pi-braid 0.40.0 → 0.40.2

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 (85) hide show
  1. package/CONTRIBUTING.md +27 -3
  2. package/README.md +39 -33
  3. package/TOOLS.md +27 -3
  4. package/build-info.json +9 -0
  5. package/dist/extensions/cognitive/metacognition/handlers/hypothesize.d.ts +9 -0
  6. package/dist/extensions/cognitive/metacognition/handlers/hypothesize.js +13 -1
  7. package/dist/extensions/cognitive/metacognition/handlers/verify.js +35 -4
  8. package/dist/extensions/core/prompt-builder/inspect.js +9 -3
  9. package/dist/extensions/enforcement/handlers/register-tool-result-formatting.d.ts +1 -1
  10. package/dist/extensions/enforcement/handlers/register-tool-result-formatting.js +1 -1
  11. package/dist/extensions/enforcement/handlers/turn-end-handler.js +1 -1
  12. package/dist/extensions/enforcement/self-state-claims.d.ts +1 -1
  13. package/dist/extensions/enforcement/self-state-claims.js +3 -2
  14. package/dist/extensions/manifest.js +2 -2
  15. package/dist/extensions/memory/context-manager/handlers/tool-batch-boundary.d.ts +22 -5
  16. package/dist/extensions/memory/context-manager/handlers/tool-batch-boundary.js +43 -11
  17. package/dist/extensions/memory/context-manager/handlers/tools/context-compress.js +174 -50
  18. package/dist/extensions/memory/context-manager/handlers/tools/context-navigate.js +4 -1
  19. package/dist/extensions/memory/memory-agent-bridge/memory-store-tool.js +16 -0
  20. package/dist/extensions/memory/memory-agent-bridge/memory-tool.js +21 -9
  21. package/dist/extensions/memory/scratchpad/format.js +8 -2
  22. package/dist/extensions/memory/scratchpad/index.js +11 -2
  23. package/dist/extensions/memory/session-recall/index.js +19 -6
  24. package/dist/extensions/memory/three-memory-stores/episodes.d.ts +1 -1
  25. package/dist/extensions/memory/three-memory-stores/episodes.js +2 -2
  26. package/dist/extensions/orchestrator/index.js +73 -8
  27. package/dist/extensions/self/health-maintenance/index.js +112 -9
  28. package/dist/lib/config/braid-settings.d.ts +21 -0
  29. package/dist/lib/config/braid-settings.js +36 -0
  30. package/dist/lib/constants/session-thresholds.d.ts +2 -3
  31. package/dist/lib/constants/session-thresholds.js +2 -3
  32. package/dist/lib/enforcement/turn-dedup.d.ts +2 -2
  33. package/dist/lib/enforcement/turn-dedup.js +2 -2
  34. package/dist/lib/harness/index.d.ts +1 -1
  35. package/dist/lib/harness/index.js +1 -1
  36. package/dist/lib/harness/types.d.ts +1 -1
  37. package/dist/lib/harness/types.js +1 -1
  38. package/dist/lib/orchestrator/index.d.ts +2 -2
  39. package/dist/lib/orchestrator/index.js +1 -1
  40. package/dist/lib/orchestrator/workers.d.ts +184 -2
  41. package/dist/lib/orchestrator/workers.js +521 -6
  42. package/dist/lib/utils/injection-tracker.js +1 -1
  43. package/docs/BEGINNERS.md +22 -0
  44. package/docs/CONTEXT-POLICY.md +13 -5
  45. package/docs/LOCAL_DEV.md +31 -5
  46. package/docs/ROADMAP.md +10 -7
  47. package/docs/SETUP.md +86 -42
  48. package/package.json +11 -8
  49. package/skills/braid-reference/SKILL.md +23 -23
  50. package/skills/release-pipeline/SKILL.md +46 -37
  51. package/skills/subagent-orchestration/SKILL.md +51 -1
  52. package/src/extensions/cognitive/metacognition/handlers/hypothesize.ts +14 -1
  53. package/src/extensions/cognitive/metacognition/handlers/verify.ts +39 -5
  54. package/src/extensions/core/prompt-builder/inspect.ts +10 -3
  55. package/src/extensions/enforcement/AGENTS.md +25 -19
  56. package/src/extensions/enforcement/handlers/register-tool-result-formatting.ts +1 -1
  57. package/src/extensions/enforcement/handlers/turn-end-handler.ts +1 -1
  58. package/src/extensions/enforcement/self-state-claims.ts +3 -2
  59. package/src/extensions/heartbeat/heartbeat-scheduler/AGENTS.md +6 -3
  60. package/src/extensions/manifest.ts +2 -2
  61. package/src/extensions/memory/AGENTS.md +13 -5
  62. package/src/extensions/memory/context-manager/handlers/tool-batch-boundary.ts +45 -11
  63. package/src/extensions/memory/context-manager/handlers/tools/context-compress.ts +202 -60
  64. package/src/extensions/memory/context-manager/handlers/tools/context-navigate.ts +4 -1
  65. package/src/extensions/memory/memory-agent-bridge/memory-store-tool.ts +16 -0
  66. package/src/extensions/memory/memory-agent-bridge/memory-tool.ts +22 -9
  67. package/src/extensions/memory/scratchpad/format.ts +8 -2
  68. package/src/extensions/memory/scratchpad/index.ts +12 -2
  69. package/src/extensions/memory/session-recall/index.ts +20 -8
  70. package/src/extensions/memory/three-memory-stores/episodes.ts +2 -2
  71. package/src/extensions/orchestrator/index.ts +100 -11
  72. package/src/extensions/scheduling/timer/AGENTS.md +7 -0
  73. package/src/extensions/self/health-maintenance/index.ts +143 -14
  74. package/src/extensions/self/watchdog/AGENTS.md +1 -0
  75. package/src/lib/AGENTS.md +21 -21
  76. package/src/lib/config/braid-settings.ts +74 -0
  77. package/src/lib/constants/session-thresholds.ts +2 -3
  78. package/src/lib/enforcement/turn-dedup.ts +2 -2
  79. package/src/lib/harness/AGENTS.md +1 -1
  80. package/src/lib/harness/index.ts +1 -1
  81. package/src/lib/harness/types.ts +1 -1
  82. package/src/lib/orchestrator/index.ts +26 -0
  83. package/src/lib/orchestrator/workers.ts +705 -7
  84. package/src/lib/utils/AGENTS.md +1 -1
  85. package/src/lib/utils/injection-tracker.ts +1 -1
@@ -210,6 +210,56 @@ When a worker signals completion, review critically before closing the pane.
210
210
  - **Worker crashes mid-task**: orchestrator detects via `agent_get` status change, spawns replacement.
211
211
  - **Orchestrator is busy**: worker retries signal with backoff (max 3, 60s apart), then writes results to a well-known path.
212
212
 
213
+ ### Gate evidence rows and the population stamp
214
+
215
+ `close-worker` takes `gates` as `{cmd, exit}` rows, one per gate the wave ran. A gate whose
216
+ subject is a file or row population records what it scanned, and at which revision, as an
217
+ optional `stamp` on its row:
218
+
219
+ ```json
220
+ {
221
+ "cmd": "node scripts/check-placement.mjs --strict",
222
+ "exit": 0,
223
+ "stamp": "scanned 914 files at <revision>"
224
+ }
225
+ ```
226
+
227
+ `n` is a non-negative integer, `unit` one lowercase word (`files`, `rows`, `cases`, `lines`),
228
+ and `revision` a 7-to-40-char hex sha or a ref name. `check-result` warns; it does not fail
229
+ when a row whose command carries a walk, list or count verb (`ls`, `find`, `fd`, `git ls-files`,
230
+ `wc`, `rg`, `grep`) carries no stamp or an unreadable one.
231
+
232
+ `verify-and-merge` reads the close's stamped `files` rows and compares each number against that
233
+ revision's tracked-entry census. A recorded number below the census refuses the merge and names
234
+ the delta: the gate certified fewer files than the commit carries, which is what a run before
235
+ `git add` looks like. Re-run the gate on the staged tree and re-issue the close. A close with no
236
+ stamped `files` row prints `no recorded population` and merges.
237
+
238
+ | Setting | Default | Effect | Read cadence |
239
+ | -------------------------------------------- | -------- | -------------------------------------------------------------------------------------------------------------------------------------------- | -------------------------------------------------- |
240
+ | `braid.orchestrator.closeGateWellFormedness` | `shadow` | `shadow` logs invalid `{cmd, exit}` rows and allows close; `enforce` refuses an invalid row; `off` skips validation. Commands are not rerun. | Re-read at use time on every `close-worker` close. |
241
+
242
+ ### Claim verifiers and the instrument-capability warn
243
+
244
+ Each claim's `verified_by` is read against a small registry of instrument shapes: search and <!-- doc-ref-exempt: seat-side result-block field; the parser lives in the orchestrator seat, not this repo -->
245
+ count verbs (`grep`, `rg`, `wc`, `count`, `hits`), artifact citations (a report, a readme, a
246
+ description), execution verbs (`boot`, `install`, `exit`, `suite`), content verbs (`read`,
247
+ `measured`, `bytes`, `entries`), and enumeration instruments (`git ls-files`, an AST or
248
+ type-aware census, a suite run, a row count). The claim's wording is read against the property
249
+ class it asserts, and `check-result` warns; it does not fail, when the pair cannot observe
250
+ what the claim asserts:
251
+
252
+ - a coverage or elimination claim ("gone", "no remaining", "every reader") grounded on a
253
+ search or count verb with no census named;
254
+ - a content claim grounded only on an artifact citation (a description is not the content);
255
+ - a classification claim ("only", "private", "leak") grounded on a bare count with no
256
+ decomposition named.
257
+
258
+ The warn names the census shape it wants (`git ls-files`, an AST or type-aware census, a suite
259
+ run, or an explicit row count) and the counter line `claim-capability: k/n checked, m warn(s)`
260
+ reports the share. The tier is warn-only: a mismatched pair leaves the block valid and the exit
261
+ code unchanged. A verifier the registry cannot classify is left to the seat's re-derivation.
262
+
213
263
  ## Completion Signal
214
264
 
215
265
  ### For orchestrate-spawned workers
@@ -239,7 +289,7 @@ braid ships a context-manager: `context_navigate` (tag / checkout) and `context_
239
289
 
240
290
  Enforce a HARD context budget BEFORE degradation begins. Mistakes and loop-risk rise sharply at ~75–100k tokens, independent of window size.
241
291
 
242
- **Trigger (absolute, not %):** when `context_inspect` reports **>80k tokens** OR **>35% of the window** (whichever comes first):
292
+ **Trigger (absolute, not %):** when `context_inspect` reports **>80k tokens** OR **>35% of the window** (whichever comes first). `context_inspect` ships deferred: if it is not in your active set, find it with `tool_search` first:
243
293
 
244
294
  1. `context_navigate action=tag name="pre-archive-<n>"` - mark your position
245
295
  2. `context_compress action=archive topic="<task>"` - archive everything before the tag
@@ -50,6 +50,18 @@ export function emptyResponseRetryOptions(lastError: string): LlmCallOptions | n
50
50
  return { temperature: 0.3, maxTokens: 4000, reasoning: "off" };
51
51
  }
52
52
 
53
+ /**
54
+ * The hypothesis count a call asks for. Argument validation converts a blank
55
+ * string to 0 for `Type.Optional(Type.Number())` and the converted value passes,
56
+ * so a handler cannot tell a blank from a typed 0; a prompt asking the model for
57
+ * zero hypotheses is never a usable request, so a non-positive or non-finite
58
+ * value is read as absent and the documented default (3) applies. Exported pure
59
+ * for the fixture, like emptyResponseRetryOptions above.
60
+ */
61
+ export function resolveMaxHypotheses(value: unknown): number {
62
+ return typeof value === "number" && Number.isFinite(value) && value > 0 ? value : 3;
63
+ }
64
+
53
65
  const FORMAT_OPTIONS = ["ranked", "tree", "list"] as const;
54
66
 
55
67
  const hypothesizeSchema = Type.Object({
@@ -133,7 +145,8 @@ export function createHypothesizeTool(): ToolDefinition<any, any, any> {
133
145
  _onUpdate: unknown,
134
146
  ctx: ExtensionContext | undefined,
135
147
  ) => {
136
- const { problem, evidence, maxHypotheses = 3, format = "ranked", verificationPlan } = params;
148
+ const { problem, evidence, format = "ranked", verificationPlan } = params;
149
+ const maxHypotheses = resolveMaxHypotheses(params.maxHypotheses);
137
150
 
138
151
  if (!ctx?.model) {
139
152
  throw new Error(
@@ -83,6 +83,23 @@ const DEFAULT_CRITERIA: Record<VerificationTarget, string[]> = {
83
83
  ],
84
84
  };
85
85
 
86
+ /** A blank string is "not provided": clients that emit every schema key send "" for the unused ones. */
87
+ function optionalString(v: unknown): string | undefined {
88
+ return typeof v === "string" && v.trim().length > 0 ? v : undefined;
89
+ }
90
+
91
+ /**
92
+ * Drop blank criteria entries before use. A list left empty names nothing
93
+ * usable, so the target's defaults stand in: the LLM tier must never receive
94
+ * an empty rubric.
95
+ */
96
+ function resolveCriteria(provided: unknown, target: VerificationTarget): string[] {
97
+ const cleaned = Array.isArray(provided)
98
+ ? provided.filter((c): c is string => typeof c === "string" && c.trim().length > 0)
99
+ : [];
100
+ return cleaned.length > 0 ? cleaned : DEFAULT_CRITERIA[target];
101
+ }
102
+
86
103
  /** Auto-detect the most recent code/reasoning/plan from session messages. */
87
104
  function autoDetectTarget(
88
105
  ctx: ExtensionContext,
@@ -325,16 +342,31 @@ export function createVerifyTool(): ToolDefinition<any, any, any> {
325
342
  const startMs = Date.now();
326
343
  const target = p.target ?? "result";
327
344
  const depth = p.depth ?? "standard";
328
- const criteria = p.criteria ?? DEFAULT_CRITERIA[target];
329
- const confidenceThreshold = p.confidenceThreshold ?? 0.5;
330
-
331
- const text = p.targetText ?? autoDetectTarget(ctx)?.text;
345
+ const criteria = resolveCriteria(p.criteria, target);
346
+ // A blank selector arrives as 0, and a value outside (0,1] can never
347
+ // gate a 0-1 confidence: both are "not provided", and the documented
348
+ // default is 0.5. The default is documented, so this substitutes -
349
+ // it never refuses the call.
350
+ const rawThreshold = p.confidenceThreshold;
351
+ const confidenceThreshold =
352
+ typeof rawThreshold === "number" &&
353
+ Number.isFinite(rawThreshold) &&
354
+ rawThreshold > 0 &&
355
+ rawThreshold <= 1
356
+ ? rawThreshold
357
+ : 0.5;
358
+
359
+ // A blank targetText is "not provided" too: the documented path
360
+ // auto-detects from recent context. Only when that finds nothing does
361
+ // the tool refuse, and the refusal names auto-detection as what ran.
362
+ const providedText = optionalString(p.targetText);
363
+ const text = providedText ?? autoDetectTarget(ctx)?.text;
332
364
  if (!text) {
333
365
  return {
334
366
  content: [
335
367
  {
336
368
  type: "text",
337
- text: "[verify] Could not auto-detect target text from recent context. Provide targetText explicitly.",
369
+ text: "[verify] Auto-detection ran over recent context and found no target text. Provide targetText explicitly.",
338
370
  },
339
371
  ],
340
372
  details: { passed: false, reason: "no-text" },
@@ -431,6 +463,8 @@ export function createVerifyTool(): ToolDefinition<any, any, any> {
431
463
  suggestions: result.suggestions,
432
464
  target: result.target,
433
465
  checksRun: result.checksRun,
466
+ confidenceThreshold,
467
+ criteria,
434
468
  verificationMs: result.verificationMs,
435
469
  blockedFinalize: !passed,
436
470
  },
@@ -30,7 +30,7 @@ const inspectSchema = Type.Object({
30
30
  truncate: Type.Optional(
31
31
  Type.Number({
32
32
  description:
33
- "Max characters (0 = no limit, default 5000); applied after tool/skill exclusion.",
33
+ "Max characters before truncation (default 5000); applied after tool/skill exclusion. A blank or non-positive value applies the default.",
34
34
  default: 5000,
35
35
  }),
36
36
  ),
@@ -122,7 +122,14 @@ export function createPromptInspectTool() {
122
122
  ): Promise<any> => {
123
123
  const showTools = params.includeTools ?? true;
124
124
  const showSkills = params.includeSkills ?? true;
125
- const maxLen = params.truncate ?? 5000;
125
+ // Argument validation converts a blank truncate to 0, and the
126
+ // description used to define 0 as "no limit" - so a blank flipped the
127
+ // tool into dumping the whole assembled prompt. A deliberate 0 is
128
+ // indistinguishable from a blank after conversion, so the documented
129
+ // default wins and the old no-limit mode is no longer reachable through
130
+ // this selector (use `match` to read the full prompt).
131
+ const maxLen =
132
+ typeof params.truncate === "number" && params.truncate > 0 ? params.truncate : 5000;
126
133
 
127
134
  if (!ctx) {
128
135
  return {
@@ -202,7 +209,7 @@ export function createPromptInspectTool() {
202
209
  }
203
210
  } else {
204
211
  // Apply truncation only when no match filter is active
205
- if (maxLen > 0 && body.length > maxLen) {
212
+ if (body.length > maxLen) {
206
213
  didTruncate = true;
207
214
  body =
208
215
  body.slice(0, maxLen - 80) +
@@ -8,25 +8,25 @@ the wake-up policy.
8
8
 
9
9
  ## Files at this level
10
10
 
11
- | File | Purpose |
12
- | ------------------------------------ | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ |
13
- | `state.ts` | Composed session-scoped state. Re-exports `ENFORCEMENT_PARAMS` and `EnforcementState` from `enforcement-types.ts`. |
14
- | `enforcement-types.ts` | The shared `ENFORCEMENT_PARAMS` and `EnforcementState` definitions. **Lives outside `state.ts` to break the circular dep** that would otherwise form between `state.ts <-> session-state.ts <-> verification-judge-state.ts`. |
15
- | `session-state.ts` | Per-turn tracking, config, topic-drift detector. |
16
- | `whisper.ts` | The `sendTrackedMessage` wrapper with the whisper cap and metadata. |
17
- | `judge-utils.ts` | Helpers for the verification judge. |
18
- | `signals.ts` | The signal taxonomy (severity, source, priority). |
19
- | `validation.ts` | Validation of whisper payloads. |
20
- | `similarity.ts` | Cosine similarity for dedup. |
21
- | `nudge-manager.ts` | Manages nudge cooldowns and dedup. |
22
- | `verification-judge.ts` | The verification judge - runs in background. |
23
- | `self-state-claims.ts` | Deterministic self-state claim detector for the judge: scans the visible response for liveness/delivery/progress claims, suppresses them when the evidence window shows a probe call (session_health, health_maintenance, v2raytun_status, telegram_send); renders the judge's SELF-STATE CLAIMS section only when candidates survive (prompt byte-unchanged otherwise). |
24
- | `task-evidence.ts` | Task-wide evidence lines (since last real user message) rebuilt from the pi session store for the judge prompt; survives turn breaks (wake-retry redelivery). Logs over the cap are compacted into LLM chunk summaries (newest chunk raw), with drop-oldest truncation as the fallback. |
25
- | `verification-judge-state.ts` | Judge/calibration helpers. |
26
- | `handler-registry.ts` | Thin orchestrator that wires all handlers. |
27
- | `register-tool-result-formatting.ts` | MI-3 Structured Output Catalyst - wraps read/grep/find/ls outputs in structured header format (opt-in). |
28
- | `event-types.ts` | Types for events that flow through the enforcement layer. |
29
- | `*-state.ts` | Sub-state files (tool-tracking, error-tracking, memory-hygiene, context-hygiene, scratchpad, thought-chain). Each owns its slice of the composed state. |
11
+ | File | Purpose |
12
+ | ------------------------------------ | --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
13
+ | `state.ts` | Composed session-scoped state. Re-exports `ENFORCEMENT_PARAMS` and `EnforcementState` from `enforcement-types.ts`. |
14
+ | `enforcement-types.ts` | The shared `ENFORCEMENT_PARAMS` and `EnforcementState` definitions. **Lives outside `state.ts` to break the circular dep** that would otherwise form between `state.ts <-> session-state.ts <-> verification-judge-state.ts`. |
15
+ | `session-state.ts` | Per-turn tracking, config, topic-drift detector. |
16
+ | `whisper.ts` | The `sendTrackedMessage` wrapper with the whisper cap and metadata. |
17
+ | `judge-utils.ts` | Helpers for the verification judge. |
18
+ | `signals.ts` | The signal taxonomy (severity, source, priority). |
19
+ | `validation.ts` | Validation of whisper payloads. |
20
+ | `similarity.ts` | Cosine similarity for dedup. |
21
+ | `nudge-manager.ts` | Manages nudge cooldowns and dedup. |
22
+ | `verification-judge.ts` | The verification judge - runs in background. |
23
+ | `self-state-claims.ts` | Deterministic self-state claim detector for the judge: scans the visible response for liveness/delivery/progress claims, suppresses them when the evidence window shows a probe call (session_health, health_maintenance, tunnel_status, v2raytun_status, telegram_send); renders the judge's SELF-STATE CLAIMS section only when candidates survive (prompt byte-unchanged otherwise). |
24
+ | `task-evidence.ts` | Task-wide evidence lines (since last real user message) rebuilt from the pi session store for the judge prompt; survives turn breaks (wake-retry redelivery). Logs over the cap are compacted into LLM chunk summaries (newest chunk raw), with drop-oldest truncation as the fallback. |
25
+ | `verification-judge-state.ts` | Judge/calibration helpers. |
26
+ | `handler-registry.ts` | Thin orchestrator that wires all handlers. |
27
+ | `register-tool-result-formatting.ts` | Structured Output Catalyst - wraps read/grep/find/ls outputs in structured header format (opt-in). |
28
+ | `event-types.ts` | Types for events that flow through the enforcement layer. |
29
+ | `*-state.ts` | Sub-state files (tool-tracking, error-tracking, memory-hygiene, context-hygiene, scratchpad, thought-chain). Each owns its slice of the composed state. |
30
30
 
31
31
  ## Sub-directories
32
32
 
@@ -67,3 +67,9 @@ allow sub-state modules to be tested in isolation.
67
67
  for the agent's natural flow.
68
68
 
69
69
  See individual handler files in `handlers/` for the per-source policy.
70
+
71
+ ## Over-correction gate setting
72
+
73
+ | Setting | Default | Effect | Read cadence |
74
+ | ------------------------------------------- | ------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ---------------------------------------------- |
75
+ | `braid.enforcement.overCorrectionGate.mode` | `"off"` | `off` leaves the gate inert; `shadow` logs and whispers without blocking; `on` blocks destructive `bash` and `write` calls while a recent accusation or hot acceptance signal is active. | Re-read at use time on each checked tool call. |
@@ -1,5 +1,5 @@
1
1
  /**
2
- * register-tool-result-formatting — Structured Output Catalyst (MI-3)
2
+ * register-tool-result-formatting — Structured Output Catalyst
3
3
  *
4
4
  * Wraps tool outputs for read/grep/find/ls in a consistent structured format
5
5
  * to improve transformer processing. Uses the tool_result event's middleware
@@ -193,7 +193,7 @@ export async function dispatchTurnEndLoop(
193
193
  const memoryUpdated = actualTools.some((t: string) => ["memory_store", "memory"].includes(t));
194
194
 
195
195
  // Truth-seeking trigger: verify_claim/conflict_qa usage
196
- // MI-2: When the agent uses verification tools, briefly surface TRUTH-SEEKING
196
+ // When the agent uses verification tools, briefly surface TRUTH-SEEKING
197
197
  // mode explicitly, confirming the correct epistemic stance.
198
198
  const truthTools = actualTools.filter((t: string) => ["verify_claim", "conflict_qa"].includes(t));
199
199
  if (truthTools.length > 0) {
@@ -29,6 +29,7 @@
29
29
  export const SELF_STATE_PROBE_TOOLS = [
30
30
  "session_health",
31
31
  "health_maintenance",
32
+ "tunnel_status",
32
33
  "v2raytun_status",
33
34
  "telegram_send",
34
35
  ] as const;
@@ -109,7 +110,7 @@ const NEGATION_HEDGE =
109
110
 
110
111
  /** Probe-tool names that, named in the claim clause itself, exempt it: citing the check ("I ran session_health: healthy") is probe awareness, not an ungrounded assertion. */
111
112
  const PROBE_NAME =
112
- /(?:session_health|health_maintenance|v2raytun_status|telegram_send|telegram_poller)/i;
113
+ /(?:session_health|health_maintenance|tunnel_status|v2raytun_status|telegram_send|telegram_poller)/i;
113
114
 
114
115
  /** Scan input: a subset of JudgeContext fields. */
115
116
  export interface SelfStateScanInput {
@@ -197,7 +198,7 @@ export function renderSelfStateClaimsSection(claims: SelfStateClaim[]): string {
197
198
  }
198
199
  lines.push(
199
200
  "",
200
- "These claims cannot be verified by build or test evidence: they need a decisive in-session probe call (session_health, health_maintenance, v2raytun_status, telegram_send), and the evidence window above shows none.",
201
+ "These claims cannot be verified by build or test evidence: they need a decisive in-session probe call (session_health, health_maintenance, tunnel_status, v2raytun_status, telegram_send), and the evidence window above shows none.",
201
202
  'Rule: flag each listed claim as an issue (category "claim") unless TASK EVIDENCE, TOOL EXECUTION LOG, or chunk summaries above show a probe call supporting it. Hedged or negative self-state ("poller not running", "reachability unverified") is honest and must NOT be flagged.',
202
203
  );
203
204
  return lines.join("\n");
@@ -73,9 +73,12 @@ per-source caps are enforced in `injection-tracker.ts`.
73
73
 
74
74
  - **Process-global**: state is per-session (via `getExtensionState(ctx, key, factory)`).
75
75
  The poller does not run when no session is active (gated by `SessionPhase`).
76
- - **Opt-in**: enable via `BRAID_MODULES=+heartbeat`. Off by default in
77
- core (timer) to avoid surprising users; on by default in
78
- `all,-heartbeat-scheduler` config.
76
+ - **Default-on**: the manifest entry carries no `disabledByDefault`, so the
77
+ scheduler loads like every other default extension. Turn it off per
78
+ extension (`braid.modules["heartbeat-scheduler"]: false`, or
79
+ `/modules-disable heartbeat-scheduler`) or by its module group
80
+ (`/modules-disable heartbeat`, which also covers `analytics`); the
81
+ in-process toggle changes the current session's set without a restart.
79
82
  - **First-tick timing**: the first heartbeat fires 5 turns after
80
83
  session_start, not on session_start itself. This avoids racing
81
84
  with the user's first message.
@@ -159,7 +159,7 @@ export const MANIFEST: ManifestEntry[] = [
159
159
  id: "session-recall",
160
160
  fn: sessionRecallFn,
161
161
  dependencies: ["three-memory-stores"],
162
- desc: "Session recall tools (session_search, session_query)",
162
+ desc: "Session recall tool (session)",
163
163
  module: "memory",
164
164
  },
165
165
  {
@@ -259,7 +259,7 @@ export const MANIFEST: ManifestEntry[] = [
259
259
  fn: errorCaptureFn,
260
260
  disabledByDefault: true,
261
261
  dependencies: ["agent-environment"],
262
- desc: "Error logging tools: error_list, error_get",
262
+ desc: "Error logging tool (error)",
263
263
  },
264
264
  ];
265
265
 
@@ -47,13 +47,21 @@ The `cancelTimersByTag` helper from `extensions/scheduling/timer/index.ts` is th
47
47
  ## Parallel-tool-call constraint
48
48
 
49
49
  `context_compress` returns its result immediately and defers the actual
50
- `ctx.compact()` to the agent-run settle point (batch-settle scheduler:
51
- armed on the compress tool's action, fired after `agent_settled` once the
52
- batch's tool-result messages have landed - see
53
- `context-manager/handlers/tools/context-compress.ts`). This removes the old
50
+ `ctx.compact()` to the next safe boundary (batch-settle scheduler in
51
+ `context-manager/handlers/tools/context-compress.ts`): the first `turn_end`
52
+ whose persisted `toolResults` account for every active tool call, or
53
+ `agent_settled`, whichever comes first. Active calls are tracked by
54
+ `toolCallId`, registered at `tool_execution_start` and released by its own
55
+ `tool_execution_end` or by a boundary event whose `toolResults` carry that
56
+ id; `agent_settled` clears whatever is left. The earlier in-flight COUNT was
57
+ decremented only by the exact end event it waited for, so one missed
58
+ `tool_execution_end` left it positive and every later boundary returned
59
+ before the armed compaction could drain; identity cannot drift that way, and
60
+ an id with no result stays protective. This removes the old
54
61
  mid-batch abort race: a still-executing parallel sibling's tool result is
55
62
  delivered BEFORE the aborting compact fires and is no longer discarded
56
- (pinned by `test/runtime/memory/test-compress-batch-settle.mjs`). The
63
+ (pinned by `test/runtime/memory/test-compress-batch-settle.mjs` and
64
+ `test/runtime/memory/test-compress-drain-boundaries.mjs`). The
57
65
  degraded path (batch-settle bookkeeping unavailable) uses a bounded
58
66
  `DEGRADED_DEFERRAL_MS` (150ms) deferral - never a raw `setTimeout(0)`.
59
67
 
@@ -19,6 +19,16 @@
19
19
  * at consumption (the armer may have read the branch before the batch's later
20
20
  * results landed).
21
21
  *
22
+ * A supplied index can land on a non-message entry (`label` from
23
+ * `context_navigate action=tag`, `custom`/`custom_message`, a model or
24
+ * thinking-level change, a context edit, a compaction): those entries carry no
25
+ * provider message, so the kept slice's opening message is the first `message`
26
+ * entry at or after the index. The leading-result check resolves that entry
27
+ * before judging the slice; judging the index itself let a tag placed between
28
+ * an assistant's tool call and its result pass, and the kept slice then opened
29
+ * with a bare tool result ("No tool call found for function call output with
30
+ * call_id ..." from Responses-style endpoints).
31
+ *
22
32
  * The readers take raw session entries (`readonly unknown[]`) because that is
23
33
  * what a live branch is at the boundaries; the shape is narrowed here rather
24
34
  * than asserted at every call site.
@@ -83,11 +93,25 @@ function toolCallIdsOf(entry: BoundaryEntry | undefined): string[] {
83
93
  return ids;
84
94
  }
85
95
 
96
+ /**
97
+ * Index of the first message entry at or after `fromIndex`; non-message
98
+ * session entries carry no provider message and are stepped over. Equals
99
+ * `entries.length` when no message remains.
100
+ */
101
+ function firstMessageIndex(entries: readonly unknown[], fromIndex: number): number {
102
+ let i = fromIndex;
103
+ while (i < entries.length && entryOf(entries[i])?.type !== "message") i++;
104
+ return i;
105
+ }
106
+
86
107
  /**
87
108
  * The structurally invalid shape of a kept slice [candidateIndex..end]:
88
- * either it begins with a tool result (its call fell outside the slice), or it
89
- * contains an assistant whose tool_calls are not fully answered inside it.
90
- * Returns null for a valid slice. Exported for tests and audit surfaces.
109
+ * either its first message is a tool result (its call fell outside the slice),
110
+ * or it contains an assistant whose tool_calls are not fully answered inside
111
+ * it. Non-message entries between the candidate and the first message are
112
+ * stepped over: they are not provider messages, so they neither open the
113
+ * slice nor hide a bare result behind them. Returns null for a valid slice.
114
+ * Exported for tests and audit surfaces.
91
115
  */
92
116
  export function findBoundaryViolation(
93
117
  entries: readonly unknown[],
@@ -95,16 +119,23 @@ export function findBoundaryViolation(
95
119
  ): {
96
120
  kind: "leading-tool-result" | "unanswered-tool-calls";
97
121
  entryId?: string;
122
+ /** Index of the offending tool-result entry (leading-tool-result). */
123
+ resultIndex?: number;
98
124
  /** Index of the offending assistant entry (unanswered-tool-calls). */
99
125
  assistantIndex?: number;
100
126
  } | null {
101
127
  if (!Array.isArray(entries) || candidateIndex < 0 || candidateIndex >= entries.length) {
102
128
  return null;
103
129
  }
104
- if (isToolResultEntry(entryOf(entries[candidateIndex]))) {
105
- return { kind: "leading-tool-result", entryId: entryOf(entries[candidateIndex])?.id };
130
+ const firstMessage = firstMessageIndex(entries, candidateIndex);
131
+ if (isToolResultEntry(entryOf(entries[firstMessage]))) {
132
+ return {
133
+ kind: "leading-tool-result",
134
+ entryId: entryOf(entries[firstMessage])?.id,
135
+ resultIndex: firstMessage,
136
+ };
106
137
  }
107
- for (let i = candidateIndex; i < entries.length; i++) {
138
+ for (let i = firstMessage; i < entries.length; i++) {
108
139
  for (const id of toolCallIdsOf(entryOf(entries[i]))) {
109
140
  let answered = false;
110
141
  for (let j = i + 1; j < entries.length; j++) {
@@ -160,8 +191,10 @@ function firstNonToolResultIndex(entries: readonly unknown[], fromIndex: number)
160
191
  /**
161
192
  * Snap a candidate first-kept index so the kept slice [result..end] cannot
162
193
  * split a tool batch:
163
- * - a candidate landing mid-batch snaps FORWARD to the first entry after
164
- * the batch's last tool result (parallel calls are one batch of N results);
194
+ * - a candidate landing mid-batch - including on a non-message entry whose
195
+ * first following message is mid-batch - snaps FORWARD to the first entry
196
+ * after the batch's last tool result (parallel calls are one batch of N
197
+ * results);
165
198
  * - when that would keep less than `MIN_KEPT_ENTRIES`, it snaps BACKWARD to
166
199
  * the batch's assistant message instead, keeping the whole batch live;
167
200
  * - a leading result whose call is no longer in the branch (summarized away)
@@ -182,12 +215,13 @@ export function snapBoundaryOutsideToolBatch(
182
215
  if (!violation) return index;
183
216
 
184
217
  if (violation.kind === "leading-tool-result") {
185
- const owner = findOwningAssistant(entries, index);
218
+ const resultIndex = violation.resultIndex ?? index;
219
+ const owner = findOwningAssistant(entries, resultIndex);
186
220
  if (owner < 0) {
187
221
  // The call is outside the branch: the only valid repair is dropping
188
222
  // the orphaned results.
189
- const after = firstNonToolResultIndex(entries, index);
190
- if (after > index && entries.length - after >= MIN_KEPT_ENTRIES) {
223
+ const after = firstNonToolResultIndex(entries, resultIndex);
224
+ if (after > resultIndex && entries.length - after >= MIN_KEPT_ENTRIES) {
191
225
  index = after;
192
226
  continue;
193
227
  }