@lmzhen/dsh-evolution-core 0.4.0 → 0.4.1

package/lib/index.js

@@ -328,7 +328,10 @@ const LOCK_BODY_RE = /^(\d+):[0-9a-f]*$/;
 /** Parse a writer-lock body into its holder pid. `null` when the body does
 * not have the `pid:token` shape at all (e.g. a user support file named
 * `*.lock`) — callers leave such files alone. A shape-matching body always
-* yields a number (possibly `0`, which `isProcessAlive` treats as dead). */
+* yields a number, possibly `0` (state corruption / hand edit): there is no
+* pid 0, so a `0` holder must be treated as DEAD by the caller —
+* `isProcessAlive(0)` signals the caller's own process group on POSIX and
+* answers true, so probe sites must guard `holder > 0` before probing. */
 function parseLockBody(body) {
 	const match = LOCK_BODY_RE.exec(body.trim());
 	return match === null ? null : Number(match[1]);
@@ -365,6 +368,14 @@ const EMPTY_LOCK_TAKEOVER_MS = 3e4;
 /** A body with no parseable pid (crash mid-write): 1h, far above any legal hold
 * and far below "forever". */
 const LOCK_TEAR_TAKEOVER_MS = 36e5;
+/** A2 (audit P1-2): even a lock whose holder pid probes ALIVE is reclaimable
+* past this age. Liveness-by-pid cannot distinguish the original holder from
+* an unrelated process the OS later assigned the same pid, so the plain alive
+* probe let one recycled pid brick every writer of one state file forever.
+* No write in this family holds a lock for more than minutes (the longest is
+* the ~120s review window), so a day-old "alive" lock is a recycled pid, not
+* a live writer. */
+const ALIVE_LOCK_TAKEOVER_MS = 864e5;
 /** P2-27 (v37): the commit-point ownership re-read. A transient read failure
 * (EACCES/EMFILE/antivirus hold) must not abort a valid RMW, so the read is
 * retried in place; only a still-unreadable lock fails the attempt. */
@@ -385,8 +396,11 @@ function isCommittedWarning(error) {
 * V27 G1.1: the lock-takeover decision as ONE pure function, so the protocol is
 * testable and exhaustive instead of being an inline expression inside the
 * acquisition loop:
-*   - `none`    the lock is fresh, or its holder is alive → wait, never steal;
-*   - `dead`    a named holder that is gone, past the dead threshold;
+*   - `none`    the lock is fresh, or its holder is alive within the alive
+*               window → wait, never steal;
+*   - `dead`    a named holder that is gone past the dead threshold, OR whose
+*               pid still probes alive but whose lock is older than the alive
+*               window (A2: a recycled pid must not hold the file forever);
 *   - `empty`   no body at all: nothing attributes it to a holder, so only the
 *               wide `emptyAfterMs` window may reclaim it;
 *   - `corrupt` a body with no parseable pid (a crash mid-write), past the 1h
@@ -401,7 +415,7 @@ function decideTakeover(probe) {
 	const namedHolder = Number.isInteger(holder) && holder > 0;
 	if (probe.body === "") return age > (probe.emptyAfterMs ?? 3e4) ? "empty" : "none";
 	if (!namedHolder) return age > (probe.corruptAfterMs ?? 36e5) ? "corrupt" : "none";
-	if (probe.alive(holder)) return "none";
+	if (probe.alive(holder)) return age > (probe.aliveAfterMs ?? 864e5) ? "dead" : "none";
 	return age > (probe.deadAfterMs ?? 1e3) ? "dead" : "none";
 }
 /**
@@ -1692,6 +1706,9 @@ async function appendEvolutionEvent(io, path, event, rotateAt = EVENT_LOG_ROTATE
 	let refuseMessage = "";
 	let parsedBody = null;
 	await transactIo(io, path, async (current) => {
+		parsedBody = null;
+		refuseMessage = "";
+		assigned = 0;
 		if (current !== null && current.trim() !== "") {
 			let shape;
 			try {
@@ -3887,9 +3904,10 @@ const SECRET_PATTERNS = [
 	["stripe key", /[sr]k_(?:live|test)_[A-Za-z0-9]{16,}/g],
 	["github fine-grained token", /github_pat_[A-Za-z0-9_]{20,}/g],
 	["google api key", /AIza[0-9A-Za-z_-]{30,}/g],
-	["bearer credential", /Bearer[\s]+[a-z0-9._~+/=\-]{16,}/gi]
+	["bearer credential", /Bearer[\s]+[a-z0-9._~+/=\-]{16,}/gi],
+	["basic credential", /\bBasic[\s]+[a-z0-9._~+/=]{16,}/gi]
 ];
-const INLINE_ASSIGNMENT_PATTERN = /* @__PURE__ */ new RegExp("(^|[^\\w-])([\\w-]*[_\\-])?((?:token|api[_-]?key|secret|password|passwd)(?:[_\\-][\\w-]*)?)\\b([\"\\']?[\\t ]*[:=][\\t ]*)([^\\r\\n]+)", "gi");
+const INLINE_ASSIGNMENT_PATTERN = /* @__PURE__ */ new RegExp("(^|[^\\w-])([\\w-]*[_\\-])?((?:token|api[_-]?key|secret|password|passwd|authorization)(?:[_\\-][\\w-]*)?)\\b([\"\\']?[\\t ]*[:=][\\t ]*)([^\\r\\n]+)", "gi");
 const URL_CREDENTIALS_PATTERN = /([a-z][a-z0-9+.-]{0,63}:\/\/[^\s:/@]+:)([^\s/@]+)@/gi;
 const PEM_PRIVATE_KEY_PATTERN = new RegExp(`-----BEGIN\\s+(?:(?:RSA|EC|DSA|OPENSSH|ENCRYPTED|PGP)\\s+)?PRIVATE\\s+KEY(?:\\s+BLOCK)?-----[\\s\\S]*?-----END\\s+(?:(?:RSA|EC|DSA|OPENSSH|ENCRYPTED|PGP)\\s+)?PRIVATE\\s+KEY(?:\\s+BLOCK)?-----`, "g");
 const BLOCK_KEY_ONLY_LINE = /(?:^|\s)([\w-]*[_\-])?((?:token|api[_-]?key|secret|password|passwd)(?:[_\-][\w-]*)?)\s*:(?:\r)?$/i;
@@ -4996,15 +5014,18 @@ function frontmatterBlock(content) {
 * strict catalog cannot load". `frontmatterCatalogInvalid` publishes it and
 * `normalizeFrontmatter` decides each rewrite with the same predicate
 * (`yamlPlainScalarNeedsQuotes`), so the audit verdict and the write path can
-* never disagree. Only single-line `key: value` entries are judged; a line with
-* embedded breaks is skipped.
-*/
-function unsafeFrontmatterEntries(block, nl) {
+* never disagree. A9 (audit P2-12): the scan is ENDING-AGNOSTIC — lines are
+* split on `\n` with the per-line CR stripped, exactly the iteration the
+* rewrite path uses. The former split on the block's FIRST-LINE ending turned
+* a mixed-ending block (first line CRLF, entries LF) into one unsplit chunk
+* whose entries were all skipped, so an unquoted ` #` value read as CLEAN
+* while the catalog silently dropped it and the next edit would have quoted
+* it. Only single-line `key: value` entries are judged.
+*/
+function unsafeFrontmatterEntries(block) {
 	const found = [];
-	for (const line of block.split(nl)) {
-		const clean = withoutCr(line);
-		if (clean.includes("\n") || clean.includes("\r")) continue;
-		const match = /^([A-Za-z0-9_-]+):\s*(.*)$/.exec(clean);
+	for (const rawLine of block.split("\n")) {
+		const match = /^([A-Za-z0-9_-]+):\s*(.*)$/.exec(withoutCr(rawLine));
 		if (!match) continue;
 		const key = match[1];
 		if (key === void 0) continue;
@@ -5153,7 +5174,7 @@ function readFrontmatterBlock(content) {
 	return {
 		frontmatter,
 		body,
-		unsafeValues: unsafeFrontmatterEntries(found.block, found.nl),
+		unsafeValues: unsafeFrontmatterEntries(found.block),
 		strictFailed: strict === null,
 		platformStringSplit: strict?.split ?? []
 	};
@@ -6199,20 +6220,28 @@ var SkillLibrary = class {
 	* Structure-health facts for one skill (rc.73 A1, 008 design): body
 	* chars/density from SKILL.md, support groups from countSupportDirs, plus
 	* optional usage counts (A2 churn dimension) when the caller has them.
-	* Derived, never persisted; null when the skill is unreadable.
+	* Derived, never persisted. CONTRACT: `null` whenever the skill cannot be
+	* read — a missing file AND any read failure (EACCES/EIO/…) both answer
+	* null, so a whole health view degrades one ROW instead of throwing out of
+	* its per-skill loop (A6, audit P2-10: the former code absorbed only
+	* missing/EISDIR and let a transient win32 hold kill the entire view).
 	*/
 	async assessHealth(rawName, thresholds = DEFAULT_HEALTH_THRESHOLDS, counts) {
-		const name = rawName.trim();
-		const content = await this.read(name);
-		if (content === null) return null;
-		return assessStructureHealth({
-			skillName: name,
-			bodyChars: content.length,
-			bodyText: content,
-			supportGroups: await this.countSupportDirs(name),
-			patchCount: counts?.patchCount,
-			readCount: counts?.readCount
-		}, thresholds);
+		try {
+			const name = rawName.trim();
+			const content = await this.read(name);
+			if (content === null) return null;
+			return assessStructureHealth({
+				skillName: name,
+				bodyChars: content.length,
+				bodyText: content,
+				supportGroups: await this.countSupportDirs(name),
+				patchCount: counts?.patchCount,
+				readCount: counts?.readCount
+			}, thresholds);
+		} catch {
+			return null;
+		}
 	}
 	/** Best-effort audit trail entry; never blocks the mutation. */
 	async audit(skillName, action, before, after, summary) {
@@ -6388,13 +6417,13 @@ var SkillLibrary = class {
 			ok: false,
 			message: `Skill "${normalized}" already exists.`
 		};
+		let markerWarning = "";
 		if (origin !== "foreground") {
-			let markerDurabilityWarning = "";
 			try {
 				await this.io.writeText(markerPath(dir, "hermes-managed"), "");
 			} catch (error) {
-				if (!isCommittedOnly(error)) throw error;
-				markerDurabilityWarning = error instanceof Error ? error.message : String(error);
+				if (isCommittedOnly(error)) markerWarning = `the hermes-managed marker landed but its directory fsync failed — durability unconfirmed: ${error instanceof Error ? error.message : String(error)}`;
+				else markerWarning = `hermes-managed marker write failed (${error instanceof Error ? error.message : String(error)}); the skill landed but is NOT lifecycle-managed — recreate or pin it manually`;
 			}
 			if (!await this.io.exists(createPath)) {
 				await this.io.remove(markerPath(dir, "hermes-managed")).catch(() => {});
@@ -6404,7 +6433,6 @@ var SkillLibrary = class {
 					message: `Skill "${normalized}" was archived concurrently while being created; the partial marker was removed — retry once the mover settles.`
 				};
 			}
-			if (markerDurabilityWarning !== "") createDurabilityWarning = createDurabilityWarning === "" ? markerDurabilityWarning : `${createDurabilityWarning}; marker: ${markerDurabilityWarning}`;
 		}
 		await this.audit(normalized, "create", null, onDisk, "created");
 		this.notifyMutation({
@@ -6412,9 +6440,10 @@ var SkillLibrary = class {
 			name: normalized,
 			skillDir: dir
 		});
+		const warnings = [createDurabilityWarning !== "" ? `the write landed but the directory fsync failed — durability unconfirmed: ${createDurabilityWarning}` : "", markerWarning].filter((warning) => warning !== "");
 		return {
 			ok: true,
-			message: `Skill "${normalized}" created.${createDurabilityWarning === "" ? "" : ` (warning: the write landed but the directory fsync failed — durability unconfirmed: ${createDurabilityWarning})`}`,
+			message: `Skill "${normalized}" created.${warnings.length === 0 ? "" : ` (warning: ${warnings.join("; ")})`}`,
 			path: dir,
 			...norm.changed ? { normalizedFrontmatterFields: norm.fields } : {}
 		};
@@ -6772,15 +6801,28 @@ var SkillLibrary = class {
 			for (const entry of entries) if (entry.endsWith(".lock")) await this.sweepLockIfStranded(join(dir, supportDir, entry));
 		}
 	}
+	/** R2 follow-up (audit of A2): the recycled-pid window from io's
+	* `ALIVE_LOCK_TAKEOVER_MS`, applied to the stranded-lock sweepers. A lock
+	* whose holder pid probes alive but whose mtime is older than the alive
+	* window is a recycled pid, not a live writer (no write in this family
+	* holds a lock for more than minutes) — without this, one recycled pid
+	* blocked whole-tree snapshot recovery forever with a "retry once the
+	* write completes" message that could never become true. A backend without
+	* `mtime` keeps the conservative refuse-on-alive posture. */
+	async lockHolderAgedOut(lockPath) {
+		const mtime = await this.io.mtime?.(lockPath).catch(() => null);
+		return typeof mtime === "number" && Date.now() - mtime > 864e5;
+	}
 	/** Remove `lockPath` only when its body has the writer-lock `pid:token`
-	* shape AND the holder pid is not alive; anything else (a user support file
-	* or a live writer's lock) is left untouched. */
+	* shape AND the holder pid is not alive (or is an aged-out recycled pid —
+	* R2 follow-up); anything else (a user support file or a live writer's
+	* lock) is left untouched. */
 	async sweepLockIfStranded(lockPath) {
 		const body = await this.io.readText(lockPath).catch(() => null);
 		if (body === null) return;
 		const pid = parseLockBody(body);
 		if (pid === null) return;
-		if (Number.isInteger(pid) && pid > 0 && isProcessAlive(pid)) return;
+		if (Number.isInteger(pid) && pid > 0 && isProcessAlive(pid) && !await this.lockHolderAgedOut(lockPath)) return;
 		await this.io.remove(lockPath).catch(() => {});
 	}
 	/** A1-4 (v18): a manifest-declared name is copied with `join(root, name)`;
@@ -6805,7 +6847,7 @@ var SkillLibrary = class {
 		if (body === null) return;
 		const pid = parseLockBody(body);
 		if (pid === null) return;
-		if (Number.isInteger(pid) && pid > 0 && isProcessAlive(pid)) throw new Error(`snapshot restore refused: ${label} is being written (write lock present); retry once the write completes`);
+		if (Number.isInteger(pid) && pid > 0 && isProcessAlive(pid) && !await this.lockHolderAgedOut(lockPath)) throw new Error(`snapshot restore refused: ${label} is being written (write lock present); retry once the write completes`);
 		await this.io.remove(lockPath).catch(() => {});
 	}
 	async archive(rawName, options = {}) {
@@ -7975,4 +8017,4 @@ function sessionAudited(ctx, sessionId, sessionScoped) {
 	return false;
 }
 //#endregion
-export { AUTHORING_DESCRIPTION_BAR, COMBINED_REVIEW_PLAN_PROMPT, COMBINED_REVIEW_PROMPT, COMPLETION_SKILL_REVIEW_PROMPT, CURATOR_DRY_RUN_BANNER, CURATOR_PROMPT, DEAD_LOCK_TAKEOVER_MS, DEDUP_MAX_PAIR_COMPARISONS, DEFAULT_ARCHIVE_AFTER_DAYS, DEFAULT_CONSOLIDATION_FAILURES, DEFAULT_CURATOR_BOOT_GRACE_SECONDS, DEFAULT_CURATOR_INTERVAL_HOURS, DEFAULT_CURATOR_MODEL, DEFAULT_CURATOR_REVIEW_MAX_TOKENS, DEFAULT_HEALTH_THRESHOLDS, DEFAULT_MAX_OPS_PER_PLAN, DEFAULT_MEMORY_CHAR_LIMIT, DEFAULT_MEMORY_REVIEW_MODEL, DEFAULT_MIN_IDLE_HOURS, DEFAULT_MUTATION_CAP, DEFAULT_REVIEW_CONTEXT_MESSAGES, DEFAULT_REVIEW_MEMORY_INTERVAL, DEFAULT_REVIEW_MESSAGE_CHARS, DEFAULT_REVIEW_SKILL_INTERVAL, DEFAULT_REVIEW_TIMEOUT_MS, DEFAULT_SKILL_CONTENT_CHARS, DEFAULT_SKILL_LIMITS, DEFAULT_SKILL_REVIEW_COMPLETION_MIN_TOOL_CALLS, DEFAULT_SKILL_REVIEW_MODEL, DEFAULT_SKILL_REVIEW_TRIGGER, DEFAULT_STALE_AFTER_DAYS, DEFAULT_SUBSTANTIVE_MIN_AGENT_CHARS, DEFAULT_SUBSTANTIVE_MIN_TOOL_CALLS, DEFAULT_SUBSTANTIVE_MIN_USER_CHARS, DEFAULT_USER_CHAR_LIMIT, DISPATCH_EVENT_TYPES, DRIFT_MAX_LINE_CHARS, DRIFT_SIGNALS_VERSION, DRIFT_SIGNAL_NOUNS, DSH_AUTHORING_STANDARDS, EMPTY_LOCK_TAKEOVER_MS, ENTRY_DELIMITER, EVENT_ARCHIVE_PREFIX, EVENT_LOG_RETAIN_ARCHIVES, EVENT_LOG_ROTATE_AT, EVENT_LOG_VERSION, EVOLUTION_WRITE_TOOLS, EvolutionGateSet, FAMILY_SESSION_TOOL_NAMES, FORBIDDEN_CONTROL_KEYS, HEALTH_STAMP_RE, INSTANCE_KEYS, LOCK_BODY_RE, LOCK_SUFFIX, LOCK_TEAR_TAKEOVER_MS, LOW_QUALITY_THRESHOLD, LostWriteLock, MAINTAIN_OUTPUT_INSTRUCTION, MAINTAIN_PROMPT, MAX_DESCRIPTION_LENGTH, MAX_RESTRUCTURE_MOVES, MAX_SKILL_CONTENT_CHARS, MAX_SKILL_FILE_BYTES, MAX_SKILL_NAME_LENGTH, MAX_TIMER_DELAY_MS, MEMORY_GUIDANCE_SECTION_ORDER, MEMORY_REVIEW_PROMPT, MIN_STAMP_BODY_CHARS, MUTATIONS_FILE_VERSION, MemoryStore, NATIVE_CALL_EVENT, NATIVE_RESULT_EVENT, PATTERN_OVERLAP, PROMPT_BUNDLE, PROMPT_BUNDLE_ID, PROMPT_BUNDLE_VERSION, PROTECTED_BUILTIN_SKILLS, PTC_DISPATCH_EVENT, PTC_DISPATCH_START_EVENT, QUALITY_WEIGHTS, RESTRUCTURE_TARGET_RE, SKILLS_GUIDANCE, SKILLS_GUIDANCE_SECTION_ORDER, SKILL_ACTION_REQUIRED_FIELDS, SKILL_NAME_RE, SKILL_REVIEW_PLAN_PROMPT, SKILL_REVIEW_PROMPT, SNAPSHOT_EXTRA_NAME_RE, SUPPORT_DIRS, SUPPRESSED_FILE_VERSION, SkillLibrary, THREAT_EXEMPTION_HINT, ToolDispatchNormalizer, ToolDispatchPayloadError, advanceReview, allowRowCollisions, appendEvolutionEvent, applyCuratorLifecycleFields, applyCuratorMetaFields, assertDispatchPayload, assertSkillsRootAliasRetired, assessStructureHealth, authoringFeedback, buildCuratorRunReport, buildLearnPrompt, bumpPatch, bumpUse, bumpView, callingScope, claimInstance, clampedNumber, clearReviewChannel, collectReadSkillNames, composePresetComposition, computeDedupGroups, computeDriftSignals, computeLifecycleTransitions, computePrefixClusters, computeQualityScores, computeScopeView, contentHash, countDispatches, createGateSet, decideTakeover, duplicateHeadings, emptyRecord, evaluateThreat, eventsFile, evolutionEventPayloadIssue, evolutionHome, evolutionIoAdapter, evolutionRoot, findDriftSignal, foldCuratorFields, foldToolDispatches, foldTurn, frontmatterBlock, frontmatterCatalogInvalid, getRecord, instanceClaimKey, instanceClaimedWriteSites, instanceHolder, isAbsent, isCommittedWarning, isMissingPath, isPresent, isProcessAlive, isProgramToolName, isReviewChannelSession, isSkillReadToolName, isSkillToolName, isUnknown, latestActivityAt, lifecycleCandidate, listEventArchives, loadMutations, loadRowOverrides, loadSuppressedNames, loadUsage, makeSerialQueue, mapProbe, markAgentCreated, markReviewChannel, markerEntryName, memoryRoot, missingSupportPointers, mutateUsage, mutationsFile, narrowNameMatches, neutralizePromptVariables, newSkillLibrary, nodeEvolutionIo, normalizeFrontmatter, normalizeUsageRecord, observeEvent, overlongLines, parseCuratorNominations, parseEvolutionEvents, parseFrontmatter, parseLockBody, pendingSelfCleanup, persistedWriteSite, persistedWriteSites, probeAbsent, probeList, probeMtime, probePresent, probeReason, probeUnknown, readDispatchSignal, readEvolutionEvents, readEvolutionTimeline, recordMutation, redactSecrets, relatedSkillNames, releaseInstance, renameWithRetry, renderCuratorReportMarkdown, resolveExecOrigins, resolveOrigins, resolveRootConfig, resolveSkillsRoot, retainEventArchives, reviewPrompt, saveSuppressedNames, saveUsage, scanContentThreats, scanMemoryThreats, scanThreats, scopedProbeReport, sessionAudited, sessionSeesFamilyTools, skillReadNameOf, skillsRoot, suppressedFile, sweepReviewChannelSessions, transactIo, transactTaskGuard, updateSuppressedNames, usageFile, usageObserved, validateFrontmatter, validateRestructureTarget, valueOr, verifyPromptBundle, writeDurableTmp, yamlPlainScalarNeedsQuotes };
+export { ALIVE_LOCK_TAKEOVER_MS, AUTHORING_DESCRIPTION_BAR, COMBINED_REVIEW_PLAN_PROMPT, COMBINED_REVIEW_PROMPT, COMPLETION_SKILL_REVIEW_PROMPT, CURATOR_DRY_RUN_BANNER, CURATOR_PROMPT, DEAD_LOCK_TAKEOVER_MS, DEDUP_MAX_PAIR_COMPARISONS, DEFAULT_ARCHIVE_AFTER_DAYS, DEFAULT_CONSOLIDATION_FAILURES, DEFAULT_CURATOR_BOOT_GRACE_SECONDS, DEFAULT_CURATOR_INTERVAL_HOURS, DEFAULT_CURATOR_MODEL, DEFAULT_CURATOR_REVIEW_MAX_TOKENS, DEFAULT_HEALTH_THRESHOLDS, DEFAULT_MAX_OPS_PER_PLAN, DEFAULT_MEMORY_CHAR_LIMIT, DEFAULT_MEMORY_REVIEW_MODEL, DEFAULT_MIN_IDLE_HOURS, DEFAULT_MUTATION_CAP, DEFAULT_REVIEW_CONTEXT_MESSAGES, DEFAULT_REVIEW_MEMORY_INTERVAL, DEFAULT_REVIEW_MESSAGE_CHARS, DEFAULT_REVIEW_SKILL_INTERVAL, DEFAULT_REVIEW_TIMEOUT_MS, DEFAULT_SKILL_CONTENT_CHARS, DEFAULT_SKILL_LIMITS, DEFAULT_SKILL_REVIEW_COMPLETION_MIN_TOOL_CALLS, DEFAULT_SKILL_REVIEW_MODEL, DEFAULT_SKILL_REVIEW_TRIGGER, DEFAULT_STALE_AFTER_DAYS, DEFAULT_SUBSTANTIVE_MIN_AGENT_CHARS, DEFAULT_SUBSTANTIVE_MIN_TOOL_CALLS, DEFAULT_SUBSTANTIVE_MIN_USER_CHARS, DEFAULT_USER_CHAR_LIMIT, DISPATCH_EVENT_TYPES, DRIFT_MAX_LINE_CHARS, DRIFT_SIGNALS_VERSION, DRIFT_SIGNAL_NOUNS, DSH_AUTHORING_STANDARDS, EMPTY_LOCK_TAKEOVER_MS, ENTRY_DELIMITER, EVENT_ARCHIVE_PREFIX, EVENT_LOG_RETAIN_ARCHIVES, EVENT_LOG_ROTATE_AT, EVENT_LOG_VERSION, EVOLUTION_WRITE_TOOLS, EvolutionGateSet, FAMILY_SESSION_TOOL_NAMES, FORBIDDEN_CONTROL_KEYS, HEALTH_STAMP_RE, INSTANCE_KEYS, LOCK_BODY_RE, LOCK_SUFFIX, LOCK_TEAR_TAKEOVER_MS, LOW_QUALITY_THRESHOLD, LostWriteLock, MAINTAIN_OUTPUT_INSTRUCTION, MAINTAIN_PROMPT, MAX_DESCRIPTION_LENGTH, MAX_RESTRUCTURE_MOVES, MAX_SKILL_CONTENT_CHARS, MAX_SKILL_FILE_BYTES, MAX_SKILL_NAME_LENGTH, MAX_TIMER_DELAY_MS, MEMORY_GUIDANCE_SECTION_ORDER, MEMORY_REVIEW_PROMPT, MIN_STAMP_BODY_CHARS, MUTATIONS_FILE_VERSION, MemoryStore, NATIVE_CALL_EVENT, NATIVE_RESULT_EVENT, PATTERN_OVERLAP, PROMPT_BUNDLE, PROMPT_BUNDLE_ID, PROMPT_BUNDLE_VERSION, PROTECTED_BUILTIN_SKILLS, PTC_DISPATCH_EVENT, PTC_DISPATCH_START_EVENT, QUALITY_WEIGHTS, RESTRUCTURE_TARGET_RE, SKILLS_GUIDANCE, SKILLS_GUIDANCE_SECTION_ORDER, SKILL_ACTION_REQUIRED_FIELDS, SKILL_NAME_RE, SKILL_REVIEW_PLAN_PROMPT, SKILL_REVIEW_PROMPT, SNAPSHOT_EXTRA_NAME_RE, SUPPORT_DIRS, SUPPRESSED_FILE_VERSION, SkillLibrary, THREAT_EXEMPTION_HINT, ToolDispatchNormalizer, ToolDispatchPayloadError, advanceReview, allowRowCollisions, appendEvolutionEvent, applyCuratorLifecycleFields, applyCuratorMetaFields, assertDispatchPayload, assertSkillsRootAliasRetired, assessStructureHealth, authoringFeedback, buildCuratorRunReport, buildLearnPrompt, bumpPatch, bumpUse, bumpView, callingScope, claimInstance, clampedNumber, clearReviewChannel, collectReadSkillNames, composePresetComposition, computeDedupGroups, computeDriftSignals, computeLifecycleTransitions, computePrefixClusters, computeQualityScores, computeScopeView, contentHash, countDispatches, createGateSet, decideTakeover, duplicateHeadings, emptyRecord, evaluateThreat, eventsFile, evolutionEventPayloadIssue, evolutionHome, evolutionIoAdapter, evolutionRoot, findDriftSignal, foldCuratorFields, foldToolDispatches, foldTurn, frontmatterBlock, frontmatterCatalogInvalid, getRecord, instanceClaimKey, instanceClaimedWriteSites, instanceHolder, isAbsent, isCommittedWarning, isMissingPath, isPresent, isProcessAlive, isProgramToolName, isReviewChannelSession, isSkillReadToolName, isSkillToolName, isUnknown, latestActivityAt, lifecycleCandidate, listEventArchives, loadMutations, loadRowOverrides, loadSuppressedNames, loadUsage, makeSerialQueue, mapProbe, markAgentCreated, markReviewChannel, markerEntryName, memoryRoot, missingSupportPointers, mutateUsage, mutationsFile, narrowNameMatches, neutralizePromptVariables, newSkillLibrary, nodeEvolutionIo, normalizeFrontmatter, normalizeUsageRecord, observeEvent, overlongLines, parseCuratorNominations, parseEvolutionEvents, parseFrontmatter, parseLockBody, pendingSelfCleanup, persistedWriteSite, persistedWriteSites, probeAbsent, probeList, probeMtime, probePresent, probeReason, probeUnknown, readDispatchSignal, readEvolutionEvents, readEvolutionTimeline, recordMutation, redactSecrets, relatedSkillNames, releaseInstance, renameWithRetry, renderCuratorReportMarkdown, resolveExecOrigins, resolveOrigins, resolveRootConfig, resolveSkillsRoot, retainEventArchives, reviewPrompt, saveSuppressedNames, saveUsage, scanContentThreats, scanMemoryThreats, scanThreats, scopedProbeReport, sessionAudited, sessionSeesFamilyTools, skillReadNameOf, skillsRoot, suppressedFile, sweepReviewChannelSessions, transactIo, transactTaskGuard, updateSuppressedNames, usageFile, usageObserved, validateFrontmatter, validateRestructureTarget, valueOr, verifyPromptBundle, writeDurableTmp, yamlPlainScalarNeedsQuotes };

package/lib/types/io.d.ts

@@ -173,7 +173,10 @@ export declare const LOCK_BODY_RE: RegExp;
 /** Parse a writer-lock body into its holder pid. `null` when the body does
  * not have the `pid:token` shape at all (e.g. a user support file named
  * `*.lock`) — callers leave such files alone. A shape-matching body always
- * yields a number (possibly `0`, which `isProcessAlive` treats as dead). */
+ * yields a number, possibly `0` (state corruption / hand edit): there is no
+ * pid 0, so a `0` holder must be treated as DEAD by the caller —
+ * `isProcessAlive(0)` signals the caller's own process group on POSIX and
+ * answers true, so probe sites must guard `holder > 0` before probing. */
 export declare function parseLockBody(body: string): number | null;
 /**
  * V27 G0.2 (EVO-IO-01): the write lock named by this claim is no longer ours
@@ -199,6 +202,14 @@ export declare const EMPTY_LOCK_TAKEOVER_MS = 30000;
 /** A body with no parseable pid (crash mid-write): 1h, far above any legal hold
  * and far below "forever". */
 export declare const LOCK_TEAR_TAKEOVER_MS = 3600000;
+/** A2 (audit P1-2): even a lock whose holder pid probes ALIVE is reclaimable
+ * past this age. Liveness-by-pid cannot distinguish the original holder from
+ * an unrelated process the OS later assigned the same pid, so the plain alive
+ * probe let one recycled pid brick every writer of one state file forever.
+ * No write in this family holds a lock for more than minutes (the longest is
+ * the ~120s review window), so a day-old "alive" lock is a recycled pid, not
+ * a live writer. */
+export declare const ALIVE_LOCK_TAKEOVER_MS = 86400000;
 /**
  * V27 G1.3: the error `commitTmp` throws when the rename landed but the parent
  * directory fsync failed — the bytes ARE visible, only their durability is
@@ -224,13 +235,17 @@ interface TakeoverProbe {
     deadAfterMs?: number;
     emptyAfterMs?: number;
     corruptAfterMs?: number;
+    aliveAfterMs?: number;
 }
 /**
  * V27 G1.1: the lock-takeover decision as ONE pure function, so the protocol is
  * testable and exhaustive instead of being an inline expression inside the
  * acquisition loop:
- *   - `none`    the lock is fresh, or its holder is alive → wait, never steal;
- *   - `dead`    a named holder that is gone, past the dead threshold;
+ *   - `none`    the lock is fresh, or its holder is alive within the alive
+ *               window → wait, never steal;
+ *   - `dead`    a named holder that is gone past the dead threshold, OR whose
+ *               pid still probes alive but whose lock is older than the alive
+ *               window (A2: a recycled pid must not hold the file forever);
  *   - `empty`   no body at all: nothing attributes it to a holder, so only the
  *               wide `emptyAfterMs` window may reclaim it;
  *   - `corrupt` a body with no parseable pid (a crash mid-write), past the 1h

package/lib/types/skill-store.d.ts

@@ -372,7 +372,11 @@ export declare class SkillLibrary {
      * Structure-health facts for one skill (rc.73 A1, 008 design): body
      * chars/density from SKILL.md, support groups from countSupportDirs, plus
      * optional usage counts (A2 churn dimension) when the caller has them.
-     * Derived, never persisted; null when the skill is unreadable.
+     * Derived, never persisted. CONTRACT: `null` whenever the skill cannot be
+     * read — a missing file AND any read failure (EACCES/EIO/…) both answer
+     * null, so a whole health view degrades one ROW instead of throwing out of
+     * its per-skill loop (A6, audit P2-10: the former code absorbed only
+     * missing/EISDIR and let a transient win32 hold kill the entire view).
      */
     assessHealth(rawName: string, thresholds?: SkillHealthThresholds, counts?: {
         patchCount?: number;
@@ -436,9 +440,19 @@ export declare class SkillLibrary {
      * stolen by the sweep. A dead-pid residue would otherwise permanently
      * refuse archive/restore. */
     private deleteStrandedLocks;
+    /** R2 follow-up (audit of A2): the recycled-pid window from io's
+     * `ALIVE_LOCK_TAKEOVER_MS`, applied to the stranded-lock sweepers. A lock
+     * whose holder pid probes alive but whose mtime is older than the alive
+     * window is a recycled pid, not a live writer (no write in this family
+     * holds a lock for more than minutes) — without this, one recycled pid
+     * blocked whole-tree snapshot recovery forever with a "retry once the
+     * write completes" message that could never become true. A backend without
+     * `mtime` keeps the conservative refuse-on-alive posture. */
+    private lockHolderAgedOut;
     /** Remove `lockPath` only when its body has the writer-lock `pid:token`
-     * shape AND the holder pid is not alive; anything else (a user support file
-     * or a live writer's lock) is left untouched. */
+     * shape AND the holder pid is not alive (or is an aged-out recycled pid —
+     * R2 follow-up); anything else (a user support file or a live writer's
+     * lock) is left untouched. */
     private sweepLockIfStranded;
     /** A1-4 (v18): a manifest-declared name is copied with `join(root, name)`;
      * only a single, non-traversing path component is safe. Dotfiles

package/package.json

@@ -1,7 +1,7 @@
 {
   "name": "@lmzhen/dsh-evolution-core",
   "description": "Shared stores, prompts, signals and lifecycle logic for the dsh-evolution plugin family (community build)",
-  "version": "0.4.0",
+  "version": "0.4.1",
   "publishConfig": {
     "access": "public"
   },