typeclaw 0.39.6 → 0.40.0

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.
Files changed (72) hide show
  1. package/package.json +1 -1
  2. package/scripts/dump-system-prompt.ts +4 -50
  3. package/src/agent/auth-reloadable.ts +24 -0
  4. package/src/agent/auth.ts +12 -0
  5. package/src/agent/index.ts +18 -58
  6. package/src/agent/todo/continuation-policy.ts +8 -0
  7. package/src/agent/todo/continuation-state.ts +41 -7
  8. package/src/agent/todo/continuation-wiring.ts +26 -0
  9. package/src/bundled-plugins/guard/policies/non-workspace-write.ts +0 -2
  10. package/src/bundled-plugins/guard/policy.ts +0 -5
  11. package/src/bundled-plugins/memory/README.md +26 -37
  12. package/src/bundled-plugins/memory/index.ts +32 -144
  13. package/src/bundled-plugins/memory/load-memory.ts +7 -121
  14. package/src/bundled-plugins/memory/memory-logger.ts +12 -6
  15. package/src/bundled-plugins/memory/search-command.ts +124 -0
  16. package/src/bundled-plugins/memory/vector/doctor.ts +1 -1
  17. package/src/bundled-plugins/memory/vector/embedder.ts +4 -5
  18. package/src/bundled-plugins/memory/vector/hybrid.ts +47 -5
  19. package/src/bundled-plugins/memory/vector/relevance-gate.ts +11 -7
  20. package/src/bundled-plugins/memory/vector/script.ts +87 -0
  21. package/src/channels/adapters/describe-error.ts +40 -0
  22. package/src/channels/adapters/discord.ts +9 -12
  23. package/src/channels/adapters/kakaotalk.ts +14 -17
  24. package/src/channels/adapters/line.ts +11 -14
  25. package/src/channels/adapters/slack-bot.ts +15 -18
  26. package/src/channels/adapters/slack.ts +45 -21
  27. package/src/channels/adapters/telegram-bot.ts +15 -18
  28. package/src/channels/adapters/webex-bot.ts +13 -14
  29. package/src/channels/adapters/webex.ts +14 -15
  30. package/src/channels/router.ts +60 -1
  31. package/src/cli/command-meta.ts +7 -16
  32. package/src/cli/compose.ts +21 -1
  33. package/src/cli/cron.ts +1 -1
  34. package/src/cli/docker-preflight.ts +42 -0
  35. package/src/cli/doctor.ts +1 -1
  36. package/src/cli/dreams.ts +1 -1
  37. package/src/cli/hostd.ts +12 -3
  38. package/src/cli/index.ts +1 -1
  39. package/src/cli/init.ts +24 -22
  40. package/src/cli/inspect.ts +1 -1
  41. package/src/cli/logs.ts +7 -0
  42. package/src/cli/memory.ts +85 -0
  43. package/src/cli/model.ts +1 -1
  44. package/src/cli/provider.ts +8 -1
  45. package/src/cli/reload.ts +6 -0
  46. package/src/cli/restart.ts +7 -0
  47. package/src/cli/shell.ts +7 -0
  48. package/src/cli/start.ts +7 -0
  49. package/src/cli/status.ts +9 -0
  50. package/src/cli/stop.ts +7 -0
  51. package/src/cli/tui.ts +6 -0
  52. package/src/config/config.ts +31 -8
  53. package/src/container/docker-app.ts +253 -0
  54. package/src/container/index.ts +11 -0
  55. package/src/container/start.ts +85 -41
  56. package/src/hostd/current-host-daemon.ts +21 -0
  57. package/src/hostd/daemon.ts +27 -3
  58. package/src/hostd/supervisor.ts +18 -3
  59. package/src/plugin/hooks.ts +3 -4
  60. package/src/plugin/types.ts +1 -1
  61. package/src/role-claim/client.ts +6 -1
  62. package/src/run/channel-session-factory.ts +3 -5
  63. package/src/run/index.ts +50 -17
  64. package/src/server/command-runner.ts +2 -11
  65. package/src/server/index.ts +8 -0
  66. package/src/skills/typeclaw-memory/SKILL.md +2 -2
  67. package/src/skills/typeclaw-plugins/SKILL.md +1 -1
  68. package/typeclaw.schema.json +1 -23
  69. package/src/bundled-plugins/guard/policies/memory-retrieval-cache-write.ts +0 -43
  70. package/src/bundled-plugins/memory/memory-retrieval.ts +0 -111
  71. package/src/bundled-plugins/memory/vector/cache-write.ts +0 -19
  72. package/src/bundled-plugins/memory/vector/config.ts +0 -28
@@ -183,7 +183,7 @@ export type SessionTurnStartEvent = {
183
183
  userPrompt: string
184
184
  origin?: SessionOrigin
185
185
  // Mutable ref: plugin writes retrieval results here; server/router reads after hook returns.
186
- // Only populated when vector.enabled and injection plan is index mode.
186
+ // Populated when a turn driver supplies a retrieval context bag.
187
187
  retrievalContext?: { results: string }
188
188
  }
189
189
 
@@ -13,6 +13,10 @@ export type ClaimSessionOptions = {
13
13
  role: string
14
14
  channel?: string
15
15
  ttlMs?: number
16
+ // Defaults to connectTimeoutMs. Split out so a loaded host can't let the
17
+ // open (transport) phase exhaust the budget meant for the connected phase
18
+ // and surface the wrong timeout error. See waitForOpen/waitForConnected.
19
+ openTimeoutMs?: number
16
20
  connectTimeoutMs?: number
17
21
  onStarted?: (payload: ClaimStartedPayload) => void
18
22
  }
@@ -28,11 +32,12 @@ const DEFAULT_CONNECT_TIMEOUT_MS = 30_000
28
32
  export async function runClaimSession(opts: ClaimSessionOptions): Promise<ClaimSessionResult> {
29
33
  const ttlMs = opts.ttlMs ?? DEFAULT_TTL_MS
30
34
  const connectTimeoutMs = opts.connectTimeoutMs ?? DEFAULT_CONNECT_TIMEOUT_MS
35
+ const openTimeoutMs = opts.openTimeoutMs ?? connectTimeoutMs
31
36
  const code = generateClaimCode()
32
37
 
33
38
  const ws = new WebSocket(opts.url)
34
39
  const displayUrl = redactUrl(opts.url)
35
- await waitForOpen(ws, displayUrl, connectTimeoutMs)
40
+ await waitForOpen(ws, displayUrl, openTimeoutMs)
36
41
  await waitForConnected(ws, displayUrl, connectTimeoutMs)
37
42
 
38
43
  try {
@@ -1,3 +1,5 @@
1
+ import { join } from 'node:path'
2
+
1
3
  import { SessionManager } from '@mariozechner/pi-coding-agent'
2
4
 
3
5
  import { createSession as defaultCreateSession } from '@/agent'
@@ -71,9 +73,6 @@ export type BuildChannelSessionFactoryDeps = {
71
73
  subagentRegistry?: SubagentRegistry
72
74
  getCreateSessionForSubagent?: () => CreateSessionForSubagent
73
75
  liveSessionRegistry?: LiveSessionRegistry
74
- // Forwarded to createSession: when true the `# Memory` section is omitted
75
- // from the system prompt (vector agents inject memory per-turn instead).
76
- suppressSystemMemory?: boolean
77
76
  }
78
77
 
79
78
  // Tight basename validation so a tampered or corrupt channels/sessions.json
@@ -141,7 +140,6 @@ export function buildChannelSessionFactory(deps: BuildChannelSessionFactoryDeps)
141
140
  ...(deps.getCreateSessionForSubagent !== undefined
142
141
  ? { createSessionForSubagent: deps.getCreateSessionForSubagent() }
143
142
  : {}),
144
- ...(deps.suppressSystemMemory !== undefined ? { suppressSystemMemory: deps.suppressSystemMemory } : {}),
145
143
  })
146
144
 
147
145
  const sessionId = sessionManager.getSessionId()
@@ -192,7 +190,7 @@ function tryReopenOrCreate(
192
190
  )
193
191
  return SessionManager.create(cwd, sessionDir)
194
192
  }
195
- const path = `${sessionDir}/${existingSessionFile}`
193
+ const path = join(sessionDir, existingSessionFile)
196
194
  if (capOptions !== null) {
197
195
  try {
198
196
  const stats = capJsonlFileInPlace(path, capOptions)
package/src/run/index.ts CHANGED
@@ -1,6 +1,7 @@
1
1
  import { SessionManager } from '@mariozechner/pi-coding-agent'
2
2
 
3
3
  import { createSession, createSessionWithDispose } from '@/agent'
4
+ import { createProviderAuthReloadable } from '@/agent/auth-reloadable'
4
5
  import { LiveSessionRegistry } from '@/agent/live-sessions'
5
6
  import { LiveSubagentRegistry } from '@/agent/live-subagents'
6
7
  import { requestContainerRestart } from '@/agent/restart'
@@ -19,8 +20,7 @@ import {
19
20
  type SubagentRegistry,
20
21
  type SubagentShared,
21
22
  } from '@/agent/subagents'
22
- import { clearTodosForOrigin } from '@/agent/todo/continuation-wiring'
23
- import { vectorEnabledFromMemoryConfig } from '@/bundled-plugins/memory/vector/config'
23
+ import { clearTodosForOrigin, markRestartAbortPendingForOrigin } from '@/agent/todo/continuation-wiring'
24
24
  import { embed, warmEmbedder } from '@/bundled-plugins/memory/vector/embedder'
25
25
  import { buildStartupVectorIndex } from '@/bundled-plugins/memory/vector/startup'
26
26
  import { resolveCapOptionsFromConfig } from '@/bundled-plugins/tool-result-cap'
@@ -211,14 +211,14 @@ async function startAgentRuntime(
211
211
  processGlobalsDisposed = true
212
212
  uninstallCodexFetchObserver()
213
213
  fatalGuard.dispose()
214
+ // onSigterm is defined later in this scope but only ever invoked after init;
215
+ // removing it here keeps a restarted startAgent() from stacking listeners.
216
+ if (sigtermHandler !== null) process.removeListener('SIGTERM', sigtermHandler)
214
217
  }
218
+ let sigtermHandler: (() => void) | null = null
215
219
  onProcessGlobalsInstalled(disposeProcessGlobals)
216
220
 
217
221
  const { config: cwdConfig, pluginConfigs: pluginConfigsByName } = loadConfigBundleSync(cwd)
218
- // Vector agents omit the system-prompt `# Memory` section and inject memory
219
- // per-turn instead. Derived once here: `memory.vector.enabled` is
220
- // restart-required, so a single boot read is coherent for the process.
221
- const suppressSystemMemory = vectorEnabledFromMemoryConfig(pluginConfigsByName.memory)
222
222
  const githubTokenBridge = createGithubTokenBridge()
223
223
  const mcpManager =
224
224
  cwdConfig.mcpServers.length > 0 ? createMcpManager(cwdConfig.mcpServers, { env: process.env }) : null
@@ -252,7 +252,7 @@ async function startAgentRuntime(
252
252
  hasGithubAppTokenResolver: githubTokenBridge.hasAppTokenResolver,
253
253
  ...(cwdConfig.roles !== undefined ? { roles: cwdConfig.roles } : {}),
254
254
  })
255
- const vectorStartupPromise = suppressSystemMemory ? runVectorStartup(cwd) : Promise.resolve()
255
+ const vectorStartupPromise = runVectorStartup(cwd)
256
256
  let pluginsLoaded: LoadPluginsResult
257
257
  try {
258
258
  const [loaded] = await Promise.all([pluginsLoadedPromise, vectorStartupPromise])
@@ -364,7 +364,6 @@ async function startAgentRuntime(
364
364
  stream,
365
365
  reloadRegistry,
366
366
  pluginRuntime,
367
- suppressSystemMemory,
368
367
  getChannelRouter: () => channelManager.router,
369
368
  rehydrateCapOptions: resolveCapOptionsFromConfig(pluginConfigsByName['tool-result-cap']),
370
369
  permissions: pluginsLoaded.permissions,
@@ -608,7 +607,6 @@ async function startAgentRuntime(
608
607
  containerName: containerNameOpt.containerName,
609
608
  sessionFactory,
610
609
  channelRouter: channelManager.router,
611
- suppressSystemMemory,
612
610
  ...mcpManagerOpt,
613
611
  }),
614
612
  subagent: (subName: string, payload?: unknown) =>
@@ -649,7 +647,6 @@ async function startAgentRuntime(
649
647
  channelRouter: channelManager.router,
650
648
  origin: cronOrigin,
651
649
  permissions: pluginsLoaded.permissions,
652
- suppressSystemMemory,
653
650
  ...(refOverride !== undefined ? { refOverride } : {}),
654
651
  ...(snap.hasAnyPluginContent
655
652
  ? {
@@ -743,6 +740,17 @@ async function startAgentRuntime(
743
740
  })
744
741
  })
745
742
 
743
+ // Registered before channels so its cache clear lands before any channel
744
+ // session teardown observes it. secrets.json provider credentials are not
745
+ // part of the typeclaw.json config diff, so a rotated key takes effect on
746
+ // `typeclaw reload` only via this dedicated scope. Live sessions captured
747
+ // their AuthStorage at creation, so teardown recreates them with fresh auth.
748
+ reloadRegistry.register(
749
+ createProviderAuthReloadable({
750
+ onProviderAuthChanged: () => channelManager.router.tearDownAllLive(),
751
+ }),
752
+ )
753
+
746
754
  reloadRegistry.register(createChannelsReloadable({ manager: channelManager }))
747
755
 
748
756
  // Two-phase channel restart-resume around adapter startup, to close the race
@@ -782,12 +790,10 @@ async function startAgentRuntime(
782
790
  // In-flight coalescing for direct ctx.spawnSubagent calls mirrors the
783
791
  // SubagentConsumer's stream-path gate (subagents.ts:441). Two queued
784
792
  // `new-session` messages for the same (name, inFlightKey) drop the second
785
- // on the consumer side; without the same gate here, two consecutive
786
- // session.prompt fires (cold-start prompt N immediately followed by prompt
787
- // N+1 on the same channel session) could both fire memory-retrieval spawns
788
- // racing to write `memory/.retrieval-cache/<sessionId>.md`. Awaiting the
789
- // spawn in the hook used to mask this; now that the hook is fire-and-forget,
790
- // the race is exposed and the gate is mandatory.
793
+ // on the consumer side; without the same gate here, two consecutive direct
794
+ // plugin spawns with the same in-flight key could race. Awaiting the spawn in
795
+ // the hook used to mask this; detached hook paths expose the race and make the
796
+ // gate mandatory.
791
797
  //
792
798
  // Same key shape as the consumer: `${name}:${inFlightKey(payload)}` when the
793
799
  // subagent declares one, else just `${name}`. Collisions resolve cleanly
@@ -894,7 +900,6 @@ async function startAgentRuntime(
894
900
  outbound,
895
901
  sessionFactory,
896
902
  channelRouter: channelManager.router,
897
- suppressSystemMemory,
898
903
  ...mcpManagerOpt,
899
904
  })
900
905
 
@@ -954,6 +959,34 @@ async function startAgentRuntime(
954
959
  }
955
960
  }
956
961
 
962
+ // Graceful shutdown on host-initiated restart (`typeclaw restart` → docker
963
+ // stop → SIGTERM, ~10s before SIGKILL). Mark every live session's todo scope
964
+ // so the turn this restart aborts does not arm the durable user-abort block —
965
+ // each scope's incomplete todos then auto-continue after the new container
966
+ // boots. Then run best-effort teardown. A hard deadline force-exits well
967
+ // inside Docker's grace window so a hung turn can't get the process SIGKILL'd
968
+ // mid-flush. Only meaningful inside a container; outside Docker SIGTERM is an
969
+ // ordinary stop with nothing to resume.
970
+ const SHUTDOWN_DEADLINE_MS = 8_000
971
+ let shuttingDown = false
972
+ const onSigterm = (): void => {
973
+ if (shuttingDown) return
974
+ shuttingDown = true
975
+ const forceExit = setTimeout(() => process.exit(0), SHUTDOWN_DEADLINE_MS)
976
+ forceExit.unref()
977
+ void (async () => {
978
+ if (containerName !== undefined) {
979
+ await markRestartAbortPendingForOrigin(cwd, { kind: 'tui', sessionId: 'tui' }).catch(() => undefined)
980
+ await channelManager.router.markRestartAbortForAllLive().catch(() => undefined)
981
+ }
982
+ await stop().catch(() => undefined)
983
+ clearTimeout(forceExit)
984
+ process.exit(0)
985
+ })()
986
+ }
987
+ sigtermHandler = onSigterm
988
+ process.on('SIGTERM', onSigterm)
989
+
957
990
  if (!attachTui) {
958
991
  return {
959
992
  server,
@@ -58,10 +58,6 @@ export type CommandRunnerOptions = {
58
58
  // wire for the handler/command path.
59
59
  channelRouter: ChannelRouter | undefined
60
60
  mcpManager?: McpManager
61
- // When true, prompt sessions spawned here omit the system-prompt `# Memory`
62
- // section (vector agents inject memory per-turn). Forwarded to createSession
63
- // so command/handler sessions stay coherent with the rest of the runtime.
64
- suppressSystemMemory?: boolean
65
61
  }
66
62
 
67
63
  type CommandHandle = {
@@ -201,7 +197,6 @@ export function createCommandRunner(opts: CommandRunnerOptions): CommandRunner {
201
197
  signal: abortController.signal,
202
198
  sessionFactory: opts.sessionFactory,
203
199
  channelRouter: opts.channelRouter,
204
- ...(opts.suppressSystemMemory !== undefined ? { suppressSystemMemory: opts.suppressSystemMemory } : {}),
205
200
  ...(opts.mcpManager !== undefined ? { mcpManager: opts.mcpManager } : {}),
206
201
  }),
207
202
  subagent: (subName, payload) =>
@@ -388,7 +383,6 @@ export async function runPromptForCommand(args: {
388
383
  // so the spawned session exposes `channel_send`.
389
384
  channelRouter?: ChannelRouter
390
385
  mcpManager?: McpManager
391
- suppressSystemMemory?: boolean
392
386
  // Test seam for the agent-session boundary. Production passes the real
393
387
  // `createSessionWithDispose`; tests inject a fake to verify wiring
394
388
  // (specifically: the sessionManager handed off must be persisted, not
@@ -418,14 +412,11 @@ export async function runPromptForCommand(args: {
418
412
  ...(args.mcpManager !== undefined ? { mcpManager: args.mcpManager } : {}),
419
413
  ...(args.runtimeVersion !== undefined ? { runtimeVersion: args.runtimeVersion } : {}),
420
414
  ...(args.containerName !== undefined ? { containerName: args.containerName } : {}),
421
- ...(args.suppressSystemMemory !== undefined ? { suppressSystemMemory: args.suppressSystemMemory } : {}),
422
415
  })
423
416
  const detachAbort = bindSignalToSession(args.signal, session)
424
417
  // Mirror the other turn drivers (TUI/channel/cron/subagent): fire
425
- // session.turn.start with a retrievalContext so a vector agent — whose
426
- // system-prompt `# Memory` section is suppressed — gets its long-term memory
427
- // injected per-turn into the user prompt here too. Without this, command and
428
- // handler prompt sessions would have no memory at all under vector mode.
418
+ // session.turn.start with a retrievalContext so vector memory is injected
419
+ // per-turn into command and handler prompts too.
429
420
  const turnEvent = { sessionId, agentDir: args.agentDir, origin: args.origin }
430
421
  const retrievalContext = { results: '' }
431
422
  try {
@@ -25,6 +25,7 @@ import type { CreateSessionForSubagent } from '@/agent/subagents'
25
25
  import { TODO_CONTINUATION_SOURCE } from '@/agent/todo/continuation'
26
26
  import {
27
27
  armRestartKickForOrigin,
28
+ clearAbortSuppressionForOrigin,
28
29
  extractTurnUsage,
29
30
  recordTurnOutcome,
30
31
  recordTurnStart,
@@ -620,6 +621,13 @@ export function createServer({
620
621
  await armRestartKickForOrigin(agentDir, origin).catch((err) =>
621
622
  logger.error(`[server] ${sessionFileId}: arm restart-kick suppression failed: ${describeErr(err)}`),
622
623
  )
624
+ // The restart may have aborted an in-flight turn, arming the
625
+ // durable user-abort block. This resume IS the restart, so
626
+ // clear it — otherwise the resumed session never auto-continues
627
+ // its incomplete todos. Gated to this restart-handoff path.
628
+ await clearAbortSuppressionForOrigin(agentDir, origin).catch((err) =>
629
+ logger.error(`[server] ${sessionFileId}: clear abort suppression failed: ${describeErr(err)}`),
630
+ )
623
631
  }
624
632
  stream.publish({
625
633
  target: { kind: 'session', sessionId: sessionFileId },
@@ -9,9 +9,9 @@ The agent's long-term memory is sharded across files in `memory/topics/<slug>.md
9
9
 
10
10
  ## Reading
11
11
 
12
- The `# Memory` section comes from topic shards only. Undreamed daily-stream events are **not** injected — call `memory_search` when you need them. When total shard bytes are above the 16 KB injection budget (or when speaking in a channel), shard bodies are dropped — only the heading + `cites=N, days=N, lastReinforced=YYYY-MM-DD` shows; call `memory_search` to fetch the bodies you need. The same `memory_search` covers both surfaces (topic shards and undreamed stream events), so one tool call reaches everything.
12
+ The `# Memory` block is injected per turn into the user prompt; the system prompt never contains long-term memory; vector memory is always on. Undreamed daily-stream events are **not** dumped wholesale into prompts — use `memory_search` when you need them. When speaking in a channel, shard bodies are dropped and only headings/slugs show; call `memory_search` to fetch the bodies you need. The same `memory_search` covers both surfaces (topic shards and undreamed stream events), so one tool call reaches everything.
13
13
 
14
- **Where the `# Memory` section lives depends on `memory.vector.enabled`.** With vector **off** (default), it's part of the system prompt, snapshotted once at session creation. With vector **on**, it is removed from the system prompt and injected fresh into each **user turn** instead: under budget you get all shard bodies, over budget you get the top-K shards/fragments most relevant to the current message (hybrid vector + keyword search). This keeps the system-prompt cache prefix stable across a session and lets retrieval track the current topic instead of a stale session-start snapshot. Either way `memory_search` remains available on demand.
14
+ Under budget you get shard bodies de-duplicated across turns; over budget you get the top-K shards/fragments/references most relevant to the current message (hybrid vector + keyword search). This keeps the system-prompt cache prefix stable across a session and lets retrieval track the current topic instead of a stale session-start snapshot. `memory_search` remains available on demand.
15
15
 
16
16
  ## Writing
17
17
 
@@ -380,7 +380,7 @@ hooks: {
380
380
 
381
381
  The `prompt` field on `SessionPromptEvent` is the system prompt as it's being composed by `createResourceLoader` (`basePrompt + IDENTITY.md + SOUL.md`), NOT the user's most recent message. Reading it as if it were the user's prompt — and feeding it to a retrieval system, classifier, or LLM — will keyword-mine TypeClaw's framing prose (`TypeClaw`, `subagent`, `AGENTS.md`) on every session.
382
382
 
383
- If you want the **user's actual prompt** (their message text), subscribe to `session.turn.start` and read `event.userPrompt`. The bundled memory plugin's `memory-retrieval` subagent learned this the hard way; see `src/bundled-plugins/memory/index.ts`'s `session.turn.start` handler.
383
+ If you want the **user's actual prompt** (their message text), subscribe to `session.turn.start` and read `event.userPrompt`. The bundled memory plugin's per-turn retrieval uses this hook so it embeds the user's text instead of the assembling system prompt; see `src/bundled-plugins/memory/index.ts`'s `session.turn.start` handler.
384
384
 
385
385
  #### CRITICAL: `session.prompt` and provider prompt caching
386
386
 
@@ -2528,11 +2528,7 @@
2528
2528
  "bufferBytes": 500000,
2529
2529
  "injectionBudgetBytes": 16384,
2530
2530
  "minIdleDeltaLines": 3,
2531
- "spawnTimeoutMs": 50000,
2532
- "retrievalSpawnTimeoutMs": 30000,
2533
- "vector": {
2534
- "enabled": false
2535
- }
2531
+ "spawnTimeoutMs": 50000
2536
2532
  },
2537
2533
  "type": "object",
2538
2534
  "properties": {
@@ -2566,12 +2562,6 @@
2566
2562
  "minimum": 1,
2567
2563
  "maximum": 9007199254740991
2568
2564
  },
2569
- "retrievalSpawnTimeoutMs": {
2570
- "default": 30000,
2571
- "type": "integer",
2572
- "minimum": 1,
2573
- "maximum": 9007199254740991
2574
- },
2575
2565
  "dreaming": {
2576
2566
  "type": "object",
2577
2567
  "properties": {
@@ -2580,18 +2570,6 @@
2580
2570
  "minLength": 1
2581
2571
  }
2582
2572
  }
2583
- },
2584
- "vector": {
2585
- "default": {
2586
- "enabled": false
2587
- },
2588
- "type": "object",
2589
- "properties": {
2590
- "enabled": {
2591
- "default": false,
2592
- "type": "boolean"
2593
- }
2594
- }
2595
2573
  }
2596
2574
  }
2597
2575
  },
@@ -1,43 +0,0 @@
1
- import path from 'node:path'
2
-
3
- import type { SessionOrigin } from '@/agent/session-origin'
4
- import type { SecuritySeverity } from '@/bundled-plugins/security/permissions'
5
-
6
- export const GUARD_MEMORY_RETRIEVAL_CACHE_WRITE = 'memoryRetrievalCacheWrite'
7
- export const GUARD_MEMORY_RETRIEVAL_CACHE_WRITE_SEVERITY: SecuritySeverity = 'low'
8
-
9
- const SESSION_ID_REGEX = /^[A-Za-z0-9._-]{1,128}$/
10
-
11
- export async function isMemoryRetrievalCacheWriteAllowed(options: {
12
- tool: string
13
- args: Record<string, unknown>
14
- agentDir: string
15
- origin?: SessionOrigin
16
- }): Promise<boolean> {
17
- const { tool, args, agentDir, origin } = options
18
- if (tool !== 'write') return false
19
- // Allow the memory-retrieval subagent (existing path) OR the in-process
20
- // vector retrieval writer (system origin with component='memory-retrieval').
21
- // The in-process path runs under the session origin, not a subagent, so we
22
- // check for the system component name as the trusted internal actor.
23
- const isMemoryRetrievalSubagent = origin?.kind === 'subagent' && origin.subagent === 'memory-retrieval'
24
- const isInProcessWriter = origin?.kind === 'system' && origin.component === 'memory-retrieval'
25
- if (!isMemoryRetrievalSubagent && !isInProcessWriter) return false
26
-
27
- const rawPath = args.path
28
- if (typeof rawPath !== 'string') return false
29
-
30
- const targetPath = path.resolve(agentDir, rawPath)
31
- const expectedDir = path.resolve(agentDir, 'memory', '.retrieval-cache')
32
- const relative = path.relative(expectedDir, targetPath)
33
- if (relative === '' || relative.startsWith('..') || path.isAbsolute(relative)) return false
34
-
35
- const parts = relative.split(path.sep).filter(Boolean)
36
- if (parts.length !== 1) return false
37
-
38
- const fileName = parts[0]!
39
- if (!fileName.endsWith('.md')) return false
40
-
41
- const sessionId = fileName.slice(0, -3)
42
- return SESSION_ID_REGEX.test(sessionId)
43
- }
@@ -1,111 +0,0 @@
1
- import { z } from 'zod'
2
-
3
- import { lsTool, readTool, type Subagent, writeTool } from '@/plugin'
4
-
5
- import { createMemorySearchTool } from './search-tool'
6
-
7
- export const memoryRetrievalPayloadSchema = z.object({
8
- parentSessionId: z.string().min(1),
9
- agentDir: z.string().min(1),
10
- recentPrompt: z.string(),
11
- cacheFilePath: z.string().min(1),
12
- origin: z.unknown().optional(),
13
- })
14
-
15
- export type MemoryRetrievalPayload = z.infer<typeof memoryRetrievalPayloadSchema>
16
-
17
- export function isMemoryRetrievalPayload(value: unknown): value is MemoryRetrievalPayload {
18
- return memoryRetrievalPayloadSchema.safeParse(value).success
19
- }
20
-
21
- export type MemoryRetrievalLogger = {
22
- info: (msg: string) => void
23
- warn: (msg: string) => void
24
- error: (msg: string) => void
25
- }
26
-
27
- export type CreateMemoryRetrievalSubagentOptions = {
28
- logger?: MemoryRetrievalLogger
29
- timeoutMs?: number
30
- }
31
-
32
- export const MEMORY_RETRIEVAL_SYSTEM_PROMPT = `You are the memory-retrieval subagent. Read the user's most recent prompt and decide what's relevant across topic shards in \`memory/topics/\` (consolidated long-term memory), references in \`memory/references/\` (verbatim artifacts), AND undreamed daily-stream events under \`memory/streams/\` (recent fragments not yet folded into shards). Use \`memory_search\` to query all three surfaces; use \`read\`/\`ls\` to pull full shard bodies when needed. Synthesize a focused ≤8 KB summary of the relevant memory. Save by \`write\`ing it to the exact path provided in your payload as \`cacheFilePath\`. Be ruthlessly concise. Do NOT write anywhere else. Do NOT delete files.
33
-
34
- Search discipline: issue ALL your \`memory_search\` queries in a SINGLE response as parallel tool calls (up to 3 at once), then wait for every result before deciding what to do next. Different angles in parallel, NEVER one search per turn — sequential searches waste a full LLM round-trip per query (~3s each) on file I/O that takes milliseconds. Pick queries that match the user's literal phrasing — not framing vocabulary, not metadata (session ids, dates), not words from your own system prompt. If the parallel batch turns up nothing relevant, write the empty-context note and stop.`
35
-
36
- export function memoryRetrievalExhaustedMessage(used: number, max: number): string {
37
- const usedKb = Math.round(used / 1024)
38
- const maxKb = Math.round(max / 1024)
39
- return [
40
- `[memory-retrieval budget exhausted: used ${usedKb}KB of ${maxKb}KB across memory_search and read]`,
41
- '',
42
- 'Stop searching. Stop reading. Every subsequent memory_search or read call will return this same notice.',
43
- 'Write the cache file at the provided cacheFilePath with whatever relevant memory you have already gathered.',
44
- 'If nothing was relevant, write a short empty-context note to the cache file and stop.',
45
- ].join('\n')
46
- }
47
-
48
- const consoleLogger: MemoryRetrievalLogger = {
49
- info: (m) => console.warn(m),
50
- warn: (m) => console.warn(m),
51
- error: (m) => console.error(m),
52
- }
53
-
54
- export function createMemoryRetrievalSubagent(
55
- options: CreateMemoryRetrievalSubagentOptions = {},
56
- ): Subagent<MemoryRetrievalPayload> {
57
- const logger = options.logger ?? consoleLogger
58
- return {
59
- systemPrompt: MEMORY_RETRIEVAL_SYSTEM_PROMPT,
60
- // Retrieval is "4 keyword searches + 1 write" — no reasoning required.
61
- // `fast` falls back to `default` (with a one-time warning) when the
62
- // operator hasn't configured it, so this is safe by construction.
63
- profile: 'fast',
64
- tools: [readTool, writeTool, lsTool],
65
- customTools: [createMemorySearchTool()],
66
- payloadSchema: memoryRetrievalPayloadSchema,
67
- inFlightKey: (payload) => payload.parentSessionId,
68
- ...(options.timeoutMs !== undefined ? { timeoutMs: options.timeoutMs } : {}),
69
- // 256 KB read + memory_search budget. Sized for one retrieval pass:
70
- // ~16 KB of memory_search hits (3 queries × ~5 KB excerpts) plus a few
71
- // shard reads (~5 KB each). A smaller budget would systematically
72
- // exhaust on any agent with rich memory; a larger budget invites the
73
- // pre-fix failure mode where the LLM kept iterating searches until it
74
- // gave up. The exhausted-message tells the subagent to write the
75
- // cache file with what it has rather than retrying forever.
76
- toolResultBudget: {
77
- maxTotalBytes: 256 * 1024,
78
- toolNames: ['read', 'memory_search'],
79
- exhaustedMessage: memoryRetrievalExhaustedMessage,
80
- },
81
- handler: async (ctx, runSession) => {
82
- const start = Date.now()
83
- logger.info(`[memory-retrieval] ${ctx.payload.parentSessionId} start cache=${ctx.payload.cacheFilePath}`)
84
- try {
85
- await runSession({ userPrompt: buildInitialPrompt(ctx.payload) })
86
- logger.info(`[memory-retrieval] ${ctx.payload.parentSessionId} done elapsed_ms=${Date.now() - start}`)
87
- } catch (err) {
88
- const message = err instanceof Error ? err.message : String(err)
89
- logger.warn(
90
- `[memory-retrieval] ${ctx.payload.parentSessionId}: run threw: ${message} elapsed_ms=${Date.now() - start}`,
91
- )
92
- throw err
93
- }
94
- },
95
- }
96
- }
97
-
98
- function buildInitialPrompt(payload: MemoryRetrievalPayload): string {
99
- return [
100
- `Parent session: ${payload.parentSessionId}`,
101
- `Agent folder: ${payload.agentDir}`,
102
- `Recent user prompt: ${payload.recentPrompt}`,
103
- `Topic shard directory: memory/topics/`,
104
- `Daily-stream directory: memory/streams/`,
105
- `Cache output path: ${payload.cacheFilePath}`,
106
- '',
107
- 'Use `memory_search` to find relevant material across BOTH topic shards and undreamed stream events (results are discriminated by `source: "topic" | "stream"`). Read any shard whose body you need in full via `read`. Write one concise retrieval summary to the cache output path exactly as provided. Keep the file ≤8 KB. If nothing is relevant, write a short empty-context note to the cache output path. Do not write any other path.',
108
- ].join('\n')
109
- }
110
-
111
- export const memoryRetrievalSubagent: Subagent<MemoryRetrievalPayload> = createMemoryRetrievalSubagent()
@@ -1,19 +0,0 @@
1
- import { mkdir, writeFile, rename } from 'node:fs/promises'
2
- import { join } from 'node:path'
3
-
4
- /**
5
- * Atomically write retrieval cache to memory/.retrieval-cache/<sessionId>.md
6
- * using tmp+rename pattern for crash safety.
7
- */
8
- export async function writeRetrievalCache(agentDir: string, sessionId: string, content: string): Promise<void> {
9
- const cacheDir = join(agentDir, 'memory', '.retrieval-cache')
10
- const cachePath = join(cacheDir, `${sessionId}.md`)
11
- const tmpPath = `${cachePath}.tmp`
12
-
13
- // Create directory if it doesn't exist
14
- await mkdir(cacheDir, { recursive: true })
15
-
16
- // Write to tmp file, then atomically rename
17
- await writeFile(tmpPath, content, 'utf8')
18
- await rename(tmpPath, cachePath)
19
- }
@@ -1,28 +0,0 @@
1
- import { z } from 'zod'
2
-
3
- import { loadPluginConfigsSync } from '@/config'
4
-
5
- export const vectorConfigSchema = z
6
- .object({
7
- enabled: z.boolean().default(false),
8
- })
9
- .default({ enabled: false })
10
-
11
- export type VectorConfig = z.infer<typeof vectorConfigSchema>
12
-
13
- // Fails closed to `false`: a memory block we can't parse is treated as opted
14
- // out. Shared by the host-side download gate and the runtime's per-turn vs
15
- // system-prompt memory-injection decision, so both read the flag identically.
16
- export function vectorEnabledFromMemoryConfig(memory: unknown): boolean {
17
- if (typeof memory !== 'object' || memory === null) return false
18
- const parsed = vectorConfigSchema.safeParse((memory as Record<string, unknown>).vector)
19
- return parsed.success && parsed.data.enabled
20
- }
21
-
22
- export function agentUsesVector(cwd: string): boolean {
23
- try {
24
- return vectorEnabledFromMemoryConfig(loadPluginConfigsSync(cwd).memory)
25
- } catch {
26
- return false
27
- }
28
- }