Two robustness fixes to `gbrain autopilot --install`/`--status`, hardening #3305.
1. Universal bun PATH (extends #3305). The install-generated wrapper
(~/.gbrain/autopilot-run.sh) execs the `#!/usr/bin/env bun` gbrain shim, so
bun must be on PATH under cron/systemd/launchd's minimal env. #3305 hardcodes
`$HOME/.bun/bin`, which only covers the default bun.sh installer. Hosts where
bun lives elsewhere (Homebrew, npm -g, Docker /usr/local/bin, custom
BUN_INSTALL, nix) still die with `env: bun: No such file or directory`,
leaving a stale lock that stalls the nightly cycle. Fix: bake the dir of the
actually-running bun (dirname(process.execPath)) onto PATH at install time,
~/.bun/bin kept as fallback, single-quote-escaped, empty execPath guarded.
2. `--status` false negative. showStatus() checked crontab.includes('gbrain
autopilot'), but --install writes a line calling the wrapper
`.../autopilot-run.sh` — no such substring. So `--status` reported
installed:false on every wrapper-based Linux host. Fix: also match
'autopilot-run.sh'.
Tests: test/autopilot-install.test.ts — universal-form + runtime-derivation +
wrapper-detection assertions (fail-before/pass-after verified).
* v0.42.67.0 fix(agent): provider-neutral help + one truthiness parser for the gateway-loop toggle (#2753)
The gbrain agent help described --model as Anthropic-only and named only
ANTHROPIC_API_KEY. It also overclaimed that any recipe works and that MCP
submitters get permission_denied.
Reviewing that turned up a live mismatch: the doctor accepted true/1/yes/on
for agent.use_gateway_loop, the subagent worker accepted only true/1. So
config set ... yes reported healthy and still refused the job. Both now share
isConfigTruthy() in src/core/config.ts.
Item 1 of the issue (registering the key) is already on master, so this scopes
to the help text, the parser, and the regression test.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
* drop VERSION/package.json/CHANGELOG bump — contributor PRs in this repo do not carry it
Checked precedent on my own merged PRs (#3253, #3248, #3241, #3236): none
touch VERSION, package.json or CHANGELOG. The version-first title + 5-file
sync rule in CLAUDE.md is the maintainer ship flow, not the contributor path.
Carrying the bump here would just hand the maintainer a guaranteed conflict
on every merge.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 5 <noreply@anthropic.com>
* skillify: make the Phase 0 gate fail closed
The gate only rejected when all three answers were no, but each
criterion's parenthetical reads as individually disqualifying
("One-off work != skill"). A one-line alias used once answers
No/No/Yes and runs the entire pipeline - up to 9 frontier eval
calls, four test layers, resolver wiring - and gets certified
properly skilled.
Any single no now stops the run, with the forbidden follow-on
work enumerated so executors cannot rationalize past it.
* skillify: add an upper-bound scope check to Phase 0
Phase 0 only guarded the lower bound (one-off, trivial), so an
entire multi-feature subsystem answered yes to all three checks
and became one mega-skill. In that shape the cross-modal eval
diagnoses the problem (every model says split it) but no phase
can act on the advice - decomposition is not a file edit - so
the only path is ship-with-KNOWN_GAPS, and Phase 4 then locks
the below-bar scope in with tests: the exact tests-cement-
mediocrity outcome the eval gate exists to prevent.
Multi-intent targets now stop in Phase 0 with a proposed split
and a question about which target to skillify first.
The check asks about the set of intents rather than the
existence of a trigger phrase, because check 3 is existential
and any one phrase ("ship it") makes a subsystem answer yes.
Anthropic released Claude Opus 5, at the same $5/$25 pricing tier as
Opus 4.8. Neither the chat recipe allowlist nor CANONICAL_PRICING knew
about it, so operators could not opt into it via models.tier.deep /
models.default without gbrain rejecting the id.
- src/core/ai/recipes/anthropic.ts: add claude-opus-5 to the models list.
- src/core/model-pricing.ts: add anthropic:claude-opus-5 { input: 5.00,
output: 25.00 } (plus cache rates, matching Opus 4.8's ratios).
- src/core/takes-quality-eval/pricing.ts: add it to SUPPORTED_MODELS so
eval takes-quality run --budget-usd doesn't reject it during preflight.
- Refreshed the stale pricing-verification date and the Opus list in
docs/architecture/KEY_FILES.md.
- Tests: pinned-value regression in test/model-pricing.test.ts, recipe
membership in test/anthropic-model-ids.test.ts, budget-pricing coverage
in test/eval-takes-quality-pricing.test.ts.
Scope: registration only. TIER_DEFAULTS / DEFAULT_ALIASES /
DEFAULT_CHAT_MODEL are untouched — default-routing bumps are the
separate, already-open #2858; this just makes the id valid/priced for
operators who opt in explicitly.
* chore(ci): refresh GitHub Actions SHA pins (checkout v4, action-gh-release v2)
Pre-ship pin staleness check per docs/RELEASING.md: both floating major
tags moved upstream; pins updated to the current tag commits.
* v0.42.65.0 chore(release): 92 verified fixes since v0.42.64.0 — changelog + version bump
Aggregates everything merged to master since the v0.42.64.0 bump commit:
community fixes, credited takeovers, batch re-lands, CI hardening, and
maintainer-approved features. Net commit list excludes revert pairs.
No new schema migrations.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(deps): clear OSV-flagged transitive dependencies via override floors
Raise the existing security-floor overrides so the lockfile resolves
patched versions of three transitive packages flagged by the OSV scan
(@hono/node-server, fast-uri, body-parser). None are on gbrain's own
runtime path (@hono/node-server is only referenced by the MCP SDK's
optional hono transport, which gbrain does not load); the floors keep
the dependency scan green. MCP/OAuth unit tests pass against the
resolved versions.
* chore(release): fold #3110 into the v0.42.65.0 entry (93 net changes)
* v0.42.66.0 chore(release): 54 verified fixes since v0.42.65.0 — changelog + version bump
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
---------
Co-authored-by: Garry Tan <garrytan@gmail.com>
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
list_pages clamps limit to max 100 (default 50) — deliberate server
protection, pinned in test/search-limit.test.ts. But the clamp was
SILENT: a caller whose limit was defaulted or clamped got a
full-looking array with no signal that rows were dropped, and with the
default updated_desc sort the dropped rows are always the OLDEST —
precisely what exhaustive consumers (audits, scans, backfills) exist
to find. Observed in the field: a source with 212 pages enumerated as
80 visible rows, hiding 26 pages from a compliance scan for days.
Fix, with no response-shape change (MCP consumers still get an array)
and no engine surface change (handler probes limit+1):
- handler probes one row past the effective limit; when the caller's
limit was NOT honored (unset -> default, or clamped to cap) and rows
were dropped, it warns on stderr for local (CLI) callers — same
operator-facing channel as the put_page unknown-type hint, but
without the isTTY gate: scripted callers are exactly the consumers
that cannot detect truncation any other way, and stderr keeps stdout
parseable. An explicit honored limit stays silent (ordinary
pagination), as does a clamped-but-complete result. Remote (MCP)
ctx never writes to stderr.
- LIST_PAGES_DESCRIPTION documents the cap and the exhaustive-listing
recipe (sort=updated_asc + updated_after cursor) — the description
is the signal channel MCP clients actually read.
- regression suite: default-limit truncation warns, honored limit
silent, clamped-but-complete silent, remote silent, and the
documented cursor recipe enumerates a corpus to completion.
Co-authored-by: paul-0320 <paul@ymyd.co.kr>
Co-authored-by: YMYD <53603073+OJ-OnJourney@users.noreply.github.com>
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
Co-authored-by: Garry Tan <garrytan@gmail.com>
* v0.42.66.0 fix(extract): make conversation backfill outcomes durable (takeover of #3293)
Versioned, snapshot-bound terminal audit rows become the durable authority
for conversation fact backfill completion; checkpoint GC can no longer
repeat completed model work, and best-effort empty results no longer mask
provider/output failures as complete.
Supersedes #3293 (rebased onto current master; only version-trio conflicts).
Co-authored-by: FloridaStyle <danwiggins@users.noreply.github.com>
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* chore: drop version-trio bump — individual fixes do not carry version bumps (release PRs do)
* merge: reconcile durable-outcome skip accounting with master's LLM fallback tests
The two fallback replay tests from #3371 asserted the legacy checkpoint
pages_skipped counter; under this PR's durable-outcome authority a
completed page is skipped via pages_skipped_completed before any parse.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
---------
Co-authored-by: Garry Tan <garrytan@gmail.com>
Co-authored-by: FloridaStyle <danwiggins@users.noreply.github.com>
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
* feat(recipes): add reranker touchpoint to OpenRouter (#2164)
OpenRouter's POST /api/v1/rerank is wire-compatible with gateway.rerank()
({query, documents, model} → {results: [{index, relevance_score}]}). This
adds a recipe-only reranker touchpoint declaring four models:
- cohere/rerank-v3.5 (default; $0.001/search)
- cohere/rerank-4-fast ($0.002/search, 32K context)
- cohere/rerank-4-pro ($0.0025/search, SOTA quality)
- nvidia/llama-nemotron-rerank-vl-1b-v2:free (multimodal)
Unlike embedding/chat, the reranker path strictly enforces the models
allowlist — the openai-compat extended-model bypass does not apply. New
rerank models must be added to this recipe before they can be called.
The cost_per_1m_tokens_usd value is a pseudo-rate for the budget tracker's
chars/4 heuristic — Cohere bills per-search, not per-token. At ~4K chars
the estimated cost is in the right ballpark.
Recipe-only change; no gateway or search-layer modifications. gateway
auto-concatenates path → .../api/v1/rerank.
Adds hermetic unit test (test/openrouter-reranker-recipe.test.ts) covering
shape, models, default_model, path, max_payload_bytes, default_timeout_ms,
and cost field. No DB, no env mutation — survives the parallel 8-shard
fan-out.
Verified: bun run verify (30/30 green); 285 targeted recipe+rerank+budget
tests pass.
Co-authored-by: Hippityy <Hippityy@users.noreply.github.com>
* test(facts): pin gateway to 1536d in facts-engine.test.ts beforeAll
Shard-composition hermeticity fix. The legacy preload's beforeEach only
re-applies the 1536-d gateway default before each TEST, not before a
file's beforeAll — so when the previous file in the shard resets the
gateway in its teardown (e.g. test/providers-test-model-base-url.test.ts
via afterEach), this file's initSchema() sized facts.embedding at the
1280-d production default and the 1536-d fixture inserts threw
'expected 1280 dimensions, not 1536' (CI shard 1 failure on #3302).
Same pattern as test/consolidate-valid-until.test.ts.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
---------
Co-authored-by: Ryan Xie <64182766+Hippityy@users.noreply.github.com>
Co-authored-by: Hippityy <Hippityy@users.noreply.github.com>
Co-authored-by: Garry Tan <garrytan@gmail.com>
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
When link_resolution.global_basename is enabled, extend basename-index
resolution to frontmatter link fields (FRONTMATTER_LINK_MAP), mirroring the
body bare-wikilink path added in #972.
Problem: a bare-title wikilink in a frontmatter list -- e.g.
sources:
- "[[2025-12-25_mentor-extraction]]"
never resolves. SlugResolver.resolve() has no '/' to hit the slug-direct
getPage, and the field's dirHint (sources -> ['source','media']) may name
folders absent from the brain, so the dir-scoped exact + fuzzy steps also
miss. The frontmatter path never consulted resolveBasenameMatches -- that was
wired only for body bare-wikilinks. On a PARA/Obsidian vault this silently
drops the bulk of sources:/related: provenance edges.
Fix: extractFrontmatterLinks takes a globalBasename flag (threaded from
extractPageLinks). On a resolve() miss, unwrap [[ ]] and fall back to
resolver.resolveBasenameMatches -- UNIQUE-MATCH-ONLY, so ambiguous basenames
(archive dupes, generic hubs like _index) stay unresolved rather than create
a wrong edge. Purely additive; resolved frontmatter edges are unchanged.
Scope: covers the db-source extract and live put_page paths (real
makeResolver). The --source fs extract uses an inline resolver without a
basename index, so it gracefully no-ops there (typeof guard).
Tested: 3 new cases (resolves-when-on, ambiguous-stays-unresolved,
gated-off-by-flag); full link-extraction suite green (130 pass).
Co-authored-by: spiky02plateau <155588579+spiky02plateau@users.noreply.github.com>
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Co-authored-by: Garry Tan <garrytan@gmail.com>
* fix(onboard): stop repeating the same auto-remediation within a run (#2854)
When the recommendation list is refreshed between remediation steps, a
remediation that doesn't clear its own health signal is reintroduced
under its stable id and attempted again, indefinitely on long runs.
Track attempted recommendation ids for the run and skip re-attempts.
Includes a behavioral regression test: a persistently-stuck signal is
attempted once, the loop terminates, and other remediations still run.
* fix(test): quarantine remediation-run-loop test as serial + complete BrainHealth fixture
Two CI failures, one root cause each:
- verify (check:test-isolation + typecheck): the new test uses mock.module
(R2) so it must live in the *.serial.test.ts quarantine, and the
BrainHealth fixture was missing the now-required linkable_page_count.
- test (6): the top-level mock.module('../src/core/ai/gateway.ts') leaked
into other files in the parallel shard process, flaking
test/ai/adaptive-embed-batch.test.ts. Serial quarantine fixes it —
run-serial-tests.sh executes each serial file in its own bun process.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
---------
Co-authored-by: Sanchal Ranjan <84386862+sanchalr@users.noreply.github.com>
Co-authored-by: Garry Tan <garrytan@gmail.com>
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
Dispatch timeout was derived as interval*2 with a 5-minute floor, tuned
for light per-interval work. A full autopilot cycle routinely needs more
than 10 minutes at common intervals, so healthy full cycles were killed
mid-run. Full-cycle dispatch now gets a 30-minute floor; lighter
dispatches keep the interval-derived budget.
Adds a regression test for the full-cycle floor.
Co-authored-by: Sanchal Ranjan <84386862+sanchalr@users.noreply.github.com>
* fix(serve): boot-readiness deadline releases PGLite lock on wedged boot (#3273)
A serve process that wedges mid-boot (e.g. a boot step blocked on an
unreachable upstream) held the PGLite write lock indefinitely — the
post-#2348 lock discipline never steals from a live holder, so every CLI
consumer timed out until the serve PID was manually killed.
runServe (stdio path) now arms a boot-readiness deadline around
startMcpServer: if the transport hasn't connected within
GBRAIN_SERVE_BOOT_TIMEOUT_SECONDS (default 60, 0 disables), it logs the
condition, awaits engine.disconnect() (raced against the existing
5s cleanup deadline so a wedged WASM close can't trap it either), and
exits non-zero so supervisors restart with backoff. A completed boot
clears the timer; the HTTP path is untouched (own lifecycle).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* test(e2e): 60s hook timeout for jsonb-parity setup/teardown
The #2339 parity guard's beforeAll runs setupDB (full migration chain)
under bun's default 5s hook timeout, which flaked on a slow CI runner
(setupDB hit 5001ms). Other e2e suites already pass explicit hook
timeouts; bring this file in line.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
---------
Co-authored-by: Garry Tan <garrytan@gmail.com>
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
`sources.config` is a jsonb OBJECT column, but a read→write cycle that
JSON.stringify'd an already-stringified value re-wrapped it into a JSON string
scalar ("{}", "\"{}\"", ...) that grew one layer per write. parseSourceConfig
only unwrapped one layer, so the corruption never healed and federation/ACL
reads saw a string instead of the settings object.
- Add normalizeSourceConfig: a bounded (10-iteration) loop that JSON.parses
while the value is a string and returns {} (with a console.warn) when the
result is not a plain object. All six `UPDATE sources SET config` writers run
their config through it before stringify, converging the stored value back to
a jsonb object on the next write.
- parseSourceConfig now does the same bounded unwrap and warns once when more
than one layer was found (one layer is the normal PGLite path).
- Add a `source_config_shape` doctor check that flags any sources row where
jsonb_typeof(config) <> 'object', with the repair path.
- Unit-test the helper (object passthrough, 1-layer, 5-layer nested, garbage
and over-bound inputs) and the doctor check (mock engine).
Co-authored-by: 1alessio <alessio.sulpizi@gmail.com>
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
gbrain list --limit 100000 silently returned 100 rows (default 50) with
no warning, and --offset was accepted but dropped at the op layer even
though PageFilters has supported it all along.
- Local CLI callers (ctx.remote === false, the same trust boundary that
already bypasses scope enforcement) get an explicit limit above 100
honored — full enumeration is a legitimate local operation.
- Remote MCP/OAuth callers keep the 100-row DoS cap, now loud: one
logger.warn (stderr, stdout stays script-clean) with both numbers,
parity with the three search-path clamp warnings.
- offset is declared as a param (so the CLI coerces it to number) and
threaded to engine.listPages for real pagination.
Claude-Session: https://claude.ai/code/session_01Vswwe1y5fQbJWfbaSK3enT
Co-authored-by: Deacon Bot Doctor <deacon@botdoctor.io>
Co-authored-by: deacon-botdoctor <291411030+deacon-botdoctor@users.noreply.github.com>
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
The idempotency row is only written inside `for (const p of proposals)`, so a
page that extracts ZERO gradeable claims never records an idempotency tuple
and is re-sent to the LLM on every cycle forever. The docstring's "unchanged
page never re-spends tokens" contract only holds for pages that produce >=1
claim; a page that legitimately has no gradeable claims (or any machine-
generated page) is a perpetual cache miss and re-spends tokens indefinitely.
Fix: when `proposals.length === 0`, write one tombstone row keyed by the same
(source_id, page_slug, content_hash, prompt_version) tuple, with
status='rejected' so it never surfaces in a pending-review query (the pending
index filters status='pending'). Content changes (new content_hash) or a
PROPOSE_TAKES_PROMPT_VERSION bump still miss the tombstone and re-extract. The
extractor-throw path `continue`s before the tombstone, so failed pages are
retried rather than cached.
Guard against a subtle regression: `parseExtractorOutput` returns [] for BOTH
a genuine empty extraction AND malformed/prose/truncated model output, so
naively tombstoning every [] would permanently suppress a page that has claims
but hit a transient parse failure. `defaultExtractor` now throws when the
output is empty-but-not-a-clean-`[]` (new `isWellFormedEmptyExtraction`
predicate), routing transient failures into the existing retry path; only a
cleanly-parsed empty array is memoized.
Adds a `tombstones_written` counter for observability.
Tests: tombstone written on genuine empty extraction; two-cycle idempotency
(no repeat LLM call on an unchanged zero-claim page); extractor error writes
no tombstone; isWellFormedEmptyExtraction discriminates clean-[] from
malformed/prose/non-empty output. propose-takes suite: 36 pass / 0 fail.
Co-authored-by: ivandebot <ivanlanlei@gmail.com>
Co-authored-by: ivandebot <187176982+ivandebot@users.noreply.github.com>
Co-authored-by: Garry Tan <garrytan@gmail.com>
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
* fix(search): honor recency decay config on the hybrid path (#2386)
The hybrid recency stage in runPostFusionStages imported
DEFAULT_RECENCY_DECAY directly, so operator overrides via the
GBRAIN_RECENCY_DECAY env var and the gbrain.yml `recency:` section were
honored only on the get_recent_salience SQL path and silently ignored on
the hot hybridSearch path. Non-default vault layouts therefore stayed on
the baked-in defaults / DEFAULT_FALLBACK (90d / 0.5) regardless of
tuning.
Call resolveRecencyDecayMap() (already used by the SQL path) so the
configured decay map reaches the boost stage. Behavior is unchanged when
no override is set — resolveRecencyDecayMap() returns DEFAULT_RECENCY_DECAY.
Adds test/hybrid-recency-config.test.ts asserting the env override
reaches the applied recency factor (fails against the prior wiring).
* test: use withEnv() in hybrid-recency-config test (check-test-isolation R1)
The test-isolation lint (shipped after #2386 was written) rejects raw
process.env mutation in non-serial test files. Wrap the
GBRAIN_RECENCY_DECAY overrides in withEnv() from test/helpers/with-env.ts;
assertions unchanged.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
---------
Co-authored-by: Richard Baker <rich@rwbaker.com>
Co-authored-by: Garry Tan <garrytan@gmail.com>
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
* fix(frontmatter): derive validate slug from brain root, not absolute path (#2340)
Single-file `frontmatter validate` derived the expected slug from the
absolute path: relative(resolve(target), file) is empty when target IS the
file, so it fell back to `|| file` (the full path), yielding "root/<abs>"
slugs and a false SLUG_MISMATCH. The pre-commit hook from install-hook
validates staged files one-by-one, so this rejected every commit in a
markdown brain (only bypassable with --no-verify).
Walk up to the brain root (nearest .git) and use relative(brainRoot, file)
|| basename(file), matching runAudit/runGenerate and sync/extract. Files
above the root fall back to basename instead of a ../-prefixed slug.
Reopens#565. Present since v0.32.0; reproduced on v0.42.51.
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* test(facts): pin embedding dims in facts-engine — kill the shard-order 1280/1536 flake
facts-engine.test.ts hardcodes Float32Array(1536) vectors (vec()) but lets
initSchema size its vector columns from process-global gateway state
(getEmbeddingDimensions(), default 1280). Whether the file passes depends
on which test files run before it in the shard; adding
test/frontmatter-validate-slug-565.test.ts reshuffled the weight-packed
shards and tripped it on this PR's CI (test (1):
'expected 1280 dimensions, not 1536' in findCandidateDuplicates cosine
ordering).
Same fix + rationale as doctor-hidden-by-search-policy.test.ts (#2801),
engine-find-trajectory.test.ts and cosine-rescore-column.test.ts:
configureGateway(1536) in beforeAll BEFORE initSchema, resetGateway in
afterAll.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
---------
Co-authored-by: alessioalionco <alessioalionco@gmail.com>
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Co-authored-by: Garry Tan <garrytan@gmail.com>
* fix(autopilot): export ~/.bun/bin onto PATH in cron wrapper (#2013)
The wrapper script that 'gbrain autopilot --install' writes to
~/.gbrain/autopilot-run.sh sources ~/.bashrc to inherit PATH for the
exec'd gbrain binary (which has a '#!/usr/bin/env bun' shebang). The
standard Debian/Ubuntu ~/.bashrc ships a non-interactive guard that
returns early when bash is launched non-interactively (cron, launchd,
systemd) — so PATH exports operators add to ~/.bashrc never reach the
wrapper subprocess.
The result: the wrapper dies silently with 'env: bun: No such file or
directory', leaves a stale lockfile, and every subsequent cron tick
hits the lockfile and bails. The nightly dream cycle hangs waiting on
a worker that never comes back, and the wrapper's own 10-min
stale-lock window is the only thing that can recover it.
This bites every operator whose bashrc is the standard distro default
(which is the default), and there is no warning at install time.
Fix: prepend ~/.bun/bin to PATH directly in the wrapper, so it is
self-contained regardless of which init file the OS loaded. Add a
regression test alongside the existing zshenv/zshrc source-order test
(v0.36.1.x #966) so this class of bug stays caught.
* fix(test): scrub real agent-fork name from regression comment (privacy check)
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
---------
Co-authored-by: klampatech <73077262+klampatech@users.noreply.github.com>
Co-authored-by: Garry Tan <garrytan@gmail.com>
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
Idempotency was keyed on atom rows alone — a page the LLM judges
un-atomizable leaves no row, so it re-entered the discovery window every
run. Two production consequences: --drain false-stopped with
no_progress once the window head was mostly zero-yield pages (remaining
frozen while batches report +0), and every nightly re-spent extraction
budget on the same pages.
Fix:
- After a SUCCESSFUL chat call that parses to zero atoms, stamp the
source page with frontmatter.atoms_scan_hash = contentHash16. LLM
failures take the catch path and stay retryable.
- discoverExtractablePages + countExtractAtomsBacklog (both variants)
exclude pages whose stamp matches the CURRENT content hash prefix —
content edits re-eligibilize, mirroring atom-row staleness semantics.
- Drain no_progress now recounts the backlog on a zero-atom batch and
only stops when it genuinely didn't shrink — tombstoning IS progress.
Tests: +2 pure-loop drain cases (shrinking backlog continues / flat
backlog stops) and +3 PGLite integration cases (stamp + exclusion /
content-change re-eligibility / failed chat does not stamp).
29 pass / 0 fail across the two files; tsc clean.
Co-authored-by: 陈源泉 <84364275+ChenyqThu@users.noreply.github.com>
Co-authored-by: 陈源泉 <chenyuanquan@chenyuanquandeMac-mini.local>
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
Co-authored-by: Garry Tan <garrytan@gmail.com>
Adds the bold-time-dash built-in pattern: **Speaker** HH:MM <dash> text
(em dash, en dash, or ASCII hyphen), valid 24-hour times only, date from
page frontmatter/date headings, multi-line continuation bodies.
Opt-in score_continuations_as_body scoring keeps long multiline messages
parseable while preserving the sparse-prose false-positive floor (needs
two anchors or a first-line anchor before candidate-only scoring kicks in).
Hardens validatePatternEntry to reject non-integer / out-of-range capture
indexes including text_group. Adds maintainer doc, JSONL fixtures, and
adversarial coverage.
Takeover of #3289 (fork branch went CONFLICTING against master on the
version trio); code applied 3-way, version/CHANGELOG bump dropped per
fleet release convention.
Co-authored-by: Garry Tan <garrytan@gmail.com>
Co-authored-by: FloridaStyle <daniel.wiggins@gmail.com>
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
On machines with neither gtimeout nor timeout on PATH, run-verify-parallel.sh
and run-unit-parallel.sh fall back to a bg-pid + sleep-watchdog cap. Both
read $? only after tearing the watchdog down (kill + wait on cap_pid), so the
sentinel .exit files recorded the killed watchdog's status — 143 — instead of
the check/shard's own exit code. Every run reported total failure (verify:
pass=0 fail=31; unit: rc=143 per shard) while every per-check/shard log
showed success.
Capture rc immediately after `wait $pid` in both scripts, and reap the
watchdog's sleep child (pkill -P, children-first — the same orphan quirk the
heartbeat cleanup documents) so the fallback stops leaking one sleep per
check/shard.
Regression tests force the fallback branch hermetically on any host via a
curated PATH with no timeout binaries: the verify dispatcher runs from a
tempdir copy with a stubbed `bun`, pinning exit 0 + all-zero sentinels when
checks pass and the check's own rc (not 143) when one fails; the unit wrapper
runs real two-shard fixture passes, pinning rc=0 sentinels and a real
failure's rc=1.
Co-authored-by: paul-0320 <paul@ymyd.co.kr>
Co-authored-by: YMYD <53603073+OJ-OnJourney@users.noreply.github.com>
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
* fix(embed): stamp gateway-resolved model in content_chunks.model, not compiled default (#2846)
upsertChunks fell back to the compile-time DEFAULT_EMBEDDING_MODEL
('zeroentropyai:zembed-1') when a ChunkInput carried no explicit `model`.
The embed pipeline (src/commands/embed.ts) builds ChunkInputs without a
`model` field, so rows whose vectors were produced by the config-resolved
model (e.g. openai:text-embedding-3-large) were mislabeled with the
hardcoded default — corrupting the provenance that signature-drift
staleness and dimension-migration logic depend on.
Both engines now resolve the gateway's runtime embedding model once per
upsert and use it as the fallback, mirroring the existing resolve-then-
default pattern used for schema sizing. Regression test added (pglite);
verified via negative control that it fails against the old fallback.
This is a write-path change (upsertChunks), not a search-path change, so
retrieval eval replay is not applicable.
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
* test: Lane A.7 pins gateway-resolved chunk model, not compiled default
#2846 changed upsertChunks' fallback from DEFAULT_EMBEDDING_MODEL to the
gateway-resolved runtime model. Lane A.7 still pinned the old fallback,
and the test preload (test/helpers/legacy-embedding-preload.ts) pins the
gateway to openai:text-embedding-3-large for every test process — so the
original #2846 landing failed this test deterministically and got batch-
reverted. The test now asserts the resolved model (the intended #2846
semantics) while keeping the CDX2-4 bare-literal regression guard.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
---------
Co-authored-by: SailorJoe6 <SailorJoe6@Gmail.com>
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
Co-authored-by: Garry Tan <garrytan@gmail.com>
DashScope's OpenAI-compatible rerank endpoint lives at
{base}/compatible-api/v1/reranks — PLURAL leaf, different base path from
the embedding surface (compatible-mode). Reusing llama-server-reranker
against DashScope forces users to hand-patch the recipe's '/rerank' leaf
in node_modules, which every upgrade silently reverts (and llama.cpp
genuinely serves singular /rerank, so changing that recipe would break
real llama.cpp users).
New dedicated recipe rides the v0.40.6.1 recipe-pluggable reranker path:
- id dashscope-rerank, base_url_default compatible-api/v1 (intl), ZE wire
- path '/reranks', default_timeout_ms 30s, 5MB payload ceiling
- models: only qwen3-rerank (live-verified 200; gte-rerank-v2 is rejected
by the compat surface with 'Unsupported model for OpenAI compatibility
mode', so it is deliberately not listed)
- separate recipe (not a reranker touchpoint on dashscope) because
provider_base_urls is keyed by recipe id and the two capabilities need
different prefixes — same topology as llama-server vs
llama-server-reranker
Tests: recipe shape smoke mirroring recipe-llama-server-reranker.test.ts
(path/timeout/payload pins, /v1/v1 concat guard, auth resolve, sibling
recipe isolation). bun test test/ai/: 322 pass / 0 fail.
Co-authored-by: Yicon <charlieyiconghuang@gmail.com>
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
The four `hybridSearch — reranker enabled (reorder)` cases stub the gateway
at 1536 dims (DIMS). Since v0.36.3.0 hybridSearch resolves the embedding
column via loadConfig(), whose precedence is
cfg.embedding_dimensions > gateway dims > default. On any machine whose
~/.gbrain/config.json sets embedding_dimensions to something other than 1536
(e.g. text-embedding-3-small at 1280), the real config outranks the stub: the
1536-d stub vector fails the gateway dim check, the error is swallowed, search
falls back to keyword-only, and the reranker never runs (rerankerFn gets 0
docs, rerank_score undefined). Green in CI only because a fresh runner has no
config file — deterministic red on a contributor's machine.
Fix (test-only): isolate GBRAIN_HOME to an empty tmpdir in beforeAll so
loadConfig() returns null and the stub's dims win, then restore it and clean
up in afterAll. Same idiom as emptyHome() in
test/ai/gateway-probe-chat-model.test.ts.
Verified with a planted ~/.gbrain/config.json at 1280 dims: 2 pass / 4 fail
before, 6 pass / 0 fail after; still green with no config file. typecheck clean.
Fixes#1527
Co-authored-by: Willisbest <132954469+Willisbest@users.noreply.github.com>
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
Reland of #2639, reverted with its batch in 68e4cebd. getHealth's
entity_pages CTE and the top-linked-pages query only match the legacy
'person' and 'company' types, so brains using the gbrain-base-v2 pack's
'entity' type report 0% entity link/timeline coverage in `gbrain health`.
Add 'entity' to both queries in both engines (PGLite + Postgres, in
lockstep per the engine-parity rule).
Reland fix (the batch-red root cause): the original PR's test expected
the entities/project-x page to appear in orphan_pages, but #3023's shared
orphan-reporting policy (landed before #2639 merged) excludes the
'entities' first segment from orphan reporting, so the test failed on
master. Orphan expectations now account for the policy exclusion.
Co-authored-by: Tyler Robinson <tylr.rob@gmail.com>
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
Replace the strictly sequential per-chunk synopsis loop with a bounded
sliding worker pool (existing runSlidingPool helper). Results land in
chunk order via index-addressed writes; code chunks still bypass the
wrapper; embedding remains one page-level batch after all synopses.
New knob GBRAIN_CONTEXTUAL_CHUNK_CONCURRENCY, default 4, clamped to
[1,16]; 1 reproduces the prior sequential behavior exactly. Each chunk
task still acquires/releases the global synopsis rate-lease, which
remains the cross-worker governor; the lease id now travels from
acquire to release instead of shared mutable state, and lease waits
are abort-responsive.
At 20-45s per synopsis call, a 120-chunk transcript page previously
needed 60-90+ min wall time and routinely outlived job timeouts.
Co-authored-by: spiky02plateau <155588579+spiky02plateau@users.noreply.github.com>
statusForVersion short-circuited on any 'complete' entry before checking
the trailing 'retry' marker, so --force-retry appended an inert row and a
version marked complete with zero work done could never be re-run without
hand-editing completed.jsonl. Check retry-latest first: an explicit
--force-retry now yields 'pending' even past an earlier 'complete', while
a stray 'partial' after 'complete' still cannot regress the version.
Co-authored-by: spiky02plateau <155588579+spiky02plateau@users.noreply.github.com>
SupabaseStorage.getSignedUrl built the download URL as `${projectUrl}${signedURL}`,
but Supabase's sign API returns `signedURL` relative to the Storage API root
(/object/sign/<bucket>/<path>?token=...), so the generated link dropped /storage/v1
and returned 404. Now prepends `${projectUrl}/storage/v1`, tolerating an
already-absolute URL or a value that already carries the prefix. `gbrain files
signed-url` links resolve again.
Co-authored-by: FloridaStyle <daniel.wiggins@gmail.com>
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
Reasoning models (MiniMax-M3, DeepSeek-R1, etc.) return <think>...</think>
tags in the content field before the actual JSON output. This caused
parseExtractorOutput to fail in two ways:
1. The fence regex /^\`\`\`(json)?...$/ requires the fence at text start;
<think> preceding it prevents matching, so the raw text (with trailing
fences) hits JSON.parse and throws.
2. When think tags contain [ or { characters, indexOf finds them inside
the reasoning block instead of the actual JSON array.
Changes:
- Strip <think>...</think> tags before any parsing (covers all reasoning models)
- Add JSON.parse fallback: truncate at last ] or } to handle trailing
noise (leftover markdown fences after stripping)
Tests: 28/28 pass (3 new cases for think tags + trailing noise).
Co-authored-by: qaz8545355 <603191978@qq.com>
Co-authored-by: qaz8545355 <junjun@openclaw.local>
The empty-fence guard counted every `row_num IS NULL AND entity_slug IS NOT NULL`
row as a pending v0_32_2 backfill, but the inline facts writer keeps producing
rows of exactly that shape post-migration: when a resolved slug has no fenceable
page (slugify-floor / stub-guard-blocked unprefixed slugs like `wingman`,
`people-jane-doe`), backstop.ts falls through to a DB-only insert with row_num
NULL. Those rows are structurally unfenceable — no page to fence onto, and the
ledger-complete migration won't re-run — so they jammed the phase forever
(~16/day observed) and the warning advised a no-op `apply-migrations --yes`.
Discriminator: a row is a genuine backfill candidate only if its entity_slug
resolves to a LIVE page in the same source (EXISTS in `pages` with deleted_at
NULL) — mirroring the migration's Phase B, which only fences slugs that map to
a writable page. Genuine pre-v0.32.2 rows (their entity page exists) still gate;
inline-writer unfenceable rows no longer do. Warning text updated to name the
"entity page present, not yet fenced" condition.
Regression tests pin both sides: unfenceable rows (no page / soft-deleted page)
do NOT gate and the phase converges; a legacy row WITH a backing page still
gates. Fails pre-fix, passes post-fix.
(#2484)
Reland note: original merge (53c90869) was batch-reverted (4b6cf32c) —
the guard-semantics change broke test/phantom-redirect.test.ts
'round 2 P1: legacy-row guard fires BEFORE phantom-redirect pass',
which seeded a legacy row WITHOUT a backing page and expected the
guard to fire. Under the new (intended) semantics such a row is
structurally unfenceable and must NOT gate. Fixed by seeding a live
backing page for the legacy row, preserving what the test pins
(guard fires before the phantom-redirect pass).
Co-authored-by: Javier Aldape <javieraldape@users.noreply.github.com>
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
Code legitimately contains tiktoken special-token strings (e.g. CLIP/GPT tokenizers embed the literal <|endoftext|>). The default encode() uses disallowed_special='all' and THROWS on those, crashing reindex-code on valid source files. Re-encode treating them as ordinary text (allowed=[], disallowed=[]); heuristic fallback if even that fails. A token COUNT needs no special-token semantics.
Co-authored-by: Jim Tang <jimruitang@gmail.com>
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* fix(write-through): guard mkdir against EEXIST on Bun+Windows
* fix(budget): resolve non-Anthropic model pricing via canonical table under --max-cost
* fix(enrich): exclude dream-generated pages from thin candidates
---------
Co-authored-by: nguyenchiviet <40517873+nguyenchiviet@users.noreply.github.com>
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
OpenRouter (and potentially other proxy providers) expose OpenAI's
text-embedding-3 models with a provider prefix in the model ID, e.g.
`openai/text-embedding-3-large` rather than bare `text-embedding-3-large`.
`dimsProviderOptions()` checks `modelId.startsWith('text-embedding-3')`
which fails for the prefixed form, so the `dimensions` parameter is never
sent. The upstream provider returns its native dimensionality (3072 for
-large) instead of the configured value (e.g. 1536), causing an immediate
"dim mismatch" error on first embed.
The default OpenRouter embedding (`text-embedding-3-small` at 1536d)
masked this because its native size happens to match the default config.
The bug surfaces when using `-large`, or `-small` with a non-1536 dim
(512, 768, 1024 — all listed in the recipe's `dims_options`).
Fix: strip the provider prefix before the `startsWith` check. The full
prefixed ID is preserved in the error message for user clarity.
Co-authored-by: Noetherly <280958447+noetherly@users.noreply.github.com>
* feat(subagent): claude-cli MessagesClient adapter (baseline, no tool use)
Closes#334 (partially — text-only baseline; tool use lands in the next
commit on this branch).
Adds a MessagesClient adapter that shells out to `claude --print
--output-format json --model <model>` instead of the Anthropic SDK. When
`GBRAIN_USE_CLAUDE_CLI=1` is set, the subagent worker registers the adapter
in place of the SDK client; the default path (Anthropic SDK with
ANTHROPIC_API_KEY) is unchanged when the env var is unset or set to
anything else.
The benefit is that Claude Max subscribers can run Minions subagents
against their existing OAuth subscription, no ANTHROPIC_API_KEY needed.
New: src/core/minions/handlers/claude-cli-adapter.ts
- Implements the MessagesClient interface exported from subagent.ts.
- Strips provider prefixes (`anthropic:`, `litellm:`) from the model id
because `claude --print` only accepts CLI-native aliases (`sonnet`,
`opus`, `haiku`, or the bare `claude-*-N-M` form).
- Flattens the Anthropic messages array into a single text prompt for
claude-cli stdin. Tool blocks (tool_use / tool_result) are stringified
as placeholders so multi-turn conversations stay coherent in this
baseline; native tool_use round-tripping is the follow-up commit.
- Spawns claude with stdio piped, captures stdout, parses the
`{type:"result", subtype:"success", result, usage, ...}` JSON envelope,
and returns it as a properly shaped Anthropic.Message with
`stop_reason: 'end_turn'`.
- Token totals propagate from the claude usage block so the subagent
handler's `ctx.updateTokens()` reports usable numbers.
- AbortSignal is wired through to SIGTERM the child so the subagent loop's
cancellation path stays correct.
Modified: src/commands/jobs.ts (worker registration)
- Conditionally constructs a MessagesClient via the new adapter when
GBRAIN_USE_CLAUDE_CLI=1.
- Passes it into makeSubagentHandler({ engine, client: subagentClient }).
- Logs `[minion worker] subagent routing via claude-cli (GBRAIN_USE_CLAUDE_CLI=1)`
on startup so the env var status is operator-visible.
Limitations of this commit (addressed in the follow-up):
- Tool use is not yet supported. Tools in params.tools are ignored; the
adapter returns a single text block with stop_reason='end_turn'.
- Token counts come from claude-cli's reporting and may not match the
Anthropic API's accounting precisely (especially for cache tiers).
Original design from #334; this commit preserves that author's attribution.
The follow-up commits on this branch carry the tool-use implementation.
* feat(subagent): tool use + context isolation + convention rename on top of jarvisdoes baseline
Builds on the previous commit (jarvisdoes's #334 baseline) by adding three
things the upstream issue called out as gaps or that surfaced during review:
1. Tool use support via system-prompt-instructed JSON emission.
2. Context isolation flags so claude-cli does not load operator-level
CLAUDE.md, skills, and local project context into every subagent call.
3. Env var rename from GBRAIN_USE_CLAUDE_CLI=1 to
GBRAIN_SUBAGENT_PROVIDER=claude-cli to match the existing
GBRAIN_<noun>_<role>=<value> convention used by GBRAIN_CHAT_MODEL,
GBRAIN_EMBEDDING_MODEL, GBRAIN_EXPANSION_MODEL.
## Tool use
The MessagesClient interface returns Anthropic.Message objects whose
content array may include tool_use blocks. The subagent handler filters
those blocks and dispatches each tool, so any backend that produces
correctly shaped tool_use blocks gets the same loop behavior as the
Anthropic SDK.
The adapter injects a system-prompt addendum describing the tool registry
plus an emission protocol:
<use_tools>
[{"id": "...", "name": "...", "input": {...}}, ...]
</use_tools>
After the response comes back, extractToolCalls() scans for the block,
parses the JSON (tolerant of optional ```json fencing), and converts each
entry into a tool_use content block. Multiple parallel tool calls in one
turn are supported via the array shape; this is the exact case that
breaks today on the codex-proxy / litellm GPT-5.x bridge where parallel
tool-call response IDs get dropped.
Defensive fallbacks:
- Malformed JSON inside the block: drop to text-only, stop_reason='end_turn'.
- Unterminated <use_tools> (no close tag): drop to text-only.
- Model omits id field: adapter synthesizes a toolu_claude_cli_<rand> id.
- Empty response: still hand the subagent loop a well-formed content
array so the .filter chain does not crash.
## Context isolation
claude-cli auto-discovers CLAUDE.md from cwd upward and injects the
operator's skills + plugins + auto-memory into the default system prompt.
On a real install that is ~42-65k tokens of contamination per subagent
call, with both cost and behavioral consequences (the subagent picks up
the operator's coding conventions, opinions, and preferences).
The maximum suppression that still preserves OAuth / Claude Max
subscription auth is:
- Spawn from a dedicated clean cwd (tmpdir-based) so LOCAL CLAUDE.md
auto-discovery has nothing to find. -13k tokens on a real gbrain
install where CLAUDE.md is substantial.
- --disable-slash-commands so skill resolution does not pull in
/skill-name handlers.
- --system-prompt <gbrain prompt> so the default system prompt is
replaced rather than appended to.
The --bare flag would also strip user-level ~/.claude/CLAUDE.md but it
forces ANTHROPIC_API_KEY auth, defeating the whole point of this adapter.
The remaining ~42k cached tokens from user-level instructions are
accepted as a cost-trivial trade-off because the Max subscription absorbs
the per-call cost. Behavioral contamination is mitigated by gbrain's
strong per-call system prompt overriding any operator-level drift.
## Env var rename
Surveyed all ~140 GBRAIN_* env vars in src/. The codebase uses three
patterns: GBRAIN_NO_<feature> (negative toggles), GBRAIN_<noun>_<role>
=<value> (routing keys), GBRAIN_ALLOW_<feature> (permissive toggles).
GBRAIN_USE_* does not appear anywhere except jarvisdoes's original
commit; it would introduce a fourth pattern.
GBRAIN_SUBAGENT_PROVIDER=claude-cli aligns with the routing-keys family
and is value-extensible — adding codex-cli / meridian-proxy / etc. later
means a new value, not a new env var. The scope ('SUBAGENT_*') is also
unambiguous about which calls the toggle covers; GBRAIN_USE_CLAUDE_CLI
was silent on whether it applied to all gbrain LLM calls or only the
subagent path.
Unknown values are rejected with a fail-fast error message naming the
two valid values rather than silently falling through to the default.
## Tests
New file: test/claude-cli-adapter.test.ts — 12 tests, 33 assertions:
- Text-only round trip (single text block, usage propagation, end_turn).
- Provider prefix stripping ('anthropic:claude-sonnet-4-6' -> 'claude-sonnet-4-6').
- Single tool_use parsing.
- Multiple parallel tool calls in one block (the case that triggered
the codex-proxy regression).
- Fenced JSON inside <use_tools> block.
- Model-omitted id gets synthesized to toolu_claude_cli_<rand>.
- Malformed JSON falls back to text.
- Unterminated block falls back to text.
- AbortSignal SIGTERMs the child.
- Error envelope rejected with informative message.
- Non-JSON output rejected with raw-output excerpt in the error.
- argv + cwd assertion: --disable-slash-commands + --system-prompt are
present and cwd is the dedicated tmpdir.
Tests use a POSIX shell stub at GBRAIN_CLAUDE_CLI_BIN that emits a
scripted --output-format json envelope, so the suite runs without
claude-cli installed and without API credits.
* feat(ai): claude-cli recipe with native gateway integration (supersedes #334 baseline)
Replaces the MessagesClient adapter + GBRAIN_USE_CLAUDE_CLI=1 env-var
gate from the previous commit on this branch with a proper gateway recipe.
The recipe path gives per-call routing as a native capability: a model
string like `claude-cli:claude-sonnet-4-6` lands here while a sibling
`litellm:gpt-5.4` continues through the litellm-proxy / codex-proxy path
in the same worker. No global env-var switch, no agent.use_gateway_loop
bypass, no MessagesClient injection at jobs.ts worker startup.
The previous commit on this branch (jarvisdoes baseline) is preserved
in the history for #334 authorship attribution. Its functional changes
are backed out here because the recipe pattern is gbrain's established
integration seam; introducing a parallel MessagesClient + env-var path
would have created two routing mechanisms competing for the same job.
New: src/core/ai/recipes/claude-cli.ts
- Recipe declaration: id 'claude-cli', tier 'native', implementation
'claude-cli', chat-only (no embedding or expansion touchpoints).
- Models: claude-opus-4-7, claude-sonnet-4-6, claude-haiku-4-5-20251001.
- supports_tools and supports_subagent_loop both true.
- supports_prompt_cache false because the CLI handles caching internally
and does not surface cache_control via the standard control plane.
- auth_env.required is the empty array because the CLI owns auth (OAuth
session managed by `claude login`).
- Friendly aliases mirror the `anthropic` recipe: `sonnet`, `haiku`,
`opus` and the same legacy-id rewrites for back-compat with stale
config strings.
New: src/core/ai/providers/claude-cli-language-model.ts
- ClaudeCliLanguageModel class implementing the ai-sdk LanguageModelV2
interface.
- doGenerate: renders the ai-sdk prompt array into a system text + user
text, injects the use_tools protocol instructions when tools are
present, spawns `claude --print --output-format json --model <X>
--disable-slash-commands --system-prompt <gbrain prompt>` from a
dedicated tmpdir (contamination suppression: no local CLAUDE.md
auto-discovery), parses the JSON envelope, extracts <use_tools>
blocks, and returns ai-sdk-shaped LanguageModelV2Content (text +
tool-call parts with stringified-JSON input matching the V2 contract).
- Tolerates fenced JSON inside use_tools blocks, malformed JSON
(falls back to text), missing close tag (falls back to text),
model-omitted ids (synthesizes toolu_claude_cli_<rand>).
- Parallel tool calls in one block round-trip cleanly: this is the
case that drops IDs on the litellm + codex-proxy bridge today.
- AbortSignal SIGTERMs the child for proper cancellation.
- doStream throws not-supported (gateway.toolLoop is non-streaming).
Modified: src/core/ai/gateway.ts
- Adds case 'claude-cli' to instantiateChat (returns ClaudeCliLanguageModel).
- Adds case 'claude-cli' to instantiateExpansion (same wrapper, reserved
for a future expansion touchpoint declaration).
- Adds case 'claude-cli' to instantiateEmbedding (throws, no embedding
model, mirrors the native-anthropic path).
- Lazy require() at the call site keeps the gateway module load cheap
for users who never use the claude-cli path.
Modified: src/core/ai/recipes/index.ts
- Registers `claudeCli` in the ALL[] array next to `anthropic`.
Modified: src/core/ai/types.ts
- Adds 'claude-cli' to the Implementation union so the gateway switch
is exhaustive at compile time.
Reverted: src/commands/jobs.ts
- Drops the GBRAIN_USE_CLAUDE_CLI=1 env-var gate the prior commit
added. Routing now happens at the gateway based on the model string.
Deleted: src/core/minions/handlers/claude-cli-adapter.ts
- The MessagesClient adapter is superseded by the recipe + LanguageModelV2
path. Two routing mechanisms competing for the same job would have
forced users to reason about which one wins; the recipe is the single
source of truth.
New file: test/claude-cli-recipe.test.ts (16 tests, 46 assertions):
- Recipe registration: getRecipe returns chat-only Recipe; aliases map
short names (sonnet/haiku/opus) to canonical model ids.
- Text round trip: single text content block, usage propagation, stop
finish reason.
- Provider prefix stripping.
- Single tool-call parsing.
- Multiple parallel tool calls in one block.
- Fenced JSON inside the block.
- Model-omitted id synthesizes toolu_claude_cli_<rand>.
- Malformed JSON falls back to text + stop reason.
- Unterminated block falls back to text + stop reason.
- Tools offered but model declines: returns text-only with stop reason
so the gateway-loop treats it as a final answer rather than wedging
for tool calls that never come.
- AbortSignal SIGTERMs the child.
- is_error envelope rejected.
- Non-JSON output rejected.
- doStream throws.
- argv + cwd assertion: --print, --disable-slash-commands,
--system-prompt are present and cwd is the dedicated tmpdir.
Tests use a POSIX shell stub at GBRAIN_CLAUDE_CLI_BIN so the suite runs
without claude-cli installed and without API credits.
End-to-end smoke verified against a real `claude --print --model haiku`
invocation: model emitted `<use_tools>` block with toolu_add_001 +
{"a":12,"b":30}, adapter parsed back into a `tool-call` content block,
finishReason 'tool-calls'.
* feat(ai/claude-cli): harden subagent isolation, env scrub, verbose + stdin robustness
Four defensive fixes to the claude-cli provider so a subagent call behaves
identically regardless of the host's ambient Claude Code config:
- Agent isolation: pass `--tools ''` and `--strict-mcp-config` so the subprocess
runs as a raw LLM with no built-in tools and no inherited user MCP servers.
Without `--strict-mcp-config`, each call boots the user's MCP servers (including
gbrain's own), causing recursion plus PGLite single-writer lock contention.
- Env scrub: drop ANTHROPIC_API_KEY / ANTHROPIC_AUTH_TOKEN / ANTHROPIC_BASE_URL
from the child env so the CLI authenticates via its own OAuth subscription
session. An inherited API key silently flips billing to per-token API usage,
the exact setup this recipe exists to replace.
- Verbose-mode compat: with `"verbose": true` in ~/.claude/settings.json,
`--print --output-format json` emits an event array instead of a bare result
object. Tolerate both shapes and select the result event.
- stdin robustness: handle the child stdin 'error' event and wrap write/end so a
missing binary (ENOENT) or early child death (EPIPE) rejects cleanly instead of
crashing the worker with an unhandled error.
Adds unit coverage for the env scrub, the isolation argv, and the verbose event
array. Verified against claude CLI 2.1.x.
* test(ai/claude-cli): cover verbose-array no-result + missing-binary reject paths
Two error branches in the hardened claude-cli provider had no coverage: the
verbose event-array path when no result event is present, and a missing binary
surfacing as a clean spawn-failed rejection. The missing-binary case is the
deterministic form of the stdin/EPIPE robustness; a synchronous stdin-write
throw is not reliably triggerable in a unit test, so the real ENOENT path the
handlers defend is exercised instead. Both reuse the existing shell-stub harness.
---------
Co-authored-by: Brett <brettdavies@users.noreply.github.com>
Co-authored-by: jarvisdoes <258486803+jarvisdoes@users.noreply.github.com>
Co-authored-by: Marco Maldonado <34176133+loweaxerium@users.noreply.github.com>
