mirror of
https://github.com/garrytan/gbrain.git
synced 2026-07-28 14:59:47 +00:00
fix/backlog-c45
2
Commits
| Author | SHA1 | Message | Date | |
|---|---|---|---|---|
|
a81f7e05e8 |
v0.42.43.0 feat(context): push-based context (#2095) + teardown-exit hardening (#2084) (#2175)
* fix(cli): exit deliberately after bounded teardown instead of riding the 10s backstop (#2084)
Root cause: bounded teardown (endPoolBounded, #2015) RESOLVES, but lingering
sockets — embedding-provider fetch keep-alive, PgBouncer txn-mode sockets the
bound raced past — keep Bun's event loop alive, so every `gbrain query` paid
a flat 10s tax exiting via the hard-deadline force-exit banner.
Three changes, one contract:
- flushStdoutThenExit (cli-force-exit.ts): when main() resolves and the
command is not a daemon, exit deliberately — after stdout AND stderr drain
(writableLength===0, 'drain'-event + poll loop, 2s unref'd guard for a
blocked pipe). Incident #1959 (force-exit truncating piped stdout) is the
regression class; pinned by a 256KB real-pipe subprocess test.
- drainThenDisconnect (cli.ts): ONE owner-disconnect helper at all 8 sites
(op-dispatch, CLI_ONLY fall-through, search dashboard, doctor remediation
x3, ze-switch, dream, read-only timeout path). Drains the background-work
registry, then disconnect (best-effort), bounded by the 10s unref'd
hard-deadline — which is now armed around the TEARDOWN window only, not
before the op handler (the old placement would have force-killed any op
slower than 10s). Closes the filed TODOS P3 drain-hoist: six sites
previously skipped the drain entirely and had no hang timer at all.
- Inner process.exit sweep: mid-handler exits in engine-owning/output-bearing
paths (status, friction, claw-test, smoke-test, eval cross-modal /
takes-quality replay / conversation-parser / whoknows-thin, status-thin)
become process.exitCode + return so they flow through the drains and the
flush-exit. Pre-engine usage/parse/refusal exits stay as-is.
BrainRegistry.disconnectAll deliberately unchanged: zero production callers
in src/, per-engine disconnects already bounded, and the kernel reclaims
sockets on exit (src/core/timeout.ts doctrine).
DAEMON_COMMANDS gains 'watch' ahead of the #2095 push transport.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* test(e2e): PgBouncer transaction-mode pooler in CI + teardown e2e (#2084)
Three consecutive waves (#1972 → #2015 → #2084) fixed pooler-teardown bugs
verified only against one production deployment — CI had no transaction-mode
pooler and could never see the class. Now it can:
- docker-compose.ci.yml: `pgbouncer` service (transaction pooling) fronting
postgres-1, mirroring the production split-pool topology (direct :5432 +
pooled :6543). AUTH_TYPE=plain (pg16 SCRAM verifiers need the plaintext
password in the userlist) + IGNORE_STARTUP_PARAMETERS for the
statement_timeout/idle_in_transaction_session_timeout startup params
gbrain's client sets (the Supabase pooler whitelists the same).
- test/e2e/pgbouncer-teardown.test.ts: schema + fixture via the DIRECT url
into a dedicated `gbrain_pgbouncer` database (never races shard TRUNCATEs),
then spawns the real CLI against the POOLED url and asserts: exit 0,
stdout intact (the #1959 truncation class), and NO
"did not return within 10000ms — force-exiting" banner (pre-#2084 it
printed on 100% of query-shaped ops on this topology). Class bound, not
exact timing. Skips gracefully without GBRAIN_PGBOUNCER_URL.
- scripts/ci-local.sh: threads GBRAIN_PGBOUNCER_URL +
GBRAIN_PGBOUNCER_DIRECT_URL into all three e2e phases.
Verified live: both tests green against pgbouncer 1.25.2 in transaction mode.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* feat(schema): context_volunteer_events table (v116) — push-context feedback log (#2095)
One row per page the brain volunteers (op / reflex / watch channels).
"Used" is DERIVED, never written: pages.last_retrieved_at > volunteered_at
(the existing bumpLastRetrievedAt write-back is the open/cite signal), so
there is no second tracking path. session_id/turn are nullable
caller-supplied attribution; rationale is a deterministic template string,
never raw conversation text.
- Migration v116 (idempotent) + mirrors in src/schema.sql +
src/core/pglite-schema.ts + regenerated schema-embedded.ts (regen also
folds in pre-existing comment-only drift from the v114 links edits).
- src/core/context/volunteer-events.ts: insertVolunteerEvents (ONE
multi-row parameterized INSERT — never per-row awaited round-trips) +
purgeStaleVolunteerEvents (90-day GC, returns 0 on pre-v116 brains).
- Dream cycle purge phase prunes stale events alongside op_checkpoints /
brainstorm checkpoints / batch-retry audit files.
- RLS on Postgres comes from the v35 auto_rls_on_create_table event
trigger (the same mechanism that covered v110 page_aliases and v115
op_checkpoint_paths); the volunteer Postgres e2e pins it.
- No ::jsonb anywhere; no bootstrap probe needed (nothing references the
table pre-creation; writers guard with try/catch).
Tests: v116 shape + columns + indexes + live insert/purge round-trip on
PGLite (test/migrate.test.ts, 161 pass); schema-bootstrap-coverage green.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* feat(context): multi-turn window extraction + confidence-scored volunteer core (#2095)
- entity-salience.ts: extractCandidatesFromWindow(turns) — runs the existing
per-turn extractor across the last N turns (oldest→newest), merges by the
normalizeAlias form with occurrence/newest-turn/user-mention metadata, and
orders by salience (recency > frequency > user-role) so the MAX_CANDIDATES
cap drops stale assistant chatter, not the entity the user just named.
Closes the filed assistant-introduced-entities recall TODO; true pronoun
coreference (never-named antecedents) stays out of scope.
- retrieval-reflex.ts: ReflexPointer gains source_id + arm + confidence +
matchedNorm. ARM_CONFIDENCE (alias 0.9 / title 0.8 / slug-suffix 0.6)
lives next to the arm definitions so identity and score can't drift.
Arm-2 provenance is classified in JS (codex D8 — the combined OR can't
report which predicate matched). Federated sourceIds[] scope (alias arm
loops per source; arm 2 uses source_id = ANY — no engine-interface
change). Suppression gains 'slug-only' mode (codex D7, REQUIRED for
windowing): the legacy title-whole-word rule would suppress every entity
merely MENTIONED in a prior window turn, breaking the feature by
construction — slugs only enter context when a pointer/page was actually
surfaced. Default stays 'slug-and-title' for the window=1 legacy path.
- volunteer.ts (new): parseWindow (lenient user:/assistant: prefixes, CRLF,
unprefixed → one user turn), volunteerContext (zero-LLM: extract →
resolve → +0.05 multi-turn/newest-turn boost → min_confidence 0.7 gate →
cap 3/5; deterministic rationale strings, never raw conversation text),
and volunteerUsageStats (per-arm/channel precision from the
last_retrieved_at join, labeled approximate — 5-min throttle false
negatives, unrelated-read false positives; codex D9).
Tests: 35 green across volunteer-context (window parsing, pronoun follow-up
via assistant-introduced entity, confidence gating, slug-only suppression,
takes-fence privacy, multi-source scope, caps, stats join math) +
retrieval-reflex back-compat + resolve-ipc.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* feat(ops): volunteer_context op — CLI (stdin) + MCP, drained event sink (#2095)
New read-scope op on the contract surface (CLI `gbrain volunteer-context`
with stdin → window, MCP tool for free): takes a rolling conversation
window, returns confidence-gated page pointers with rationales + synopses.
`window` is optional-unless-stats (validated in the handler, codex D9);
`stats: true` returns the volunteered-vs-used precision summary, labeled
APPROXIMATE (the 5-min last-retrieved throttle and unrelated reads both
bias the join). Source scope threads through sourceScopeOpts — federated
grants narrow the volunteer to the granted sources.
Event logging is fire-and-forget through a new `volunteer-events`
background-work sink (volunteer-events.ts, mirrors last-retrieved: tracked
dangling promise set + bounded drain + snapshot-drop on timeout so a
long-lived process never accumulates ghosts). ONE batched INSERT per call,
drained on every exit path by the commit-1 drain hoist; failure never
fails the op (pinned by an injected failing-engine test).
cli formatResult renders both shapes (pointer lines with confidence/arm/
rationale; the stats summary with per-arm precision).
Tests: op contract surface, window-required validation, sink round-trip
with session_id/turn attribution, failing-engine fail-open, federated
grant scoping, stats mode (26 green on PGLite) + a real-Postgres e2e
proving the op + sink + stats join AND that context_volunteer_events has
RLS enabled (keeps the auto-RLS event-trigger mechanism honest for v116).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* feat(context): reflex consumes the rolling window + ambient-channel logging (#2095)
The default-on retrieval reflex now extracts entities from the last N turns
(retrieval_reflex_window_turns, default 4; env
GBRAIN_RETRIEVAL_REFLEX_WINDOW_TURNS; window=1 reproduces the legacy
current-turn-only behavior exactly). assemble() passes the recent
user/assistant turns (hard cap 12); the reflex slices to the configured
window. Assistant-introduced entities and "what did she invest in?"
follow-ups whose antecedent was NAMED in the window now surface pointers —
the issue's "zero agent-initiated queries" success criterion on the
ambient path.
Under windowing, suppression switches to slug-only (codex D7): the legacy
title-whole-word rule would suppress every entity merely MENTIONED in a
prior window turn, breaking the feature by construction. Slugs only enter
prior context when a pointer/page was actually surfaced, so
already-surfaced pages still suppress. The suppression mode flows through
all three resolver rungs (host opts, serve IPC request, direct Postgres).
Ambient-channel feedback (codex D11): the server-side resolver paths
(serve IPC + direct Postgres) log volunteered pointers with
channel: 'reflex' through the drained volunteer-events sink, so
`gbrain volunteer-context --stats` measures the default-on path where most
volunteering happens. Host-injected resolvers (no gbrain engine) can't
log — documented gap. Precision gates, 1.5s ceiling, fail-open, and the
pointer cap are unchanged.
Tests: prev-assistant-turn entity fires; window=1 legacy parity; slug-only
vs already-surfaced suppression; throwing resolver stays fail-open
(16 green).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* feat(cli): gbrain watch — push transport over stdin (#2095)
The issue's headline: the brain volunteers pages as the conversation flows,
instead of waiting to be asked. `some-transcript-feed | gbrain watch` reads
turns line-by-line ('user:'/'assistant:' prefixes set the role; unprefixed
lines are user turns), keeps a rolling window (--window-turns, default 4),
and streams confidence-gated pointers with rationales to stdout (--json for
JSONL). Session dedupe rides the core's slug-only suppression — a slug is
volunteered at most once per session. Events log on channel 'watch' with
session_id + turn through the drained sink.
Lifecycle: watch BLOCKS in the stdin iteration (like `jobs work`) — an
interactive TTY stays alive until Ctrl-C/Ctrl-D, piped input ends at EOF —
so it is deliberately NOT in DAEMON_COMMANDS (reverts the commit-1
placeholder): when main() resolves the work is over, the CLI_ONLY finally
drains volunteer events via drainThenDisconnect, and the entrypoint
flush-exit ends the process. Keeping it in the daemon set would have made
the piped EOF path hang on lingering sockets — the exact #2084 class.
SIGINT closes the stream and flows through the same drain path instead of
killing mid-write. Per-turn resolution failures are fail-open (the stream
never dies on a transient DB error).
Full wiring (eng-review D12): CLI_ONLY + CLI_ONLY_SELF_HELP (WATCH_HELP) +
THIN_CLIENT_REFUSED_COMMANDS (thin clients use the volunteer_context MCP
op) + main --help entry.
Tests: 18 green — help, per-turn volunteering + clean EOF return, rolling
window via assistant-introduced entity, session dedupe, --json shape with
turn attribution, channel-watch event rows, --min-confidence gate, CRLF/
blank tolerance, daemon-gate semantics. Live smoke: piped `gbrain watch`
on a fresh PGLite brain exits 0 at EOF with no force-exit banner.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* docs: KEY_FILES + push-context guide + TODOS for the #2084/#2095 wave
- docs/architecture/KEY_FILES.md (current-state): context entries gain the
window extractor, arm provenance/confidence, suppression modes, volunteer
+ volunteer-events modules; background-work entry now lists FIVE sinks and
the drainThenDisconnect owner-disconnect contract; new entries for
src/core/cli-force-exit.ts (the exit contract) and src/commands/watch.ts.
- docs/guides/push-context.md (new): the three channels (reflex/op/watch),
the confidence model, CLI usage, config keys, and the approximate-stats
caveat. Linked from CLAUDE.md's reference map.
- CLAUDE.md: ops line mentions volunteer_context + the guide link;
bun run build:llms regenerated in the same commit (freshness test green).
- TODOS.md: #2095 deferrals filed (SSE push channel, policy skill + doctor
check, structured messages[] param); the #1981 entity-detection TODO
narrowed (window extraction covered assistant-introduced entities +
named-antecedent follow-ups); the drain-hoist P3 marked DONE by the
#2084 wave.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* test(e2e): truncate context_volunteer_events in setupDB (#2095)
The new feedback-log table wasn't in ALL_TABLES, so volunteered-event rows
persisted across e2e runs on a reused database and poisoned count/stats
assertions in volunteer-context-postgres on the second run. No FK to pages
(slug join), so position before pages is for hygiene only.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(cli): own the exit verdict — never trust ambient process.exitCode (#2084)
Caught by the full unit suite: `gbrain apply-migrations` on PGLite started
exiting 99. Root cause: PGLite's Emscripten runtime writes the WASM
backend's proc_exit status into process.exitCode (initdb at create-time,
the postmaster at close-time — `exitCode=status` in pglite's dist), and
the writes land ASYNCHRONOUSLY, outside any snapshot/restore window around
create/close (a guarded attempt verified this). The pre-#2084 success path
never read process.exitCode, so the pollution was invisible; the new
deliberate flush-exit propagated it faithfully.
Fix: gbrain records its own verdict. setCliExitCode(n)/getCliExitCode() in
cli-force-exit.ts — every gbrain-owned exit-code assignment routes through
the setter (still mirrored to process.exitCode for outside readers), and
both exit paths (entrypoint flushStdoutThenExit + the drainThenDisconnect
hard-deadline backstop) read the getter. Swept all assignment sites:
cli.ts (op error, friction, claw-test, smoke-test, eval runners, status,
import errors) + reindex/transcripts/brainstorm/frontmatter/autopilot.
Also updates the v0.42.20 structural pins to the drainThenDisconnect shape
(ordering invariant asserted INSIDE the helper + >=8 helper call sites,
superseding the two-inline-pairs assertion).
Verified: apply-migrations spawn test green; `init --migrate-only` exits 0;
an errored op still exits 1.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* test: re-pin the teardown-arming invariant at its post-#2084 home
Master's v0.42.41.0 triage wave and the #2084 wave fixed the same
pre-armed-timer bug independently; the merge keeps #2084's shape (arming
inside the shared drainThenDisconnect helper, covering all 8 exit paths).
The structural pin now asserts the same invariant — no pre-try arming;
gated, unref'd, before-drain, cleared — at the helper.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* test: coverage for ambient reflex-channel logging + watch window/cap flags
Ship coverage audit (85%, gate PASS) named five gaps; the two substantive
cheap ones close here: the codex-D11 logChannel='reflex' path now has a
behavioral pin (events land on channel 'reflex' through the drained sink;
no logChannel → no events), and gbrain watch's --window-turns / --max-pages
flags are exercised (turn-1 attribution under window=1; cap to one page).
Remaining flagged-not-blocking: the wallclock-timeout branch (untestable
without >10s real-clock flake — same rationale as the arming pin),
formatResult's volunteer case (module-private), and the cycle purge wiring.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* test: close the remaining plan-audit gaps — formatResult rendering + watch SIGINT
formatResult exported for tests (same import-safety contract as cliAliases);
test/cli-format-volunteer.test.ts pins the pointer lines, empty-gate message,
and approximate stats summary. test/watch-command.test.ts gains a real
subprocess SIGINT test: piped stdin that never reaches EOF, SIGINT mid-stream,
assert exit 0 with no force-exit banner — the drain-then-exit lifecycle under
the actual signal, not just the shared exit path.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix: doctor's FAIL verdict was zeroed by the owned exit — sweep stragglers + class pin
The merged-state suite caught it: doctor --fast --json reported FAIL but
exited 0. Master's v0.42.41.0 brought raw `process.exitCode =` writes
(doctor.ts hasFail ternary, extract.ts) that the #2084 verdict-owning exit
silently zeroes — getCliExitCode() deliberately never reads ambient
process.exitCode (the PGLite-Emscripten pollution defense), so any setter
that bypasses setCliExitCode reports success on failure.
Swept both sites and added the structural class pin: a test greps src/ for
raw `process.exitCode =` outside cli-force-exit.ts, so the next merge that
introduces one fails loudly instead of lying about exit codes.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* chore: bump version and changelog (v0.42.43.0)
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* test: quarantine the watch SIGINT subprocess test to the serial lane
The parallel unit shards flake on concurrent CLI subprocess spawns (failed
at 7ms in-suite, green solo) — same isolation rationale as
apply-migrations-pglite-spawn.serial.test.ts and #2141's R3 quarantine.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* docs: update project documentation for v0.42.43.0
Post-ship doc verification against the release diff (#2095 push-based
context + #2084 superset hardening), with a cross-model doc review:
- push-context.md: version tag corrected to v0.42.43.0; per-call knobs
now cover prior_context/days and watch's flag surface accurately;
feedback-log writes described as best-effort; synopsis fence-strip
described as unconditional.
- CLAUDE.md: stale operation count (~47 -> ~90); volunteer_context
release reference corrected to v0.42.43.0.
- KEY_FILES.md: ci-local entry rewritten to current topology (4-shard
parallel default, four Postgres services, transaction-mode PgBouncer
+ GBRAIN_PGBOUNCER_URL/_DIRECT_URL exports); stale E2E file counts
dropped from the selector entry.
- TESTING.md: inventory entries for the new #2084 structural pins
(cli-exit-verdict-pin, cli-pipe-truncation), the push-context test
suite (volunteer-context, watch-command, watch-sigint.serial,
cli-format-volunteer), migrate v117 coverage, and the two new E2E
files (pgbouncer-teardown env gating, volunteer-context-postgres RLS
pin); check:all row corrected (not a superset of verify).
- AGENTS.md + RELEASING.md: ci:local descriptions updated to the
sharded + pooler topology.
- CHANGELOG (wording only, entry preserved): "retrieved" instead of
"opened" for the used-signal, pooler scoped to the local CI gate,
feedback log labeled best-effort.
- llms-config.ts: index the new push-context guide; bundles
regenerated (build:llms) and freshness test green.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* docs(test): correct the v116 reference — the table shipped as migration v117
* fix: pre-landing review hardening — federated alias parallelism, trust-boundary clamps, shared protocol helpers (#2095)
Five specialist reviewers (testing/maintainability/security/performance/
data-migration) on the reconciled diff; every finding applied:
Performance: the alias arm now resolves all granted sources CONCURRENTLY
(a federated caller paid M sequential RTTs per turn — ~355ms at 5 sources
cross-region, inside the reflex's 1.5s budget); watch's session dedupe is
O(1) Set membership instead of a monotonically growing priorContext string
(O(T²) over a long-lived session); getWindowTurns iterates from the tail
(per-turn cost no longer grows with session length); the resolver's
provenance maps fold into the existing candidate pass.
Security: volunteer_context clamps caller-supplied attribution at the trust
boundary — session_id capped at 256 chars (a read-scoped token could bank
~1MiB TEXT per request, retained 90 days), turn logged only when a safe
integer (a non-integer threw inside the batched INSERT and silently dropped
the whole batch). The privacy comments now state precisely what rationale
may contain (the matched entity's surface form — which by construction
resolved to an existing alias/title/slug — never free conversation text).
Maintainability: TURN_PREFIX_RE + formatVolunteeredPage exported from
volunteer.ts and shared by watch/cli (the two surfaces can no longer
drift); volunteerEventRowsFrom is the single VolunteerEventRow assembly
site for all three channels; watch's window default now honors the same
retrieval_reflex_window_turns config knob the reflex reads; the stale
pre-v116 comments swept to pre-v117.
Testing: the two flake-class CRITICALs fixed (pipe test asserts the
backstop banner instead of a cold-CI-hostile 9s wall bound; the SIGINT test
waits on watch's new machine-readable ready line instead of a fixed 15s
sleep — 2.5s and deterministic now); new coverage for the sink's timeout
branch + ghost-reference drop, watch per-turn fail-open, untrusted knob
clamps (min_confidence/max_pages/days), window-cap ordering (newest user
mention survives), serve-IPC suppression passthrough + channel=reflex
logging, windowTurnCount edge semantics, and structural pins for the sink
registration + cycle purge wiring. The exit-verdict pin's grep is now
operator/whitespace-tolerant.
Deferred with TODOs: resolver index shapes for the per-turn query;
batched first-prune after a long dream-cycle gap.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(context): red-team hardening — pre-cap dedupe, delivery-side reflex logging, window clamp
Four red-team findings on the #2095 push-context surface:
- RT1 starvation: watch's session-dedupe Set filtered AFTER volunteerContext's
cap, so a recurring already-pushed entity burned cap slots every turn and
starved fresh pages behind it. VolunteerOpts.excludeSlugs now skips inside
the pointer loop BEFORE the confidence gate and the cap.
- RT3 honest stats: reflex-channel event logging moved from inside the
resolver to the DELIVERY point — serve's resolve-IPC onDelivered hook fires
only after the response write succeeds, and buildReflexAddition logs only
after the per-turn timeout admits the block. A block the client's 250ms
budget abandoned was never injected and no longer counts as volunteered.
(logChannel resolver opt removed; logDeliveredReflexPointers is the seam.)
- RT5 unbounded window: --window-turns is clamped to [1, 64] so a config typo
can't reintroduce the re-scan-everything-per-turn cost class.
- RT2/RT4 documented + filed: PGLite watch connection monopoly (WATCH_HELP,
push-context guide, TODO to route watch via serve IPC); host-resolver
suppression contract at ResolveEntitiesFn (TODO for a capability gate).
Tests: starvation guard (watch + volunteerContext unit), window clamp floor +
ceiling, delivery-side logging (helper writes channel=reflex through the
drained sink; bare resolver writes nothing; empty list no-op), IPC wiring test
rewired to onDelivered. KEY_FILES.md + push-context.md updated; build:llms run.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(context): env-plane window knob works config-less; harden two gateway-state-leak victims
Three CI-only check failures, two root causes:
1. windowTurnCount ignored GBRAIN_RETRIEVAL_REFLEX_WINDOW_TURNS when
loadConfig() returned null (no config file AND no DATABASE_URL — a clean
CI shard with no brain). loadConfig drops its env→config mapping in that
case, so the documented escape hatch silently died and the window fell
back to 4 → windowed extraction widened when the test set window=1 →
prior-turn entity leaked. Fixed: read the env var directly in
windowTurnCount, mirroring reflexEnabled's direct process.env read. This
is a real product bug, not just a test artifact — any config-less host
using the env hatch was affected. Regression test pins it.
2. sync-cost-preview + doctor-federation-health failed only IN-SHARD: a
sibling test configured a non-legacy (ZeroEntropy 1280-d / $0.05) gateway
and never reset it. The legacy-embedding preload only restores the
OpenAI/1536 default when the gateway slot is EMPTY, so a non-empty foreign
config survives into the next file — and a file's beforeAll runs BEFORE
the preload's restoring beforeEach, so federation-health built a
vector(1280) column and its 1536-d fixture hit CheckExpectedDim. My new
test files reshuffled the deterministic file→shard assignment, exposing
this latent ordering bug. Hardened both victims to establish the gateway
state they assert (sync-cost-preview resets to the unconfigured fallback;
federation-health pins legacy 1536 before initSchema) so they're
order-independent. Verified against a simulated leaker run before them.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* test(context): use withEnv() in the window env-hatch test (test-isolation guard)
The regression test added in
|
||
|
|
d4211f4176 |
v0.42.11.0 feat(skillopt): held-out eval gate, honest receipts, ENFORCE + ablation opts (#1759)
* feat(skillopt): wire held-out gate, honest receipts, ENFORCE + ablation opts Wire the F11 held-out gate into the orchestrator at checkpoint acceptance (runHeldOutGate was dead code); parse + thread --held-out through CLI, batch, fleet, background job, and the run_skillopt MCP op. Populate the real receipt.baseline_sel_score (was hardcoded 0) and add a final-test eval (test_score + baseline_test_score) via a shared scoreSkillOnTasks primitive. Fix the --no-mutate proposed.md write (was a stub) and enforce maxRuntimeMin. D16 ENFORCE in core mutation policy (assertBundledMutationHeldOut): mutating a bundled skill in place requires a non-empty (>=5), benchmark-disjoint held-out set or hard-refuses. Add three eval-internal ablation opts (reflectMode, disableValidationGate, optimizerMode='one-shot-rewrite') recorded in the receipt + audit; ROLLOUT_SUCCESS_THRESHOLD named constant. Security: run_skillopt MCP op validates skill_name (kebab-only) and confines caller-supplied benchmark/held-out paths to the skills dir for remote callers. * test(skillopt): held-out gate, ENFORCE, one-shot rewrite, runtime + receipt honesty New test/skillopt/rollout.test.ts (rollout had zero coverage). Held-out ENFORCE unit cases + one-shot-rewrite fence handling (whole-response unwrap, embedded-fence preserved, error path). E2E: F11 held-out BLOCKS/ALLOWS, bundled no-mutate write, reflectMode/disableValidationGate/optimizerMode, maxRuntimeMin abort, receipt baseline/test-score honesty, held-out/benchmark disjointness, D2 no-DB-pollution. * chore: bump version and changelog (v0.42.9.0) Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * docs: document skillopt held-out gate + bundled mutation requirement for v0.42.9.0 Wire --held-out into the skill-optimizer SKILL.md, guide flags/safety tables, and the tutorial's bundled-skill step: mutating a bundled skill in place now requires --allow-mutate-bundled AND --held-out (>=5 benchmark-disjoint tasks) or it hard-refuses. Add the --held-out flag row + F11 held-out gate to the guide; update the receipt contract to the honest baseline/test-score fields. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * fix(gateway): AI SDK v6 toolLoop compat — multi-turn tool calls work again The ai@6.x bump tightened ModelMessage + tool-schema validation, which silently broke every multi-turn tool loop. Both `gbrain skillopt` rollouts and production background `subagent` jobs route through `chat()`/`toolLoop` and crashed the moment the model called a tool ("messages do not match the ModelMessage[] schema" / "schema is not a function"). Surfaced end-to-end by the SkillOpt real-LLM eval. Three fixes: - chat(): wrap tool defs with the SDK's `jsonSchema()` helper instead of a bare `{jsonSchema}` object (v6 asSchema() treated the bare object as a thunk and threw). - chat(): new exported pure `toModelMessages()` converts gbrain's provider-neutral ChatMessage[] into v6 ModelMessage[] — tool results ride a dedicated `role:'tool'` message with structured `{type,value}` output; null output preserved as json null. Load-bearing for the production subagent path, not just skillopt. - rollout.ts: replace the inline params→schema mapper (dropped `items` on array params) with the shared `paramDefToSchema` single source of truth. Pinned by test/gateway-model-messages.test.ts (8 cases). Folds into the open v0.42.9.0 PR (#1759) — these complete the eval-readiness wave by making skillopt actually run against a live model. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * fix(skillopt): budget no-pricing for Haiku silently scored every rollout 0 Surfaced by the SkillOpt real-LLM eval (Track B). Two coupled bugs that made a budget-capped Haiku run report a vacuous "0/N" measurement in ~2ms with zero LLM calls — indistinguishable from a real deficient-skill score: 1. Claude Haiku 4.5's canonical dateless id (`claude-haiku-4-5`) was missing from anthropic-pricing.ts (only the dated `-20251001` was present). With `--max-cost` set, BudgetTracker.reserve() threw no_pricing on the FIRST chat() of every rollout. Added the dateless entry (sonnet already had its dateless form). 2. runValidationGate swallowed that BUDGET_EXHAUSTED error — runWithLimit settled it as {ok:false}, which the gate turned into median:0. A pricing/cap crash became a fake score. The gate now scans settled results for isMustAbortError() and re-throws so the caller aborts loudly; ordinary (non-abort) rollout errors still fail-open to 0 (judge-hiccup posture kept). Pinned by test/skillopt/validate-gate-abort.test.ts (3 cases). Folds into the open v0.42.9.0 PR (#1759). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * fix(ci): llms-full.txt over size budget — drop what-schemas-unlock from full bundle The toolLoop + budget bug-fix annotations grew CLAUDE.md, pushing llms-full.txt to 756KB over the 750KB FULL_SIZE_BUDGET (the `build-llms > size budget` test failed, failing the `test` CI job). CLAUDE.md stays inlined by design (it's the point of the one-fetch bundle), so per the budget comment's own guidance ("ship with includeInFull=false exclusions") this excludes docs/what-schemas-unlock.md (15.4KB value-explainer, not load-bearing operational reference) from llms-full.txt; it stays linked in llms.txt. Bundle now 740KB with ~9KB headroom. No budget bump — 750KB is near the ~190k-token-context fit ceiling. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * chore(ci): re-admit policy docs into ci-cache-hash before doc relocation docs/**/*.md is deny-listed from the CI cache hash (test-irrelevant). The CLAUDE.md restructure moves test/release POLICY into docs/TESTING.md + docs/RELEASING.md, which DO carry contracts the test suite reads. Without re-admitting them, a policy-only edit would produce the same cache hash and skip the test shard that runs the build-llms + doc-history guards (false-pass). Adds an ALLOW_PATTERNS re-admit step after the deny, scoped to the named policy docs (not a blanket docs un-deny). Lands FIRST, before any doc moves. Pinned by 3 new cases in test/scripts/ci-cache-hash.test.ts: TESTING.md + RELEASING.md edits MUST change the hash; docs/guide.md still must not. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * refactor(docs): relocate Key files / thin-client / Testing out of CLAUDE.md (verbatim) CLAUDE.md had grown to 592KB / ~147k tokens auto-loaded every session (~77% of the llms-full.txt single-fetch bundle). The per-file index was append-only by mandate. This is the exact thin-dispatcher-vs-fat-blob anti-pattern gbrain exists to fix, so CLAUDE.md becomes a thin orientation + resolver that points at on-demand docs. This commit is the VERBATIM move (content-preserving — the next commit compresses): - docs/architecture/KEY_FILES.md <- ## Key files + the calibration key-files cluster + Schema Cathedral v3 impl detail - docs/architecture/thin-client.md <- ## Thin-client routing - docs/TESTING.md <- ## Testing - ## Commands DROPPED (18 'added in vX.Y' history blocks; current surface is gbrain 0.41.38.0 -- personal knowledge brain USAGE gbrain <command> [options] SETUP init [--pglite|--supabase|--url] Create brain (PGLite default, no server) migrate --to <supabase|pglite> Transfer brain between engines upgrade Self-update check-update [--json] Check for new versions doctor [--json] [--fast] Health check (resolver, skills, pgvector, RLS, embeddings) integrations [subcommand] Manage integration recipes (senses + reflexes) PAGES get <slug> Read a page put <slug> [< file.md] Write/update a page delete <slug> Delete a page list [--type T] [--tag T] [-n N] List pages SEARCH search <query> Keyword search (tsvector) query <question> [--no-expand] Hybrid search (RRF + expansion) ask <question> [--no-expand] Alias for query IMPORT/EXPORT import <dir> [--no-embed] Import markdown directory sync [--repo <path>] [flags] Git-to-brain incremental sync sync --watch [--interval N] Continuous sync (loops until stopped) sync --install-cron Install persistent sync daemon export [--dir ./out/] Export to markdown export --restore-only [--repo <p>] Restore missing supabase-only files [--type T] [--slug-prefix S] With optional filters FILES files list [slug] List stored files files upload <file> --page <slug> Upload file to storage files upload-raw <file> --page <s> Smart upload (size routing + .redirect.yaml) files signed-url <path> Generate signed URL (1-hour) files sync <dir> Bulk upload directory files verify Verify all uploads EMBEDDINGS embed [<slug>|--all|--stale] Generate/refresh embeddings LINKS link <from> <to> [--type T] Create typed link unlink <from> <to> Remove link backlinks <slug> Incoming links graph <slug> [--depth N] Traverse link graph (returns nodes) graph-query <slug> [--type T] Edge-based traversal with type/direction filters [--depth N] [--direction in|out|both] TAGS tags <slug> List tags tag <slug> <tag> Add tag untag <slug> <tag> Remove tag TIMELINE timeline [<slug>] View timeline timeline-add <slug> <date> <text> Add timeline entry TOOLS extract <links|timeline|all> Extract links/timeline (idempotent) [--source fs|db] fs (default) walks .md files; db iterates engine pages [--dir <brain>] brain dir for fs source [--type T] [--since DATE] filters (db source) [--dry-run] [--json] publish <page.md> [--password] Shareable HTML (strips private data, optional AES-256) check-backlinks <check|fix> [dir] Find/fix missing back-links across brain lint <dir|file> [--fix] Catch LLM artifacts, placeholder dates, bad frontmatter orphans [--json] [--count] Find pages with no inbound wikilinks salience [--days N] [--kind P] v0.29: pages ranked by emotional + activity salience anomalies [--since D] [--sigma N] v0.29: cohort-based statistical anomalies (tag, type) transcripts recent [--days N] v0.29: recent raw .txt transcripts (local-only) dream [--dry-run] [--json] Run the overnight maintenance cycle once (cron-friendly). See also: autopilot --install (continuous daemon). check-resolvable [--json] [--fix] Validate skill tree (reachability/MECE/DRY) report --type <name> --content ... Save timestamped report to brain/reports/ BRAIN (capture / ideate / explore — v0.37/v0.38) capture [content] [--file PATH] Single entrypoint for getting content into the brain [--stdin] [--slug s] [--type t] Inline content / file / stdin; writes to inbox/ by default [--source ID] [--quiet|--json] Multi-source brains: route to a non-default source brainstorm <question> [--json] Bisociation idea generator (hybrid search + far-set + judge) [--save|--no-save] [--limit N] lsd <question> [--json] Lateral Synaptic Drift: inverted-judge brainstorm [--save|--no-save] [--limit N] rewarding far-from-obvious + axiomatic inversions SOURCES (multi-repo / multi-brain) sources list Show registered sources sources add <id> --path <p> Register a source (id = short name, e.g. 'wiki') sources remove <id> Remove a source + its pages sync --all Sync all sources with a local_path sync --source <id> Sync one specific source repos ... DEPRECATED alias for 'sources' (v0.19.0) CODE INDEXING (v0.19.0 / v0.20.0 Cathedral II) code-def <symbol> [--lang l] Find the definition of a symbol across code pages code-refs <symbol> [--lang l] Find all references to a symbol (JSON-first) code-callers <symbol> Who calls this symbol? (v0.20.0 A1) code-callees <symbol> What does this symbol call? (v0.20.0 A1) query <q> --lang <l> Filter hybrid search to one language (v0.20.0) query <q> --symbol-kind <k> Filter to symbol type (function|class|method|...) (v0.20.0) reconcile-links [--dry-run] Batch-recompute doc↔impl edges (v0.20.0) reindex-code [--source id] [--yes] Explicit code-page reindex (v0.20.0) sync --strategy code Sync code files into the brain JOBS (Minions) jobs submit <name> [--params JSON] Submit background job [--follow] [--dry-run] jobs list [--status S] [--limit N] List jobs jobs get <id> Job details + history jobs cancel <id> Cancel job jobs retry <id> Re-queue failed/dead job jobs prune [--older-than 30d] Clean old jobs jobs stats Job health dashboard jobs work [--queue Q] Start worker daemon (Postgres only) ADMIN stats Brain statistics health Brain health dashboard history <slug> Page version history revert <slug> <version-id> Revert to version features [--json] [--auto-fix] Scan usage + recommend unused features autopilot [--repo] [--interval N] Self-maintaining brain daemon config [show|get|set] <key> [val] Brain config storage status [--repo <path>] Storage tier status and health [--json] (git-tracked vs supabase-only) serve MCP server (stdio) serve --http [--port N] HTTP MCP server with OAuth 2.1 --token-ttl N Access token TTL in seconds (default: 3600) --enable-dcr Enable Dynamic Client Registration --public-url URL Public issuer URL (required behind proxy/tunnel) call <tool> '<json>' Raw tool invocation version Version info --tools-json Tool discovery (JSON) Run gbrain <command> --help for command-specific help. + the per-command KEY_FILES entries; content stays in git) CLAUDE.md gains: a Reference map (resolver), a Maintaining section (the anti-disease rule), and a Cross-cutting invariants subsection under Architecture so the must-never-violate rules (trust fail-closed, sourceScopeOpts isolation, JSONB trap, engine parity, contract-first, migrations, multi-source) still auto-load after the index moved out. Result: CLAUDE.md 592KB -> 61KB; llms-full.txt 740KB -> 210KB (new docs link-only until compressed). build-llms drift + budget test green; verify 29/29 green. The pre-move content is recoverable at git show <this^>:CLAUDE.md. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * refactor(docs): compress relocated docs to current-state + add recurrence guard Compresses the verbatim-relocated reference docs from append-only release-history to current-state-only (the disease cure), then makes recurrence structurally impossible via a CI guard. Compression (fan-out subagents + adversarial verify, audited mechanically): - KEY_FILES.md 453KB -> 356KB; TESTING.md 42KB -> 38KB; thin-client.md already clean. - 393/393 entries preserved; every src/test/scripts path from the verbatim original survives (mechanical comm-check); zero bolded **v0. markers remain. - Conservative ratio (~22%) because the content is invariant-dense — correctness over brevity. Dropped: **vX.Y.Z (#NNN):** clauses, codex/review tags, contributor credits, PR-numbers-as-ids, pre-fix/then/was-now history deltas. Kept: every exported symbol, invariant, and Pinned-by reference. Verbatim original recoverable at git show <relocation-commit>:docs/architecture/KEY_FILES.md. Recurrence guard (scripts/check-key-files-current-state.sh, wired into verify + check:all): - HARD: bans the bolded **v0.<digit> marker in the reference docs (scoped — plain 'as of pgvector 0.7' prose is fine, no false positives). - HARD: CLAUDE.md size cap (90KB; currently 61KB) — the structural backstop. - Pinned by test/scripts/check-key-files-current-state.test.ts (7 cases). Content contracts (test/build-llms.test.ts, +5 cases per codex outside-voice): CLAUDE.md keeps inline ship IRON RULES (version format, document-release, never-hand-roll); AGENTS.md keeps its boot order; llms indexes the new docs; KEY_FILES stays link-only (not inlined). Privacy: scrubbed the relocated 'wintermute/chat/' source-boost examples + the literal harvest-lint regex to generic placeholders (legitimate in allowlisted CLAUDE.md; genericized for the new public docs per the privacy rule). Reverts the |