Compare commits

..
Author SHA1 Message Date
Garry TanandClaude Fable 5 7b8676be3e ci: raise unit-test matrix shard timeout 15 -> 20 min
Shard 4 runs ~14.5 min on master (dream.test.ts at ~29s/test dominates
its wallclock) and hit the 15-min job timeout twice on this PR with
1328 tests passing and 0 failing — the gate then failed on
'gated job did not succeed (got cancelled)'. Real fix is re-mining
scripts/test-weights.json to rebalance the shards; this unblocks CI.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-22 10:59:25 -07:00
SinabinaandClaude Fable 5 b8376f7327 fix(doctor): stop treating ClawHub workspace skills as required gbrain-routable skills (#1767)
ClawHub-installed skills (detected via .clawhub/origin.json) are external
runtime integrations, not gbrain resolver skills. The derived manifest now
skips them, so doctor resolver_health no longer hard-fails with
unreachable/mece_gap on e.g. an email integration skill. A ClawHub skill
opts back into strict checking by declaring triggers: in its SKILL.md
frontmatter (the same surface that makes it routable); an explicit
manifest.json listing also keeps strict checks. Also keeps dry-fix from
rewriting externally-managed skill files.

Fixes #1767

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-21 14:27:57 -07:00
15 changed files with 131 additions and 166 deletions
+5 -1
View File
@@ -206,7 +206,11 @@ jobs:
needs: cache-check
if: needs.cache-check.outputs.hit != 'true'
runs-on: ubuntu-latest
timeout-minutes: 15
# 20 (was 15): shard 4 runs ~14.5 min on master (dream.test.ts ~29s/test
# dominates it) and hits the 15-min ceiling on slower runners, cancelling
# mid-run with 0 test failures. Rebalancing via
# scripts/mine-shard-weights.ts is the real fix; this stops the bleeding.
timeout-minutes: 20
strategy:
fail-fast: false
matrix:
-4
View File
@@ -908,10 +908,6 @@ export const KNOWN_CONFIG_KEYS: readonly string[] = [
'facts.extraction_model',
// #2113: output-token cap for the per-turn facts extractor (default 4000).
'facts.extraction_max_tokens',
// Owner opt-in: let the local stdio MCP pipe read this owner's private facts
// (find_trajectory / recall). Default off; HTTP transport ignores it. See
// src/core/facts/reader-trust.ts.
'facts.trust_local_reads',
// Dream cycle config
'dream.synthesize.session_corpus_dir',
'dream.synthesize.meeting_transcripts_dir',
-2
View File
@@ -591,8 +591,6 @@ export interface TrajectoryOpts {
sourceIds?: string[];
/** When true, filters to visibility='world' only. Set by MCP layer from ctx.remote. */
remote?: boolean;
/** Owner opt-in: read private facts despite `remote`. See facts/reader-trust.ts. */
trustedFactReads?: boolean;
/** Metric filter. When set, only facts with this canonical metric label participate. */
metric?: string;
/**
+5 -8
View File
@@ -17,7 +17,6 @@
import type { OperationContext } from './../operations.ts';
import type { FactRow } from './../engine.ts';
import { effectiveConfidence } from './decay.ts';
import { readableFactVisibilities } from './reader-trust.ts';
const DEFAULT_TTL_MS = 30_000;
const DEFAULT_TOP_K = 10;
@@ -51,13 +50,7 @@ export async function getBrainHotMemoryMeta(
const sessionId = (ctx as { source_session?: string }).source_session
?? null;
const allowListHash = hashAllowList(ctx.takesHoldersAllowList);
// Visibility tier: untrusted remote → world-only; trusted local +
// owner-trusted reads → all rows. Folded into the cache key (the header's
// "cache entries don't bleed across tiers" invariant): trustedFactReads is
// re-read from config per call, so a mid-session opt-out must not keep
// serving a private-inclusive cached payload for the TTL window.
const visibility = readableFactVisibilities(ctx);
const cacheKey = `${sourceId}::${sessionId ?? '_'}::${allowListHash}::${visibility ? 'world' : 'all'}`;
const cacheKey = `${sourceId}::${sessionId ?? '_'}::${allowListHash}`;
const ttl = Math.max(1000, opts.ttlMs ?? DEFAULT_TTL_MS);
const topK = Math.max(1, Math.min(opts.topK ?? DEFAULT_TOP_K, 25));
@@ -68,6 +61,10 @@ export async function getBrainHotMemoryMeta(
return cached.payload;
}
// Build a fresh payload. Visibility tier: remote → world-only;
// local → all rows.
const visibility = ctx.remote === false ? undefined : ['world'] as ('world' | 'private')[];
let rows: FactRow[] = [];
if (sessionId) {
rows = await ctx.engine.listFactsBySession(sourceId, sessionId, {
-48
View File
@@ -1,48 +0,0 @@
/**
* Fact-read visibility trust.
*
* Fact rows are tagged `private` | `world`. Remote/untrusted callers
* (`remote === true`) see only `world` rows — the posture that keeps a
* published or HTTP-served brain from leaking private claims to strangers.
*
* But the stdio MCP server is an unauthenticated LOCAL pipe: on a single-owner
* machine the caller IS the owner, yet it still defaults `remote: true` for
* safety, so the owner's own agent is denied the owner's own private facts
* (e.g. `find_trajectory` returns empty over MCP even though the facts exist).
*
* `trustedFactReads` is a narrow, READ-ONLY trust elevation, deliberately
* DECOUPLED from `remote` so every other remote protection — file_upload
* confinement, source isolation, fence stripping, takes-holder scoping — stays
* fully in force. The stdio MCP server sets it ONLY when the brain owner opts
* in via the `facts.trust_local_reads` config (default off). The HTTP/published
* transport never sets it, so a served brain stays world-only regardless.
*/
export interface FactReaderTrust {
/**
* Mirrors OperationContext.remote. FAIL-CLOSED: anything not strictly
* `false` is treated as remote/untrusted (CLAUDE.md trust invariant).
*/
remote?: boolean;
/** Owner opt-in: this remote caller may read private facts. */
trustedFactReads?: boolean;
}
/**
* True when the reader is restricted to `visibility = 'world'` rows.
* Fail-closed: an unset/undefined `remote` is untrusted — only an explicit
* `remote: false` (trusted local CLI) or an explicit owner opt-in
* (`trustedFactReads: true`) reads private rows.
*/
export function factsWorldOnly(t: FactReaderTrust): boolean {
return t.remote !== false && t.trustedFactReads !== true;
}
/**
* Visibility filter for list-style fact reads: `['world']` when the reader is
* world-only, `undefined` (no filter — all rows) when it is trusted.
*/
export function readableFactVisibilities(
t: FactReaderTrust,
): ('private' | 'world')[] | undefined {
return factsWorldOnly(t) ? ['world'] : undefined;
}
+7 -14
View File
@@ -18,7 +18,6 @@ import { captureEvalCandidate, isEvalCaptureEnabled, isEvalScrubEnabled } from '
import type { HybridSearchMeta } from './types.ts';
import { extractPageLinks, isAutoLinkEnabled, isAutoTimelineEnabled, isGlobalBasenameEnabled, parseTimelineEntries, makeResolver, type UnresolvedFrontmatterRef } from './link-extraction.ts';
import { isFactsBackstopEligible } from './facts/eligibility.ts';
import { readableFactVisibilities } from './facts/reader-trust.ts';
import { stripTakesFence } from './takes-fence.ts';
import { stripFactsFence } from './facts-fence.ts';
import { getContentFlag } from './quarantine.ts';
@@ -333,15 +332,6 @@ export interface OperationContext {
* remote/untrusted (defense in depth in case the type is bypassed via cast).
*/
remote: boolean;
/**
* Owner opt-in (`facts.trust_local_reads`): allow this remote caller to read
* `private` facts. A NARROW, read-only trust elevation decoupled from
* `remote` — every other remote protection (file confinement, source
* isolation, fence stripping, takes scoping) stays in force. Set ONLY by the
* stdio MCP server when the config is on; the HTTP transport never sets it.
* Consulted via `src/core/facts/reader-trust.ts`.
*/
trustedFactReads?: boolean;
/**
* 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
@@ -3639,7 +3629,6 @@ const find_trajectory: Operation = {
entitySlug: p.entity_slug,
...scope,
remote: ctx.remote === true,
trustedFactReads: ctx.trustedFactReads === true,
metric,
kind,
since,
@@ -4000,9 +3989,13 @@ const recall: Operation = {
const includeExpired = p.include_expired === true;
const grep = typeof p.grep === 'string' ? p.grep.toLowerCase() : null;
// Visibility filter: world-only for untrusted remote callers; all rows for
// trusted local CLI and owner-trusted reads (facts.trust_local_reads).
const visibility = readableFactVisibilities(ctx);
// Visibility filter: remote callers see world-only unless their token
// grants elevated visibility (future-proofing; v0.31 ships world-only
// for remote, all for local CLI).
const visibility =
ctx.remote === false
? undefined
: ['world'] as ('private' | 'world')[];
let rows: Awaited<ReturnType<typeof ctx.engine.listFactsByEntity>> = [];
+1 -5
View File
@@ -18,7 +18,6 @@ import type {
} from './engine.ts';
import { MAX_SEARCH_LIMIT, clampSearchLimit } from './engine.ts';
import { withRetry, BULK_RETRY_OPTS, resolveBulkRetryOpts, computeNextDelay, type BatchAuditSite } from './retry.ts';
import { factsWorldOnly } from './facts/reader-trust.ts';
import { logBatchRetry as auditLogBatchRetry, logBatchExhausted as auditLogBatchExhausted } from './audit/batch-retry-audit.ts';
import { runMigrations } from './migrate.ts';
import { PGLITE_SCHEMA_SQL, getPGLiteSchema } from './pglite-schema.ts';
@@ -4319,10 +4318,7 @@ export class PGLiteEngine implements BrainEngine {
const useArray = Array.isArray(opts.sourceIds) && opts.sourceIds.length > 0;
const sourceIds = useArray ? opts.sourceIds! : null;
const sourceId = opts.sourceId ?? 'default';
// Direct-engine contract: unset `remote` here means a trusted in-process
// caller (CLI/tests) — the fail-closed default lives in the op layer, which
// always passes explicit booleans. Normalize before the fail-closed helper.
const remoteFilter = factsWorldOnly({ remote: opts.remote === true, trustedFactReads: opts.trustedFactReads === true });
const remoteFilter = opts.remote === true;
// Build SQL dynamically. PGLite uses $N positional params; we
// assemble the WHERE clauses + params array in tandem to keep them
+1 -5
View File
@@ -14,7 +14,6 @@ import type {
SourceRow,
} from './engine.ts';
import { withRetry, BULK_RETRY_OPTS, resolveBulkRetryOpts, computeNextDelay, type BatchAuditSite } from './retry.ts';
import { factsWorldOnly } from './facts/reader-trust.ts';
import { logBatchRetry as auditLogBatchRetry, logBatchExhausted as auditLogBatchExhausted } from './audit/batch-retry-audit.ts';
import type {
DomainBankSampleOpts, CorpusSampleOpts, DomainBankRow,
@@ -4533,10 +4532,7 @@ export class PostgresEngine implements BrainEngine {
const useArray = Array.isArray(opts.sourceIds) && opts.sourceIds.length > 0;
const sourceIds = useArray ? opts.sourceIds! : null;
const sourceId = opts.sourceId ?? 'default';
// Direct-engine contract: unset `remote` here means a trusted in-process
// caller (CLI/tests) — the fail-closed default lives in the op layer, which
// always passes explicit booleans. Normalize before the fail-closed helper.
const remoteFilter = factsWorldOnly({ remote: opts.remote === true, trustedFactReads: opts.trustedFactReads === true });
const remoteFilter = opts.remote === true;
// Source-scope predicate: array path (federated) wins over scalar.
// Engine.ts contract: returns chronological points; regressions +
+34 -1
View File
@@ -23,6 +23,15 @@
* hold conventions and shared rule files, not skills. Files like
* `_brain-filing-rules.md` live at the root and are not considered
* skills by either loader.
*
* ClawHub-installed workspace skills (#1767): a skill dir carrying
* `.clawhub/origin.json` is an externally-managed runtime integration
* (e.g. an email or catalog skill), not a gbrain-routable skill. The
* derive path SKIPS those so `gbrain doctor` resolver_health doesn't
* hard-fail on them — UNLESS the skill's SKILL.md frontmatter declares
* `triggers:`, which is the explicit opt-in to gbrain routing (and the
* same surface that makes it reachable). An explicit manifest.json that
* lists a ClawHub skill also keeps strict checking (verbatim path).
*/
import { existsSync, readFileSync, readdirSync, statSync } from 'fs';
@@ -60,9 +69,27 @@ function parseSkillName(skillMdPath: string): string | null {
}
}
/**
* Does the SKILL.md frontmatter declare a `triggers:` key? A ClawHub-
* installed skill that ships gbrain `triggers:` has explicitly opted in
* to gbrain routing and gets full resolver checks (#1767).
*/
function declaresTriggers(skillMdPath: string): boolean {
try {
const content = readFileSync(skillMdPath, 'utf-8');
const fmMatch = content.match(/^---\n([\s\S]*?)\n---/);
if (!fmMatch) return false;
return /^triggers:/m.test(fmMatch[1]);
} catch {
return false;
}
}
/**
* Walk skillsDir, return every `<skillsDir>/<dir>/SKILL.md` as a
* ManifestEntry. Dotfile and underscore-prefixed dirs are skipped.
* ManifestEntry. Dotfile and underscore-prefixed dirs are skipped, as
* are ClawHub-installed external skills that haven't opted in to gbrain
* routing via `triggers:` frontmatter (#1767).
*/
function deriveManifest(skillsDir: string): ManifestEntry[] {
const out: ManifestEntry[] = [];
@@ -93,6 +120,12 @@ function deriveManifest(skillsDir: string): ManifestEntry[] {
const skillMd = join(subdirAbs, 'SKILL.md');
if (!existsSync(skillMd)) continue;
// ClawHub-installed external skill (#1767): skip unless it opts in
// to gbrain routing by declaring `triggers:` in its frontmatter.
if (existsSync(join(subdirAbs, '.clawhub', 'origin.json')) && !declaresTriggers(skillMd)) {
continue;
}
const frontmatterName = parseSkillName(skillMd);
const name = frontmatterName && frontmatterName !== '' ? frontmatterName : entry;
out.push({ name, path: `${entry}/SKILL.md` });
-8
View File
@@ -30,13 +30,6 @@ export interface ToolResult {
export interface DispatchOpts {
/** Defaults to true (remote/untrusted). Local CLI callers (`gbrain call`) pass false. */
remote?: boolean;
/**
* Owner opt-in (`facts.trust_local_reads`): let this remote caller read
* `private` facts. Set ONLY by the stdio MCP server; the HTTP transport
* leaves it unset so a served brain stays world-only. See
* `src/core/facts/reader-trust.ts`.
*/
trustedFactReads?: boolean;
/** Override the default stderr logger (e.g. CLI uses console.* directly). */
logger?: OperationContext['logger'];
/**
@@ -210,7 +203,6 @@ export function buildOperationContext(
logger: opts.logger || stderrLogger,
dryRun: !!params.dry_run,
remote: opts.remote ?? true,
trustedFactReads: opts.trustedFactReads === true,
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.
-11
View File
@@ -35,16 +35,6 @@ export async function startMcpServer(engine: BrainEngine) {
// shape and cast through `any` (the SDK accepts it via the ServerResult union).
server.setRequestHandler(CallToolRequestSchema, async (request: any): Promise<any> => {
const { name, arguments: params } = request.params;
// Owner opt-in: the stdio pipe is local + unauthenticated, so on a
// single-owner machine its caller is the owner. When facts.trust_local_reads
// is on, let fact reads (find_trajectory / recall) see this owner's own
// private facts. Narrow + read-only — every other remote protection stays
// on (remote stays true). HTTP transport never sets this. Best-effort: a
// config read blip falls back to the safe world-only default.
let trustedFactReads = false;
try {
trustedFactReads = (await engine.getConfig('facts.trust_local_reads')) === 'true';
} catch { /* keep world-only default */ }
// v0.28: stdio MCP has no per-token auth (local pipe). Default the
// takes-holder allow-list to ['world'] so agent-facing callers don't
// see private hunches via takes_list / takes_search / query. Operators
@@ -52,7 +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,
trustedFactReads,
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
+29
View File
@@ -382,6 +382,35 @@ describe("DRY detection — checkResolvable", () => {
});
});
describe("#1767 — ClawHub workspace skills are not resolver-required", () => {
let dir: string;
afterEachCleanup(() => dir && rmSync(dir, { recursive: true, force: true }));
test("ClawHub skill without gbrain metadata produces no unreachable/mece_gap", () => {
dir = mkdtempSync(join(tmpdir(), "gbrain-clawhub-"));
// Native gbrain skill: routable via frontmatter triggers. No manifest.json
// (the OpenClaw derive path from the issue repro).
mkdirSync(join(dir, "query"), { recursive: true });
writeFileSync(
join(dir, "query", "SKILL.md"),
`---\nname: query\ndescription: test\ntriggers:\n - "what do we know"\n---\n\n# query\n`
);
// ClawHub-installed integration: no triggers, no resolver row.
mkdirSync(join(dir, "agentmail", ".clawhub"), { recursive: true });
writeFileSync(
join(dir, "agentmail", ".clawhub", "origin.json"),
JSON.stringify({ registry: "https://clawhub.ai", slug: "agentmail" })
);
writeFileSync(join(dir, "agentmail", "SKILL.md"), `---\nname: agentmail\ndescription: email integration\n---\n\n# agentmail\n`);
const report = checkResolvable(dir);
const agentmailIssues = report.issues.filter(i => i.skill === "agentmail");
expect(agentmailIssues).toEqual([]);
expect(report.ok).toBe(true);
expect(report.summary.total_skills).toBe(1);
});
});
describe("v0.22.4 regression — actual repo skills/ has 0 errors", () => {
test("repo skills/ pass check-resolvable cleanly (zero errors AND zero warnings)", () => {
// The v0.22.4 (Part A) contract was zero warnings AND zero errors.
-16
View File
@@ -165,22 +165,6 @@ describe('findTrajectory — visibility filter (D-CDX-1 / R6)', () => {
const all = await engine.findTrajectory({ entitySlug: 'traj-vis-default' });
expect(all.length).toBe(2);
});
test('remote=true + trustedFactReads bypasses world-only (owner-trusted reads)', async () => {
await insertTyped({ entity_slug: 'traj-vis-trusted', metric: 'mrr', value: 50000, visibility: 'private', valid_from: new Date('2026-01-15') });
await insertTyped({ entity_slug: 'traj-vis-trusted', metric: 'mrr', value: 99999, visibility: 'world', valid_from: new Date('2026-04-12') });
// Untrusted remote: world only.
const untrusted = await engine.findTrajectory({ entitySlug: 'traj-vis-trusted', remote: true });
expect(untrusted.length).toBe(1);
expect(untrusted[0].value).toBe(99999);
// Owner-trusted remote: sees the private point too. remote stays true — only
// fact-read visibility is elevated.
const trusted = await engine.findTrajectory({ entitySlug: 'traj-vis-trusted', remote: true, trustedFactReads: true });
expect(trusted.length).toBe(2);
expect(trusted.map(p => p.value).sort((a, b) => (a! - b!))).toEqual([50000, 99999]);
});
});
describe('findTrajectory — metric + since + until filters', () => {
-43
View File
@@ -1,43 +0,0 @@
import { describe, test, expect } from 'bun:test';
import {
factsWorldOnly,
readableFactVisibilities,
} from '../src/core/facts/reader-trust.ts';
describe('factsWorldOnly', () => {
test('FAIL-CLOSED: unset remote is untrusted (world-only)', () => {
expect(factsWorldOnly({})).toBe(true);
expect(factsWorldOnly({ remote: undefined })).toBe(true);
expect(factsWorldOnly({ trustedFactReads: false })).toBe(true);
});
test('explicit remote=false (trusted local CLI) sees all', () => {
expect(factsWorldOnly({ remote: false })).toBe(false);
expect(factsWorldOnly({ remote: false, trustedFactReads: false })).toBe(false);
});
test('untrusted remote callers are world-only', () => {
expect(factsWorldOnly({ remote: true })).toBe(true);
expect(factsWorldOnly({ remote: true, trustedFactReads: false })).toBe(true);
});
test('owner-trusted remote reads bypass the world-only filter', () => {
expect(factsWorldOnly({ remote: true, trustedFactReads: true })).toBe(false);
});
test('trustedFactReads is a no-op for an already-trusted local caller', () => {
expect(factsWorldOnly({ remote: false, trustedFactReads: true })).toBe(false);
});
});
describe('readableFactVisibilities', () => {
test("world-only readers get the ['world'] filter (incl. unset remote)", () => {
expect(readableFactVisibilities({ remote: true })).toEqual(['world']);
expect(readableFactVisibilities({})).toEqual(['world']);
});
test('trusted readers get undefined (no filter — all rows)', () => {
expect(readableFactVisibilities({ remote: false })).toBeUndefined();
expect(readableFactVisibilities({ remote: true, trustedFactReads: true })).toBeUndefined();
});
});
+49
View File
@@ -166,6 +166,55 @@ describe('loadOrDeriveManifest', () => {
expect(r.skills.map(s => s.name)).toEqual(['apple', 'mango', 'zebra']);
});
// #1767 — ClawHub-installed workspace skills are external integrations,
// not gbrain-routable skills. The derive path skips them unless they
// opt in via `triggers:` frontmatter.
it('skips ClawHub-origin skills without triggers frontmatter (#1767)', () => {
const dir = scratch();
writeSkill(dir, 'query', 'query');
writeSkill(dir, 'agentmail', 'agentmail');
mkdirSync(join(dir, 'agentmail', '.clawhub'), { recursive: true });
writeFileSync(
join(dir, 'agentmail', '.clawhub', 'origin.json'),
JSON.stringify({ registry: 'https://clawhub.ai', slug: 'agentmail' })
);
const r = loadOrDeriveManifest(dir);
expect(r.derived).toBe(true);
expect(r.skills.map(s => s.name)).toEqual(['query']);
});
it('includes ClawHub-origin skills that opt in via triggers frontmatter (#1767)', () => {
const dir = scratch();
writeSkill(dir, 'agentmail', 'agentmail');
mkdirSync(join(dir, 'agentmail', '.clawhub'), { recursive: true });
writeFileSync(
join(dir, 'agentmail', '.clawhub', 'origin.json'),
JSON.stringify({ registry: 'https://clawhub.ai', slug: 'agentmail' })
);
writeFileSync(
join(dir, 'agentmail', 'SKILL.md'),
`---\nname: agentmail\ndescription: test\ntriggers:\n - "send email"\n---\n\n# agentmail\n`
);
const r = loadOrDeriveManifest(dir);
expect(r.derived).toBe(true);
expect(r.skills.map(s => s.name)).toEqual(['agentmail']);
});
it('keeps ClawHub-origin skills listed in an explicit manifest.json (#1767)', () => {
// Explicit manifest.json is a deliberate declaration — strict checking stays.
const dir = scratch();
writeSkill(dir, 'agentmail', 'agentmail');
mkdirSync(join(dir, 'agentmail', '.clawhub'), { recursive: true });
writeFileSync(
join(dir, 'agentmail', '.clawhub', 'origin.json'),
JSON.stringify({ registry: 'https://clawhub.ai', slug: 'agentmail' })
);
writeManifest(dir, { skills: [{ name: 'agentmail', path: 'agentmail/SKILL.md' }] });
const r = loadOrDeriveManifest(dir);
expect(r.derived).toBe(false);
expect(r.skills.map(s => s.name)).toEqual(['agentmail']);
});
it('treats dirs without SKILL.md as not-a-skill', () => {
const dir = scratch();
writeSkill(dir, 'query', 'query');