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 64 additions and 166 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 \
+12 -42
View File
@@ -161,7 +161,7 @@ interface ResolveAIOptionsArgs {
nonInteractive: boolean; // --non-interactive (forces D3 fail-loud, no picker)
}
export interface ResolvedAIOptions {
interface ResolvedAIOptions {
embedding_model?: string;
embedding_dimensions?: number;
expansion_model?: string;
@@ -170,41 +170,6 @@ export interface ResolvedAIOptions {
noEmbedding?: boolean;
}
/**
* Seed init's AI options from persisted config, falling back to the raw env
* vars when loadConfig() returned null (#1058). On a cold install (no
* config.json AND no DATABASE_URL) loadConfig short-circuits BEFORE its env
* merge, so GBRAIN_EMBEDDING_MODEL / GBRAIN_EMBEDDING_DIMENSIONS /
* GBRAIN_EXPANSION_MODEL / GBRAIN_CHAT_MODEL were silently ignored by init
* and Tier-3 detection auto-picked by API key instead. Exported for unit
* tests (env injectable).
*/
export function seedAIOptionsFromConfig(
cfg: GBrainConfig | null,
env: NodeJS.ProcessEnv = process.env,
): ResolvedAIOptions {
const envDims = env.GBRAIN_EMBEDDING_DIMENSIONS
? parseInt(env.GBRAIN_EMBEDDING_DIMENSIONS, 10)
: NaN;
const seed = cfg ?? {
embedding_disabled: undefined,
embedding_model: env.GBRAIN_EMBEDDING_MODEL,
embedding_dimensions: Number.isFinite(envDims) ? envDims : undefined,
expansion_model: env.GBRAIN_EXPANSION_MODEL,
chat_model: env.GBRAIN_CHAT_MODEL,
};
const out: ResolvedAIOptions = {};
if (seed.embedding_disabled) {
out.noEmbedding = true;
} else if (seed.embedding_model) {
out.embedding_model = seed.embedding_model;
if (seed.embedding_dimensions) out.embedding_dimensions = seed.embedding_dimensions;
}
if (seed.expansion_model) out.expansion_model = seed.expansion_model;
if (seed.chat_model) out.chat_model = seed.chat_model;
return out;
}
/**
* Resolve AI provider options for `gbrain init`.
*
@@ -238,13 +203,18 @@ async function resolveAIOptions(opts: ResolveAIOptionsArgs): Promise<ResolvedAIO
// user already opted into deferred mode.
try {
const { loadConfig } = await import('../core/config.ts');
// #1058: loadConfig() returns null on a cold install (no config.json AND
// no DATABASE_URL) — before it ever reaches its env merge. The seed helper
// falls back to the same GBRAIN_* env vars directly in that case.
Object.assign(out, seedAIOptionsFromConfig(loadConfig()));
const cfg = loadConfig();
if (cfg?.embedding_disabled) {
out.noEmbedding = true;
} else if (cfg?.embedding_model) {
out.embedding_model = cfg.embedding_model;
if (cfg.embedding_dimensions) out.embedding_dimensions = cfg.embedding_dimensions;
}
if (cfg?.expansion_model) out.expansion_model = cfg.expansion_model;
if (cfg?.chat_model) out.chat_model = cfg.chat_model;
} catch {
// loadConfig threw — treat as first-time install, fall through to env
// detection.
// loadConfig throws when no brain configured — first-time install, fall
// through to env detection.
}
// --- Tier 1+2: explicit flags ---------------------------------------------
+3 -19
View File
@@ -332,15 +332,6 @@ export interface OperationContext {
* remote/untrusted (defense in depth in case the type is bypassed via cast).
*/
remote: boolean;
/**
* Transport marker for auth-less remote surfaces (#1061). The stdio MCP
* dispatch sets 'stdio' — it is deliberately `remote: true` (agent-facing,
* untrusted) but has no per-token auth (local pipe), so identity ops like
* whoami need a way to distinguish "known auth-less transport" from "a
* transport bug forgot to thread ctx.auth". Trust decisions MUST NOT key
* off this field — only `ctx.remote === false` grants trust.
*/
transport?: 'stdio';
/**
* Subagent runtime context (v0.16+). Set by the subagent tool dispatcher when
* dispatching an op as a tool call from an LLM loop. Used to enforce per-op
@@ -3722,10 +3713,9 @@ const whoami: Operation = {
'Introspect the calling identity. Returns one of three transport shapes: ' +
'{transport: "oauth", client_id, client_name, scopes, expires_at}, ' +
'{transport: "legacy", token_name, scopes, expires_at: null}, or ' +
'{transport: "local", scopes: []}, or {transport: "stdio", scopes: []} ' +
'for the auth-less stdio MCP pipe. Throws unknown_transport when the ' +
'context is ambiguous (remote=true without auth and no transport marker) ' +
'— fail-closed posture mirroring the v0.26.9 trust-boundary contract.',
'{transport: "local", scopes: []}. Throws unknown_transport when the ' +
'context is ambiguous (remote=true without auth) — fail-closed posture ' +
'mirroring the v0.26.9 trust-boundary contract.',
params: {},
scope: 'read',
handler: async (ctx) => {
@@ -3737,12 +3727,6 @@ const whoami: Operation = {
if (ctx.remote === false) {
return { transport: 'local', scopes: [] };
}
// #1061: stdio MCP is remote/untrusted by design but has no per-token
// auth (local pipe) — a known transport, not a bug. Report it instead of
// throwing. Empty scopes: nothing here may be used to gate anything.
if (!ctx.auth && ctx.transport === 'stdio') {
return { transport: 'stdio', scopes: [] };
}
if (!ctx.auth) {
throw new OperationError(
'unknown_transport',
+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;
}
/**
-7
View File
@@ -32,12 +32,6 @@ export interface DispatchOpts {
remote?: boolean;
/** Override the default stderr logger (e.g. CLI uses console.* directly). */
logger?: OperationContext['logger'];
/**
* #1061: transport marker for auth-less remote surfaces. The stdio MCP
* server passes 'stdio' so identity ops (whoami) can report the transport
* instead of throwing unknown_transport. Never used for trust decisions.
*/
transport?: OperationContext['transport'];
/**
* v0.28: per-token allow-list for the takes.holder field. Threaded by
* the HTTP/stdio transport from `access_tokens.permissions.takes_holders`.
@@ -209,7 +203,6 @@ export function buildOperationContext(
logger: opts.logger || stderrLogger,
dryRun: !!params.dry_run,
remote: opts.remote ?? true,
transport: opts.transport,
takesHoldersAllowList: opts.takesHoldersAllowList,
// v0.34 D4: sourceId is REQUIRED at the type level. Auto-fill 'default'
// for single-source brains and any caller who didn't resolve a sourceId.
-4
View File
@@ -42,10 +42,6 @@ export async function startMcpServer(engine: BrainEngine) {
// `gbrain call <op>` (sets remote=false in src/cli.ts).
return dispatchToolCall(engine, name, params, {
remote: true,
// #1061: mark the transport so whoami can report {transport: 'stdio'}
// instead of throwing unknown_transport. Trust posture unchanged —
// stdio stays remote/untrusted.
transport: 'stdio',
takesHoldersAllowList: ['world'],
// v0.31: source defaults to 'default' for stdio (no per-token scope).
// Operators who want a different source on stdio MCP should set
+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');
+1 -45
View File
@@ -12,7 +12,7 @@
*/
import { describe, test, expect } from 'bun:test';
import { groupReadyByProvider, findEnvKeyTypos, seedAIOptionsFromConfig } from '../src/commands/init.ts';
import { groupReadyByProvider, findEnvKeyTypos } from '../src/commands/init.ts';
describe('groupReadyByProvider — embedding touchpoint', () => {
test('OPENAI_API_KEY alone → openai is ready', async () => {
@@ -149,47 +149,3 @@ describe('findEnvKeyTypos', () => {
expect(got.find(t => t.userSet === 'COMPLETELY_UNRELATED_KEY')).toBeUndefined();
});
});
describe('seedAIOptionsFromConfig — #1058 cold-install env fallback', () => {
test('null config (no config.json, no DATABASE_URL) falls back to GBRAIN_* env vars', () => {
const got = seedAIOptionsFromConfig(null, {
GBRAIN_EMBEDDING_MODEL: 'voyage:voyage-3-large',
GBRAIN_EMBEDDING_DIMENSIONS: '1024',
GBRAIN_EXPANSION_MODEL: 'openai:gpt-5-mini',
GBRAIN_CHAT_MODEL: 'anthropic:claude-sonnet-4-6',
});
expect(got.embedding_model).toBe('voyage:voyage-3-large');
expect(got.embedding_dimensions).toBe(1024);
expect(got.expansion_model).toBe('openai:gpt-5-mini');
expect(got.chat_model).toBe('anthropic:claude-sonnet-4-6');
});
test('null config + no env vars → empty seed (Tier-3 detection takes over)', () => {
const got = seedAIOptionsFromConfig(null, {});
expect(got).toEqual({});
});
test('persisted config wins (loadConfig already merged env when non-null)', () => {
const got = seedAIOptionsFromConfig(
{ engine: 'pglite', embedding_model: 'openai:text-embedding-3-small', embedding_dimensions: 1536 } as any,
{ GBRAIN_EMBEDDING_MODEL: 'voyage:voyage-3-large' },
);
expect(got.embedding_model).toBe('openai:text-embedding-3-small');
expect(got.embedding_dimensions).toBe(1536);
});
test('embedding_disabled sentinel honored on re-init', () => {
const got = seedAIOptionsFromConfig({ engine: 'pglite', embedding_disabled: true } as any, {});
expect(got.noEmbedding).toBe(true);
expect(got.embedding_model).toBeUndefined();
});
test('non-numeric GBRAIN_EMBEDDING_DIMENSIONS ignored, model still seeds', () => {
const got = seedAIOptionsFromConfig(null, {
GBRAIN_EMBEDDING_MODEL: 'voyage:voyage-3-large',
GBRAIN_EMBEDDING_DIMENSIONS: 'not-a-number',
});
expect(got.embedding_model).toBe('voyage:voyage-3-large');
expect(got.embedding_dimensions).toBeUndefined();
});
});
+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);
-29
View File
@@ -94,35 +94,6 @@ describe('whoami op contract', () => {
expect(result.expires_at).toBeNull();
});
// #1061: stdio MCP is remote/untrusted by design but has no per-token auth
// (local pipe). The stdio dispatch marks ctx.transport='stdio'; whoami
// reports it instead of throwing unknown_transport.
test('stdio transport (remote=true, no auth, transport marker) reports stdio', async () => {
const result = (await whoami.handler(
ctxWith({ remote: true, auth: undefined, transport: 'stdio' }),
{},
)) as any;
expect(result.transport).toBe('stdio');
expect(result.scopes).toEqual([]);
});
test('stdio marker does not mask real auth (auth still wins)', async () => {
const result = (await whoami.handler(
ctxWith({
remote: true,
transport: 'stdio',
auth: {
token: 'gbrain_at_xxx',
clientId: 'gbrain_cl_abc',
scopes: ['read'],
expiresAt: 1,
} as AuthInfo,
}),
{},
)) as any;
expect(result.transport).toBe('oauth');
});
// Q3: ambiguous transport — fail-closed. The footgun this guards against
// is a future transport that lands without threading auth, where a buggy
// caller might trust whoami's output to gate sensitive ops.