claude-code-session-manager 0.8.0 → 0.8.2

package/src/main/scheduler.cjs

@@ -24,7 +24,8 @@
  *     frees in that group.
  *
  * Execution:
- *   - `claude -p "<PRD body>" --dangerously-skip-permissions` per PRD.
+ *   - `claude -p "<PRD body>" --model sonnet --dangerously-skip-permissions`
+ *     per PRD. Backlog runs on Sonnet; interactive sessions stay on Opus.
  *   - Stdout/stderr → runs/<ts>/<slug>.log; meta json gets exit + duration.
  *   - PRD frontmatter `cwd` → child cwd. Default: PROJECT_CWD const below.
  *
@@ -46,11 +47,20 @@ const os = require('node:os');
 const { spawn } = require('node:child_process');
 const { ipcMain } = require('electron');
 const billing = require('./usage.cjs');
+const { cleanChildEnv } = require('./lib/cleanEnv.cjs');
+const {
+  POLL_INTERVAL_MS,
+  USAGE_REFRESH_INTERVAL_MS,
+  MAX_JOB_DURATION_MS,
+} = require('./lib/schedulerConfig.cjs');
 
 const ROOT = path.join(os.homedir(), '.claude', 'session-manager', 'scheduled-plans');
 const PRDS_DIR = path.join(ROOT, 'prds');
 const RUNS_DIR = path.join(ROOT, 'runs');
 const QUEUE_PATH = path.join(ROOT, 'queue.json');
+const SCHEDULER_STATE_PATH = path.join(os.homedir(), '.claude', 'session-manager', 'scheduler-state.json');
+const HEARTBEAT_PATH = path.join(os.homedir(), '.claude', 'session-manager', 'scheduler-heartbeat.log');
+const HEARTBEAT_MAX_BYTES = 1024 * 1024;
 const DEFAULT_PROJECT_CWD = path.join(os.homedir(), 'Projects', 'session-manager');
 
 const DEFAULT_CONFIG = {
@@ -61,7 +71,7 @@ const DEFAULT_CONFIG = {
   defaultCwd: DEFAULT_PROJECT_CWD,
   // 'when-available' = poll usage and fire whenever utilization < threshold.
   // 'on-reset'        = fire offsetMinutes after the next 5h reset (legacy).
-  // 'manual'          = only fire on explicit Run-now click.
+  // 'manual'          = only fire on explicit Run now click.
   firePolicy: 'when-available',
   // For 'when-available'. Fire only when five_hour utilization < this percent.
   utilizationThreshold: 90,
@@ -76,11 +86,61 @@ function ensureDirs() {
 }
 
 function atomicWriteJson(p, data) {
-  const tmp = `${p}.${process.pid}.tmp`;
+  const tmp = `${p}.${process.pid}.${Date.now()}.${Math.random().toString(36).slice(2, 8)}.tmp`;
   fs.writeFileSync(tmp, JSON.stringify(data, null, 2));
   fs.renameSync(tmp, p);
 }
 
+// ---------- scheduler-state.json (sidecar) ----------
+
+function loadSchedulerState() {
+  try {
+    const raw = fs.readFileSync(SCHEDULER_STATE_PATH, 'utf8');
+    const s = JSON.parse(raw);
+    if (s.lastObservedReset) cachedNextReset = s.lastObservedReset;
+    if (typeof s.consecutiveFailures === 'number') consecutiveFailures = s.consecutiveFailures;
+    if (typeof s.backoffMs === 'number') backoffMs = s.backoffMs;
+    if (typeof s.pauseClearedManuallyAt === 'number') pauseClearedManuallyAt = s.pauseClearedManuallyAt;
+    if (typeof s.lastPollAt === 'number') lastPollAt = s.lastPollAt;
+  } catch { /* first boot or corrupt — start fresh */ }
+}
+
+function persistSchedulerState() {
+  try {
+    atomicWriteJson(SCHEDULER_STATE_PATH, {
+      version: 1,
+      lastObservedReset: cachedNextReset,
+      lastResetObservedAt: cachedNextReset ? Date.now() : null,
+      lastPollAt,
+      consecutiveFailures,
+      backoffMs,
+      pausedReason: null,
+      pausedSince: null,
+      pauseClearedManuallyAt,
+    });
+  } catch (e) {
+    console.warn('[scheduler] failed to persist scheduler state', e?.message);
+  }
+}
+
+// ---------- heartbeat log ----------
+
+function appendHeartbeat(entry) {
+  try {
+    const line = JSON.stringify(entry) + '\n';
+    let size = 0;
+    try { size = fs.statSync(HEARTBEAT_PATH).size; } catch { /* new file */ }
+    if (size >= HEARTBEAT_MAX_BYTES) {
+      const rotated = HEARTBEAT_PATH + '.1';
+      try { fs.unlinkSync(rotated); } catch { /* */ }
+      try { fs.renameSync(HEARTBEAT_PATH, rotated); } catch { /* */ }
+    }
+    fs.appendFileSync(HEARTBEAT_PATH, line);
+  } catch (e) {
+    console.warn('[scheduler] heartbeat write failed', e?.message);
+  }
+}
+
 function readQueue() {
   try {
     const raw = fs.readFileSync(QUEUE_PATH, 'utf8');
@@ -102,6 +162,25 @@ function writeQueue(state) {
   atomicWriteJson(QUEUE_PATH, state);
 }
 
+// ---------- serialized mutation queue ----------
+
+// All read-modify-write operations on queue.json go through mutate() so
+// concurrent job completions in a parallel wave cannot lose each other's
+// status updates. mutateTail is always a resolved promise even when the
+// preceding mutate threw, so the chain never deadlocks.
+let mutateTail = Promise.resolve();
+
+function mutate(fn) {
+  const next = mutateTail.then(async () => {
+    const state = readQueue();
+    const ret = await fn(state);
+    writeQueue(state);
+    return ret;
+  });
+  mutateTail = next.catch(() => {}); // keep chain alive on errors
+  return next;
+}
+
 // ---------- PRD parsing ----------
 
 /**
@@ -220,32 +299,41 @@ function reconcile(state) {
 
 // ---------- next-reset detection ----------
 
-let cachedNextReset = null;
+let cachedNextReset = null; // bare ISO string or null
 let cachedUtilization = null; // five_hour utilization %, 0–100, or null if unknown
 
+/** Fetches latest usage from billing API. Throws on any error — callers handle it. */
 async function refreshNextReset() {
-  try {
-    const r = await billing.fetchUsage();
-    cachedNextReset = r?.usage?.five_hour?.resets_at ?? null;
-    cachedUtilization = r?.usage?.five_hour?.utilization ?? cachedUtilization;
-    return cachedNextReset;
-  } catch {
-    return cachedNextReset;
-  }
+  const r = await billing.fetchUsage();
+  if (r.kind !== 'ok') throw new Error(`usage fetch failed (${r.kind}): ${r.message ?? ''}`);
+  cachedNextReset = r.data?.usage?.five_hour?.resets_at ?? null;
+  cachedUtilization = r.data?.usage?.five_hour?.utilization ?? cachedUtilization;
+  return cachedNextReset;
 }
 
 function getNextResetCached() {
   return cachedNextReset;
 }
 
+// ---------- health / poll state ----------
+
+let bootedAt = Date.now();
+let lastPollAt = null;
+let lastPollOk = false;
+let consecutiveFailures = 0;
+let backoffMs = 0;
+let backoffNextAt = null;
+let firstFailureAt = null;
+let pauseClearedManuallyAt = null;
+
 // ---------- timer ----------
 
 let mainWindow = null;
 let fireTimer = null;
 let resumeTimer = null;
-let pollTimer = null;
+let pollLoopTimer = null;
 let rescheduleInterval = null;
-let initialPollTimeout = null;
+let heartbeatInterval = null;
 let isExecuting = false;
 let cancelToken = { cancelled: false };
 let claudeBinPathCached = null;
@@ -288,65 +376,83 @@ function computeFireAt(state, nextResetIso) {
 
 async function rescheduleTimer() {
   clearFireTimer();
-  const state = readQueue();
-  reconcile(state);
-  const nextResetIso = await refreshNextReset();
-  const fireAt = computeFireAt(state, nextResetIso);
-  if (!fireAt) {
-    state.scheduledFor = null;
-    writeQueue(state);
-    broadcast();
-    return;
+  // Wrap in try/catch — on failure use the cached value so the on-reset
+  // timer can still be armed from the last known reset.
+  let nextResetIso;
+  try {
+    nextResetIso = await refreshNextReset();
+  } catch {
+    nextResetIso = cachedNextReset;
   }
-
-  state.scheduledFor = new Date(fireAt).toISOString();
-  writeQueue(state);
+  const fireAt = await mutate((state) => {
+    reconcile(state);
+    const fa = computeFireAt(state, nextResetIso);
+    state.scheduledFor = fa ? new Date(fa).toISOString() : null;
+    return fa;
+  });
   broadcast();
+  if (!fireAt) return;
 
   const delay = Math.max(1000, fireAt - Date.now());
-  // setTimeout caps at int32 ms (~24.8 days) — well above our 5h horizon, so
-  // a single timer is fine. If reset_at is wildly in the future we'd still
-  // re-anchor on the next billing refresh.
   fireTimer = setTimeout(() => { runDueJobs().catch(() => {}); }, delay);
-  console.log(`[scheduler] next fire in ${Math.round(delay / 1000)}s @ ${state.scheduledFor}`);
+  console.log(`[scheduler] next fire in ${Math.round(delay / 1000)}s @ ${new Date(fireAt).toISOString()}`);
 }
 
 // ---------- pause / resume ----------
 
-function setPaused(reason, resumeAtIso) {
-  const s = readQueue();
-  if (s.paused && s.paused.reason === reason) {
-    if (resumeAtIso) s.paused.resumeAt = resumeAtIso;
-  } else {
-    s.paused = { reason, since: new Date().toISOString(), resumeAt: resumeAtIso || null };
+async function setPaused(reason, resumeAtIso) {
+  // Honor manual-override cooldown: if the user cleared a pause within the
+  // last 5 minutes, suppress auto-pause re-engagement on the same condition.
+  if (pauseClearedManuallyAt && Date.now() - pauseClearedManuallyAt < 300_000) {
+    console.log(`[scheduler] setPaused(${reason}) suppressed by manual override cooldown`);
+    return;
+  }
+
+  // For 'network' with no explicit resumeAt, auto-resume after 30 minutes.
+  let effectiveResumeAt = resumeAtIso;
+  if (reason === 'network' && !resumeAtIso) {
+    effectiveResumeAt = new Date(Date.now() + 30 * 60_000).toISOString();
   }
-  writeQueue(s);
+
+  await mutate((s) => {
+    if (s.paused && s.paused.reason === reason) {
+      if (effectiveResumeAt) s.paused.resumeAt = effectiveResumeAt;
+    } else {
+      s.paused = { reason, since: new Date().toISOString(), resumeAt: effectiveResumeAt || null };
+    }
+  });
   broadcast();
   cancelToken.cancelled = true;
   if (resumeTimer) { clearTimeout(resumeTimer); resumeTimer = null; }
-  if (!resumeAtIso) return;
+  if (!effectiveResumeAt) return;
+
   // Resume 30s after the reset to give the auth/billing endpoint time to flip.
-  const delay = Math.max(30_000, new Date(resumeAtIso).getTime() - Date.now() + 30_000);
+  const delay = Math.max(30_000, new Date(effectiveResumeAt).getTime() - Date.now() + 30_000);
   if (delay > 0x7fffffff) {
     console.warn(`[scheduler] paused (${reason}); resumeAt too far for setTimeout (${delay}ms)`);
     return;
   }
-  resumeTimer = setTimeout(() => {
-    clearPause('resume-timer');
+  resumeTimer = setTimeout(async () => {
+    await clearPause('resume-timer');
     runDueJobs().catch(() => {});
   }, delay);
-  console.log(`[scheduler] paused (${reason}); auto-resume in ${Math.round(delay/1000)}s`);
+  console.log(`[scheduler] paused (${reason}); auto-resume in ${Math.round(delay / 1000)}s`);
 }
 
-function clearPause(source) {
+async function clearPause(source) {
   if (resumeTimer) { clearTimeout(resumeTimer); resumeTimer = null; }
-  const s = readQueue();
-  if (s.paused) {
+  const wasPaused = await mutate((s) => {
+    if (!s.paused) return false;
     console.log(`[scheduler] clearPause (${source || 'manual'})`);
     s.paused = null;
-    writeQueue(s);
-    broadcast();
+    return true;
+  });
+  // Track manual clears for the auto-pause cooldown.
+  if (source === 'manual' || source === 'run-now') {
+    pauseClearedManuallyAt = Date.now();
+    persistSchedulerState();
   }
+  if (wasPaused) broadcast();
 }
 
 /** Mutate a job in place to "pending" with cleared run metadata. */
@@ -357,6 +463,7 @@ function resetJobFields(job, errorMsg) {
   job.finishedAt = null;
   job.exitCode = null;
   job.error = errorMsg ?? null;
+  delete job.runtime;
 }
 
 /** Scan the tail of a job's log for the canonical rate-limit signal. We look
@@ -407,15 +514,34 @@ function pickRunDir() {
   return { runId: ts, dir };
 }
 
-async function executeJob(job, runDir, defaultCwd) {
+/**
+ * Execute a single PRD job. Writes stdout/stderr to a log file and a meta
+ * JSON sidecar. Accepts an optional onPid(pid) callback called synchronously
+ * after spawn so callers can persist the pid before the job finishes.
+ */
+async function executeJob(job, runDir, defaultCwd, onPid) {
   const logPath = path.join(runDir, `${job.slug}.log`);
   const metaPath = path.join(runDir, `${job.slug}.meta.json`);
   const cwd = job.cwd || defaultCwd;
   const startedAt = Date.now();
 
   const fd = fs.openSync(logPath, 'a');
+  let fdClosed = false;
+  const closeFd = () => { if (fdClosed) return; fdClosed = true; fs.closeSync(fd); };
+
   fs.writeSync(fd, `[scheduler] starting ${job.slug} at ${new Date().toISOString()}\n[scheduler] cwd=${cwd}\n\n`);
 
+  // Dead-cwd guard: verify the target directory exists and is traversable
+  // before handing it to the child process.
+  try { fs.accessSync(cwd, fs.constants.X_OK); }
+  catch {
+    const errMsg = `cwd no longer exists: ${cwd}`;
+    fs.writeSync(fd, `[scheduler] ${errMsg}\n`);
+    closeFd();
+    atomicWriteJson(metaPath, { slug: job.slug, cwd, exitCode: -1, error: errMsg, startedAt, finishedAt: Date.now(), durationMs: 0 });
+    return { exitCode: -1, durationMs: 0, error: errMsg };
+  }
+
   // Read full PRD body fresh from disk (queue stored only the preview).
   let prompt;
   try {
@@ -423,37 +549,54 @@ async function executeJob(job, runDir, defaultCwd) {
     prompt = parsed.body;
   } catch (e) {
     fs.writeSync(fd, `[scheduler] failed to read PRD: ${e?.message}\n`);
-    fs.closeSync(fd);
+    closeFd();
     return { exitCode: -1, durationMs: 0, error: e?.message };
   }
 
   return await new Promise((resolve) => {
     const claudeBin = resolveClaudeBin();
+    // Strip Claude Code env and secrets that leak in when session-manager is
+    // launched from a `claude` shell. CLAUDE_EFFORT=xhigh forces Opus and
+    // overrides `--model sonnet`, so scheduled jobs burn Opus credits silently.
+    const childEnv = cleanChildEnv();
     const child = spawn(claudeBin, [
       '-p', prompt,
+      '--model', 'sonnet',
       '--dangerously-skip-permissions',
       '--output-format', 'stream-json',
       '--verbose',
     ], {
       cwd,
-      env: process.env,
+      env: childEnv,
       stdio: ['ignore', fd, fd],
     });
 
     fs.writeSync(fd, `[scheduler] spawned pid=${child.pid}\n\n`);
 
+    // Fire-and-forget pid persistence — best effort.
+    if (onPid) onPid(child.pid).catch(() => {});
+
+    // Kill the child if it runs past the maximum allowed duration.
+    const watchdog = setTimeout(() => {
+      fs.writeSync(fd, `\n[scheduler] watchdog SIGKILL after ${MAX_JOB_DURATION_MS}ms\n`);
+      try { child.kill('SIGKILL'); } catch { /* already dead */ }
+    }, MAX_JOB_DURATION_MS);
+    if (watchdog.unref) watchdog.unref();
+
     child.on('error', (err) => {
+      clearTimeout(watchdog);
       const durationMs = Date.now() - startedAt;
-      fs.writeSync(fd, `\n[scheduler] spawn error: ${err.message}\n`);
-      fs.closeSync(fd);
+      fs.writeSync(fd, `\n[scheduler] error: ${err.message}\n`);
+      closeFd();
       atomicWriteJson(metaPath, { slug: job.slug, cwd, exitCode: -1, error: err.message, startedAt, finishedAt: Date.now(), durationMs });
       resolve({ exitCode: -1, durationMs, error: err.message });
     });
 
     child.on('exit', (code) => {
+      clearTimeout(watchdog);
       const durationMs = Date.now() - startedAt;
       fs.writeSync(fd, `\n[scheduler] exit code=${code} duration=${Math.round(durationMs / 1000)}s\n`);
-      fs.closeSync(fd);
+      closeFd();
       const rateLimited = code !== 0 && detectRateLimitInLog(logPath);
       atomicWriteJson(metaPath, { slug: job.slug, cwd, exitCode: code, rateLimited, startedAt, finishedAt: Date.now(), durationMs });
       resolve({ exitCode: code, durationMs, rateLimited });
@@ -477,7 +620,6 @@ async function runDueJobs() {
       return;
     }
     const { runId, dir: runDir } = pickRunDir();
-    state.lastRunAt = new Date().toISOString();
 
     // Group by parallelGroup, ascending. Each group runs serially after the
     // previous group completes.
@@ -489,7 +631,7 @@ async function runDueJobs() {
     }
     const groupKeys = Array.from(groups.keys()).sort((a, b) => a - b);
 
-    writeQueue(state);
+    await mutate((s) => { s.lastRunAt = new Date().toISOString(); });
     broadcast();
 
     for (const gk of groupKeys) {
@@ -501,41 +643,59 @@ async function runDueJobs() {
       const inFlight = new Set();
 
       const launch = (job) => {
-        const s = readQueue();
-        const idx = s.jobs.findIndex((x) => x.slug === job.slug);
-        if (idx >= 0) {
-          s.jobs[idx].status = 'running';
-          s.jobs[idx].runId = runId;
-          s.jobs[idx].startedAt = new Date().toISOString();
-          writeQueue(s);
-          broadcast();
-        }
-        const promise = executeJob(job, runDir, state.config.defaultCwd).then(async (res) => {
-          // Rate-limit OR pause-already-set means we treat this job as
-          // unfinished — bounce it back to 'pending' so the next run
-          // (after token reset) picks it up.
-          if (res.rateLimited) {
-            const resetIso = await refreshNextReset();
-            setPaused('rate_limit', resetIso);
-          }
-          const sn = readQueue();
-          const i2 = sn.jobs.findIndex((x) => x.slug === job.slug);
-          if (i2 >= 0) {
-            const treatAsPending = res.rateLimited || (sn.paused && sn.paused.reason === 'rate_limit');
-            if (treatAsPending) {
-              resetJobFields(sn.jobs[i2], res.rateLimited ? 'paused: rate limit' : 'paused: queue halted');
-            } else {
-              sn.jobs[i2].status = res.exitCode === 0 ? 'completed' : 'failed';
-              sn.jobs[i2].finishedAt = new Date().toISOString();
-              sn.jobs[i2].exitCode = res.exitCode;
-              sn.jobs[i2].error = res.error || null;
+        const promise = (async () => {
+          try {
+            // Mark job running.
+            await mutate((s) => {
+              const idx = s.jobs.findIndex((x) => x.slug === job.slug);
+              if (idx >= 0) {
+                s.jobs[idx].status = 'running';
+                s.jobs[idx].runId = runId;
+                s.jobs[idx].startedAt = new Date().toISOString();
+              }
+            });
+            broadcast();
+
+            // Execute — onPid persists the child PID into the running state.
+            const res = await executeJob(job, runDir, state.config.defaultCwd, async (pid) => {
+              await mutate((s) => {
+                const idx = s.jobs.findIndex((x) => x.slug === job.slug);
+                if (idx >= 0) {
+                  s.jobs[idx].runtime = { pid, runId, startedAt: s.jobs[idx].startedAt };
+                }
+              });
+            });
+
+            // Rate-limit: pause before writing terminal status so the status
+            // mutate below can read the pause state.
+            if (res.rateLimited) {
+              const resetIso = await refreshNextReset().catch(() => cachedNextReset);
+              await setPaused('rate_limit', resetIso);
             }
-            writeQueue(sn);
+
+            // Write terminal status; strip runtime regardless of outcome.
+            await mutate((s) => {
+              const i2 = s.jobs.findIndex((x) => x.slug === job.slug);
+              if (i2 >= 0) {
+                const treatAsPending = res.rateLimited || (s.paused && s.paused.reason === 'rate_limit');
+                if (treatAsPending) {
+                  resetJobFields(s.jobs[i2], res.rateLimited ? 'paused: rate limit' : 'paused: queue halted');
+                } else {
+                  s.jobs[i2].status = res.exitCode === 0 ? 'completed' : 'failed';
+                  s.jobs[i2].finishedAt = new Date().toISOString();
+                  s.jobs[i2].exitCode = res.exitCode;
+                  s.jobs[i2].error = res.error || null;
+                  delete s.jobs[i2].runtime;
+                }
+              }
+            });
             broadcast();
+          } catch (e) {
+            console.error('[scheduler] launch error', job.slug, e);
           }
-          inFlight.delete(promise);
-        });
+        })();
         inFlight.add(promise);
+        promise.then(() => inFlight.delete(promise), () => inFlight.delete(promise));
       };
 
       // Prime up to cap
@@ -555,45 +715,97 @@ async function runDueJobs() {
     isExecuting = false;
     // No longer auto-disable after a run. The firePolicy now governs whether
     // the next batch fires automatically. Just clear the one-shot scheduledFor.
-    const s = readQueue();
-    s.scheduledFor = null;
-    writeQueue(s);
+    await mutate((s) => { s.scheduledFor = null; });
     broadcast();
   }
 }
 
-// ---------- when-available poll loop ----------
+// ---------- when-available launch logic ----------
 
-async function pollWhenAvailable() {
+async function maybeLaunchWhenAvailable(state) {
+  if (state.config.firePolicy !== 'when-available') return;
+  if (state.paused) return;
+  if (isExecuting) return;
+  const pending = state.jobs.filter((j) => j.status === 'pending');
+  if (pending.length === 0) return;
+  if (cachedUtilization === null || cachedUtilization === undefined) return;
+  if (cachedUtilization >= state.config.utilizationThreshold) {
+    broadcast();
+    return;
+  }
+  console.log(`[scheduler] when-available: util=${cachedUtilization}%, ${pending.length} pending — firing`);
+  runDueJobs().catch((e) => console.error('[scheduler] runDueJobs error', e));
+}
+
+// ---------- poll loop with exponential backoff ----------
+
+async function pollLoop() {
   try {
-    const state = readQueue();
-    if (state.config.firePolicy !== 'when-available') return;
-    if (state.paused) return;
-    if (isExecuting) return;
-    const pending = state.jobs.filter((j) => j.status === 'pending');
-    if (pending.length === 0) return;
+    const r = await billing.fetchUsage();
 
-    // Refresh utilization. If the call fails, don't fire blindly — wait
-    // for the next tick.
-    let util;
-    try {
-      const r = await billing.fetchUsage();
-      util = r?.usage?.five_hour?.utilization ?? null;
-      cachedUtilization = util;
-      cachedNextReset = { at: r?.usage?.five_hour?.resets_at ?? cachedNextReset.at, fetchedAt: Date.now() };
-    } catch {
-      return;
-    }
-    if (util === null || util === undefined) return;
-    if (util >= state.config.utilizationThreshold) {
-      // Tokens too high — broadcast so the UI shows current util but don't fire.
+    if (r.kind === 'ok') {
+      cachedNextReset = r.data?.usage?.five_hour?.resets_at ?? cachedNextReset;
+      cachedUtilization = r.data?.usage?.five_hour?.utilization ?? cachedUtilization;
+      consecutiveFailures = 0;
+      backoffMs = 0;
+      backoffNextAt = null;
+      firstFailureAt = null;
+      lastPollAt = Date.now();
+      lastPollOk = true;
+      persistSchedulerState();
+
+      // If a 'network' pause resolved, clear it now that we have a good reading.
+      const cur = readQueue();
+      if (cur.paused?.reason === 'network') {
+        await clearPause('network-recovered');
+      }
+      // If 'reset_failure' was set and we now have a valid reset, clear it.
+      if (cur.paused?.reason === 'reset_failure' && cachedNextReset) {
+        await clearPause('reset-recovered');
+      }
+
+      await maybeLaunchWhenAvailable(cur);
       broadcast();
-      return;
+    } else {
+      lastPollAt = Date.now();
+      lastPollOk = false;
+      consecutiveFailures++;
+      if (!firstFailureAt) firstFailureAt = Date.now();
+
+      if (r.kind === 'auth') {
+        console.error(`[scheduler] auth failure (HTTP ${r.httpStatus}): ${r.message}`);
+        await setPaused('auth', null);
+      } else {
+        // transient or config — apply exponential backoff.
+        backoffMs = backoffMs ? Math.min(backoffMs * 2, 480_000) : 30_000;
+        const totalFailureMs = Date.now() - firstFailureAt;
+        console.log(`[scheduler] transient failure #${consecutiveFailures}: ${r.kind} ${r.message ?? ''}; retry in ${backoffMs / 1000}s`);
+
+        // After 30 minutes of consecutive failures, set 'network' pause.
+        if (totalFailureMs > 30 * 60_000) {
+          const cur2 = readQueue();
+          if (!cur2.paused || cur2.paused.reason === 'network') {
+            await setPaused('network', null);
+          }
+        }
+      }
+
+      backoffNextAt = Date.now() + backoffMs;
+      persistSchedulerState();
     }
-    console.log(`[scheduler] when-available: util=${util}%, ${pending.length} pending — firing`);
-    runDueJobs().catch((e) => console.error('[scheduler] runDueJobs error', e));
   } catch (e) {
-    console.error('[scheduler] poll error', e);
+    // Unexpected error (e.g., IPC transport failure)
+    lastPollAt = Date.now();
+    lastPollOk = false;
+    consecutiveFailures++;
+    if (!firstFailureAt) firstFailureAt = Date.now();
+    backoffMs = backoffMs ? Math.min(backoffMs * 2, 480_000) : 30_000;
+    backoffNextAt = Date.now() + backoffMs;
+    persistSchedulerState();
+  } finally {
+    const delay = backoffMs || POLL_INTERVAL_MS;
+    pollLoopTimer = setTimeout(() => { pollLoop().catch(() => {}); }, delay);
+    if (pollLoopTimer.unref) pollLoopTimer.unref();
   }
 }
 
@@ -618,44 +830,85 @@ function registerScheduleHandlers() {
     };
   });
 
-  ipcMain.handle('schedule:set-config', async (_e, partial) => {
+  ipcMain.handle('schedule:health', async () => {
     const state = readQueue();
-    state.config = { ...state.config, ...(partial || {}) };
-    if (typeof state.config.concurrencyCap === 'number') {
-      state.config.concurrencyCap = Math.max(1, Math.min(20, Math.floor(state.config.concurrencyCap)));
+    const runningJobs = [];
+    for (const j of state.jobs) {
+      if (j.status === 'running' && j.runtime) {
+        runningJobs.push({
+          slug: j.slug,
+          startedAt: j.startedAt ? Date.parse(j.startedAt) : 0,
+          pid: j.runtime.pid ?? 0,
+        });
+      }
     }
-    if (typeof state.config.offsetMinutes === 'number') {
-      state.config.offsetMinutes = Math.max(0, Math.min(180, Math.floor(state.config.offsetMinutes)));
+    return {
+      bootedAt,
+      lastPollAt,
+      lastPollOk,
+      consecutiveFailures,
+      backoffNextAt,
+      nextResetCached: cachedNextReset,
+      pausedSince: state.paused ? Date.parse(state.paused.since) : null,
+      pauseReason: state.paused?.reason ?? null,
+      runningJobs,
+    };
+  });
+
+  ipcMain.handle('schedule:set-config', async (_e, partial) => {
+    const { schemas: s } = require('./ipcSchemas.cjs');
+    let validated;
+    try {
+      validated = s.setConfigSchema.parse(partial || {});
+    } catch (e) {
+      return { ok: false, error: e?.message ?? 'invalid config' };
     }
-    writeQueue(state);
+    const config = await mutate((state) => {
+      state.config = { ...state.config, ...validated };
+      return state.config;
+    });
     await rescheduleTimer();
-    return { ok: true, config: state.config };
+    return { ok: true, config };
   });
 
-  ipcMain.handle('schedule:reset-job', async (_e, { slug }) => {
-    const state = readQueue();
-    const idx = state.jobs.findIndex((j) => j.slug === slug);
-    if (idx < 0) return { ok: false, error: 'not found' };
-    resetJobFields(state.jobs[idx]);
-    writeQueue(state);
+  ipcMain.handle('schedule:reset-job', async (_e, payload) => {
+    const { schemas: s } = require('./ipcSchemas.cjs');
+    let slug;
+    try {
+      ({ slug } = s.scheduleSlug.parse(payload));
+    } catch (e) {
+      return { ok: false, error: 'invalid slug' };
+    }
+    // Containment check after path.join.
+    const resolved = path.resolve(path.join(PRDS_DIR, `${slug}.md`));
+    if (!resolved.startsWith(PRDS_DIR + path.sep)) {
+      return { ok: false, error: 'invalid slug' };
+    }
+    const found = await mutate((state) => {
+      const idx = state.jobs.findIndex((j) => j.slug === slug);
+      if (idx < 0) return false;
+      resetJobFields(state.jobs[idx]);
+      return true;
+    });
+    if (!found) return { ok: false, error: 'not found' };
     broadcast();
     return { ok: true };
   });
 
   ipcMain.handle('schedule:run-now', async () => {
     // Manual run-now overrides any auto-pause. Clear it first.
-    clearPause('run-now');
+    await clearPause('run-now');
     runDueJobs().catch((e) => console.error('[scheduler] runDueJobs error', e));
     return { ok: true };
   });
 
   ipcMain.handle('schedule:resume', async () => {
-    clearPause('manual');
+    await clearPause('manual');
     return { ok: true };
   });
 
   ipcMain.handle('schedule:refresh-reset', async () => {
-    const at = await refreshNextReset();
+    const at = await refreshNextReset().catch(() => cachedNextReset);
     await rescheduleTimer();
     return { ok: true, nextReset: at };
   });
@@ -666,39 +919,123 @@ function registerScheduleHandlers() {
     return { ok: true };
   });
 
-  ipcMain.handle('schedule:read-prd', async (_e, { slug }) => {
+  ipcMain.handle('schedule:read-prd', async (_e, payload) => {
+    const { schemas: s } = require('./ipcSchemas.cjs');
+    let slug;
     try {
-      const text = await fsp.readFile(path.join(PRDS_DIR, `${slug}.md`), 'utf8');
+      ({ slug } = s.scheduleSlug.parse(payload));
+    } catch {
+      return { ok: false, error: 'invalid slug' };
+    }
+    const filePath = path.resolve(path.join(PRDS_DIR, `${slug}.md`));
+    if (!filePath.startsWith(PRDS_DIR + path.sep)) {
+      return { ok: false, error: 'invalid slug' };
+    }
+    try {
+      const text = await fsp.readFile(filePath, 'utf8');
       return { ok: true, text };
     } catch (e) {
       return { ok: false, error: e?.message };
     }
   });
 
-  ipcMain.handle('schedule:read-log', async (_e, { runId, slug }) => {
+  ipcMain.handle('schedule:read-log', async (_e, payload) => {
+    const { schemas: s } = require('./ipcSchemas.cjs');
+    let slug, runId;
+    try {
+      ({ slug, runId } = s.scheduleReadLog.parse(payload));
+    } catch {
+      return { ok: false, error: 'invalid slug or runId' };
+    }
+    const logPath = path.resolve(path.join(RUNS_DIR, runId, `${slug}.log`));
+    if (!logPath.startsWith(RUNS_DIR + path.sep)) {
+      return { ok: false, error: 'invalid slug or runId' };
+    }
     try {
-      const p = path.join(RUNS_DIR, runId, `${slug}.log`);
-      const text = await fsp.readFile(p, 'utf8');
+      const text = await fsp.readFile(logPath, 'utf8');
       return { ok: true, text };
     } catch (e) {
       return { ok: false, error: e?.message };
     }
   });
+
+  const PRD_WRITE_MAX_BYTES = 256 * 1024;
+  const SLUG_RE = /^[A-Za-z0-9._-]{1,128}$/;
+
+  ipcMain.handle('schedule:write-prd', async (_e, { slug, body }) => {
+    if (!SLUG_RE.test(slug)) throw new Error(`invalid slug: ${slug}`);
+    if (typeof body !== 'string') throw new Error('body must be string');
+    if (Buffer.byteLength(body, 'utf8') > PRD_WRITE_MAX_BYTES) throw new Error('body too large');
+    const file = path.join(PRDS_DIR, `${slug}.md`);
+    const resolved = path.resolve(file);
+    if (!resolved.startsWith(PRDS_DIR + path.sep)) throw new Error('path escape');
+    const tmp = `${resolved}.${process.pid}.${Date.now()}.tmp`;
+    await fsp.writeFile(tmp, body, { encoding: 'utf8', mode: 0o644 });
+    await fsp.rename(tmp, resolved);
+    const stat = await fsp.stat(resolved);
+    return { ok: true, bytesWritten: stat.size };
+  });
+
+  ipcMain.handle('schedule:list-prds', async () => {
+    ensureDirs();
+    let entries;
+    try {
+      entries = await fsp.readdir(PRDS_DIR);
+    } catch {
+      return [];
+    }
+    const out = [];
+    for (const name of entries) {
+      if (!name.endsWith('.md') || name.startsWith('.')) continue;
+      const filePath = path.join(PRDS_DIR, name);
+      try {
+        const parsed = parsePrd(filePath);
+        const stat = await fsp.stat(filePath);
+        out.push({
+          slug: parsed.slug,
+          parallelGroup: parsed.parallelGroup,
+          title: parsed.title,
+          cwd: parsed.cwd || '',
+          estimateMinutes: parsed.estimateMinutes,
+          mtimeMs: stat.mtimeMs,
+        });
+      } catch (e) {
+        console.warn('[scheduler] list-prds: skipping unparseable', name, e?.message);
+      }
+    }
+    out.sort((a, b) => a.slug.localeCompare(b.slug, undefined, { numeric: true }));
+    return out;
+  });
 }
 
 async function init() {
   ensureDirs();
-  // Ensure queue.json exists with defaults so the renderer can read it.
-  if (!fs.existsSync(QUEUE_PATH)) writeQueue({ config: { ...DEFAULT_CONFIG }, jobs: [], scheduledFor: null, lastRunAt: null, paused: null });
+
+  // Hydrate cached state from the sidecar before any scheduling decisions.
+  loadSchedulerState();
+  bootedAt = Date.now();
+
+  // Boot reconciliation: mark any job that was 'running' when the app died as
+  // 'failed'. mutate() creates queue.json from defaults if it doesn't exist.
+  await mutate((state) => {
+    for (const j of state.jobs) {
+      if (j.status === 'running') {
+        j.status = 'failed';
+        j.error = 'orphaned: app restarted while running';
+        j.finishedAt = new Date().toISOString();
+        delete j.runtime;
+      }
+    }
+  });
 
   // If we boot up while paused with a resumeAt in the past, clear it. This
   // happens when the app was closed across the reset window.
   const boot = readQueue();
   if (boot.paused && boot.paused.resumeAt && new Date(boot.paused.resumeAt).getTime() <= Date.now()) {
-    clearPause('boot-elapsed');
+    await clearPause('boot-elapsed');
   } else if (boot.paused && boot.paused.resumeAt) {
     // Re-arm the resume timer (lost across restart).
-    setPaused(boot.paused.reason, boot.paused.resumeAt);
+    await setPaused(boot.paused.reason, boot.paused.resumeAt);
   }
 
   await rescheduleTimer();
@@ -706,13 +1043,49 @@ async function init() {
   // resets early or the auth token rotates. Tracked so re-init doesn't leak.
   if (rescheduleInterval) clearInterval(rescheduleInterval);
   rescheduleInterval = setInterval(() => { rescheduleTimer().catch(() => {}); }, 10 * 60_000);
-  // when-available poll loop. Tick every 2 minutes; the function itself is
-  // a no-op when policy != 'when-available' or queue is empty/paused.
-  if (pollTimer) clearInterval(pollTimer);
-  pollTimer = setInterval(() => { pollWhenAvailable().catch(() => {}); }, 2 * 60_000);
-  // First tick fires after a short delay so billing is warmed up.
-  if (initialPollTimeout) clearTimeout(initialPollTimeout);
-  initialPollTimeout = setTimeout(() => { pollWhenAvailable().catch(() => {}); }, 15_000);
+
+  // Self-rescheduling poll loop with exponential backoff. Replaces the
+  // old fixed-interval pollTimer + initialPollTimeout.
+  if (pollLoopTimer) clearTimeout(pollLoopTimer);
+  // First tick fires after the standard warmup delay so billing is ready.
+  pollLoopTimer = setTimeout(() => { pollLoop().catch(() => {}); }, USAGE_REFRESH_INTERVAL_MS);
+  if (pollLoopTimer.unref) pollLoopTimer.unref();
+
+  // Heartbeat: once per minute, log queue state for 24h visibility.
+  if (heartbeatInterval) clearInterval(heartbeatInterval);
+  heartbeatInterval = setInterval(() => {
+    const s = readQueue();
+    const counts = { pending: 0, running: 0, completed: 0, failed: 0 };
+    for (const j of s.jobs) counts[j.status] = (counts[j.status] || 0) + 1;
+    appendHeartbeat({
+      ts: Date.now(),
+      counts,
+      paused: s.paused ? { reason: s.paused.reason, resumeAt: s.paused.resumeAt } : null,
+      nextReset: cachedNextReset,
+      utilization: cachedUtilization,
+      consecutiveFailures,
+    });
+  }, 60_000);
+  if (heartbeatInterval.unref) heartbeatInterval.unref();
+
+  // Wake-from-sleep: immediately re-poll and re-evaluate the queue.
+  try {
+    const { powerMonitor } = require('electron');
+    powerMonitor.on('resume', () => {
+      console.log('[scheduler] system resumed; re-polling and re-evaluating queue');
+      if (pollLoopTimer) { clearTimeout(pollLoopTimer); pollLoopTimer = null; }
+      backoffMs = 0;
+      backoffNextAt = null;
+      // Clear any paused-but-resumeAt-elapsed state immediately.
+      const wakeState = readQueue();
+      if (wakeState.paused?.resumeAt && new Date(wakeState.paused.resumeAt).getTime() <= Date.now()) {
+        clearPause('boot-elapsed').then(() => { runDueJobs().catch(() => {}); }).catch(() => {});
+      }
+      pollLoop().catch(() => {});
+    });
+  } catch (e) {
+    console.warn('[scheduler] powerMonitor unavailable', e?.message);
+  }
 }
 
 module.exports = { registerScheduleHandlers, attachWindow, init, ROOT, PRDS_DIR };