pi-braid 0.40.1 → 0.40.3
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/CONTRIBUTING.md +27 -3
- package/README.md +13 -13
- package/TOOLS.md +27 -3
- package/build-info.json +9 -0
- package/dist/extensions/cognitive/metacognition/handlers/claim-verification.js +7 -1
- package/dist/extensions/cognitive/metacognition/handlers/finalize.js +20 -0
- package/dist/extensions/cognitive/metacognition/handlers/hypothesize.d.ts +9 -0
- package/dist/extensions/cognitive/metacognition/handlers/hypothesize.js +13 -1
- package/dist/extensions/cognitive/metacognition/handlers/verify.js +35 -4
- package/dist/extensions/cognitive/topic-drift/index.js +19 -2
- package/dist/extensions/core/prompt-builder/inspect.js +9 -3
- package/dist/extensions/core/prompt-builder/top-level-extensions.d.ts +6 -5
- package/dist/extensions/core/prompt-builder/top-level-extensions.js +64 -29
- package/dist/extensions/enforcement/handlers/register-tool-result-formatting.d.ts +1 -1
- package/dist/extensions/enforcement/handlers/register-tool-result-formatting.js +1 -1
- package/dist/extensions/enforcement/handlers/turn-end-handler.js +1 -1
- package/dist/extensions/enforcement/self-state-claims.d.ts +1 -1
- package/dist/extensions/enforcement/self-state-claims.js +3 -2
- package/dist/extensions/manifest.js +2 -2
- package/dist/extensions/memory/context-manager/handlers/tool-batch-boundary.d.ts +22 -5
- package/dist/extensions/memory/context-manager/handlers/tool-batch-boundary.js +43 -11
- package/dist/extensions/memory/context-manager/handlers/tools/context-compress.js +174 -50
- package/dist/extensions/memory/context-manager/handlers/tools/context-inspect.js +7 -1
- package/dist/extensions/memory/context-manager/handlers/tools/context-navigate.js +4 -1
- package/dist/extensions/memory/memory-agent-bridge/explore-paths.d.ts +2 -24
- package/dist/extensions/memory/memory-agent-bridge/explore-paths.js +33 -2
- package/dist/extensions/memory/memory-agent-bridge/memory-store-tool.js +16 -0
- package/dist/extensions/memory/memory-agent-bridge/memory-tool.js +21 -9
- package/dist/extensions/memory/scratchpad/format.js +8 -2
- package/dist/extensions/memory/scratchpad/index.js +11 -2
- package/dist/extensions/memory/session-recall/index.js +19 -6
- package/dist/extensions/memory/three-memory-stores/episodes.d.ts +1 -1
- package/dist/extensions/memory/three-memory-stores/episodes.js +2 -2
- package/dist/extensions/orchestrator/index.js +73 -8
- package/dist/extensions/scheduling/timer/index.js +9 -1
- package/dist/extensions/self/error-tracking/index.js +8 -1
- package/dist/extensions/self/health-maintenance/handlers/session-health-tool.js +9 -2
- package/dist/extensions/self/health-maintenance/index.js +112 -9
- package/dist/lib/config/braid-settings.d.ts +21 -0
- package/dist/lib/config/braid-settings.js +36 -0
- package/dist/lib/constants/session-thresholds.d.ts +2 -3
- package/dist/lib/constants/session-thresholds.js +2 -3
- package/dist/lib/enforcement/turn-dedup.d.ts +2 -2
- package/dist/lib/enforcement/turn-dedup.js +2 -2
- package/dist/lib/harness/index.d.ts +1 -1
- package/dist/lib/harness/index.js +1 -1
- package/dist/lib/harness/types.d.ts +1 -1
- package/dist/lib/harness/types.js +1 -1
- package/dist/lib/orchestrator/index.d.ts +2 -2
- package/dist/lib/orchestrator/index.js +1 -1
- package/dist/lib/orchestrator/workers.d.ts +184 -2
- package/dist/lib/orchestrator/workers.js +521 -6
- package/dist/lib/utils/injection-tracker.js +1 -1
- package/docs/CONTEXT-POLICY.md +13 -5
- package/docs/LOCAL_DEV.md +31 -5
- package/docs/ROADMAP.md +10 -7
- package/docs/SETUP.md +3 -2
- package/package.json +11 -8
- package/skills/braid-reference/SKILL.md +23 -23
- package/skills/release-pipeline/SKILL.md +46 -37
- package/skills/subagent-orchestration/SKILL.md +51 -1
- package/src/extensions/cognitive/metacognition/handlers/claim-verification.ts +8 -1
- package/src/extensions/cognitive/metacognition/handlers/finalize.ts +21 -0
- package/src/extensions/cognitive/metacognition/handlers/hypothesize.ts +14 -1
- package/src/extensions/cognitive/metacognition/handlers/verify.ts +39 -5
- package/src/extensions/cognitive/topic-drift/index.ts +20 -2
- package/src/extensions/core/prompt-builder/AGENTS.md +1 -1
- package/src/extensions/core/prompt-builder/inspect.ts +10 -3
- package/src/extensions/core/prompt-builder/top-level-extensions.ts +68 -29
- package/src/extensions/enforcement/AGENTS.md +25 -19
- package/src/extensions/enforcement/handlers/register-tool-result-formatting.ts +1 -1
- package/src/extensions/enforcement/handlers/turn-end-handler.ts +1 -1
- package/src/extensions/enforcement/self-state-claims.ts +3 -2
- package/src/extensions/heartbeat/heartbeat-scheduler/AGENTS.md +6 -3
- package/src/extensions/manifest.ts +2 -2
- package/src/extensions/memory/AGENTS.md +13 -5
- package/src/extensions/memory/context-manager/handlers/tool-batch-boundary.ts +45 -11
- package/src/extensions/memory/context-manager/handlers/tools/context-compress.ts +202 -60
- package/src/extensions/memory/context-manager/handlers/tools/context-inspect.ts +8 -1
- package/src/extensions/memory/context-manager/handlers/tools/context-navigate.ts +4 -1
- package/src/extensions/memory/memory-agent-bridge/explore-paths.ts +37 -3
- package/src/extensions/memory/memory-agent-bridge/memory-store-tool.ts +16 -0
- package/src/extensions/memory/memory-agent-bridge/memory-tool.ts +22 -9
- package/src/extensions/memory/scratchpad/format.ts +8 -2
- package/src/extensions/memory/scratchpad/index.ts +12 -2
- package/src/extensions/memory/session-recall/index.ts +20 -8
- package/src/extensions/memory/three-memory-stores/episodes.ts +2 -2
- package/src/extensions/orchestrator/index.ts +100 -11
- package/src/extensions/scheduling/timer/AGENTS.md +7 -0
- package/src/extensions/scheduling/timer/index.ts +10 -1
- package/src/extensions/self/error-tracking/index.ts +9 -1
- package/src/extensions/self/health-maintenance/handlers/session-health-tool.ts +10 -2
- package/src/extensions/self/health-maintenance/index.ts +143 -14
- package/src/extensions/self/watchdog/AGENTS.md +1 -0
- package/src/lib/AGENTS.md +21 -21
- package/src/lib/config/braid-settings.ts +74 -0
- package/src/lib/constants/session-thresholds.ts +2 -3
- package/src/lib/enforcement/turn-dedup.ts +2 -2
- package/src/lib/harness/AGENTS.md +1 -1
- package/src/lib/harness/index.ts +1 -1
- package/src/lib/harness/types.ts +1 -1
- package/src/lib/orchestrator/index.ts +26 -0
- package/src/lib/orchestrator/workers.ts +705 -7
- package/src/lib/utils/AGENTS.md +1 -1
- package/src/lib/utils/injection-tracker.ts +1 -1
package/docs/ROADMAP.md
CHANGED
|
@@ -51,16 +51,19 @@ provenance; they are no longer work.
|
|
|
51
51
|
2. **Version-drift cleanup** - RESOLVED: the four workspace packages declare
|
|
52
52
|
the pi peer as `"*"`; the root manifest pins the installed version
|
|
53
53
|
(`package.json`, the `@earendil-works/*` deps).
|
|
54
|
-
3. **Tag-scheme unification** - RESOLVED: two tag roles
|
|
55
|
-
the root `CHANGELOG.md` - iteration tags `v0.40.0-main.N
|
|
56
|
-
`0.40.0-main.N
|
|
54
|
+
3. **Tag-scheme unification** - RESOLVED: two tag roles, one namespace per package, recorded in
|
|
55
|
+
the root `CHANGELOG.md` - core keeps the bare `v` (iteration tags `v0.40.0-main.N`,
|
|
56
|
+
`package.json` `0.40.0-main.N`; release tag `v0.40.0`), every sibling tags as
|
|
57
|
+
`<workspace>-v<version>` (`telegram-v0.1.4`).
|
|
57
58
|
4. **CI private-data gate** - RESOLVED: `check-private-data.mjs` is a CI step
|
|
58
59
|
and scans the whole tree by default (tracked plus untracked); the
|
|
59
60
|
staged-file checks are a local pre-commit convenience.
|
|
60
|
-
5. **Skills pack** - finish validating the
|
|
61
|
+
5. **Skills pack** - finish validating the five skills still marked
|
|
61
62
|
`verdict: UNVALIDATED` (`braid-reference`, `misdirection-audit`,
|
|
62
|
-
`placement-discipline`, `probe-evaluation`, `
|
|
63
|
-
|
|
63
|
+
`placement-discipline`, `probe-evaluation`, `voice-replies`); keep the pack
|
|
64
|
+
catalog fresh as decision rows land. The machine-local `tunnel-control`
|
|
65
|
+
skill is `KEEP` with its v2RayTun arm validated; its Happ and Incy arms
|
|
66
|
+
stay unvalidated until a host installs those clients.
|
|
64
67
|
6. **Extension surface** - ongoing de-bloat of tool/extension counts;
|
|
65
68
|
subagent-RPC hardening. Worker-stats reporting shipped (`list-workers`
|
|
66
69
|
`includeStats`); the orchestrator's metrics-emission half (writing
|
|
@@ -79,7 +82,7 @@ knowledge base.
|
|
|
79
82
|
8. **Small-correctness batch** - RESOLVED: the `prompt_inspect` exclusion is
|
|
80
83
|
pinned by `test-prompt-inspect-exclusion.mjs`; the watchdog tunnel-counter
|
|
81
84
|
canary is a getMe-based liveness probe
|
|
82
|
-
(`telegram/src/lib/inbound-poller.ts`); the
|
|
85
|
+
(`telegram/src/lib/inbound-poller.ts`); the tunnel status description is
|
|
83
86
|
complete in the shipped skill. Effort S.
|
|
84
87
|
9. **Cross-provider fallback on quota death** - RESOLVED: the route-order
|
|
85
88
|
engine walks to the next route, with bounded turn-level fail-back
|
package/docs/SETUP.md
CHANGED
|
@@ -303,7 +303,7 @@ The user-global file always exists; the project one is optional and belongs to p
|
|
|
303
303
|
| `braid.modules.<ref>` | boolean | per manifest | Force-enable (`true`) or disable (`false`) an extension id or module group (`memory`, `heartbeat`); a persisted `false` removes a default-enabled extension from the boot set (see [§ Module toggling](#module-toggling)). |
|
|
304
304
|
| `braid.suppressNudges` | boolean | `false` | Silence all cognitive nudges (enforcement whispers, heartbeat suggestions, feedback reminders). |
|
|
305
305
|
| `braid.timezone` | IANA name | `Europe/Moscow` | Time zone for all braid-facing time rendering. |
|
|
306
|
-
| `braid.enforcement.nudgeConfigs.<type>` | object | built-in per-type defaults | Per-nudge-type overrides with fields `baseCooldownMs` (number), `escalateAfter` (number), `deduplicate` (boolean). Types: `validation`, `recordFailure`, `malformed`, `memory`, `scratchpad`, `context`, `sessionSignals`. Invalid entries are logged and dropped.
|
|
306
|
+
| `braid.enforcement.nudgeConfigs.<type>` | object | built-in per-type defaults | Per-nudge-type overrides with leaf shape `braid.enforcement.nudgeConfigs.<type>.<key>`; accepted fields are `baseCooldownMs` (number), `escalateAfter` (number), and `deduplicate` (boolean). Types: `validation`, `recordFailure`, `malformed`, `memory`, `scratchpad`, `context`, `sessionSignals`. Invalid entries are logged and dropped. |
|
|
307
307
|
| `braid.memory.writeGate` | object | gate ON, 20-min cooldown, `enforce` | Review-before-write gate on episode writes; `mode: "warn"` logs decisions and writes anyway; `sites` opts individual writers out. |
|
|
308
308
|
| `braid.timers.maxWakeAttempts` | integer 1..10 | `3` | Redelivery cap for a failed timer wake before it is dropped with a visible notice. |
|
|
309
309
|
| `braid.timers.wakeAlert` | object | ON, `outcomeTimeouts: 2` | Dead-wake page: repeated outcome timeouts / exhausted redeliveries feed the watchdog's wake-delivery condition, which pages the owner out of band (telegram) once per episode plus one recovery line. `mode: "off"` disables the feed (delivery itself is unchanged); `outcomeTimeouts` is an integer 2..10 (1 would page on every normal long turn). |
|
|
@@ -315,6 +315,7 @@ The user-global file always exists; the project one is optional and belongs to p
|
|
|
315
315
|
| `braid.telegram.notifyMaxChars` | positive number | `1200` | Craft target and guard for turn-end notifications. |
|
|
316
316
|
| `braid.telegram.pauseStateFile` | path string | built-in default | Override for the `/telegram` pause-state file path. |
|
|
317
317
|
| `braid.contrib.<name>` | boolean or object | enabled / per-extension | Per-contrib-extension config - see [§ Contrib extensions](#contrib-extensions). |
|
|
318
|
+
| `braid.contrib.<extensionName>.enabled` | string[] | unset | Generic per-extension tool-enable list. The shared helper reads and writes string entries; the `/tools` surface stores its shared list in `braid.contrib._global.enabled`. |
|
|
318
319
|
|
|
319
320
|
### Worked example
|
|
320
321
|
|
|
@@ -348,7 +349,7 @@ The 12 contrib extension names: `browser`, `codebase`, `diff`, `healthcheck`, `n
|
|
|
348
349
|
- `braid.contrib.browserAdvanced: true` - the 7 advanced browser tools (opt-in).
|
|
349
350
|
- `braid.contrib.pathSecurity.block: true` - hard-block dangerous operations instead of soft-warn.
|
|
350
351
|
- `braid.contrib.observatory.record: true` - record session lifecycle events.
|
|
351
|
-
- `braid.contrib.networkMonitor` - `{ profileName, reachabilityUrl, curlTimeoutSecs, disconnectDelayMs, connectTimeoutMs, pollIntervalMs, cooldownMs, autoReconnect, autoWatch, probeIntervalMs, maxAttemptsPerOutage, reconnectCooldownMs }`.
|
|
352
|
+
- `braid.contrib.networkMonitor` - `{ client, profileName, reachabilityUrl, curlTimeoutSecs, disconnectDelayMs, connectTimeoutMs, pollIntervalMs, cooldownMs, autoReconnect, autoWatch, probeIntervalMs, outageThreshold, maxAttemptsPerOutage, reconnectCooldownMs, logFile }`. The watcher toggles only after `outageThreshold` consecutive failed probes (default 3, probes spaced by `probeIntervalMs`); the toggle mechanism comes from the client adapter (app deep links, or the NE profile via `scutil --nc stop|start`), and settings are re-read while the extension runs.
|
|
352
353
|
- `braid.contrib["rate-limiter"]` - `{ rpm, tpm, providers: { <provider>: { rpm, tpm } } }` ceilings; note the hyphenated key (the bundle toggle name for the extension itself is `rateLimiter`).
|
|
353
354
|
- `braid.contrib._global.enabled: [...]` - global tool list managed by the tool-config extension.
|
|
354
355
|
|
package/package.json
CHANGED
|
@@ -1,6 +1,6 @@
|
|
|
1
1
|
{
|
|
2
2
|
"name": "pi-braid",
|
|
3
|
-
"version": "0.40.
|
|
3
|
+
"version": "0.40.3",
|
|
4
4
|
"description": "Metacognitive extension suite for pi-coding-agent: enforcement, memory, reasoning, self-improvement, and structured cognitive signals.",
|
|
5
5
|
"type": "module",
|
|
6
6
|
"main": "./dist/index.js",
|
|
@@ -19,7 +19,8 @@
|
|
|
19
19
|
"LICENSE",
|
|
20
20
|
"AGENTS.md",
|
|
21
21
|
"CONTRIBUTING.md",
|
|
22
|
-
"TOOLS.md"
|
|
22
|
+
"TOOLS.md",
|
|
23
|
+
"build-info.json"
|
|
23
24
|
],
|
|
24
25
|
"repository": {
|
|
25
26
|
"type": "git",
|
|
@@ -56,11 +57,13 @@
|
|
|
56
57
|
"check:instruction-strings": "node scripts/check-instruction-strings.mjs",
|
|
57
58
|
"check:doc-index": "node scripts/check-doc-index.mjs",
|
|
58
59
|
"check:schema": "node scripts/check-braid-schema.mjs",
|
|
60
|
+
"check:settings-docs": "node scripts/check-settings-doc-coverage.mjs",
|
|
59
61
|
"ci:check": "npm run check:docs && npm run check:doc-drift && npm run check:instruction-strings && npm run check:doc-index && npm run check:deps && npm run check:labels && npx tsc --noEmit && npm test",
|
|
60
62
|
"release:local": "node scripts/publish-local.mjs",
|
|
61
63
|
"publish:local": "node scripts/publish-local.mjs",
|
|
62
64
|
"dev:link": "npm link && cd ~/.pi/agent/local-packages && npm link pi-braid",
|
|
63
|
-
"prepack": "npm run build"
|
|
65
|
+
"prepack": "npm run build && node scripts/pack-provenance.mjs write",
|
|
66
|
+
"postpack": "node scripts/pack-provenance.mjs remove"
|
|
64
67
|
},
|
|
65
68
|
"devDependencies": {
|
|
66
69
|
"@eslint/js": "10.0.1",
|
|
@@ -73,11 +76,11 @@
|
|
|
73
76
|
"typescript-eslint": "8.59.1"
|
|
74
77
|
},
|
|
75
78
|
"peerDependencies": {
|
|
76
|
-
"@earendil-works/pi-agent-core": "
|
|
77
|
-
"@earendil-works/pi-ai": "
|
|
78
|
-
"@earendil-works/pi-coding-agent": "
|
|
79
|
-
"@earendil-works/pi-tui": "
|
|
80
|
-
"typebox": "
|
|
79
|
+
"@earendil-works/pi-agent-core": "^1.2.0",
|
|
80
|
+
"@earendil-works/pi-ai": "^1.2.0",
|
|
81
|
+
"@earendil-works/pi-coding-agent": "^1.2.0",
|
|
82
|
+
"@earendil-works/pi-tui": "^1.2.0",
|
|
83
|
+
"typebox": "^1.1.38"
|
|
81
84
|
},
|
|
82
85
|
"keywords": [
|
|
83
86
|
"pi-package",
|
|
@@ -1,7 +1,7 @@
|
|
|
1
1
|
---
|
|
2
2
|
name: braid-reference
|
|
3
3
|
description: "Find braid's own documentation: context policy and compaction, timers and wakes, workers and the orchestrator registry, the Telegram channel, prompt and discipline rules, memory stores, the skills catalog, release and install paths, health maintenance, extensions and the polyfill. Use when a question is about braid internals, before editing braid, or when a braid behavior needs its owning doc named instead of guessed."
|
|
4
|
-
validated: 2026-10-
|
|
4
|
+
validated: 2026-10-18
|
|
5
5
|
verdict: KEEP
|
|
6
6
|
---
|
|
7
7
|
|
|
@@ -21,7 +21,7 @@ Braid ships as separate packages, and docs live under each package root. Find th
|
|
|
21
21
|
| `skills` | the on-demand pack: `skills/<name>/SKILL.md`, `README.md`, `AGENTS.md` |
|
|
22
22
|
| `telegram` | `README.md`, `docs/` |
|
|
23
23
|
|
|
24
|
-
This file sits at `<core>/skills/braid-reference/SKILL.md`, so the core root is two levels up. The core skill set ships with the core package; the `skills` package root carries only the on-demand pack; machine-local skills such as `
|
|
24
|
+
This file sits at `<core>/skills/braid-reference/SKILL.md`, so the core root is two levels up. The core skill set ships with the core package; the `skills` package root carries only the on-demand pack; machine-local skills such as `tunnel-control` live in `~/.pi/agent/skills/`.
|
|
25
25
|
|
|
26
26
|
```bash
|
|
27
27
|
# 1. core root — two levels above this SKILL.md
|
|
@@ -40,27 +40,27 @@ In the monorepo the roots are `packages/core`, `packages/skills`, `packages/tele
|
|
|
40
40
|
|
|
41
41
|
Paths below are checkout paths: strip the leading `packages/<name>/` and prepend that package's root. The `Shipped` column says whether the file is in the published tarball (`npm`) or exists only in a checkout (`checkout`).
|
|
42
42
|
|
|
43
|
-
| Topic | Doc
|
|
44
|
-
| --------------------------------------------------------------------------------- |
|
|
45
|
-
| What braid is, install paths, module toggling, storage | `packages/core/README.md`, `packages/core/docs/SETUP.md`
|
|
46
|
-
| First walkthrough: anatomy of an extension, writing one | `packages/core/docs/BEGINNERS.md`
|
|
47
|
-
| Context policy: window, reserve, bands, caps, offload | `packages/core/docs/CONTEXT-POLICY.md`
|
|
48
|
-
| Context map: contexts, owners, seam registry, where state lives | `packages/core/docs/CONTEXT-MAP.md`
|
|
49
|
-
| Context automation: proposals, policy engine, cooldown | `packages/core/src/extensions/memory/context-lifecycle/AGENTS.md`
|
|
50
|
-
| What compaction loses and which guards exist | `packages/core/docs/COMPACTION-LOSS-INVESTIGATION.md`
|
|
51
|
-
| Timers, wakes, wake retry, wake-delivery alerts | `packages/core/src/extensions/scheduling/timer/AGENTS.md`
|
|
52
|
-
| Workers, the `orchestrate` tool, locks, merge verification | `packages/core/skills/subagent-orchestration/SKILL.md`
|
|
53
|
-
| Telegram channel: live update, pause control | `packages/telegram/docs/live-update.md`, `packages/telegram/docs/pause-control.md`
|
|
54
|
-
| Telegram package overview: tools, env vars, components; turn-complete auto-notify | `packages/telegram/README.md`, `packages/telegram/docs/turn-complete-notify.md`
|
|
55
|
-
| System prompt composition, identity, whisper sections | `packages/core/src/extensions/core/prompt-builder/AGENTS.md`
|
|
56
|
-
| Enforcement: whisper cap, dedup, nudges, verification judge | `packages/core/src/extensions/enforcement/AGENTS.md`
|
|
57
|
-
| Memory stores: episodes, semantic, heuristics | `packages/core/src/extensions/memory/three-memory-stores/AGENTS.md`, `packages/core/src/extensions/memory/AGENTS.md`
|
|
58
|
-
| Skills: the core set and the on-demand pack, catalog, adding a skill | `packages/skills/README.md`, `packages/skills/AGENTS.md`
|
|
59
|
-
| Release, publish, install paths, upstream pi updates | `packages/core/skills/release-pipeline/SKILL.md
|
|
60
|
-
| Health maintenance, checkpoints, self-repair | `packages/core/src/extensions/self/health-maintenance/AGENTS.md`
|
|
61
|
-
| Extension loading, manifest, module toggles, the polyfill | `packages/core/AGENTS.md`, `packages/core/src/AGENTS.md`, `packages/core/src/extensions/core/AGENTS.md`
|
|
62
|
-
| Every tool and command braid adds | `packages/core/TOOLS.md`
|
|
63
|
-
| Roadmap (forward-looking) and history | `packages/core/docs/ROADMAP.md`, `packages/core/docs/HISTORY.md`
|
|
43
|
+
| Topic | Doc | Shipped |
|
|
44
|
+
| --------------------------------------------------------------------------------- | ----------------------------------------------------------------------------------------------------------------------- | ------- |
|
|
45
|
+
| What braid is, install paths, module toggling, storage | `packages/core/README.md`, `packages/core/docs/SETUP.md` | npm |
|
|
46
|
+
| First walkthrough: anatomy of an extension, writing one | `packages/core/docs/BEGINNERS.md` | npm |
|
|
47
|
+
| Context policy: window, reserve, bands, caps, offload | `packages/core/docs/CONTEXT-POLICY.md` | npm |
|
|
48
|
+
| Context map: contexts, owners, seam registry, where state lives | `packages/core/docs/CONTEXT-MAP.md` | npm |
|
|
49
|
+
| Context automation: proposals, policy engine, cooldown | `packages/core/src/extensions/memory/context-lifecycle/AGENTS.md` | npm |
|
|
50
|
+
| What compaction loses and which guards exist | `packages/core/docs/COMPACTION-LOSS-INVESTIGATION.md` | npm |
|
|
51
|
+
| Timers, wakes, wake retry, wake-delivery alerts | `packages/core/src/extensions/scheduling/timer/AGENTS.md` | npm |
|
|
52
|
+
| Workers, the `orchestrate` tool, locks, merge verification | `packages/core/skills/subagent-orchestration/SKILL.md` | npm |
|
|
53
|
+
| Telegram channel: live update, pause control | `packages/telegram/docs/live-update.md`, `packages/telegram/docs/pause-control.md` | npm |
|
|
54
|
+
| Telegram package overview: tools, env vars, components; turn-complete auto-notify | `packages/telegram/README.md`, `packages/telegram/docs/turn-complete-notify.md` | npm |
|
|
55
|
+
| System prompt composition, identity, whisper sections | `packages/core/src/extensions/core/prompt-builder/AGENTS.md` | npm |
|
|
56
|
+
| Enforcement: whisper cap, dedup, nudges, verification judge | `packages/core/src/extensions/enforcement/AGENTS.md` | npm |
|
|
57
|
+
| Memory stores: episodes, semantic, heuristics | `packages/core/src/extensions/memory/three-memory-stores/AGENTS.md`, `packages/core/src/extensions/memory/AGENTS.md` | npm |
|
|
58
|
+
| Skills: the core set and the on-demand pack, catalog, adding a skill | `packages/skills/README.md`, `packages/skills/AGENTS.md` | npm |
|
|
59
|
+
| Release, publish, install paths, upstream pi updates | `packages/core/skills/release-pipeline/SKILL.md` (local release and remote npm publication through GitHub Actions OIDC) | npm |
|
|
60
|
+
| Health maintenance, checkpoints, self-repair | `packages/core/src/extensions/self/health-maintenance/AGENTS.md` | npm |
|
|
61
|
+
| Extension loading, manifest, module toggles, the polyfill | `packages/core/AGENTS.md`, `packages/core/src/AGENTS.md`, `packages/core/src/extensions/core/AGENTS.md` | npm |
|
|
62
|
+
| Every tool and command braid adds | `packages/core/TOOLS.md` | npm |
|
|
63
|
+
| Roadmap (forward-looking) and history | `packages/core/docs/ROADMAP.md`, `packages/core/docs/HISTORY.md` | npm |
|
|
64
64
|
|
|
65
65
|
Related skills in the core set: `dox` (the AGENTS.md tree), `placement-discipline` (which layer owns a fact), `release-pipeline` (the local publish flow), `skill-usage` (adoption measurement), `pi-reference` (pi core, the layer braid extends).
|
|
66
66
|
|
|
@@ -1,21 +1,21 @@
|
|
|
1
1
|
---
|
|
2
2
|
name: release-pipeline
|
|
3
|
-
description: "
|
|
3
|
+
description: "Release pipeline for braid packages: local builds, packing, installs, and verification; remote npm publication through GitHub Actions OIDC."
|
|
4
4
|
trigger: "releasing packages, npm publish, local package install, building packages, packing npm packages, package verification"
|
|
5
|
-
validated: 2026-
|
|
5
|
+
validated: 2026-10-18
|
|
6
6
|
verdict: IMPROVE
|
|
7
7
|
---
|
|
8
8
|
|
|
9
9
|
# Release Pipeline
|
|
10
10
|
|
|
11
|
-
|
|
11
|
+
Local package release and install flow, with remote npm publication through GitHub Actions OIDC.
|
|
12
12
|
|
|
13
13
|
## Principles
|
|
14
14
|
|
|
15
15
|
1. **Isolation** - workers operate on git worktrees, never the live repo
|
|
16
16
|
2. **Hygiene** - worktrees and .tgz files are cleaned up after release
|
|
17
17
|
3. **Traceability** - every release has a version, commit, and timestamp
|
|
18
|
-
4. **
|
|
18
|
+
4. **Separation** - local package installs use this pipeline; npm publication runs in a separate GitHub Actions OIDC workflow
|
|
19
19
|
|
|
20
20
|
## Worktree Management
|
|
21
21
|
|
|
@@ -48,12 +48,9 @@ clean apart from a symlinked `node_modules` - and remove any resurrected path (`
|
|
|
48
48
|
|
|
49
49
|
### Step 1: Build
|
|
50
50
|
|
|
51
|
-
|
|
52
|
-
|
|
53
|
-
|
|
54
|
-
```
|
|
55
|
-
|
|
56
|
-
Verify: `npm run typecheck` (or `npx tsc --noEmit`).
|
|
51
|
+
From `<package-dir>`, run `npm run build` when the package defines a build
|
|
52
|
+
script. Run `npm run typecheck` or `npx tsc --noEmit` when the package has a
|
|
53
|
+
TypeScript check.
|
|
57
54
|
|
|
58
55
|
### Step 2: Pack
|
|
59
56
|
|
|
@@ -144,25 +141,30 @@ Rules the script enforces:
|
|
|
144
141
|
`npm --prefix <global-prefix> install -g @earendil-works/pi-coding-agent@<version>`.
|
|
145
142
|
- **Publish, do not repack.** `--apply` composes `publish-local.mjs` per braid
|
|
146
143
|
package; it never packs inline.
|
|
147
|
-
- **Verify
|
|
148
|
-
|
|
149
|
-
|
|
150
|
-
|
|
144
|
+
- **Verify does not start a pi session.** `--verify` reads installed package
|
|
145
|
+
files and settings, evaluates every item, and exits non-zero if any item
|
|
146
|
+
fails. It resolves the current upstream version with `npm view`; live-session
|
|
147
|
+
checks remain a human checklist. It is standalone and refuses
|
|
148
|
+
`--apply`/`--runtime`.
|
|
151
149
|
|
|
152
150
|
## Remote Publication
|
|
153
151
|
|
|
154
|
-
|
|
152
|
+
Remote publication uses `.github/workflows/publish.yml`, separate from the
|
|
153
|
+
local install path. A bare core release tag (`v<core version>`) runs the
|
|
154
|
+
workflow. To publish one sibling package, dispatch it from `main` with
|
|
155
|
+
`only=<package>`; sibling tags do not run this release workflow, and iteration
|
|
156
|
+
tags are refused by its job condition. A release-tag run publishes each stable
|
|
157
|
+
workspace version not already present on npm through trusted publishing (OIDC);
|
|
158
|
+
the workflow uses no npm token.
|
|
155
159
|
|
|
156
|
-
|
|
157
|
-
npm publish
|
|
158
|
-
```
|
|
160
|
+
Preparation:
|
|
159
161
|
|
|
160
|
-
|
|
162
|
+
- Commit the intended stable package versions. Refresh `package-lock.json` as
|
|
163
|
+
a release convention, as described in `CONTRIBUTING.md`.
|
|
164
|
+
- Configure trusted publishing for the package on npmjs.com.
|
|
161
165
|
|
|
162
|
-
|
|
163
|
-
|
|
164
|
-
- CHANGELOG.md updated
|
|
165
|
-
- Git tag created (`git tag v<version>`)
|
|
166
|
+
Local CLI publishing is not a fallback: `npm publish` needs 2FA, and this
|
|
167
|
+
account's passkey second factor leaves the CLI OTP prompt nothing to answer.
|
|
166
168
|
|
|
167
169
|
## Disk Hygiene
|
|
168
170
|
|
|
@@ -175,17 +177,24 @@ After release:
|
|
|
175
177
|
|
|
176
178
|
## Version Convention
|
|
177
179
|
|
|
178
|
-
-
|
|
179
|
-
|
|
180
|
-
|
|
181
|
-
-
|
|
182
|
-
|
|
183
|
-
|
|
184
|
-
|
|
185
|
-
|
|
186
|
-
|
|
187
|
-
|
|
188
|
-
|
|
189
|
-
|
|
190
|
-
|
|
191
|
-
|
|
180
|
+
- `publish-local.mjs` normally derives `<base>-<branch>.<iteration>` from the
|
|
181
|
+
workspace version; if that version already carries a prerelease, it uses it
|
|
182
|
+
verbatim. The script restores the original manifest after packing.
|
|
183
|
+
- Commit stable package versions before remote npm publication. Local snapshots
|
|
184
|
+
use the derived iteration version instead.
|
|
185
|
+
- Use `npm version` for semver bumps in the target workspace.
|
|
186
|
+
- A bare core release tag (`v<core version>`) runs the publish workflow. Publish
|
|
187
|
+
a sibling package by dispatching from `main` with `only=<package>`.
|
|
188
|
+
`tag-prerelease.mjs` creates package-scoped iteration tags; they do not
|
|
189
|
+
publish.
|
|
190
|
+
|
|
191
|
+
## Timer and Artifact Locations
|
|
192
|
+
|
|
193
|
+
Timer files are project-scoped at
|
|
194
|
+
`<braid-home>/projects/<canonical-sanitized-cwd>/pending-timers.jsonl`; the
|
|
195
|
+
default braid home is `~/.pi/agent/braid`, and `BRAID_HOME` can override it.
|
|
196
|
+
|
|
197
|
+
`publish-local.mjs` creates a tarball in the workspace directory, copies it
|
|
198
|
+
to `~/.pi/agent/local-packages` (or `LOCAL_PKG`), then removes the workspace
|
|
199
|
+
copy. The installed package and retained tarball use the destination, not the
|
|
200
|
+
project CWD.
|
|
@@ -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
|
|
@@ -27,6 +27,11 @@ import { consolidatedClaimCritique } from "../../../../lib/ai/consolidated-claim
|
|
|
27
27
|
|
|
28
28
|
const log = createLogger("metacognition:claim-verification");
|
|
29
29
|
|
|
30
|
+
/** A blank string is "not provided": a client that emits every schema key sends "" for the unused ones. */
|
|
31
|
+
function optionalString(v: unknown): string | undefined {
|
|
32
|
+
return typeof v === "string" && v.trim().length > 0 ? v : undefined;
|
|
33
|
+
}
|
|
34
|
+
|
|
30
35
|
function truncate(text: string, maxLen: number): string {
|
|
31
36
|
return text.length > maxLen ? text.slice(0, maxLen) + "..." : text;
|
|
32
37
|
}
|
|
@@ -166,7 +171,9 @@ export function createVerifyClaimTool(): ToolDefinition<any, any, any> {
|
|
|
166
171
|
if (!p.claim || !Array.isArray(p.context)) {
|
|
167
172
|
throw new Error("verify_claim requires 'claim' (string) and 'context' (string array)");
|
|
168
173
|
}
|
|
169
|
-
|
|
174
|
+
// A blank claimId is no id: it would become the tracking key of every
|
|
175
|
+
// such call for the rest of the session.
|
|
176
|
+
const claimId = optionalString(p.claimId) ?? `claim-${Date.now()}`;
|
|
170
177
|
const issues: Record<string, unknown>[] = [];
|
|
171
178
|
|
|
172
179
|
// PRIMARY PATH: consolidated single-LLM-call
|
|
@@ -528,6 +528,27 @@ finalize(completed=true, summary="Fixed the bug", tasksRemaining=["write tests"]
|
|
|
528
528
|
};
|
|
529
529
|
}
|
|
530
530
|
|
|
531
|
+
// A blank summary names nothing. The `=== undefined` check above lets it
|
|
532
|
+
// through, and the summary lands in the memory episode, the session label
|
|
533
|
+
// and the deliverables manifest -- an unfindable task recorded as done.
|
|
534
|
+
if (typeof params.summary !== "string" || params.summary.trim().length === 0) {
|
|
535
|
+
return {
|
|
536
|
+
content: [
|
|
537
|
+
{
|
|
538
|
+
type: "text" as const,
|
|
539
|
+
text: `**FINALIZE REQUIRES A SUMMARY**
|
|
540
|
+
|
|
541
|
+
finalize was called with summary ${JSON.stringify(params.summary ?? null)}, which names nothing to record.
|
|
542
|
+
Provide the summary text: what was accomplished, or what is still open with completed: false.
|
|
543
|
+
|
|
544
|
+
Example:
|
|
545
|
+
finalize(completed=true, summary="Fixed the parser offset clamp", tasksRemaining=["add fixture"])`,
|
|
546
|
+
},
|
|
547
|
+
],
|
|
548
|
+
details: { finalized: false, blocked: true, error: "missing_summary" },
|
|
549
|
+
};
|
|
550
|
+
}
|
|
551
|
+
|
|
531
552
|
// Pre-flight check: verify must have passed
|
|
532
553
|
if (_ctx && !params.completed) {
|
|
533
554
|
const lastVerification = getLastVerification(_ctx);
|
|
@@ -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,
|
|
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
|
|
329
|
-
|
|
330
|
-
|
|
331
|
-
|
|
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]
|
|
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
|
},
|
|
@@ -29,6 +29,11 @@ import { BRAID_INJECTED, BRAID_SOURCE } from "../../../lib/constants/injection-d
|
|
|
29
29
|
|
|
30
30
|
const log = createLogger("topic-drift");
|
|
31
31
|
|
|
32
|
+
/** A blank string is "not provided": a client that emits every schema key sends "" for the unused ones. */
|
|
33
|
+
function optionalString(v: unknown): string | undefined {
|
|
34
|
+
return typeof v === "string" && v.trim().length > 0 ? v : undefined;
|
|
35
|
+
}
|
|
36
|
+
|
|
32
37
|
interface DriftState {
|
|
33
38
|
/** The original request text (used by critics directly - no token extraction needed). */
|
|
34
39
|
originalRequest: string;
|
|
@@ -348,7 +353,20 @@ export default function topicDriftExtension(pi: ExtensionAPI) {
|
|
|
348
353
|
switch (params?.action) {
|
|
349
354
|
case "check": {
|
|
350
355
|
const state = getExtensionState(ctx, "topic-drift", createDriftState);
|
|
351
|
-
const output = params.output
|
|
356
|
+
const output = optionalString(params.output);
|
|
357
|
+
if (output === undefined) {
|
|
358
|
+
// "" would be measured as an empty response and answered
|
|
359
|
+
// "No topic drift" -- a pass for a check that never ran.
|
|
360
|
+
return {
|
|
361
|
+
content: [
|
|
362
|
+
{
|
|
363
|
+
type: "text" as const,
|
|
364
|
+
text: `topic_drift action=check requires a non-empty 'output': the response text to test against the anchored request.`,
|
|
365
|
+
},
|
|
366
|
+
],
|
|
367
|
+
details: { error: "missing_output", action: "check" },
|
|
368
|
+
};
|
|
369
|
+
}
|
|
352
370
|
let drifted = false;
|
|
353
371
|
let similarity = 1.0;
|
|
354
372
|
let missingAnchors: string[] = [];
|
|
@@ -397,7 +415,7 @@ export default function topicDriftExtension(pi: ExtensionAPI) {
|
|
|
397
415
|
}
|
|
398
416
|
case "reset": {
|
|
399
417
|
const state = getExtensionState(ctx, "topic-drift", createDriftState);
|
|
400
|
-
const request = params.newRequest ?? state.originalRequest;
|
|
418
|
+
const request = optionalString(params.newRequest) ?? state.originalRequest;
|
|
401
419
|
resetAnchors(state, request);
|
|
402
420
|
return {
|
|
403
421
|
content: [
|
|
@@ -25,7 +25,7 @@ Every file in this folder either:
|
|
|
25
25
|
| `identity.ts` | Framing constants: `IDENTITY_INTRO` (generalist identity), `DISCOVERY` (D1 guarded exploration stance), `WORK_ETHIC` (methodical precision, signal-not-noise), `REPAIR_DISCIPLINE` (root cause named falsifiably before repair; re-derivation of claims that are not yours; weighted repair decisions), `SELF_MANAGEMENT` (memory workspace, evidence-carrying done items, known-open-check done gate), `COMMUNICATION` (conclusion-first, output style), `EPISTEMIC_STANCE` (identity-core: claim gate, self-state gate, retrieved-context preference), `SELF_AUDIT` (symmetric doubt; handoff ends with the unverified), `PHILOSOPHY` (character-level stance - silence, calm strategist, precise self-irony). |
|
|
26
26
|
| `surface-discipline-section.ts` | `SURFACE_DISCIPLINE` (signal discipline across code, commits, tests, docs, reports, logs) + `surfaceDisciplineSection()` - its own section, rendered between the identity and identity-core blocks. Text pinned verbatim by test-prompt-budget.mjs. |
|
|
27
27
|
| `instance-composition.ts` | **v3.8** Instance (consolidated): agent identity + pi-core version + root (pi-core npm package folder) + cwd + userland settings dir + os/pid + session (id + elapsed/last-event/turns/tool-avg) + top-level pi extension packages (settings-driven, scope-tagged `project` or `user-global`). Full block rendered on demand via the `system_info` tool; the prompt carries only the compact orientation line. |
|
|
28
|
-
| `top-level-extensions.ts` | **v3.8** Discovery logic for top-level pi extension PACKAGES. Reads pi-core's `packages` field from `<cwd>/.pi/settings.json` (project) and `~/.pi/agent/settings.json` (user-global); project wins on collision. Resolves `npm:` specs
|
|
28
|
+
| `top-level-extensions.ts` | **v3.8** Discovery logic for top-level pi extension PACKAGES. Reads pi-core's `packages` field from `<cwd>/.pi/settings.json` (project) and `~/.pi/agent/settings.json` (user-global); project wins on collision. Resolves `npm:` specs in pi's managed install tree (`<agentDir>/npm/node_modules`) first, falling back to the global npm root (`npm root -g`); resolves local paths in place. Source tags: `project` / `user-global`. |
|
|
29
29
|
| `project-docs-section.ts` | Project docs (AGENTS.md / CLAUDE.md) pointer-only rendering as plain-text bullets (`- <basename> at <path>`). Mirrors the skills discoverability model but uses plain text (no XML tags) so the system prompt stays prose. Includes fingerprint cache and `RESOURCES_DISCOVER` invalidation. |
|
|
30
30
|
| `memory-section.ts` | Memory store pointer block (episodes + semantic + heuristics) as plain-text bullets (`- episodes at PATH`; count-free - per-store counts go stale at print time and are not rendered). Mirrors project-docs. Cached per session (static content, no fingerprint needed); invalidated on `RESOURCES_DISCOVER`. |
|
|
31
31
|
| `context-section.ts` | **v3.6** body fragment for the unified context block: context %, session health, steps-since-checkpoint. Returns `""` when nothing to surface. Composer (`index.ts`) joins this with project-docs and memory bodies. No heading. (Memory stores render pointer-only: per-store entry counts are not rendered; the whisper carries token counts, not memory counts.) |
|