Compare commits

..
Author SHA1 Message Date
06eeb2890b fix(skillopt): emit proposed.md in no-mutate mode (#2635)
writeProposed now writes both best.md (optimizer's current-best pointer)
and proposed.md (the stable review artifact documented by --no-mutate);
the orchestrator returns the real proposed.md path instead of aliasing
best.md. Fix lands in the shared helper so both accept-branch call sites
are covered.

Takeover of #2719.

Co-authored-by: RerankerGuo <RerankerGuo@users.noreply.github.com>
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-21 14:30:19 -07:00
11 changed files with 57 additions and 270 deletions
@@ -233,13 +233,14 @@ keep it or `git checkout` to throw it away. Nothing is committed for you.
**For a skill that ships with gbrain** (anything under the gbrain repo's own
`skills/`): SkillOpt refuses to overwrite it by default and writes the winner to
`skills/<name>/skillopt/best.md` instead, so an optimization pass can never
silently mutate a skill other people depend on. Two ways to handle that:
`skills/<name>/skillopt/proposed.md` instead (while keeping `best.md` as the
optimizer's current-best pointer), so an optimization pass can never silently
mutate a skill other people depend on. Two ways to handle that:
```bash
# See the proposed improvement without touching SKILL.md (works for ANY skill):
gbrain skillopt meeting-prep --split 1:1:1 --no-mutate
# → writes skills/meeting-prep/skillopt/best.md (the proposed rewrite), prints its path. Copy what you want.
# → writes skills/meeting-prep/skillopt/proposed.md, updates best.md, and prints the proposal path.
# Actually rewrite a bundled skill (explicit opt-in + an independent held-out set):
gbrain skillopt brain-ops --split 1:1:1 --allow-mutate-bundled \
+1 -42
View File
@@ -38,7 +38,6 @@ import { logSelfUpgrade } from '../core/audit/self-upgrade-audit.ts';
import { detectInstallMethod } from './upgrade.ts';
import { evaluateQuietHours } from '../core/minions/quiet-hours.ts';
import { inspectLock } from '../core/db-lock.ts';
import { registerCleanup } from '../core/process-cleanup.ts';
/**
* v0.37.7.0 #1162 — classify autopilot reconnect-loop errors.
@@ -434,37 +433,6 @@ export async function runAutopilot(engine: BrainEngine, args: string[]) {
let stopping = false;
let childSupervisor: ChildWorkerSupervisor | null = null;
// #1872: graceful engine shutdown. On PGLite the cycle steps run INLINE in
// this process, so a hard `process.exit` mid-write (systemctl stop →
// SIGTERM) kills WASM Postgres with the WAL dirty and can corrupt the
// brain. Two exit paths must both close the engine:
// - autopilot's own shutdown() below (owns SIGINT + internal stops like
// max_crashes / cycle-failure-cap), and
// - process-cleanup's SIGTERM handler (installed at cli.ts module load;
// it runs the cleanup registry with a 3s deadline and then exits) —
// which is why closeEngine is ALSO registered there.
// closeEngine aborts the in-flight inline cycle (runCycle checks the
// signal between phases and threads it into phase sub-work), gives it a
// short bounded window to wind down, then disconnects. PGLite's
// disconnect() drains the pending query and checkpoints before closing;
// a second call is a no-op (disconnect snapshots + nulls the handle), so
// both paths firing is safe.
const shutdownAbort = new AbortController();
let inflightInlineCycle: Promise<unknown> | null = null;
const closeEngine = async () => {
shutdownAbort.abort(new Error('autopilot shutdown'));
if (inflightInlineCycle) {
// ponytail: 2s cap keeps us inside process-cleanup's 3s deadline; a
// between-phase abort resolves instantly, a mid-phase one may not.
await Promise.race([
inflightInlineCycle.catch(() => { /* cycle errors already logged by the loop */ }),
new Promise((r) => setTimeout(r, 2_000)),
]);
}
try { await engine.disconnect(); } catch { /* best-effort */ }
};
const deregisterEngineClose = registerCleanup('autopilot-engine-close', closeEngine);
if (spawnManagedWorker) {
const cliPath = resolveGbrainCliPath();
// Cgroup-aware auto-sized RSS watchdog cap (issue #1678). The old flat
@@ -552,10 +520,6 @@ export async function runAutopilot(engine: BrainEngine, args: string[]) {
childSupervisor.killChild('SIGKILL');
}
}
// #1872: abort the in-flight inline cycle and close the engine BEFORE
// process.exit — a hard exit mid-write corrupts PGLite's WASM Postgres.
await closeEngine();
deregisterEngineClose();
try { unlinkSync(lockPath); } catch { /* already gone */ }
process.exit(0);
};
@@ -1044,21 +1008,16 @@ export async function runAutopilot(engine: BrainEngine, args: string[]) {
// path's phase set). Now both converge on the same primitive.
try {
const { runCycle } = await import('../core/cycle.ts');
// #1872: track the promise so closeEngine can drain it on shutdown,
// and pass the abort signal so the cycle winds down between phases.
const cyclePromise = runCycle(engine, {
const report = await runCycle(engine, {
brainDir: repoPath,
// Autopilot daemon path: pulls by default (matches
// pre-v0.17 autopilot behavior). CLI dream defaults false
// for cron safety; that choice is scoped to dream only.
pull: true,
signal: shutdownAbort.signal,
yieldBetweenPhases: async () => {
await new Promise(r => setImmediate(r));
},
});
inflightInlineCycle = cyclePromise;
const report = await cyclePromise.finally(() => { inflightInlineCycle = null; });
// Only 'failed' (every attempted phase failed) trips the autopilot
// circuit breaker. 'partial' means at least one phase warned or
// failed while others ran — that's a soft signal, not a fatal
+3 -23
View File
@@ -26,7 +26,6 @@
import type { BrainEngine } from '../core/engine.ts';
import {
runCycle,
resolveSourceForDir,
ALL_PHASES,
type CyclePhase,
type CycleReport,
@@ -381,9 +380,9 @@ Options:
--source <id> Scope the cycle to one source so doctor's
cycle_freshness check sees a fresh stamp on
completion. When omitted, gbrain derives the
source from --dir / the configured checkout
when it matches a source's local_path (#1869).
completion. Without this, gbrain dream's
timestamp never lands and federated brains
see "stale cycle" forever.
--source-id <id> Alias for --source. Matches the v0.37.7.0+
naming used by import/extract/graph-query.
@@ -635,25 +634,6 @@ export async function runDream(engine: BrainEngine | null, args: string[]): Prom
);
process.exit(1);
}
// #1869: a path-scoped run (--dir, or the configured sync.repo_path) whose
// directory matches a registered source's local_path IS that source's cycle
// — derive the source id so runCycle writes last_source_cycle_at /
// last_full_cycle_at on success and doctor's cycle_freshness check stops
// reading perpetually stale. Explicit --source still wins (resolved above).
// Fixed here at the command level, NOT in runCycle's stamp gate, so legacy
// global callers (autopilot-global-maintenance runs GLOBAL_PHASES with a
// brainDir and no sourceId) can't falsely stamp per-source freshness.
// A derived match on an archived source is skipped silently (falls back to
// legacy unscoped behavior) — stamping it would mask staleness on restore,
// mirroring the explicit --source archived guard above.
if (resolvedSourceId === undefined && engine !== null && brainDir !== null) {
const derived = await resolveSourceForDir(engine, brainDir);
if (derived !== undefined) {
const src = await fetchSource(engine, derived);
if (src?.archived !== true) resolvedSourceId = derived;
}
}
// ─── issue #1678: bounded single-hold extract_atoms drain ──────────
if (opts.drain) {
if (engine === null) {
+1 -9
View File
@@ -855,16 +855,8 @@ interface SyncPhaseResult extends PhaseResult {
* Resolve the source id for a brain directory by looking up the sources
* table. Returns undefined when no registered source matches (falls back
* to pre-v0.18 global config.sync.* keys).
*
* Exported for dream.ts (#1869): a `gbrain dream --dir <path>` run whose
* path matches a registered source's local_path is a per-source cycle in
* everything but name, so dream derives the source id up front and passes
* it as opts.sourceId — landing the freshness stamp without changing
* runCycle's stamp/lock semantics for legacy global callers (the
* autopilot-global-maintenance handler runs GLOBAL_PHASES with a brainDir
* and MUST NOT stamp per-source freshness; see rejected PR #2549).
*/
export async function resolveSourceForDir(
async function resolveSourceForDir(
engine: BrainEngine,
brainDir: string | null,
): Promise<string | undefined> {
+10 -4
View File
@@ -93,7 +93,13 @@ import { resolveLrSchedule } from './lr-schedule.ts';
import { preflight, formatPreflightReport } from './preflight.ts';
import { isRejected, loadRejectedBuffer, makeRejectedEntry, saveRejectedBuffer } from './rejected-buffer.ts';
import { runReflect, runOneShotRewrite, describeJudges } from './reflect.ts';
import { acceptCandidate, bestPath, revertAllPending, skillPath, writeProposed } from './version-store.ts';
import {
acceptCandidate,
proposedPath as proposedFilePath,
revertAllPending,
skillPath,
writeProposed,
} from './version-store.ts';
import { runValidationGate, scoreSkillOnTasks } from './validate-gate.ts';
import { ROLLOUT_SUCCESS_THRESHOLD } from './types.ts';
import type { SkillOptOpts, EditOp, RunReceipt, BenchmarkTask } from './types.ts';
@@ -702,9 +708,9 @@ async function runOptimizationLoop(
// to the catch's assignment values only (it can't prove the async callback ran).
const finalOutcome = outcome as 'accepted' | 'no_improvement' | 'aborted' | 'errored';
if (!mutateDecision.mutate && finalOutcome === 'accepted') {
// best.md was written by writeProposed() in the accept branch (no-mutate
// path); it doubles as proposed.md for human review. SKILL.md untouched.
proposedPath = bestPath(skillsDir, skillName);
// writeProposed() emitted both the best pointer and the stable review
// artifact in the accept branch. SKILL.md remains untouched.
proposedPath = proposedFilePath(skillsDir, skillName);
} else if (mutateDecision.mutate) {
mutatedSkillFile = finalOutcome === 'accepted';
}
+15 -9
View File
@@ -23,6 +23,7 @@
*
* history.json
* best.md
* proposed.md
* versions/
* v0001_e1_s1.md
* v0002_e1_s2.md
@@ -52,6 +53,10 @@ export function bestPath(skillsDir: string, skillName: string): string {
return path.join(skilloptDir(skillsDir, skillName), 'best.md');
}
export function proposedPath(skillsDir: string, skillName: string): string {
return path.join(skilloptDir(skillsDir, skillName), 'proposed.md');
}
export function skillPath(skillsDir: string, skillName: string): string {
return path.join(skillsDir, skillName, 'SKILL.md');
}
@@ -171,17 +176,18 @@ export function acceptCandidate(input: AcceptInput): AcceptResult {
}
/**
* Write the candidate to `best.md` (which doubles as `proposed.md`) WITHOUT
* touching SKILL.md or the history ledger. Used by the `--no-mutate` /
* bundled-without-allow paths: the optimizer found a better candidate but the
* caller opted out of in-place mutation, so we surface it for human review.
* Returns the path written. Atomic (.tmp + rename).
* Write the candidate to both `best.md` and `proposed.md` WITHOUT touching
* SKILL.md or the history ledger. `best.md` remains the optimizer's current
* best pointer; `proposed.md` is the stable human-review artifact promised by
* `--no-mutate`. Returns the proposal path. Each write is atomic (.tmp + rename).
*/
export function writeProposed(skillsDir: string, skillName: string, candidateText: string): string {
const p = bestPath(skillsDir, skillName);
fs.mkdirSync(path.dirname(p), { recursive: true });
atomicWrite(p, candidateText);
return p;
const best = bestPath(skillsDir, skillName);
const proposed = proposedPath(skillsDir, skillName);
fs.mkdirSync(path.dirname(best), { recursive: true });
atomicWrite(best, candidateText);
atomicWrite(proposed, candidateText);
return proposed;
}
/**
@@ -1,63 +0,0 @@
/**
* #1872 — autopilot SIGTERM/SIGINT must close the engine before exit.
*
* On PGLite the cycle steps run INLINE in the autopilot process, so a hard
* `process.exit` mid-write (systemctl stop → SIGTERM) kills WASM Postgres
* with the WAL dirty and can corrupt the brain. Two exit paths must both
* close the engine:
*
* - autopilot's own shutdown() (owns SIGINT + internal stops like
* max_crashes / cycle-failure-cap), and
* - process-cleanup's SIGTERM handler (installed at cli.ts module load,
* which exits within its 3s cleanup deadline) — reached via the
* registered 'autopilot-engine-close' cleanup callback.
*
* Because the shutdown path is deep inside `runAutopilot()` (a long-running
* daemon loop that ends in process.exit), a behavioral test would have to
* spawn + signal a real daemon. Following the established precedent
* (test/autopilot-supervisor-wiring.test.ts, test/autopilot-fanout-wiring.test.ts),
* these static-shape regressions pin the load-bearing wiring instead.
*/
import { describe, expect, it } from 'bun:test';
import { readFileSync } from 'fs';
import { join } from 'path';
const AUTOPILOT_SRC = readFileSync(
join(import.meta.dir, '..', 'src', 'commands', 'autopilot.ts'),
'utf8',
);
describe('autopilot.ts graceful engine shutdown (#1872)', () => {
it('registers an engine-close callback in the process-cleanup registry (SIGTERM path)', () => {
// process-cleanup owns SIGTERM (installed at cli.ts:10) and hard-exits
// after its cleanup pass; without this registration the engine is never
// closed on `systemctl stop`.
expect(AUTOPILOT_SRC).toContain(
"import { registerCleanup } from '../core/process-cleanup.ts';",
);
expect(AUTOPILOT_SRC).toContain(
"registerCleanup('autopilot-engine-close', closeEngine)",
);
});
it('closeEngine aborts the in-flight inline cycle then disconnects the engine', () => {
// Abort first (runCycle checks the signal between phases and threads it
// into phase sub-work), bounded drain, then disconnect.
expect(AUTOPILOT_SRC).toMatch(
/const closeEngine = async \(\) => \{[\s\S]{0,900}shutdownAbort\.abort\([\s\S]{0,900}engine\.disconnect\(\)/,
);
});
it('the inline runCycle call carries the shutdown abort signal and is tracked as in-flight', () => {
// PGLite / --inline path: the cycle runs in-process, so shutdown must be
// able to (a) signal it to wind down and (b) await it before closing.
expect(AUTOPILOT_SRC).toMatch(/signal:\s*shutdownAbort\.signal/);
expect(AUTOPILOT_SRC).toMatch(/inflightInlineCycle\s*=\s*cyclePromise/);
});
it('shutdown() awaits closeEngine() before process.exit(0) (SIGINT + internal-stop path)', () => {
expect(AUTOPILOT_SRC).toMatch(
/await closeEngine\(\);[\s\S]{0,400}process\.exit\(0\)/,
);
});
});
-99
View File
@@ -1,99 +0,0 @@
/**
* #1869 — `gbrain dream --dir <path>` stamps cycle freshness when the path
* matches a registered source's local_path.
*
* Pre-fix, only `--source <id>` runs wrote last_source_cycle_at /
* last_full_cycle_at (runCycle's stamp gate reads opts.sourceId, and dream
* never derived one from --dir), so a path-scoped brain showed doctor's
* cycle_freshness as perpetually stale.
*
* The fix lives in dream.ts (derive the source id from the resolved brain
* dir via resolveSourceForDir), NOT in runCycle's stamp gate — a runCycle-
* wide change would make the autopilot-global-maintenance handler (global
* phases, brainDir set, no sourceId) falsely stamp per-source freshness
* (the #2194 poisoning class; see rejected PR #2549).
*
* Same real-PGLite/no-mocks discipline as test/dream.test.ts; same
* GBRAIN_HOME isolation as test/cycle-last-full-cycle-at.test.ts (the
* cycle's PGLite file lock lives under ~/.gbrain).
*/
import { describe, test, expect, beforeAll, afterAll, beforeEach, afterEach } from 'bun:test';
import { mkdtempSync, rmSync } from 'fs';
import { join } from 'path';
import { tmpdir } from 'os';
import { PGLiteEngine } from '../src/core/pglite-engine.ts';
import { resetPgliteState } from './helpers/reset-pglite.ts';
import { runDream } from '../src/commands/dream.ts';
import { withEnv } from './helpers/with-env.ts';
let engine: PGLiteEngine;
let brainDir: string;
let gbrainHome: string;
beforeAll(async () => {
engine = new PGLiteEngine();
await engine.connect({});
await engine.initSchema();
}, 60_000);
afterAll(async () => {
await engine.disconnect();
});
beforeEach(async () => {
await resetPgliteState(engine);
brainDir = mkdtempSync(join(tmpdir(), 'gbrain-dream-stamp-'));
gbrainHome = mkdtempSync(join(tmpdir(), 'gbrain-dream-stamp-home-'));
}, 60_000);
afterEach(() => {
rmSync(brainDir, { recursive: true, force: true });
rmSync(gbrainHome, { recursive: true, force: true });
});
async function seedSource(id: string, archived = false): Promise<void> {
await engine.executeRaw(
`INSERT INTO sources (id, name, local_path, config, archived, created_at)
VALUES ($1, $2, $3, '{}'::jsonb, $4, NOW())`,
[id, id, brainDir, archived],
);
}
async function readLastFullCycleAt(sourceId: string): Promise<string | null> {
const rows = await engine.executeRaw<{ config: Record<string, unknown> | null }>(
`SELECT config FROM sources WHERE id = $1`,
[sourceId],
);
const raw = rows[0]?.config?.last_full_cycle_at;
return typeof raw === 'string' ? raw : null;
}
describe('gbrain dream --dir <path> freshness stamp (#1869)', () => {
test('--dir matching a source local_path stamps last_full_cycle_at', async () => {
await withEnv({ GBRAIN_HOME: gbrainHome }, async () => {
await seedSource('path-scoped');
expect(await readLastFullCycleAt('path-scoped')).toBeNull();
const report = await runDream(engine, ['--dir', brainDir, '--phase', 'lint', '--json']);
expect(report).toBeTruthy();
if (report) expect(['ok', 'clean']).toContain(report.status);
// Pre-fix this stays null forever: dream never passed a sourceId, so
// runCycle's stamp gate skipped the write.
expect(await readLastFullCycleAt('path-scoped')).not.toBeNull();
});
}, 60_000);
test('--dir matching an ARCHIVED source does not stamp it', async () => {
await withEnv({ GBRAIN_HOME: gbrainHome }, async () => {
await seedSource('mothballed', true);
const report = await runDream(engine, ['--dir', brainDir, '--phase', 'lint', '--json']);
expect(report).toBeTruthy();
// Stamping an archived source would mask data staleness when it is
// later restored (mirrors the explicit --source archived guard).
expect(await readLastFullCycleAt('mothballed')).toBeNull();
});
}, 60_000);
});
+4 -14
View File
@@ -562,22 +562,12 @@ describe('runDream — --source / --source-id (v0.41.13)', () => {
// ─── Back-compat: bare `gbrain dream` does NOT write per-source stamp ─
test('gbrain dream (no --source) stamps only the source whose local_path matches --dir (#1869)', async () => {
// Pre-#1869 this asserted NO source was ever stamped without an explicit
// --source — which is exactly the bug: a path-scoped `gbrain dream --dir`
// run never landed a freshness stamp and doctor's cycle_freshness stayed
// stale forever. New truth: the source whose local_path matches the
// resolved brain dir is derived and stamped; unrelated sources stay
// untouched (cross-source isolation).
await seedSource('alpha'); // local_path = repo → derived + stamped
await engine.executeRaw(
`INSERT INTO sources (id, name, local_path, config, archived, created_at)
VALUES ($1, $2, $3, '{}'::jsonb, false, NOW())`,
['beta', 'beta', '/somewhere/else'],
);
test('gbrain dream (no --source) leaves all sources untouched (back-compat regression)', async () => {
await seedSource('alpha');
await seedSource('beta');
const report = await runDream(engine, ['--dir', repo, '--phase', 'lint', '--json']);
expect(report).toBeTruthy();
expect(await readLastFullCycleAt('alpha')).not.toBeNull();
expect(await readLastFullCycleAt('alpha')).toBeNull();
expect(await readLastFullCycleAt('beta')).toBeNull();
}, 60_000);
+4 -4
View File
@@ -39,6 +39,7 @@ import { runSkillOpt } from '../../src/core/skillopt/orchestrator.ts';
import {
bestPath,
loadHistory,
proposedPath,
skillPath,
} from '../../src/core/skillopt/version-store.ts';
import { loadRejectedBuffer } from '../../src/core/skillopt/rejected-buffer.ts';
@@ -741,7 +742,7 @@ describe('skillopt T3 — F11 held-out gate, ablation opts, no-DB-pollution', ()
} finally { fixture.cleanup(); }
});
test('--no-mutate writes proposed.md (best.md), leaves SKILL.md untouched', async () => {
test('--no-mutate writes proposed.md and best.md, leaves SKILL.md untouched', async () => {
const fixture = setupFixture(SKILL_PEOPLE_ONLY, CITATIONS_BENCHMARK);
try {
installStub({
@@ -753,10 +754,9 @@ describe('skillopt T3 — F11 held-out gate, ablation opts, no-DB-pollution', ()
const result = await runOnce(fixture, { noMutate: true });
expect(result.outcome).toBe('accepted');
expect(result.mutatedSkillFile).toBe(false);
expect(result.proposedPath).toBeDefined();
// proposed.md (best.md) exists and carries the improvement.
expect(fs.existsSync(result.proposedPath!)).toBe(true);
expect(result.proposedPath).toBe(proposedPath(fixture.skillsDir, SKILL));
expect(fs.readFileSync(result.proposedPath!, 'utf8')).toContain('## Citations');
expect(fs.readFileSync(bestPath(fixture.skillsDir, SKILL), 'utf8')).toContain('## Citations');
// SKILL.md on disk is UNCHANGED (still People-only).
const skill = fs.readFileSync(skillPath(fixture.skillsDir, SKILL), 'utf8');
expect(skill).not.toContain('## Citations');
+15
View File
@@ -12,9 +12,11 @@ import {
bestPath,
historyPath,
loadHistory,
proposedPath,
revertAllPending,
skillPath,
versionsDir,
writeProposed,
} from '../../src/core/skillopt/version-store.ts';
let tmpDir: string;
@@ -79,6 +81,19 @@ describe('acceptCandidate (D8 two-phase commit)', () => {
});
});
describe('writeProposed', () => {
test('writes distinct best and proposed artifacts without mutating SKILL.md (#2635)', () => {
const candidate = '---\nname: test\n---\nproposed body\n';
const written = writeProposed(tmpDir, SKILL, candidate);
expect(written).toBe(proposedPath(tmpDir, SKILL));
expect(fs.readFileSync(bestPath(tmpDir, SKILL), 'utf8')).toBe(candidate);
expect(fs.readFileSync(proposedPath(tmpDir, SKILL), 'utf8')).toBe(candidate);
expect(fs.readFileSync(skillPath(tmpDir, SKILL), 'utf8')).toContain('baseline body');
});
});
describe('revertAllPending (D8 crash recovery)', () => {
test('no-op when no pending rows', () => {
const reverted = revertAllPending(tmpDir, SKILL);