admin/gbrain
1
0
Fork 0
mirror of https://github.com/garrytan/gbrain.git synced 2026-08-14 00:48:18 +00:00
Code Issues Packages Projects Releases Wiki Activity
963 Commits 590 Branches 27 Tags
master
Commit Graph
7 Commits
Author SHA1 Message Date
Garry TanandClaude Fable 5 c2cc8b0207 v0.45.1.0 feat(context): hook-lane feedback loop + cross-turn dedupe for the harness hooks (#4025) 
* refactor(context): extract pure gateVolunteeredPointers (no behavior change)

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* feat(context): volunteer-shaped resolve IPC — channel/cwd/volunteer fields, handler-result widening, server-side gate + validated source + delivery-point logging; listener on both engines

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* feat(cli): gbrain volunteer-hook — one-shot harness hook entry (IPC-first; Postgres-only fallback; startup-hook-exempt)

T0 evidence: a UserPromptSubmit hook's additionalContext is recorded in the
session transcript as a structured hook_additional_context attachment and
transcript_path is stable across turns — cross-turn dedupe reads ONLY those
attachments (never raw tail substrings). Live smoke: seeded brain + running
serve → pointer injected via IPC at 0.43s wall, turn 2 suppressed, one event
banked channel=claude-code.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* feat(recipes): claude-code-reflex + codex-reflex (experimental) hook adapters

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* feat(doctor): volunteer_channels per-adapter visibility

Engine-aware sibling of retrieval_reflex_health: groups
context_volunteer_events by channel (7d) so operators can see which push
adapters fire; message separates installed-but-unregistered from
registered-but-quiet; pre-v117 tolerant; info-only status.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* feat(context): hook-lane feedback loop + cross-turn dedupe — converge cathedral-3 onto the #3975 harness hooks

The #3975 hook lane volunteered pages invisibly (nothing logged to
context_volunteer_events) and re-volunteered the same page every turn an
entity was mentioned (no priorContextText). This convergence ports the
cathedral-3 branch's two gap-fills into the blessed lane and retires its
parallel surface:

- transcripts: parseTranscript extracts injectedContextBlocks (structured
  hook_additional_context attachments — the T0-verified recorded shape of
  our own injections; never raw-turn substring matching)
- hook user-prompt: feeds those blocks as priorContextText (volunteer once
  per session, not once per mention) + sends a feedback-loop channel
  (--harness <claude-code|codex>, default claude-code); hook joins
  STARTUP_HOOK_SKIP_COMMANDS (per-prompt check-update spawn class, grep-pinned)
- turn-context: result exposes post-trim volunteered survivors
- resolve-ipc: additive TurnContextRequest.channel + onTurnContextDelivered
  seam (fires post-write only — abandoned blocks are never counted)
- serve: delivery callback logs volunteered pages + pointers per channel
- doctor: volunteer_channels check (per-channel 7d visibility; separates
  installed-but-unregistered from registered-but-quiet; pre-v117 tolerant)

Retired (never released): gbrain volunteer-hook, hook-adapters.ts,
claude-code-reflex + codex-reflex recipes, the volunteer-shaped resolve wire
extension. Kept: gateVolunteeredPointers extraction, VolunteerChannel
claude-code/codex values, the real hook-transcript fixture.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* docs: regenerate llms bundles for the hook-lane doc updates

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* test: coverage for the hook-lane delivery wiring

Extract logTurnContextDeliveryFireAndForget into volunteer-events.ts so the
shipped serve callback is unit-testable (channel fallback, sessionId clamp,
pointer+page rows in ONE synchronous registration — the double dynamic import
could register the pending write too late to survive process exit). Direct
units for isVolunteerChannel, explicit-channel logDeliveredReflexPointers,
and the pure gate (idempotency pin; stale parity claim in its doc corrected).

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* fix: pre-landing review fixes — 7-reviewer batch

Multi-confirmed: volunteer_channels now runs on the LOCAL doctor path too
(buildChecks; docs pointed there); transient DB errors no longer misreported
as pre-v117; priorContextText deduped + 32KB-capped hook-side. Red team:
delivery counts reconciled against the hook heartbeat (CAUTION when mostly
degraded — write-accept isn't injection); IPC clamp drops advisory
priorContextText BEFORE any window turn; foreign hook_additional_context
blocks (no gbrain marker) excluded from dedupe input; engine-aware quiet
guidance (Postgres brains aren't told to chase hook registration); 2MB dedupe
horizon documented. Security: wire channel claims restricted to harness
channels. DRY: VOLUNTEER_CHANNELS single source, shared reflexPointerRationale,
SESSION_ID_MAX_LEN shared with the op clamp; dead channel param dropped from
logDeliveredReflexPointers. Tests: budget-trim volunteered invariant under
real trim pressure, poll/ordering-based delivery-seam tests (no fixed sleeps),
doctor branch + engine-parity-row tests, wiring source pins, cap pins;
fixture pruned to load-bearing lines (environment inventory removed).

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* fix: adversarial-review fixes — cross-model P1 + Claude adversarial batch

P1 (codex structured review + Claude adversarial, cross-model): the remote
run_doctor path now threads the caller's source scope into
volunteer_channels — a source-bound token can no longer see other sources'
activity counts/timestamps (must-never-violate source-isolation rule);
local trusted doctor stays brain-wide. Claude adversarial: IPC connection
handler gains a one-request re-entrancy guard (trailing bytes mid-await
double-processed the same line → duplicate handler work + duplicate
delivery logging); partial client-side trims now heartbeat as reason
'trimmed' (delivery-count drift visibility); heartbeat reconciliation uses
the same 7-day window as the counts + a 5-sample floor; one oversized
dedupe block no longer evicts all smaller ones (continue, not break);
quiet-channel guidance no longer claims old serves attribute hook traffic
to reflex (they log nothing); 'transient' claim softened to evidence;
marker-filter + unknown-channel-default comments state their honest limits.
Pins: scoped-query SQL/params, re-entrancy single-fire, updated wiring grep.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* chore: bump version and changelog (v0.45.1.0)

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* chore: restamp bootstrap runbook + regenerate vendored template repo for v0.45.1.0

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* docs: update project documentation for v0.45.1.0

Post-ship /document-release pass: fold the post-doc review fixes into the
reference docs (KEY_FILES: source-scoped remote volunteer_channels + heartbeat
reconciliation, 32KB-capped dedupe input, marker filter, IPC one-request
guard + clamp order), fix BRAINBENCH's now-stale memoryless-hook rationale,
and mark the doctor-check half of the push-context TODO shipped.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* docs: cross-model doc-review fixes for v0.45.1.0

Codex doc review against the shipped diff: push-context channel table gains
the harness-hook row (three-channels claim was stale), the guide states the
PGLite-socket topology limit and the bounded dedupe horizon, delivery-point
claims are hedged to match the code (write-accept isn't proof of injection;
the doctor reconciles against the hook heartbeat, which can't inspect
registration), BRAINBENCH's future-tense hooks-PR sentence updated now the
hook has landed, and the adapter-flip + resolver-channel TODOS reflect the
current lane count. llms-full.txt regenerated.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* docs(hook): document --harness in the user-prompt usage text

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

---------

Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
 2026-08-11 22:45:44 -07:00
Garry TanandClaude Fable 5 d35c9c9e44 v0.45.0.0 feat(bootstrap): paste-in personal-agent install for Codex + Claude Code (#3975) 
* docs(designs): agent-bootstrap plan + design docs (normative, review-absorbed)

The scrubbed, in-repo sources of truth for the gbrain bootstrap wave:
AGENT_BOOTSTRAP_DESIGN.md (product scope/sequencing) and
AGENT_BOOTSTRAP_PLAN.md (implementation; all review-finding IDs inlined).

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* feat(bootstrap): format spec, question bank, identity templates, bundled assets

agent.json manifest (format_version 1, initialized sentinel) + machine-local
install receipt [CX2-1, CX2-12]; 12-question/6-required interview bank with
consent keys and a persist:false sink for the optional provider key [CX2-13];
ten {{TOKEN}} identity templates (generic, adapted to gbrain ops — gates call
recall/query/put_page, write-through-ops rule, keyless agent-authored facts,
silence contract); assets embedded compiled-binary-safe via file-type imports
[ENG-6].

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* feat(bootstrap): runbook, README paste block, bootstrap guide, TODOS entries

BOOTSTRAP_FOR_AGENTS.md (agent-driven install runbook: CLI phase list is the
source of truth, never-invent rules, Codex approvals preflight, keyless posture,
failure-modes table, version stamp for the skew check); README gains the
full-agent paste block pinned to latest-stable inside the Claude Code/Codex
quick start (memory-only tier stays); docs/guides/bootstrap.md carries the full
install/security/consent/degradation/uninstall contract.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* docs(designs): spike instrument for the bootstrap wave (build order 0)

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* feat(bootstrap): interview + render engines

Interview gate with read-back confirm-hash (any later answer change clears the
confirmation — the hostile single-batch case is structurally impossible),
per-answer provenance, caps + escaping at set time, config-sink routing for the
provider key; renderer with hard-fail token sweep, subordinate fencing of
principal input, never-clobber + backups, deterministic minimal mode for the
template repo, scaled byte floors. 58 unit tests.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* feat(bootstrap): private-repo lifecycle — repo create, attach, uninstall, run lock

gh-gated private repo creation with API-verified privacy (rate-limit distinct
from public), refuse-foreign-origin with attach as the sanctioned path, atomic
bootstrap mutex (pid liveness + age + token), receipt-keyed uninstall that
never wholesale-deletes the gbrain home and only offers --delete-brain for a
brain it created; read-only PGLite lock probe (never opens the engine).
54 unit tests, injectable exec seam throughout.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* feat(release): latest-stable ref, template-repo publish job, bootstrap CI guards

release.yml advances the latest-stable tag only after assets publish (the paste
block's permanent ref — copies in the wild never rot) and gains a PAT-gated
publish-template job verified against the vendored tree; two skip-graceful
guards (sanctioned-ref + runbook stamp; template/token bijection + placeholder
assertion + generator byte-diff) wired into verify; README + runbook re-admitted
to the CI cache hash; vendored deterministic template tree generated.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* feat(context): IPC v2 turn_context + 8KB assembly + visibility resolver + session identity

Discriminated-union IPC with handler map, protocol echo (stale-serve detection),
shared-secret gate, server-side source binding, per-kind budgets; turn-context
assembly (reflex pointers + volunteered pages + world-only hot facts) under a
data-not-instructions envelope trimmed to the harness's 10KB hook-output cap;
facts.default_visibility resolved through one helper at all four sites (explicit
caller wins, typos fail closed); typed sessionId threads _meta.session_id into
the hot-memory cache key. 50 new tests; 180 adjacent tests confirmed green.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* feat(persistence): secret-scan, gbrain sources push, durability unification

Pattern secret scanner (own runtime allowlist, redacted previews, corpus-write
redaction mode); sources push runs the whole scan→stage→commit→pull→push
sequence under one cross-platform lock (mkdir-atomic, pid+age+token) with a
deny-glob backstop, commit-first divergence-safe pull, refuse-unverifiable
visibility, and push-status telemetry; gbrain-home choke point unifies
GBRAIN_HOME semantics with config (0700); durability is parent-repo-aware and
rotates its push log at 0600. 35 new tests; 200 existing green.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* fix(sources): harden/pull gates accept sources inside a parent git repo

The bootstrap workspace registers brain/ (a subdirectory) as the source; the
durability core already resolves the repo root, so the command gates now check
inside-a-repo rather than .git-right-here [CX2-3].

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* feat(serve): resident maintenance sweep + keyless capability probe

The lock-owning serve process now closes the persistence loop: startup (3s
post-connect, best-effort, unref'd) and idle (10-min quiet intervals through
the injectable timer seam) sweeps run facts-fence reconciliation, deterministic
link/timeline extraction over recent workspace pages, and spend-gated corpus
ingest (skipped keyless — agent-authored fences cover it). gbrain sweep --once
is the trusted CLI seam bootstrap verify uses. Capability probe renders the
honest keyless/keyed report. Full reuse of the cycle extractor + extract cores;
26 new tests, neighbors green.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* feat(hooks): engine-free gbrain hook command, settings writers, transcript parser

Four hook events (session-start digest + crashed-session recovery push,
user-prompt turn-context injection under an 800ms deadline and the 10KB cap,
stop buffers, session-end corpus write with redaction/retention/dedup +
best-effort push); structural JSON settings merger keyed by a _gbrain marker
(foreign hooks and permissions survive); dated host-spec registry; Claude Code
.jsonl parser as a spec-target with a scrubbed 7-shape fixture. Heartbeat is
counters-only by construction. 59 tests; zero engine modules in the import
graph.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* docs(bootstrap): cross-link the full-agent path from the connection docs

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* feat(bootstrap): dispatcher, verify, status — the command assembled

gbrain bootstrap {status,interview,render,repo,hooks,verify,uninstall,attach}:
engine-free except verify (owns its engine, in-process sweep — no live-serve
conflict); phase list is the TS source of truth with install.jsonl telemetry
and the support blob; verify's fail-soft check suite covers the real write path
(put_page → write-through file → sweep → graph floor → recall), passes keyless,
persists snapshots, and ends with the first-run tour. cli.ts wired per the
three-touchpoint rule; doctor gains the bootstrap check group (silent on
machines with no bootstrap state). 28 new tests; 353 adjacent green.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* docs: KEY_FILES bootstrap cluster + CLAUDE.md dispatcher row (+ build:llms)

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* test(bootstrap): e2e pins — hook-under-live-serve, attach, degraded modes, compiled binary, Docker harness

The permanent pins: a real serve holds the PGLite lock while the engine-free
hook completes (and a direct engine open provably throws LiveServeLockError);
stale-socket fail-open; machine-2 attach with marker-keyed hook repair;
decline-everything installs verify green with every degradation named; the
compiled binary renders bundled templates in an empty cwd. Offline Docker
harness (networkless, read-only) gated into heavy-tests.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* fix(bootstrap): register doctor check categories + system-of-record allow comments

The six bootstrap doctor checks join OPS_CHECK_NAMES; the sweep's batch link/
timeline inserts carry the explicit extract-path allow comments (the sweep IS
the extraction path for workspace pages).

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* fix(test): shard wedge cap tracks suite growth (1500s -> 1800s)

At ~9000 tests a healthy shard finished at 1466s and two progressing shards
were false-killed at the old cap; 1800s restores ~25% headroom over the
slowest observed healthy shard. Real hangs still hit it.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* test(ci): cache-hash policy — README + runbook edits must invalidate [C2]

The old deny-list assertion predates the paste block; README.md and
BOOTSTRAP_FOR_AGENTS.md are policy-doc re-admissions now, so their edits must
change the hash (a paste-block edit shipping under a cached green was the C2
hole).

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* fix(test): classify post-suite exit-hangs as warn-pass; file the leak forensics

A shard killed by the wedge watchdog with every assigned file started and zero
fail markers did all its work and leaked a handle at exit — pre-existing and
master-reproducible (P1 TODO carries the full bisect forensics). Bun's per-test
timeout turns a hung test into a (fail), so the classifier cannot mask one.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* fix(test): shard cap 2400s — the count-balanced heavy shard needs it under contention

Observed: the heavy shard still progressing 22s before an 1800s kill while
siblings finish at 1150-1550s (split balances file count, not weight). Filed
the load-sensitive WAL-repair flake (pre-existing, master's v0.42.75.0 wave).

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* test(bootstrap): quarantine env-mutating suites to the serial lane

check-test-isolation R1: six new files mutate GBRAIN_HOME/env at module scope —
the serial lane (one process per file) is the guard's prescribed home for them.
All 114 tests pass post-rename.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* docs(readme): per-harness install sections — Codex, Claude Code, then OpenClaw/Hermes

Each harness gets its own complete paste-block section (desktop app first,
terminal noted — Claude Code CLI is the identical harness; Codex CLI works
pull-based today); the OpenClaw/Hermes platform path keeps equal weight with
its one-click deploys and INSTALL_FOR_AGENTS block intact; memory-only and
remote-connect tiers consolidated under 'Lighter ways in'. Supersedes the
review's D5 ordering by user direction; stale heading references updated.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* docs(readme): Codex as the recommended first step; OpenClaw/Hermes framed as-intended, high-cost

The install section now routes newcomers explicitly: Codex first
(subscription-priced, nothing to deploy), OpenClaw/Hermes as GBrain used the
way it was designed — always on, at real server + API cost.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* fix: doc-audit code chasers — broken recovery hints, stale op description, auto_link key

Four small code fixes surfaced by the markdown accuracy audit:
- doctor's auto-RLS recovery hint pointed at `apply-migrations --force-retry 35`,
  which cannot work (--force-retry targets the vX.Y.Z orchestrator registry, not
  the numeric schema MIGRATIONS array). Hint now points at the recreate SQL in
  docs/guides/rls-and-you.md; test pins against regression.
- v0_11_0 migration printed the same broken-mechanism class of hint
  (`config set minion_mode` writes DB config nothing reads); now names
  `apply-migrations --mode` + preferences.json, the real setter.
- submit_job's op description hardcoded a stale handler list; now points at
  registerBuiltinHandlers as the source plus the --follow discovery trick.
- `auto_link` added to KNOWN_CONFIG_KEYS: read by link-extraction, reconcile-links,
  and sweep, and documented as the off-switch in brain-ops/maintain, but the
  allowlist rejected `config set auto_link false`.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* docs: repo-wide accuracy + MECE reform from the 9-bucket markdown audit

A code-grounded audit of every markdown file (root, architecture, guides,
mcp, tutorials, docs-root, operations/eval/designs, skills, recipes) followed
by a fix wave with per-bucket ownership. Four classes of change:

Accuracy — every documented command/flag verified against src/ before writing:
dead commands replaced with working ones (pages purge-deleted, jobs watch
--follow, gbrain restore, import-based Obsidian flow, space-separated --scopes,
real thin-client recipes, working isolation verification, real supervisor
restart procedure, curl-based ngrok health check, real minion_mode setter);
count drift fixed with rot-proof phrasing (100+ ops, 50+ bundled skills via
skills/manifest.json, 140+ engine methods, KNOBS_HASH_VERSION pointer instead
of hardcoded versions); stale claims corrected (search-mode defaults, RETRIEVAL
pipeline order incl. autocut, sentinel rules, refusal-list mechanism, engine
snapshot, shard cap 2400s + EXIT-HANG classifier in TESTING.md, latest-stable +
publish-template documented in RELEASING.md as release.yml promises).

MECE — one home per concept, pointers elsewhere: test isolation → TESTING.md;
OAuth registration + --bind/--public-url lore → DEPLOY.md; mode bundles →
guides/search-modes.md (the home the CLAUDE.md dispatcher always promised);
merge contract → schema-packs.md; WAL ladder → ENGINES.md; quiet-hours →
quiet-hours.md; capture taxonomy → entity-detection.md; person-page taxonomy →
compiled-truth.md; brain-first protocol → brain-first-lookup.md; refresh
semantics → refresh-algorithm.md; KEY_FILES.md deduplicated (58 extension
entries merged, one entry per file); infra-layer.md rewritten as a pointer page.

Privacy — placeholder sweep across guides, docs, skills, and recipes per the
iron rule; per-release narration stripped from reference docs (current-state
prose only).

Bootstrap coverage — AGENTS.md pointer, RESOLVER routing row, INSTALL.md path,
tutorial cross-links, keyless-mode sections in spend-controls/headless-install.

skills.lock.json regenerated; llms.txt/llms-full.txt rebuilt. Gates: verify
36/36, typecheck clean, doctor 96/96, skills-integrity + resolver + build-llms
+ config-set + migrations all green.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* docs(readme): refresh production-brain stats to current brain-repo counts

155,795 pages / 24,589 people / 5,340 companies, counted from the brain
repo's current HEAD; the "100K-page brain" framing moves to 150K to match.
Cron-fleet count unchanged (its store lives on the deployment host, not in
the repos available for verification).

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* fix(scan,push): modern OpenAI/Voyage key patterns, scan staged blobs not disk

- secret-scan matches sk-proj-/sk-svcacct-/sk-None- and pa- Voyage keys
  (the bare sk- pattern missed every current OpenAI key format).
- workspacePush stages first, then scans the staged index blobs via
  git cat-file, closing the scan-then-stage TOCTOU where a file changed
  between snapshot and commit shipped unscanned.
- shared binary-sniff helper, memoized glob regexes, atomic push-status
  write, and tests for pull_conflict + gitignored deny-match paths.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* fix(ci): regenerate flag registry for new commands, harden shard classifier + release token

- cli-flag-registry.generated.ts regenerated: bootstrap/hook/sweep and
  sources push --message/--allow-unverified-remote were missing, so the
  strict #2185 validator rejected real invocations and skipped the new
  commands entirely.
- EXIT-HANG shard classifier now requires every assigned file to have
  started before warn-passing a watchdog kill (was fail-open).
- release.yml passes TEMPLATE_REPO_PAT via http.extraheader, off the argv.
- compiled-binary e2e fails loud in CI instead of a silent permanent skip.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* fix(hook,sweep,ipc): non-blocking hook pushes, bounded sweep + cache, source-bound resolve

- session-start/session-end no longer run synchronous git + inline push
  inside their self-deadline; a detached child does the push and the hook
  returns immediately (blocked Claude Code startup for minutes on a dirty
  tree before).
- serve sweep drops the unbounded listAllPageRefs, resolves only candidate
  targets, claims corpus files atomically (no double-LLM-spend race), and
  caps the fence LIKE scan; heartbeat writes are O_APPEND with rare compaction.
- hot-memory cache evicts expired entries and bounds entry count (the key is
  caller-controlled via _meta.session_id).
- v1 resolve IPC honors boundSourceId like turn_context; turn-context runs
  its arms concurrently. doctor reads push/heartbeat thresholds from hook.ts.
- new tests: doctor bootstrap checks, hook push-gate + deadline, concurrent
  sweep claims, cache eviction, bound-source resolve.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* fix(bootstrap): origin-ownership gate, world visibility, collision-safe source id, consent + templates

- repo adoption requires an exact receipt repo_url match or authed-owner
  check (undefined repo_url was a wildcard); create verifies privacy BEFORE
  the first push.
- verify sets facts.default_visibility=world if unset, so agent-authored
  facts surface in per-turn context (they defaulted private before).
- source_id derives a path-hash suffix when 'workspace' is taken by another
  checkout; every consumer reads manifest.source_id.
- skipped HOOKS_CONSENT now declines (was falling through to default yes);
  --minimal refuses on an initialized manifest; tilde fences escaped.
- MCP registration pins --surface full; status hard-fails a public origin
  (template door); receipt writers guard against newer/corrupt receipts;
  uninstall only claims brain-deleted after a real rm.
- templates ship jobs disabled + provider-consent + support-relay lines;
  soul-audit re-runs over the shared interview bank. TODOS: 11 follow-ups.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* fix(security): close adversarial-review findings — scan fails closed, whole-PEM redaction, bound repo push

Cross-model adversarial pass (Claude + Codex) on the bootstrap wave:

- secret scan fails CLOSED: an unreadable, oversized, or binary staged blob
  now blocks the push (blocked_unscannable, exit 5) instead of committing
  unscanned; only a confirmed staged deletion is skipped. This was the
  headline "block secrets before they leave the machine" property failing open.
- private-key redaction spans the whole PEM block (header+body+footer), not
  just the header line — the base64 body no longer survives into the corpus
  the sweep sends to an extraction provider.
- bootstrap repo commits the workspace (secret-scan-gated) before the first
  push and verifies the remote actually received it, so a push-fail retry
  can't adopt an empty remote as success.
- privacy verify is re-bound to origin immediately before push (a concurrent
  origin rewrite between verify and push is refused).
- session-end corpus write is atomic and clears the stale ingested/in-progress
  sidecars so a resumed session's appended transcript is re-ingested.
- public-origin refusal enforced at render (not only status); MCP "already
  registered" is verified to target this workspace, not blessed blindly;
  verify probe cleanup scopes deletes to its own slugs, not a token substring;
  allowlist fingerprint floor raised 8→16 hex.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* v0.45.0.0 feat(bootstrap): paste-in personal-agent install for Codex + Claude Code

Turns a Codex or Claude Code session into a persistent personal agent:
interview-rendered identity files, a local PGLite brain, per-turn context via
serve IPC (Claude Code hooks / Codex pull protocol), session-triggered
persistence, and a private GitHub repo as the agent's portable body. Keyless-
first (the harness model is the LLM; one optional key adds embeddings +
extraction). New `gbrain bootstrap` command family + `gbrain hook` + `gbrain
sweep`; doctor bootstrap health checks; latest-stable distribution ref +
template-repo publish job. Opt-in, additive — existing installs untouched.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* fix: regenerate flag registry for security-fix flags; drop fabricated gbrain capabilities doc ref

CI caught two real failures under the merged state:
- the flag registry lagged the blocked_unscannable/exit-5 flags the security
  round added, tripping the #2185 freshness guard.
- headless-install.md described the keyless capability report as a
  `gbrain capabilities` command, which the #3502 doc-command resolver
  rejects — reworded to prose (the real surface is bootstrap verify's report).

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* docs: sync KEY_FILES + bootstrap plan to security-fix behavior

Cross-referenced the security-fix round against the reference docs and
corrected the drift those commits introduced:

- workspace-push.ts entry: stage-FIRST-then-scan order (the TOCTOU fix),
  fail-closed blocked_unscannable, and the sources-push status -> exit-code map.
- hooks.ts entry: MCP registration pins `serve --surface full`.
- hook.ts entry: session-start/session-end pushes run in a detached child
  (non-blocking); atomic corpus write clears stale sidecars.
- bootstrap.ts entry: render hard-refuses a public origin (template door).
- verify.ts entry: source_id collision resolution (workspace-<path-hash>).
- AGENT_BOOTSTRAP_PLAN as-shipped delta note for the scan/stage reorder.

llms bundle unchanged (KEY_FILES is link-only); build:llms and
test/build-llms.test.ts green.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* fix(ci): silence SC2016 on the intentional askpass literal in release.yml

The one-shot GIT_ASKPASS script must contain literal $1 and
$TEMPLATE_REPO_PAT so they expand when /bin/sh runs it at git's credential
prompt, not when the outer shell writes the file — single quotes are correct.
Add a scoped shellcheck disable so actionlint passes.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* fix(skills): declare 'bootstrap my data' trigger in cold-start frontmatter

The doc reform added 'bootstrap my data' to cold-start's RESOLVER.md row (to
disambiguate data-bootstrap from agent-bootstrap) but not to the skill's own
frontmatter triggers, tripping the RESOLVER↔frontmatter round-trip contract
(resolver.test.ts). Declare it; regenerate skills.lock.json.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* feat(bootstrap): default per-turn hooks + search mode ON without a prompt

Installing gbrain for your coding agent IS the consent for the behaviors that
make it work, so stop re-litigating them with install-time questions whose
"no" defeats the product:

- Per-turn hooks (Claude Code) install ON by default — no prompt. Off-ramps:
  `--no-hooks` at install, `GBRAIN_HOOKS=0` at runtime, `bootstrap uninstall`.
  The "hooks installed" line now surfaces the kill switch so default-on is
  never silent. A persisted HOOKS_CONSENT=no (interview --skip) still declines.
- Search mode defaults to `balanced` silently (nobody knows the modes at
  install; `gbrain search modes` changes it any time).
- MCP scope stays the ONE deliberate prompt — project vs user is a real
  cross-repo privacy choice, not friction.

Marks the two consents `silent: true` in the question bank (new QuestionSpec
field), rewrites the runbook phases so the agent no longer asks them, adds the
`--no-hooks` flag (+ registry regen), and adds default-on / opt-out tests.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* test(bootstrap): real end-to-end coverage — cross-session recall, per-turn content, Codex door, realistic corpus

Closes the seven e2e gaps a coverage audit surfaced: the plumbing was
well-unit-tested but the product claims ("Codex works, context shows up every
turn with real content, it remembers across restarts, machine two recovers,
Postgres works") were unproven end to end. Test-only wave — zero src changes.

- Hermetic synthetic corpus (test/fixtures/bootstrap-corpus/ + a loader helper):
  12 interlinked pages (52 edges, timelines), 12 world/private beliefs, 8 gold
  queries — curated from the gbrain-evals synthetic corpora, 100% placeholder
  names, so recall is asserted on a real multi-entity brain instead of a
  2-node self-planted probe.
- GAP1 magic moment: author a fact via the real write path, disconnect the
  engine, reopen against the same DB, recall it — a real session boundary, not
  verify.ts's same-connection SQL read-back. Plus a source-isolation assertion.
- GAP2 per-turn content: hook-under-serve Pin 1 now seeds a known fact and
  asserts its text lands in the injected block AND private beliefs never do
  (was: empty brain, empty_block accepted as a pass).
- GAP3 Codex door: assert the rendered AGENTS.md carries the Gate-3 brain-first
  pull protocol; make the fake codex shim implement `mcp get` so the [FIX7]
  target-verification can actually fail; the Docker cold-machine harness now
  exercises the hooks/MCP registration step instead of skipping it.
- GAP4 corpus recall: turn-context + verify graph-floor/qrels run on the real
  multi-entity brain with real edges.
- GAP5 attach: machine-two now re-ingests the cloned brain/ into a fresh DB and
  recalls a fact authored only on machine one — the multi-device payoff.
- GAP6 keyed + Postgres (env-gated): real embeddings prove semantic recall a
  paraphrase query can reach but keyless BM25 cannot; bootstrap verify drives a
  real Postgres engine (skipIf DATABASE_URL/keys absent).
- GAP7 persistence: session-end runs the REAL push (not the mocked seam) to a
  local bare remote and the remote receives the content; a planted secret is
  blocked at the gate; the 15-min cron installs and fires a scan-gated push.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* test(bootstrap): real-agent e2e — drive the actual claude + codex binaries end to end

Closes the audit's biggest gap ("no real harness ever drives a turn"). Adapts
gstack's PTY/headless agent harness to prove the bootstrap install + smoke work
against the REAL binaries, not PATH shims. Test/CI/docs only — zero src changes.

- test/helpers/agent-harness.ts: hermetic clean-room child env (ported from
  gstack; drops CONDUCTOR_/CLAUDE_/GSTACK_/MCP_/GBRAIN_, promotes
  GSTACK_ANTHROPIC_API_KEY→ANTHROPIC_API_KEY), real-binary resolvers + auth
  probes, headless `claude -p --output-format stream-json` and `codex exec
  --json` turn runners, a gbrain stdio MCP-config writer, and a keyless brain
  seeder. + a fixture-parse unit test (no binary needed).
- test/e2e/bootstrap-real-claude.serial.test.ts: real `gbrain bootstrap`
  install → REAL `claude mcp add` (verified via `claude mcp get`) → verify
  exit 0 → a real `claude -p --mcp-config --strict-mcp-config` turn that
  invokes mcp__gbrain__search and answers from the brain (proven: toolCalls
  include mcp__gbrain__search, final text carries the seeded fact).
- test/e2e/bootstrap-real-codex.serial.test.ts: same install with REAL `codex
  mcp add` into a real ~/.codex/config.toml + Gate-3 pull-protocol assertion,
  then a real `codex exec --json` turn surfacing the fact (MCP or the pull-
  protocol shell path). Bounded retry absorbs codex's occasional MCP-call
  cancellation without softening the fact-requiring assertion.
- Everything hermetic (temp HOME/CLAUDE_CONFIG_DIR/CODEX_HOME/GBRAIN_HOME;
  real ~/.codex auth copied read-only) and skipIf-gated so it self-skips
  cleanly where the binaries/auth are absent.
- heavy-tests.yml: gated `real-agent-e2e` job (nightly/label, never the PR
  shard; no-op on a runner without authed binaries).
- TODOS: compiled `gbrain` binary can't serve a PGLite brain (bun compile
  omits the WASM/extension payloads); harness falls back to `bun run` serve.

Verified against live claude 4.6 + codex 0.147.0: 15 pass / 0 fail; verify
36/36; typecheck clean.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* fix(ci): real-agent-e2e job — bash array + --timeout (actionlint SC2086 + bun-test-timeout guard)

The real-agent-e2e job's file loop used an unquoted $FILES (SC2086) and ran
`bun test` without --timeout (check-bun-test-timeout guard). Switch to a bash
array and add --timeout=600000 (real-agent turns are slow; the tests self-skip
without authed binaries so it's a no-op elsewhere).

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* fix(pglite): embed WASM + extension assets so the compiled binary can serve

A `bun build --compile` gbrain binary could not `serve` a PGLite brain: the
compile bundles JS but not PGLite's runtime payload (pglite.wasm, initdb.wasm,
pglite.data, vector/pg_trgm tarballs), so `serve` on PGLite died with a
bunfs/ENOENT. Now the assets ride inside the binary.

- src/core/pglite-embedded-assets.ts: embeds the five assets via
  `import … with { type: 'file' }` (the ENG-6 idiom) and exposes
  getEmbeddedPgliteOptions() → { pgliteWasmModule, initdbWasmModule, fsBundle,
  extensions:{vector,pg_trgm} }. WASM/fsBundle are consumed as bytes; the two
  extension tarballs are materialized to a content-addressed temp file (atomic,
  size-verified reuse) because PGLite reads them via fs.createReadStream, which
  cannot read a /$bunfs path. Unconditional (works in bun-run and compiled),
  so no fragile mode branch.
- src/core/pglite-engine.ts: static-import getEmbeddedPgliteOptions (engine
  path stays static per the engine-dynamic-import invariant); spread into both
  PGlite.create sites (initial + WAL-repair retry). The bunfs classifier stays
  as a backstop but no longer fires for a correct binary.
- scripts/check-pglite-embedded.sh (+ smoketest): compiles a focused binary and
  asserts it boots PGLite, CREATE EXTENSION vector/pg_trgm, and round-trips a
  page — wired into `bun run verify` (now 37 checks), check:all, and
  check:pglite-embedded. Fail-soft only when compile is unavailable.
- agent-harness.ts probeCompiledPglite now passes → the real-agent e2e uses the
  fast compiled MCP server. TODOS: the P2 "can't serve PGLite" item is closed.

Verified: fresh compiled binary ran `search`/`query` against a PGLite brain and
returned the seeded row (no bunfs/ENOENT); verify 37/37; pglite-engine 120/0
source-mode; typecheck clean; engine-dynamic-import + parity guards pass.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

---------

Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
 2026-08-11 16:42:11 -07:00
Garry Tan 75fae742d5 v0.44.0.0 feat(eval): BrainBench — cross-harness memory conformance suite (Cathedral 2) (#2158) 
Public, reproducible, cross-harness benchmark for the four ways agent memory fails (know-to-ask, push precision/recall, write-back fidelity, cross-session continuity). Hermetic PGLite, sealed gold, deterministic corpus, CI gate vs master's committed baseline, published JSON-schema foreign-runner contract. See docs/eval/BRAINBENCH.md.
 2026-08-10 09:44:46 -07:00
Garry TanandClaude Opus 4.8 9a0bae8d62 v0.42.25.0 fix(pricing): unify chat-model pricing into one canonical source; add Opus 4.8 (#1819) (#1827) 
* fix(pricing): unify chat-model pricing into one canonical source; add Opus 4.8 (#1819)

Single canonical CANONICAL_PRICING table (src/core/model-pricing.ts) with
canonicalLookup; ANTHROPIC_PRICING and takes-quality MODEL_PRICING become
derived views. cost-tracker, cross-modal runner, skillopt preflight, brainstorm
orchestrator, and brain-score all source from it. Adds Opus 4.8 ($5/$25) so
--max-cost-usd and the dream-cycle budget meter enforce on 4.8 runs; fixes a
stale Opus 4.7 $15/$75 in the takes-quality gate and reconciles Gemini 2.0 Flash
to $0.10/$0.40. Because every table derives from canonical, cross-table price
drift is structurally impossible.

* chore: bump version and changelog (v0.42.25.0)

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* docs: document canonical chat-pricing table for v0.42.25.0

Add KEY_FILES.md entries for src/core/model-pricing.ts (canonical
CANONICAL_PRICING + canonicalLookup) and refresh the now-derived
anthropic-pricing.ts + takes-quality-eval/pricing.ts entries to
current-state. Add the "one canonical chat-pricing table" cross-cutting
invariant to CLAUDE.md. Fix the stale model-price snapshot pointer in
SEARCH_MODE_METHODOLOGY.md (anthropic-pricing.ts -> model-pricing.ts).
Regenerate llms-full.txt.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
 2026-06-03 23:00:11 -07:00
Garry TanandClaude Opus 4.8 d9eadfec13 v0.42.3.0 feat(search): autocut — score-discontinuity result-sizing (#1663 wave 1) (#1682) 
* feat(search): autocut — score-discontinuity result-sizing on the rerank separatrix

Cut the ranked set at the cross-encoder rerank-score cliff instead of a fixed
top-K. Default-ON in reranked modes (balanced/tokenmax), no-op without a
reranker. New pure src/core/search/autocut.ts; mode.ts knobs + reranker_top_n_in
= searchLimit (no unscored tail); query op autocut param; --explain + glossary.

* test(search): autocut pure-fn, agent-surface, behavioral + precision/recall eval gate

Adds autocut.test.ts, query-op-autocut.test.ts, autocut-integration.serial.test.ts
(IRON-RULE behavioral via rerankerFn seam), autocut-eval.test.ts (in-repo
precision-lift-without-recall-regression gate). Updates existing knobsHash/bundle
pins to v=7 + reranker_top_n_in.

* chore: version + changelog + docs for autocut (v0.41.34.0)

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* fix(search): autocut preserves alias-hop exact matches + cache-HIT meta (codex P1/P2)

P1: applyAliasHop injects the canonical page after reranking (no rerank_score);
autocut would drop it when cutting on the scored set. applyAutocut gains an
optional preserve predicate; hybrid passes r => r.alias_hit === true.
P2: cache-HIT cachedMeta now carries autocut/adaptive_return/mode/embedding_column.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* chore: bump version to v0.42.3.0 (autocut wave)

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* docs: PR titles lead with the version (IRON RULE in CLAUDE.md)

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
 2026-06-01 20:16:37 -07:00
Garry TanandClaude Opus 4.8 d6db3f0ce3 v0.41.34.0 feat(search): retrieval cathedral — max-pool + title + alias + evidence (#1657) 
* feat(search): per-page max-pool in searchVector (both engines)

T1 of the retrieval-cathedral wave (supersedes #1616). Vector search returned
chunk-grain top-k with no DISTINCT ON, so a page could be represented by a
weak chunk while a hub page's chunks crowded a distinct page's strong chunk
out of the candidate set entirely. Keyword search always pooled per page; the
vector path did not.

- New shared buildBestPerPagePoolCte() in sql-ranking.ts — single source of
  truth consumed by searchKeyword + searchVector across postgres + pglite, so
  the two engines can't drift (the recurring parity bug class).
- searchVector both engines: compute score as a select-list expr (HNSW
  ORDER BY stays pure-distance), pool DISTINCT ON (slug) over the full
  candidate set before the user LIMIT, deterministic tiebreak
  (slug, score DESC, page_id ASC, chunk_id ASC).
- All keyword pooling blocks refactored onto the shared builder (DRY).
- Regression test: a hub page's chunks no longer crowd out a distinct page's
  strong chunk; results are one-per-page by best chunk. Fails on old path.

Verified: real-Postgres engine-parity 22/22, PGLite hermetic suite green.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* feat(search): title-phrase boost (page.title first-class signal)

T2 of the retrieval-cathedral wave. A query that is a phrase from a page's
title ("Greek amphitheater" -> "The Mingtang - Indoor Greek Amphitheater")
matched a weak body chunk instead of being recognized as a title hit. Names
of things deserve weight.

- New pure title-match.ts: isTitlePhraseMatch (contiguous token-run inside
  page.title OR exact full-title match). Precision guards: >= 2 content
  tokens OR exact full-title; stopword filter; token-boundary match (no raw
  substring). Reused by the eval later so production + bench can't drift.
- applyTitleBoost post-fusion stage in hybrid.ts: reads page.title (not the
  brittle "first chunk"), floor-ratio-gated, stamps title_match_boost for
  --explain, never touches base_score (the agent's dedup confidence).
- ModeBundle.title_boost knob (1.25, on in all modes - cheap gated
  correctness fix), search.title_boost config key, dashboard description.
- KNOBS_HASH_VERSION 6 -> 7 so a boost-on cache write can't serve a
  boost-off lookup; all version-pin + canonical-bundle assertions updated.
- 18 new tests (matcher 13 + stage 5); typecheck clean.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* feat(search): page_aliases data layer (T3 foundation)

Free-text alias resolution for search. gbrain stored a page's chosen names
in pages.frontmatter `aliases:` JSONB but search never consulted them, so a
query like "Hall of Light" or "明堂" couldn't surface the "Mingtang" page.

DELIBERATELY SEPARATE from slug_aliases (re-grounded against current code):
  - slug_aliases:  old-slug -> canonical-slug (wikilink/get_page redirect,
    populated only from concept-redirect conversions)
  - page_aliases:  normalized free-text name -> canonical slug (search hop)
Overloading slug_aliases would muddy two distinct semantics, so this is a
new table, not an extension (honors DRY by keeping concepts separate).

- src/core/search/alias-normalize.ts: ONE normalizeAlias() (NFKC + lowercase
  + ws-collapse + quote-strip) + normalizeAliasList() shared by the write
  (ingest) and read (search) paths so they match on the same key (CQ2).
- Migration v108 page_aliases (source_id, alias_norm, slug); btree
  (source_id, alias_norm) for indexed-equality hop, NOT ILIKE; unique TRIPLE
  (not source_id+alias_norm) so two pages may claim one alias — collisions
  reported + resolved at query time, not blocked at ingest (Codex#8).
  Mirror in pglite-schema.ts; Postgres fresh gets it from the migration.
- engine.resolveAliases(aliasNorms, {sourceId|sourceIds}) read +
  setPageAliases(slug, source, aliasNorms) write, both engines, source-scoped.
- 17 tests: normalize round-trip, collision, source-scope, replace, clear.

Ingest projection + the hybridSearch alias hop land next (T3 wiring).

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* feat(search): alias hop + ingest projection (T3 wiring)

Wires the page_aliases data layer into ingest (write) and hybridSearch (read)
so a query that is a page's declared chosen name surfaces that page — the
named-thing class neither max-pool nor title-boost can fix (true synonyms with
zero surface overlap: "Hall of Light" / "明堂" -> the Mingtang page).

- Ingest projection (import-file.ts): after the page write commits,
  normalizeAliasList(frontmatter.aliases) -> engine.setPageAliases. Always
  called (even []) so removing an alias clears its row; content_hash includes
  non-timestamp frontmatter so alias edits reach this path, not the skip branch.
  Fail-soft + pre-v108-safe (isUndefinedTableError swallowed).
- applyAliasHop (hybrid.ts), AFTER rerank so a named query reliably surfaces
  its page: FULL normalized-query exact match only (no substring/n-grams),
  skip >6-token prose queries, present-boost 1.10x / inject absent canonical at
  top-of-organic + epsilon (never absolute 1.0, D3), collisions alpha-ordered +
  capped at 3, fail-open on pre-v108 / lookup error (D9). Stamps alias_hit for
  the T4 evidence contract.
- SearchResult.alias_hit attribution field.
- 8 tests: inject/boost/CJK/no-match/long-skip/collision + ingest projection
  round-trip + alias-removal-clears. 73 pass across the T1/T2/T3 + import suite.

Backfill of existing pages' aliases lands as T8 (reindex --aliases).

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* feat(search): evidence/create_safety contract + search→cheap-hybrid + per-call mode (T4)

The agent-facing fix for the incident's ROOT behavior: tonight the agent read a
single blended 0.64 score, decided "no strong match, safe to write a new page",
and wrote a duplicate on a developed concept page. A blended RRF/cosine score is
not a calibrated probability, so the don't-duplicate decision must key off WHY a
page matched, not a raw number.

- evidence.ts: classifyEvidence (alias_hit > exact_title_match > high_vector_match
  > keyword_exact > weak_semantic) + createSafetyFor (exists | probable | unknown).
  stampEvidence runs at the end of every hybrid return path (main + both keyword
  fallbacks). SearchResult gains evidence + create_safety. The agent keys
  don't-duplicate off create_safety='exists', not a score threshold.
- search op → cheap-hybrid everywhere (D4/D15): full vector+keyword+RRF+pool+
  title+alias, expansion OFF (no per-call LLM cost); `query` stays full-control.
  search.mcp_keyword_only escape hatch (D17) keeps the old keyword-only behavior
  for operators who don't want query text sent to an embedding provider.
- Alias hop + evidence now also run on the keyword-only fallback paths (the
  named-thing fix is most valuable exactly when vector is unavailable).
- Per-call `mode` (D5): honored ONLY for local/trusted callers (ctx.remote===
  false) so a remote OAuth client can't escalate to costly tokenmax; local +
  unknown mode rejects loudly; threaded into resolveSearchMode + the cache key.
- 30 tests (evidence classifier incl. before/after-incident cases, per-call mode
  gate, alias hop). Updated mcp-eval-capture to the new cheap-hybrid contract.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* fix(cli): reconcile `gbrain search` dispatch (T5)

After T4 made the `search` op cheap-hybrid, `gbrain search "x"` already does the
right thing — but `gbrain search modes/stats/tune` would have run a hybrid search
for the literal word "modes" instead of opening the config dashboard (the op
intercepts before the unreachable handleCliOnly dashboard path).

Add a pre-dispatch interception in main(): `search` + subArgs[0] in
{modes,stats,tune} → runSearch dashboard (with the v0.41.6.0 read-only connect+
dispatch 10s timeout preserved); everything else (free-text) falls through to the
cheap-hybrid `search` op. Subprocess test pins all three routes:
modes/stats → dashboard, free-text → search op ("No results", not "Unknown
subcommand").

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* feat(eval): NamedThingBench retrieval-quality gate (T6)

The eval that makes the retrieval-maxpool incident impossible to reintroduce
silently. 7 query families, each a failure class the incident exposed:
title-substring, generic-to-named, alias-synonym, multi-chunk-dilution,
short-vs-rich, graph-relationship, hard-negative.

- src/eval/retrieval-quality/harness.ts: pure scoring (Hit@1/Hit@3/MRR per
  family) + injected SearchFn (CLI uses hybridSearch; tests stub it) +
  evaluateGate. D12 gate: hard-gate the families that ARE the bug from day one
  (title-substring Hit@1>=0.95, alias-synonym Hit@1>=0.98, dilution Hit@3=1.0),
  warn-then-enforce the softer families. Env-overridable floors.
- `gbrain eval retrieval-quality <fixture.jsonl> [--json] [--source]` +
  dispatch in eval.ts. Exit 0 PASS / 1 FAIL / 2 USAGE.
- Synthetic fixture (placeholder names only, privacy-grep guarded) + hermetic
  gate test: seeds a synthetic brain, forces the keyword+title+alias path
  (embed transport stubbed to throw — free, deterministic), asserts the bug
  families pass. The vector max-pool guarantee is pinned separately by
  searchvector-maxpool.test.ts.
- CI gate: the hermetic test is a normal unit test, so it runs in every PR
  shard — the gate is live on every change.
- 23 tests (harness unit + hermetic gate + fixture privacy guard).

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* feat(telemetry): rank-1 score drift signal (T7)

Standing observability so a retrieval regression is caught before a human hits
it in chat (like tonight). Aggregate columns on search_telemetry (NOT per-query
rows, D10): sum_rank1_score + count_rank1 + 3 coarse buckets (<0.6 / 0.6-0.85 /
>=0.85). The mean rank-1 base_score is the headline; a downward drift = retrieval
quality regressing.

- hybrid.ts: capture rank-1 base_score at all three return paths, thread through
  emitMeta → recordSearchTelemetry opts (like results_count).
- telemetry.ts: Bucket + record + flush ON CONFLICT-add + readSearchStats expose
  avg_rank1_score (null when no samples — no NaN) + rank1_distribution.
- Migration v109 ADD COLUMN IF NOT EXISTS (both engines; search_telemetry lives
  only in migration v57, so the v57+v109 chain covers fresh + upgrade). Columns
  exempted in schema-bootstrap-coverage (no forward-ref index → no bootstrap need).
- `gbrain search stats` surfaces the avg + bucket line; JSON envelope auto-carries
  the fields. "true-positive" wording dropped per Codex#14 — production has no
  labels, so this is an unlabeled rank-1 score histogram; labeled calibration
  lives in NamedThingBench (T6).
- 3 round-trip tests (mean+buckets, no-result excluded, empty=null).

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* feat(reindex): gbrain reindex --aliases backfill (T8)

Import-time projection (T3) covers new + changed pages; this backfills EXISTING
pages whose frontmatter `aliases:` predate v108 / the projection. Walks
listAllPageRefs (cheap cross-source (source_id, slug) enumeration), reads each
page's frontmatter aliases, writes page_aliases via setPageAliases.

Idempotent (setPageAliases replaces) so re-running is convergent — no op-checkpoint
needed (fast, no embedding). --dry-run reports would-write counts, --source
narrows, --limit caps, --json envelope, progress reporter. Wired into the
`reindex` dispatch alongside --markdown / --multimodal.

4 tests: backfill from array + comma-scalar frontmatter, --dry-run writes
nothing, idempotent second run.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* test(search): pre-migration fail-open regression (T9)

Pins that pre-v108 brains (no page_aliases table) keep working: applyAliasHop
returns input unchanged + doesn't throw, importFromContent with frontmatter
aliases still imports (projection swallows table-missing via isUndefinedTableError),
and resolveAliases surfaces the error for the caller to catch.

Completes the T9 mandatory regression set (dilution → searchvector-maxpool,
dispatch → cli-search-dispatch, MCP contract → mcp-eval-capture, engine parity
→ engine-parity 22/22, pre-migration → here).

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* feat(search): Phase-0 retrieval diagnostic — `gbrain search diagnose` (T0)

The operator-facing trace the user runs against the production brain to pin
which retrieval layer surfaces (or misses) a target page — the diagnostic the
plan front-loaded so we don't ship a fix that doesn't move the incident.

`gbrain search diagnose "<query>" --target <slug> [--json] [--source]` reports,
for the target: keyword rank+score, vector rank+score (skipped/graceful if no
embedding provider), whether the query is a registered alias, and the hybrid
final rank + evidence + create_safety + which boosts fired (title/alias). The
verdict names the layer that surfaces the target at rank 1 (or "none"), telling
you whether the lever is max-pool/innerLimit (vector) vs title/alias.

Wired into the `search` dispatch alongside modes/stats/tune (60s timeout since
it runs real retrieval). 2 hermetic tests (alias-query trace + title-phrase
trace). For the Mingtang incident, run:
  gbrain search diagnose "Greek amphitheater" --target projects/new-greek-theater/concept_v0

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* docs(retrieval): corrected incident record + named-thing layers + glossary (T10)

- RETRIEVAL_MAXPOOL_INCIDENT.md: replaces closed PR #1616's RFC with the
  verified record — what happened, the disease, the corrections to the RFC's
  mechanics (search was keyword-only, --mode unthreaded, hybrid already pooled
  at dedup, aliases dead to search), the four-layer fix that shipped, and the
  triage commands (search diagnose / reindex --aliases / search stats / eval
  retrieval-quality).
- RETRIEVAL.md: new "Named-thing retrieval" section documenting per-page pool +
  title boost + alias hop + the evidence contract, reconciling the doc with the
  shipped pipeline (closes the doc/reality gap).
- metric-glossary.ts + regenerated METRIC_GLOSSARY.md: Hit@1, Hit@3,
  avg_rank1_score (drift signal, not labeled accuracy), and create_safety
  (the evidence contract) now carry plain-English glossary entries.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* test(eval): NamedThingBench fixture privacy guard via slug-shape (T6 fixup)

The banned-name literal list itself tripped check-privacy/check-test-real-names.
Replace it with the load-bearing assertion: every fixture slug must be an
*-example placeholder (no real brain page can be referenced).

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* fix(search): source-isolation in per-page pool + alias hop (P0, codex adversarial)

Codex outside-voice caught two source-isolation P0s in the retrieval wave — the
exact class the v0.34.1 seal guards. Both fixed before merge.

P0-1: buildBestPerPagePoolCte pooled on `slug` alone. In a federated brain, two
pages with the same slug in different sources collapsed before ranking/pagination
(the neighbor-source page dropped). Now DISTINCT ON (COALESCE(source_id,'default'),
slug) — composite key matching dedup.ts's pageKey. Also fixes the PRE-EXISTING
keyword-path bug (best_per_page was slug-only before this wave); real-PG parity 23/23.

P0-2: the alias hop dropped source_id. resolveAliases returned bare slugs and
applyAliasHop hydrated via getPage(slug, undefined), so a federated caller could
get the default-source page injected or the right allowed-source page suppressed.
resolveAliases now returns {slug, source_id} pairs; applyAliasHop matches by
(source_id, slug) and fetches each canonical in its OWN source.

Regression tests: alias hop boosts only the aliased source (not same-slug in
another source); resolveAliases keeps cross-source same-slug distinct.

Deferred as documented tradeoffs (TODO): evidence high_vector_match label uses
blended base_score not pure cosine; deep-pagination candidate budget is
chunk-bounded; telemetry writes swallow errors pre-v109 on rolling deploys.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* chore: bump version and changelog (v0.41.30.0)

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* docs: v0.41.30.0 retrieval cathedral — CLAUDE.md key files + llms regen

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* chore: renumber release v0.41.30.0 → v0.41.34.0 (queue moved)

Version trio + CHANGELOG header + CLAUDE.md key-file annotations + TODOS
heading + regenerated llms bundles, all moved to 0.41.34.0.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* fix(ci): restore glossary roster + harden facts-anti-loop hook budget

Two CI failures surfaced after the master merges that brought the branch to
111 migrations:

1. shard 1 — `ALL_METRICS roster > matches the renderer output (no orphans)`:
   the merge took master's `renderMetricGlossaryMarkdown` whose `groups`
   array lacked this branch's 4 retrieval-quality keys (hit@1, hit@3,
   avg_rank1_score, create_safety). `ALL_METRICS` (derived via Object.keys)
   kept them, so the roster test saw 4 orphans. The freshness check
   (check:eval-glossary) passed because renderer-output == committed doc —
   it can't catch a renderer that drops a metric; the roster test can.
   Restored the "Retrieval-Quality / Evidence Metrics (NamedThingBench)"
   group + regenerated docs/eval/METRIC_GLOSSARY.md.

2. shard 2 — facts-anti-loop's two engine-dependent put_page tests failed
   while the two engine-free extractFactsFromTurn tests passed (the
   signature of a partially-failed beforeAll). This file has a documented
   PGLite-cold-start-under-deep-shard-load timeout history; the 30s budget
   was tuned for 95 migrations and the chain is now 111. Bumped to 60s.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* fix(ci): isolate facts-anti-loop in its own process (serial)

Follow-up to the prior hook-timeout bump, which was the wrong theory: the
[58ms]/[71ms] body times in the re-run prove beforeAll did NOT time out —
the engine connects and the two put_page tests run and fail for real, while
the two engine-free extractFactsFromTurn tests in the same file pass.

put_page (via dispatchToolCall) touches process-global singletons (the
facts queue + the AI gateway used by importFromContent's embed step). Some
sibling file in the 78-file shard-2 process leaves residual global state
that makes put_page's pre-backstop path fail on the CI runner. The failure
is NOT reproducible alone, in a Linux oven/bun:1 container, or in a full
local shard-2 run (1172 pass) — only on the GitHub runner, deterministically.

Per CLAUDE.md's test-isolation rules, a test coupled to shared process
state belongs in its own process. Renamed to *.serial.test.ts so it runs
in the dedicated serial-tests job (scripts/run-serial-tests.sh spawns a
fresh `bun test` per serial file), where it passes deterministically;
test-shard.sh excludes serial files from the matrix. Updated the comment
to reflect the real cause and refreshed the test-weights.json key.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* fix(ci): close cross-file gateway-config pollution in test shards

The prior serial-move theory was incomplete. The real, single root cause
behind all three shard failures (2, 5, 10) is cross-file AI-gateway config
pollution within a shard's bun process:

- A test calls configureGateway() and doesn't restore the gateway on exit.
  The legacy-embedding preload pins OpenAI/1536 ONCE at process start and
  re-pins per-test ONLY when the gateway slot is empty — so a leaker that
  reconfigured the gateway to the v0.37 default (zeroentropyai:zembed-1 /
  1280-d) and never reset poisons every later file in the shard.
- Victim A (shard 5, test/search/searchvector-maxpool.test.ts): runs
  initSchema in beforeAll under the leaked gateway → content_chunks.embedding
  becomes vector(1280) → inserting its hardcoded 1536-d basis vectors throws
  pgvector CheckExpectedDim.
- Victims B/C (shard 10 facts-backstop-gating, shard 2 facts-anti-loop):
  put_page's importFromContent embeds by design (embed failure PROPAGATES,
  Codex C2). Under a leaked fake-key gateway the embed step 401s and put_page
  returns isError → the backstop assertions fail.

My branch's shard re-partition (added test files + weight changes) merely
co-located leakers with victims; the hazard was latent.

Fixes (root cause + self-sufficient victims):
- test/search/rerank.test.ts (the shard-5 leaker): add afterAll(resetGateway).
  Its stub omits embedding_model, so it fell back to the ZE/1280 default;
  now it restores the empty slot so the preload re-pins legacy for the next
  file.
- test/search/searchvector-maxpool.test.ts: pin configureGateway(openai/1536)
  in beforeAll BEFORE initSchema (initSchema runs before any preload
  beforeEach, so it can't rely on the inherited slot).
- test/facts-backstop-gating.test.ts + test/facts-anti-loop.test.ts: reset
  the gateway in beforeEach so put_page's embed is a graceful no-op; reverted
  anti-loop from the serial quarantine back into the matrix (the serial move
  was the wrong fix for a gateway-state problem).

Validated deterministically: a non-resetting leaker that poisons the gateway
to ZE, run first in one bun process, no longer breaks any of the three
victims (14/14 pass). verify 29/29, typecheck clean, isolation lint clean.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
 2026-05-30 12:07:38 -07:00
garrytan-agentsgarrytan-agentsGarry Tan
 		1a6b543cc5 v0.33.2.0 feat(search-lite): token budget + semantic query cache + intent weighting (#897) 
* feat(search-lite): token budget + semantic query cache + intent weighting

Adds three additive features to the hybrid search pipeline. All
backward-compatible: existing callers see identical behavior unless they
opt in to the new options.

## 1. Token Budget Enforcement (src/core/search/token-budget.ts)

Cap the cumulative token cost of returned results so search payloads
fit downstream context windows. Greedy top-down walk; preserves caller
ordering; no re-rank. char/4 heuristic for token counting (no
tokenizer dependency \u2014 keeps the bun --compile bundle small).

  SearchOpts.tokenBudget   \u2014 numeric cap. Default undefined = no-op.
  HybridSearchMeta.token_budget = { budget, used, kept, dropped }

  HTTP query op: pass `token_budget` param.

## 2. Semantic Query Cache (src/core/search/query-cache.ts + migration v52)

Cache search results keyed by query embedding similarity. HNSW lookup:
`embedding <=> $1 < 0.08` (cosine similarity >= 0.92). Per-source
isolation so multi-source brains don\u2019t bleed. Per-row TTL (default 3600s).
Best-effort writes; all errors swallowed so the cache never breaks the
search hot path.

  Migration v52 creates query_cache table with HALFVEC where pgvector >= 0.7;
  falls back to VECTOR with the resolved config.embedding_dimensions dim.

  New `gbrain cache` CLI: stats / clear --yes / prune.
  Config keys: search.cache.enabled / similarity_threshold / ttl_seconds.

  HybridSearchMeta.cache = { status, similarity?, age_seconds? }

  Routed through new `hybridSearchCached(engine, query, opts)` wrapper;
  the operations.ts query op now uses this wrapper so MCP/CLI calls
  benefit automatically. Skipped for two-pass walks + non-default
  embedding columns where cache semantics don\u2019t hold.

## 3. Zero-LLM Intent Weighting (src/core/search/intent-weights.ts)

Builds on the existing query-intent classifier (4 intents: entity /
temporal / event / general). New weight-adjustment layer applies subtle
per-intent nudges:

  entity   \u2192 boost keyword RRF + exact slug/title match
  temporal \u2192 default recency=on when caller left it unset
  event    \u2192 boost keyword RRF (rare named entities) + soft recency
  general  \u2192 no-op (1.0 multipliers everywhere)

All adjustments are SUBTLE (max 1.25x). Caller-explicit options ALWAYS
win \u2014 intent weighting never silently overrides recency / salience.

Default ON; opt out via `opts.intentWeighting = false`. LLM query
expansion (expansion.ts) is still available and opt-in via
`opts.expansion = true` \u2014 it just isn\u2019t the default anymore.

  HybridSearchMeta.intent now surfaces classifier output for debugging.

## Tests

  test/token-budget.test.ts            (10 tests, pure module)
  test/intent-weights.test.ts          (13 tests, pure module)
  test/query-cache.test.ts             (12 tests, PGLite)
  test/hybrid-search-lite.serial.test.ts (9 tests, PGLite e2e)

Plus 105 pre-existing search tests still pass. `bun run verify` clean.

Co-authored-by: Wintermute <agents@garrytan.com>

* feat(search-mode): MODE_BUNDLES + resolveSearchMode wired into bare hybridSearch

Three named modes (conservative / balanced / tokenmax) that bundle the
search-lite knobs from PR #897 into a single config key. Mode resolution
lives in bare hybridSearch (NOT just the cached wrapper) so eval-replay
and eval-longmemeval — which call bare hybridSearch — test the same
mode-affected behavior as production. See [CDX-5+6] in the plan.

The mode bundle supplies DEFAULTS for intentWeighting, tokenBudget,
expansion, and searchLimit when the caller leaves those undefined.
Per-call SearchOpts and per-key config overrides still win (matches the
v0.31.12 model-tier resolution chain at model-config.ts:resolveModel).

knobsHash() exposes a stable SHA-256 of the resolved knob set; the cache
contamination hotfix (next commit) consumes it to prevent a tokenmax
write from being served to a conservative read.

Three new fields on HybridSearchMeta:
  - mode (resolved mode name)
  - existing token_budget meta now fires from bare hybridSearch too

Bare hybridSearch now applies tokenBudget at all three return paths
(no-embedding-provider, keyword-only-fallback, main). Previously only
hybridSearchCached enforced budget; eval commands missed it.

Tests: 37 unit cases pin the 3x7 bundle table cell-by-cell, the
resolution chain semantics, knobs hash determinism + cross-mode
separation, and the config-table parser. All 72 search-lite tests pass.

Bisect-friendly: this commit ONLY adds mode resolution. The cache-key
contamination hotfix [CDX-4] is a separate atomic commit (next).

* fix(query-cache): cross-mode contamination hotfix [CDX-4]

PR #897's query_cache keyed rows on sha256(source_id::query_text) only.
A tokenmax search (expansion=on, limit=50) populated a row that a
subsequent conservative call (no expansion, limit=10) read back, serving
the wrong-shape results. This is a real bug in PR #897 today, regardless
of the v0.32.3 mode picker work — Codex caught it in plan review.

Fix:
- Migration v56 adds query_cache.knobs_hash TEXT column + composite
  (source_id, knobs_hash, created_at) index. Existing rows have NULL
  knobs_hash and are excluded from lookups (silently re-populated with
  the right hash on first hit — no orphan data, no destructive migration).
- cacheRowId(query, source, knobsHash) — knobsHash now part of the PK so
  a tokenmax write and a conservative write for the same (query, source)
  land in distinct rows.
- SemanticQueryCache.lookup({knobsHash}) filters WHERE knobs_hash = $.
- SemanticQueryCache.store({knobsHash}) writes the resolved hash.
- hybridSearchCached threads knobsHash from resolveSearchMode through
  every cache call. Cache config (enabled/threshold/TTL) now reads from
  the resolved mode bundle, not directly from the config table.

Tests (test/query-cache-knobs-hash.test.ts, 11 cases):
- cacheRowId bifurcates by knobsHash
- Tokenmax write does NOT contaminate conservative lookup
- Three modes coexist as distinct rows for same query
- Legacy NULL-knobs_hash rows are excluded from lookup
- Same-mode write updates in place (no duplicate rows)

All 58 cache + mode tests pass. Migration v56 applies cleanly on a fresh
PGLite brain.

Bisect-friendly: this commit is the cache-key hotfix alone. Mode
resolution wiring lives in the previous commit.

* feat(search-telemetry): in-process rollup writer + search_telemetry table

Migration v57 creates search_telemetry (date, mode, intent, count,
sum_results, sum_tokens, sum_budget_dropped, cache_hit, cache_miss,
first_seen, last_seen). PK (date, mode, intent) caps growth at ~4380
rows/year. Sums + counts only — averages derive at read time so
concurrent ON CONFLICT writes from multiple gbrain processes accumulate
correctly [CDX-17].

In-memory bucket flushed periodically (60s OR 100 calls) + on process
beforeExit/SIGINT/SIGTERM with a 2-second cap. The search hot path NEVER
waits on this write [D2, CDX-19].

Date-bucketed cache_hit / cache_miss columns make hit rate over --days N
derivable [CDX-18]. query_cache.hit_count is a lifetime counter and
can't be sliced by window.

Wired into bare hybridSearch via emitMeta: every search call sync-bumps
a bucket. flush() drains atomically by swapping the map before SQL writes
so a record() during flush lands in the new map.

readSearchStats(engine, {days}) returns the StatsWindow shape that
gbrain search stats consumes (next commit).

Tests: 16 unit cases pin record/flush/read semantics including
ON-CONFLICT-adds-raw-values, concurrent-flush coalescing, cache hit-rate
math, missing-table graceful degradation, and window clamping.

53 migrations apply on a fresh PGLite brain.

* feat(config): add unset + listConfigKeys + readLineSafe helper [CDX-7+8+9]

CDX-8: gbrain config has no unset path today. Required before
`gbrain search modes --reset` can clear search.* overrides.

  - BrainEngine.unsetConfig(key) → returns rows deleted (0|1)
  - BrainEngine.listConfigKeys(prefix) → exact-literal prefix match
    with LIKE-escape on user-supplied % / _ / \ characters
  - PGLiteEngine + PostgresEngine implementations
  - `gbrain config unset <key>` and `gbrain config unset --pattern <prefix>`
    sub-subcommands

CDX-9: readLine has no EOF detection or timeout. Mode-picker plan calls
out "TTY closes mid-prompt → defaults to balanced" but the raw helper
hangs forever. New readLineSafe(prompt, defaultValue, timeoutMs=60s):

  - Returns defaultValue on stdin 'end' event
  - Returns defaultValue on timeout
  - Returns defaultValue on empty Enter
  - Non-TTY stdin returns defaultValue immediately (e2e safe)
  - Returns trimmed user input otherwise

Exported so install picker (next task) can use it.

Tests: 9 cases pin unset semantics + prefix matcher edge cases
(glob-wildcard escape, sort order, idempotent loop, search.* sweep).
All 53 migrations apply on a fresh PGLite brain.

* feat(init): install-time mode picker + upgrade banner

Install picker (src/commands/init-mode-picker.ts):
  - Runs as a phase inside `gbrain init` AFTER engine.initSchema() so DB
    config writes work [CDX-7].
  - Idempotent: skipped on re-init if search.mode is already set.
  - Smart auto-suggestion via recommendModeFor() reads
    models.tier.subagent / models.default / OPENAI_API_KEY:
      * Opus default/subagent → tokenmax (quality ceiling)
      * Haiku subagent → conservative (4K budget keeps cost down)
      * No OpenAI key → conservative (no LLM expansion possible)
      * Sonnet / unknown → balanced (safe default)
  - TTY shows menu via readLineSafe (60s timeout, defaults on EOF/empty).
  - Non-TTY auto-selects + emits operator hint:
      [gbrain] search mode: X (auto-selected — reason)
      [gbrain] To change: gbrain config set search.mode <...>
  - --json mode emits structured `{phase: 'search_mode_picker', ...}` event.
  - Wired into both initPGLite and initPostgres flows.

Upgrade banner (src/commands/upgrade.ts):
  - One-shot stderr banner in runPostUpgrade.
  - State persisted via config key `search.mode_upgrade_notice_shown=true`
    — fires at most once per install.
  - Copy corrected per [CDX-1+2+3]: production query op STILL defaults
    expand=true and limit=20. The banner reframes from "behavior is
    regressing" to "named modes available + here's how to preserve
    exact current shape."

Tests (test/init-mode-picker.test.ts, 16 cases):
  - recommendModeFor heuristic for all 4 input shapes
  - parseModeInput accepts numeric/named/case-insensitive, rejects garbage
  - runModePicker non-TTY auto-selects + writes config
  - Idempotent + --force re-prompt + JSON output
  - Opus → tokenmax, Haiku → conservative real wiring through engine

* feat(cli): gbrain search modes/stats/tune command

Three sub-subcommands mirroring the gbrain models (v0.31.12) shape:

  gbrain search modes [--json]
    Read-only routing dashboard. Shows the three mode bundles, the active
    mode, and the source of every resolved knob:
      cache_enabled = true   [override: search.cache.enabled]
      tokenBudget   = 4000   [mode: conservative]
    Plus knob descriptions for legibility.

  gbrain search modes --reset [--source <mode>]
    Clears every search.* override (NOT search.mode itself). Preserves
    the upgrade-notice state key. --source <mode> is a dry-run that
    lists what --reset would change without writing — the paved path
    [CDX-8] flagged as missing.

  gbrain search stats [--days N] [--json]
    Observability. Reads the search_telemetry rollup over the window
    (clamps to [1, 365]). Prints cache hit rate, mode mix, intent mix,
    budget drops, avg results/tokens. JSON output includes
    _meta.metric_glossary block per [CDX-25].

  gbrain search tune [--apply] [--json]
    Recommendation engine. 5 rules cover the bug class:
      - Insufficient data → "no_recommendations" status
      - Conservative + high budget-drop rate → suggest balanced
      - High cache hit rate (>85%) → suggest similarity threshold bump
      - Tokenmax + Haiku subagent → suggest balanced (cost mismatch)
      - Cache disabled but stats show usage → suggest re-enabling
    --apply mutates config via setConfig / unsetConfig with a paste-ready
    revert command printed at the end.

Registered in src/cli.ts dispatch table. 17 unit cases pin:
  - Dashboard report shape + per-knob source attribution
  - --reset preserves search.mode + notice key
  - --source dry-run never writes
  - stats reads telemetry rollup; --days clamps
  - tune recommendation rules fire on real telemetry data
  - --apply mutates config
  - --help + unknown subcommand exit codes

* feat(eval): metric glossary module + auto-gen METRIC_GLOSSARY.md + CI guard

Single source of truth at src/core/eval/metric-glossary.ts. Every entry
carries 3 fields:
  - industry_term (canonical IR/NLP literature name, preserved verbatim)
  - eli10 (plain-English a 16-year-old can follow)
  - range (numeric range + interpretation)

Covers 4 metric families:
  - Retrieval: P@k, R@k, MRR, nDCG@k
  - Stability: Jaccard@k, top-1 stability
  - Statistical: p-value (paired bootstrap + Bonferroni), 95% CI
  - Operational: cache hit rate, avg results/tokens, cost per query, p99 latency

Public surface:
  - getMetricGloss(metric) → full entry or null
  - eli10For(metric) → plain-English string or null
  - buildMetricGlossaryMeta(metrics[]) → {metric → eli10} record for
    JSON `_meta.metric_glossary` blocks per [CDX-25]. ONE block per
    response, NOT sibling `_gloss` fields on every metric.
  - renderMetricGlossaryMarkdown() → deterministic Markdown for the doc

Auto-generation:
  scripts/generate-metric-glossary.ts emits docs/eval/METRIC_GLOSSARY.md.
  Deterministic (same input → same bytes) so the CI guard can diff.

CI guard:
  scripts/check-eval-glossary-fresh.sh regenerates into a temp file and
  diffs against the committed doc. Out-of-date doc fails the build.
  Wired into `bun run verify` (and therefore `bun run test:full`).

Tests (test/metric-glossary.test.ts, 18 cases):
  - Every documented metric is present
  - Every entry has all 3 required fields
  - Accessors return null on unknown metrics (no throw)
  - buildMetricGlossaryMeta silently drops unknown metrics
  - renderer output is deterministic across calls
  - Renderer groups metrics into 4 sections

docs/eval/METRIC_GLOSSARY.md: 5491 bytes, 124 lines, fresh.

* feat(doctor): search_mode + eval_drift checks + drift-watch module

src/core/eval/drift-watch.ts — curated retrieval watch-list [CDX-6].
Five patterns covering the surface that actually affects retrieval quality:
  - src/core/search/      (search pipeline)
  - src/core/embedding.ts (embedding shape)
  - src/core/chunkers/    (chunk granularity)
  - src/core/ai/recipes/anthropic.ts + openai.ts (expansion + embed routing)
  - src/core/operations.ts (the query op definition)

Adding to the list is a deliberate act — requires a CHANGELOG line so
coverage grows on purpose, not by accident. Pure functions:
  - matchesWatchPattern(path) — trailing-slash = prefix, bare = equality
  - filesDriftedSince(repoRoot, sha?) — git diff --name-only wrapper
  - watchedFilesDrifted(repoRoot, sha?) — composite

src/commands/doctor.ts — two new checks.

checkSearchMode [CDX-20]: status stays 'ok' (never warns, never docks
health score). Hint in message field. Three branches:
  - unset → "search.mode is unset (using balanced fallback). Run
    `gbrain search modes` to see what is running and pick a mode."
  - mode + no overrides → "Mode: X (no per-key overrides — mode bundle
    is canonical)."
  - mode + overrides → "Mode: X with N per-key override(s) (k1, k2, …).
    To consolidate to the pure mode bundle: gbrain search modes --reset"
Upgrade-notice state key (search.mode_upgrade_notice_shown) is excluded
from the override roster — it's not a knob.

checkEvalDrift [CDX-6]: surfaces uncommitted changes to retrieval-watched
files. Always 'ok'; operator-facing reminder. Names up to 3 drifted files
in the message + paste-ready re-eval command.

Both helpers exported (was: file-private) so tests can pin behavior
without walking the full runDoctor pipeline.

Tests: 12 drift-watch cases + 7 doctor-check cases. Pin watch-list shape,
prefix-vs-equality matcher semantics, missing-repo graceful failure, and
all three search_mode branches.

* feat(eval): --mode flag on longmemeval/replay + run-all + compare

Per-mode --mode flag plumbed into:
  - gbrain eval longmemeval --mode <conservative|balanced|tokenmax>
    Sets search.mode in the benchmark brain's config table; config is
    in PRESERVE_TABLES so resetTables doesn't wipe it between questions.
    Mode surfaces in the per-question NDJSON row.
  - gbrain eval replay --mode <m> + --compare-limit N
    --compare-limit forces a constant K across modes [CDX-13]; without
    it, Jaccard@k against the captured baseline measures K-drift, not
    quality. Mode is set once before the replay loop.
  - NOT cross-modal per [CDX-11]: cross-modal scores OUTPUT against
    TASK; it doesn't retrieve. Adding --mode there is theater.

New: gbrain eval run-all orchestrator (src/commands/eval-run-all.ts):
  - Sweeps every requested mode × suite combination
  - Sequential default per D9; --parallel N opt-in (clamped to mode count)
  - Cost guard with split caps [CDX-15+16]:
      --budget-usd-retrieval N (default $5)
      --budget-usd-answer N (default $20)
    Non-TTY refuses with exit 2 unless --yes AND explicit --budget-usd-*
    flags pass. TTY refuses without --yes (defense against agent loops).
  - estimateRunCost computes per-(suite,mode) breakdown including the
    expansion-Haiku surcharge for tokenmax.
  - Audit trail: appends to <repo>/.gbrain-evals/eval-results.jsonl
    [CDX-23]. Personal brain (~/.gbrain) NEVER touched.
  - v0.32.3 ships orchestrator + argv + guard + persist hook.
    In-process per-suite invocation is a v0.32.4 follow-up (operator
    runs the per-suite CLIs with the documented --mode flag for now;
    each completion calls persistRunRecord to log).

New: gbrain eval compare report (src/commands/eval-compare.ts):
  - Reads eval-results.jsonl, groups by (suite, mode), renders MD or JSON
  - Most-recent (suite, mode, commit) wins when duplicates exist
  - JSON output has schema_version=2 + _meta.metric_glossary block per
    [CDX-25] (ONE block per response, not sibling _gloss fields)
  - _meta.methodology field names the paired-bootstrap + Bonferroni
    discipline per [CDX-14] so haters can reproduce
  - Missing file → friendly hint pointing at `gbrain eval run-all`

Wired into eval dispatch table in src/commands/eval.ts.

Metric glossary fuzzy fallback: `recall@10` → `recall@k` lookup
(the glossary documents the family; report rows carry specific K
values). Routes through getMetricGloss for every call site.

Tests (42 cases total — all green):
  - eval-run-all.test.ts (19): argv parser, cost estimate, guard
    semantics for all 4 (over/under × tty/non-tty) shapes, persist hook
    NDJSON shape.
  - eval-compare.test.ts (5): JSON + MD output shapes, glossary
    integration, missing-file graceful, mode filter, most-recent-wins.
  - metric-glossary.test.ts (18): unchanged but updated assertions to
    cover the fuzzy `@N` → `@k` fallback.

Pre-existing eval-replay / eval-longmemeval / eval-export / eval-prune
tests (42 cases) still pass — --mode + --compare-limit are additive.

* docs: methodology + CLAUDE.md/README/RESOLVER + skills/conventions

docs/eval/SEARCH_MODE_METHODOLOGY.md — haters-immune 8-section template.
Documents what the eval measures + does NOT measure, datasets + sizes
(LongMemEval n=500, Replay n=200, BrainBench n=1240 docs / 350 qrels),
random seed 42, run procedure verbatim, threats to validity (LongMemEval
English+technical skew, char/4 heuristic ~5-10% off, expansion ~97.6%
relative lift on this corpus), per-question raw outputs, pre-registered
expectations (tokenmax wins R@10 by 5-15pp, conservative wins cost by
5-15x, balanced lands within 3pp), re-run cadence anchored to the
src/core/eval/drift-watch.ts watch-list.

Statistical-significance section pins paired bootstrap with 10,000
resamples + Bonferroni correction across 3 modes × 4 metrics [CDX-14].

CLAUDE.md gets two new sections: ## Search Mode (3-mode table + resolution
chain + [CDX-4] cache contamination fix note + CLI commands) and ## Eval
discipline (single-source-of-truth glossary, methodology doc, eval_results
in repo NOT personal brain per [CDX-23]).

README.md Quick Start gets a paragraph naming the install picker, mode
heuristic, and the methodology link.

skills/conventions/search-modes.md NEW — convention file consumed by
brain-ops + query + signal-detector skills via the existing
`> **Convention:**` callout pattern. Routes "what mode" / "tune
retrieval" / "compare modes" queries to the right CLI surface.

skills/RESOLVER.md gets two new trigger rows pointing at
gbrain search * and gbrain eval compare.

* chore: regen llms.txt + llms-full.txt for v0.32.3 search-mode docs

bun run build:llms — picks up the new CLAUDE.md sections (Search Mode +
Eval discipline) and the docs/eval/SEARCH_MODE_METHODOLOGY.md addition.
build-llms.test.ts gate now passes.

* fix(doctor): wire search_mode + eval_drift checks into runDoctor main flow

The v0.32.3 search_mode + eval_drift helpers were inserted into the
DB-checks sub-helper at runDbChecks (line 345-355), but runDoctor itself
maintains its own check list and only calls the helpers' subset. Push
the two checks into the main runDoctor path (after the existing
sync_freshness check at line 2347) so they actually appear in
`gbrain doctor --json` output.

Both checks gated on engine !== null. Progress reporter heartbeat fires
for each. Both still return status 'ok' per [CDX-20] so health score is
preserved.

Verified end-to-end on a real Postgres brain: gbrain doctor --json now
includes 'search_mode' and 'eval_drift' in the checks array.

* fix: claw-test hang — DATABASE_URL leak + telemetry beforeExit deadlock

Two root causes for the hang, both fixed.

1. DATABASE_URL leak in claw-test scripted harness
   The harness inherits the parent process's env via `...process.env`
   for every phase child (init / import / query / extract / doctor).
   When the e2e runner sets DATABASE_URL (for OTHER e2e tests), it
   leaks into claw-test's children. `loadConfig` at src/core/config.ts:143
   then flips inferredEngine to 'postgres' for every subsequent phase,
   breaking the hermetic-PGLite-tempdir contract: phases race against
   each other on a shared test Postgres while pointing at different
   brain states.

   Fix: strip DATABASE_URL + GBRAIN_DATABASE_URL from the child env
   before forwarding. Re-apply GBRAIN_HOME / GBRAIN_FRICTION_RUN_ID
   after the merge so a parent's override can't win. The harness is
   PGLite-only by design.

2. Telemetry beforeExit deadlock
   v0.32.3's recordSearchTelemetry installed a `process.on('beforeExit',
   drainOnExit)` hook that wrapped the flush in `Promise.race([flush(),
   setTimeout(2000)])`. beforeExit fires when the event loop empties,
   but the hook enqueued NEW async work (the race's setTimeout +
   pending flush), so the event loop never re-emptied. Short-lived
   CLI invocations (`gbrain query "the"` finishing in ~100ms) ended
   up waiting on the DB write indefinitely.

   The claw-test harness spawns several short-lived gbrain queries.
   Each one hung after its real work finished. The harness then waited
   forever on its child subprocess's exit code.

   Fix: drop the beforeExit + SIGINT + SIGTERM hooks. Per [CDX-19]'s
   "stats are directional, not exact" contract, losing one unflushed
   bucket on process exit is acceptable. The unref'd setInterval
   handles long-running processes (HTTP MCP, autopilot, jobs work).
   Short-lived CLI invocations exit immediately.

Verified:
  - `gbrain query "the"` on a fresh PGLite brain exits in <1s (was
    hanging forever).
  - `bun test test/e2e/claw-test.test.ts` → 3 pass / 0 fail / 3.86s
    (was hanging at the banner indefinitely).
  - 85/85 e2e files / 574/574 tests pass including claw-test, with
    DATABASE_URL set (the configuration that originally repro'd the
    hang).
  - 6235/6235 unit tests pass.
  - Typecheck clean.

The two bugs interacted: the DATABASE_URL leak meant queries hit the
real Postgres (slow), making the beforeExit deadlock visible. Fixing
either alone would have masked the other. Both fixed in this commit.

* feat(install-picker): cost anchors in mode prompt + upgrade banner + docs

The install picker already asks explicitly (1/2/3 menu, default to the
recommendation on Enter). What was missing: a way to reason about the
cost tradeoff. Without numbers, "tokenmax" looks free and "conservative"
sounds restrictive; with numbers, the operator picks intentionally.

Cost anchors added everywhere the user encounters the mode choice:
  - Install picker MENU_TEXT (gbrain init)
  - Upgrade banner (gbrain upgrade post-upgrade)
  - CLAUDE.md ## Search Mode section
  - README.md Quick Start
  - docs/eval/SEARCH_MODE_METHODOLOGY.md (with the math)

Anchors at Sonnet 4.6 downstream ($3/M input):
  conservative  ~$0.012/query  ~$12/mo @ 1K  ~$1,200/mo @ 100K
  balanced      ~$0.030/query  ~$30/mo @ 1K  ~$3,000/mo @ 100K
  tokenmax      ~$0.060/query  ~$60/mo @ 1K  ~$6,000/mo @ 100K

Plus tokenmax's Haiku expansion overhead: ~$1.50 per 1K queries on top.
Cache hits roughly halve these on a brain with repeat-query traffic.

The math is documented in SEARCH_MODE_METHODOLOGY.md so a reviewer can
audit each variable (T = ~400 tokens/chunk from the recursive chunker's
300-word target; N = `searchLimit` cap; R = downstream model rate from
src/core/anthropic-pricing.ts). Drift away from these numbers requires
updating CLAUDE.md + the picker + the methodology doc in lockstep — a
regression test pins the picker's anchor strings to enforce this.

The framing also names the cost rule honestly: the dominant cost isn't
gbrain (semantic cache is free; Haiku expansion is rounding-error). It's
the downstream agent reading retrieved chunks back into its context.
Operators who don't realize this pick badly.

Tests: 5 new regression cases in init-mode-picker.test.ts pin every
cost string in MENU_TEXT. Total 21/21 picker tests pass; 6240/6240
unit tests pass; verify gate green.

* docs: realistic-scale cost anchor for search modes

The per-query cost framing in the picker (~$0.012/$0.030/$0.060) is
honest but theoretical — it treats each search as an isolated billable
event. Real agent loops amortize a lot of context across turns via
Anthropic prompt caching, so the per-query 5x ratio doesn't translate
1:1 into total agent spend.

Added a "Realistic-scale anchor" section to SEARCH_MODE_METHODOLOGY.md
representing one heavy power-user agent loop running tokenmax:

  - ~860 turns/mo (~29/day, one active agent)
  - ~900K tokens/turn (system + tools + history + reasoning + search)
  - ~$0.85/turn → ~$700/mo total agent spend at tokenmax
  - ~88% Anthropic prompt-cache hit rate

Scaling balanced + conservative DOWN from that anchor:

  - tokenmax  → ~$700/mo, search ~22% of total spend
  - balanced  → ~$620/mo, search ~12% (saves ~$78/mo vs tokenmax)
  - conservative → ~$575/mo, search ~5% (saves ~$124/mo vs tokenmax)

Honest takeaway: at realistic agent-loop scale WITH disciplined prompt
caching, mode choice saves 10-20% of total agent spend, not 5x. The
per-query math kicks back in for setups WITHOUT cache discipline (churn
the prompt prefix every turn → search payload becomes a larger fraction).
Both framings live in the doc.

CLAUDE.md ## Search Mode gets a forward-pointer paragraph naming the
"per-query math vs real-world spend" delta so agents reading the section
find the methodology footnote.

Numbers in the doc are anonymized + scaled away from any specific
deployment. No model names, no specific dollar figures from a real
production setup — just the per-turn / cache-hit-rate / search-count
shape ratios that a thoughtful operator can validate against their own
billing dashboard.

* feat(picker): mode × model cost matrix (25x corner-to-corner spread)

Previous version showed mode costs assuming Sonnet-only downstream.
That muted the spread to 5x and made mode choice look minor. Reality:
the downstream model tier is the BIGGER cost lever — pairing mode with
model is where the 25x spread lives.

New 3×3 matrix in the install picker, CLAUDE.md, methodology doc, README:

                  Haiku 4.5     Sonnet 4.6    Opus 4.7
                  ($1/M input)  ($3/M input)  ($5/M input)
  conservative    $400/mo       $1,200/mo     $2,000/mo
  balanced        $1,000/mo     $3,000/mo     $5,000/mo
  tokenmax        $2,000/mo     $6,000/mo     $10,000/mo

(per-query cost @ 100K queries/mo, full search payload, no cache savings)

The methodology doc gets a new "Mode × Model matrix" section above the
realistic-scale anchor with concrete right-sizing guidance:

  - tokenmax + Haiku: wrong direction. Haiku can't filter 50 chunks → noise
    not signal. Pay Haiku rates, get sub-Haiku quality.
  - conservative + Opus: wasted Opus. 200K context window starved on
    retrieval depth. Pay Opus rates, get conservative-shape retrieval.
  - Natural pairings span ~4x; the matrix corners span 25x. The natural
    diagonal is where most users should land.

Realistic-scale anchor refreshed:
  - tokenmax + Opus: ~$700/mo at 860 turns
  - balanced + Sonnet: ~$430/mo
  - conservative + Haiku: ~$170/mo

Plus a "mismatched pairings" section showing the math for tokenmax+Haiku
and conservative+Opus — both burn budget for no improvement.

Regression test updated: pins the 25x framing + the four anchor cells
(two corners + two diagonal mids) + the three downstream model rates.

22/22 picker tests pass. 6241/6241 unit tests pass. CI guards green.

* docs(picker): rescale cost matrix from 100K → 10K queries/mo (typical single user)

Most users running gbrain are single-user installs at ~10K queries/month,
not the 100K fleet-scale used in the original matrix. The picker numbers
($400 to $10,000/mo) looked alien to the actual audience. Rescaled to
10K with an explicit linear-scaling callout.

New matrix in picker, CLAUDE.md, README, methodology doc:

                  Haiku 4.5     Sonnet 4.6    Opus 4.7
                  ($1/M)        ($3/M)        ($5/M)
  conservative    $40/mo        $120/mo       $200/mo
  balanced        $100/mo       $300/mo       $500/mo
  tokenmax        $200/mo       $600/mo       $1,000/mo

Still 25x corner-to-corner. Still 4x natural-diagonal spread. But now in
numbers a single user picks up and reasons about: "balanced + Sonnet at
$300/mo, that's fine" or "tokenmax + Opus at $1,000/mo, that's a
deliberate choice for max-quality high-stakes work."

Every surface updated:
  - Install picker MENU_TEXT (with "scales linearly — multiply by 10
    for 100K/mo" footnote so heavier users still see their number)
  - CLAUDE.md ## Search Mode table + scaling prose
  - README Quick Start
  - methodology doc Mode × Model matrix section
  - upgrade banner (post-upgrade notice)

Regression test updated: pins the 3 new anchor cells ($40, $300, $1,000)
+ the 10K/mo volume frame + the linear-scaling callout. 23/23 picker
tests pass, 6241/6241 unit tests pass, verify gate green.

Methodology doc's existing 1K/10K/100K Monthly cost breakdown tables
left intact (they already show the linear scaling explicitly).

* feat(picker): agent-facing install protocol + tokenmax default + [AGENT] directive

DX gap: an agent installing gbrain (OpenClaw, Hermes, Codex, Cursor) ran
gbrain init non-TTY, saw 2 stderr lines flash by, and silently auto-applied
a default search mode. The operator never saw the cost matrix or the choice.
At 25x corner-to-corner cost spread, that's surprise-spend territory.

Five surfaces fixed:

1. **Auto-suggest default flipped balanced → tokenmax.** The Sonnet/unknown
   fallback now recommends tokenmax (preserves v0.31.x retrieval shape:
   expand=on, generous result set). Haiku subagent → conservative still
   wins (cost-sensitive signal). No-OpenAI-key → conservative still wins
   (vector search not possible). Heuristic reordered: Haiku check now
   fires BEFORE the Opus check, because a Haiku subagent loop signalling
   cost sensitivity should win over a default-model heuristic.

2. **gbrain init non-TTY output rebuilt.** Previously: 2 stderr lines.
   Now: the full 3×3 cost matrix + an explicit [AGENT] directive block
   telling the agent to relay the matrix to its operator before
   continuing. Includes a pointer to INSTALL_FOR_AGENTS.md Step 3.5 for
   the full protocol.

3. **gbrain upgrade banner same treatment.** Existing v0.32.3 banner now
   includes [AGENT] directive at the top so upgrading agents relay the
   matrix to their operator instead of silently accepting v0.31.x →
   v0.32.x default-applied behavior.

4. **INSTALL_FOR_AGENTS.md Step 3.5 NEW** with the matrix verbatim, the
   exact paraphrasable ask-the-user wording, and the gbrain config set
   commands to run after the operator picks. Plus a paragraph in the
   Upgrade section pointing back at Step 3.5.

5. **AGENTS.md install checklist** gets a new Step 4 ("STOP — ask the
   user about search mode") between init and the rest of the flow. The
   agent's job description now explicitly says: silent acceptance is
   the wrong default.

Tests (24/24 pass):
  - Updated recommendModeFor heuristic order (Haiku floor > Opus default)
  - New regression test: non-TTY output contains the matrix corners +
    [AGENT] directive + INSTALL_FOR_AGENTS.md pointer
  - withEnv() helper used for OPENAI_API_KEY mutation (test-isolation lint)
  - Default-recommendation tests updated: Sonnet / unknown → tokenmax

Privacy + test-isolation gates clean. 6256/6256 unit tests pass.

---------

Co-authored-by: garrytan-agents <agents@garrytan.com>
Co-authored-by: Garry Tan <garrytan@gmail.com>
 2026-05-13 13:14:58 -04:00