taskflow-core 0.2.1 → 0.2.3

package/dist/store.js

@@ -20,6 +20,7 @@ import * as path from "node:path";
 import { parseJsonc } from "./jsonc.js";
 import { getAgentDir } from "./paths.js";
 import { parseStrict } from "./interpolate.js";
+import { directoryIdentity } from "./cwd-bridge.js";
 /** Build a single-line, user-facing message from a failed `LoadResult`. */
 export function describeLoadFailure(r, what) {
     return r.reason === "missing"
@@ -54,11 +55,16 @@ const LOCK_STALE_MS = 30_000;
 const LOCK_POLL_MS = 50;
 /** Default acquisition timeout before throwing. */
 const LOCK_TIMEOUT_MS = 10_000;
+/** Retention is opportunistic: never stall a foreground save behind cleanup. */
+const CLEANUP_LOCK_TIMEOUT_MS = 250;
 // ---------------------------------------------------------------------------
 // Cleanup throttle
 // ---------------------------------------------------------------------------
 /** Minimum ms between opportunistic cleanup runs (called inside saveRun). */
 const CLEANUP_INTERVAL_MS = 60_000;
+/** Bound the project-keyed throttle so a long-lived multi-project host cannot
+ * retain one map entry for every directory it has ever visited. */
+const CLEANUP_THROTTLE_MAX_ROOTS = 256;
 /** Retain at most this many terminal runs by default. */
 const DEFAULT_MAX_KEPT_TERMINAL = 100;
 /** Remove terminal runs older than this (days). */
@@ -66,8 +72,9 @@ const DEFAULT_MAX_AGE_DAYS = 30;
 // Re-exported for use in TaskflowSettings defaults (agents.ts).
 export const DEFAULT_KEPT_RUNS = DEFAULT_MAX_KEPT_TERMINAL;
 export const DEFAULT_RUN_AGE_DAYS = DEFAULT_MAX_AGE_DAYS;
-/** Last cleanup timestamp — module-level so it persists across calls. */
-let lastCleanupAt = 0;
+/** Per-runs-root cleanup timestamps. A process-global scalar lets one busy
+ * project suppress retention in every other project served by the same host. */
+const lastCleanupAtByRoot = new Map();
 /** Shared buffer for Atomics.wait in acquireLock busy-wait (Finding 6). */
 const LOCK_WAIT_BUF = new Int32Array(new SharedArrayBuffer(4));
 // ---------------------------------------------------------------------------
@@ -120,27 +127,220 @@ function lockPathForRun(runsRoot, flowName, runId) {
  * Legitimate runIds are produced by newRunId() and contain only [A-Za-z0-9._-].
  */
 export function validateRunId(runId) {
-    return (typeof runId === "string" &&
-        runId.length > 0 &&
-        !runId.includes("/") &&
-        !runId.includes("\\") &&
-        !runId.includes("\0") &&
-        !runId.includes(".."));
+    // A single leading/trailing dot is safe once the id is used as
+    // `${runId}.json`, and `newRunId()` has historically produced leading-dot
+    // ids for valid flow names such as `.ci`. Reject separators and dot-dot
+    // traversal, but retain compatibility with those already-persisted runs.
+    return typeof runId === "string" && runId.length > 0 && runId.length <= 160 &&
+        /^[A-Za-z0-9._-]+$/.test(runId) && !runId.includes("..");
+}
+/**
+ * Validate an index path before it is joined to the runs root.
+ *
+ * Index files live in a project-controlled directory and may be stale or
+ * manually edited. Only the two layouts Taskflow itself has ever emitted are
+ * accepted: `<flowDir>/<runId>.json` and the legacy `<runId>.json`. Keeping
+ * this check independent from the host OS also makes an index written on one
+ * platform safe to consume on another (generated index paths always use `/`).
+ */
+function isSafeRunIndexRelPath(relPath, runId) {
+    if (!validateRunId(runId) || relPath.length === 0 || relPath.length > 420)
+        return false;
+    if (path.isAbsolute(relPath) || relPath.includes("\\"))
+        return false;
+    const parts = relPath.split("/");
+    if (parts.some((part) => part.length === 0 || part === "." || part === ".."))
+        return false;
+    if (parts.length === 1)
+        return parts[0] === `${runId}.json`;
+    if (parts.length !== 2)
+        return false;
+    const [flowDir, fileName] = parts;
+    return flowDir.length <= 255 && safeFlowDirName(flowDir) === flowDir && fileName === `${runId}.json`;
+}
+/** Join an already-validated, platform-neutral index path to the runs root. */
+function runIndexFilePath(runsRoot, relPath) {
+    return path.join(runsRoot, ...relPath.split("/"));
+}
+/** Canonical, bounded-LRU throttle for opportunistic per-project cleanup. */
+function shouldRunCleanup(runsRoot, now) {
+    let key;
+    try {
+        key = fs.realpathSync(runsRoot);
+    }
+    catch {
+        key = path.resolve(runsRoot);
+    }
+    const previous = lastCleanupAtByRoot.get(key);
+    if (previous !== undefined && now - previous < CLEANUP_INTERVAL_MS)
+        return false;
+    // Refresh insertion order for simple LRU eviction.
+    lastCleanupAtByRoot.delete(key);
+    lastCleanupAtByRoot.set(key, now);
+    while (lastCleanupAtByRoot.size > CLEANUP_THROTTLE_MAX_ROOTS) {
+        const oldest = lastCleanupAtByRoot.keys().next().value;
+        if (oldest === undefined)
+            break;
+        lastCleanupAtByRoot.delete(oldest);
+    }
+    return true;
+}
+/**
+ * Resolve a physical directory only when it is a non-symlink descendant of
+ * runsRoot. Retention performs destructive operations, so lexical containment
+ * alone is insufficient: a checked-in `runs/<flow>` symlink could otherwise
+ * redirect the run lock and unlink to an arbitrary directory.
+ */
+function physicalDirectoryInsideRunsRoot(runsRoot, candidate) {
+    try {
+        const rootReal = fs.realpathSync(runsRoot);
+        const stat = fs.lstatSync(candidate);
+        if (!stat.isDirectory() || stat.isSymbolicLink())
+            return null;
+        const realPath = fs.realpathSync(candidate);
+        const rel = path.relative(rootReal, realPath);
+        if (rel === ".." || rel.startsWith(`..${path.sep}`) || path.isAbsolute(rel))
+            return null;
+        return { realPath, device: stat.dev, inode: stat.ino };
+    }
+    catch {
+        return null;
+    }
+}
+function physicalDirectoryStillMatches(runsRoot, candidate, snapshot) {
+    const current = physicalDirectoryInsideRunsRoot(runsRoot, candidate);
+    return Boolean(current && current.realPath === snapshot.realPath &&
+        current.device === snapshot.device && current.inode === snapshot.inode);
+}
+/** Remove a Taskflow-owned artifact tree without following a project symlink. */
+function removeArtifactDirectoryInsideRunsRoot(runsRoot, target) {
+    const parent = path.dirname(target);
+    const parentSnapshot = physicalDirectoryInsideRunsRoot(runsRoot, parent);
+    if (!parentSnapshot)
+        return;
+    try {
+        const stat = fs.lstatSync(target);
+        if (!stat.isDirectory() || stat.isSymbolicLink())
+            return;
+        const targetSnapshot = physicalDirectoryInsideRunsRoot(runsRoot, target);
+        if (!targetSnapshot || !physicalDirectoryStillMatches(runsRoot, parent, parentSnapshot))
+            return;
+        if (!physicalDirectoryStillMatches(runsRoot, target, targetSnapshot))
+            return;
+        fs.rmSync(target, { recursive: true, force: true });
+    }
+    catch { /* missing / concurrently changed */ }
+}
+/** Accept a persisted cwd for control-record cleanup only when it resolves to
+ * the same project run store currently being retained. */
+function controlCwdForRunsRoot(runsRoot, candidate) {
+    const fallback = path.dirname(path.dirname(path.dirname(runsRoot)));
+    if (typeof candidate !== "string" || candidate.length === 0)
+        return fallback;
+    try {
+        if (fs.realpathSync(runsDir(candidate)) === fs.realpathSync(runsRoot))
+            return candidate;
+    }
+    catch { /* malformed, missing, or from another project */ }
+    return fallback;
 }
 // ---------------------------------------------------------------------------
 // File-lock primitives — zero-dependency, using O_CREAT|O_EXCL (atomic)
 // ---------------------------------------------------------------------------
+function readLockOwner(lockPath) {
+    try {
+        const stat = fs.lstatSync(lockPath);
+        if (!stat.isFile() || stat.isSymbolicLink() || stat.size > 4_096)
+            return null;
+        const raw = fs.readFileSync(lockPath, "utf-8");
+        if (Buffer.byteLength(raw, "utf-8") > 4_096)
+            return null;
+        const parsed = JSON.parse(raw);
+        if (!Number.isSafeInteger(parsed.pid) || typeof parsed.ts !== "number")
+            return null;
+        return {
+            pid: parsed.pid,
+            ts: parsed.ts,
+            ...(typeof parsed.token === "string" ? { token: parsed.token } : {}),
+        };
+    }
+    catch {
+        return null;
+    }
+}
+function sameLockOwner(left, right) {
+    return Boolean(left && left.pid === right.pid && left.ts === right.ts && left.token === right.token);
+}
+/**
+ * Serialize stale-lock stealers for one observed lock generation. Without this
+ * claim, contender B can replace a dead lock and contender C — acting on its
+ * earlier observation — can then rename B's fresh live lock.
+ */
+function tryStealDeadLock(lockPath, observed, owner) {
+    if (probeProcess(owner.pid) !== "dead")
+        return false;
+    const generation = crypto.createHash("sha256")
+        .update(`${observed.dev}\0${observed.ino}\0${owner.pid}\0${owner.ts}\0${owner.token ?? ""}`)
+        .digest("hex")
+        .slice(0, 16);
+    const claimPath = `${lockPath}.steal.${generation}`;
+    let claimFd;
+    try {
+        claimFd = fs.openSync(claimPath, "wx");
+    }
+    catch {
+        return false;
+    }
+    const claimToken = crypto.randomBytes(16).toString("hex");
+    let claimHandle;
+    try {
+        fs.writeFileSync(claimFd, JSON.stringify({ pid: process.pid, ts: Date.now(), token: claimToken }));
+        const claimStat = fs.fstatSync(claimFd);
+        claimHandle = { device: claimStat.dev, inode: claimStat.ino, token: claimToken };
+        fs.closeSync(claimFd);
+        claimFd = -1;
+        const current = fs.lstatSync(lockPath);
+        if (current.dev !== observed.dev || current.ino !== observed.ino)
+            return false;
+        if (!sameLockOwner(readLockOwner(lockPath), owner))
+            return false;
+        const grave = `${lockPath}.stale.${process.pid}.${crypto.randomBytes(4).toString("hex")}`;
+        fs.renameSync(lockPath, grave);
+        try {
+            fs.unlinkSync(grave);
+        }
+        catch { /* best-effort grave cleanup */ }
+        return true;
+    }
+    catch {
+        return false;
+    }
+    finally {
+        if (claimFd >= 0) {
+            try {
+                fs.closeSync(claimFd);
+            }
+            catch { /* ignore */ }
+        }
+        if (claimHandle)
+            releaseLock(claimPath, claimHandle);
+        else {
+            // Initialization did not finish; only this exclusive creator can own it.
+            try {
+                fs.unlinkSync(claimPath);
+            }
+            catch { /* ignore */ }
+        }
+    }
+}
 /**
  * Acquire a file lock by atomically creating a lock file.
  *
  * Uses O_CREAT|O_EXCL (`wx` flag) which is atomic on POSIX and NTFS.
- * Stale locks (> LOCK_STALE_MS) are stolen via an atomic rename rather than a
- * naive unlink-then-create: a plain `unlinkSync` + `openSync('wx')` has a
- * TOCTOU window where two processes both unlink the same stale lock and both
- * then create a fresh one, yielding two simultaneous holders (risk-reviewer
- * v0.0.9 audit, L1). `rename` is atomic and removes the *specific* inode the
- * caller observed: only one racing process can win the rename of that exact
- * stale file, so at most one process proceeds to re-create the lock.
+ * Stale locks (> LOCK_STALE_MS) are stolen only after their recorded owner PID
+ * is definitively dead, then moved via an atomic rename. Age alone cannot prove
+ * abandonment: stealing from a slow but live holder creates two simultaneous
+ * critical sections, and the old holder can later unlink the new holder's lock.
  * Throws on timeout.
  */
 function acquireLock(lockPath, timeoutMs = LOCK_TIMEOUT_MS) {
@@ -151,33 +351,38 @@ function acquireLock(lockPath, timeoutMs = LOCK_TIMEOUT_MS) {
     while (true) {
         try {
             const fd = fs.openSync(lockPath, "wx");
-            fs.writeFileSync(fd, JSON.stringify({ pid: process.pid, ts: Date.now() }));
-            fs.closeSync(fd);
-            return; // lock acquired
+            const token = crypto.randomBytes(16).toString("hex");
+            try {
+                fs.writeFileSync(fd, JSON.stringify({ pid: process.pid, ts: Date.now(), token }));
+                const stat = fs.fstatSync(fd);
+                return { device: stat.dev, inode: stat.ino, token };
+            }
+            catch (error) {
+                // Creation succeeded but initialization did not. Remove only the inode
+                // this process created so a partial lock cannot become unstealable.
+                try {
+                    const held = fs.fstatSync(fd);
+                    const current = fs.lstatSync(lockPath);
+                    if (current.dev === held.dev && current.ino === held.ino)
+                        fs.unlinkSync(lockPath);
+                }
+                catch { /* best-effort rollback */ }
+                throw error;
+            }
+            finally {
+                fs.closeSync(fd);
+            }
         }
         catch (e) {
             if (e.code !== "EEXIST")
                 throw e;
             // Lock file exists — check if stale.
             try {
-                const stat = fs.statSync(lockPath);
+                const stat = fs.lstatSync(lockPath);
                 if (Date.now() - stat.mtimeMs > LOCK_STALE_MS) {
-                    // Stale lock — steal it via atomic rename so only one racing
-                    // stealer can win (L1). The "graveyard" name is unique per
-                    // process+attempt; the winner unlinks it, losers see ENOENT
-                    // on their own rename and simply retry the acquire loop.
-                    const grave = `${lockPath}.stale.${process.pid}.${crypto.randomBytes(4).toString("hex")}`;
-                    try {
-                        fs.renameSync(lockPath, grave);
-                        // We won the steal — discard the graveyard copy and retry
-                        // the loop, where openSync('wx') will create a fresh lock.
-                        try {
-                            fs.unlinkSync(grave);
-                        }
-                        catch { /* ignore */ }
-                    }
-                    catch { /* lost the steal race (ENOENT) — just retry */ }
-                    continue;
+                    const owner = readLockOwner(lockPath);
+                    if (owner && tryStealDeadLock(lockPath, stat, owner))
+                        continue;
                 }
             }
             catch {
@@ -194,25 +399,31 @@ function acquireLock(lockPath, timeoutMs = LOCK_TIMEOUT_MS) {
     }
 }
 /**
- * Release a file lock by deleting the lock file.  Ignores ENOENT (already
- * released by another process or stolen due to staleness).
+ * Release a file lock only when the path still carries this holder's inode and
+ * random ownership token. A dead-owner steal must never let an old finally
+ * block delete the successor's lock.
  */
-function releaseLock(lockPath) {
+function releaseLock(lockPath, handle) {
     try {
+        const stat = fs.lstatSync(lockPath);
+        if (!stat.isFile() || stat.dev !== handle.device || stat.ino !== handle.inode)
+            return;
+        if (readLockOwner(lockPath)?.token !== handle.token)
+            return;
         fs.unlinkSync(lockPath);
     }
-    catch { /* ENOENT or other — ignore */ }
+    catch { /* ENOENT, replaced, or corrupt — never unlink another owner's lock */ }
 }
 /**
  * Execute `fn` while holding a file lock.  Guarantees release even on throw.
  */
-export function withLock(lockPath, fn) {
-    acquireLock(lockPath);
+export function withLock(lockPath, fn, timeoutMs = LOCK_TIMEOUT_MS) {
+    const handle = acquireLock(lockPath, timeoutMs);
     try {
         return fn();
     }
     finally {
-        releaseLock(lockPath);
+        releaseLock(lockPath, handle);
     }
 }
 // ---------------------------------------------------------------------------
@@ -220,15 +431,20 @@ export function withLock(lockPath, fn) {
 // ---------------------------------------------------------------------------
 /**
  * Extract a RunIndexEntry from a RunState + computed relative path.
+ * Exported for tests; pure.
  */
-function extractIndexEntry(state, relPath) {
+export function extractIndexEntry(state, relPath) {
     return {
         runId: state.runId,
         flowName: state.flowName,
+        cwd: state.cwd,
         status: state.status,
         createdAt: state.createdAt,
         updatedAt: state.updatedAt,
         relPath,
+        ...(state.host !== undefined ? { host: state.host } : {}),
+        ...(state.packageVersion !== undefined ? { packageVersion: state.packageVersion } : {}),
+        ...(state.parentRunId !== undefined ? { parentRunId: state.parentRunId } : {}),
     };
 }
 /** Read the index file; return [] on any error (missing, corrupt, etc.). */
@@ -238,8 +454,10 @@ function readIndex(runsRoot) {
         const parsed = JSON.parse(raw);
         if (!Array.isArray(parsed))
             return [];
-        // Validate each entry minimally.
-        return parsed.filter((e) => e && typeof e.runId === "string" && typeof e.relPath === "string");
+        // Treat the project-controlled index as untrusted input. In particular,
+        // never let a crafted relPath escape runsRoot during load or retention.
+        return parsed.filter((e) => e && typeof e.runId === "string" && typeof e.relPath === "string" &&
+            isSafeRunIndexRelPath(e.relPath, e.runId));
     }
     catch {
         return [];
@@ -302,14 +520,10 @@ function rebuildIndex(runsRoot) {
             continue;
         }
         for (const file of files) {
-            try {
-                const raw = fs.readFileSync(path.join(dirPath, file), "utf-8");
-                const state = JSON.parse(raw);
-                if (state && typeof state.runId === "string") {
-                    entries.set(state.runId, extractIndexEntry(state, `${dirName}/${file}`));
-                }
+            const loaded = tryReadRunFile(runsRoot, path.join(dirPath, file));
+            if (loaded.ok && validateRunId(loaded.value.runId) && file === `${loaded.value.runId}.json`) {
+                entries.set(loaded.value.runId, extractIndexEntry(loaded.value, `${dirName}/${file}`));
             }
-            catch { /* skip corrupt */ }
         }
     }
     // Scan legacy flat files (runs/*.json, skip index.json).
@@ -323,14 +537,11 @@ function rebuildIndex(runsRoot) {
     for (const file of flatFiles) {
         if (entries.has(file.replace(/\.json$/, "")))
             continue; // prefer subdir entry
-        try {
-            const raw = fs.readFileSync(path.join(runsRoot, file), "utf-8");
-            const state = JSON.parse(raw);
-            if (state && typeof state.runId === "string" && !entries.has(state.runId)) {
-                entries.set(state.runId, extractIndexEntry(state, file));
-            }
+        const loaded = tryReadRunFile(runsRoot, path.join(runsRoot, file));
+        if (loaded.ok && validateRunId(loaded.value.runId) && file === `${loaded.value.runId}.json` &&
+            !entries.has(loaded.value.runId)) {
+            entries.set(loaded.value.runId, extractIndexEntry(loaded.value, file));
         }
-        catch { /* skip corrupt */ }
     }
     const scanned = Array.from(entries.values());
     // Persist the rebuilt index under the index lock. Re-read the current
@@ -351,127 +562,203 @@ function rebuildIndex(runsRoot) {
 // TTL / cap cleanup
 // ---------------------------------------------------------------------------
 /**
- * Remove excess and expired terminal (completed/failed) runs.
+ * Remove excess and expired inactive runs.
  *
  * Called opportunistically at the end of saveRun.  Throttled to at most once
- * per CLEANUP_INTERVAL_MS.  Active runs (running/paused/blocked) are never
- * touched.
+ * per CLEANUP_INTERVAL_MS. Only actually executing (`running`) runs are never
+ * touched. Paused and blocked runs remain resumable/inspectable inside the
+ * configured retention window, but no longer bypass it forever.
  *
  * The index read-modify-write is performed under the index lock so it cannot
  * race a concurrent updateIndexEntry and clobber a freshly-added entry (M1).
  * We re-read the index *inside* the lock (rather than trusting a snapshot read
  * before locking) so the rewrite reflects the latest committed state. File and
- * directory unlinks happen after the lock is released to keep the critical
- * section short; deleting a file that is no longer in the index is harmless.
+ * directory unlinks happen after the index lock is released, but each candidate
+ * is revalidated while holding its own run lock. This prevents cleanup from
+ * unlinking a run that was resumed or otherwise saved after selection.
  */
 function cleanupTerminalRuns(runsRoot, maxKeep = DEFAULT_MAX_KEPT_TERMINAL, maxAgeDays = DEFAULT_MAX_AGE_DAYS) {
-    const cleanupStarted = Date.now();
-    const now = cleanupStarted;
-    if (now - lastCleanupAt < CLEANUP_INTERVAL_MS)
+    const now = Date.now();
+    if (!shouldRunCleanup(runsRoot, now))
         return;
-    lastCleanupAt = now;
     const maxAgeMs = maxAgeDays * 86_400_000;
     let toRemove = [];
-    withLock(indexLockPath(runsRoot), () => {
-        const entries = readIndex(runsRoot);
-        const terminal = [];
-        const active = [];
-        for (const e of entries) {
-            if (e.status === "completed" || e.status === "failed") {
-                terminal.push(e);
-            }
-            else {
-                active.push(e);
-            }
-        }
-        // Sort terminal by updatedAt desc (newest first).
-        // Filter out entries with corrupt updatedAt (non-numeric/NaN) BEFORE sorting
-        // to prevent NaN from corrupting sort order. Corrupt entries cannot be
-        // reliably aged, so they are always moved to toRemove.
-        const cleanTerminal = [];
-        for (const e of terminal) {
-            if (typeof e.updatedAt === "number" && !Number.isNaN(e.updatedAt)) {
-                cleanTerminal.push(e);
+    try {
+        withLock(indexLockPath(runsRoot), () => {
+            const entries = readIndex(runsRoot);
+            const terminal = [];
+            const active = [];
+            for (const e of entries) {
+                if (e.status !== "running") {
+                    terminal.push(e);
+                }
+                else {
+                    active.push(e);
+                }
             }
-            else {
-                toRemove.push(e);
+            // Sort terminal by updatedAt desc (newest first).
+            // Filter out entries with corrupt updatedAt (non-numeric/NaN) BEFORE sorting
+            // to prevent NaN from corrupting sort order. Corrupt entries cannot be
+            // reliably aged, so they are always moved to toRemove.
+            const cleanTerminal = [];
+            for (const e of terminal) {
+                if (typeof e.updatedAt === "number" && !Number.isNaN(e.updatedAt)) {
+                    cleanTerminal.push(e);
+                }
+                else {
+                    toRemove.push(e);
+                }
             }
-        }
-        cleanTerminal.sort((a, b) => b.updatedAt - a.updatedAt);
-        for (let i = 0; i < cleanTerminal.length; i++) {
-            const e = cleanTerminal[i];
-            const expiredByAge = now - e.updatedAt > maxAgeMs;
-            const excessByCount = i >= maxKeep;
-            if (expiredByAge || excessByCount) {
-                toRemove.push(e);
+            cleanTerminal.sort((a, b) => b.updatedAt - a.updatedAt);
+            for (let i = 0; i < cleanTerminal.length; i++) {
+                const e = cleanTerminal[i];
+                const expiredByAge = maxAgeDays > 0 && now - e.updatedAt > maxAgeMs;
+                const excessByCount = maxKeep > 0 && i >= maxKeep;
+                if (expiredByAge || excessByCount) {
+                    toRemove.push(e);
+                }
             }
-        }
-        if (toRemove.length === 0)
-            return;
-        // Commit the pruned index while holding the lock so a concurrent
-        // updateIndexEntry cannot interleave and lose entries.
-        const remaining = cleanTerminal.filter((e) => !toRemove.includes(e));
-        writeIndex(runsRoot, [...active, ...remaining]);
-    });
+            if (toRemove.length === 0)
+                return;
+            // Commit the pruned index while holding the lock so a concurrent
+            // updateIndexEntry cannot interleave and lose entries.
+            const removalSet = new Set(toRemove);
+            const remaining = cleanTerminal.filter((e) => !removalSet.has(e));
+            writeIndex(runsRoot, [...active, ...remaining]);
+        }, CLEANUP_LOCK_TIMEOUT_MS);
+    }
+    catch {
+        // Retention is opportunistic. A busy index must not stall or fail saveRun.
+        return;
+    }
     if (toRemove.length === 0)
         return;
-    console.warn(`[taskflow] Cleaning up ${toRemove.length} old run(s) ` +
-        `(max ${maxKeep} runs, ${maxAgeDays} day age limit). ` +
-        `Configure 'taskflow.maxKeptRuns' / 'taskflow.maxRunAgeDays' in settings.json (0 = keep all).`);
-    // Delete run files + lock files (outside the index lock).
+    // Delete artifacts outside the index lock, but under the exact per-run lock
+    // used by saveRun. Only the entries whose on-disk snapshots still match the
+    // selected index entries are safe to remove.
+    const removed = [];
     for (const e of toRemove) {
-        const filePath = path.join(runsRoot, e.relPath);
-        // Race guard: skip files modified after cleanup started (Finding 2).
-        try {
-            if (fs.statSync(filePath).mtimeMs > cleanupStarted)
-                continue;
-        }
-        catch {
-            continue;
-        }
-        try {
-            fs.unlinkSync(filePath);
-        }
-        catch { /* already gone */ }
-        // Also remove any orphaned lock file.
-        try {
-            fs.unlinkSync(filePath + ".lock");
-        }
-        catch { /* ignore */ }
-        // Also remove the deterministic-replay trace (sibling .trace.jsonl +
-        // its append lock) so pruned runs don't accumulate trace files.
-        try {
-            fs.unlinkSync(filePath.replace(/\.json$/, ".trace.jsonl"));
-        }
-        catch { /* ignore */ }
-        try {
-            fs.unlinkSync(filePath.replace(/\.json$/, ".trace.jsonl.lock"));
-        }
-        catch { /* ignore */ }
-        // Also remove the per-run Shared Context Tree directory (C6). Orphaned
-        // ctx dirs would otherwise accumulate under runs/ctx/ over many runs.
-        try {
-            fs.rmSync(path.join(runsRoot, "ctx", e.runId), { recursive: true, force: true });
-        }
-        catch { /* ignore */ }
-        // Also remove the per-run isolated-workspace dir tree (cwd:"dedicated").
-        // `dedicated` workspaces are persistent by design; reclaim them once the
-        // run is pruned. The dir name uses the same sanitization as workspace.ts.
-        try {
-            const wsSeg = e.runId.replace(/[^A-Za-z0-9._-]/g, "_").replace(/^\.+/, "_").slice(0, 100) || "phase";
-            fs.rmSync(path.join(runsRoot, "ws", wsSeg), { recursive: true, force: true });
-        }
-        catch { /* ignore */ }
+        if (cleanupRunArtifactsIfSnapshotMatches(runsRoot, e))
+            removed.push(e);
+    }
+    if (removed.length > 0) {
+        console.warn(`[taskflow] Cleaning up ${removed.length} old run(s) ` +
+            `(max ${maxKeep} runs, ${maxAgeDays} day age limit). ` +
+            `Configure 'taskflow.maxKeptRuns' / 'taskflow.maxRunAgeDays' in settings.json (0 = keep all).`);
     }
     // Remove empty flow subdirectories.
-    for (const e of toRemove) {
-        const dirPath = path.dirname(path.join(runsRoot, e.relPath));
+    for (const e of removed) {
+        const dirPath = path.dirname(runIndexFilePath(runsRoot, e.relPath));
         try {
             fs.rmdirSync(dirPath);
         }
         catch { /* ENOTEMPTY or ENOENT — ignore */ }
     }
 }
+/**
+ * Remove one retention candidate if it is still the exact state selected from
+ * the index. Returning false is deliberately fail-open: the run remains on
+ * disk, and a changed valid snapshot is restored to the index.
+ */
+function cleanupRunArtifactsIfSnapshotMatches(runsRoot, entry) {
+    if (!isSafeRunIndexRelPath(entry.relPath, entry.runId))
+        return false;
+    const filePath = runIndexFilePath(runsRoot, entry.relPath);
+    const fileDir = path.dirname(filePath);
+    const directorySnapshot = physicalDirectoryInsideRunsRoot(runsRoot, fileDir);
+    if (!directorySnapshot)
+        return false;
+    const restore = (state) => {
+        try {
+            updateIndexEntry(runsRoot, extractIndexEntry(state, entry.relPath));
+        }
+        catch { /* best effort */ }
+    };
+    try {
+        return withLock(`${filePath}.lock`, () => {
+            if (!physicalDirectoryStillMatches(runsRoot, fileDir, directorySnapshot))
+                return false;
+            // saveRun holds this same run lock until its index upsert commits. If a
+            // save won the race after cleanup pruned the old entry, that fresh entry
+            // is now visible and owns the file even when Date.now() reused the same
+            // millisecond/status values. Never remove a re-indexed candidate.
+            const wasReindexed = withLock(indexLockPath(runsRoot), () => readIndex(runsRoot).some((current) => current.runId === entry.runId), CLEANUP_LOCK_TIMEOUT_MS);
+            if (wasReindexed)
+                return false;
+            let state;
+            let fileSnapshot;
+            try {
+                const stat = fs.lstatSync(filePath);
+                if (!stat.isFile() || stat.isSymbolicLink())
+                    return false;
+                fileSnapshot = { device: stat.dev, inode: stat.ino };
+                const loaded = tryReadRunFile(runsRoot, filePath);
+                if (!loaded.ok)
+                    return false;
+                state = loaded.value;
+            }
+            catch {
+                return false;
+            }
+            if (state.runId !== entry.runId)
+                return false;
+            if (state.status === "running" || state.status !== entry.status || state.updatedAt !== entry.updatedAt) {
+                restore(state);
+                return false;
+            }
+            try {
+                if (!physicalDirectoryStillMatches(runsRoot, fileDir, directorySnapshot)) {
+                    restore(state);
+                    return false;
+                }
+                const currentFile = fs.lstatSync(filePath);
+                if (!currentFile.isFile() || currentFile.isSymbolicLink() ||
+                    currentFile.dev !== fileSnapshot.device || currentFile.ino !== fileSnapshot.inode) {
+                    restore(state);
+                    return false;
+                }
+                fs.unlinkSync(filePath);
+            }
+            catch {
+                restore(state);
+                return false;
+            }
+            // Remove deterministic-replay trace while respecting its append lock.
+            const tracePath = filePath.replace(/\.json$/, ".trace.jsonl");
+            try {
+                withLock(`${tracePath}.lock`, () => { try {
+                    fs.unlinkSync(tracePath);
+                }
+                catch { /* missing */ } }, CLEANUP_LOCK_TIMEOUT_MS);
+            }
+            catch { /* best effort */ }
+            // Remove per-run Shared Context Tree and isolated-workspace artifacts.
+            removeArtifactDirectoryInsideRunsRoot(runsRoot, path.join(runsRoot, "ctx", entry.runId));
+            const wsSeg = entry.runId.replace(/[^A-Za-z0-9._-]/g, "_").replace(/^\.+/, "_").slice(0, 100) || "phase";
+            removeArtifactDirectoryInsideRunsRoot(runsRoot, path.join(runsRoot, "ws", wsSeg));
+            // Remove user-private detached control records left by an interrupted worker.
+            const controlCwd = controlCwdForRunsRoot(runsRoot, state.cwd);
+            const controlDir = detachedControlDir(controlCwd);
+            try {
+                fs.unlinkSync(path.join(controlDir, `${entry.runId}.cancel.json`));
+            }
+            catch { /* ignore */ }
+            try {
+                fs.unlinkSync(path.join(controlDir, `${entry.runId}.processes.json`));
+            }
+            catch { /* ignore */ }
+            try {
+                fs.rmdirSync(controlDir);
+            }
+            catch { /* other runs / missing */ }
+            return true;
+        }, CLEANUP_LOCK_TIMEOUT_MS);
+    }
+    catch {
+        // Retention is opportunistic and must never make saveRun fail.
+        return false;
+    }
+}
 // ---------------------------------------------------------------------------
 // Original helpers (unchanged)
 // ---------------------------------------------------------------------------
@@ -718,6 +1005,22 @@ export function runsDir(cwd) {
     const projDir = findProjectFlowsDir(cwd, true);
     return path.join(projDir, "runs");
 }
+/**
+ * User-private control directory for one invocation root.
+ *
+ * Cancellation and process-registry markers must not live below a repository:
+ * a checked-out `.pi` tree can contain symlinks controlled by project content.
+ * Bind the control plane to the canonical directory identity and keep only its
+ * digest in the user agent directory, so sibling worktrees remain isolated.
+ */
+export function detachedControlDir(cwd) {
+    const identity = directoryIdentity(cwd);
+    const source = identity
+        ? `${identity.canonicalPath}\0${identity.device}\0${identity.inode}`
+        : path.resolve(cwd);
+    const key = crypto.createHash("sha256").update(source).digest("hex");
+    return path.join(getAgentDir(), "taskflow-control", key);
+}
 /** Root dir for the cross-run memoization cache (sibling of `runs`). */
 export function cacheDir(cwd) {
     const projDir = findProjectFlowsDir(cwd, true);
@@ -793,8 +1096,9 @@ export function loadRunDiagnosed(cwd, runId) {
     let corrupt = null;
     const probe = (filePath) => {
         const r = tryReadRunFile(root, filePath);
+        // A filename/index record must never alias a different run's state.
         if (r.ok)
-            return r.value;
+            return r.value.runId === runId ? r.value : undefined;
         if (r.reason === "unparseable" && !corrupt)
             corrupt = r;
         return undefined;
@@ -897,31 +1201,30 @@ export function listRuns(cwd, limit = 20) {
         const runIdFromName = file.replace(/\.json$/, "");
         if (indexRunIds.has(runIdFromName))
             continue;
-        try {
-            const raw = fs.readFileSync(path.join(root, file), "utf-8");
-            const state = JSON.parse(raw);
-            if (state && typeof state.runId === "string" && !indexRunIds.has(state.runId)) {
-                entries.push(extractIndexEntry(state, file));
-                indexRunIds.add(state.runId);
-            }
+        const loaded = tryReadRunFile(root, path.join(root, file));
+        if (loaded.ok && validateRunId(loaded.value.runId) && file === `${loaded.value.runId}.json` &&
+            !indexRunIds.has(loaded.value.runId)) {
+            entries.push(extractIndexEntry(loaded.value, file));
+            indexRunIds.add(loaded.value.runId);
         }
-        catch { /* skip corrupt */ }
     }
-    // Sort by updatedAt desc, slice to limit.
+    // Sort by updatedAt desc. Invalid/unreadable files do not consume the limit:
+    // otherwise one corrupt or replaced high-ranked entry could hide healthy
+    // history that follows it.
     // Filter out entries with non-numeric/NaN updatedAt BEFORE sorting to
     // prevent NaN from corrupting V8's sort order (which can displace valid
     // entries when a limit is applied).
     const valid = entries.filter((e) => typeof e.updatedAt === "number" && !Number.isNaN(e.updatedAt));
     valid.sort((a, b) => b.updatedAt - a.updatedAt);
-    const sliced = valid.slice(0, limit);
+    const targetLimit = Number.isFinite(limit) ? Math.max(0, Math.floor(limit)) : valid.length;
     // Read full RunState for each entry.
     const runs = [];
-    for (const e of sliced) {
-        try {
-            const raw = fs.readFileSync(path.join(root, e.relPath), "utf-8");
-            runs.push(JSON.parse(raw));
-        }
-        catch { /* file may have been deleted since index was built — skip */ }
+    for (const e of valid) {
+        if (runs.length >= targetLimit)
+            break;
+        const loaded = tryReadRunFile(root, runIndexFilePath(root, e.relPath));
+        if (loaded.ok && loaded.value.runId === e.runId)
+            runs.push(loaded.value);
     }
     // F-010: filter out records with non-numeric/NaN updatedAt.
     return runs.filter((r) => typeof r.updatedAt === "number" && !Number.isNaN(r.updatedAt));
@@ -930,20 +1233,30 @@ export function listRuns(cwd, limit = 20) {
 export function hashInput(...parts) {
     return crypto.createHash("sha256").update(parts.join("\u0000")).digest("hex").slice(0, 16);
 }
-/**
- * Check whether a process with the given PID is still alive.
- * Uses signal 0 (no signal sent) — succeeds if the process exists and we have
- * permission to signal it, throws ESRCH if it doesn't exist.
- */
-export function isProcessAlive(pid) {
+export function probeProcess(pid, signalZero = (candidate, signal) => process.kill(candidate, signal)) {
+    // Node/libuv rejects values outside the signed 32-bit PID range before an
+    // OS probe occurs; those are definitively not live process identifiers.
+    if (!Number.isSafeInteger(pid) || pid <= 0 || pid > 0x7fffffff)
+        return "dead";
     try {
-        process.kill(pid, 0);
-        return true;
+        signalZero(pid, 0);
+        return "alive";
     }
-    catch {
-        return false;
+    catch (error) {
+        const code = error && typeof error === "object" && "code" in error
+            ? String(error.code ?? "")
+            : "";
+        if (code === "ESRCH")
+            return "dead";
+        // EPERM proves that a process exists but its identity is not observable.
+        // Unknown platform errors must likewise never terminalize a live run.
+        return "unknown";
     }
 }
+/** Back-compatible boolean probe. Unknown is conservatively treated as alive. */
+export function isProcessAlive(pid) {
+    return probeProcess(pid) !== "dead";
+}
 /**
  * Write a file atomically: write to a unique temp file in the same directory,
  * then rename over the target (rename is atomic on the same filesystem). Prevents