCrashCount: 0,
                 snapshot: {
                     name: agent.name,
                     channel: agent.channel,
@@ -76,6 +113,10 @@ class DaemonProcessManager {
                     startedAt: null,
                     lastCrashAt: null,
                     backoffMs: this.initialBackoffMs,
+                    lastExitCode: null,
+                    lastSignal: null,
+                    errorReason: null,
+                    fixHint: null,
                 },
             });
         }
@@ -94,14 +135,62 @@ class DaemonProcessManager {
         this.clearRestartTimer(state);
         state.stopRequested = false;
         state.snapshot.status = "starting";
+        if (this.configCheckFn) {
+            const result = this.configCheckFn(agent);
+            if (!result.ok) {
+                state.snapshot.status = "crashed";
+                // Surface the error and fix to the snapshot so sibling agents can
+                // read it via the pulse. Without this, the diagnosis stayed
+                // trapped in the nerves event and stderr — visible to humans
+                // running `ouro status` or grepping logs, but invisible to
+                // peer agents trying to coordinate around the broken state.
+                state.snapshot.errorReason = result.error ?? "agent config validation failed";
+                state.snapshot.fixHint = result.fix ?? null;
+                (0, runtime_1.emitNervesEvent)({
+                    level: "error",
+                    component: "daemon",
+                    event: "daemon.agent_config_invalid",
+                    message: result.error ?? "agent config validation failed",
+                    meta: { agent, fix: result.fix ?? null },
+                });
+                process.stderr.write(`[daemon] ${agent}: ${result.error}\n` +
+                    (result.fix ? `  Fix: ${result.fix}\n` : ""));
+                this.notifySnapshotChange(state.snapshot);
+                return;
+            }
+            // Config check passed — clear any prior error so the pulse stops
+            // reporting the broken state. This is the recovery path: the user
+            // fixed their secrets/config, the next startAgent attempt sees a
+            // valid config, and the pulse goes quiet.
+            state.snapshot.errorReason = null;
+            state.snapshot.fixHint = null;
+        }
         const runCwd = (0, identity_1.getRepoRoot)();
         const entryScript = path.join((0, identity_1.getRepoRoot)(), "dist", state.config.entry);
-        const args = [entryScript, "--agent", agent, ...(state.config.args ?? [])];
+        if (this.existsSyncFn && !this.existsSyncFn(entryScript)) {
+            state.snapshot.status = "crashed";
+            (0, runtime_1.emitNervesEvent)({
+                level: "error",
+                component: "daemon",
+                event: "daemon.agent_entry_missing",
+                message: "agent entry script does not exist — cannot spawn. Run 'ouro daemon install' from the correct location.",
+                meta: { agent, entryScript },
+            });
+            this.notifySnapshotChange(state.snapshot);
+            return;
+        }
+        const args = [entryScript, "--agent", state.config.agentArg ?? agent, ...(state.config.args ?? [])];
         const child = this.spawnFn("node", args, {
             cwd: runCwd,
             env: state.config.env ? { ...process.env, ...state.config.env } : process.env,
             stdio: ["ignore", "ignore", "ignore", "ipc"],
         });
+        /* v8 ignore next 7 -- defensive: spawn should always return a ChildProcess @preserve */
+        if (!child) {
+            state.snapshot.status = "crashed";
+            (0, runtime_1.emitNervesEvent)({ level: "error", component: "daemon", event: "daemon.agent_spawn_failed", message: "spawn returned null", meta: { agent } });
+            return;
+        }
         state.process = child;
         state.snapshot.status = "running";
         state.snapshot.pid = child.pid ?? null;
@@ -112,6 +201,18 @@ class DaemonProcessManager {
             message: "daemon started managed agent process",
             meta: { agent, pid: child.pid ?? null, cwd: runCwd },
         });
+        this.notifySnapshotChange(state.snapshot);
+        /* v8 ignore start — child process error handler; requires real spawn to trigger */
+        child.on("error", (err) => {
+            (0, runtime_1.emitNervesEvent)({
+                level: "warn",
+                component: "daemon",
+                event: "daemon.agent_process_error",
+                message: "managed agent process emitted error",
+                meta: { agent, error: err.message },
+            });
+        });
+        /* v8 ignore stop */
         child.once("exit", (code, signal) => {
             this.onExit(state, code, signal);
         });
@@ -119,10 +220,12 @@ class DaemonProcessManager {
     async stopAgent(agent) {
         const state = this.requireAgent(agent);
         this.clearRestartTimer(state);
+        this.clearCooldownTimer(state);
         state.stopRequested = true;
         if (!state.process) {
             state.snapshot.status = "stopped";
             state.snapshot.pid = null;
+            this.notifySnapshotChange(state.snapshot);
             return;
         }
         const child = state.process;
@@ -141,6 +244,7 @@ class DaemonProcessManager {
                 meta: { agent },
             });
         }
+        this.notifySnapshotChange(state.snapshot);
     }
     async restartAgent(agent) {
         await this.stopAgent(agent);
@@ -179,6 +283,8 @@ class DaemonProcessManager {
             return;
         state.process = null;
         state.snapshot.pid = null;
+        state.snapshot.lastExitCode = code;
+        state.snapshot.lastSignal = signal;
         const crashed = !state.stopRequested && code !== 0;
         const now = this.now();
         const startedAt = state.snapshot.startedAt ? Date.parse(state.snapshot.startedAt) : now;
@@ -195,13 +301,45 @@ class DaemonProcessManager {
             if (runDuration >= this.stabilityThresholdMs) {
                 state.snapshot.backoffMs = this.initialBackoffMs;
             }
+            this.notifySnapshotChange(state.snapshot);
             return;
         }
         state.snapshot.lastCrashAt = new Date(now).toISOString();
+        // Fast-crash detection: if the agent dies within 5 seconds of starting, it's likely
+        // a configuration issue (missing credentials, bad provider, etc.) not a transient failure.
+        // After 3 consecutive fast crashes, stop retrying and mark as config-failed.
+        const FAST_CRASH_THRESHOLD_MS = 5000;
+        const FAST_CRASH_MAX = 3;
+        if (runDuration < FAST_CRASH_THRESHOLD_MS) {
+            state.fastCrashCount = state.fastCrashCount + 1;
+            if (state.fastCrashCount >= FAST_CRASH_MAX) {
+                state.snapshot.status = "crashed";
+                // Capture the fast-crash diagnosis on the snapshot so it surfaces
+                // via the pulse. The error message is prescriptive: it tells the
+                // user (and their sibling agents) exactly what to do.
+                state.snapshot.errorReason = `agent crashed ${FAST_CRASH_MAX} times within ${FAST_CRASH_THRESHOLD_MS}ms of startup — likely a configuration issue (missing credentials, bad provider).`;
+                state.snapshot.fixHint = `Fix the config and run \`ouro up\` to restart, or check daemon logs for the underlying error.`;
+                (0, runtime_1.emitNervesEvent)({
+                    level: "error",
+                    component: "daemon",
+                    event: "daemon.agent_config_failure",
+                    message: `agent crashed ${FAST_CRASH_MAX} times within ${FAST_CRASH_THRESHOLD_MS}ms of startup — likely a configuration issue (missing credentials, bad provider). Not retrying. Fix the config and run \`ouro up\` to restart.`,
+                    meta: { agent: state.config.name, fastCrashCount: state.fastCrashCount, avgRunDurationMs: runDuration },
+                });
+                this.notifySnapshotChange(state.snapshot);
+                return; // Don't schedule cooldown recovery — this needs human/agent intervention
+            }
+        }
+        else {
+            // Reset fast-crash counter on a stable run
+            state.fastCrashCount = 0;
+        }
         state.crashTimestamps = state.crashTimestamps.filter((crashTs) => crashTs >= startOfHour(now));
         state.crashTimestamps.push(now);
         if (state.crashTimestamps.length > this.maxRestartsPerHour) {
             state.snapshot.status = "crashed";
+            state.snapshot.errorReason = `agent exceeded restart limit (${this.maxRestartsPerHour}/hr) — entering cooldown`;
+            state.snapshot.fixHint = "investigate why the agent keeps crashing; cooldown will retry shortly";
             (0, runtime_1.emitNervesEvent)({
                 level: "error",
                 component: "daemon",
@@ -209,6 +347,8 @@ class DaemonProcessManager {
                 message: "managed agent exceeded restart limit and is marked crashed",
                 meta: { agent: state.config.name, maxRestartsPerHour: this.maxRestartsPerHour },
             });
+            this.notifySnapshotChange(state.snapshot);
+            this.scheduleCooldownRecovery(state);
             return;
         }
         state.snapshot.status = "starting";
@@ -219,6 +359,7 @@ class DaemonProcessManager {
         state.restartTimer = this.setTimeoutFn(() => {
             void this.startAgent(state.config.name);
         }, waitMs);
+        this.notifySnapshotChange(state.snapshot);
     }
     clearRestartTimer(state) {
         if (state.restartTimer === null)
@@ -226,6 +367,45 @@ class DaemonProcessManager {
         this.clearTimeoutFn(state.restartTimer);
         state.restartTimer = null;
     }
+    scheduleCooldownRecovery(state) {
+        if (state.cooldownRetryCount >= this.maxCooldownRetries) {
+            (0, runtime_1.emitNervesEvent)({
+                level: "error",
+                component: "daemon",
+                event: "daemon.agent_permanent_failure",
+                message: "managed agent exhausted all cooldown retries — permanently stopped",
+                meta: { agent: state.config.name, cooldownRetryCount: state.cooldownRetryCount, maxCooldownRetries: this.maxCooldownRetries },
+            });
+            return;
+        }
+        this.clearCooldownTimer(state);
+        state.cooldownTimer = this.setTimeoutFn(() => {
+            state.cooldownRetryCount += 1;
+            state.crashTimestamps = [];
+            state.snapshot.backoffMs = this.initialBackoffMs;
+            state.snapshot.status = "starting";
+            state.snapshot.restartCount += 1;
+            (0, runtime_1.emitNervesEvent)({
+                component: "daemon",
+                event: "daemon.agent_cooldown_recovery",
+                message: "attempting cooldown recovery for managed agent",
+                meta: { agent: state.config.name, cooldownRetryCount: state.cooldownRetryCount },
+            });
+            void this.startAgent(state.config.name);
+        }, this.cooldownRecoveryMs);
+        (0, runtime_1.emitNervesEvent)({
+            component: "daemon",
+            event: "daemon.agent_cooldown_scheduled",
+            message: `scheduled cooldown recovery in ${this.cooldownRecoveryMs}ms`,
+            meta: { agent: state.config.name, cooldownRecoveryMs: this.cooldownRecoveryMs, cooldownRetryCount: state.cooldownRetryCount },
+        });
+    }
+    clearCooldownTimer(state) {
+        if (state.cooldownTimer === null)
+            return;
+        this.clearTimeoutFn(state.cooldownTimer);
+        state.cooldownTimer = null;
+    }
     requireAgent(agent) {
         const state = this.agents.get(agent);
         if (!state) {