@yugabytedb/meko-mcp 2.1.0 → 2.3.0

package/assets/hooks-handlers/lib/checkpoint-timer.js

@@ -1,126 +1,530 @@
 #!/usr/bin/env node
 /**
- * checkpoint-timer.js — Background daemon for periodic conversation capture.
+ * checkpoint-timer.js — Per-session daemon: live checkpoints + recovery scheduler.
  *
- * Spawned by session-start.sh at session start. Runs capture.js in checkpoint
- * mode every MEKO_CHECKPOINT_INTERVAL seconds (default: 600 = 10 minutes).
+ * Spawned by session-start.sh (and Cursor native SessionStart) at session start.
+ * Every MEKO_CHECKPOINT_INTERVAL seconds (default 600):
+ *   1. Drain this live session via `capture.js checkpoint`
+ *   2. Uncapped scan → capped fair recovery of eligible dead/closing sessions
+ *   3. Rebuild aggregate health cache
  *
- * Writes PID to ~/.claude/meko-capture/<session-id>.timer.pid for cleanup
- * by session-end.sh.
+ * Recovery also runs immediately on daemon start, then repeats while work remains.
  *
  * Usage: node checkpoint-timer.js <transcript_path>
  *
- * Environment:
- *   MEKO_CHECKPOINT_INTERVAL  Seconds between ticks (default: 600)
- *   MEKO_TIMER_MAX_AGE        Hard self-exit age in seconds (default: 86400 = 24h)
- *   MEKO_TIMER_IDLE_EXIT      Self-exit after the transcript is untouched this
- *                             long, in seconds (default: 7200 = 2h)
- *   MEKO_TIMER_STARTUP_GRACE  Wait this long for the transcript to appear
- *                             before self-exiting (default: 1800 = 30min)
- *   MEKO_WATERMARK_DIR        Watermark directory (default: ~/.claude/meko-capture)
- *   MEKO_MCP_URL              Inherited by capture.js
- *   MEKO_API_KEY              Inherited by capture.js
+ * Environment: see docs/plans/meko-capture-v2-contracts.md
  */
 
+"use strict";
+
 const { execFile } = require("child_process");
 const fs = require("fs");
 const path = require("path");
+const crypto = require("crypto");
+const {
+  checkpointInterval,
+  checkpointTimeout,
+  recoveryMaxSessions,
+  recoveryMaxExchanges,
+  maxCaptureBatchSize,
+  timerMaxAge,
+  timerPidPath,
+  watermarkDir,
+  scanStates,
+  scanCompleteForDropClamp,
+  sortOldestFirst,
+  computeAggregateHealth,
+  writeHealthCache,
+  importLegacyCaptureErrors,
+  legacyErrorsSafeToRemove,
+  isResumableState,
+  readState,
+} = require("./capture-state");
 
-const INTERVAL_SECS = parseInt(process.env.MEKO_CHECKPOINT_INTERVAL || "600", 10);
-// Self-exit backstops so an orphaned daemon (its pid file removed out of band,
-// a missed SessionEnd, a blocked SIGTERM) cannot tick forever. Both are
-// generous enough not to interrupt a genuinely long, active session.
-const MAX_AGE_SECS = parseInt(process.env.MEKO_TIMER_MAX_AGE, 10) || 86400;
+const INTERVAL_SECS = Math.max(1, checkpointInterval());
+const CAPTURE_TIMEOUT_MS = Math.max(1, checkpointTimeout()) * 1000;
+const MAX_AGE_SECS = Math.max(1, timerMaxAge());
 const IDLE_EXIT_SECS = parseInt(process.env.MEKO_TIMER_IDLE_EXIT, 10) || 7200;
 const STARTUP_GRACE_SECS =
   parseInt(process.env.MEKO_TIMER_STARTUP_GRACE, 10) || 1800;
-const startedAt = Date.now();
-const WATERMARK_DIR =
-  process.env.MEKO_WATERMARK_DIR ||
-  path.join(process.env.HOME || "~", ".claude", "meko-capture");
 
+const startedAt = Date.now();
+const STATE_DIR = watermarkDir();
 const transcriptPath = process.argv[2];
+
 if (!transcriptPath) {
-  process.stderr.write("[meko-timer] Usage: checkpoint-timer.js <transcript_path>\n");
+  process.stderr.write(
+    "[meko-timer] Usage: checkpoint-timer.js <transcript_path>\n",
+  );
   process.exit(1);
 }
 
 const sessionId = path.basename(transcriptPath, ".jsonl");
-const pidFile = path.join(WATERMARK_DIR, `${sessionId}.timer.pid`);
+const pidFile = timerPidPath(sessionId);
 const captureScript = path.join(__dirname, "capture.js");
 
-// Write PID file
-fs.mkdirSync(WATERMARK_DIR, { recursive: true });
-fs.writeFileSync(pidFile, String(process.pid));
+let tickInFlight = false;
+let recoveryInFlight = false;
+let workRemains = true; // force immediate recovery on start
+let startupRecoveryPending = true;
+let startupPersistentQueue = null;
 
-function cleanup() {
-  try { fs.unlinkSync(pidFile); } catch { /* already gone */ }
-  process.exit(0);
+function isProcessAlive(pid) {
+  if (!Number.isInteger(pid) || pid <= 0) return false;
+  try {
+    process.kill(pid, 0);
+    return true;
+  } catch {
+    return false;
+  }
 }
 
-process.on("SIGTERM", cleanup);
-process.on("SIGINT", cleanup);
-
-function tick() {
-  // Self-exit backstop: never outlive the hard max age. Guards against orphans
-  // whose pid file was removed out of band (so session-end.sh can't kill them)
-  // or whose SessionEnd never fired (force-quit / terminal closed).
-  if (Date.now() - startedAt >= MAX_AGE_SECS * 1000) {
+/**
+ * Atomic PID publish (PR #254): write populated candidate inode, then linkSync.
+ * Avoids empty-lock races from create-then-write on the published path.
+ */
+function acquirePidFile() {
+  fs.mkdirSync(STATE_DIR, { recursive: true, mode: 0o700 });
+  const candidateFile = path.join(
+    STATE_DIR,
+    `${sessionId}.timer.pid.${process.pid}.${Date.now()}.${crypto
+      .randomBytes(4)
+      .toString("hex")}.tmp`,
+  );
+  try {
+    fs.writeFileSync(candidateFile, String(process.pid), {
+      flag: "wx",
+      mode: 0o600,
+    });
+  } catch (err) {
     process.stderr.write(
-      `[meko-timer] Max age (${MAX_AGE_SECS}s) reached; self-exiting to avoid orphan.\n`,
+      `[meko-timer] Could not stage PID file: ${err.message}\n`,
     );
-    cleanup();
-    return;
+    return false;
   }
 
-  // Exit if the transcript is gone (session ended but we weren't killed yet).
-  // SessionStart can precede transcript creation by up to 20 minutes, so a
-  // missing file during the startup grace period is expected.
-  let stat;
+  let acquired = false;
   try {
-    stat = fs.statSync(transcriptPath);
+    for (let attempt = 0; attempt < 2; attempt++) {
+      try {
+        fs.linkSync(candidateFile, pidFile);
+        acquired = true;
+        break;
+      } catch (err) {
+        if (!err || err.code !== "EEXIST") {
+          process.stderr.write(
+            `[meko-timer] PID publish failed: ${err.message}\n`,
+          );
+          break;
+        }
+        let existingPid = NaN;
+        try {
+          existingPid = parseInt(fs.readFileSync(pidFile, "utf8"), 10);
+        } catch {
+          existingPid = NaN;
+        }
+        if (isProcessAlive(existingPid)) {
+          process.stderr.write(
+            `[meko-timer] Timer already active for ${sessionId} (pid ${existingPid})\n`,
+          );
+          break;
+        }
+        try {
+          fs.unlinkSync(pidFile);
+        } catch {
+          /* ignore */
+        }
+      }
+    }
+  } finally {
+    try {
+      fs.unlinkSync(candidateFile);
+    } catch {
+      /* ignore */
+    }
+  }
+  return acquired;
+}
+
+function cleanup() {
+  try {
+    const recorded = parseInt(fs.readFileSync(pidFile, "utf8"), 10);
+    if (recorded === process.pid) fs.unlinkSync(pidFile);
   } catch {
-    if (Date.now() - startedAt < STARTUP_GRACE_SECS * 1000) {
-      return;
+    /* already gone */
+  }
+  process.exit(0);
+}
+
+function runCaptureChild(mode, payload, timeoutMs) {
+  return new Promise((resolve) => {
+    const child = execFile(
+      process.execPath,
+      [captureScript, mode],
+      { timeout: timeoutMs, maxBuffer: 4 * 1024 * 1024 },
+      (err, stdout, stderr) => {
+        let result = null;
+        const text = (stdout || "").trim();
+        if (text) {
+          try {
+            const line = text
+              .split(/\r?\n/)
+              .find((l) => l.trim().startsWith("{"));
+            if (line) result = JSON.parse(line);
+          } catch {
+            result = null;
+          }
+        }
+        resolve({
+          ok: !err,
+          error: err ? err.message : null,
+          result,
+          stderr: stderr || "",
+        });
+      },
+    );
+    child.stdin.on("error", () => {});
+    child.stdin.end(JSON.stringify(payload || {}));
+  });
+}
+
+// MEKO-635: settle OpenCode pending turns past the 30-minute default.
+function sweepOpenCodeSettles() {
+  try {
+    const opencode = require("./opencode-handler");
+    if (typeof opencode.sweepExpiredSettles === "function") {
+      opencode.sweepExpiredSettles();
     }
-    cleanup();
-    return;
+  } catch (err) {
+    process.stderr.write(
+      `[meko-timer] OpenCode settle sweep failed: ${err.message}\n`,
+    );
   }
+}
 
-  // Exit if the transcript has been idle far longer than a checkpoint window —
-  // the session is effectively over even though we were never signalled.
-  if (Date.now() - stat.mtimeMs >= IDLE_EXIT_SECS * 1000) {
+function rebuildHealth() {
+  try {
+    const scan = scanStates({ includeBlocked: true });
+    const health = computeAggregateHealth(scan);
+    writeHealthCache(health, {
+      clampDropped: scanCompleteForDropClamp(scan),
+    });
+  } catch (err) {
     process.stderr.write(
-      `[meko-timer] Transcript idle > ${IDLE_EXIT_SECS}s; self-exiting.\n`,
+      `[meko-timer] Health rebuild failed: ${err.message}\n`,
     );
-    cleanup();
-    return;
   }
+}
 
-  const hookInput = JSON.stringify({ transcript_path: transcriptPath });
-
-  // Run capture ASYNCHRONOUSLY. execFileSync would block this daemon's event
-  // loop for up to 30s, during which SIGTERM/SIGINT (clean shutdown) can't be
-  // handled. execFile keeps the loop responsive so cleanup() fires promptly.
-  const child = execFile(
-    "node",
-    [captureScript, "checkpoint"],
-    { timeout: 30000 },
-    (err) => {
-      if (err) {
-        // Best-effort: log and continue to the next tick.
-        process.stderr.write(`[meko-timer] Checkpoint tick failed: ${err.message}\n`);
+function migrateLegacyErrors() {
+  try {
+    importLegacyCaptureErrors();
+    const uncapped = scanStates({ includeBlocked: true });
+    for (const legacyPath of legacyErrorsSafeToRemove(uncapped)) {
+      try {
+        fs.unlinkSync(legacyPath);
+      } catch {
+        /* best effort; a future uncapped scan retries */
       }
-    },
+    }
+  } catch (err) {
+    process.stderr.write(
+      `[meko-timer] Legacy capture error import failed: ${err.message}\n`,
+    );
+  }
+}
+
+function refreshStartupPersistentQueue(scan, forcePersistentRetry) {
+  const forcedIds = new Set(
+    (scan.entries || [])
+      .filter(
+        (entry) =>
+          entry &&
+          entry.ok &&
+          entry.eligibility &&
+          entry.eligibility.persistent_startup_override,
+      )
+      .map((entry) => entry.session_id),
   );
-  // The stdin stream can emit EPIPE if the child exits or fails to spawn before
-  // we finish writing. Without a listener that error is unhandled and crashes
-  // the daemon — so swallow it here; the execFile callback above reports the
-  // real failure.
-  child.stdin.on("error", (err) => {
-    process.stderr.write(`[meko-timer] Checkpoint stdin error: ${err.message}\n`);
-  });
-  child.stdin.end(hookInput);
+  if (forcePersistentRetry && startupPersistentQueue === null && scan.ok) {
+    // Snapshot the daemon-start backlog. A bounded cycle may attempt only a
+    // subset; retain the rest for follow-up cycles without also forcing new
+    // persistent failures that happen later in this daemon's lifetime.
+    startupPersistentQueue = new Set(forcedIds);
+  } else if (startupPersistentQueue) {
+    for (const queuedId of startupPersistentQueue) {
+      if (!forcedIds.has(queuedId)) startupPersistentQueue.delete(queuedId);
+    }
+  }
+  return forcedIds;
 }
 
-setInterval(tick, INTERVAL_SECS * 1000);
+/**
+ * Uncapped discovery, then capped fair recovery (oldest first).
+ * Returns true if more recoverable work may remain for a later cycle.
+ */
+async function runRecoveryCycle() {
+  if (recoveryInFlight) return workRemains;
+  recoveryInFlight = true;
+  const forcePersistentRetry = startupRecoveryPending;
+  let moreWork = false;
+  try {
+    // Finalize any resumable session whose settling window has expired before
+    // the scan snapshots eligibility. Otherwise recovery keeps draining a
+    // pending user (queued_exchanges === 1) whose assistant will never arrive.
+    sweepOpenCodeSettles();
+    const scan = scanStates({
+      includeBlocked: true,
+      forcePersistentRetry,
+    });
+    refreshStartupPersistentQueue(scan, forcePersistentRetry);
+    const eligible = sortOldestFirst(
+      (scan.entries || []).filter(
+        (e) =>
+          e &&
+          e.ok &&
+          e.eligibility &&
+          e.eligibility.eligible_for_recovery &&
+          (!e.eligibility.persistent_startup_override ||
+            (startupPersistentQueue && startupPersistentQueue.has(e.session_id))),
+      ),
+    );
+
+    const maxSessions = recoveryMaxSessions();
+    const maxExchanges = recoveryMaxExchanges();
+    const selected = eligible.slice(0, maxSessions);
+    moreWork = eligible.length > selected.length;
+
+    let budget = maxExchanges;
+    for (const entry of selected) {
+      if (budget <= 0) {
+        moreWork = true;
+        break;
+      }
+      // Skip the live session this daemon owns — live tick handles it.
+      if (entry.session_id === sessionId) continue;
+
+      if (
+        entry.eligibility &&
+        entry.eligibility.persistent_startup_override &&
+        startupPersistentQueue
+      ) {
+        // One forced attempt per session per daemon start. Remove before
+        // launch so a child crash cannot make this daemon hammer the session.
+        startupPersistentQueue.delete(entry.session_id);
+      }
+
+      const state = entry.state || {};
+      const resumable = isResumableState(state);
+      const payload = {
+        session_id: entry.session_id,
+        transcript_path: state.transcript_path || undefined,
+        lifecycle: resumable ? "active" : "closing",
+        confirmed_dead: !resumable,
+        recorded_owner_only: true,
+        max_exchanges: Math.min(budget, maxCaptureBatchSize()),
+        force_retry: Boolean(
+          entry.eligibility &&
+            entry.eligibility.persistent_startup_override,
+        ),
+      };
+      const outcome = await runCaptureChild(
+        "recover",
+        payload,
+        CAPTURE_TIMEOUT_MS,
+      );
+      const result = outcome.result;
+      if (!outcome.ok) {
+        process.stderr.write(
+          `[meko-timer] Recovery drain failed for ${entry.session_id}: ${outcome.error}\n`,
+        );
+        moreWork = true;
+        // Stop hammering an unhealthy server this cycle.
+        break;
+      }
+      const captured = Number((result && result.captured) || 0);
+      budget -= captured;
+      const queued = Number((result && result.queued_remaining) || 0);
+      if (
+        (queued > 0 || (result && result.status === "retry_wait")) &&
+        !(resumable && captured === 0 && queued <= 1 && result.status !== "retry_wait")
+      ) {
+        moreWork = true;
+      }
+      if (result && (result.status === "partial" || result.status === "retry_wait" || result.status === "blocked")) {
+        // Continue other sessions unless this looks like a systemic outage.
+        if (result.failure_class === "transient" && captured === 0) {
+          moreWork = true;
+          break;
+        }
+      }
+    }
+
+    startupRecoveryPending =
+      startupPersistentQueue === null || startupPersistentQueue.size > 0;
+    if (startupRecoveryPending) moreWork = true;
+
+    // Also consider pending entries we skipped due to retry_wait not due yet.
+    if (!moreWork) {
+      moreWork = (scan.entries || []).some(
+        (e) =>
+          e &&
+          e.ok &&
+          e.eligibility &&
+          e.session_id !== sessionId &&
+          Number(e.eligibility.queued_exchanges || 0) > 0 &&
+          e.eligibility.retry_due === false &&
+          !(isResumableState(e.state) &&
+            Number(e.eligibility.queued_exchanges || 0) <= 1 &&
+            e.state && e.state.delivery === "pending"),
+      );
+    }
+
+    rebuildHealth();
+  } catch (err) {
+    process.stderr.write(`[meko-timer] Recovery cycle error: ${err.message}\n`);
+    moreWork = true;
+  } finally {
+    recoveryInFlight = false;
+  }
+  workRemains = moreWork;
+  return moreWork;
+}
+
+// MEKO-635: for the daemon's own session, look up whether its captured intent
+// declared a resumable pending-turn policy. If so, timer-idle and
+// transcript-gone paths must drain without finalizing the session as dead.
+function selfIsResumable() {
+  const read = readState(sessionId);
+  return read.ok ? isResumableState(read.state) : false;
+}
+
+async function tick() {
+  if (tickInFlight) return;
+  tickInFlight = true;
+  try {
+    if (Date.now() - startedAt >= MAX_AGE_SECS * 1000) {
+      process.stderr.write(
+        `[meko-timer] Max age (${MAX_AGE_SECS}s) reached; self-exiting (outbox left pending).\n`,
+      );
+      cleanup();
+      return;
+    }
+
+    let stat;
+    try {
+      stat = fs.statSync(transcriptPath);
+    } catch {
+      if (Date.now() - startedAt < STARTUP_GRACE_SECS * 1000) {
+        // Still start recovery even if our transcript isn't ready yet.
+        await runRecoveryCycle();
+        return;
+      }
+      // Transcript gone after grace. For a resumable session drain without
+      // finalizing so a later snapshot can complete its pending turn.
+      const resumable = selfIsResumable();
+      await runCaptureChild(
+        "drain",
+        {
+          session_id: sessionId,
+          transcript_path: transcriptPath,
+          lifecycle: resumable ? "active" : "closing",
+          confirmed_dead: !resumable,
+          recorded_owner_only: true,
+        },
+        CAPTURE_TIMEOUT_MS,
+      );
+      await runRecoveryCycle();
+      cleanup();
+      return;
+    }
+
+    if (Date.now() - stat.mtimeMs >= IDLE_EXIT_SECS * 1000) {
+      process.stderr.write(
+        `[meko-timer] Transcript idle > ${IDLE_EXIT_SECS}s; final drain then exit.\n`,
+      );
+      const resumable = selfIsResumable();
+      await runCaptureChild(
+        "drain",
+        {
+          session_id: sessionId,
+          transcript_path: transcriptPath,
+          lifecycle: resumable ? "active" : "closing",
+          confirmed_dead: !resumable,
+          recorded_owner_only: true,
+        },
+        CAPTURE_TIMEOUT_MS,
+      );
+      await runRecoveryCycle();
+      cleanup();
+      return;
+    }
+
+    // Live session checkpoint. Recovery skips the daemon's own session so it
+    // cannot accidentally finalize a live transcript; consume that session's
+    // one-shot persistent retry here instead.
+    let forceLivePersistentRetry = false;
+    if (startupRecoveryPending) {
+      const startupScan = scanStates({
+        includeBlocked: true,
+        forcePersistentRetry: true,
+      });
+      const forcedIds = refreshStartupPersistentQueue(startupScan, true);
+      forceLivePersistentRetry = Boolean(
+        forcedIds.has(sessionId) &&
+          startupPersistentQueue &&
+          startupPersistentQueue.has(sessionId),
+      );
+      if (forceLivePersistentRetry) {
+        // Remove before launch so a child crash still consumes the one-shot.
+        startupPersistentQueue.delete(sessionId);
+      }
+    }
+    const live = await runCaptureChild(
+      "checkpoint",
+      {
+        transcript_path: transcriptPath,
+        force_retry: forceLivePersistentRetry,
+      },
+      CAPTURE_TIMEOUT_MS,
+    );
+    if (!live.ok) {
+      process.stderr.write(
+        `[meko-timer] Checkpoint tick failed: ${live.error}\n`,
+      );
+    }
+
+    await runRecoveryCycle();
+  } finally {
+    tickInFlight = false;
+  }
+}
+
+if (!acquirePidFile()) {
+  process.exit(0);
+}
+
+process.on("SIGTERM", cleanup);
+process.on("SIGINT", cleanup);
+
+// Immediate recovery on start, then periodic ticks.
+migrateLegacyErrors();
+tick().catch((err) => {
+  process.stderr.write(`[meko-timer] Startup tick failed: ${err.message}\n`);
+});
+setInterval(() => {
+  tick().catch((err) => {
+    process.stderr.write(`[meko-timer] Tick failed: ${err.message}\n`);
+  });
+}, INTERVAL_SECS * 1000);
+
+// While backlog remains, also schedule a shorter recovery follow-up without
+// waiting for the full checkpoint interval (self-draining).
+setInterval(() => {
+  if (!workRemains || recoveryInFlight || tickInFlight) return;
+  runRecoveryCycle().catch((err) => {
+    process.stderr.write(
+      `[meko-timer] Follow-up recovery failed: ${err.message}\n`,
+    );
+  });
+}, Math.min(30, INTERVAL_SECS) * 1000);