opencode-dejavu 2.2.0 → 2.3.0

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
package/CHANGELOG.md CHANGED
@@ -1,5 +1,39 @@
1
1
  # Changelog
2
2
 
3
+ ## 2.3.0 — 2026-08-24
4
+
5
+ ### Multi-process hardening (several OpenCode windows = several plugin processes on one store)
6
+ - Remind→block session state is persisted ON THE GATE (`remindedSessions`/`failedSessions`) instead of per-process memory: the escalation chain now survives process restarts and is visible to every window serving the session (before: two windows or a restart reset it to "remind forever, never block"). Enforcement reads fresh gate state under the store lock.
7
+ - Session state rots after 24h and is capped per gate; `session.deleted` cleans it from disk.
8
+
9
+ ### Performance (hot path runs on every tool call)
10
+ - `fuzzySimilar`: O(1) length-band pre-filter (triangle inequality — zero false negatives) and a `FUZZY_MAX_LEN` cap — kills the Levenshtein explosion on long signatures (was 150-600ms/tool-call at scale).
11
+ - `GateStore`: O(1) key index + cached blocking subset for lookups; 1s TTL on the mtime cache so the hot path stops paying a `stat` per call (saves refresh the cache directly).
12
+ - Global log rotates at 2MB instead of 512KB — with several projects the aggregate forensics no longer vanish within a day.
13
+
14
+ ### Fixed
15
+ - Init-storm TOCTOU: index orphan-pruning now keeps a 24h grace window, so a just-promoted gate's index entry cannot be pruned by a concurrent startup.
16
+ - After-hook escalation state is written under the store lock and follows the gate to the global store on escalation (previously lost in both cases).
17
+
18
+ ### Added
19
+ - `doctor.ts` reports LOCK DEGRADATIONS (count of `degraded` log events) as an observability note — the evidence signal for whether the storage backend ever needs revisiting.
20
+
21
+ ## 2.2.1 — 2026-08-24
22
+
23
+ ### Fixed (adversarial-review round)
24
+ - Escalation order: the gate is written to the global store BEFORE being removed from the project store — a crash between the two writes leaves a duplicate (healed by migrate), never a hole.
25
+ - `dejavu:proceed` inside quoted strings no longer bypasses gates (`echo "dejavu:proceed" && gated-cmd` stays enforced); the marker is honored only outside quotes.
26
+ - Concurrent first-encounter race: calls dispatched in the same burst as a REMINDER (within 500ms) are reminded too instead of slipping through as a "retry".
27
+ - CRLF/CR commands normalize identically to LF; `splitChain` splits on CR — no more line-ending fragmentation.
28
+ - `normalizeCommand` is fully idempotent: quoted spans are parameterized BEFORE path rules (a `<str>` substitution inserts spaces that would expose an adjacent `/` to the path rule only on a second pass), fingerprint payloads are trimmed, and already-parameterized payloads are never re-fingerprinted.
29
+ - Interpreter flags glued to their payload (`node -e"code"`) fingerprint identically to the spaced form.
30
+ - Session state maps: inner key sets are capped — long sessions no longer grow unbounded.
31
+
32
+ ### Added
33
+ - Lock degradation (contention > 3s) emits a `degraded` log event — the only window where concurrent writes can lose updates is now visible.
34
+ - `test/property.ts` — property-based tests for the normalization pipeline (idempotency, no nested tokens, output bound, one-liner distinctness, marker neutrality, splitChain atomicity).
35
+ - `test/fuzz.ts` — seeded mutation fuzzer with a metamorphic oracle and case shrinking; both harnesses run in CI. The harnesses caught the idempotency, marker-neutrality, glued-flag and nested-token-detector bugs above before production did.
36
+
3
37
  ## 2.2.0 — 2026-08-23
4
38
 
5
39
  ### Added
package/README.md CHANGED
@@ -66,7 +66,8 @@ Restart OpenCode. Gates appear automatically as failures recur — nothing to co
66
66
  - **Intended non-zero exits** — exit 1 from diagnostics is NOT a failure (that is their normal "found nothing / found issues" outcome). Exit ≥ 2 always counts.
67
67
  - **Aborted ≠ failed** — cancelled/aborted tool executions ("Tool execution aborted") are infrastructure noise and are never counted as failures.
68
68
  - **File content is not command output** — text failure signatures are scanned for `bash` only; `read`/`edit`/`write` failures come exclusively from the event channel (a file containing "TypeError" is not a failure).
69
- - **Concurrency** — gates.json mutations run under an exclusive lockfile; log appends and rotation take their own lock (every OpenCode window shares the global log); writes are tmp+rename with EPERM/EACCES/EBUSY retry (Windows AV/indexer). NT long paths get the `\\?\` prefix.
69
+ - **Concurrency** — gates.json mutations run under an exclusive lockfile; log appends and rotation take their own lock (every OpenCode window shares the global log); writes are tmp+rename with EPERM/EACCES/EBUSY retry (Windows AV/indexer). NT long paths get the `\\?\` prefix. If a lock cannot be acquired within 3s the critical section degrades to unlocked (the tool pipeline must never hang) and emits a `degraded` log event — the only window where updates can be lost is visible.
70
+ - **Multi-window safe** — the remind→block escalation chain is persisted on the gate itself (`remindedSessions`/`failedSessions`), not in process memory: several OpenCode windows on one store — and process restarts — all see the same chain. Enforcement always reads fresh gate state under the store lock.
70
71
  - **Near-duplicate consolidation** — new failures merge into existing patterns via normalized Levenshtein ≤ 0.3 with an absolute floor of 3 edits (replaces token Jaccard, which collapsed all `<str>` placeholders; the floor stops `git push` vs `git pull`-style merges).
71
72
  - **Bounded memory** — per-session maps are capped (200 sessions) and freed on `session.deleted`; handled part IDs evict FIFO; TTL expiry re-runs every 6 h in long-lived processes.
72
73
  - **Migration** — gates outside the blocking policy are demoted to `watching` automatically; project copies of already-global gates are merged into the global gate (evidence is consolidated, never deleted).
package/index.ts CHANGED
@@ -23,8 +23,10 @@ const TTL_DAYS = 60
23
23
  const TTL_INTERVAL_MS = 6 * 60 * 60 * 1000
24
24
  /** a gate firing this often without killing the error gets flagged for review */
25
25
  const REVIEW_FIRES = 10
26
- /** per-session state maps are capped to bound memory in long-lived processes */
27
- const SESSION_MAP_CAP = 200
26
+ /** a "retry" arriving this soon after a reminder was dispatched concurrently with it
27
+ * (same tool-call burst) and never saw the reminder — it gets reminded as well.
28
+ * A true agent retry needs a full model turn (≥1s in practice), so 500ms separates both. */
29
+ const REMINDER_RACE_WINDOW_MS = 500
28
30
  /** handled part IDs are capped FIFO-style */
29
31
  const HANDLED_CAP = 5000
30
32
  const HANDLED_KEEP = 2500
@@ -34,24 +36,6 @@ const PENDING_CAP = 1000
34
36
  /** Sentinel: intentional gate/reminder throws (rethrown); our own bugs are swallowed. */
35
37
  class GateSignal extends Error {}
36
38
 
37
- function addToSetMap(map: Map<string, Set<string>>, outer: string, inner: string): void {
38
- let set = map.get(outer)
39
- if (!set) {
40
- set = new Set()
41
- map.set(outer, set)
42
- }
43
- set.add(inner)
44
- }
45
-
46
- /** Drop oldest entries (Map preserves insertion order) to bound memory. */
47
- function capMap(map: Map<string, Set<string>>, cap: number): void {
48
- while (map.size > cap) {
49
- const oldest = map.keys().next()
50
- if (oldest.done) break
51
- map.delete(oldest.value)
52
- }
53
- }
54
-
55
39
  function scrubbedArgs(args: Record<string, unknown>): Record<string, unknown> {
56
40
  if (typeof args.command === "string") return { ...args, command: scrubSecrets(args.command) }
57
41
  if (typeof args.pattern === "string") return { ...args, pattern: scrubSecrets(args.pattern) }
@@ -89,10 +73,6 @@ export const Dejavu: Plugin = async ({ directory, client }) => {
89
73
  : null
90
74
  const stores = new Stores(globalStore, projectStore)
91
75
 
92
- /** sessions in which a gate key was already reminded about */
93
- const reminded = new Map<string, Set<string>>()
94
- /** sessions in which a reminded pattern failed again — next attempt is blocked */
95
- const failedAfterReminder = new Map<string, Set<string>>()
96
76
  /** callID -> signature fallback when the after-hook does not receive args */
97
77
  const pendingCalls = new Map<string, string>()
98
78
  /** message part IDs already counted as tool-level errors */
@@ -169,7 +149,10 @@ export const Dejavu: Plugin = async ({ directory, client }) => {
169
149
  const session = typeof input.sessionID === "string" ? input.sessionID : "unknown"
170
150
 
171
151
  // Explicit escape hatch — checked only in the actionable text field,
172
- // with word boundaries, so unrelated args cannot bypass gates.
152
+ // with word boundaries, so unrelated args cannot bypass gates. Quoted
153
+ // spans are stripped first: `echo "dejavu:proceed" && gated-cmd` must
154
+ // NOT bypass the gate on the chained command — the marker is a
155
+ // comment-style annotation, not data.
173
156
  const commandText =
174
157
  typeof rawArgs.command === "string"
175
158
  ? rawArgs.command
@@ -178,33 +161,50 @@ export const Dejavu: Plugin = async ({ directory, client }) => {
178
161
  : typeof rawArgs.filePath === "string"
179
162
  ? rawArgs.filePath
180
163
  : ""
181
- if (/\bdejavu:proceed\b/.test(commandText)) {
164
+ if (/\bdejavu:proceed\b/.test(commandText.replace(/"[^"]*"|'[^']*'/g, " "))) {
182
165
  await stores.logAll({ type: "override", key: gate.key, tool: gate.tool, session, project: directory })
183
166
  return
184
167
  }
185
168
 
186
- // Repeat offense: reminded in this session, retried, failed again -> hard block.
187
- if (failedAfterReminder.get(session)?.has(gate.key)) {
188
- gate.blockedCount += 1
189
- if (gate.blockedCount >= REVIEW_FIRES) gate.review = true
190
- await found.store.save()
191
- await stores.logAll({ type: "blocked", key: gate.key, tool: gate.tool, session, project: directory, via })
192
- throw new GateSignal(blockMessage(gate, found.store.dir))
193
- }
169
+ // Enforce from FRESH gate state under the store lock. The remind→block
170
+ // chain lives on the gate itself (remindedSessions/failedSessions), so
171
+ // it survives process restarts and is visible to every window serving
172
+ // this session per-process maps lost it on both.
173
+ const target = found
174
+ let signal: GateSignal | null = null
175
+ await target.store.runLocked(async () => {
176
+ const fresh = (await target.store.load(true)).find((g) => g.key === gate.key)
177
+ if (fresh === undefined) return // gate deleted between find and lock
178
+
179
+ // Repeat offense: reminded, retried, failed again -> hard block.
180
+ if (fresh.failedSessions !== undefined && fresh.failedSessions.includes(session)) {
181
+ fresh.blockedCount += 1
182
+ if (fresh.blockedCount >= REVIEW_FIRES) fresh.review = true
183
+ await target.store.save()
184
+ await stores.logAll({ type: "blocked", key: fresh.key, tool: fresh.tool, session, project: directory, via })
185
+ signal = new GateSignal(blockMessage(fresh, target.store.dir))
186
+ return
187
+ }
194
188
 
195
- // First encounter this session -> remind (the call is aborted; agent may retry corrected).
196
- if (!reminded.get(session)?.has(gate.key)) {
197
- addToSetMap(reminded, session, gate.key)
198
- addToSetMap(reminded, session, patternKey(signature)) // exact key too: retry may fuzzy-match differently
199
- capMap(reminded, SESSION_MAP_CAP)
200
- gate.remindedCount += 1
201
- await found.store.save()
202
- await stores.logAll({ type: "reminded", key: gate.key, tool: gate.tool, session, project: directory, via })
203
- throw new GateSignal(remindMessage(gate))
204
- }
189
+ // First encounter this session -> remind (the call is aborted; agent may retry corrected).
190
+ // Race guard: calls dispatched in the same burst all arrive before the agent can
191
+ // have seen any reminder, so a "retry" within REMINDER_RACE_WINDOW_MS of the
192
+ // remind is itself a concurrent first encounter and gets reminded too.
193
+ const remindedAt = fresh.remindedSessions?.[session]
194
+ if (remindedAt === undefined || Date.now() - remindedAt < REMINDER_RACE_WINDOW_MS) {
195
+ if (fresh.remindedSessions === undefined) fresh.remindedSessions = {}
196
+ fresh.remindedSessions[session] = Date.now()
197
+ fresh.remindedCount += 1
198
+ await target.store.save()
199
+ await stores.logAll({ type: "reminded", key: fresh.key, tool: fresh.tool, session, project: directory, via })
200
+ signal = new GateSignal(remindMessage(fresh))
201
+ return
202
+ }
205
203
 
206
- // Already reminded, no repeated failure yet -> allow one retry.
207
- await stores.logAll({ type: "retry-allowed", key: gate.key, tool: gate.tool, session, project: directory, via })
204
+ // Already reminded, no repeated failure yet -> allow one retry.
205
+ await stores.logAll({ type: "retry-allowed", key: fresh.key, tool: fresh.tool, session, project: directory, via })
206
+ })
207
+ if (signal !== null) throw signal
208
208
  } catch (error) {
209
209
  if (error instanceof GateSignal) throw error
210
210
  // Our own bugs must never break the user's tool calls.
@@ -288,21 +288,29 @@ export const Dejavu: Plugin = async ({ directory, client }) => {
288
288
  await logClient("info", `dejavu: gate went global — "${result.gate.signature}"`)
289
289
  }
290
290
 
291
- // Metric: failure of an already-enforced pattern (the event that
292
- // promoted the gate does not count the gate did not exist yet).
293
- if (result.gate.status === "blocking" && !result.promoted) {
294
- result.gate.recurredAfterGate += 1
295
- await result.store.save()
296
- await stores.logAll({ type: "recurred-after-gate", key, tool: input.tool, session, project: directory })
297
- }
298
-
299
- // Same-session repeat after a reminder -> escalate to hard block.
300
- if (reminded.get(session)?.has(key) || reminded.get(session)?.has(result.gate.key)) {
301
- addToSetMap(failedAfterReminder, session, result.gate.key)
302
- capMap(failedAfterReminder, SESSION_MAP_CAP)
303
- result.gate.recurredAfterReminder += 1
304
- await result.store.save()
305
- }
291
+ // Persist escalation state on the gate itself (under the store lock) so
292
+ // every window serving this session sees the same remind→block chain.
293
+ const ownerStore = result.wentGlobal ? stores.globalStore : result.store
294
+ await ownerStore.runLocked(async () => {
295
+ const fresh = (await ownerStore.load(true)).find((g) => g.key === result.gate.key)
296
+ if (fresh === undefined) return
297
+ let changed = false
298
+ // Metric: failure of an already-enforced pattern (the event that
299
+ // promoted the gate does not count the gate did not exist yet).
300
+ if (fresh.status === "blocking" && !result.promoted) {
301
+ fresh.recurredAfterGate += 1
302
+ changed = true
303
+ await stores.logAll({ type: "recurred-after-gate", key: fresh.key, tool: input.tool, session, project: directory })
304
+ }
305
+ // Same-session repeat after a reminder -> escalate to hard block.
306
+ if (fresh.remindedSessions?.[session] !== undefined) {
307
+ if (fresh.failedSessions === undefined) fresh.failedSessions = []
308
+ if (!fresh.failedSessions.includes(session)) fresh.failedSessions.push(session)
309
+ fresh.recurredAfterReminder += 1
310
+ changed = true
311
+ }
312
+ if (changed) await ownerStore.save()
313
+ })
306
314
  } catch {
307
315
  // detection failures must never break the tool pipeline
308
316
  }
@@ -312,12 +320,11 @@ export const Dejavu: Plugin = async ({ directory, client }) => {
312
320
  try {
313
321
  const type = (event as { type?: unknown }).type
314
322
 
315
- // Free per-session state when a session is deleted.
323
+ // Free the persisted per-session state when a session is deleted.
316
324
  if (type === "session.deleted") {
317
325
  const props = (event as { properties?: unknown }).properties as { sessionID?: unknown } | undefined
318
326
  if (typeof props?.sessionID === "string") {
319
- reminded.delete(props.sessionID)
320
- failedAfterReminder.delete(props.sessionID)
327
+ stores.forgetSession(props.sessionID).catch(() => {})
321
328
  }
322
329
  return
323
330
  }
package/package.json CHANGED
@@ -1,6 +1,6 @@
1
1
  {
2
2
  "name": "opencode-dejavu",
3
- "version": "2.2.0",
3
+ "version": "2.3.0",
4
4
  "description": "Cross-session memory prosthesis for OpenCode: detects recurring tool-call failures and promotes them into enforced gates. Remind first, block on same-session repeat offense.",
5
5
  "type": "module",
6
6
  "main": "index.ts",
package/src/AGENTS.md CHANGED
@@ -24,13 +24,16 @@ Two dependency-free modules: `patterns.ts` (pure functions — call identity, no
24
24
  ## INVARIANTS (do not break)
25
25
 
26
26
  - Rule order in `PARAM_RULES` matters: quoted strings first, specific tokens (uuid/sha/ip/url/date), generic numbers last — reordering fragments signatures
27
+ - Rule order in `normalizeCommand` matters too: quoted strings are parameterized BEFORE path rules — a `<str>` substitution inserts spaces that would expose an adjacent `/` to the path rule on a second pass (idempotency); interpreter payload hashing runs while the payload is still raw
27
28
  - `scrubSecrets()` runs on every string before it touches disk; `recordFailure` re-scrubs defensively
28
29
  - `canBlock(tool, sig)` = bash && non-diagnostic && not a bare one-liner shape — the ONLY path to `blocking`; probe tools use `PROMOTE_COUNT_PROBE` and never block
29
30
  - `DIAGNOSTIC_VERBS` serves two callers (exit-1 allowlist + blocking policy) — one list, two uses; edit knowing both move
30
31
  - Lock order is always project → global, gates → index (see `recordFailure` escalation) — reversing deadlocks; the log lock is separate and leaf-level
31
32
  - Cross-project evidence lives ONLY in the global `index.json` — a gate's own `projects` array sees one store and never drives escalation alone
33
+ - Escalation writes the global gate FIRST, then removes the project copy — a crash between the two writes must leave a duplicate (healed by migrate), never a hole
32
34
  - Inside `runLocked` always `load(true)`; unlocked `load()` peeks are routing hints only, never a basis for mutation
33
- - `GateStore.load` caches by mtime after external edits the cache refreshes on next stat; `save()` refreshes it manually
35
+ - Hot-path reads use the 1s TTL cache + key index (`byKey`/`blockingOnly`); mutations inside locks use `load(true)`; `save()` refreshes the cache directly
36
+ - The remind→block chain is persisted ON THE GATE (`remindedSessions`/`failedSessions`) and enforced under the store lock — process memory holds nothing authoritative, so several windows and restarts share one escalation
34
37
  - Log appends and rotation take the log lock — every OpenCode window shares the global log; unlocked appends interleave into broken JSON
35
38
  - Every gate read from disk crosses `coerceGateShape` + `repairGate` in `load()` — enforcement never sees raw state; hopeless records are dropped, repairable ones coerced
36
39
  - Quarantine preserves bytes: unparseable files are renamed to `*.corrupt-*`, never deleted; every repair emits a `repaired`/`quarantined` log event
package/src/patterns.ts CHANGED
@@ -51,16 +51,24 @@ export function scrubSecrets(text: string): string {
51
51
  * Secrets are scrubbed before hashing so they neither persist nor fragment.
52
52
  */
53
53
  const INTERPRETER_ONELINER =
54
- /(?:^|[|;&(\n]\s*)(?:\S+[\\/])?(python3?|node|bun|deno|perl|ruby|pwsh|powershell)(?:\.exe)?(?:\s+-\w+)*\s+(-c|-e|--eval|-command)\s+/i
54
+ /(?:^|[|;&(\n]\s*)(?:\S+[\\/])?(python3?|node|bun|deno|perl|ruby|pwsh|powershell)(?:\.exe)?(?:\s+-\w+)*\s+(-c|-e|--eval|-command)\s*/i
55
55
 
56
56
  function hashInterpreterPayload(command: string): string {
57
57
  const match = INTERPRETER_ONELINER.exec(command)
58
58
  if (!match) return command
59
59
  const payload = command.slice(match.index + match[0].length)
60
60
  if (payload.trim() === "") return command
61
+ // Already fingerprinted (re-normalization) — keep the existing token so
62
+ // normalizeCommand stays idempotent.
63
+ if (/^<code:[0-9a-f]+>$/.test(payload.trim())) return command
64
+ // Already-parameterized placeholders are data, not code — never hash them
65
+ // (idempotency: a second pass must not fingerprint a <str>).
66
+ if (/^(?:<(?:str|path|n|hash|uuid|sha|md5|ip|url|email|date)>\s*)+$/.test(payload.trim())) return command
61
67
  // For whole (unchained) commands the payload runs to end of string; chain
62
68
  // segments are normalized separately, so segment keys stay exact.
63
- const fingerprint = createHash("sha1").update(scrubSecrets(payload)).digest("hex").slice(0, 8)
69
+ // Trim before hashing: trailing whitespace (e.g. a stripped override marker)
70
+ // is not part of the code's identity.
71
+ const fingerprint = createHash("sha1").update(scrubSecrets(payload.trim())).digest("hex").slice(0, 8)
64
72
  return `${command.slice(0, match.index + match[0].length)}<code:${fingerprint}>`
65
73
  }
66
74
 
@@ -70,11 +78,18 @@ function hashInterpreterPayload(command: string): string {
70
78
  * away so that "same failure, different instance" collapses into one pattern.
71
79
  */
72
80
  export function normalizeCommand(command: string): string {
73
- let s = command.replace(COMMENT_LINE, "$1").toLowerCase()
81
+ // CRLF/CR commands (Windows pastes, agent multi-line) normalize to LF —
82
+ // otherwise the same command fragments across line-ending styles.
83
+ let s = command.replace(/\r\n?/g, "\n")
84
+ s = s.replace(COMMENT_LINE, "$1").toLowerCase()
74
85
  s = hashInterpreterPayload(s)
86
+ // Quoted spans come out FIRST: they are data, and removing them before the
87
+ // path rules keeps normalization idempotent — a <str> replacement inserts
88
+ // spaces that would otherwise expose an adjacent "/" to the path rule only
89
+ // on a second pass.
90
+ s = s.replace(/"[^"]*"|'[^']*'/g, " <str> ")
75
91
  s = s.replace(/[a-z]:[\\/][^\s"']+/gi, " <path> ")
76
92
  s = s.replace(/(^|\s)\/[^\s"']+/g, "$1<path> ")
77
- s = s.replace(/"[^"]*"|'[^']*'/g, " <str> ")
78
93
  // lookbehind: never re-parameterize the <code:...> fingerprint hex
79
94
  s = s.replace(/(?<!<code:)\b[0-9a-f]{7,64}\b/gi, " <hash> ")
80
95
  s = s.replace(/(?<!<code:)\b\d[\d.]*\b/g, " <n> ")
@@ -213,7 +228,7 @@ export function splitChain(command: string): string[] {
213
228
  continue
214
229
  }
215
230
  if (depth === 0) {
216
- if (ch === ";" || ch === "\n") {
231
+ if (ch === ";" || ch === "\n" || ch === "\r") {
217
232
  flush()
218
233
  i += 1
219
234
  continue
@@ -317,6 +332,11 @@ export function levenshtein(a: string, b: string): number {
317
332
  /** Code fingerprints are IDENTITY, not data — they must match exactly. */
318
333
  const CODE_FINGERPRINTS = /<code:[0-9a-f]+>/g
319
334
 
335
+ /** Signatures longer than this match exactly only: a 300-char normalized
336
+ * command is already specific enough that "30% near" is meaningless, and
337
+ * Levenshtein on long signatures is the hot-path cost cliff. */
338
+ export const FUZZY_MAX_LEN = 300
339
+
320
340
  /**
321
341
  * Near-duplicate match: normalized edit distance <= 30% AND absolute distance
322
342
  * >= 3. Unlike token-set Jaccard, this does not collapse commands that merely
@@ -335,6 +355,11 @@ export function fuzzySimilar(a: string, b: string): boolean {
335
355
  }
336
356
  const maxLen = Math.max(a.length, b.length)
337
357
  if (maxLen === 0) return true
358
+ if (maxLen > FUZZY_MAX_LEN) return false
359
+ // Triangle inequality: distance >= |lenA - lenB|. If even that floor
360
+ // exceeds the ratio threshold, no Levenshtein result can pass — an O(1)
361
+ // pre-filter with zero false negatives that skips most DP computations.
362
+ if (Math.abs(a.length - b.length) / maxLen > 0.3) return false
338
363
  const distance = levenshtein(a, b)
339
364
  return distance >= 3 && distance / maxLen <= 0.3
340
365
  }
package/src/store.ts CHANGED
@@ -1,10 +1,10 @@
1
1
  import { appendFile, mkdir, readFile, rename, stat, unlink, writeFile } from "node:fs/promises"
2
2
  import { dirname, join } from "node:path"
3
- import { canBlock, fuzzySimilar, scrubSecrets } from "./patterns"
3
+ import { canBlock, fuzzySimilar, FUZZY_MAX_LEN, scrubSecrets } from "./patterns"
4
4
  import { coerceGateShape, repairGate } from "./validate"
5
5
 
6
6
  /** Bumped on behavior changes; stamped into init log events so stale sessions are visible. */
7
- export const PLUGIN_VERSION = "2.2.0"
7
+ export const PLUGIN_VERSION = "2.3.0"
8
8
 
9
9
  export interface Gate {
10
10
  /** sha1 signature prefix — the pattern identity */
@@ -32,6 +32,12 @@ export interface Gate {
32
32
  recurredAfterGate: number
33
33
  /** flagged for manual review when the gate fires often but errors stopped */
34
34
  review?: boolean
35
+ /** sessions currently reminded about this gate: sessionID -> remind time (ms).
36
+ * Persisted on the gate so the remind→block chain survives process restarts
37
+ * and is visible to every window serving the session. */
38
+ remindedSessions?: Record<string, number>
39
+ /** sessions that failed again after a reminder — their next attempt blocks */
40
+ failedSessions?: string[]
35
41
  }
36
42
 
37
43
  interface GatesFile {
@@ -62,6 +68,7 @@ export type LogEventType =
62
68
  | "init"
63
69
  | "repaired"
64
70
  | "quarantined"
71
+ | "degraded"
65
72
 
66
73
  export interface LogEvent {
67
74
  type: LogEventType
@@ -83,8 +90,12 @@ export interface LogEvent {
83
90
  const MAX_SESSIONS = 50
84
91
  const MAX_PROJECTS = 20
85
92
  const LOG_ROTATE_BYTES = 512 * 1024
93
+ /** the global log aggregates every project — rotate it later or forensics vanish in a day */
94
+ const GLOBAL_LOG_ROTATE_BYTES = 2048 * 1024
86
95
  const LOG_ROTATE_KEEP_LINES = 1000
87
96
  const DAY_MS = 24 * 60 * 60 * 1000
97
+ /** trust the loaded-gates cache this long without re-statting (hot path: every tool call) */
98
+ const LOAD_CACHE_TTL_MS = 1000
88
99
 
89
100
  /** failures required before a pattern becomes an enforced gate */
90
101
  export const PROMOTE_COUNT = 3
@@ -137,7 +148,7 @@ const LOCK_WAIT_MS = 3000
137
148
  * degradation: if the lock cannot be acquired within LOCK_WAIT_MS the
138
149
  * critical section runs unlocked rather than hanging the tool pipeline.
139
150
  */
140
- async function withLock<T>(lockTarget: string, fn: () => Promise<T>): Promise<T> {
151
+ async function withLock<T>(lockTarget: string, fn: () => Promise<T>, onDegrade?: () => void): Promise<T> {
141
152
  const lock = `${lockTarget}.lock`
142
153
  await mkdir(ntPath(dirname(lock)), { recursive: true })
143
154
  const started = Date.now()
@@ -157,7 +168,11 @@ async function withLock<T>(lockTarget: string, fn: () => Promise<T>): Promise<T>
157
168
  } catch {
158
169
  continue // lock vanished between attempts
159
170
  }
160
- if (Date.now() - started > LOCK_WAIT_MS) break
171
+ if (Date.now() - started > LOCK_WAIT_MS) {
172
+ // The only window where concurrent writes can lose updates — make it visible.
173
+ if (onDegrade) onDegrade()
174
+ break
175
+ }
161
176
  await new Promise((resolve) => setTimeout(resolve, 50))
162
177
  }
163
178
  }
@@ -175,6 +190,10 @@ async function withLock<T>(lockTarget: string, fn: () => Promise<T>): Promise<T>
175
190
  export class GateStore {
176
191
  private gates: Gate[] | null = null
177
192
  private mtimeMs = 0
193
+ /** hot-path caches: valid until LOAD_CACHE_TTL_MS / invalidated on mutation */
194
+ private cacheUntilMs = 0
195
+ private keyIndex: Map<string, Gate> | null = null
196
+ private blockingCache: Gate[] | null = null
178
197
  private index: IndexFile | null = null
179
198
  private indexMtimeMs = 0
180
199
 
@@ -194,7 +213,9 @@ export class GateStore {
194
213
 
195
214
  /** Run a load→mutate→save section under the store's exclusive lock. */
196
215
  async runLocked<T>(fn: () => Promise<T>): Promise<T> {
197
- return withLock(this.gatesPath, fn)
216
+ return withLock(this.gatesPath, fn, () => {
217
+ this.log({ type: "degraded", key: "gates.lock", snippet: `lock contention exceeded ${LOCK_WAIT_MS}ms; critical section ran unlocked` }).catch(() => {})
218
+ })
198
219
  }
199
220
 
200
221
  /**
@@ -203,9 +224,16 @@ export class GateStore {
203
224
  * dropped, repairable ones coerced — enforcement never sees raw state.
204
225
  */
205
226
  async load(force = false): Promise<Gate[]> {
227
+ // TTL fast path: the hot path (every tool call) must not pay a stat per
228
+ // call. Gates change rarely (promotion, manual edit); 1s staleness is
229
+ // invisible to enforcement and our own saves refresh the cache directly.
230
+ if (!force && this.gates !== null && Date.now() < this.cacheUntilMs) {
231
+ return this.gates
232
+ }
206
233
  try {
207
234
  const info = await stat(ntPath(this.gatesPath))
208
235
  if (!force && this.gates !== null && info.mtimeMs === this.mtimeMs) {
236
+ this.cacheUntilMs = Date.now() + LOAD_CACHE_TTL_MS
209
237
  return this.gates
210
238
  }
211
239
  const raw = await readFile(ntPath(this.gatesPath), "utf8")
@@ -219,15 +247,39 @@ export class GateStore {
219
247
  gates.push(gate)
220
248
  }
221
249
  this.gates = gates
250
+ this.keyIndex = new Map(gates.map((g) => [g.key, g]))
251
+ this.blockingCache = gates.filter((g) => g.status === "blocking")
222
252
  this.mtimeMs = info.mtimeMs
253
+ this.cacheUntilMs = Date.now() + LOAD_CACHE_TTL_MS
223
254
  return this.gates
224
255
  } catch {
225
256
  // missing or unreadable gates.json — treat as an empty store
226
- if (this.gates === null) this.gates = []
257
+ if (this.gates === null) {
258
+ this.gates = []
259
+ this.keyIndex = new Map()
260
+ this.blockingCache = []
261
+ }
262
+ this.cacheUntilMs = Date.now() + LOAD_CACHE_TTL_MS
227
263
  return this.gates
228
264
  }
229
265
  }
230
266
 
267
+ /** O(1) exact lookup over the cached gates (call load() first to refresh). */
268
+ byKey(key: string): Gate | undefined {
269
+ if (this.keyIndex === null) {
270
+ this.keyIndex = new Map((this.gates ?? []).map((g) => [g.key, g]))
271
+ }
272
+ return this.keyIndex.get(key)
273
+ }
274
+
275
+ /** Cached blocking subset — the fuzzy scan iterates this, not all gates. */
276
+ blockingOnly(): Gate[] {
277
+ if (this.blockingCache === null) {
278
+ this.blockingCache = (this.gates ?? []).filter((g) => g.status === "blocking")
279
+ }
280
+ return this.blockingCache
281
+ }
282
+
231
283
  async save(): Promise<void> {
232
284
  if (this.gates === null) return
233
285
  await mkdir(ntPath(this.dir), { recursive: true })
@@ -238,6 +290,9 @@ export class GateStore {
238
290
  } catch {
239
291
  // mtime refresh is best-effort
240
292
  }
293
+ // We know the content we just wrote — refresh the TTL cache directly.
294
+ // (keyIndex/blockingCache hold references into this.gates, still valid.)
295
+ this.cacheUntilMs = Date.now() + LOAD_CACHE_TTL_MS
241
296
  }
242
297
 
243
298
  /** Cross-project pattern index; meaningful only on the global store. */
@@ -273,7 +328,9 @@ export class GateStore {
273
328
 
274
329
  /** Run an index load→mutate→save section under the index's own lock. */
275
330
  async runLockedIndex<T>(fn: () => Promise<T>): Promise<T> {
276
- return withLock(this.indexPath, fn)
331
+ return withLock(this.indexPath, fn, () => {
332
+ this.log({ type: "degraded", key: "index.lock", snippet: `lock contention exceeded ${LOCK_WAIT_MS}ms; critical section ran unlocked` }).catch(() => {})
333
+ })
277
334
  }
278
335
 
279
336
  /**
@@ -303,15 +360,19 @@ export class GateStore {
303
360
  extract(keys: Set<string>): Gate[] {
304
361
  if (this.gates === null) return []
305
362
  const removed = this.gates.filter((g) => keys.has(g.key))
306
- if (removed.length > 0) this.gates = this.gates.filter((g) => !keys.has(g.key))
363
+ if (removed.length > 0) {
364
+ this.gates = this.gates.filter((g) => !keys.has(g.key))
365
+ this.keyIndex = null
366
+ this.blockingCache = null
367
+ }
307
368
  return removed
308
369
  }
309
370
 
310
- async rotateLog(): Promise<void> {
371
+ async rotateLog(rotateBytes: number = LOG_ROTATE_BYTES): Promise<void> {
311
372
  await withLock(this.logPath, async () => {
312
373
  try {
313
374
  const info = await stat(ntPath(this.logPath))
314
- if (info.size < LOG_ROTATE_BYTES) return
375
+ if (info.size < rotateBytes) return
315
376
  const raw = await readFile(ntPath(this.logPath), "utf8")
316
377
  const lines = raw.split("\n").filter((l) => l.trim() !== "")
317
378
  const kept = lines.slice(-LOG_ROTATE_KEEP_LINES)
@@ -462,7 +523,8 @@ export class Stores {
462
523
  /** True if a pattern with this key exists in any scope (chain attribution). */
463
524
  async hasKey(key: string): Promise<boolean> {
464
525
  for (const store of this.scopes()) {
465
- if ((await store.load()).some((g) => g.key === key)) return true
526
+ await store.load()
527
+ if (store.byKey(key) !== undefined) return true
466
528
  }
467
529
  return false
468
530
  }
@@ -473,13 +535,15 @@ export class Stores {
473
535
  signature: string,
474
536
  ): Promise<{ gate: Gate; store: GateStore; via: "exact" | "fuzzy" } | null> {
475
537
  for (const store of this.scopes()) {
476
- const exact = (await store.load()).find((g) => g.key === key)
538
+ await store.load()
539
+ const exact = store.byKey(key)
477
540
  if (exact) return { gate: exact, store, via: "exact" }
478
541
  }
542
+ // Over-long signatures match exactly only — see FUZZY_MAX_LEN.
543
+ if (signature.length > FUZZY_MAX_LEN) return null
479
544
  let best: { gate: Gate; store: GateStore; score: number } | null = null
480
545
  for (const store of this.scopes()) {
481
- for (const gate of await store.load()) {
482
- if (gate.status !== "blocking") continue
546
+ for (const gate of store.blockingOnly()) {
483
547
  if (!fuzzySimilar(signature, gate.signature)) continue
484
548
  const score = Math.abs(signature.length - gate.signature.length)
485
549
  if (best === null || score < best.score) best = { gate, store, score }
@@ -492,9 +556,8 @@ export class Stores {
492
556
  async blockingGates(): Promise<Gate[]> {
493
557
  const result: Gate[] = []
494
558
  for (const store of this.scopes()) {
495
- for (const gate of await store.load()) {
496
- if (gate.status === "blocking") result.push(gate)
497
- }
559
+ await store.load()
560
+ for (const gate of store.blockingOnly()) result.push(gate)
498
561
  }
499
562
  return result.sort((a, b) => b.count - a.count)
500
563
  }
@@ -532,7 +595,31 @@ export class Stores {
532
595
 
533
596
  async rotateLogs(): Promise<void> {
534
597
  for (const store of this.scopes()) {
535
- await store.rotateLog()
598
+ // The global log aggregates every project — give it more room.
599
+ await store.rotateLog(store === this.globalStore ? GLOBAL_LOG_ROTATE_BYTES : LOG_ROTATE_BYTES)
600
+ }
601
+ }
602
+
603
+ /** Forget per-session enforcement state when a session dies. */
604
+ async forgetSession(sessionID: string): Promise<void> {
605
+ for (const store of this.scopes()) {
606
+ await store.runLocked(async () => {
607
+ const gates = await store.load(true)
608
+ let changed = false
609
+ for (const gate of gates) {
610
+ if (gate.remindedSessions && gate.remindedSessions[sessionID] !== undefined) {
611
+ delete gate.remindedSessions[sessionID]
612
+ if (Object.keys(gate.remindedSessions).length === 0) delete gate.remindedSessions
613
+ changed = true
614
+ }
615
+ if (gate.failedSessions !== undefined && gate.failedSessions.includes(sessionID)) {
616
+ gate.failedSessions = gate.failedSessions.filter((s) => s !== sessionID)
617
+ if (gate.failedSessions.length === 0) delete gate.failedSessions
618
+ changed = true
619
+ }
620
+ }
621
+ if (changed) await store.save()
622
+ })
536
623
  }
537
624
  }
538
625
 
@@ -658,7 +745,11 @@ export class Stores {
658
745
  const index = await this.globalStore.loadIndex(true)
659
746
  let pruned = 0
660
747
  for (const key of Object.keys(index.keys)) {
661
- if (!knownKeys.has(key)) {
748
+ const entry = index.keys[key]
749
+ // Young entries may belong to a promotion in flight in another window
750
+ // (its gates were snapshotted after this key was written) — only prune
751
+ // orphans that have been stale for a day.
752
+ if (entry && !knownKeys.has(key) && Date.now() - Date.parse(entry.lastSeen) > DAY_MS) {
662
753
  delete index.keys[key]
663
754
  pruned += 1
664
755
  }
@@ -776,9 +867,8 @@ export class Stores {
776
867
 
777
868
  let wentGlobal = false
778
869
  if (store !== this.globalStore && this.projectStore && indexProjects >= input.globalProjects) {
779
- const idx = gates.findIndex((g) => g.key === moved.key)
780
- if (idx >= 0) gates.splice(idx, 1)
781
- await store.save()
870
+ // Global FIRST, then remove the local copy: a crash between the two
871
+ // writes must leave a duplicate (healed by migrate), never a hole.
782
872
  await this.globalStore.runLocked(async () => {
783
873
  const globalGates = await this.globalStore.load(true)
784
874
  const existing = globalGates.find((g) => g.key === moved.key)
@@ -789,6 +879,9 @@ export class Stores {
789
879
  }
790
880
  await this.globalStore.save()
791
881
  })
882
+ const idx = gates.findIndex((g) => g.key === moved.key)
883
+ if (idx >= 0) gates.splice(idx, 1)
884
+ await store.save()
792
885
  wentGlobal = true
793
886
  }
794
887
 
package/src/validate.ts CHANGED
@@ -11,6 +11,10 @@ import { canBlock, scrubSecrets } from "./patterns"
11
11
  const KEY_SHAPE = /^[0-9a-f]{12}$/
12
12
  /** detection truncates snippets at 200 chars on ingest */
13
13
  const SNIPPET_MAX = 200
14
+ /** per-session enforcement state rots after a day — sessions do not live longer */
15
+ const SESSION_STATE_TTL_MS = 24 * 60 * 60 * 1000
16
+ /** bound per-gate session state so long-lived gates cannot bloat */
17
+ const SESSION_STATE_CAP = 50
14
18
 
15
19
  /**
16
20
  * Structural parse of one persisted gate object. Returns a well-shaped Gate
@@ -50,6 +54,17 @@ export function coerceGateShape(raw: unknown): Gate | null {
50
54
  }
51
55
  if (typeof r.correction === "string") gate.correction = r.correction
52
56
  if (r.review === true) gate.review = true
57
+ if (r.remindedSessions !== null && typeof r.remindedSessions === "object" && !Array.isArray(r.remindedSessions)) {
58
+ const sessions: Record<string, number> = {}
59
+ for (const [session, at] of Object.entries(r.remindedSessions as Record<string, unknown>)) {
60
+ if (typeof at === "number" && Number.isFinite(at)) sessions[session] = at
61
+ }
62
+ if (Object.keys(sessions).length > 0) gate.remindedSessions = sessions
63
+ }
64
+ if (Array.isArray(r.failedSessions)) {
65
+ const sessions = r.failedSessions.filter((x): x is string => typeof x === "string")
66
+ if (sessions.length > 0) gate.failedSessions = sessions
67
+ }
53
68
  return gate
54
69
  }
55
70
 
@@ -87,6 +102,33 @@ export function repairGate(gate: Gate): boolean {
87
102
  changed = true
88
103
  }
89
104
  }
105
+ // Per-session enforcement state hygiene: rot stale entries, bound the rest.
106
+ if (gate.remindedSessions !== undefined) {
107
+ const now = Date.now()
108
+ const reminded = gate.remindedSessions
109
+ for (const session of Object.keys(reminded)) {
110
+ if (now - (reminded[session] ?? 0) > SESSION_STATE_TTL_MS) {
111
+ delete reminded[session]
112
+ changed = true
113
+ }
114
+ }
115
+ const sessions = Object.keys(reminded)
116
+ if (sessions.length > SESSION_STATE_CAP) {
117
+ sessions.sort((a, b) => (reminded[a] ?? 0) - (reminded[b] ?? 0))
118
+ for (const session of sessions.slice(0, sessions.length - SESSION_STATE_CAP)) {
119
+ delete reminded[session]
120
+ }
121
+ changed = true
122
+ }
123
+ if (Object.keys(reminded).length === 0) delete gate.remindedSessions
124
+ }
125
+ if (gate.failedSessions !== undefined) {
126
+ if (gate.failedSessions.length > SESSION_STATE_CAP) {
127
+ gate.failedSessions = gate.failedSessions.slice(-SESSION_STATE_CAP)
128
+ changed = true
129
+ }
130
+ if (gate.failedSessions.length === 0) delete gate.failedSessions
131
+ }
90
132
  // Policy is the single source of truth: a blocking gate that cannot block
91
133
  // is a leftover from an older policy and must be demoted.
92
134
  if (gate.status === "blocking" && !canBlock(gate.tool, gate.signature)) {
@@ -97,11 +139,11 @@ export function repairGate(gate: Gate): boolean {
97
139
  }
98
140
 
99
141
  /**
100
- * Corruption fingerprint of tokens re-parameterized inside other tokens
101
- * (e.g. `<code: <n> >` — a fingerprint eaten by the number rule). Such a
102
- * signature is stable under re-normalization, so only an explicit shape
103
- * check catches it.
142
+ * Corruption fingerprint of a placeholder re-parameterized inside another
143
+ * token (`<code: <n> >` — a fingerprint eaten by the number rule). Only the
144
+ * `<code:` token carries nested content, so the check is scoped to it —
145
+ * shell text like heredoc `<<eof:` must NOT trip the detector.
104
146
  */
105
147
  export function hasNestedTokens(signature: string): boolean {
106
- return /<[a-z]+:\s*</.test(signature)
148
+ return /<code:\s*</.test(signature)
107
149
  }