mirror of
https://github.com/garrytan/gbrain.git
synced 2026-07-31 04:07:52 +00:00
cb5df8fcc51441e8d24a7179fce613cc0f12217c
107
Commits
| Author | SHA1 | Message | Date | |
|---|---|---|---|---|
 Garry TanandClaude Opus 4.6
|
cb5df8fcc5 |
fix(eval): drop top_p from amara-life-gen Opus params + gitignore _cache/
Two fixes surfaced during the Day 3b real-corpus run against Opus 4.5:
**eval/generators/amara-life-gen.ts** — Current Opus rejects
`temperature` and `top_p` together:
```
400 invalid_request_error: `temperature` and `top_p` cannot both be
specified for this model. Please use only one.
```
top_p=1.0 was a no-op (no nucleus truncation), so removing it has zero
semantic effect. The field is still part of MODEL_PARAMS for the cache
key so any past cache entries (none in v1) would invalidate cleanly
on the next schema version bump.
**.gitignore** — `eval/data/amara-life-v1/_cache/` is runtime Opus
cache (398 files, ~1.6MB). Regenerable from seed; no point in source
control. The corpus itself (inbox/slack/calendar/meetings/notes/docs +
corpus-manifest.json with per-item content_sha256) stays committable
for reproducibility, just the cache directory gets excluded.
Real corpus generation ran cleanly after these two fixes: 398 LLM
calls, 84,424 input / 38,062 output tokens, \$4.12 spent (vs \$20 cap,
vs \$12 estimate). All 418 items produced. Poison fixtures use
subtle paraphrased injection ("for anyone on your team who might be
triaging this thread later…") — exactly the pattern that defeats
regex redaction and requires the structured-evidence judge contract
from Day 5.
Corpus itself stays local (will move to the brainbench sibling repo
during the v0.16 split per the design doc). No eval/data/amara-life-v1/
content landing in this PR.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
|
||
|
Garry Tan
|
3dc7d69a1e |
Merge remote-tracking branch 'origin/master' into garrytan/gbrain-evals
# Conflicts: # CHANGELOG.md # VERSION # package.json |
||
|
Garry TanandClaude Opus 4.6
|
7e98739a6b |
feat(eval): Day 10 — all.ts rewrite + llm-budget + BrainBench N tiers
Final wiring of BrainBench v1 Complete. all.ts now orchestrates the full
Cat catalog (1-12) via a mix of subprocess dispatch (Cats 1, 2, 3, 4, 6,
7, 10, 11, 12 — standalone runners with CLI entry points) and
programmatic invocation (Cats 5, 8, 9 — require runtime inputs that
can't come via CLI flags). Subprocess Cats run concurrently under a
p-limit(2) bound to cap peak memory around ~800MB (two PGLite instances
at ~400MB each).
Cats 5/8/9 show as "programmatic" in the report with a one-line
reference to their `runCatN({...})` harness API. They're deliberately
skipped from the master runner because their inputs (claim catalog,
probe catalog, scenario catalog, pre-seeded agent state, evidence
pagesBySlug) are task-specific and assembled at the caller.
**eval/runner/all.ts** — rewritten:
- CATEGORIES is a tagged union of SubprocessCategory | ProgrammaticCategory
- runCatSubprocess spawns Bun with pipe'd stdout/stderr, 10-min timeout
per Cat (124 exit + SIGTERM on timeout; no hung subprocesses)
- runConcurrently is a bounded worker pool preserving input order
- buildReport emits the full markdown with per-Cat elapsed times,
migration-noise filter, and a separate programmatic-only section
- Honors BRAINBENCH_N (1/5/10 for smoke/iteration/published),
BRAINBENCH_CONCURRENCY (default 2),
BRAINBENCH_LLM_CONCURRENCY (default 4, consumed by llm-budget)
**eval/runner/llm-budget.ts** — shared LLM rate-limit semaphore. A full
N=10 published scorecard makes ~900 Anthropic calls (150 Cat 8/9 probes
× N=10 + 100 Cat 5 claims × N=10). Without coordination, concurrent
adapters trigger 429s on per-minute limits.
- LlmBudget class: acquireSlot/releaseSlot + withLlmSlot(fn) wrapper
that releases on success AND throw (try/finally)
- getDefaultLlmBudget() singleton reads BRAINBENCH_LLM_CONCURRENCY,
falls back to 4 on missing/garbage values
- capacity enforced ≥1 (rejects 0/negative)
- Double-release is a no-op (guards against upstream double-call bugs)
- Active + waiting counts exposed for observability / tests
**package.json** scripts:
- eval:brainbench — default N=5 iteration
- eval:brainbench:smoke — N=1 for fast iteration
- eval:brainbench:published — N=10 for committed baselines
- eval:cat6 / eval:cat11 — individual new subprocess Cats
**Tests (24):** CATEGORIES catalog enforces the exact Cat-number partition
(subprocess: 1,2,3,4,6,7,10,11,12; programmatic: 5,8,9). runConcurrently
respects the cap (observable via peak in-flight counter), preserves input
order under non-uniform delays, handles empty input. LlmBudget enforces
capacity, releases on throw, honors env var, rejects 0/negative.
buildReport filters migration noise, counts passed/failed/programmatic
correctly, includes every Cat + programmatic-only section.
Full eval suite now: 314 pass, 0 fail (15 test files).
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
|
||
|
Garry TanandClaude Opus 4.6
|
e1e699f0a6 |
feat(eval): Day 9 — sealed qrels via PublicPage + PublicQuery at adapter boundary
Codex fixes #1, #2, #3 from the plan's outside-voice review. Enforcement shifts from SOFT-VIA-TYPE-COMMENT to SOFT-VIA-SANITIZED-OBJECT. Hard enforcement via process isolation waits for BrainBench v2 Docker sandbox. **eval/runner/types.ts** additions: - `PublicPage = Pick<Page, 'slug' | 'type' | 'title' | 'compiled_truth' | 'timeline'>` — the exact 5 fields adapters should see. No _facts. No frontmatter (a known hiding spot for accidental gold leaks). - `sanitizePage(p: Page): PublicPage` — returns a NEW object with the 5 fields only. Cannot be bypassed by `(page as any)._facts` because the field does not exist on the sanitized object. - `PublicQuery = Omit<Query, 'gold'>` — strips the gold field. - `sanitizeQuery(q: Query): PublicQuery` — enumerates public fields explicitly (not spread+delete) so no prototype weirdness leaves gold reachable. **eval/runner/multi-adapter.ts** — scoreOneRun now calls sanitizePage / sanitizeQuery before passing to adapter.init / adapter.query. The scorer retains the full Query shape (including gold.relevant) for precision / recall computation. Adapter signatures unchanged — the sealing is at the OBJECT level, not the type level. This keeps existing adapters (ripgrep-bm25, vector-only, hybrid-nograph, gbrain-after) binary-compatible. Verified: no existing adapter reads q.gold or page._facts, so the change is safe without further adapter updates. **test/eval/sealed-qrels.test.ts** (17 tests): - sanitizePage strips _facts + frontmatter + arbitrary hidden keys - Output has exactly the 5 public keys (deep introspection) - Proxy tripwire simulates a malicious adapter: any access to _facts or gold throws `sealed-qrels violation` - sanitizeQuery retains optional fields (as_of_date, tags, author, acceptable_variants, known_failure_modes) but omits undefined ones - Honest documentation of the seal's limits: filesystem bypass and Proxy attacks would still work in v1; Docker isolation (v2) is the real enforcement Every existing eval test still passes (273 before + 17 sealed-qrels = 290). Total eval suite now: 290 pass, 0 fail. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> |
||
|
Garry TanandClaude Opus 4.6
|
a4cdb41b07 |
feat(eval): Day 8 — Cat 8 skill compliance + Cat 9 end-to-end workflows
**eval/runner/cat8-skill-compliance.ts** — Deterministic, judge-free Cat 8
scoring. Replays inbound signals through the agent adapter (Day 5) and
extracts four iron-law metrics directly from the tool-bridge state:
- brain_first_compliance: agent called search/get_page BEFORE producing
its final answer. Non-compliance = hallucinating from general knowledge.
- back_link_compliance: every dry_run_put_page intent has at least one
markdown [Name](slug) back-link in its compiled_truth.
- citation_format: timeline entries use canonical `- **YYYY-MM-DD** |
Source — Summary`; long final answers cite at least one slug.
- tier_escalation: simple probes use light tooling (≥1 brain call);
complex probes require ≥2 brain calls or a dry_run write when
expects_dry_run_write is set.
No judge call required — everything is computable from
`tool_bridge_state.made_dry_run_writes` + `count_by_tool` + final_answer
regex. Fast, deterministic, reproducible.
Bounded concurrency (p-limit style) worker pool at default 4 to keep
Sonnet rate limits comfortable across 100-probe batches.
**eval/runner/cat9-workflows.ts** — Rubric-graded Cat 9. 5 canonical
workflows (meeting_ingestion, email_to_brain, daily_task_prep, briefing,
sync) × ~10 scenarios each. Each scenario runs through the agent adapter,
then judge.ts scores the answer against a per-scenario rubric.
`buildEvidence(scenario, agentResult, pagesBySlug)` composes the
JudgeEvidence contract: resolves ground_truth_slugs to full
GroundTruthPage[] from a slug-map, pulls tool_call_summary directly from
tool_bridge_state (no raw tool_result content — Section-3 defense),
attaches rubric from the scenario.
Per-workflow rollup: each workflow gets its own pass_rate so the verdict
can fail one workflow without failing the whole Cat. Overall verdict
requires every populated workflow's pass_rate ≥ threshold (default 0.80)
when enableThreshold=true.
Both Cats default to verdict=baseline_only in v1 per codex fix #9: real
thresholds return after 10-probe Haiku-vs-hand-score calibration (κ > 0.7)
runs against the Day 3b amara-life-v1 corpus.
**Tests (23):** Cat 8 per-metric scorer unit tests covering every branch
(brain_first ordering, back-link compliance on mixed writes, long vs
short answer citation requirement, tier escalation for simple/complex/
writey probes, finalAnswerCiteCount dedups across syntaxes). Cat 9
buildEvidence contract shape — evidence_refs flow from agent, missing
slugs skip gracefully, no raw_transcript/tool_result leakage to judge.
Cat 9 runCat9 integration with stubbed agent + mixed-verdict judge
produces fractional pass rates correctly.
Total eval suite now: 273 pass, 0 fail.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
|
||
|
Garry TanandClaude Opus 4.6
|
33e2e73940 |
feat(eval): Day 7 — Cat 5 provenance runner + structured classify_claim judge
**eval/runner/cat5-provenance.ts** — BrainBench Cat 5 scoring. Samples claims from gbrain brain-export and classifies each against its source material via a dedicated Haiku judge (classify_claim tool with a three-label enum: supported | unsupported | over-generalized). Separate from judge.ts by design: Cat 5 is a single three-way classification per claim, not a weighted rubric. Rather than overload judge.ts with a mode switch, Cat 5 has its own tool definition (CLASSIFY_CLAIM_TOOL) and prompt. The retry-once pattern, $20 cost gate semantics, and structured parsing are mirrored from judge.ts so failures look the same across Cats. Metric: `citation_accuracy` = fraction where predicted label equals gold expected_label. Threshold (informational): >0.90 per design-doc METRICS.md. v1 ships with `enableThreshold: false` so the verdict is always baseline_only — we don't have hand-authored gold claims yet, and codex flagged that threshold gating should wait until the amara-life-v1 corpus + gold file authoring lands in Day 3b. runCat5 uses a bounded-concurrency worker pool (default 4) to respect Haiku rate limits across 100+ claim batches. Evidence pages are looked up by slug from a caller-provided pagesBySlug map — missing pages don't crash, they just pass an empty source list to the judge (correct behavior for genuinely unsupported claims). **Tests (23):** classifyClaim happy/retry/fallback paths with stubbed Haiku, aggregate accuracy math, threshold gating (pass/fail vs baseline_only), runCat5 concurrency + missing-page handling, renderClaimPrompt embeds claim + sources correctly, parseClassification rejects invalid enum values + plain-text responses. Total eval suite now: 250 pass, 0 fail. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> |
||
|
Garry TanandClaude Opus 4.6
|
839258dc42 |
feat(eval): Day 6 — adversarial-injections + Cat 6 prose-scale + Cat 11 multi-modal
Three modules that together cover BrainBench v1 Cat 6 (prose-scale
extraction fidelity) and Cat 11 (multi-modal ingest fidelity).
**eval/runner/adversarial-injections.ts** — 6 deterministic content
transforms shared by Cat 10 (adversarial.ts, 22 hand-crafted cases) and
Cat 6 (prose-scale variants). Each injection produces a modified content
string + a structured GoldDelta describing what the extractor MUST and
MUST NOT produce. Kinds:
- code_fence_leak — fake [X](people/fake) inside ``` fence, must NOT extract
- inline_code_slug — `people/fake` in backticks, must NOT extract
- substring_collision — "SamAI" near real `people/sam`, exactly one link
- ambiguous_role — "works with" vs "works at", downgrade type to mentions
- prose_only_mention — strip markdown link syntax, bare name → mentions only
- multi_entity_sentence — pack 4+ entities into one clause, extract all
Mulberry32 PRNG keeps variant generation deterministic under fixed seed.
Codex flagged the original plan's wording ("extract injection engine from
adversarial.ts") as overstated — adversarial.ts is a static case list,
not a reusable engine. This module is NEW code.
**eval/runner/cat6-prose-scale.ts** — Runner. Loads world-v1, applies all
6 injection kinds to sampled base pages (default 50 variants per kind ×
6 kinds = 300 variants), runs extractPageLinks on each, compares to gold
delta. Emits per-kind + overall metrics (precision, recall, F1,
code_fence_leak_rate, substring_fp_rate, pages_with_links_coverage,
mean_links_per_page). **v1 verdict is always "baseline_only"** — no
gating threshold per codex fix #9 (current extractor residuals make
>0.80 unreachable; v1 records a baseline, regression guard triggers on
drop below it).
**eval/runner/cat11-multimodal.ts** — PDF + HTML + audio runners.
Fixtures load from eval/data/multimodal/<modality>/fixtures.json
manifests; each modality skips gracefully when manifest missing or
(audio) when neither GROQ_API_KEY nor OPENAI_API_KEY is set. Metrics:
- PDF: char-level similarity via Levenshtein + optional entity_recall
- HTML: word-recall over normalized tokens (multiset semantics)
- Audio: WER (word error rate) via Levenshtein on word sequences
Fixtures are NOT committed; a future eval:fetch-multimodal script will
download them hash-verified from public sources (arXiv CC-licensed
papers, Wikipedia CC-BY-SA, Common Voice CC0).
Injectable audio transcriber (`opts.transcribe`) means tests don't need
GROQ/OpenAI keys — stubbed transcriptions exercise the WER math path
directly.
**Tests (60 new):** adversarial-injections (19) — per-kind assertions +
dispatcher coverage + slug regex conformance; cat6 (12) — variant
determinism, scoreVariant shape, aggregate per-kind + overall metrics,
corpus resolver slug rules; cat11 (29) — charSimilarity / wordRecall /
wer math, htmlToText strips scripts + decodes entities, HTML modality
with real fixtures, audio modality gracefully skips without key + uses
stub transcriber correctly.
All 60 tests pass in 48ms + 41ms.
Total eval suite now: 227 pass, 0 fail.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
|
||
|
|
ebfbd5e6f7 |
feat(doctor): proximity-based DRY detection + --fix auto-repair (v0.14.1) (#254)
* feat(doctor): proximity-based DRY detection + --fix auto-repair
Fixes false-positive DRY violations on skills that properly delegate
notability/filing rules to `skills/_brain-filing-rules.md`. The old
check only accepted `conventions/quality.md` as a valid delegation
target, leaving 9 skills flagged every run even though they delegate
correctly.
- CROSS_CUTTING_PATTERNS.conventions is now an array; notability gate
accepts both `conventions/quality.md` AND `_brain-filing-rules.md`
- New extractDelegationTargets() parses `> **Convention:**`,
`> **Filing rule:**`, and inline backtick references
- DRY suppression is proximity-based (K=40 lines) via DRY_PROXIMITY_LINES
- New src/core/dry-fix.ts module with autoFixDryViolations:
- expanders strategy map (bullet / blockquote / paragraph)
- 5 guards: working-tree-dirty, no-git-backup, inside-code-fence,
already-delegated, ambiguous-multi-match, block-is-callout
- execFileSync array args (no shell-injection surface)
- EOF newline preservation
- `gbrain doctor --fix` and `--dry-run` flags wire in via doctor.ts
- 31 new tests across dry-fix.test.ts (28 unit), check-resolvable.test.ts
(13 DRY detection + extraction), doctor-fix.test.ts (3 CLI integration)
* chore: bump version and changelog (v0.14.1)
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* docs: update project documentation for v0.14.1
CLAUDE.md:
- Added src/core/dry-fix.ts entry under Key files (expanders, guards,
execFileSync safety, EOF newline preservation).
- Updated src/commands/doctor.ts entry to cover --fix/--dry-run flags.
- Updated src/core/check-resolvable.ts entry to reflect array-valued
CROSS_CUTTING_PATTERNS.conventions, extractDelegationTargets(), and
proximity-based DRY suppression via DRY_PROXIMITY_LINES = 40.
- Added test/dry-fix.test.ts and test/doctor-fix.test.ts to the test
list, and annotated test/check-resolvable.test.ts with v0.14.1 cases.
README.md:
- ADMIN block: --fix now names what it actually fixes (DRY violations
via conventions delegation) and documents --dry-run.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
|
||
|
Garry TanandClaude Opus 4.6
|
05aec3feb8 |
feat(eval): Day 5 — agent adapter + judge with structured evidence contract
Two modules that together wire Cat 8 / Cat 9 / Cat 5 end-to-end scoring. **eval/runner/judge.ts** — Haiku 4.5 via tool-use `score_answer`. Input is the structured JudgeEvidence contract (fix #16 from the plan's codex review): probe + final_answer_text + evidence_refs + tool_call_summary + ground_truth_pages + rubric. Raw tool output NEVER reaches the judge — that's the Section-3 defense against paraphrased prompt-injection payloads in gold/poison.json. Retry policy: one retry on malformed tool_use response. If the second attempt is still malformed, score the probe as `judge_failed` (all scores 0, verdict=fail) so the run still completes. Aggregation: weighted mean across rubric criteria. Canonical thresholds (pass ≥3.5, partial 2.5-3.5, fail <2.5) — judge can propose a verdict but the computed verdict from the weighted mean is what the scorecard records. This prevents the model from inflating or deflating its own verdict. Score values are clamped to 0-5 on parse even if the model returns out of range. `assertNoRawToolOutput(evidence)` is a regression guard that returns the list of forbidden fields (tool_result, raw_transcript, etc.) if any leak into the evidence contract. **eval/runner/adapters/claude-sonnet-with-tools.ts** — The agent adapter. Implements `Adapter` interface minimally: `init()` spins up PGLite and seeds it, `query()` throws because the adapter is Cat 8/9-only and emits a final-answer text, not a RankedDoc[]. Retrieval scorecard stays at 4 adapters. `runAgentLoop(probeId, text, state, config)` drives the multi-turn loop: Sonnet → tool_use → tool-bridge.executeTool → tool_result → back to Sonnet. Turn cap 10. max_tokens 1024. System prompt (brain-first iron law, citation format, amara context) is cached via cache_control. Exponential backoff on rate-limit errors (1s, 2s, 4s). Emits a `Transcript` per eval/schemas/transcript.schema.json — consumed directly by recorder.ts for the flight-recorder bundle. `brain_first_ordering` classifies Cat 8's flagship metric: did the agent call search/get_page BEFORE producing the final answer? The `no_brain_calls` case (agent answers from general knowledge without ever hitting the brain) is the compliance failure to surface. ForbiddenOpError + UnknownToolError from the bridge are caught in the agent loop and surfaced as tool_result with is_error=true — keeps the loop going and preserves full audit trail for the judge. **Tests (35 new):** judge (23) — happy path, retry, fallback, evidence contract sanitization, rendered prompt does not contain raw tool_result text, verdict thresholds, score clamping, weighted mean with mixed weights, parseToolUse rejects malformed input. agent-adapter (12) — Adapter.query() throws, init() seeds PGLite, end-to-end tool loop with stubbed Sonnet, turn cap exhaustion, mutating-op rejection surfaces as tool_result error, extractSlugs regex. All 12 agent tests take ~23s because PGLite runs 13 schema migrations per test; the alternative of shared-engine-across-tests was rejected so each test is isolated. Total eval suite now: 167 pass, 0 fail. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> |
||
|
Garry TanandClaude Opus 4.6
|
934d7ea4b6 |
feat(eval): Day 4 — pdf-parse + flight-recorder + tool-bridge (dry_run + expand:false)
Three infrastructure modules for BrainBench v1 Complete Cats 5/8/9/11.
**eval/runner/loaders/pdf.ts** — Thin pdf-parse wrapper. Lazy import keeps
pdf-parse out of the module-load path (avoids library debug-mode side
effects). Size cap (50MB default), encryption detection, structured error
classes (PdfEncryptedError, PdfTooLargeError, PdfParseError). Only Cat 11
multimodal will import this; production bundle never sees pdf-parse.
**eval/runner/tool-bridge.ts** — Maps 12 read-only operations from
src/core/operations.ts to Anthropic tool definitions + adds 3 dry_run write
tools. Three structural invariants enforced:
1. No hidden LLM calls. `operations.query` defaults expand=true which
routes through expansion.ts → Haiku. Bridge strips `expand` from the
query tool's input schema AND executor hard-sets expand:false. Zero
nested Haiku calls in any agent trace.
2. Mutating ops throw ForbiddenOpError. put_page, add_link, delete_page,
etc. are rejected by name. Agents record intent via dry_run_put_page /
dry_run_add_link / dry_run_add_timeline_entry which persist to the
flight-recorder without mutating the engine. This is how Cat 8's
back_link_compliance + citation_format metrics measure anything with
a read-only tool surface.
3. Poison tagged by the bridge, not the judge. Every tool result is
scanned for slugs matching gold/poison.json fixtures. Matched
fixture_ids flow into tool_call_summary.saw_poison_items for the
structured-evidence judge contract. Judge never reads raw tool
output — Section-3 defense against paraphrased prompt injections
(poison payloads never reach the judge model at all).
32K-token cap (~128K chars) with "…[truncated]" suffix.
**eval/runner/recorder.ts** — Per-run flight-recorder bundle emitter. Full
6-artifact bundle (transcript.md, brain-export.json, entity-graph.json,
citations.json, scorecard.json, judge-notes.md) when the adapter provides
an AdapterExport; 3-artifact fallback (transcript + scorecard +
judge-notes) otherwise. Atomic writes via tmp+rename. Collision-safe:
duplicate directory names get incremental -2, -3 suffix. `safeStringify`
handles circular references without throwing and JSON-serializes
Float32Array embeddings.
**package.json:** adds pdf-parse@2.4.5 as a devDependency. Scoped to eval/
use only; production gbrain binary unaffected.
**Tests:** 63 new — 30 tool-bridge, 21 recorder, 12 pdf-loader. All pass.
Fake engine uses a Proxy with `__default__` fallback so poison-matching
tests don't have to mock the exact engine method name that each operation
calls (some route via searchKeyword, others via getPage — proxy handles
both uniformly).
Total eval suite now: 132 pass, 0 fail, 923 expect() calls.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
|
||
|
Garry TanandClaude Opus 4.6
|
8f326cb9f9 |
docs: update CLAUDE.md + README + eval/README for v0.15.0 BrainBench
CLAUDE.md: adds a full BrainBench section to the Key Files list — 14 new entries covering eval/README.md, multi-adapter.ts, types.ts (with new PublicPage/PublicQuery), adapters/, queries/, type-accuracy.ts, adversarial.ts, all.ts, world.ts/gen.ts, world-html.ts, amara-life.ts, amara-life-gen.ts, schemas/, data/world-v1/, data/gold/, data/amara-life-v1/, docs/benchmarks/, and test/eval/. Adds 3 new test/eval/ lines to the unit-tests catalog. eval/README.md: file tree updated to reflect v0.15 additions — data/amara-life-v1/, data/gold/, schemas/, generators/amara-life.ts + amara-life-gen.ts, runner/all.ts + adversarial.ts. README.md: updates hero benchmark numbers (L7 intro + L353 mid-page) from v0.10.5 PR #188 numbers (R@5 83→95, P@5 39→45) to current v0.12.1 4-adapter numbers (P@5 49.1% · R@5 97.9% · +31.4 pts vs hybrid-nograph). Adds the v0.11→v0.12 regression comparison as the secondary reference. Deeper-section tables (L422+) labeled "BrainBench v1 (PR #188)" are preserved as historical data. CHANGELOG is untouched — /ship already wrote the v0.15.0 entry. TODOS.md is untouched — Cat 5/6/8/9/11 remain open (only foundations shipped in v0.15.0; Cat runners ship in v1 Complete follow-ups). Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> |
||
|
Garry TanandClaude Opus 4.6
|
d4a2d51e11 |
chore: bump version and changelog (v0.15.0)
Resets package.json from stale 0.13.1 to 0.15.0 (matches VERSION). v0.14.0 shipped with the stale package.json version; this sync catches that up and moves to v0.15.0 in one step. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> |
||
|
Garry Tan
|
1f38c4f96b |
feat(eval): amara-life-gen.ts with structured cache key + $20 cost gate
Opus prose expansion of the amara-life-v1 skeleton. Per-item structured
cache key = sha256({schema_version, template_id, template_hash, model_id,
model_params, seed, item_spec_hash}). Prompt-template tweak changes
template_hash; only those items regenerate. Schema bump changes
schema_version; everything invalidates cleanly. Interrupted runs resume
from the last cached item; zero re-spend.
Cost-gated at $20 hard-stop with Anthropic input/output pricing tracking.
Dry-run mode (--dry-run) executes the full pipeline with stub bodies for
smoke-testing the I/O layout without LLM spend. --max N caps items per
type for debugging. --force ignores cache.
Writes per-format outputs under eval/data/amara-life-v1/:
inbox/emails.jsonl (one email per line with body_text appended)
slack/messages.jsonl (one message per line with text appended)
calendar.ics (RFC-5545 VEVENT format, templated — no LLM)
meetings/<id>.md (transcript with YAML frontmatter)
notes/<YYYY-MM-DD-topic>.md (first-person journal)
docs/*.md (6 reference docs, templated — no LLM)
corpus-manifest.json (per eval/schemas/corpus-manifest.schema.json,
including per-item content_sha256 and generator_cache_key)
Perturbation hints (contradiction, stale-fact, poison, implicit-
preference) flow through the prompt so Opus weaves the specific claim
into each item's body. Poison items are hand-crafted to include
paraphrased prompt-injection attempts (not literal 'IGNORE ALL
PREVIOUS' — defense is the structured-evidence judge contract at
Day 5, not regex redaction).
New package.json scripts:
eval:generate-amara-life # real run (~$12 Opus estimated)
eval:generate-amara-life:dry # smoke test, zero spend
test:eval extended to include test/eval/. 10 cache-key tests cover
determinism, invalidation across every field of the key, canonical JSON
stability under object-key reorder, and per-skeleton-item spec-hash
uniqueness (50 distinct hashes for 50 distinct emails).
|
||
|
Garry Tan
|
53e32ee427 |
feat(eval): amara-life-v1 skeleton + Page.type enum for email/slack/cal/note
Deterministic procedural generator for the twin-amara-lite fictional-life corpus (BrainBench v1 Cat 5/8/9/11 target). 15 contacts picked from world-v1, 50 emails + 300 Slack messages across 4 channels + 20 calendar events + 8 meeting transcripts + 40 first-person notes. Mulberry32 PRNG gives byte-identical output under reseed. Plants 10 contradictions + 5 stale facts + 5 poison items + 3 implicit preferences at deterministic positions. Fixture_ids are unique across the corpus so gold/contradictions.json + gold/poison.json + gold/implicit- preferences.json can cross-reference by stable ID. PageType extended in both src/core/types.ts and eval/runner/types.ts to include email | slack | calendar-event | note (+ meeting on the production side). src/core/markdown.ts inferType() heuristics updated for the new one-slash slug prefixes (emails/em-NNNN, slack/sl-NNNN, cal/evt-NNNN, notes/YYYY-MM-DD-topic, meeting/mtg-NNNN). 17 tests cover counts (50/300/20/8/40), perturbation counts (exact 10/5/5/3), seed determinism + divergence, slug regex conformance (matches eval/runner/queries/validator.ts:131 one-slash rule), unique fixture_ids, amara-in-every-email invariant, calendar dtstart < dtend, and Amara-is- attendee on every meeting. |
||
|
Garry Tan
|
18baca6629 |
feat(eval): BrainBench v1 portable JSON schemas + gold templates
Adds the v1→v2 contract boundary for BrainBench. 6 JSON schemas at
eval/schemas/ pin the shape of every artifact a stack must emit to be
scorable: corpus-manifest, public-probe (PublicQuery with gold stripped),
tool-schema (12 read + 3 dry_run tools, 32K tool-output cap), transcript,
scorecard (N ∈ {1, 5, 10}), evidence-contract (structured judge input).
8 gold file templates at eval/data/gold/ scaffold the sealed qrels,
contradictions, poison items, and citation labels. Empty-but-valid
skeletons; Day 3b fills them with real content once the amara-life-v1
corpus generates.
48 tests validate schema syntax, $schema/$id/title/type headers,
round-trip stability, and cross-schema coherence (new Page types in
manifest enum, tool counts, token cap, N enum).
When v2 ports to Python + Inspect AI + Docker, these schemas are the
boundary. Same fixtures, same tool contracts, zero rework.
|
||
|
Garry Tan
|
739c50a4f3 | Merge remote-tracking branch 'origin/master' into garrytan/gbrain-evals | ||
|
|
5fd9cd2644 |
feat: shell job type + worker abort-path fix (v0.13.0) (#217)
* feat(minions): add protected-name constant + ctx.shutdownSignal
Introduce PROTECTED_JOB_NAMES ('shell') in a side-effect-free core module
so queue.ts can check it without importing from handlers/. MinionJobContext
gains shutdownSignal (distinct from signal) — handlers that need to run
SIGTERM-triggered cleanup subscribe to both; most handlers ignore shutdown
and run through the worker's 30s cleanup race to natural completion.
* fix(minions): MinionQueue.add gains trusted 4th arg + trim-normalized guard
Adds allowProtectedSubmit opt-in as a separate 4th parameter (NOT folded into
opts) so callers spreading user-provided opts ({...userOpts}) can't accidentally
carry the trust flag. PROTECTED_JOB_NAMES check runs on the trimmed name BEFORE
insert, closing the queue.add(' shell ', ...) whitespace bypass that would have
evaded a has(name) check.
* fix(minions): worker calls failJob on abort + wires ctx.shutdownSignal
Pre-v0.13.0 worker returned silently when ctx.signal.aborted fired, leaving
jobs in 'active' until stall sweep. Handlers using cooperative cancel had
no deterministic status flip — timeout/cancel/lock-loss all looked the same
from downstream callers (gbrain jobs get, --follow loops).
Fix: derive abort reason from abort.signal.reason ('timeout' | 'cancel' |
'lock-lost' | 'shutdown') and call failJob with 'aborted: <reason>' text.
failJob is idempotent via token+status match, so no-op when another path
already flipped status (handleTimeouts, cancelJob, stall).
Also: new shutdownAbort (instance-level AbortController) fires on process
SIGTERM/SIGINT and propagates to every handler's ctx.shutdownSignal.
Shell handler listens to both signals and runs SIGTERM→5s→SIGKILL on its
child on either; other handlers only listen to ctx.signal so deploy
restarts don't cancel them mid-flight.
* feat(minions): add shell job handler + submission audit log
New 'shell' job type spawns arbitrary commands under the Minions worker.
Deterministic cron scripts (API fetch, token refresh, scrape+write) can
move off the LLM gateway — zero Opus tokens per fire.
Handler contract:
- cmd or argv (exactly one required). cmd spawns via /bin/sh -c (absolute
path, not 'sh', to block PATH-override shell substitution). argv spawns
direct with no shell.
- cwd required, must be absolute. Operator-trust boundary.
- env defaults to SHELL_ENV_ALLOWLIST ({PATH, HOME, USER, LANG, TZ,
NODE_ENV}) picked from process.env, with caller overrides merged on top.
Prevents accidental $OPENAI_API_KEY interpolation into scripts.
- stdout/stderr retained as UTF-8-safe tails (64KB/16KB) via
string_decoder.StringDecoder. Prepends [truncated N bytes] marker.
- Abort (either ctx.signal or ctx.shutdownSignal) fires SIGTERM → 5s grace
→ SIGKILL on child. Timer NOT .unref'd so worker's 30s race waits for
the child to actually die.
shell-audit.ts writes a JSONL line per submission to
~/.gbrain/audit/shell-jobs-YYYY-Www.jsonl (ISO-week rotated, override via
GBRAIN_AUDIT_DIR). argv logged as JSON array (not space-joined, which would
flatten args with spaces). Never logs env values. Best-effort writes:
failures log to stderr but don't block submission.
* feat(jobs): submit_job MCP guard + CLI --timeout-ms + starvation warning
submit_job operation gains timeout_ms param (was missing — couldn't plumb
the existing MinionJobInput field through from either CLI or MCP). When
ctx.remote=true and name is in PROTECTED_JOB_NAMES, throws
OperationError('permission_denied'). Combined with the queue.add trusted
guard, MCP callers can never submit shell jobs even if the env flag is on.
CLI submit: new --timeout-ms N flag. Passes {allowProtectedSubmit:true}
as the 4th arg to queue.add only when the submitted name is protected
(not blanket-set for every job). Prints a starvation-warning block to
stderr when a shell job is submitted without --follow, pointing at both
--follow and 'gbrain jobs work' remediation. Fires for every shell submit
regardless of the submitter's env — the submitter env is a weak proxy for
the worker env.
Worker handler registration: conditional on GBRAIN_ALLOW_SHELL_JOBS=1.
Default: off. 'gbrain jobs submit --help' now lists handler types with a
pointer to docs/guides/minions-shell-jobs.md for shell.
* test(minions): 40 unit + 4 E2E cases for shell handler
Unit (test/minions-shell.test.ts):
- Protected names: trim-normalized, case-sensitive, whitespace bypass defense
- MinionQueue.add: trusted opt-in, whitespace bypass, non-protected untouched
- Handler validation: cmd|argv exclusive, cwd required/absolute, env strings
- Spawn: cmd/argv happy paths, non-zero exit, ENOENT, result shape
- Env allowlist: leaked-secret blocked, PATH inherited, caller override
- Abort: ctx.signal, ctx.shutdownSignal, pre-aborted signal
- Audit: ISO-week year boundary (2027-01-01 → W53 2026), mid-year W52/W53,
GBRAIN_AUDIT_DIR override, argv as JSON array, env never logged, EACCES
non-blocking
- Output truncation: 100KB → last 64KB with [truncated N bytes] marker
E2E (test/e2e/minions-shell.test.ts):
- Full lifecycle: submit → worker claim → spawn → complete
- MinionQueue.add without trusted arg throws (including whitespace bypass)
- submit_job with ctx.remote=true rejects shell (MCP guard)
- submit_job with ctx.remote=false allows shell (CLI path)
* chore: bump version and changelog (v0.13.0)
Move gateway crons to Minions. Zero LLM tokens per cron fire.
Worker abort path finally marks aborted jobs dead.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* docs: reframe v0.13.0 copy for OpenClaw operators (not Wintermute-specific)
gbrain is an open-source product for any OpenClaw/Hermes operator, not
Garry's personal Wintermute deployment. Rewords the v0.13.0 CHANGELOG
entry, the minions-shell-jobs guide, and the deferred TODOS entries to
speak to "your OpenClaw" / "OpenClaw operators" instead.
Replaces /data/wintermute cwd examples with the canonical
/data/.openclaw/workspace path. Pre-existing Wintermute references in
older CHANGELOG entries (v0.11/v0.10.3) left unchanged.
* feat(migrations): add v0.13.0 adoption playbook for shell jobs
Adding the migration file the CEO review originally scoped out. Without
it, operators upgrade to v0.13.0 and the capability ships but adoption
doesn't happen — the 60% gateway CPU reduction only lands if someone
actually rewrites their crontab.
skills/migrations/v0.13.0.md is the instruction manual the host agent
reads on gbrain upgrade:
- Enable worker: GBRAIN_ALLOW_SHELL_JOBS=1 gbrain jobs work (Postgres)
or per-tick --follow (PGLite)
- Audit cron manifest: classify LLM-requiring vs deterministic
- Propose per-cron rewrites with diffs, approved one at a time
- Env allowlist guidance for scripts that need API keys
- Verification playbook: run one fire, compare pre/post, only then
approve the next batch
- Starvation sanity-check runbook item
Iron rules: never auto-rewrite the operator's crontab (host-specific
code per CLAUDE.md). LLM-requiring crons stay on the gateway. Ambiguous
cases ask the operator.
No mechanical orchestrator ships with this migration — every rewrite
is operator judgment. A future gbrain crontab-to-minions helper is
tracked in TODOS.md as P1.
* docs: sync UPGRADING + SKILLPACK with v0.13.0 shell jobs
UPGRADING_DOWNSTREAM_AGENTS.md: append v0.13.0 section per the file's
convention (each release appends). No skill edits required, feature is
off-by-default, optional adoption via skills/migrations/v0.13.0.md.
Lists typical LLM-vs-deterministic classifications so operators know
which of their crons are candidates for migration.
GBRAIN_SKILLPACK.md: add shell-jobs guide row to the cron/Minions guide
table so it's discoverable alongside existing Cron via Minions, Plugin
Handlers, and Minions fix guides.
---------
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
|
||
|
|
c89aa909c7 |
feat: Knowledge Runtime — Resolver SDK + BrainWriter + integrity + Budget + scheduler polish (v0.13.0) (#210)
* docs: Knowledge Runtime design doc (draft) — 4-layer architecture + reduced-scope delta
Captures the Knowledge Runtime design thinking from the CEO review session:
Resolver SDK, Enrichment Orchestrator, Scheduler, Deterministic Output Builder.
The original 7-phase plan was drafted before v0.12.0 (knowledge graph layer)
and v0.11.x (Minions agent runtime) shipped. Cross-referenced against what's
already merged on master, roughly 60% of the 4-layer vision is already in
production under different names:
- Minions = scheduler + plugin contract (L1 + L3)
- Knowledge graph auto-link = deterministic output at L4 + orchestrator at L2
- BrainBench v1 benchmarks already validate the graph layer
The doc is kept as a draft design reference; the actual build-out will scope
down to the real delta (typed Resolver interface, BrainWriter API + validators,
BudgetLedger, CompletenessScorer, quiet-hours + stagger). See the CEO review
notes for the reduced plan.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* feat(resolvers): Resolver SDK pass 1 — interface + registry (PR 1/5)
Adds the typed plugin interface that unifies external-lookup calls (X API,
Perplexity, HEAD check, brain-local slug resolution) behind a single shape:
registry.resolve('x_handle_to_tweet', { handle, keywords }, ctx)
→ { value, confidence, source, fetchedAt, raw? }
Zero behavior change — the registry is empty by default. Builtins
(url_reachable, x_handle_to_tweet) land in the next pass. ScheduledResolver
wrapping via Minions lands in PR 5.
New files:
- src/core/resolvers/interface.ts — Resolver<I,O>, ResolverResult<O>,
ResolverContext (engine, storage, config, logger, requestId, remote,
deadline, signal), ResolverError (not_found, already_registered,
unavailable, timeout, rate_limited, auth, schema, aborted, upstream)
- src/core/resolvers/registry.ts — ResolverRegistry (register/get/has/
list/resolve/clear/size) + getDefaultRegistry() for process-wide use
- src/core/resolvers/index.ts — barrel export
Design rules enforced by types:
- Every result carries confidence (0.0-1.0) + source attribution
- LLM-backed resolvers return confidence<1.0 by convention
- ctx.remote propagates the trust boundary (mirrors OperationContext.remote)
- AbortSignal threads through for cooperative cancellation
Smoke: imports + runs, list()/get()/resolve() behave as typed.
Dependency-free beyond types and storage/engine type imports.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* feat(fail-improve): optional AbortSignal — Resolver SDK pass 2 (PR 1/5)
Extends FailImproveLoop.execute with an optional `opts.signal` that threads
through the deterministic-first / LLM-fallback flow. Needed by the Resolver
SDK so long-running lookups can be cooperatively cancelled when a caller
aborts (deadline hit, Minion job timeout, user ctrl-c).
Additive and backwards-compatible:
- execute() signature widens callbacks to (input, signal?) => ...; existing
two-arg callbacks are structurally compatible and ignore the extra arg.
- opts is optional; callers that omit it get pre-extension behavior.
- Aborts throw a DOM-style AbortError (name='AbortError'), matching what
fetch() throws, so downstream `err.name === 'AbortError'` branches work
unchanged.
- Aborted runs are NOT logged to the failure JSONL — not informative and
would pollute pattern analysis.
Abort check fires in three places:
- Before the deterministic call (pre-flight)
- Between deterministic miss and LLM call (mid-flight)
- Inside llmFallbackFn if the implementation respects signal itself
Smoke tests: 5 scenarios (existing sig, llm fallback, pre-abort, mid-flight
abort, signal threaded to fallback) — all pass. Existing test/fail-improve.test.ts
(13 tests, 27 expects) unchanged and passing.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* feat(resolvers): url_reachable + x_handle_to_tweet — SDK pass 3 (PR 1/5)
Two reference resolver implementations that validate the interface against
real-world requirements: a deterministic free-cost check and a rate-limited
paid-backend lookup.
src/core/resolvers/builtin/url-reachable.ts
HEAD-check a URL, follow redirects (max 5), detect dead links. Reused
isInternalUrl() from the wave-3 SSRF hardening; re-validates every redirect
hop against the same filter. Falls back from HEAD to GET on 405/501.
Composes caller's AbortSignal with a per-request timeout via
AbortSignal.any (with manual-propagation fallback). Confidence=1 when the
backend answers; confidence=0 only on transport failure (DNS/connect/timeout).
src/core/resolvers/builtin/x-api/handle-to-tweet.ts
Find a tweet by handle + free-text keyword hint. Used by the upcoming
`gbrain integrity --auto` loop to repair the 1,424 bare-tweet citations
in Garry's brain. Confidence buckets align with the three-bucket contract:
- >=0.8 auto-repair (single strong match, or dominant in small candidate set)
- 0.5-0.8 review queue (ambiguous but promising)
- <0.5 skip (many candidates or weak match)
Scoring: normalized keyword-token overlap against tweet text, with margin
boost for dominant matches. Strict handle regex (X's username rules).
Retries on 429 up to 2x with Retry-After honor. Terminal 401/403 surfaces
as auth ResolverError so the caller stops hammering. Bearer token read
from ctx.config.x_api_bearer_token or X_API_BEARER_TOKEN env — never logged.
Smoke: registry accepts both, SSRF blocks localhost + file://, available()
returns false when token missing, schema validator rejects bad handles.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* feat(resolvers): tests + gbrain resolvers CLI — SDK pass 4 (PR 1/5 complete)
Closes out PR 1. 43 new tests in test/resolvers.test.ts covering registry
contract, both reference builtins, all three confidence buckets, and every
ResolverError subcode.
test/resolvers.test.ts
- ResolverRegistry: register, duplicate-id rejection, get/has, list with
cost+backend filters, resolve, unavailable propagation, clear, default
singleton lifecycle.
- url_reachable: available(), SSRF guard on localhost + RFC1918 + 169.254
metadata + file:// scheme, empty-url schema error, 200/404 status
propagation, HEAD→GET fallback on 405, redirect chain, per-hop SSRF
re-validation, network failure → reachable=false, AbortSignal mid-flight.
- x_handle_to_tweet: token gate via env AND via ctx.config, invalid/long
handle schema errors, zero-candidate + single-strong + single-weak +
many-ambiguous confidence buckets (gates >=0.5 url emission), 401/403
auth error, 500 upstream error, 429 retry-then-rate_limited, X operator
stripping (prompt injection defense).
src/commands/resolvers.ts
- `gbrain resolvers list [--cost | --backend | --json]` pretty table
or JSON.
- `gbrain resolvers describe <id>` schema + availability detail.
- registerBuiltinResolvers() is idempotent; ready to be called from
future entry points (gbrain integrity, MCP server).
src/cli.ts wires `resolvers` into CLI_ONLY + dispatches to runResolvers.
Full suite: 1343 pass / 0 fail / 141 skip (E2E without DATABASE_URL).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* feat(output): BrainWriter + Scaffolder + SlugRegistry — PR 2 pass 1/4
Lands the transactional writer library that the rest of the Knowledge
Runtime sits on top of. No callers routed through it yet — publish.ts /
backlinks.ts / put_page migrations are pass 4 and PR 2.5.
src/core/output/scaffold.ts
Deterministic URL / citation / link builders. Callers pass typed inputs
(handle + tweetId, account + messageId, slug + display text) and get
canonical markdown bytes out. LLM-generated URLs never touch disk.
- tweetCitation({handle, tweetId, dateISO?})
- emailCitation({account, messageId, subject, dateISO?})
- sourceCitation(resolverResult, {url?, label?})
- entityLink({slug, displayText, relativePrefix?})
- timelineLine({dateISO, summary, citation?})
ScaffoldError with codes for invalid_handle / invalid_tweet_id /
invalid_slug / invalid_message_id / invalid_date / empty.
src/core/output/slug-registry.ts
Solves the "Marc Benioff vs Marc-Benioff both slug to marc-benioff" bug.
create() probes engine.getPage and either returns the desired slug or
disambiguates (alice-smith → alice-smith-2). isFree() + suggestDisambiguators()
for interactive UX. Errors: collision, disambiguator_exhausted, invalid_slug.
src/core/output/writer.ts
BrainWriter.transaction(fn, ctx) wraps engine.transaction. The `fn`
callback receives a WriteTx with createEntity / appendTimeline /
setCompiledTruth / setFrontmatterField / putRawData / addLink (the last
creates both forward + reverse back-link atomically). On commit, per-page
validators run against all touchedSlugs. Strict mode throws on
error-severity findings, rolling back the outer tx. Lint mode (default for
PR 2 rollout) returns the report but commits regardless. Pages with
`validate: false` frontmatter skip validators entirely (grandfather hook
for PR 2 migration).
Integration smoke against PGLite: createEntity → disambiguator (2nd call
with same desired slug), addLink writes both forward + back-link,
strict-mode validator failure rolls back the transaction bit-identically.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* feat(output): 4 pre-commit validators + tests — PR 2 pass 2/4
Lands the validator suite that BrainWriter runs before committing a
transaction. Paragraph-level deterministic checks, markdown-aware, skip
legacy pages via validate:false frontmatter.
src/core/output/validators/citation.ts
Every factual paragraph in compiled_truth carries at least one citation
marker: [Source: ...] or a linked URL. Splits paragraphs on blank lines,
strips fenced code / inline code / HTML comments before checking.
Ignores headings, key-value lines ("**Status:** Active"), table rows,
pure wikilink bullets (## See Also), and short labels without a factual
verb. Deterministic — no LLM, no semantic judgment.
src/core/output/validators/link.ts
Every [text](path) wikilink resolves to a page that exists (unless it's
an external http(s) URL, which this validator doesn't check; that's
url_reachable's job in PR 3). Strips relative prefix and .md extension.
Batches engine.getPage lookups per unique target. mailto/anchor/other
schemes flagged as warning. Links inside fenced code blocks are skipped.
src/core/output/validators/back-link.ts
Iron Law: if page X → page Y, then Y → X. Reads engine.getLinks(ctx.slug),
and for each target checks engine.getLinks(target) for a reverse edge.
Missing reverses flagged as warning (runAutoLink is the authoritative
enforcer on put_page; this is defense-in-depth for pages edited outside
the main write path).
src/core/output/validators/triple-hr.ts
Catches hygiene issues on the compiled_truth / timeline split: bare `---`
in compiled_truth would re-split on round-trip through parseMarkdown;
headings in the timeline section signal authoring mistakes. Both warn
(not error) — legacy pages legitimately use thematic breaks.
src/core/output/validators/index.ts
registerBuiltinValidators(writer) wires all four.
test/writer.test.ts
57 tests: Scaffolder (all 5 helpers + error paths), SlugRegistry (create,
disambiguator, collision throw, invalid-slug, isFree, suggestDisambiguators),
BrainWriter (happy path, disambiguate, addLink + reverse, strict rollback,
lint proceeds with report, off skips validators, validate:false grandfather,
setCompiledTruth, setFrontmatterField merge, registered validators list),
citation validator (all 11 shape cases), link validator (normalizeToSlug
including ../../, external URL skip, mailto warning, code-fence skip),
back-link validator (no outbound, missing reverse → warning, bidirectional
clean), triple-hr validator (clean, bare --- warning, fenced --- skipped,
heading in timeline warning, ## Timeline header allowed).
Full suite: 1400 pass / 0 fail / 141 skip.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* feat(migrations): v0.13.0 grandfather validate:false — PR 2 pass 3/4
Adds the TS migration that makes BrainWriter's strict-mode rollout safe:
every existing page gets `validate: false` in frontmatter so the new
citation / link / back-link / triple-HR validators skip legacy content.
gbrain integrity --auto (PR 3) clears the flag per-page once real citations
are repaired.
src/commands/migrations/v0_13_0_add_validate_false.ts
Four-phase orchestrator following the v0_12_0 pattern:
A. connect — loadConfig + createEngine. Does NOT write config (prior
learning: gbrain init --migrate-only semantics; never
flip Postgres users to PGLite via bare init).
B. snapshot — engine.getAllSlugs() upfront (prior learning:
listpages-pagination-mutation; OFFSET iteration is
self-invalidating when each write bumps updated_at).
C. grandfather — per slug, skip if frontmatter.validate already set,
else append-log pre-mutation snapshot to
~/.gbrain/migrations/v0_13_0-rollback.jsonl and
putPage with validate:false merged in. Batched 100
at a time so interruption losses are bounded.
D. verify — SQL count of pages with validate=false ≥ expectedTouched.
Idempotent: second run is a no-op. Reversible: rollback log is
append-only JSONL; future `gbrain apply-migrations --rollback v0.13.0`
replays it. Safe on empty brains (returns complete with 0 touched).
src/commands/migrations/index.ts
Registers v0_13_0 after v0_12_0 in semver order.
test/migrations-v0_13_0.test.ts
Registry integration (v0.13.0 present, semver-after-v0.12.0, pitch
metadata well-formed), orchestrator handles no-config gracefully,
dryRun skips the connect phase.
test/apply-migrations.test.ts
Updated two assertions that hard-coded the v0.12.0 skippedFuture list
to also include v0.13.0 (now skippedFuture when installed < 0.13.0).
Full suite: 1405 pass / 0 fail / 141 skip.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* feat(integrity): gbrain integrity — bare-tweet repair + dead-link scan (PR 3)
Ships the user-visible milestone for the Knowledge Runtime delta: a
command that finds brain-integrity issues and repairs them through the
BrainWriter + Resolver SDK infrastructure from PRs 1 and 2.
Targets the two quantified pain points from brain/CITATIONS.md:
- 1,424 of 3,115 people pages have bare tweet references without URLs
- An unknown fraction of existing URL citations have rotted
Subcommands:
gbrain integrity check Read-only report, optional --json
gbrain integrity auto Three-bucket repair loop
gbrain integrity review Print review-queue path + count
gbrain integrity reset-progress Clear the progress file
Three-bucket contract (matches x_handle_to_tweet resolver's confidence
scoring):
>=0.8 → auto-repair via BrainWriter transaction. Appends a timeline
entry on the page with a Scaffolder-built tweet citation (URL
from the API response, never from LLM text).
0.5-0.8 → append to ~/.gbrain/integrity-review.md with all candidates
sorted by match score, for batch human review.
<0.5 → log reason to ~/.gbrain/integrity.log.jsonl and skip.
Resumable: every processed slug hits ~/.gbrain/integrity-progress.jsonl
so an interrupted run resumes from the last slug. --fresh clears it.
Bare-tweet detection patterns (regex, deterministic, skip code fences
and already-cited lines):
- "tweeted about"
- "in/on a (recent|viral) tweet"
- "wrote a tweet/post"
- "posted on X"
- "via X" (but not "via X/handle" — already cited)
- possessive "his/her/their tweet"
External-link detection extracts all [text](https?://...) pairs (code
fences skipped) for optional dead-link probing via url_reachable.
Dead links are surfaced, not auto-repaired — no "correct" replacement
exists without human judgment.
Wiring: runIntegrity dispatches subcommands, registers builtin resolvers
into the default registry, connects to the brain engine, and uses
BrainWriter in strict-off mode (integrity is the repair path, not the
write-gate path).
Unit tests: 21 cover bare-tweet regex (all 9 phrase shapes + code-fence
skip + URL-already-present skip + per-line dedup), external-link
extraction (http+https, line numbers, fenced skip), frontmatter handle
extraction (x_handle, twitter, twitter_handle, x; preference order;
leading @ strip; null paths). End-to-end auto flow verified manually
via the resolver SDK tests + BrainWriter tests it composes.
src/cli.ts wires `integrity` into CLI_ONLY + dispatches to runIntegrity.
Full suite: 1426 pass / 0 fail / 141 skip.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* feat(enrichment): BudgetLedger + CompletenessScorer — PR 4
Two layer-2 primitives that slot under the resolver SDK and BrainWriter:
cost-aware spend caps and evidence-weighted per-page completeness scoring.
Schema migration v11 adds two tables:
budget_ledger (scope, resolver_id, local_date) PK — midnight rollover by
date column means a new calendar day upserts a new row; no rollover
thread, no race.
budget_reservations (reservation_id) — TTL-bounded held reservations
(default 60s) so process death between reserve() and commit() doesn't
strand money.
Rollback plan: DROP TABLE. Budget data is regenerable from resolver call
logs; no durable product value lives in the ledger.
src/core/enrichment/budget.ts
BudgetLedger.reserve({resolverId, estimateUsd, capUsd?, ttlSeconds?})
serializes concurrent reserves on {scope, resolver_id, local_date} via
SELECT ... FOR UPDATE. Returns {kind:'held', reservationId, ...} or
{kind:'exhausted', reason, spent, pending, cap} — never over-spends.
commit(id, actualUsd) moves money from reserved_usd to committed_usd and
marks the reservation status='committed'. rollback(id) zeros out the
reservation without touching committed. Commit-after-commit throws
already_finalized; rollback-after-commit is a no-op (callers don't need
to guard). commit-unknown-id throws reservation_not_found.
cleanupExpired() sweeps held reservations past expires_at and rolls them
back; reserve() opportunistically reclaims the target row's expired
reservations before acquiring its own lock.
IANA timezone config via opts.tz (default America/Los_Angeles); midnight
rollover is naturally expressed as a date column + Intl.DateTimeFormat
with en-CA locale (YYYY-MM-DD). DST is handled by the formatter.
src/core/enrichment/completeness.ts
Seven per-type rubrics (person, company, project, deal, concept, source,
media) + default. Each rubric's dimension weights sum to 1.0, checked at
module load. scorePage(page) returns {score, dimensionScores, rubric}
where score is 0.000–1.000.
Person rubric dimensions: has_role_and_company, has_source_urls,
has_timeline_entries, has_citations, has_backlinks, recency_score,
non_redundancy. The last two are the explicit fix for the two pathologies
called out in the codex review of the earlier design: stale pages that
never decay (30-day re-enrich forever) and Wilco-style repeated blocks
that pass Wintermute's length heuristic.
Pure functions. No engine calls — BrainWriter invokes scorePage after a
transaction and caches the result in frontmatter.completeness.
test/enrichment.test.ts — 23 tests:
BudgetLedger: under-cap held, over-cap exhausted, commit moves money,
rollback clears, commit-rollback no-op, commit-commit throws, commit-
unknown throws, invalid input, empty state null, scope isolation,
parallel reserves respect cap (10 parallel, cap 1.0, est 0.3 each →
≤ 3 held; state.reservedUsd ≤ 1.0), cleanupExpired reclaims TTL=0.
CompletenessScorer: all 8 rubrics sum to 1.0, empty person scores <0.3,
fully-enriched person >0.8, dimension scores exposed, role detection,
company/concept/source/media/default routing, recency decay with age,
non_redundancy penalizes repeated lines.
Full suite: 1449 pass / 0 fail / 141 skip.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* feat(minions): quiet-hours + stagger + claim-time gate — PR 5
Closes the scheduler gap per CEO plan: Minions v7 shipped a durable
runtime but nothing about when jobs should NOT run. This wires
quiet-hours enforcement at claim time (the codex correction — dispatch-
time is wrong because a queued job can become claimable after its window
opens) plus deterministic stagger slots to prevent cron-boundary storms.
Schema migration v12 adds two columns to minion_jobs:
quiet_hours JSONB — {start, end, tz, policy} window config
stagger_key TEXT — partitioning key for deterministic offset
Plus a partial index on stagger_key for later slot-assignment queries.
src/core/minions/quiet-hours.ts
evaluateQuietHours(cfg, now?) → 'allow' | 'skip' | 'defer'. Pure,
deterministic, no engine. Handles straight-line and wrap-around windows
(e.g. 22→7 spans midnight). IANA timezone via Intl.DateTimeFormat;
unknown tz fails open (allow) — safer than hard-blocking every job.
'skip' policy drops the event; 'defer' (default) re-queues for later.
src/core/minions/stagger.ts
staggerMinuteOffset(key) → 0–59, FNV-1a hash. Same key → same slot.
Pure; no module-level state. Used by scheduled resolvers that want to
avoid cron-boundary collisions ("10 jobs all fire at minute 0").
src/core/minions/worker.ts
MinionWorker.tick now consults evaluateQuietHours on every claimed job.
Verdict 'defer' → UPDATE status='delayed', delay_until = now() + 15m
(prevents immediate re-claim loops when the claim query re-runs).
Verdict 'skip' → UPDATE status='cancelled', error_text='skipped_quiet_hours'.
Both paths clear lock_token and require lock_token match in the WHERE
clause so a concurrent stall recovery can't race us.
test/minions-quiet-hours.test.ts — 25 tests:
evaluateQuietHours: null/undefined/invalid config paths (allow fail-open),
straight-line in/out + exclusive-end, wrap-around in (before midnight +
after), skip vs defer policy, timezone-offset propagation (winter PST
vs summer PDT), localHour parity with Date.getUTCHours.
staggerMinuteOffset: deterministic same key → same offset, different
keys spread across buckets (10 keys → ≥5 unique buckets), empty/non-
string edge cases.
Schema v12: quiet_hours and stagger_key columns exist on minion_jobs,
idx_minion_jobs_stagger_key index present.
Full suite: 1474 pass / 0 fail / 141 skip.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* feat(output): post-write validator lint hook — PR 2.5
Minimal integration of BrainWriter validators into the main write path,
feature-flag-gated and non-blocking. The CEO plan explicitly scoped PR 2.5
as a pre-soak landing step: the hook plugs in now, observability lands,
but strict-mode rejection is deferred to a follow-on release gated on the
7-day soak + BrainBench regression ≤1pt.
src/core/output/post-write.ts
runPostWriteLint(engine, slug, opts?) invokes the four BrainWriter
validators (citation, link, back-link, triple-hr) against a freshly
written page and returns a PostWriteLintResult. Skips cleanly when:
- config `writer.lint_on_put_page` is not truthy (default OFF; opts.force overrides)
- the page is not found (shouldn't happen in normal put_page flow)
- the page has frontmatter.validate === false (grandfathered)
Findings are logged to:
- ~/.gbrain/validator-lint.jsonl (capped at 20 findings per line)
- engine.logIngest (ingest_log table) for durable agent-inspectable history
Validator-level exceptions are swallowed so a buggy validator never
breaks put_page.
src/core/operations.ts put_page handler
After importFromContent + runAutoLink, imports runPostWriteLint and
invokes it. Result returns writer_lint: {error_count, warning_count} or
{skipped: reason}. Try/catch wraps the whole hook so an import or
runtime error never blocks the main write.
Enable locally:
gbrain config set writer.lint_on_put_page true
Then every put_page emits a writer_lint summary + appends structured
findings to the ingest log for analysis before the strict-mode flip.
test/post-write-lint.test.ts — 11 tests:
Flag reader (default off, true/1/on, other values false, explicit false)
Hook behavior (flag-off skip, page-not-found skip, validate:false
grandfather skip, force=true overrides flag, dirty page yields citation
error, clean page yields zero findings).
Full suite: 1485 pass / 0 fail / 141 skip.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* test(migrations-v0_13_0): drop flaky no-config assertion
The 'does not succeed when no brain is configured' test assumed loadConfig
would return null when HOME is empty, but it also reads DATABASE_URL from
the environment. When .env.testing sources DATABASE_URL into the shell
(normal E2E lifecycle), the orchestrator connects successfully and runs
to completion — the test's assertion was unreachable.
The dry-run path is still covered by the remaining test in the same
describe block; registry integration and semver ordering are covered by
the sibling describe.
Full suite with DATABASE_URL live: 1574 pass / 0 fail.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* fix(minions): wire quiet_hours + stagger_key into MinionJobInput + queue.add
Codex adversarial review caught that PR 5 (claim-time quiet-hours gate) was
cosmetic: the schema v12 column existed, the worker read it via
`readQuietHoursConfig(job)`, but `MinionJobInput` never accepted it,
`queue.add()` never inserted it, and `rowToMinionJob()` never mapped it out.
Result: every scheduled job saw `quiet_hours: null`, so the gate was a
no-op. Stagger_key had the same broken wiring.
- MinionJob (types.ts): add `quiet_hours` and `stagger_key` fields.
- MinionJobInput: add matching optional fields so callers can submit them.
- rowToMinionJob: parse both columns (JSONB handled the same way as `data`).
- MinionQueue.add: include both columns in the INSERT (idempotent + normal
paths), bound as $19/$20. The `$19::jsonb` cast matches the JSONB column
shape; the wire format is the same native-JS object path that fixed the
JSONB double-encode bug in v0.12.1.
After this, `await queue.add('x', {}, { quiet_hours: {start:22,end:7,
tz:"America/Los_Angeles",policy:"defer"} })` actually stores the window
and the worker's claim-time gate defers the job inside it.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* fix(minions): route quiet-hours 'skip' through cancelJob to rollup parents
Codex flagged that handleQuietHoursDefer with verdict='skip' directly set
status='cancelled' via raw UPDATE — bypassing MinionQueue.cancelJob, which
means:
- Parent jobs in 'waiting-children' never get rolled up.
- Descendant jobs don't cascade-cancel.
- Child-done inbox notification is skipped.
Result: a parent waiting on a child that fell inside quiet hours with
policy='skip' stays stuck forever.
Fix: release the lock, then delegate to queue.cancelJob(job.id) which
handles the recursive CTE + parent rollup + inbox posting correctly.
Falls back to a direct UPDATE only if cancelJob errors — even then, the
status transition is status-guarded to avoid stomping terminal states.
Defer path unchanged (no parent rollup needed since the job hasn't reached
a terminal state).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* fix(budget): commit() re-checks cap + rejects negative actuals
Codex caught two cap-bypass bugs in BudgetLedger.commit():
1. reserve({estimateUsd: 0.01, capUsd: 1.0}) + commit(id, 100) silently
charged $100 to a $1-cap bucket. Cap is an advertised invariant that
the code was not enforcing.
2. Negative actuals (commit(id, -5)) were accepted, letting callers
artificially reduce committed_usd below the real spend. Refunds need
a dedicated API, not a side-channel on commit.
Fix:
- Reject non-finite AND negative actualUsd at entrypoint.
- Lock the ledger row FOR UPDATE during commit (same serialization as
reserve).
- Compute effective cap headroom = cap - other_committed - other_reserved
(excluding this reservation from the reserved pool since we're about to
finalize it).
- When actualUsd would exceed available, clamp committed_usd to max
available and throw BudgetError with the overage reported. The
reservation is still marked 'committed' (API call already happened;
don't retry-loop), but the cap is honored.
After this, a $1/day cap actually means $1/day.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* fix(integrity): --dry-run no longer writes progress, poisoning resume
Codex caught that 'gbrain integrity auto --dry-run' appended progress
entries (status='repaired', 'reviewed', 'skipped', 'error') despite doing
no actual writes. The follow-on real run with default --resume would then
skip those slugs — the dry-run silently consumed the work queue.
Fix: gate every appendProgress() call in cmdAuto on !dryRun. Dry-run
still logs to the skip log / review queue (so the user sees what WOULD
happen), but the progress file stays untouched.
Behavior:
--dry-run → buckets counted + summary printed + review-queue
+ log populated, but progress file unchanged.
(default) → progress file tracks every processed slug, so
Ctrl-C + re-run resumes from the right place.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* chore: bump version and changelog (v0.13.0.0)
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* fix(resolvers): DNS-rebinding defense + X rate-limit header parity
Two non-blocking codex findings on PR #210 rolled into one bisectable
commit because their tests share an import line.
url_reachable: hostname-string SSRF guard is vulnerable to DNS rebinding
(attacker-controlled DNS returns a public IP at validate time and
169.254.169.254 at fetch time). Add checkDnsRebinding() that resolves
the hostname via dns.lookup({all:true}) and rejects any result whose
A/AAAA record lands in a private range (v4 via isPrivateIpv4, v6
loopback/link-local/unique-local/IPv4-mapped). Applied on the initial
URL and on every redirect target. Null on DNS failure so genuine
network problems surface via fetch.
x_handle_to_tweet: rate-limit backoff only honored Retry-After and
ignored X's proprietary x-rate-limit-reset header. computeBackoffMs()
parses both (Retry-After = seconds or HTTP-date; x-rate-limit-reset =
epoch seconds), takes MAX, and clamps to [2s, 60s]. Exported for
testability; callers use it uniformly on every 429.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* fix(writer): advisory lock on desiredSlug prevents cross-process TOCTOU
BrainWriter's createEntity checks engine.getPage(slug) and falls back
to putPage(), which upserts. Two putPage('people/alice') calls from
separate processes (a Claude Code session + a Minions worker, say) can
both read "free" from SlugRegistry and both call putPage, silently
overwriting each other with no disambiguation.
Take a transaction-scoped advisory lock keyed on hashtext(desiredSlug)
before the registry check. Concurrent writers for the same slug now
serialize at the DB level: the second observes the first's commit and
disambiguates to alice-2. PGLite is single-process so this is a
harmless no-op there. Wrapped in try/catch so engines/test doubles
that don't support advisory locks fall through to the existing
within-process check.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* fix(validators): empty [Source:] no longer satisfies citation check
Regex /\[Source:[^\]]*\]/ matched decorative markers like [Source:]
and [Source: ] that carry zero provenance. Tighten to require at
least one non-whitespace character before the closing bracket. The
inline URL form ](https://...) already requires a scheme+host so it
stays as-is.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* fix(auto-link): advisory lock serializes concurrent reconciliation
runAutoLink wraps getLinks + addLink/removeLink in a transaction, but
row-level locks alone don't prevent the union-of-writes race: two
concurrent put_page calls on the same slug can both read the same
existingKeys BEFORE either mutates a row, then proceed to add links
the other side's rewrite no longer mentions.
Take a transaction-scoped advisory lock on hashtext("auto_link:" ||
slug) at the start of the reconciliation. Concurrent writers on the
same slug now fully serialize; writers on different slugs still run
in parallel. No-op on engines without advisory locks (PGLite).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* test: expand coverage on abort-signal threading + integrity CLI dispatch
fail-improve: four new AbortSignal cases — pre-start abort, between
deterministic and LLM, signal forwarded into both callbacks, and
LLM-thrown AbortError propagates without logging a failure entry.
integrity: three new CLI dispatch cases — --help, no-subcommand (help),
and unknown subcommand (stderr + exit 1). Non-engine paths so they
exercise routing without spinning up a DB.
Coverage-only; no source changes.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* feat(doctor): fold integrity sample scan into default health check
Expose scanIntegrity(engine, opts) as a pure library function — same
logic cmdCheck uses — and call it from doctor in non-fast mode with
a 500-page sampling limit. Surfaces bare-tweet phrase count and
external-link count as an 'integrity' check, warn-status when bare
tweets are present with a one-liner pointing at 'gbrain integrity
check' for the full report and 'integrity auto' for repair.
Read-only: no network, no writes, no resolver calls. Pages with
validate:false frontmatter are skipped (grandfathered). --fast mode
skips it entirely so the existing health-snapshot contract holds.
Users no longer need to remember three separate commands (doctor,
lint, integrity check) to audit brain health — doctor surfaces the
integrity signal by default, full scan stays available for deep dives.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* feat(put_page): auto-extract timeline entries alongside auto-link
put_page already chunks, embeds, reconciles tags, and extracts
auto-links on every write. Timeline extraction has lived in a
separate command (gbrain extract timeline) that users had to remember
to run. Fold it into the write path: after the page commits, parse
timeline entries from compiled_truth + timeline body and insert via
addTimelineEntriesBatch. ON CONFLICT DO NOTHING keeps it idempotent
across re-writes.
Mirrors auto-link shape: best-effort post-hook, skipped for remote
(MCP) callers, gated by auto_timeline config (default TRUE). Response
includes auto_timeline: { created } alongside auto_links.
Side effect: a one-shot `gbrain put` now produces a complete page —
chunks, embeddings, links, AND timeline — instead of three commands
the user has to chain manually.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* feat(migrate): verify target health after engine migration
After a PGLite↔Postgres migration, the user was left to run 'gbrain
doctor' themselves to confirm the target is good. Not great, because
the failure modes (partial copy, missing embeddings, schema drift)
all surface at next CLI use when the migration itself looks like it
succeeded.
Add verifyTarget() — inline doctor-lite that checks page count
matches the source, embedding coverage is above 90%, and schema
version is at latest. Prints a 3-line status table at the end of
migrate and points at 'gbrain doctor' for the full check. Non-fatal:
warns on discrepancies instead of failing the command so the user
sees the full picture.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* docs(bench): add v0.13 knowledge runtime benchmark deltas
Two new benchmark scripts + one consolidated markdown comparing this
branch against master (
|
||
|
|
c22ca84772 |
feat: v0.13 frontmatter relationship indexing — YAML becomes typed graph edges (#231)
* feat(schema): links provenance + engine plumbing (v0.13)
Adds link_source, origin_page_id, origin_field columns with
UNIQUE NULLS NOT DISTINCT constraint + CHECK constraint. New indexes
on link_source + origin_page_id.
migrate.ts v11 handles idempotent upgrade path for existing brains.
Both engines: addLink/addLinksBatch threads new columns (4→7 col
unnest). removeLink gains linkSource filter. getLinks/getBacklinks
return new columns.
New engine method findByTitleFuzzy(name, dirPrefix?, minSim?) uses
pg_trgm % operator + similarity(). Drives the v0.13 resolver's
fuzzy-match step with zero LLM/embedding cost.
* feat(graph): frontmatter edge extraction + slug resolver (v0.13)
Canonical FRONTMATTER_LINK_MAP: field → type + direction + dir-hint
for 10 frontmatter patterns (company/companies, key_people, investors,
attendees, partner, lead, founded, sources, source, related/see_also).
Direction semantics: "incoming" means resolved value is the FROM side
so subject-of-verb reads naturally (pedro → meeting, not backwards).
makeResolver(engine, {mode}) — two-mode resolver:
batch (migration): slug → dir-hint → pg_trgm. NEVER hits search.
live (put_page): + optional search fallback with expand=false
(dodges hidden Haiku per operations-query learning).
Per-run cache: same name → single DB lookup.
extractFrontmatterLinks handles arrays-of-objects (investors:
[{name: 'Sequoia', role: 'lead'}]), skips bad types silently,
tracks unresolved names for the summary report.
extractPageLinks is now async. LinkCandidate gains fromSlug,
linkSource, originSlug, originField. Returns {candidates, unresolved}.
22 new tests: field-map coverage, direction semantics, source vs
sources, resolver fallback chain (batch + live), cache hit, bad
types skipped, context enrichment, FRONTMATTER_LINK_MAP integrity.
* feat(auto-link): bidirectional reconciliation + unresolved response
put_page auto-link post-hook now handles incoming-direction frontmatter
edges. Reconciliation splits candidates into out (fromSlug === slug)
and in (fromSlug !== slug — frontmatter fields like key_people on a
company page emit person → company edges).
Safe reconciliation via origin_page_id scoping: we only touch
link_source='frontmatter' edges where origin_slug = the page being
written. Markdown + manual edges survive untouched. Edges created
by OTHER pages' frontmatter also survive.
put_page response extends auto_links with unresolved: Array<{field,
name}>. Agents writing attendees: [Pedro, Alex] where Alex doesn't
resolve see it in the response and can queue for enrichment.
Additive — existing agents unaffected.
extract.ts: delete the local 5-field extractFrontmatterLinks + local
inferLinkType. FS-source now calls canonical link-extraction.ts via
a synthetic resolver backed by the allSlugs Set. --include-frontmatter
flag (default OFF in v0.13 for back-compat; migration explicitly
enables for the one-time backfill). Top-20 unresolved names summary
when active.
* feat(migration): v0.13.0 orchestrator
3-phase orchestrator (schema → backfill → verify → record) follows
the v0_12_2.ts pattern. Phase A triggers migrate.ts v11 via
gbrain init --migrate-only. Phase B runs:
gbrain extract links --source db --include-frontmatter
to backfill frontmatter edges for every existing page. Uses the
batch-mode resolver (pg_trgm only, no LLM calls, zero API cost).
Ignores auto_link=false config — migration is canonical, the
auto_link flag controls per-write post-hook not one-time schema
work.
Idempotent + resumable via ON CONFLICT DO NOTHING + origin_page_id
scoping. Wall-clock budget: 2-5 min on 46K-page brains.
Registered in migrations/index.ts. apply-migrations test updated
to include v0.13.0 in skippedFuture for older installed versions.
* feat(release): upgrade-errors.jsonl trail + doctor surfacing
upgrade.ts catches post-upgrade subprocess failures as best-effort
today (line 65 comment: "post-upgrade is best-effort, don't fail
the upgrade"). When that chain silently fails, users end up with
half-upgraded brains and no signal.
v0.13: on post-upgrade failure, append a structured record to
~/.gbrain/upgrade-errors.jsonl with ts, phase, versions, error
message, and a paste-ready recovery hint.
doctor.ts reads the jsonl and surfaces the latest entry with a
warn-status check. User runs gbrain doctor, sees exactly what
failed, pastes the recovery command, files an issue if needed.
Applies to every future release — doctor grows with the codebase
without per-release edits. The CHANGELOG pattern ("To take advantage
of v[version]" block) mirrors this in user-facing form.
* chore: bump version and changelog (v0.13.0)
v0.13.0 — Frontmatter Relationship Indexing.
Adds the "To take advantage of v[version]" block pattern to
CHANGELOG format (CLAUDE.md documents the requirement going
forward). Pairs with the upgrade-errors.jsonl + doctor surfacing
to close the "half-upgraded brain, no signal" loop.
UPGRADING_DOWNSTREAM_AGENTS.md gets a v0.13 section: no-action-
required verdict for most skills, optional diffs for meeting-
ingestion / enrich / idea-ingest if they want to consume
auto_links.unresolved.
skills/migrations/v0.13.0.md is the user-facing upgrade skill.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* fix(v0.13): adversarial review P0s
Codex + Claude adversarial review caught 4 critical issues in the
v0.13 implementation. Fixing before ship.
1. findByTitleFuzzy SET LOCAL was a no-op. postgres.js auto-commits
each sql`` so SET LOCAL pg_trgm.similarity_threshold committed
before the `%` operator ran against it. Resolver used server
default (0.3, not 0.55) → way too many fuzzy matches, wrong
links on a 46K-page brain. Switched to inline
`similarity(title, $1) >= $N` which has no transaction scoping.
Added `ORDER BY sim DESC, slug ASC` for deterministic
tie-breaking (prevents reconciliation churn on re-runs).
2. v11 migration now checks Postgres ≥ 15 before applying
UNIQUE NULLS NOT DISTINCT. Old Supabase projects on PG14 would
have dropped the old unique constraint and failed to add the
new one, corrupting the uniqueness invariant. The check raises
a clear error with the actual PG version, leaving the old
constraint in place.
3. v11 migration now backfills NULL link_source → 'markdown' for
pre-v0.13 legacy rows. Without this, reconciliation's existKey
comparison treats NULL and 'markdown' as equivalent but the
unique constraint sees them as distinct (NULLS NOT DISTINCT
only collapses NULL with NULL, not NULL with 'markdown'). Result
was duplicate edges accumulating forever. Treating legacy as
markdown is the accurate best-guess — pre-v0.13 auto-link only
emitted markdown edges.
4. v0_13_0.ts orchestrator now uses process.execPath, not a bare
`gbrain` on PATH. After `gbrain upgrade` rewrites the binary,
alias shadowing / PATH caching / multiple installs could
resolve a stale `gbrain` binary. process.execPath is always
the binary that loaded this migration module.
Phase C verify clarified: reports page + link counts and points to
Phase B's own stdout as the authoritative signal for backfill
results (extract.ts already prints `Links: created N from M pages`).
* docs: scrub real names from public docs + add privacy rule to CLAUDE.md
Public artifacts (CHANGELOG, skills, docs) should never reveal real
contacts, companies, funds, or private agent-fork names from any
user's brain. When a doc copies a query like `gbrain graph diana-hu`
or names a fork like `Wintermute`, that real name gets indexed,
cross-referenced, and distributed with every release.
CLAUDE.md gains a "Privacy rule: scrub real names from public docs"
section with:
- What counts as public (CHANGELOG, README, docs/, skills/, PR bodies,
commit messages, code comments)
- Name mapping table (agent forks → your agent fork; example person →
alice-example; example fund → fund-a; etc.)
- Distinction between illustrative API examples with household brands
(Stripe, Brex) and queries that reveal real relationships
Applied the rule to v0.13 scope:
- CHANGELOG v0.13 entry: Pedro/Diana/Wintermute/Sequoia/Benchmark/a16z
all replaced with alice/charlie/fund-a/acme/agent-fork placeholders
- skills/migrations/v0.13.0.md: same
- docs/UPGRADING_DOWNSTREAM_AGENTS.md: Wintermute references scrubbed
throughout (pre-v0.13 and v0.13 sections)
- CLAUDE.md: "Brain skills (from Wintermute)" → "(ported from an
upstream agent fork)", internal Wintermute provenance notes
genericized, "Garry finds fragile upgrade paths" → "the gbrain
maintainers find fragile upgrade paths" in the template
Pre-v0.13 historical CHANGELOG entries (v0.10-v0.12) left alone —
those are shipped releases; rewriting changes public history.
---------
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
|
||
|
|
013b348c28 |
v0.12.3: Reliability wave — sync deadlock, search timeout scoping, wikilinks, orphans (#216)
* fix(sync): remove nested transaction that deadlocks > 10 file syncs sync.ts wraps the add/modify loop in engine.transaction(), and each importFromContent inside opens another one. PGLite's _runExclusiveTransaction is a non-reentrant mutex — the second call queues on the mutex the first is holding, and the process hangs forever in ep_poll. Reproduced with a 15-file commit: unpatched hangs, patched runs in 3.4s. Fix drops the outer wrap; per-file atomicity is correct anyway (one file's failure should not roll back the others). (cherry picked from commit |
||
|
Garry Tan
|
d0d3cf078a |
Merge remote-tracking branch 'origin/master' into garrytan/gbrain-evals
# Conflicts: # package.json |
||
|
Garry Tan
|
679a3f9a3c |
docs(bench): 2026-04-19 gbrain v0.11.1 vs v0.12.1 regression comparison
Runs the same eval harness against two gbrain src/ trees on the same 240-page corpus and 145 queries. Patches the v0.11 copy's gbrain-after adapter to use getLinks/getBacklinks (v0.11 has no traversePaths) with identical direction+linkType semantics. gbrain-after P@5 22.1% -> 49.1% (+27 pts); R@5 54.6% -> 97.9% (+43 pts); correct-in-top-5 99 -> 248 (+149). hybrid-nograph flat at 17.8% / 65.1% on both (v0.12 didn't touch hybridSearch / chunking). Driver is extraction quality, not graph presence: v0.12 emits 499 typed links (v0.11: 136, x3.7) and 2,208 timeline entries (v0.11: 27, x82) on the same 240 pages. Sharpens the April-18 "graph layer does the work" claim -- on v0.11 that architecture only beat hybrid-nograph by 4.3 points; the 31-point lead in the multi-adapter scorecard comes from graph + high-quality extract in combination. |
||
|
Garry Tan
|
4c885d01d9 |
docs(bench): 2026-04-19 multi-adapter scorecard
Reproducibility run of the 4-adapter side-by-side at commit
|
||
|
Garry Tan
|
c6b308a1c8 |
fix(eval): teardown PGLite engines so bun run eval:run exits 0
The multi-adapter runner left PGLite engines alive after each run. GbrainAfterAdapter and HybridNoGraphAdapter both instantiate a PGLiteEngine in init() but never disconnect it; Bun's shutdown path exits with code 99 when embedded-Postgres workers outlive main(). Added optional `teardown?(state)` to the Adapter interface, implemented it on both engine-backed adapters, and call it from scoreOneRun after the N=5 loop. ripgrep-bm25 and vector-only hold no DB resources and don't need a teardown. Verified: gbrain-after, hybrid-nograph, ripgrep-bm25, vector-only all exit 0 at N=1. Full test:eval passes (57 tests). No metric change. |
||
|
|
c0b621923b |
fix: JSONB double-encode + splitBody wiki + parseEmbedding (v0.12.1) (#196)
* fix: splitBody and inferType for wiki-style markdown content - splitBody now requires explicit timeline sentinel (<!-- timeline -->, --- timeline ---, or --- directly before ## Timeline / ## History). A bare --- in body text is a markdown horizontal rule, not a separator. This fixes the 83% content truncation @knee5 reported on a 1,991-article wiki where 4,856 of 6,680 wikilinks were lost. - serializeMarkdown emits <!-- timeline --> sentinel for round-trip stability. - inferType extended with /writing/, /wiki/analysis/, /wiki/guides/, /wiki/hardware/, /wiki/architecture/, /wiki/concepts/. Path order is most-specific-first so projects/blog/writing/essay.md → writing, not project. - PageType union extended: writing, analysis, guide, hardware, architecture. Updates test/import-file.test.ts to use the new sentinel. Co-Authored-By: @knee5 (PR #187) Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * fix: JSONB double-encode bug on Postgres + parseEmbedding NaN scores Two related Postgres-string-typed-data bugs that PGLite hid: 1. JSONB double-encode (postgres-engine.ts:107,668,846 + files.ts:254): ${JSON.stringify(value)}::jsonb in postgres.js v3 stringified again on the wire, storing JSONB columns as quoted string literals. Every frontmatter->>'key' returned NULL on Postgres-backed brains; GIN indexes were inert. Switched to sql.json(value), which is the postgres.js-native JSONB encoder (Parameter with OID 3802). Affected columns: pages.frontmatter, raw_data.data, ingest_log.pages_updated, files.metadata. page_versions.frontmatter is downstream via INSERT...SELECT and propagates the fix. 2. pgvector embeddings returning as strings (utils.ts): getEmbeddingsByChunkIds returned "[0.1,0.2,...]" instead of Float32Array on Supabase, producing [NaN] cosine scores. Adds parseEmbedding() helper handling Float32Array, numeric arrays, and pgvector string format. Throws loud on malformed vectors (per Codex's no-silent-NaN requirement); returns null for non-vector strings (treated as "no embedding here"). rowToChunk delegates to parseEmbedding. E2E regression test at test/e2e/postgres-jsonb.test.ts asserts jsonb_typeof = 'object' AND col->>'k' returns expected scalar across all 5 affected columns — the test that should have caught the original bug. Runs in CI via the existing pgvector service. Co-Authored-By: @knee5 (PR #187 — JSONB triple-fix) Co-Authored-By: @leonardsellem (PR #175 — parseEmbedding) Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * feat: extract wikilink syntax with ancestor-search slug resolution extractMarkdownLinks now handles [[page]] and [[page|Display Text]] alongside standard [text](page.md). For wiki KBs where authors omit leading ../ (thinking in wiki-root-relative terms), resolveSlug walks ancestor directories until it finds a matching slug. Without this, wikilinks under tech/wiki/analysis/ targeting [[../../finance/wiki/concepts/foo]] silently dangled when the correct relative depth was 3 × ../ instead of 2. Co-Authored-By: @knee5 (PR #187) Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * feat: gbrain repair-jsonb + v0.12.1 migration + CI grep guard - New gbrain repair-jsonb command. Detects rows where jsonb_typeof(col) = 'string' and rewrites them via (col #>> '{}')::jsonb across 5 affected columns: pages.frontmatter, raw_data.data, ingest_log.pages_updated, files.metadata, page_versions.frontmatter. Idempotent — re-running is a no-op. PGLite engines short-circuit cleanly (the bug never affected the parameterized encode path PGLite uses). --dry-run shows what would be repaired; --json for scripting. - New v0_12_1.ts migration orchestrator. Phases: schema → repair → verify. Modeled on v0_12_0 pattern, registered in migrations/index.ts. Runs automatically via gbrain upgrade / apply-migrations. - CI grep guard at scripts/check-jsonb-pattern.sh fails the build if anyone reintroduces the ${JSON.stringify(x)}::jsonb interpolation pattern. Wired into bun test via package.json. Best-effort static analysis (multi-line and helper-wrapped variants are caught by the E2E round-trip test instead). - Updates apply-migrations.test.ts expectations to account for the new v0.12.1 entry in the registry. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * chore: bump version and changelog (v0.12.1) Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * docs: update project documentation for v0.12.1 - CLAUDE.md: document repair-jsonb command, v0_12_1 migration, splitBody sentinel contract, inferType wiki subtypes, CI grep guard, new test files (repair-jsonb, migrations-v0_12_1, markdown) - README.md: add gbrain repair-jsonb to ADMIN command reference - INSTALL_FOR_AGENTS.md: fix verification count (6 -> 7), add v0.12.1 upgrade guidance for Postgres brains - docs/GBRAIN_VERIFY.md: add check #8 for JSONB integrity on Postgres-backed brains - docs/UPGRADING_DOWNSTREAM_AGENTS.md: add v0.12.1 section with migration steps, splitBody contract, wiki subtype inference - skills/migrate/SKILL.md: document native wikilink extraction via gbrain extract links (v0.12.1+) Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com> |
||
|
|
699db50a3d |
fix(extract+migrate): kill N+1 hang + v0.12.0 migration timeout (v0.12.1) (#198)
* feat(engine): add addLinksBatch + addTimelineEntriesBatch via unnest()
Multi-row INSERT...SELECT FROM unnest() JOIN pages ON CONFLICT DO NOTHING
RETURNING 1. 4 array-typed bound parameters (links) or 5 (timeline)
regardless of batch size, sidesteps Postgres's 65535-parameter cap.
Returns count of rows actually inserted (excluding ON CONFLICT no-ops
and JOIN-dropped rows whose slugs don't exist).
Per-row addLink / addTimelineEntry signatures and SQL behavior unchanged.
All 10 existing call sites compile and behave identically.
Tests: 11 PGLite cases (empty batch, missing optionals, within-batch dedup,
JOIN drops missing slug, half-existing batch, batch of 100) + 9 E2E
postgres-engine cases against real Postgres+pgvector.
* fix(migrate): pre-create btree helper in v8 + v9 dedup; bump phaseASchema timeout
Production bug: v0.12.0 schema migration timed out at Supabase Management API's
60s ceiling on brains with 80K+ duplicate timeline rows. The DELETE...USING
self-join was O(n²) without an index on the dedup columns.
Fix: pre-create idx_links_dedup_helper / idx_timeline_dedup_helper on the
dedup columns BEFORE the DELETE, drop after. Turns O(n²) into O(n log n).
On 80K+ rows the migration completes in <1s instead of timing out.
Also bumps the v0.12.0 orchestrator's phaseASchema timeout 60s -> 600s as
belt-and-suspenders for unforeseen slowness.
Exports MIGRATIONS for structural test assertions.
Tests: 2 structural assertions (helper-index DDL must appear in v8/v9 SQL
in the right order — catches regression even at 0-row scale) + 2 behavioral
regression tests (1000-row dedup completes <5s).
* perf(extract): kill N+1 dedup pre-load; switch to batched writes
Production bug: gbrain extract hung 10+ minutes producing zero output on
47K-page brains. The pre-load loop called engine.getLinks(slug) (or
getTimeline) once per page across engine.listPages({limit: 100000}) — 47K
serial round-trips over the Supabase pooler before the first file was read.
Both engines already enforced uniqueness at the SQL layer
(UNIQUE(from, to, link_type) on links, idx_timeline_dedup on timeline_entries).
The in-memory dedup Set was redundant insurance that became the bottleneck.
Fix: delete the pre-load entirely. Buffer 100 candidates per file walk,
flush via engine.addLinksBatch / engine.addTimelineEntriesBatch. ~99% fewer
DB round-trips per re-extract.
Also fixes counter accuracy: 'created' now counts rows actually inserted
(via batch RETURNING 1 row count). Re-run on a fully-extracted brain
prints 'Done: 0 links' instead of lying.
Dry-run mode keeps a per-run dedup Set so duplicate candidates from N
markdown files print exactly once, not N times.
Batch errors are visible in BOTH json and human modes — silent loss of
100 rows is worse than per-row error visibility.
Tests: extract-fs.test.ts (idempotency + truthful counter + dry-run dedup
+ perf regression guard <2s).
* chore: bump version + changelog (v0.12.1)
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* docs: update CLAUDE.md for v0.12.1 (batch engine API, test counts)
Reflect what shipped in v0.12.1:
- New engine methods addLinksBatch + addTimelineEntriesBatch (PGLite via
unnest() + manual $N, postgres-engine via INSERT...SELECT FROM
unnest($1::text[], ...) JOIN pages ON CONFLICT DO NOTHING).
- extract.ts no longer pre-loads dedup set; candidates are buffered 100
at a time and flushed via the new batch methods.
- v0.12.0 orchestrator phaseASchema timeout bumped 60s to 600s.
- Test counts 1297 unit / 105 E2E to 1412 unit / 119 E2E.
- New test/extract-fs.test.ts covers the N+1 regression guard.
- BrainEngine method count 37/38 to 40.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
|
||
|
Garry TanandClaude Opus 4.7
|
b81373d4c2 |
docs(eval): Phase 3 contributor docs + CI workflow for eval/ tests
Ships the contributor-onboarding surface promised in the plan. With this
commit, external researchers have a self-serve path from clone to PR in
under 5 minutes.
Added:
eval/README.md — 5-minute quickstart,
directory map, methodology
one-pager, adapter scorecard
eval/CONTRIBUTING.md — three contributor paths:
1. Write Tier 5.5 queries
2. Submit an external adapter
3. Reproduce a scorecard
eval/RUNBOOK.md — operational troubleshooting:
generation failures, runner
failures, query validation,
world.html rendering, CI
eval/CREDITS.md — contributor attribution
(synthetic-outsider-v1 labeled
as placeholder; real submissions
land here)
.github/PULL_REQUEST_TEMPLATE/tier5-queries.md — structured PR template
for Tier 5.5 submissions
.github/workflows/eval-tests.yml — CI: validates queries,
runs all eval unit tests,
renders world.html on every PR
touching eval/** or
src/core/link-extraction.ts
CI scope (intentionally narrow):
- Triggers on paths: eval/**, src/core/link-extraction.ts, src/core/search/**
- Runs: bun run eval:query:validate (80 queries), test:eval (57 tests),
eval:world:render (smoke-test the HTML renderer)
- Pinned actions by commit SHA (matches existing .github/workflows/test.yml)
- Zero API calls — all Opus/OpenAI paths stubbed or skipped in unit tests
- Fast: ~30s total wall clock
Contributor TTHW (clone → first merged PR):
- Path 1 (Tier 5.5 queries): ~5 min
- Path 2 (external adapter): ~30 min for a simple adapter
- Path 3 (reproduce scorecard): ~15 min wall clock (N=5 run)
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
|
||
|
Garry TanandClaude Opus 4.7
|
f0649e2f32 |
feat(eval): Phase 3 world.html explorer + eval:* CLI surface
Contributor DX magical moment. Static HTML explorer renders the full
canonical world (240 entities) as an explorable tree, opens in any browser,
zero install. Every string HTML-entity-encoded (XSS-safe — direct vuln
class per eng pass 2, confidence 9/10).
Added:
eval/generators/world-html.ts — renderer (~240 LOC; single-file
HTML with inline CSS + minimal JS)
eval/generators/world-html.test.ts — 16 tests (XSS + rendering correctness)
eval/cli/world-view.ts — render + open in default browser
eval/cli/query-validate.ts — CLI wrapper for queries/validator
eval/cli/query-new.ts — scaffold a query template
Modified:
package.json — 7 new eval:* scripts
.gitignore — ignore generated world.html
package.json scripts shipped:
bun run test:eval all eval unit tests (57 pass)
bun run eval:run full 4-adapter N=5 side-by-side
bun run eval:run:dev N=1 fast dev iteration
bun run eval:world:view render world.html + open in browser
bun run eval:world:render render only (CI-friendly, --no-open)
bun run eval:query:validate validate built-in T5+T5.5 (or a file path)
bun run eval:query:new scaffold a new Query JSON template
bun run eval:type-accuracy per-link-type accuracy report
XSS safety:
escapeHtml() encodes the 5 critical chars (& < > " '). Tested directly
with representative Opus-generated attacks:
<img src=x onerror=alert('xss')> → <img src=x onerror=alert('xss')>
<script>fetch('/steal')</script> → <script>fetch('/steal')</script>
Ledger metadata (generated_at, model) also escaped — covers the less
obvious attack surface where Opus could emit tag-like content into the
metadata file.
world.html structure:
- Left rail: entities grouped by type with counts (companies, people,
meetings, concepts), alphabetical within type
- Right pane: per-entity cards with title + slug + compiled_truth +
timeline + canonical _facts as collapsed JSON
- URL fragment deep-links (#people/alice-chen)
- Sticky rail on desktop; responsive stack on mobile
- Vanilla JS for active-link highlighting on scroll (no framework)
Generated file: ~1MB for 240 entities (full prose). Gitignored; rebuild
with `bun run eval:world:view`. Regeneration is ~50ms.
Contributor TTHW (Tier 5.5 query authoring):
1. bun run eval:world:view # see entities
2. bun run eval:query:new --tier externally-authored --author "@me"
3. edit template with real slug + query text
4. bun run eval:query:validate path/to/file.json
5. submit PR
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
|
||
|
Garry TanandClaude Opus 4.7
|
e2a5dc4a2b |
feat(eval): Phase 2 query validator + Tier 5 Fuzzy + Tier 5.5 synthetic + N=5 tolerance bands
Closes multiple Phase 2 items in one commit since they form a cohesive
package: query schema enforcement + new query tiers + per-query-set
statistical rigor.
Added:
eval/runner/queries/validator.ts — hand-rolled Query schema validator
eval/runner/queries/validator.test.ts — 24 unit tests, all pass
eval/runner/queries/tier5-fuzzy.ts — 30 hand-authored Tier 5 Fuzzy/Vibe queries
eval/runner/queries/tier5_5-synthetic.ts — 50 SYNTHETIC-labeled outsider-style queries (author: "synthetic-outsider-v1")
eval/runner/queries/index.ts — aggregator + validateAll()
Modified:
eval/runner/multi-adapter.ts — N=5 runs per adapter (BRAINBENCH_N override), page-order shuffle, mean±stddev reporting
Query validator (hand-rolled, no zod dep to match gbrain codebase style):
- Temporal verb regex enforces as_of_date (per eng pass 2 spec):
/\\b(is|was|were|current|now|at the time|during|as of|when did)\\b/i
- Validates tier enum, expected_output_type enum, gold shape per type
- gold.relevant must be non-empty slug[] for cited-source-pages queries
- abstention requires gold.expected_abstention === true
- externally-authored tier requires author field
- batch validation catches duplicate IDs
Tier 5 Fuzzy/Vibe (30 queries, hand-authored):
- Vague recall: "Someone who was a senior engineer at a biotech company..."
- Trait-based: "The engineer who pushed back on microservices"
- Cultural/epithet: "Who is known as a 'systems builder' in security?"
- Abstention bait: "Which Layer 1 project did the crypto guy leave?" (prose
mentions but never names; good systems abstain)
- Addresses Codex's circularity critique — vague queries where graph-heavy
systems shouldn't inherently win.
Tier 5.5 Synthetic Outsider (50 queries, AI-authored placeholder):
- Clearly labeled author: "synthetic-outsider-v1"
- Phrasing variety not in the 4 template families:
* fragment style ("crypto founder Goldman Sachs background")
* polite/natural ("Can you pull up what we have on...")
* comparison ("What is the difference between X and Y?")
* follow-up ("And who else advises Orbit Labs?")
* typos/misspellings ("adam lopez bioinformatcis")
* similarity ("Find me someone like Alice Davis...")
* imperative ("Pull up Alice Davis")
- Real Tier 5.5 from outside researchers supersedes synthetic via
PRs to eval/external-authors/ (docs ship in follow-up commit).
N=5 tolerance bands:
- Default N=5, override via BRAINBENCH_N env var (e.g. BRAINBENCH_N=1 for dev loops)
- Per-run seeded Fisher-Yates shuffle of page ingest order (LCG seed = run_idx+1)
- Surfaces order-dependent adapter bugs (tie-break-by-first-seen etc.)
- Reports mean ± sample-stddev per metric
- "stddev = 0" is honest signal that the adapter is deterministic, not a bug.
LLM-judge metrics (future) will naturally produce non-zero stddev.
Validation: all 80 Tier 5 + 5.5 queries pass validateAll(). 24 validator
unit tests pass.
Next commit: world.html contributor explorer (Phase 3).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
|
||
|
Garry TanandClaude Opus 4.7
|
bfa8564fc8 |
feat(eval): Phase 2 EXT-3 hybrid-without-graph adapter — graph isolated
Third and closest-to-gbrain external baseline. Runs gbrain's full hybrid
search (vector + keyword + RRF fusion + dedup) WITHOUT the knowledge-graph
layer. Same engine, same embedder, same chunking, same hybrid fusion —
only traversePaths + typed-link extraction turned off.
This is the decisive comparator for "does the knowledge graph do useful
work?" Same everything-else, only graph differs. Any lead gbrain-after has
over EXT-3 is 100% attributable to the graph layer.
Files:
eval/runner/adapters/hybrid-nograph.ts — ~110 LOC
Implementation:
- New PGLiteEngine per run; auto_link set to 'false' (belt).
- importFromContent() used instead of bare putPage() so chunks +
embeddings get populated (hybridSearch needs them).
- NO runExtract() call — typed links/timeline stay empty (suspenders).
- hybridSearch(engine, q.text) answers every query. Aggregate chunks
to page-level by best chunk score.
FOUR-adapter side-by-side on 240-page rich-prose corpus, 145 relational queries:
| Adapter | P@5 | R@5 | Correct/Gold |
|-----------------|--------|--------|--------------|
| gbrain-after | 49.1% | 97.9% | 248/261 |
| hybrid-nograph | 17.8% | 65.1% | 129/261 |
| ripgrep-bm25 | 17.1% | 62.4% | 124/261 |
| vector-only | 10.8% | 40.7% | 78/261 |
The headline delta nobody can hand-wave away:
gbrain-after → hybrid-nograph = +31.4 P@5, +32.9 R@5
hybrid-nograph → ripgrep-bm25 = +0.7 P@5, +2.7 R@5
Hybrid search (vector+keyword+RRF) over pure BM25 gains ~1 point. The
knowledge graph layer over hybrid gains ~31 points. The graph is doing
the work; adding it to a retrieval stack is what actually moves the needle
on relational queries. The vector/keyword/BM25 debate is a footnote.
Timing: hybrid-nograph init is ~2 min (embeds 240 pages once); query loop
is fast. gbrain-after is ~1.5s total because traversePaths doesn't need
embeddings. Runs at ~$0.02 Opus-equivalent in embedding cost.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
|
||
|
Garry TanandClaude Opus 4.7
|
633be384c9 |
feat(eval): Phase 2 EXT-2 vector-only RAG adapter
Second external baseline for BrainBench. Pure cosine-similarity ranking using the SAME text-embedding-3-large model gbrain uses internally — apples-to-apples on the embedding layer so any gbrain lead reflects the graph + hybrid fusion, not a better embedder. Files: eval/runner/adapters/vector-only.ts ~130 LOC eval/runner/adapters/vector-only.test.ts 6 unit tests (cosine math) Design: - One vector per page (title + compiled_truth + timeline, capped 8K chars). - No chunking (intentional; chunked vector RAG would be EXT-2b later). - No keyword fallback (that's EXT-3 hybrid-without-graph). - Embeddings in batches of 50 via existing src/core/embedding.ts (retry+backoff). - Cost on 240 pages: ~$0.02/run. Three-adapter side-by-side on 240-page rich-prose corpus, 145 relational queries: | Adapter | P@5 | R@5 | Correct top-5 | |---------------|--------|--------|---------------| | gbrain-after | 49.1% | 97.9% | 248/261 | | ripgrep-bm25 | 17.1% | 62.4% | 124/261 | | vector-only | 10.8% | 40.7% | 78/261 | Interesting finding: vector-only scores WORSE than BM25 on relational queries like "Who invested in X?" — exact entity match matters more than semantic similarity for these templates. BM25 nails the entity-name term; vector-only returns topically-similar-but-not-mentioning pages. This is the known failure mode of pure-vector RAG on precise relational/identity queries. Real-world vector RAG systems always add keyword fallback; EXT-3 (hybrid-without-graph) will be that fairer comparator. gbrain's lead widens in vector-only comparison: +38.4 pts P@5, +57.2 pts R@5. The graph layer is doing the heavy lifting for relational traversal; pure vector RAG can't express "traverse 'attended' edges from this meeting page." Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> |
||
|
Garry TanandClaude Opus 4.7
|
629ba853ff |
feat(eval): Phase 2 adapter interface + EXT-1 ripgrep+BM25 baseline
Phase 2 credibility unlock: BrainBench now compares gbrain to external
baselines on the same corpus and queries. Transforms the benchmark from
internal ablation ("gbrain-graph beats gbrain-grep") to category comparison
("gbrain-graph beats classic BM25 by 32 pts P@5"). This is the #1 fix
from the 4-review arc — addresses Codex's core critique that v1's
before/after was self-referential.
Added:
eval/runner/types.ts — Adapter interface (v1.1 spec)
eval/runner/adapters/ripgrep-bm25.ts — EXT-1 classic IR baseline
eval/runner/adapters/ripgrep-bm25.test.ts — 11 unit tests, all pass
eval/runner/multi-adapter.ts — side-by-side scorer
Adapter interface (eng pass 2 spec):
- Thin 3-method Strategy: init(rawPages, config), query(q, state), snapshot(state)
- BrainState is opaque to runner (never inspected)
- Raw pages passed in-memory; gold/ never crosses adapter boundary
(structural ingestion-boundary enforcement)
- PoisonDisposition enum reserved for future poison-resistance scoring
EXT-1 ripgrep+BM25:
- Classic Lucene-variant IDF + k1/b tuned at standard 1.5/0.75
- Title tokens double-weighted for entity-page slug-match bias
- Stopword filter, alphanumeric tokenization, stable lexicographic tie-break
- Pure in-memory inverted index — no external deps, ~100 LOC core
First side-by-side results on 240-page rich-prose corpus, 145 relational queries:
| Adapter | P@5 | R@5 | Correct top-5 |
|---------------|--------|--------|---------------|
| gbrain-after | 49.1% | 97.9% | 248/261 |
| ripgrep-bm25 | 17.1% | 62.4% | 124/261 |
| Delta | +32.0 | +35.5 | +124 |
gbrain-after is the hybrid graph+grep config from PR #188. Ripgrep+BM25 is
a genuinely strong classic-IR baseline (BM25 is what Lucene/Elasticsearch
ship). gbrain's ~+32-point lead on relational queries reflects real work
by the knowledge graph layer: typed links + traversePaths surface the
correct answers in top-K that BM25 only pulls in via partial-text overlap.
Next in Phase 2: EXT-2 vector-only RAG + EXT-3 hybrid-without-graph
adapters. Both plug into the same Adapter interface.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
|
||
|
Garry TanandClaude Opus 4.7
|
52ba00f5f8 |
feat(eval): type-accuracy runner on rich-prose corpus + wire into all.ts
New Category 2 in BrainBench: per-link-type accuracy measured directly on the
240-page rich-prose world-v1 corpus. Distinct from Cat 1's retrieval metrics,
this measures whether inferLinkType() correctly classifies extracted edges
when the prose varies (the 58% works_at and 41% advises residuals that v0.10.5
regexes targeted).
How it works:
1. Loads all pages from eval/data/world-v1/
2. Derives GOLD expected edges from each page's _facts metadata
(founders → founded, investors → invested_in, advisors → advises,
employees → works_at, attendees → attended, primary_affiliation +
role drives person-page outbound type)
3. Runs extractPageLinks() on each page → INFERRED edges
4. Per (from, to) pair, compares inferred type vs gold type
5. Emits per-link-type table: correct / mistyped / missed / spurious +
type accuracy + recall + precision + strict F1 (triple match)
6. Full confusion matrix rows=gold, cols=inferred
v0.10.5 validation on 240-page corpus (up from pre-v0.10.5 baselines):
- works_at: 58% → 100.0% (+42 pts) — 10/10 correct, 0 mistyped
- advises: 41% → 88.2% (+47 pts) — 15/17 correct
- attended: — → 100.0% 131/134 recall
- founded: 100% → 100.0% 40/40
- invested_in: 89% → 92.0% 69/75
- Overall: 88.5% → 95.7% type accuracy (conditional on edge found)
Strict F1 overall: 53.7%. Lower because the _facts-based gold set only
captures core relationships; rich prose extracts many peripheral mentions
(190 spurious "mentions" edges) that aren't bugs but are correctly-typed
prose references without a _facts counterpart. Spurious counts are signal
for future type-precision tuning, not failure.
Wired into eval/runner/all.ts as Cat 2 so every full benchmark run includes
the rich-prose type accuracy table alongside retrieval metrics.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
|
||
|
Garry TanandClaude Opus 4.7
|
109b7163a3 |
fix(link-extraction): v0.10.5 drive works_at + advises accuracy on rich prose
Extends inferLinkType patterns to cover rich-prose phrasings that miss with
v0.10.4 regexes. Targets the residuals called out in TODOS.md: works_at at
58% type accuracy, advises at 41%.
WORKS_AT_RE additions:
- Rank-prefixed: "senior engineer at", "staff engineer at", "principal/lead"
- Discipline-prefixed: "backend/frontend/full-stack/ML/data/security engineer at"
- Possessive time: "his/her/their/my time at"
- Leadership beyond "leads engineering": "heads up X at", "manages engineering at",
"runs product at", "leads the [team] at"
- Role nouns: "role at", "position at", "tenure as", "stint as"
- Promotion patterns: "promoted to staff/senior/principal at"
ADVISES_RE additions:
- Advisory capacity: "in an advisory capacity", "advisory engagement/partnership/contract"
- "as an advisor": "joined as an advisor", "serves as technical advisor"
- Prefixed advisor nouns: "strategic/technical/security/product/industry advisor to|at"
- Consulting: "consults for", "consulting role at|with"
New EMPLOYEE_ROLE_RE page-level prior: fires when the page describes the subject
as an employee (senior/staff/principal engineer, director, VP, CTO/CEO/CFO) at
some company. Biases outbound company refs toward works_at when per-edge context
is possessive or narrative without an explicit work verb. Scoped to person -> company
links only. Precedence: investor > advisor > employee (investors often hold board
seats which would otherwise mis-classify as advise/works_at).
ADVISOR_ROLE_RE broadened from "full-time/professional/advises multiple" to catch
any page that self-identifies the subject as an advisor ("is an advisor",
"serves as advisor", possessive "her advisory work/role/engagement").
Tests: 65 pass (16 new v0.10.5 coverage tests + 4 regression guards against
v0.10.4 tightenings). Templated benchmark still 88.9% type_accuracy (10/10 on
works_at and advises). Rich-prose measurement requires the multi-axis report
upgrade (next commit) to validate retroactively.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
|
||
|
|
81b3f7afac |
feat: knowledge graph layer — auto-link, typed relationships, graph-query (v0.10.3) (#188)
* feat(schema): graph layer migrations v5/v6/v7 + GraphPath/health types
Schema foundation for v0.10.3 knowledge graph layer:
- v5: links UNIQUE constraint widened to (from, to, link_type) so the same
person can both works_at AND advises the same company as separate rows.
Idempotent for fresh + upgrade (drops both old constraint names first).
- v6: timeline_entries gets UNIQUE index on (page_id, date, summary) for
ON CONFLICT DO NOTHING idempotency at DB level.
- v7: drops trg_timeline_search_vector trigger. Structured timeline entries
are now graph data, not search text. Markdown timeline still feeds search
via the pages trigger. Side benefit: extraction pagination is no longer
self-invalidating (trigger used to bump pages.updated_at on every insert).
Types: new GraphPath (edge-based traversal result), PageFilters.updated_after,
BrainHealth gets link_coverage / timeline_coverage / most_connected. Postgres
schema regenerated via build:schema.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* feat(graph): auto-link on put_page + extract --source db + security hardening
Core graph layer wired into the operation surface:
- New src/core/link-extraction.ts: extractEntityRefs (canonical extractor used
by both backlinks.ts and the new graph code), extractPageLinks (combines
markdown refs + bare-slug scan + frontmatter source, dedups within-page),
inferLinkType (deterministic regex heuristics for attended/works_at/
invested_in/founded/advises/source/mentions), parseTimelineEntries (parses
multiple date format variants from page content), isAutoLinkEnabled
(engine config flag, defaults true, accepts false/0/no/off case-insensitive).
- put_page operation auto-link post-hook: extracts entity refs from freshly
written content, reconciles links table (adds new, removes stale). Returns
auto_links: { created, removed, errors } in response so MCP callers see
outcomes. Runs in a transaction so concurrent put_page on same slug can't
race the reconciliation. Default on; opt out with auto_link=false config.
- traverse_graph operation extended with link_type and direction params.
Returns GraphPath[] (edges) when filters set, GraphNode[] (nodes) for
backwards compat. Depth hard-capped at TRAVERSE_DEPTH_CAP=10 for remote
callers; without this, depth=1e6 from MCP burns memory on the recursive CTE.
- gbrain extract <links|timeline|all> --source db: walks pages from the
engine instead of from disk. Works for live brains with no local checkout
(MCP-driven Wintermute / OpenClaw). Filesystem mode (--source fs) is
unchanged. New --type and --since filters with date validation upfront
(invalid --since used to silently no-op the filter and reprocess everything).
- Security: auto-link skipped for ctx.remote=true (MCP). Bare-slug regex
matches `people/X` anywhere in page text including code fences and quoted
strings. Without this gate an untrusted MCP caller could plant arbitrary
outbound links by writing pages with intentional slug references; combined
with the new backlink boost, attacker-placed targets would surface higher
in search.
- Postgres orphan_pages aligned to PGLite definition (no inbound AND no
outbound). Comment used to claim alignment but code disagreed; engines
drifted silently when users migrated.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* feat(cli): graph-query command + skill updates + v0.10.3 migration file
Agent-facing surface for the graph layer:
- New `gbrain graph-query <slug>` command with --type, --depth, --direction
in|out|both. Maps to traverse_graph operation with the new filters. Renders
the result as an indented edge tree.
- skills/migrations/v0.10.3.md: agent runs this post-upgrade to discover the
graph layer. Tells the agent to run `gbrain extract links --source db`,
then timeline, verify with stats, try graph-query, and lists the inferred
link types so they can be used in subsequent traversals.
- skills/brain-ops/SKILL.md Phase 2.5: documents that put_page now auto-links.
No more manual add_link calls in the Iron Law back-linking path.
- skills/maintain/SKILL.md: graph population phase. Shows the right command
to backfill links + timeline from existing pages.
- cli.ts: register graph-query in CLI_ONLY + handleCliOnly switch. Update help
text to describe `gbrain extract --source fs|db` and the new graph-query.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* test(graph): unit + e2e + 80-page A/B/C benchmark for graph layer
Coverage for the v0.10.3 graph layer (260+ new test assertions):
- test/link-extraction.test.ts (46 tests): extractEntityRefs both formats,
extractPageLinks dedup + frontmatter source, inferLinkType heuristics
(meeting/CEO/invested/founded/advises/default), parseTimelineEntries
multiple date formats + invalid date rejection, isAutoLinkEnabled
case-insensitive truthy/falsy parsing.
- test/extract-db.test.ts (12 tests): `gbrain extract <links|timeline|all>
--source db` happy paths, --type filter, --dry-run JSON output,
idempotency via DB constraint, type inference from CEO context.
- test/graph-query.test.ts (5 tests): direction in/out/both, type filter,
non-existent slug, indented tree output.
- test/pglite-engine.test.ts (+26 tests): getAllSlugs, listPages
updated_after filter, multi-type links via v5 migration, removeLink with
and without linkType, addTimelineEntry skipExistenceCheck flag,
getBacklinkCounts for hybrid search boost, traversePaths in/out/both with
cycle prevention via visited array, getHealth graph metrics
(link_coverage / timeline_coverage / most_connected).
- test/e2e/graph-quality.test.ts (6 tests): full pipeline against PGLite
in-memory. Auto-link via put_page operation handler. Reconciliation
removes stale links on edit. auto_link=false config skip.
- test/benchmark-graph-quality.ts: A/B/C comparison on 80 fictional pages,
35 queries across 7 categories. Hard thresholds: link_recall > 90%,
link_precision > 95%, timeline_recall > 85%, type_accuracy > 80%,
relational_recall > 80%. Currently passing all 9.
Built test-first: benchmark caught WORKS_AT_RE matching "founder" inside
slug names (frank-founder), "worked at" past-tense missing from regex,
PGLite Date object vs ISO string comparison bug. All fixed before merge.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* chore: bump version and changelog (v0.10.3)
CHANGELOG: knowledge graph layer headline. Auto-link on every page write.
Typed relationships (works_at, attended, invested_in, founded, advises).
gbrain extract --source db. graph-query CLI. Backlink boost in hybrid search.
Schema migrations v5/v6/v7 applied automatically.
Security hardening caught during /ship adversarial review: traverse_graph
depth capped at 10 from MCP, auto-link skipped for ctx.remote=true, runAutoLink
reconciliation in transaction, --since validates dates upfront.
TODOS.md: 2 P2 follow-ups (auto-link redundant SQL on skipped writes;
extract --source db not gated on auto_link config).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* docs: sync CLAUDE.md with v0.10.3 graph layer
Updated key files list (extract.ts now describes --source fs|db, added
graph-query.ts and link-extraction.ts), test inventory (extract-db,
link-extraction, graph-query unit tests; e2e/graph-quality), and
test count (51 unit + 7 e2e, 1151 + 105 assertions).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* docs(v0.10.3): wire graph layer into install flow + README + benchmark
Existing brains upgrading to v0.10.3 had no clear path to backfill the new
links/timeline tables. New installs had no instruction to run extract --source db
after import. This wires the knowledge graph into every install touchpoint so the
v0.10.3 features actually reach the user.
- README: headline now sells self-wiring graph + 94% benchmark numbers; new
Knowledge Graph section between Knowledge Model and Search; LINKS+GRAPH command
block expanded; Benchmarks docs group added
- INSTALL_FOR_AGENTS.md: new Step 4.5 (graph backfill) + Upgrade section now runs
gbrain init + post-upgrade and points to migrations/v<N>.md
- skills/setup/SKILL.md Phase C: new step 5 for graph backfill (idempotent,
skip-if-empty); existing file migration becomes step 6
- src/commands/init.ts: post-init hint detects existing brain (page_count > 0)
and prints extract commands for both PGLite and Postgres engines
- docs/GBRAIN_VERIFY.md: new Check #7 (knowledge graph wired) with backfill
fallback + graph-query smoke test
- docs/benchmarks/2026-04-18-graph-quality.md: checked-in benchmark report
matching the existing search-quality format (94% recall, 100% precision,
100% relational recall, idempotent both ways)
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* docs(claude): require PR descriptions to cover the whole branch
Adds a rule to CLAUDE.md so future PR bodies always cover the full diff
against the base branch, not just the most recent commit. Includes the
git log + gh pr view incantation to check what's actually in a PR.
This is a reaction to PR #189 being created with a body that described
only the last commit instead of the 7 commits it actually contained.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* feat(upgrade): post-upgrade prints full body + --execute mode + downstream skill upgrade doc
PR #188 review caught two install-flow gaps that this commit closes:
1. `gbrain post-upgrade` only printed the migration headline + description
from YAML frontmatter, never the markdown body that contains the
step-by-step backfill instructions. Agents saw "Knowledge graph layer —
your brain now wires itself" and had no idea to run `gbrain extract
links --source db`. Now prints the full body after the headline.
2. New `--execute` flag reads a structured `auto_execute:` list from
migration frontmatter and runs the safe commands sequentially. Without
`--yes` it prints the plan only (preview mode). With `--yes` it actually
runs them. Stops on first failure with a clear error.
3. Downstream agents (Wintermute etc.) keep local skill forks that gbrain
can't push updates to. New `docs/UPGRADING_DOWNSTREAM_AGENTS.md` lists
the exact diffs each release needs applied to those forks. v0.10.3
diffs for brain-ops, meeting-ingestion, signal-detector, enrich.
Changes:
- src/commands/upgrade.ts:
- runPostUpgrade(args) accepts flags
- Prints full body via extractBody()
- Parses auto_execute: list via extractAutoExecute() (hand-rolled, no yaml dep)
- --execute previews, --execute --yes runs
- Fix cosmetic bug: `recipe: null` no longer prints "show null" message
- src/cli.ts: pass args to runPostUpgrade
- skills/migrations/v0.10.3.md:
- Add auto_execute: list (gbrain init + extract links/timeline + stats)
- Fix typo: completion record version was 0.10.1, now 0.10.3
- test/upgrade.test.ts: 5 new tests covering body printing, plan preview,
actual execution, no-auto_execute case, and --help output
- docs/UPGRADING_DOWNSTREAM_AGENTS.md: NEW
- CLAUDE.md: key files list updated
Test: 13 upgrade tests pass (was 8, +5 new). Full unit suite: 1078 pass,
zero regressions, 32 expected E2E skips (no DATABASE_URL).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* bench(graph): add Configuration A baseline (no graph) vs C comparison
Previous benchmark showed C numbers only (94.4% link recall, 100% relational
recall, etc.) but never quantified what a pre-v0.10.3 brain actually loses.
Reviewer caught this gap.
Adds measureBaselineRelational() that simulates a no-graph fallback:
- Outgoing queries: regex-extract entity refs from the seed page content
- Incoming queries: grep-style scan of all pages for the seed slug
This is what an agent without the structured links table can do today.
Honest result on the 5 relational queries in the benchmark:
- Recall: 100% A vs 100% C (+0%) — markdown contains the refs either way
- Precision: 58.8% A vs 100.0% C (+70%) — without typed links, you get the
right answers buried in 41% noise
Per-query breakdown shows the divergence is concentrated in INCOMING queries:
"Who works at startup-0?" returns 5 candidates without graph (2 employees +
3 noise pages that mention startup-0) vs exactly 2 with graph. For an LLM
agent, that's ~3x less reading work per relational question.
Also documented what the benchmark deliberately doesn't test (multi-hop,
search ranking with backlink boost, aggregate queries, type-disagreement
queries) so future benchmark work has a roadmap.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* bench(graph): add 4 missing categories — multi-hop, aggregate, type-disagreement, ranking
The previous benchmark commit (
|
||
|
|
d8613366a5 |
Minions v7 + v0.11.1 canonical migration + skillify (#130)
* feat: add minion_jobs schema, migration v5, and executeRaw to BrainEngine Foundation for the Minions job queue system. Adds: - minion_jobs table (20 columns) with CHECK constraints, partial indexes, and RLS. Inspired by BullMQ's job model, adapted for Postgres. - Migration v5 creates the table for existing databases. - executeRaw<T>() method on BrainEngine interface for raw SQL access, needed by the Minions module for claim queries (FOR UPDATE SKIP LOCKED), token-fenced writes, and atomic stall detection. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * feat: Minions job queue — queue, worker, backoff, types BullMQ-inspired Postgres-native job queue built into GBrain. No Redis. No external dependencies. Postgres transactions replace Lua scripts. - MinionQueue: submit, claim (FOR UPDATE SKIP LOCKED), complete/fail (token-fenced), atomic stall detection (CTE), delayed promotion, parent-child resolution, prune, stats - MinionWorker: handler registry, lock renewal, graceful SIGTERM, exponential backoff with jitter, UnrecoverableError bypass - MinionJobContext: updateProgress(), log(), isActive() for handlers - 8-state machine: waiting/active/completed/failed/delayed/dead/ cancelled/waiting-children Patterns stolen from: BullMQ (lock tokens, stall detection, flows), Sidekiq (dead set, backoff formula), Inngest (checkpoint/resume). Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * test: 43 tests for Minions job queue Full coverage of the Minions module against PGLite in-memory: - Queue CRUD (9): submit, get, list, remove, cancel, retry, duplicate - State machine (6): waiting→active→completed/failed, retry→delayed→waiting - Backoff (4): exponential, fixed, jitter range, attempts_made=0 edge - Stall detection (3): detect stalled, counter increment, max→dead - Dependencies (5): parent waits, fail_parent, continue, remove_dep, orphan - Worker lifecycle (5): register, start-without-handlers, claim+execute, non-Error throws, UnrecoverableError bypass - Lock management (3): renewal, token mismatch, claim sets lock fields - Claim mechanics (4): empty queue, priority ordering, name filtering, delayed promotion timing - Cancel & retry (2): cancel active, retry dead Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * feat: Minions CLI commands and MCP operations Wire Minions into the GBrain CLI and MCP layer: CLI (gbrain jobs): submit <name> [--params JSON] [--follow] [--dry-run] list [--status S] [--queue Q] [--limit N] get <id> — detailed view with attempt history cancel/retry/delete <id> prune [--older-than 30d] stats — job health dashboard work [--queue Q] [--concurrency N] — Postgres-only worker daemon 6 MCP operations (contract-first, auto-exposed via MCP server): submit_job, get_job, list_jobs, cancel_job, retry_job, get_job_progress Built-in handlers: sync, embed, lint, import. --follow runs inline. Worker daemon blocked on PGLite (exclusive file lock). Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * docs: update project documentation for Minions job queue CLAUDE.md: added Minions files to key files, updated operation count (36), BrainEngine method count (38), test file count (45), added jobs CLI commands. CHANGELOG.md: added Minions entry to v0.10.0 (background jobs, retry, stall detection, worker daemon). Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * feat: Minions v2 — agent orchestration primitives (pause/resume, inbox, tokens, replay) Adds the foundation for Minions as universal agent orchestration infrastructure. GBrain's Postgres-native job queue now supports durable, observable, steerable background agents. The OpenClaw plugin (separate repo) will consume these via library import, not MCP, for zero-latency local integration. ## New capabilities - **Concurrent worker** — Promise pool replaces sequential loop. Per-job AbortController for cooperative cancellation. Graceful shutdown waits for all in-flight jobs via Promise.allSettled. - **Pause/resume** — pauseJob clears the lock and fires AbortSignal on active jobs. Handlers check ctx.signal.aborted and exit cleanly. resumeJob returns paused jobs to waiting. Catch block skips failJob when signal.aborted. - **Inbox (separate table)** — minion_inbox table for sidechannel messages. sendMessage with sender validation (parent job or admin). readInbox is token-fenced and marks read_at atomically. Separate table avoids row bloat from rewriting JSONB on every send. - **Token accounting** — tokens_input/tokens_output/tokens_cache_read columns. updateTokens accumulates; completeJob rolls child tokens up to parent. USD cost computed at read time (no cost_usd column — pricing too volatile). - **Job replay** — replayJob clones a terminal job with optional data overrides. New job, fresh attempts, no parent link. ## Handler contract additions MinionJobContext now provides: - `signal: AbortSignal` — cooperative cancellation - `updateTokens(tokens)` — accumulate token usage - `readInbox()` — check for sidechannel messages - `log()` — now accepts string or TranscriptEntry ## MCP operations added pause_job, resume_job, replay_job, send_job_message — all auto-generate CLI commands and MCP server endpoints. ## Library exports package.json exports map adds ./minions and ./engine-factory paths so plugins can `import { MinionQueue } from 'gbrain/minions'` for direct library use. ## Instruction layer (the teaching) - skills/minion-orchestrator/SKILL.md — when/how to use Minions, decision matrix, lifecycle management, anti-patterns - skills/conventions/subagent-routing.md — cross-cutting rule: all background work goes through Minions - RESOLVER.md — trigger entries for agent orchestration - manifest.json — registered ## Schema migration v6 Additive: 3 token columns, paused status, minion_inbox table with unread index. Full Postgres + PGLite support. No backfill needed. ## Tests 65 tests (was 43): pause/resume (5), inbox (6), tokens (4), replay (4), concurrent worker context (3), plus all existing coverage. ## What's NOT in this commit Deferred to follow-up PRs: - LISTEN/NOTIFY subscribe (needs real Postgres E2E) - Resource governor (depends on concurrent worker stress testing) - Routing eval harness (needs API keys + benchmark data) - OpenClaw plugin (separate @gbrain/openclaw-minions-plugin repo) See docs/designs/MINIONS_AGENT_ORCHESTRATION.md for full CEO-approved design. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * feat(minions): migration v7 — agent_parity_layer schema Adds columns on minion_jobs (depth, max_children, timeout_ms, timeout_at, remove_on_complete, remove_on_fail, idempotency_key) plus the new minion_attachments table. Three partial indexes for bounded scans: idx_minion_jobs_timeout, idx_minion_jobs_parent_status, and uniq_minion_jobs_idempotency. Check constraints enforce non-negative depth and positive child cap / timeout. Additive migration — existing installs pick it up via ensureSchema on next use. No user action required. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com> * feat(minions): extend types for v7 parity layer Extends MinionJob with depth/max_children/timeout_ms/timeout_at/ remove_on_complete/remove_on_fail/idempotency_key. Extends MinionJobInput with the same options plus max_spawn_depth override. Adds MinionQueueOpts (maxSpawnDepth default 5, maxAttachmentBytes default 5 MiB). Adds AttachmentInput/Attachment shapes and ChildDoneMessage in the InboxMessage union. rowToMinionJob updated to pick up the new columns. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com> * feat(minions): attachments validator New module validateAttachment() gates every attachment write. Rejects empty filenames, path traversal (.., /, \), null bytes, oversized content (5 MiB default, per-queue override), invalid base64, and implausible content_type headers. Returns normalized { filename, content_type, content (Buffer), sha256, size } on success. The DB also enforces UNIQUE (job_id, filename) as defense-in-depth for concurrent addAttachment races — JS-only checks are not sufficient. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com> * feat(minions): queue v7 — depth, child cap, timeouts, cascade, idempotency, child_done Wraps completeJob and failJob in engine.transaction() so parent hook invocations (resolveParent, failParent, removeChildDependency) fold into the same transaction as the child update. A process crash between child and parent can't strand the parent in waiting-children anymore. Adds v7 behaviors: - Depth tracking. add() computes depth = parent.depth + 1 and rejects past maxSpawnDepth (default 5). - Per-parent child cap. add() takes SELECT ... FOR UPDATE on the parent, counts non-terminal children, rejects when count >= max_children. NULL max_children = no cap. - Per-job wall-clock timeout. claim() populates timeout_at when timeout_ms is set. New handleTimeouts() dead-letters expired rows with error_text='timeout exceeded'. Terminal — no retry. - Cascade cancel. cancelJob() walks descendants via recursive CTE with depth-100 runaway cap. Returns the root row. Re-parented descendants (parent_job_id NULL) are naturally excluded. - Idempotency. add() uses INSERT ... ON CONFLICT (idempotency_key) DO NOTHING RETURNING; falls back to SELECT when RETURNING is empty. Same key always yields the same job id. - child_done inbox. completeJob inserts {type:'child_done', child_id, job_name, result} into the parent's inbox in the same transaction as the token rollup, guarded by EXISTS so terminal/deleted parents skip without FK violation. New readChildCompletions(parent_id, lock_token, since?) helper; token-fenced like readInbox. - removeOnComplete / removeOnFail. Deletes the row after the parent hook fires, so parent policy sees consistent state. - Attachment methods. addAttachment validates via validateAttachment then INSERTs; UNIQUE (job_id, filename) backs the JS dup check. listAttachments, getAttachment, deleteAttachment round out the API. Fixes pre-existing inverted status bug: add() now puts children in waiting/delayed (not waiting-children) and atomically flips the parent to waiting-children in the same transaction. Tests no longer need manual UPDATE workarounds. Two correctness fixes: - Sibling completion race. Under READ COMMITTED, two grandchildren completing concurrently each saw the other as still-active in the pre-commit snapshot and neither flipped the parent. Fixed by taking SELECT ... FOR UPDATE on the parent row at the start of completeJob and failJob transactions, serializing siblings on the parent lock. - JSONB double-encode. postgres.js conn.unsafe(sql, params) auto- JSON-encodes parameters. Calling JSON.stringify(obj) first stored a JSON string literal (jsonb_typeof=string) and broke payload->>'key' queries silently. Removed JSON.stringify from three call sites (child_done inbox post, updateProgress, sendMessage). PGLite tolerated both forms so unit tests missed it — real-PG E2E caught it. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com> * feat(minions): worker — timeout safety net + handleTimeouts tick Worker tick now calls handleStalled() first, then handleTimeouts() — stall requeue wins over timeout dead-letter when both could fire in the same cycle. handleTimeouts() guards on lock_until > now() so stalled jobs take the retryable path. launchJob schedules a per-job setTimeout(timeout_ms) that fires ctx.signal as a best-effort handler interrupt. The timer is always cleared in .finally so process exit isn't delayed by a dangling timer. Handlers that respect AbortSignal stop cleanly; handlers that ignore it still get dead-lettered by the DB-side handleTimeouts. Removed post-completeJob and post-failJob parent-hook calls from the worker — those are now inside the queue method transactions. Worker becomes simpler and crash-safer. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com> * test(minions): 33 new unit tests for v7 parity layer Covers depth cap, per-parent child cap, timeout dead-letter, cascade cancel (including the re-parent edge case), removeOnComplete / removeOnFail, idempotency (single + concurrent), child_done inbox (posted in txn + survives child removeOnComplete + since cursor), attachment validation (oversize, path traversal, null byte, duplicates, base64), AbortSignal firing on pause mid-handler, catch-block skipping failJob when aborted, worker in-flight bookkeeping, token-rollup guard when parent already terminal, and setTimeout safety-net cleanup. Existing tests updated to remove the inverted-status manual UPDATE workarounds that the add() fix made obsolete. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com> * test(e2e): Minions v7 concurrency + OpenClaw resilience coverage minions-concurrency.test.ts spins two MinionWorker instances against the test Postgres, submits 20 jobs, and asserts zero double-claims (every job runs exactly once). This is the only test that actually proves FOR UPDATE SKIP LOCKED under real concurrency — PGLite runs on a single connection and can't exercise the race. minions-resilience.test.ts covers the six OpenClaw daily pains: 1. Spawn storm caps enforce under concurrent submit. 2. Agent stall → handleStalled() requeues; handleTimeouts() skips (lock_until guard). 3. Forgotten dispatches recoverable via child_done inbox. 4. Cascade cancel stops grandchildren mid-flight. 5. Deep tree fan-in (parent → 3 children → 2 grandchildren each) completes with the full inbox chain. 6. Parent crash/recovery resumes from persisted state. helpers.ts extends ALL_TABLES with minion_attachments, minion_inbox, and minion_jobs (FK dependents first) so E2E teardown doesn't leak rows between runs. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com> * chore: release v0.11.0 — Minions v7 agent orchestration primitives Bumps VERSION / package.json to 0.11.0. Adds CHANGELOG entry covering depth tracking, max_children, per-job timeouts, cascade cancel, idempotency keys, child_done inbox, removeOnComplete/Fail, attachments, migration v7, plus the two correctness fixes (sibling completion race and JSONB double-encode). TODOS.md captures the four v7 follow-ups: per-queue rate limiting, repeat/cron scheduler, worker event emitter, and waitForChildren convenience helpers. 1066 unit + 105 E2E = 1171 tests passing. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com> * fix(minions): unify JSONB inserts, tighten nullish coalescing Three non-blocker cleanups from post-ship review of v0.11.0: - queue.ts add() and completeJob(): pre-stringifying with JSON.stringify while other sites pass raw objects with $n::jsonb casts. postgres.js double-encodes if you stringify first — works on PGLite (text→JSONB auto-cast), fails silently on real PG. Unify on raw object + explicit $n::jsonb cast. - queue.ts readChildCompletions: since clause used sent_at > $2 relying on PG's implicit text→TIMESTAMPTZ coercion. Explicit $2::timestamptz is safer and clearer. - types.ts rowToMinionJob: parent_job_id used || which coerces 0 to null. Harmless today (SERIAL IDs start at 1) but ?? is semantically correct. All 110 unit tests pass. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com> * fix(minions): updateProgress missed $1::jsonb cast in unification Residual from |
||
|
|
7bbfc3e36a |
security: fix wave 3 — 9 vulns (file_upload, SSRF, recipe trust, prompt injection) (#174)
* feat(engine): add cap parameter to clampSearchLimit (H6) clampSearchLimit(limit, defaultLimit, cap = MAX_SEARCH_LIMIT) — third arg is a caller-specified cap so operation handlers can enforce limits below MAX_SEARCH_LIMIT. Backward compatible: existing two-arg callers still cap at MAX_SEARCH_LIMIT. This fixes a Codex-caught semantics bug: the prior signature took (limit, defaultLimit) where the second arg was misread as a cap. clampSearchLimit(x, 20) was actually allowing values up to 100, not 20. * feat(integrations): SSRF defense + recipe trust boundary (B1, B2, Fix 2, Fix 4, B3, B4) - B1: split loadAllRecipes into trusted (package-bundled) and untrusted (cwd/recipes, $GBRAIN_RECIPES_DIR) tiers. Only package-bundled recipes get embedded=true. Closes the fake trust boundary that let any cwd-local recipe bypass health-check gates. - B2: hard-block string health_checks for non-embedded recipes (was previously only blocked when isUnsafeHealthCheck regex matched, which the cwd recipe exploit bypassed). Embedded recipes still get the regex defense. - Fix 2: gate command DSL health_checks on isEmbedded. Non-embedded recipes cannot spawnSync. - Fix 4 + B3 + B4: gate http DSL health_checks on isEmbedded; for embedded recipes, validate URLs via new isInternalUrl() before fetch: - Scheme allowlist (http/https only): blocks file:, data:, blob:, ftp:, javascript: - IPv4 range check covering hex/octal/decimal/single-integer bypass forms - IPv6 loopback ::1 + IPv4-mapped ::ffff: (canonicalized hex hextets handled) - Metadata hostnames (AWS, GCP, instance-data) blocked - fetch with redirect: 'manual' + per-hop re-validation up to 3 hops Original PRs #105-109 by @garagon. Wave 3 collector branch reimplemented the fixes after Codex outside-voice review found that PRs #106/#108 alone did not actually gate cwd-local recipes (B1) and that PR #108 missed redirect-following SSRF (B3) and non-http schemes (B4). * feat(file_upload): path/slug/filename validation + remote-caller confinement (Fix 1, B5, H5, M4, Fix 5) - Fix 1 + B5 + H1: validateUploadPath uses realpathSync + path.relative to defeat symlink-parent traversal. lstatSync alone (the original PR #105 approach) only catches final-component symlinks; a symlinked parent dir still followed to /etc/passwd. Now the entire path chain is resolved. - H5: validatePageSlug uses an allowlist regex (alphanumeric + hyphens, slash-separated segments). Closes URL-encoded traversal (%2e%2e%2f), Unicode lookalikes, backslashes, control chars implicitly. - M4: validateFilename allowlist regex. Rejects control chars, backslash, RTL override (\u202E), leading dot/dash. Filename flows into storage_path so this matters for every storage backend. - Fix 5: clamp list_pages and get_ingest_log limits at the operation layer via new clampSearchLimit cap parameter (list_pages caps at 100, get_ingest_log at 50). Internal bulk commands bypass the operation layer and remain uncapped. - New OperationContext.remote flag distinguishes trusted local CLI from untrusted MCP callers. file_upload uses strict cwd confinement when remote=true (default), loose mode when remote=false (CLI). MCP stdio server sets remote=true; cli.ts and handleToolCall (gbrain call) set remote=false. Original PR #105 by @garagon. Issue #139 reported by @Hybirdss. * feat(search): query sanitization + structural prompt boundary (Fix 3, M1, M2, M3) - M1: restructure callHaikuForExpansion to use a system message that declares the user query as untrusted data, plus an XML-tagged <user_query> boundary in the user message. Layered defense with the existing tool_choice constraint (3 layers vs 1). - Fix 3 (regex sanitizer, defense-in-depth): sanitizeQueryForPrompt strips triple-backtick code fences, XML/HTML tags, leading injection prefixes, and caps at 500 chars. Original query is still used for downstream search; only the LLM-facing copy is sanitized. - M2: sanitizeExpansionOutput validates the model's alternative_queries array before it flows into search. Strips control chars, caps length, dedupes case-insensitively, drops empty/non-string items, caps to 2 items. - M3: console.warn on stripped content NEVER logs the query text — privacy-safe debug signal only. Original PR #107 by @garagon. M1/M2/M3 are wave 3 hardening per Codex review. * chore: bump version and changelog (v0.10.2) Security wave 3: 9 vulnerabilities closed across file_upload, recipe trust boundary, SSRF defense, prompt injection, and limit clamping. See CHANGELOG for full details. Contributors: - @garagon (PRs #105-109) - @Hybirdss (Issue #139) Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * docs: sync documentation with v0.10.2 security wave 3 - CLAUDE.md: document OperationContext.remote, new security helpers (validateUploadPath, validatePageSlug, validateFilename, isInternalUrl, parseOctet, hostnameToOctets, isPrivateIpv4, getRecipeDirs, sanitizeQueryForPrompt, sanitizeExpansionOutput), updated clampSearchLimit signature, recipe trust boundary, new test files - docs/integrations/README.md: replace string-form health_check example with typed DSL (string checks now hard-block for non-embedded recipes); add recipe trust boundary subsection - docs/mcp/DEPLOY.md: document file_upload remote-caller cwd confinement, symlink rejection, slug/filename allowlists Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com> |
||
|
|
b7e3005b5b |
fix: sync pipeline, extract, features, autopilot (v0.10.1) (#129)
* feat: migrate 8 existing skills to conformance format Add YAML frontmatter (name, version, description, triggers, tools, mutating), Contract, Anti-Patterns, and Output Format sections to all existing skills. Rename Workflow to Phases. Ingest becomes thin router delegating to specialized ingestion skills (Phase 2). Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * feat: add RESOLVER.md, conventions directory, and output rules RESOLVER.md is the skill dispatcher modeled on Wintermute's AGENTS.md. Categorized routing table: Always-on, Brain ops, Ingestion, Thinking, Operational, Setup, Identity. Conventions directory extracts cross-cutting rules (quality, brain-first lookup, model routing, test-before-bulk). Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * test: add skills conformance and resolver validation tests skills-conformance.test.ts validates every skill has YAML frontmatter with required fields, Contract, Anti-Patterns, and Output Format sections, and manifest.json coverage. resolver.test.ts validates routing table categories, skill path existence, and manifest-to-resolver coverage. 50 new tests. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * feat: add 9 brain skills from Wintermute (Phase 2) Generalized from Wintermute's battle-tested skills: - signal-detector: always-on idea+entity capture on every message - brain-ops: brain-first lookup, read-enrich-write loop, source attribution - idea-ingest: links/articles/tweets with author people page mandatory - media-ingest: video/audio/PDF/book with entity extraction (absorbs video/youtube/book) - meeting-ingestion: transcripts with attendee enrichment chaining - citation-fixer: audit and fix citation formatting - repo-architecture: filing rules by primary subject - skill-creator: create skills with conformance standard + MECE check - daily-task-manager: task lifecycle with priority levels All Garry-specific references generalized. Core workflows preserved. Updated RESOLVER.md and manifest.json. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * feat: add operational infrastructure + identity layer (Phase 3) Operational skills: - daily-task-prep: morning prep with calendar context and open threads - cross-modal-review: quality gate via second model with refusal routing - cron-scheduler: schedule staggering, quiet hours, wake-up override, idempotency - reports: timestamped reports with keyword routing - testing: skill validation framework (conformance checks) - soul-audit: 6-phase interview generating SOUL.md, USER.md, ACCESS_POLICY.md, HEARTBEAT.md - webhook-transforms: external events to brain signals with dead-letter queue Identity layer: - SOUL.md template (agent identity, generated by soul-audit) - USER.md template (user profile, generated by soul-audit) - ACCESS_POLICY.md template (4-tier access control) - HEARTBEAT.md template (operational cadence) - cross-modal.yaml convention (review pairs, refusal routing chain) Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * docs: update CLAUDE.md with 24 skills, RESOLVER.md, conventions, templates GBrain is now a GStack mod for agent platforms. Updated architecture description, key files listing (16 new skill files, RESOLVER.md, conventions, templates), skills section (24 skills organized by resolver categories), and testing section (new conformance and resolver tests). Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * feat: add GStack detection + mod status to gbrain init (Phase 4) After brain initialization, gbrain init now reports: - Number of skills loaded (from manifest.json) - GStack detection (checks known host paths, uses gstack-global-discover if available) - GStack install instructions if not found - Resolver and soul-audit pointers Also adds installDefaultTemplates() for SOUL.md/USER.md/ACCESS_POLICY.md/HEARTBEAT.md deployment, and detectGStack() using gstack-global-discover with fallback to known paths (DRY: doesn't reimplement GStack's host detection logic). Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * docs: v0.10.0 release documentation - CHANGELOG: 24 skills, signal detector, RESOLVER.md, soul-audit, access control, conventions, conformance standard, GStack detection in init - README: updated skill section with 24 skills, resolver, conventions - TODOS: added runtime MCP access control (P1) - VERSION: 0.9.2 → 0.10.0 - package.json + manifest.json version bumped Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * docs: add skill table to CHANGELOG v0.10.0 16-row table detailing every new skill, what it does, and why it matters. Written to sell the upgrade, not document the implementation. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * fix: restore package.json version after merge conflict resolution Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * docs: zero-based README rewrite for GStackBrain v0.10.0 Lead with GStack mod identity. 24 skills table organized by category. Install block references RESOLVER.md and soul-audit. GBrain+GStack relationship explained. Removed redundancy (733 -> 406 lines). All essential content preserved: install, recipes, architecture, search, commands, engines, voice, knowledge model. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * docs: extract install block to INSTALL_FOR_AGENTS.md, simplify README The 30-line copy-paste install block becomes one line: "Retrieve and follow INSTALL_FOR_AGENTS.md" Benefits: agent always gets latest instructions (no stale copy-paste), README stays clean, install details live where agents read them. README now leads with what GBrain does ("gives your agent a brain") instead of GStack relationship. Removed "requires frontier model" note. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * fix: 3 bugs in init.ts from merge conflict resolution 1. llstatSync typo (merge corruption) → lstatSync 2. __dirname undefined in ESM module → fileURLToPath polyfill 3. require('fs') in ESM → use imported readFileSync All three would crash gbrain init at runtime. Caught by /review. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * feat: add checkResolvable shared core function for resolver validation Shared function at src/core/check-resolvable.ts validates that all skills are reachable from RESOLVER.md, detects MECE overlaps (with whitelist for always-on/router skills), finds gaps in frontmatter triggers, and scans for DRY violations. Returns structured ResolvableIssue objects with machine-parseable fix objects alongside human-readable action strings. Three call sites: bun test, gbrain doctor, skill-creator skill. Cleans up test/resolver.test.ts: removes stale 9-line skip list, imports from production check-resolvable.ts instead of reimplementing parsing. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * feat: expand doctor with resolver validation, filesystem-first architecture Doctor now runs filesystem checks (resolver health, skill conformance) before connecting to DB. New --fast flag skips DB checks. Falls back to filesystem-only when DB is unavailable. Adds schema_version: 2 to JSON output, composite health score (0-100), and structured issues array with action strings for agent parsing. Resolver health check calls checkResolvable() and surfaces actionable fix instructions. Link integrity check uses engine.getHealth() dead_links count. CLI routing split: doctor dispatched before connectEngine() so filesystem checks always run. Fixes Codex-identified blocker where doctor required DB. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * feat: add adaptive load-aware throttling and fail-improve loop backoff.ts: System load checking (CPU via os.loadavg, memory via os.freemem), exponential backoff with 20-attempt max guard, active hours multiplier (2x slower during waking hours), concurrent process limit (max 2). Windows-safe: defaults to "proceed" when os.loadavg returns zeros. fail-improve.ts: Deterministic-first, LLM-fallback pattern with JSONL failure logging. Cascade failure handling: when both paths fail, throws LLM error and logs both. Log rotation at 1000 entries. Call count tracking for deterministic hit rate metrics. Auto-generates test cases from successful LLM fallbacks. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * feat: add transcription service and enrichment-as-a-service transcription.ts: Groq Whisper (default) with OpenAI fallback. Files >25MB segmented via ffmpeg. Provider auto-detection from env vars. Clear error messages for missing API keys and unsupported formats. enrichment-service.ts: Global enrichment service callable from any ingest pathway. Entity slug generation (people/jane-doe, companies/acme-corp), mention counting via searchKeyword, tier auto-escalation (Tier 3→2→1 based on mention frequency and source diversity), batch enrichment with backoff throttling, regex-based entity extraction from text. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * feat: add data-research skill with recipe system, extraction, dedup, tracker New skill: data-research — one parameterized pipeline for any email-to- structured-data workflow (investor updates, donations, company metrics). 7-phase pipeline: define recipe, search, classify, extract (with extraction integrity rule), archive, deduplicate, update tracker. data-research.ts: Recipe validation, MRR/ARR/runway/headcount regex extraction (battle-tested patterns), dedup with configurable tolerance, markdown tracker parsing/appending, quarterly/monthly date windowing, 6-phase HTML email stripping with 500KB ReDoS cap. Registers data-research in manifest.json (25th skill) and RESOLVER.md. Fixes backoff test robustness for high-load systems. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * docs: update project documentation for v0.10.0 infrastructure additions CLAUDE.md: added 6 new core files (check-resolvable, backoff, fail-improve, transcription, enrichment-service, data-research), 6 new test files, updated skill count to 25, test file count to 34. README.md: updated skill count to 25, added data-research to skills table. CHANGELOG.md: added Infrastructure section documenting resolver validation, doctor expansion, adaptive throttling, fail-improve loop, voice transcription, enrichment service, and data-research skill. TODOS.md: anonymized personal references. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * fix: doctor.ts use ES module imports, harden backoff test Replace require('fs') with ES module import in doctor.ts for consistency with the rest of the file. Backoff test made resilient to parallel test execution leaking module-level state. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * fix: sync --watch routing, dead_links parity, doctor command, embed --slugs - Move sync to CLI_ONLY so --watch flag reaches runSync() (was routed through operation layer which only calls performSync single-pass) - Hide sync_brain from CLI help (MCP still exposes it) - Fix performFullSync missing sync state persistence (C1) - Align Postgres dead_links query to match PGLite (count dangling links, not empty-content chunks) (C3) - Fix doctor recommending nonexistent 'gbrain embed refresh' (C4) - Refactor doctor outputResults to not call process.exit directly - Add --slugs flag to embed for targeted page embedding - Add sync auto-extract + auto-embed after performSync - Add noExtract to SyncOpts - Route extract, features, autopilot in CLI_ONLY - Update help text with new commands Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * feat: extract, features, and autopilot commands - gbrain extract <links|timeline|all> — batch extraction of links and timeline entries from brain markdown files. Broad regex for all .md links (C7: filters external URLs). Frontmatter field parsing (company, investors, attendees). Directory-based link type inference. JSONL progress on stderr for agents. Sync integration hooks (extractLinksForSlugs, extractTimelineForSlugs). - gbrain features [--json] [--auto-fix] — scan brain usage, pitch unused features with the user's own numbers. Priority 1 (data quality): missing embeddings, dead links. Priority 2 (unused features): zero links, zero timeline, low coverage, unconfigured integrations, no sync. Embedded recipe metadata for binary-safe integration detection. Persistence in ~/.gbrain/feature-offers.json. Doctor teaser hook. Upgrade hook. - gbrain autopilot [--repo] [--interval N] — self-maintaining brain daemon. Pipeline: sync → extract → embed. Health-based adaptive scheduling (brain_score >= 90 doubles interval, < 70 halves it). --install/--uninstall for launchd (macOS) and crontab (Linux). Signal handling. Consecutive error tracking (stops at 5). Log to ~/.gbrain/autopilot.log. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * feat: hook features scan into post-upgrade flow After gbrain post-upgrade completes, automatically run gbrain features to show the user what's new and what to fix. Best-effort (doesn't fail the upgrade). Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * feat: brain_score (0-100) in BrainHealth Weighted composite score computed in getHealth() for both Postgres and PGLite: embed_coverage: 0.35, link_density: 0.25, timeline_coverage: 0.15, no_orphans: 0.15, no_dead_links: 0.10 Returns 0 for empty brains. Agents use brain_score as a health gate. Autopilot uses it for adaptive scheduling (>=90 slows down, <70 speeds up). Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * test: extract and features unit tests 25 tests covering: - extractMarkdownLinks: relative links, external URL filtering, edge cases - extractLinksFromFile: slug resolution, frontmatter parsing, directory-based type inference (works_at, deal_for, invested_in) - extractTimelineFromContent: bullet format, header format with detail, em/en dash handling, empty content - features: module exports, brain_score calculation weights, CLI routing Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * docs: instruction layer for extract, features, autopilot Agent-facing tools are invisible without instruction-layer coverage. - RESOLVER.md: add routing for extract, features, autopilot - maintain/SKILL.md: add link graph extraction, timeline extraction, autopilot check sections Without these, agents reading skills/ will never discover or run the new commands. This is the #1 DX finding from the devex review. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * chore: bump version and changelog (v0.10.1) Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * docs: sync CLAUDE.md with v0.10.1 additions Add extract.ts, features.ts, autopilot.ts to key files. Add extract.test.ts, features.test.ts to test list. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * fix: adversarial review fixes — 7 issues - #3: autopilot extract step was a no-op (imported but never called) - #6: PGLite orphan_pages query aligned with Postgres (check both inbound+outbound) - #8: embedPage throws instead of process.exit (was killing sync/autopilot) - #9: dead-links set auto_fixable=false (needs repo path we may not have) - #10: JSON auto-fix output was dead code (unreachable !jsonMode check) - #14: autopilot lock file prevents concurrent instances - #20: --dir without value no longer crashes extract Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * security: fix command injection + plaintext API key in daemon install - #1: Crontab install used echo pipe with shell-interpolated values. Now uses a temp file via crontab(1) and single-quote escaping on all interpolated paths. No shell expansion possible. - #2: OPENAI_API_KEY was baked as plaintext into the launchd plist (readable by any local process, backed up by Time Machine). Now uses a wrapper script (~/.gbrain/autopilot-run.sh) that sources ~/.zshrc at runtime. No secrets in plist or crontab. - #16: extract.ts used a custom 20-line YAML parser that only handled single-line key:value pairs. Multi-line arrays (attendees list with - items) were silently ignored. Now uses the project's gray-matter parser via parseMarkdown() from src/core/markdown.ts. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com> |
||
|
|
e5a9f0126a |
feat: GStackBrain — 16 new skills, resolver, conventions, identity layer (v0.10.0) (#120)
* feat: migrate 8 existing skills to conformance format Add YAML frontmatter (name, version, description, triggers, tools, mutating), Contract, Anti-Patterns, and Output Format sections to all existing skills. Rename Workflow to Phases. Ingest becomes thin router delegating to specialized ingestion skills (Phase 2). Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * feat: add RESOLVER.md, conventions directory, and output rules RESOLVER.md is the skill dispatcher modeled on Wintermute's AGENTS.md. Categorized routing table: Always-on, Brain ops, Ingestion, Thinking, Operational, Setup, Identity. Conventions directory extracts cross-cutting rules (quality, brain-first lookup, model routing, test-before-bulk). Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * test: add skills conformance and resolver validation tests skills-conformance.test.ts validates every skill has YAML frontmatter with required fields, Contract, Anti-Patterns, and Output Format sections, and manifest.json coverage. resolver.test.ts validates routing table categories, skill path existence, and manifest-to-resolver coverage. 50 new tests. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * feat: add 9 brain skills from Wintermute (Phase 2) Generalized from Wintermute's battle-tested skills: - signal-detector: always-on idea+entity capture on every message - brain-ops: brain-first lookup, read-enrich-write loop, source attribution - idea-ingest: links/articles/tweets with author people page mandatory - media-ingest: video/audio/PDF/book with entity extraction (absorbs video/youtube/book) - meeting-ingestion: transcripts with attendee enrichment chaining - citation-fixer: audit and fix citation formatting - repo-architecture: filing rules by primary subject - skill-creator: create skills with conformance standard + MECE check - daily-task-manager: task lifecycle with priority levels All Garry-specific references generalized. Core workflows preserved. Updated RESOLVER.md and manifest.json. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * feat: add operational infrastructure + identity layer (Phase 3) Operational skills: - daily-task-prep: morning prep with calendar context and open threads - cross-modal-review: quality gate via second model with refusal routing - cron-scheduler: schedule staggering, quiet hours, wake-up override, idempotency - reports: timestamped reports with keyword routing - testing: skill validation framework (conformance checks) - soul-audit: 6-phase interview generating SOUL.md, USER.md, ACCESS_POLICY.md, HEARTBEAT.md - webhook-transforms: external events to brain signals with dead-letter queue Identity layer: - SOUL.md template (agent identity, generated by soul-audit) - USER.md template (user profile, generated by soul-audit) - ACCESS_POLICY.md template (4-tier access control) - HEARTBEAT.md template (operational cadence) - cross-modal.yaml convention (review pairs, refusal routing chain) Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * docs: update CLAUDE.md with 24 skills, RESOLVER.md, conventions, templates GBrain is now a GStack mod for agent platforms. Updated architecture description, key files listing (16 new skill files, RESOLVER.md, conventions, templates), skills section (24 skills organized by resolver categories), and testing section (new conformance and resolver tests). Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * feat: add GStack detection + mod status to gbrain init (Phase 4) After brain initialization, gbrain init now reports: - Number of skills loaded (from manifest.json) - GStack detection (checks known host paths, uses gstack-global-discover if available) - GStack install instructions if not found - Resolver and soul-audit pointers Also adds installDefaultTemplates() for SOUL.md/USER.md/ACCESS_POLICY.md/HEARTBEAT.md deployment, and detectGStack() using gstack-global-discover with fallback to known paths (DRY: doesn't reimplement GStack's host detection logic). Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * docs: v0.10.0 release documentation - CHANGELOG: 24 skills, signal detector, RESOLVER.md, soul-audit, access control, conventions, conformance standard, GStack detection in init - README: updated skill section with 24 skills, resolver, conventions - TODOS: added runtime MCP access control (P1) - VERSION: 0.9.2 → 0.10.0 - package.json + manifest.json version bumped Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * docs: add skill table to CHANGELOG v0.10.0 16-row table detailing every new skill, what it does, and why it matters. Written to sell the upgrade, not document the implementation. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * fix: restore package.json version after merge conflict resolution Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * docs: zero-based README rewrite for GStackBrain v0.10.0 Lead with GStack mod identity. 24 skills table organized by category. Install block references RESOLVER.md and soul-audit. GBrain+GStack relationship explained. Removed redundancy (733 -> 406 lines). All essential content preserved: install, recipes, architecture, search, commands, engines, voice, knowledge model. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * docs: extract install block to INSTALL_FOR_AGENTS.md, simplify README The 30-line copy-paste install block becomes one line: "Retrieve and follow INSTALL_FOR_AGENTS.md" Benefits: agent always gets latest instructions (no stale copy-paste), README stays clean, install details live where agents read them. README now leads with what GBrain does ("gives your agent a brain") instead of GStack relationship. Removed "requires frontier model" note. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * fix: 3 bugs in init.ts from merge conflict resolution 1. llstatSync typo (merge corruption) → lstatSync 2. __dirname undefined in ESM module → fileURLToPath polyfill 3. require('fs') in ESM → use imported readFileSync All three would crash gbrain init at runtime. Caught by /review. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * feat: add checkResolvable shared core function for resolver validation Shared function at src/core/check-resolvable.ts validates that all skills are reachable from RESOLVER.md, detects MECE overlaps (with whitelist for always-on/router skills), finds gaps in frontmatter triggers, and scans for DRY violations. Returns structured ResolvableIssue objects with machine-parseable fix objects alongside human-readable action strings. Three call sites: bun test, gbrain doctor, skill-creator skill. Cleans up test/resolver.test.ts: removes stale 9-line skip list, imports from production check-resolvable.ts instead of reimplementing parsing. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * feat: expand doctor with resolver validation, filesystem-first architecture Doctor now runs filesystem checks (resolver health, skill conformance) before connecting to DB. New --fast flag skips DB checks. Falls back to filesystem-only when DB is unavailable. Adds schema_version: 2 to JSON output, composite health score (0-100), and structured issues array with action strings for agent parsing. Resolver health check calls checkResolvable() and surfaces actionable fix instructions. Link integrity check uses engine.getHealth() dead_links count. CLI routing split: doctor dispatched before connectEngine() so filesystem checks always run. Fixes Codex-identified blocker where doctor required DB. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * feat: add adaptive load-aware throttling and fail-improve loop backoff.ts: System load checking (CPU via os.loadavg, memory via os.freemem), exponential backoff with 20-attempt max guard, active hours multiplier (2x slower during waking hours), concurrent process limit (max 2). Windows-safe: defaults to "proceed" when os.loadavg returns zeros. fail-improve.ts: Deterministic-first, LLM-fallback pattern with JSONL failure logging. Cascade failure handling: when both paths fail, throws LLM error and logs both. Log rotation at 1000 entries. Call count tracking for deterministic hit rate metrics. Auto-generates test cases from successful LLM fallbacks. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * feat: add transcription service and enrichment-as-a-service transcription.ts: Groq Whisper (default) with OpenAI fallback. Files >25MB segmented via ffmpeg. Provider auto-detection from env vars. Clear error messages for missing API keys and unsupported formats. enrichment-service.ts: Global enrichment service callable from any ingest pathway. Entity slug generation (people/jane-doe, companies/acme-corp), mention counting via searchKeyword, tier auto-escalation (Tier 3→2→1 based on mention frequency and source diversity), batch enrichment with backoff throttling, regex-based entity extraction from text. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * feat: add data-research skill with recipe system, extraction, dedup, tracker New skill: data-research — one parameterized pipeline for any email-to- structured-data workflow (investor updates, donations, company metrics). 7-phase pipeline: define recipe, search, classify, extract (with extraction integrity rule), archive, deduplicate, update tracker. data-research.ts: Recipe validation, MRR/ARR/runway/headcount regex extraction (battle-tested patterns), dedup with configurable tolerance, markdown tracker parsing/appending, quarterly/monthly date windowing, 6-phase HTML email stripping with 500KB ReDoS cap. Registers data-research in manifest.json (25th skill) and RESOLVER.md. Fixes backoff test robustness for high-load systems. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * docs: update project documentation for v0.10.0 infrastructure additions CLAUDE.md: added 6 new core files (check-resolvable, backoff, fail-improve, transcription, enrichment-service, data-research), 6 new test files, updated skill count to 25, test file count to 34. README.md: updated skill count to 25, added data-research to skills table. CHANGELOG.md: added Infrastructure section documenting resolver validation, doctor expansion, adaptive throttling, fail-improve loop, voice transcription, enrichment service, and data-research skill. TODOS.md: anonymized personal references. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * fix: doctor.ts use ES module imports, harden backoff test Replace require('fs') with ES module import in doctor.ts for consistency with the rest of the file. Backoff test made resilient to parallel test execution leaking module-level state. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * docs: README rewrite with production brain stats, sample output, new infrastructure Lead with the flex: 17,888 pages, 4,383 people, 723 companies, 526 meeting transcripts built in 12 days. Show sample query output so readers see what they'll get. Document self-improving infrastructure (tier auto-escalation, fail-improve loop, doctor trajectory). Add data-research recipes to Getting Data In. Update commands section with doctor --fix, transcribe, research init/list. Fix stale "24" references to "25". Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * docs: README lead with YC President origin and production agent deployments Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * docs: README lead with skill philosophy and link to Thin Harness Fat Skills Skills section now explains: skill files are code, they encode entire workflows, they call deterministic TypeScript for the parts that shouldn't be LLM judgment. Links to the tweet and the architecture essay. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * docs: link GStack repo, add 70K stars and 30K daily users Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * docs: remove meeting transcript count from README (sensitive) Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * docs: README lead with YC President origin and production agent deployments Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * fix: rename political-donations recipe to expense-tracker (sensitivity) Renamed the built-in data-research recipe from political-donations to expense-tracker across README, CHANGELOG, SKILL.md, and reports routing. Same extraction patterns (amounts, dates, recipients), neutral framing. Also renamed social-radar keyword route to social-mentions. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com> |
||
|
|
d547a64600 |
feat: search quality boost — compiled truth ranking + detail parameter (v0.8.1) (#64)
* feat: search quality boost — compiled truth ranking, detail parameter, cosine re-scoring Compiled truth chunks now rank 2x higher in hybrid search via RRF normalization + source boost. New --detail flag (low/medium/high) controls timeline inclusion. Cosine re-scoring blends query-chunk similarity before dedup for query-specific ranking. Also: remove DISTINCT ON from keyword search (dedup handles per-page capping), add chunk_id + chunk_index to SearchResult, add getEmbeddingsByChunkIds to BrainEngine interface. Inspired by Ramp Labs' "Latent Briefing" paper (April 2026). * feat: RRF normalization, source-aware dedup, detail param in operations RRF scores normalized to 0-1 before 2.0x compiled truth boost. Source-aware dedup guarantees compiled truth chunk per page. Detail parameter added to query operation, dedupResults added to bare search operation. Debug logging via GBRAIN_SEARCH_DEBUG=1. * chore: bump version and changelog (v0.8.1) Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> * fix: CJK word count in query expansion CJK text is not space-delimited. A query like "向量搜索优化" was counted as 1 word and silently skipped expansion. Now counts characters for CJK queries instead of space-separated tokens. Co-Authored-By: YIING99 <yiing99@users.noreply.github.com> * feat: retrieval evaluation harness — P@k, R@k, MRR, nDCG@k + gbrain eval Full IR evaluation framework: precisionAtK, recallAtK, mrr, ndcgAtK metrics with runEval() orchestrator. gbrain eval CLI with single-run table and A/B comparison mode (--config-a / --config-b) for parameter tuning. HybridSearchOpts now accepts rrfK and dedupOpts overrides. Co-Authored-By: 4shut0sh <4shut0sh@users.noreply.github.com> * test: search quality tests — RRF boost, dedup guarantee, cosine similarity, E2E benchmark 42 new tests across 3 files: - test/search.test.ts: RRF normalization, compiled truth 2x boost, dedup key collision prevention, cosine similarity edge cases, CJK word count detection - test/dedup.test.ts: source-aware compiled truth guarantee, layer interactions, custom maxPerPage, empty/single result edge cases - test/e2e/search-quality.test.ts: full pipeline against PGLite with basis vector embeddings — chunk_id/chunk_index fields, detail parameter filtering, getEmbeddingsByChunkIds, keyword multi-chunk, vector ordering Also: export rrfFusion + cosineSimilarity for unit testing, fix PGLite getEmbeddingsByChunkIds to parse string vectors from pgvector. * test: search quality benchmark with A/B comparison (baseline vs PR#64) Benchmark measures P@1, MRR, nDCG@5, and source accuracy across 8 queries against 5 seeded pages. Key finding: boost helps entity lookups but over-corrects temporal queries. Validates the --detail parameter as the right control mechanism. Output at docs/benchmarks/2026-04-13.md. * feat: query intent classifier — auto-selects detail level, 100% source accuracy Zero-latency heuristic classifier detects query intent from text patterns: - "Who is Pedro?" → entity → detail=low (compiled truth only) - "When did we last meet?" → temporal → detail=high (no boost, natural ranking) - "Variant fund announcement" → event → detail=high - General queries → detail=medium (default with boost) The key insight: skip the 2.0x compiled truth boost for detail=high queries. Temporal/event queries want natural ranking where timeline entries can win. Benchmark results (source accuracy = does the top chunk match expected type): - Baseline: 100% (already good, no boost needed) - Boost only: 71.4% (boost over-corrects temporal queries) - Boost + intent classifier: 100% (best of both worlds) 35 unit tests for the classifier. 590 total tests pass. * feat: query intent classifier — auto-selects detail level, 100% source accuracy Heuristic classifier detects query intent from text patterns (zero latency, no LLM call). Maps temporal queries ("when did we last meet") to detail=high, entity queries ("who is X") to detail=low, events to detail=high. Benchmark results (29 pages, 20 queries, graded relevance): - Baseline: P@1=0.947, MRR=0.974, source accuracy=89.5% - Boost only: P@1=0.895, MRR=0.939, source accuracy=63.2% (over-correction) - Boost + intent: P@1=0.947, MRR=0.974, source accuracy=89.5% (fully recovered) The intent classifier eliminates the boost's over-correction on temporal queries while preserving its benefits for entity lookups. 35 unit tests for the classifier. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> * test: search quality benchmark with A/B comparison (baseline vs PR#64) Rich benchmark: 29 pages, 58 chunks, 20 queries with graded relevance. Now measures CHUNK-LEVEL quality, not just page-level retrieval. Key findings (C. Boost+Intent vs A. Baseline): - Unique pages in top-10: 7.2 → 8.7 (+21% broader coverage) - Compiled truth ratio: 51.6% → 66.8% (+15pp more signal) - CT-first rate: 100% (compiled truth leads for entity queries) - Timeline accessible: 100% (temporal queries still find dates) - Source accuracy: 89.5% maintained (intent classifier prevents regression) The boost alone (B) causes -26pp source accuracy regression. Intent classifier (C) recovers it fully. * docs: clean benchmark report — ELI10 search quality analysis for PR#64 Replaces two drafts with one clean report. Explains what changed, why it matters, and what the numbers mean. All fictional data, no private info. Key findings: 21% more page coverage per query, 29% more compiled truth in results. Intent classifier prevents boost from burying timeline for temporal queries. Full per-query breakdown with before/after comparison. * chore: remove auto-generated benchmark file (clean version is 2026-04-14-search-quality.md) * docs: update project documentation for search quality boost CLAUDE.md: added search/intent.ts, search/eval.ts, commands/eval.ts to key files. Added 5 new test files (search, dedup, intent, eval, e2e/search-quality). Updated test count from 23+4 to 28+5. Added docs/benchmarks/ to key files. README.md: updated search pipeline diagram with intent classifier, RRF normalization, compiled truth boost, cosine re-scoring, and 5-layer dedup. Added --detail flag explanation and benchmark instructions. CHANGELOG.md: added search quality entries to v0.9.3 (intent classifier, --detail flag, gbrain eval, CJK fix). Credited @4shut0sh and @YIING99. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> * docs: headline benchmark gains in changelog * docs: add community attribution rule to CHANGELOG voice section --------- Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com> Co-authored-by: YIING99 <yiing99@users.noreply.github.com> Co-authored-by: 4shut0sh <4shut0sh@users.noreply.github.com> |
||
|
|
f82978d38d |
security: fix wave 2 — 5 vulns + typed health check DSL (v0.9.3) (#95)
* security: path traversal, query bounds, marker injection fixes LocalStorage: contained() method validates all paths stay within storage root. file-resolver: resolveFile validates filePath within brainRoot, marker prefix rejects ../, absolute paths, bare '..'. file_list: LIMIT 100 on slug-filtered branch + FILE_LIST_LIMIT constant for both branches. Co-Authored-By: Gus <garagon@users.noreply.github.com> * security: symlink hardening in all file walkers All 4 walkers in files.ts (collectFiles, findRedirects, findAndClean, scan) plus init.ts counter now use lstatSync + isSymbolicLink skip. Tests import production collectFiles instead of reimplementing it. node_modules skipped. CLI file list and verify queries bounded with LIMIT. Co-Authored-By: Gus <garagon@users.noreply.github.com> * feat: typed health check DSL + recipe migration 4 DSL types: http, env_exists, command, any_of. Replaces raw execSync on recipe YAML. All 7 first-party recipes migrated from shell strings to typed objects. String health_checks still accepted with deprecation warning + metachar validation for non-embedded recipes. isUnsafeHealthCheck blocks shell injection for user-created recipes. Co-Authored-By: Gus <garagon@users.noreply.github.com> * chore: bump version and changelog (v0.9.3) Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> * test: E2E test for file_list LIMIT enforcement against real Postgres Inserts 150 file rows for one slug, verifies file_list returns at most 100 (both slug-filtered and unfiltered branches). Proves the LIMIT works at the database level, not just in unit tests. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> --------- Co-authored-by: Gus <garagon@users.noreply.github.com> Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com> |
||
|
Francisco MaranchelloandGitHub
|
adb02b7826 | fix: create PGLite data dir before lock (#85) | ||
|
Garry TanandClaude Opus 4.6
|
004ac6c66f |
fix: statement_timeout scoped to search, upload-raw writes pointer, publish inlines marked.js
1. statement_timeout: 8s moved from global connection config to searchKeyword/searchVector only. Prevents DoS on search without killing embed --all or bulk imports that need longer than 8s. 2. upload-raw now writes the .redirect.yaml pointer file to disk (was creating the pointer object but never calling writeFileSync). 3. publish inlines marked.js from node_modules instead of loading from cdn.jsdelivr.net. Generated HTML is now truly self-contained with no external dependencies. 4. v0.9.1 migration doc updated with slug authority breaking change warning for brains that use frontmatter slug: overrides. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> |
||
|
Anurag GoelandGitHub
|
fa62e61994 | Fix path to OpenClaw repo (#79) | ||
|
Garry TanandClaude Opus 4.6
|
784b582c6d |
docs: fix install block from agent feedback
- Anthropic key is optional (graceful fallback), not required - Remove fragile shell sourcing, add PATH export + restart note - Integrations: ask user which ones, not "set up EVERY recipe" - Note gbrain integrations doctor needs at least one configured Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> |
||
|
Garry TanandClaude Opus 4.6
|
7d49b8b696 |
docs: rewrite install block for agent success
Clone-based install (repo is home base, upgrades via git pull). Repo URL prominent. API keys called out explicitly. Brain repo vs tool repo confusion eliminated. Shell compatibility fixed. Cron jobs named with frequencies. Dream cycle highlighted as the thing that makes the brain compound. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> |
||
|
Garry Tan
|
0ddb63e646 |
Merge remote-tracking branch 'origin/garrytan/v001-skill-sync'
# Conflicts: # CHANGELOG.md # VERSION # package.json |
||
|
|
13773be071 |
fix: community fix wave — 10 PRs, 7 contributors (v0.9.1) (#65)
* fix: security hardening — search DoS, slug hijack, symlink traversal, content bombs, stdin guard 4 security vulnerabilities closed: - Search limit clamped to 100 (MAX_SEARCH_LIMIT) with statement_timeout 8s - Frontmatter slug authority enforced (path-derived, mismatch rejected) - Symlink traversal blocked (lstatSync in walker + importFromFile) - Content size guard on importFromContent (Buffer.byteLength, 5MB) - Stdin size guard in parseOpArgs (5MB cap) Search pagination added (--offset param on search + query operations). Clamp warning emitted when limit is capped. Co-Authored-By: garagon <garagon@users.noreply.github.com> * fix: PGLite concurrent access lock — prevent Aborted() crash File-based advisory lock using atomic mkdir with PID tracking and 5-minute stale detection. Clear error messages show which process holds the lock and how to recover. Co-Authored-By: danbr <danbr@users.noreply.github.com> * fix: 12 data integrity fixes + stale embedding prevention CTE searchKeyword rewrite (SQL-level LIMIT, not JS splice). Write validation on addLink/addTag/addTimelineEntry/putRawData/createVersion. Health metrics now measure real problems (stale_pages, orphan_pages, dead_links). Orphan chunk cleanup on empty pages. Embedding error logging. contentHash now covers all PageInput fields. Stale embedding NULL'd when chunk_text changes (prevents wrong vector on new text). hybridSearch stops double-embedding query. MCP param validation. type/exclude_slugs search filters now work. pgcrypto extension for Postgres <13. Co-Authored-By: win4r <win4r@users.noreply.github.com> * perf: 30x embedAll speedup + O(n²) fix + ask alias Sliding worker pool (concurrency 20, tunable via GBRAIN_EMBED_CONCURRENCY). O(n²) chunk lookup in embedPage replaced with Map. gbrain ask alias for query (CLI-only, not in MCP tools-json). .idea added to .gitignore. Co-Authored-By: stephenhungg <stephenhungg@users.noreply.github.com> Co-Authored-By: sharziki <sharziki@users.noreply.github.com> Co-Authored-By: hnshah <hnshah@users.noreply.github.com> Co-Authored-By: doguabaris <doguabaris@users.noreply.github.com> * chore: bump version and changelog (v0.9.1) Community fix wave: 10 PRs, 7 contributors. 4 security fixes, PGLite crash fix, 12 data integrity fixes, 30x embed speedup, search pagination, ask alias. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> --------- Co-authored-by: garagon <garagon@users.noreply.github.com> Co-authored-by: danbr <danbr@users.noreply.github.com> Co-authored-by: win4r <win4r@users.noreply.github.com> Co-authored-by: stephenhungg <stephenhungg@users.noreply.github.com> Co-authored-by: sharziki <sharziki@users.noreply.github.com> Co-authored-by: hnshah <hnshah@users.noreply.github.com> Co-authored-by: doguabaris <doguabaris@users.noreply.github.com> Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com> |
||
|
|
baf3517868 |
feat: v0.9.0 -- smart file storage, publish, production-grade skills (#62)
* feat: battle-tested skill patterns from production deployment Backport production-learned brain-operations patterns: - Iron Law of Back-Linking (mandatory bidirectional linking) - Brain filing rules (file by primary subject, not format) - Enrichment protocol (7-step pipeline, 3-tier system, person/company templates) - Media ingest workflows (articles, videos, podcasts, PDFs, screenshots) - Citation requirements (mandatory [Source: ...] on every fact) - Test Before Bulk operating principle - Voice recipe: unicode crash fix, PII scrub, identity-first prompt, DIY STT+LLM+TTS - X-to-Brain recipe: image OCR, Filtered Stream, tweet rating rubric, cron stagger * chore: bump version and changelog (v0.8.1) Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> * docs: add _brain-filing-rules.md to CLAUDE.md key files * feat: smart file upload with TUS resumable and .redirect.yaml pointers - Supabase Storage auto-selects upload method by file size: < 100 MB standard POST, >= 100 MB TUS resumable (6 MB chunks + retry) - Signed URL generation for private bucket access (1-hour expiry) - New `upload-raw` command with size routing: small text stays in git, large/media files go to cloud with .redirect.yaml pointer - New `signed-url` command for generating access links - File resolver supports both .redirect.yaml (v0.9+) and .redirect (legacy) - Redirect format upgraded: 10 fields with full metadata - All migration commands (mirror, redirect, restore, clean) handle both formats * feat: skills reference actual gbrain file commands - Filing rules document upload-raw, signed-url, and .redirect.yaml format - Ingest skill uses gbrain files upload-raw for raw source preservation - Maintain skill adds file storage health checks - Setup skill adds storage configuration phase with migration guidance - Voice recipe uses upload-raw for call audio storage - Migration v0.9.0 with complete storage setup instructions * chore: bump version and changelog (v0.9.0) Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> * feat: gbrain publish -- shareable HTML with password protection First code+skill pair: deterministic code does the work (strip private data, encrypt with AES-256-GCM, generate self-contained HTML), the skill tells the agent when and how to use it. 34 new tests. See: https://x.com/garrytan/status/2042925773300908103 * feat: backlinks check/fix, page lint, and report commands Three new deterministic tools (zero LLM calls): - gbrain backlinks check/fix -- scans brain for entity mentions without back-links, creates them. Enforces the Iron Law from the skills. - gbrain lint [--fix] -- catches LLM preambles, code fence wrapping, placeholder dates, missing frontmatter, broken citations, empty sections. --fix auto-strips fixable artifacts. - gbrain report --type <name> -- saves timestamped reports to brain/reports/{type}/YYYY-MM-DD-HHMM.md for audit trails. 33 new tests (409 total, 0 fail). * feat: v0.9.0 migration tells agents to swap scripts for built-in commands Migration file now: - Lists all 5 new deterministic commands with usage examples - Includes a script-to-command replacement table (old -> new) - Tells the agent to find custom script references in AGENTS.md, skills, and cron jobs and replace with gbrain commands - Adds recommended cron jobs for daily backlink fix + weekly lint - References the Thin Harness, Fat Skills thread * fix: CLI routing bugs found during DX review - Fixed subArgs reference error in handleCliOnly (used wrong variable name) - Renamed gbrain backlinks check/fix to gbrain check-backlinks to avoid conflict with existing backlinks operation (per-page incoming links) - Added TOOLS section to --help output showing publish, check-backlinks, lint, report - Added upload-raw and signed-url to FILES section in --help - Updated all docs/migration references to use check-backlinks * fix: security hardening from adversarial review - XSS: sanitize marked.parse() output (strip script/iframe/on* attrs) - Path traversal: validate report --type against [a-z0-9-] pattern - TUS: HEAD request before retry to get server's actual offset (TUS spec) - Pointer: upload-raw now includes pointer content in JSON output - Symlinks: use lstatSync in all walkers to prevent directory escape --------- Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com> |
||
|
Garry Tan
|
87bb2a59eb |
fix: security hardening from adversarial review
- XSS: sanitize marked.parse() output (strip script/iframe/on* attrs) - Path traversal: validate report --type against [a-z0-9-] pattern - TUS: HEAD request before retry to get server's actual offset (TUS spec) - Pointer: upload-raw now includes pointer content in JSON output - Symlinks: use lstatSync in all walkers to prevent directory escape |