Compare commits

..
Author SHA1 Message Date
Garry TanandClaude Fable 5 76589d3ddc fix(facts): fold visibility tier into hot-memory meta cache key
trustedFactReads is re-read from config on every stdio tool call, so the
reader's visibility tier can now flip within one server process. The
meta-hook cache key (source, session, allow-list hash) didn't capture it,
so a mid-session opt-out of facts.trust_local_reads could keep serving a
private-inclusive cached payload for the 30s TTL. Compute the visibility
filter before the cache lookup and append the tier to the key, restoring
the header's cross-tier no-bleed invariant.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-22 11:10:57 -07:00
d6b0253c59 feat(facts): owner-trusted local fact reads via facts.trust_local_reads (takeover of #2427)
Salvages the owner-trusted-reads half of #2427 with the trust regression
repaired. (The other half of #2427 — dating conversation facts at claim
time — already landed on master via #2958 and is dropped here.)

Fact rows are private|world; remote callers are world-only. The stdio MCP
server is a local unauthenticated pipe on a single-owner machine, yet it
defaults remote:true — so the owner's own agent got empty find_trajectory
and world-only recall over MCP. This adds `trustedFactReads`, a narrow
READ-ONLY trust elevation decoupled from `remote` (file confinement,
source isolation, fence stripping, takes scoping all stay in force),
opted in via the new `facts.trust_local_reads` config (default off).
Only the stdio server sets it; HTTP dispatch never does.

Repair vs #2427 as submitted: its `factsWorldOnly` was fail-OPEN
(`remote === true && …`) — an unset `remote` read private rows — and it
replaced two fail-closed sites (recall's visibility filter and the facts
meta-hook), with a test codifying the regression. Now fail-closed
(`remote !== false && trustedFactReads !== true`) per the CLAUDE.md trust
invariant; the truth-table test pins it. The two engine findTrajectory
sites keep their direct-engine default-local contract by normalizing to
explicit booleans before the helper (both engines in lockstep).

Tests: test/facts-reader-trust.test.ts (fail-closed truth table);
test/engine-find-trajectory.test.ts new case (remote=true +
trustedFactReads returns private points; plain remote=true stays
world-only).

Takeover of #2427.

Co-authored-by: jeanpierre121 <jeanpierre121@users.noreply.github.com>
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-21 14:34:48 -07:00
13 changed files with 166 additions and 104 deletions
+2 -50
View File
@@ -3,11 +3,7 @@
*
* Usage:
* gbrain migrate --to supabase [--url <connection_string>]
* (--url is persisted to config.json, mode 0600, so the migrated brain
* works without env — #1271)
* gbrain migrate --to pglite [--path <db_path>]
* (an explicit --path destination is bootstrapped with its own
* <path>/.gbrain/config.json so GBRAIN_HOME=<path> just works — #1271)
* gbrain migrate --to <engine> --force (overwrite non-empty target)
*/
@@ -15,9 +11,9 @@ import { createEngine } from '../core/engine-factory.ts';
import { loadConfig, saveConfig, toEngineConfig, gbrainPath, effectiveEnvDatabaseUrl, type GBrainConfig } from '../core/config.ts';
import type { BrainEngine } from '../core/engine.ts';
import type { EngineConfig } from '../core/types.ts';
import { writeFileSync, readFileSync, existsSync, unlinkSync, mkdirSync, chmodSync } from 'fs';
import { writeFileSync, readFileSync, existsSync, unlinkSync } from 'fs';
import { createHash } from 'crypto';
import { resolve, join } from 'path';
import { resolve } from 'path';
import { createProgress } from '../core/progress.ts';
import { getCliOptions, cliOptsToProgressOptions } from '../core/cli-options.ts';
@@ -63,31 +59,6 @@ export interface MigrateManifest {
started_at: string;
}
/**
* #1271 Finding 1: make an explicit `--to pglite --path P` destination usable
* as a standalone brain. Writes `P/.gbrain/config.json` (mode 0600, plus a
* `*` .gitignore) so `GBRAIN_HOME=P` resolves without a manual `gbrain init`.
* Never clobbers an existing config at the destination. Returns the written
* config path, or null when skipped.
*/
export function bootstrapDestinationConfig(dbPath: string): string | null {
const abs = resolve(dbPath);
const dir = join(abs, '.gbrain');
const file = join(dir, 'config.json');
if (existsSync(file)) return null;
mkdirSync(dir, { recursive: true });
const cfg: GBrainConfig = { engine: 'pglite', database_path: abs };
writeFileSync(file, JSON.stringify(cfg, null, 2) + '\n', { mode: 0o600 });
try { chmodSync(file, 0o600); } catch { /* platform-specific */ }
// Same worktree-safety pattern as saveConfig()'s ensureGitignore, scoped
// to the destination home. Don't clobber a user-customized .gitignore.
const gitignore = join(dir, '.gitignore');
if (!existsSync(gitignore)) {
writeFileSync(gitignore, '*\n', { mode: 0o600 });
}
return file;
}
export function migrationTargetId(config: EngineConfig): string {
const locator = config.engine === 'postgres'
? config.database_url ?? ''
@@ -381,25 +352,6 @@ export async function runMigrateEngine(sourceEngine: BrainEngine, args: string[]
};
saveConfig(newConfig);
// #1271 Finding 2 (by design, but say it out loud): the connection string
// is persisted so the migrated brain works without env. Mode 0600.
if (opts.targetEngine === 'postgres' && opts.targetUrl) {
console.error('Note: the --url connection string (including credentials) is persisted to config.json (mode 0600).');
}
// #1271 Finding 1: an explicit --path destination doubles as a standalone
// GBRAIN_HOME. Best-effort — never fail a completed migration over it.
if (opts.targetEngine === 'pglite' && opts.targetPath) {
try {
const written = bootstrapDestinationConfig(opts.targetPath);
if (written) {
console.log(`Destination bootstrapped: ${written} (usable via GBRAIN_HOME=${resolve(opts.targetPath)})`);
}
} catch (e) {
console.warn(` WARN could not bootstrap destination config: ${e instanceof Error ? e.message : String(e)}`);
}
}
// Clean up
clearManifest();
+4
View File
@@ -908,6 +908,10 @@ export const KNOWN_CONFIG_KEYS: readonly string[] = [
'facts.extraction_model',
// #2113: output-token cap for the per-turn facts extractor (default 4000).
'facts.extraction_max_tokens',
// Owner opt-in: let the local stdio MCP pipe read this owner's private facts
// (find_trajectory / recall). Default off; HTTP transport ignores it. See
// src/core/facts/reader-trust.ts.
'facts.trust_local_reads',
// Dream cycle config
'dream.synthesize.session_corpus_dir',
'dream.synthesize.meeting_transcripts_dir',
+2
View File
@@ -591,6 +591,8 @@ export interface TrajectoryOpts {
sourceIds?: string[];
/** When true, filters to visibility='world' only. Set by MCP layer from ctx.remote. */
remote?: boolean;
/** Owner opt-in: read private facts despite `remote`. See facts/reader-trust.ts. */
trustedFactReads?: boolean;
/** Metric filter. When set, only facts with this canonical metric label participate. */
metric?: string;
/**
+8 -5
View File
@@ -17,6 +17,7 @@
import type { OperationContext } from './../operations.ts';
import type { FactRow } from './../engine.ts';
import { effectiveConfidence } from './decay.ts';
import { readableFactVisibilities } from './reader-trust.ts';
const DEFAULT_TTL_MS = 30_000;
const DEFAULT_TOP_K = 10;
@@ -50,7 +51,13 @@ export async function getBrainHotMemoryMeta(
const sessionId = (ctx as { source_session?: string }).source_session
?? null;
const allowListHash = hashAllowList(ctx.takesHoldersAllowList);
const cacheKey = `${sourceId}::${sessionId ?? '_'}::${allowListHash}`;
// 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 ttl = Math.max(1000, opts.ttlMs ?? DEFAULT_TTL_MS);
const topK = Math.max(1, Math.min(opts.topK ?? DEFAULT_TOP_K, 25));
@@ -61,10 +68,6 @@ 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
@@ -0,0 +1,48 @@
/**
* 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;
}
+14 -7
View File
@@ -18,6 +18,7 @@ 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';
@@ -332,6 +333,15 @@ 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
@@ -3629,6 +3639,7 @@ const find_trajectory: Operation = {
entitySlug: p.entity_slug,
...scope,
remote: ctx.remote === true,
trustedFactReads: ctx.trustedFactReads === true,
metric,
kind,
since,
@@ -3989,13 +4000,9 @@ const recall: Operation = {
const includeExpired = p.include_expired === true;
const grep = typeof p.grep === 'string' ? p.grep.toLowerCase() : null;
// 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')[];
// 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);
let rows: Awaited<ReturnType<typeof ctx.engine.listFactsByEntity>> = [];
+5 -1
View File
@@ -18,6 +18,7 @@ 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';
@@ -4318,7 +4319,10 @@ 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';
const remoteFilter = opts.remote === true;
// 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 });
// Build SQL dynamically. PGLite uses $N positional params; we
// assemble the WHERE clauses + params array in tandem to keep them
+5 -1
View File
@@ -14,6 +14,7 @@ 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,
@@ -4532,7 +4533,10 @@ 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';
const remoteFilter = opts.remote === true;
// 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 });
// Source-scope predicate: array path (federated) wins over scalar.
// Engine.ts contract: returns chronological points; regressions +
+8
View File
@@ -30,6 +30,13 @@ 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'];
/**
@@ -203,6 +210,7 @@ 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
@@ -35,6 +35,16 @@ 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
@@ -42,6 +52,7 @@ 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
+16
View File
@@ -165,6 +165,22 @@ 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
@@ -0,0 +1,43 @@
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();
});
});
@@ -1,40 +0,0 @@
import { describe, expect, test } from 'bun:test';
import { mkdtempSync, mkdirSync, readFileSync, statSync, writeFileSync } from 'fs';
import { tmpdir } from 'os';
import { join, resolve } from 'path';
import { bootstrapDestinationConfig } from '../src/commands/migrate-engine.ts';
import { loadConfigFileOnly } from '../src/core/config.ts';
import { withEnv } from './helpers/with-env.ts';
describe('migrate --to pglite destination bootstrap (#1271)', () => {
test('writes <path>/.gbrain/config.json so GBRAIN_HOME=<path> resolves a brain', async () => {
const dest = mkdtempSync(join(tmpdir(), 'gbrain-dest-'));
const written = bootstrapDestinationConfig(dest);
const file = join(dest, '.gbrain', 'config.json');
expect(written).toBe(file);
const cfg = JSON.parse(readFileSync(file, 'utf-8'));
expect(cfg.engine).toBe('pglite');
expect(cfg.database_path).toBe(resolve(dest));
expect(statSync(file).mode & 0o777).toBe(0o600);
// worktree safety: destination home is git-ignored like saveConfig()'s home
expect(readFileSync(join(dest, '.gbrain', '.gitignore'), 'utf-8')).toBe('*\n');
// The exact failure mode from #1271: config resolution under
// GBRAIN_HOME=<path> used to find nothing ("No brain configured").
await withEnv({ GBRAIN_HOME: dest }, () => {
const loaded = loadConfigFileOnly();
expect(loaded?.engine).toBe('pglite');
expect(loaded?.database_path).toBe(resolve(dest));
});
});
test('never clobbers an existing destination config', () => {
const dest = mkdtempSync(join(tmpdir(), 'gbrain-dest-'));
mkdirSync(join(dest, '.gbrain'), { recursive: true });
writeFileSync(join(dest, '.gbrain', 'config.json'), '{"engine":"postgres"}\n');
expect(bootstrapDestinationConfig(dest)).toBe(null);
expect(JSON.parse(readFileSync(join(dest, '.gbrain', 'config.json'), 'utf-8')).engine).toBe('postgres');
});
});