Compare commits

..
Author SHA1 Message Date
Garry TanandClaude Fable 5 c80b8b6757 fix(search): honor sources.config.federated in unqualified local CLI search/query (#2561)
A source registered with `gbrain sources add --federated` was invisible to
an unqualified `gbrain search`/`gbrain query`: the local CLI always emitted
a scalar {sourceId} scope, and nothing on the read path ever consulted
sources.config.federated — contradicting docs/guides/multi-source-brains.md
('Source participates in unqualified gbrain search results').

Fix, at the trusted-local boundary only:
- src/cli.ts makeContext resolves the source WITH its tier and, when the
  tier is non-explicit (local_path / brain_default / sole_non_default /
  seed_default), computes ctx.localFederatedSourceIds = [resolved source,
  ...other config.federated=true sources] (archived excluded).
- New federatedSearchScope (operations.ts) delegates to
  resolveRequestedScope, then widens an unqualified trusted-local scalar
  scope to that set. Used by the search + query handlers only.
- Expansion NEVER applies when ctx.remote !== false (fail-closed source
  isolation), when a per-call source_id/__all__ is passed, when an OAuth
  grant (allowedSources) is present, or when --source/GBRAIN_SOURCE/dotfile
  named the source explicitly.

Deliberately NOT inside sourceScopeOpts: code-intel ops reject multi-source
scopes (resolveCodeIntelScope) and non-search reads keep their scalar
behavior. Cache contamination is already handled — cacheScopeKey folds
sourceIds sets into the query-cache key.

Fixes #2561

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-21 14:43:36 -07:00
12 changed files with 295 additions and 471 deletions
+11 -45
View File
@@ -466,11 +466,6 @@ async function main() {
const result = JSON.parse(JSON.stringify(rawResult, bigintToStringReplacer));
const output = formatResult(op.name, result);
if (output) process.stdout.write(output);
// #1484 — invisible-miss hint: a bare query/search that hit zero results
// on a multi-source brain tells the user (stderr) which source it
// actually searched and how to widen the scope.
const hint = await sourceScopeHint(op.name, params, ctx.sourceId, engine, result);
if (hint) console.error(hint);
} catch (e: unknown) {
// v0.42.20.0 (codex D4): on error, set exitCode + return so the `finally`
// STILL runs (drains every background-work sink + disconnects). A bare
@@ -813,12 +808,20 @@ async function makeContext(engine: BrainEngine, params: Record<string, unknown>)
// 'default'. Wrapped in try/catch so a doctor / single-source brain that
// never set up sources still returns 'default' silently.
let sourceId: string | undefined;
// #2561: when the source resolved via a NON-explicit tier (path-match /
// brain default / sole-non-default / seed default), unqualified search-shaped
// reads span every `config.federated = true` source. Computed here (the
// trusted local boundary) and consumed by federatedSearchScope in
// operations.ts, which additionally gates on ctx.remote === false.
let localFederated: string[] | undefined;
try {
const { resolveSourceId } = await import('./core/source-resolver.ts');
const { resolveSourceWithTier, localFederatedSourceIds } = await import('./core/source-resolver.ts');
// params.source is set when a CLI flag was parsed for the op (rare; most
// CLI ops don't take --source). Falls through to env/dotfile/path-match.
const explicit = (params.source as string | undefined) ?? null;
sourceId = await resolveSourceId(engine, explicit);
const resolved = await resolveSourceWithTier(engine, explicit);
sourceId = resolved.source_id;
localFederated = await localFederatedSourceIds(engine, resolved.source_id, resolved.tier);
} catch {
// Source resolution failed (e.g. sources table doesn't exist on a fresh
// pre-init brain). Leave sourceId unset; engine read methods fall through
@@ -839,47 +842,10 @@ async function makeContext(engine: BrainEngine, params: Record<string, unknown>)
// table). Matches dispatch.ts's auto-fill so the contract holds across
// every transport.
sourceId: sourceId ?? 'default',
...(localFederated ? { localFederatedSourceIds: localFederated } : {}),
};
}
/**
* #1484 — a bare `gbrain query`/`search` silently scopes to the resolved
* source (usually 'default'); on a multi-source brain a zero-hit run looks
* identical to "the brain doesn't know this" even when the answer lives in
* another source. Returns a stderr hint when (a) the op is query/search,
* (b) it returned zero results, (c) the caller did NOT scope explicitly
* (--source / --source-id / --all-sources), and (d) the brain has >1
* registered source. Best-effort: any lookup failure returns null.
*
* Exported for tests (same import-safety contract as formatResult).
*/
export async function sourceScopeHint(
opName: string,
params: Record<string, unknown>,
sourceId: string,
engine: BrainEngine,
result: unknown,
): Promise<string | null> {
if (opName !== 'query' && opName !== 'search') return null;
if (!Array.isArray(result) || result.length > 0) return null;
// Explicit scoping (flag tier) = user intent; don't second-guess it.
if (params.source || params.source_id || params.all_sources) return null;
if (sourceId === '__all__') return null;
try {
const rows = await engine.executeRaw<{ n: number }>(
`SELECT count(*)::int AS n FROM sources`,
);
const n = Number(rows[0]?.n ?? 0);
if (n <= 1) return null;
return (
`Hint: this brain has ${n} sources; you searched only "${sourceId}". ` +
`Retry with --source-id __all__ (all sources) or --source-id <id>.`
);
} catch {
return null; // hint is best-effort; never fail the query over it
}
}
// Exported for tests (same import-safety contract as cliAliases/printOpHelp).
export function formatResult(opName: string, result: unknown): string {
switch (opName) {
+20 -49
View File
@@ -1000,24 +1000,9 @@ const voyageCompatFetch = (async (input: RequestInfo | URL, init?: RequestInit)
// Voyage diverges from OpenAI in two places that break the parser:
// - `embedding` is a base64 string (SDK schema expects `number[]`)
// - `usage` lacks `prompt_tokens` (SDK schema requires it when usage present)
//
// #1610: read the body ONCE via text() and JSON.parse it. The pre-fix
// `await resp.clone().json()` truncated large bodies on bun < 1.1.27
// (oven-sh/bun#6348) — the parse threw, the catch fell back to the raw
// response, and multi-chunk pages died with "Invalid JSON response".
// Every JSON return path below rebuilds the Response so a stale
// Content-Length/Content-Encoding header from the original can't lie
// about the rewritten body.
const bodyText = await resp.text();
const rebuild = (body: string) => {
const headers = new Headers(resp.headers);
headers.delete('content-length');
headers.delete('content-encoding');
return new Response(body, { status: resp.status, statusText: resp.statusText, headers });
};
try {
const json: any = JSON.parse(bodyText);
if (!json || typeof json !== 'object') return rebuild(bodyText);
const json: any = await resp.clone().json();
if (!json || typeof json !== 'object') return resp;
let modified = false;
if (Array.isArray(json.data)) {
for (const item of json.data) {
@@ -1052,19 +1037,22 @@ const voyageCompatFetch = (async (input: RequestInfo | URL, init?: RequestInit)
: 0;
modified = true;
}
if (!modified) return rebuild(bodyText);
return rebuild(JSON.stringify(json));
if (!modified) return resp;
return new Response(JSON.stringify(json), {
status: resp.status,
statusText: resp.statusText,
headers: resp.headers,
});
} catch (err) {
// OOM-cap throws MUST propagate. The catch is here for "Voyage returned
// JSON I can't reshape" (parse error, unexpected schema) — falling back
// to the original body is correct in that case. Letting the
// to the original response is correct in that case. Letting the
// too-large response through here would defeat the entire purpose of
// Layer 2 (the per-embedding cap that fires when Content-Length wasn't
// available to Layer 1).
if (err instanceof VoyageResponseTooLargeError) throw err;
// If parsing/transformation fails, pass the original body through
// (rebuilt — resp's body stream is already consumed by text()).
return rebuild(bodyText);
// If parsing/transformation fails, fall back to the original response.
return resp;
}
}) as unknown as typeof fetch;
@@ -1204,21 +1192,9 @@ const zeroEntropyCompatFetch = (async (input: RequestInfo | URL, init?: RequestI
// validates. Also map usage.total_tokens → prompt_tokens (SDK requires
// prompt_tokens when `usage` is present — same divergence Voyage hit at
// gateway.ts:655).
//
// #1610: read the body ONCE via text() + JSON.parse — `resp.clone().json()`
// truncated large bodies on bun < 1.1.27 (oven-sh/bun#6348), so the parse
// threw and the catch fell back to the RAW ZE `{results: ...}` shape, which
// the AI SDK schema rejects → "Invalid JSON response" on multi-chunk pages.
const bodyText = await resp.text();
const rebuild = (body: string) => {
const headers = new Headers(resp.headers);
headers.delete('content-length');
headers.delete('content-encoding');
return new Response(body, { status: resp.status, statusText: resp.statusText, headers });
};
try {
const json: any = JSON.parse(bodyText);
if (!json || typeof json !== 'object') return rebuild(bodyText);
const json: any = await resp.clone().json();
if (!json || typeof json !== 'object') return resp;
let modified = false;
if (Array.isArray(json.results) && !Array.isArray(json.data)) {
// Layer 2 OOM cap — per-embedding size. ZE returns float[] arrays,
@@ -1252,25 +1228,20 @@ const zeroEntropyCompatFetch = (async (input: RequestInfo | URL, init?: RequestI
// SDK also expects total_tokens; ZE provides it directly.
modified = true;
}
if (!modified) return rebuild(bodyText);
return rebuild(JSON.stringify(json));
if (!modified) return resp;
return new Response(JSON.stringify(json), {
status: resp.status,
statusText: resp.statusText,
headers: resp.headers,
});
} catch (err) {
// OOM-cap throws MUST propagate. Voyage's pattern: instanceof check on
// its own tagged class. Same here — only rethrow our own cap class.
if (err instanceof ZeroEntropyResponseTooLargeError) throw err;
return rebuild(bodyText);
return resp;
}
}) as unknown as typeof fetch;
/**
* Test-only seams (#1610): the compat shims are module-private closures;
* exporting them lets tests drive the response-rewrite paths behaviorally
* (truncating clone(), stale Content-Length) without a live provider.
* Same pattern as __getShrinkStateForTests.
*/
export const __voyageCompatFetchForTests = voyageCompatFetch;
export const __zeroEntropyCompatFetchForTests = zeroEntropyCompatFetch;
/**
* Generic asymmetric-embedding shim for openai-compatible recipes that
* ship no compat fetch of their own (llama-server, litellm, ollama, ...).
+62 -14
View File
@@ -424,6 +424,23 @@ export interface OperationContext {
* satisfied even on single-source brains.
*/
sourceId: string;
/**
* #2561 — federated read scope for UNQUALIFIED local CLI reads.
*
* Set ONLY by the local CLI's context builder (src/cli.ts makeContext), and
* only when the source resolved via a non-explicit tier (local_path /
* brain_default / sole_non_default / seed_default — NOT --source, NOT
* GBRAIN_SOURCE, NOT a .gbrain-source dotfile). Contains the resolved
* source first, then every other `config.federated = true` source, so an
* unqualified `gbrain search "X"` spans federated sources as
* docs/guides/multi-source-brains.md promises.
*
* Consumed exclusively by `federatedSearchScope` and ONLY when
* `ctx.remote === false` — a remote caller's scope stays governed by
* `ctx.auth.allowedSources` / scalar `ctx.sourceId` (source-isolation
* invariant, fail-closed).
*/
localFederatedSourceIds?: string[];
}
/**
@@ -499,7 +516,7 @@ export function linkReadScopeOpts(ctx: OperationContext): { sourceId?: string; s
* FAIL-CLOSED: anything not strictly `ctx.remote === false` is untrusted.
*
* This is the SINGLE resolver for every read op that accepts a per-call
* `source_id` / `all_sources` parameter (query, search, code_callers, code_callees,
* `source_id` / `all_sources` parameter (query, code_callers, code_callees,
* get_page, search_by_image, code_blast, code_flow). Inlining the `__all__`
* branch per handler is the bug class that leaked cross-source reads (#1924,
* #1371): a remote client could pass `source_id: '__all__'` to opt out of its
@@ -539,6 +556,45 @@ export function resolveRequestedScope(
return sourceScopeOpts(ctx);
}
/**
* #2561 — source scope for the search-shaped read ops (`search`, `query`).
*
* Delegates to `resolveRequestedScope` (the single trust+grant resolver), then
* widens an UNQUALIFIED trusted-local scalar scope to the CLI-computed
* federated set (`ctx.localFederatedSourceIds`, resolved source first). This is
* what makes `sources add --federated` mean something for local search: a
* federated source participates in unqualified `gbrain search "X"` results.
*
* The expansion NEVER applies when:
* - the caller is not strictly trusted-local (`ctx.remote !== false`) —
* remote scope stays grant-governed (fail-closed source isolation);
* - a per-call `source_id` was passed (explicit wins, including `__all__`);
* - the resolver already produced a federated array (OAuth grant);
* - the CLI resolved the source from an explicit signal (--source / env /
* dotfile) — makeContext leaves `localFederatedSourceIds` unset then.
*
* Deliberately NOT inside `sourceScopeOpts`: code-intel ops collapse a
* multi-element scope to an error (`resolveCodeIntelScope`), and non-search
* reads (get_page, get_links, …) keep their long-standing scalar behavior.
*/
export function federatedSearchScope(
ctx: OperationContext,
sourceIdParam?: string,
): { sourceId?: string; sourceIds?: string[] } {
const scope = resolveRequestedScope(ctx, sourceIdParam);
if (
ctx.remote === false &&
sourceIdParam === undefined &&
scope.sourceId !== undefined &&
scope.sourceIds === undefined &&
ctx.localFederatedSourceIds !== undefined &&
ctx.localFederatedSourceIds.length > 1
) {
return { sourceIds: ctx.localFederatedSourceIds };
}
return scope;
}
/**
* Code-intel adapter for `resolveRequestedScope`. Graph traversal
* (code_callers/code_callees/code_blast/code_flow) is single-source by design —
@@ -1442,24 +1498,14 @@ const search: Operation = {
limit: { type: 'number', description: 'Max results (default 20)' },
offset: { type: 'number', description: 'Skip first N results (for pagination)' },
mode: { type: 'string', description: 'Search mode (conservative|balanced|tokenmax). Local callers only.' },
source_id: {
type: 'string',
description:
"Scope search to a single source. Defaults to OperationContext.sourceId. Pass '__all__' to span every source for trusted local callers; for remote callers '__all__' spans only your granted sources.",
},
all_sources: { type: 'boolean', description: "Span sources (equivalent to source_id=__all__): every source locally, your grant remotely." },
},
handler: async (ctx, p) => {
const startedAt = Date.now();
const queryText = p.query as string;
const limit = (p.limit as number) || 20;
const offset = (p.offset as number) || 0;
// #1484 follow-up: route through the canonical fail-closed resolver so
// `--source-id __all__` / `all_sources` behave the same as on `query`
// (the zero-hit CLI hint advises exactly that retry). Without a per-call
// param, `search` silently ignored --source-id — the retry looked like
// a genuine miss.
const scope = resolveRequestedScope(ctx, p.source_id as string | undefined, p.all_sources === true);
// #2561: unqualified trusted-local search spans federated sources.
const scope = federatedSearchScope(ctx);
// T4/D5 — per-call mode honored ONLY for trusted/local callers so a remote
// OAuth client can't escalate to the costly tokenmax bundle. Local + unknown
@@ -1621,7 +1667,9 @@ const query: Operation = {
// is spread into BOTH the image-similarity searchVector path and the text
// hybridSearch path below, so both honor the same grant.
const sourceIdParam = typeof p.source_id === 'string' ? p.source_id : undefined;
const querySourceScope = resolveRequestedScope(ctx, sourceIdParam);
// #2561: unqualified trusted-local query spans federated sources (per-call
// source_id / remote grants still resolve through resolveRequestedScope).
const querySourceScope = federatedSearchScope(ctx, sourceIdParam);
// v0.27.1: image-similarity branch. Bypasses hybridSearch (which is
// text-only); embeds the image via embedMultimodal and runs a direct
+2 -12
View File
@@ -1323,18 +1323,8 @@ export async function hybridSearch(
if (effectiveModality === 'both' && imageVectorList !== null) {
vectorLists = [...vectorLists, imageVectorList];
}
} catch (err) {
// Embedding/vector failure is non-fatal fall back to keyword-only
// but say WHY (#1626): this arm only runs when the embedding provider
// probed available, so a throw here is a real failure (embed timeout,
// transient pooler error on the searchVector fan-out). Pre-fix the bare
// catch made a cross-source `--source __all__` run silently collapse to
// keyword-only/"No results" with zero diagnostics.
warnOncePerProcess(
'hybrid-vector-arm-failed',
`[gbrain] vector arm failed (fail-open, keyword-only fallback): ` +
`${err instanceof Error ? err.message : String(err)}`,
);
} catch {
// Embedding failure is non-fatal, fall back to keyword-only
}
}
+39
View File
@@ -353,6 +353,45 @@ export async function resolveSourceWithTier(
return { source_id: 'default', tier: 'seed_default' };
}
/**
* #2561 — compute the federated read scope for an UNQUALIFIED local CLI call.
*
* `sources add --federated` promises that a `config.federated = true` source
* "participates in unqualified `gbrain search` results"
* (docs/guides/multi-source-brains.md). This helper turns that promise into a
* scope: given the resolved source and WHICH tier resolved it, return
* `[resolvedSource, ...other federated source ids]` — or `undefined` when the
* expansion must not apply:
*
* - explicit tiers (`flag` / `env` / `dotfile`): the user named a source;
* scalar scope stands (that IS the qualified case);
* - no other federated source exists: keep the scalar fast path unchanged.
*
* Archived sources are excluded (same rationale as pickSoleNonDefaultSource);
* the archived column is v34+, so fall back to the un-archived query on older
* brains. Callers put the result on `OperationContext.localFederatedSourceIds`
* — consumed only by `federatedSearchScope` and only when `remote === false`.
*/
export async function localFederatedSourceIds(
engine: BrainEngine,
sourceId: string,
tier: SourceTier,
): Promise<string[] | undefined> {
if (tier === 'flag' || tier === 'env' || tier === 'dotfile') return undefined;
let rows: Array<{ id: string }>;
try {
rows = await engine.executeRaw<{ id: string }>(
`SELECT id FROM sources WHERE config->>'federated' = 'true' AND archived = false ORDER BY id`,
);
} catch {
rows = await engine.executeRaw<{ id: string }>(
`SELECT id FROM sources WHERE config->>'federated' = 'true' ORDER BY id`,
);
}
const ids = [sourceId, ...rows.map((r) => r.id).filter((id) => id !== sourceId)];
return ids.length > 1 ? ids : undefined;
}
/** Exposed for tests. */
export const __testing = {
readDotfileWalk,
@@ -1,115 +0,0 @@
/**
* #1610 — Voyage/ZeroEntropy compat shims must read the response body ONCE
* via text() instead of `resp.clone().json()`.
*
* On bun < 1.1.27, Response.clone() truncates large bodies (oven-sh/bun#6348):
* the clone().json() parse threw, the shim's catch fell back to the ORIGINAL
* response — whose wire shape (ZE `{results: ...}`, Voyage base64 embeddings)
* the AI SDK's openai-compatible Zod schema rejects — and multi-chunk pages
* failed with "Invalid JSON response".
*
* These tests simulate the truncating clone() and assert the shims still
* return the fully rewritten body. They also pin that the rewritten Response
* does NOT carry the original (now stale) Content-Length header, which lied
* about the rewritten body's size (gateway.ts previously copied
* `headers: resp.headers` verbatim).
*/
import { afterEach, describe, expect, test } from 'bun:test';
import {
__voyageCompatFetchForTests,
__zeroEntropyCompatFetchForTests,
} from '../../src/core/ai/gateway.ts';
const origFetch = globalThis.fetch;
afterEach(() => {
globalThis.fetch = origFetch;
});
/** Build a Response whose clone() truncates the body (bun < 1.1.27 behavior). */
function truncatingCloneResponse(body: string): Response {
const headers = {
'content-type': 'application/json',
// Deliberately stale after any rewrite: the original wire body's length.
'content-length': String(Buffer.byteLength(body)),
};
const resp = new Response(body, { status: 200, headers });
(resp as any).clone = () =>
new Response(body.slice(0, 32), { status: 200, headers });
return resp;
}
describe('voyageCompatFetch — single body read (#1610)', () => {
test('rewrites base64 embeddings even when clone() truncates the body', async () => {
const floats = new Float32Array([0.5, 0.25, -1]);
const b64 = Buffer.from(floats.buffer).toString('base64');
const wireBody = JSON.stringify({
object: 'list',
data: [{ object: 'embedding', embedding: b64, index: 0 }],
model: 'voyage-3',
usage: { total_tokens: 7 },
});
globalThis.fetch = (async () => truncatingCloneResponse(wireBody)) as unknown as typeof fetch;
const out = await __voyageCompatFetchForTests('https://api.voyageai.com/v1/embeddings', {
method: 'POST',
body: JSON.stringify({ input: ['hello'], model: 'voyage-3' }),
headers: { 'content-type': 'application/json' },
});
const json: any = await out.json();
expect(Array.from(json.data[0].embedding)).toEqual([0.5, 0.25, -1]);
expect(json.usage.prompt_tokens).toBe(7);
// Stale Content-Length from the wire body must not survive the rewrite.
expect(out.headers.get('content-length')).toBeNull();
expect(out.headers.get('content-encoding')).toBeNull();
});
});
describe('zeroEntropyCompatFetch — single body read (#1610)', () => {
test('rewrites {results} → {data} even when clone() truncates the body', async () => {
const wireBody = JSON.stringify({
results: [{ embedding: [0.1, 0.2] }, { embedding: [0.3, 0.4] }],
usage: { total_bytes: 42, total_tokens: 9 },
});
let fetchedUrl = '';
globalThis.fetch = (async (url: string | URL | Request) => {
fetchedUrl = String(url);
return truncatingCloneResponse(wireBody);
}) as unknown as typeof fetch;
const out = await __zeroEntropyCompatFetchForTests('https://api.zeroentropy.dev/v1/embeddings', {
method: 'POST',
body: JSON.stringify({ input: ['hello'], model: 'zembed-1' }),
headers: { 'content-type': 'application/json' },
});
expect(fetchedUrl.endsWith('/v1/models/embed')).toBe(true);
const json: any = await out.json();
// The AI SDK schema requires {data: [{embedding, index}]} — the raw ZE
// {results} fallback is exactly the pre-fix "Invalid JSON response".
expect(json.results).toBeUndefined();
expect(json.data).toHaveLength(2);
expect(json.data[0]).toEqual({ object: 'embedding', embedding: [0.1, 0.2], index: 0 });
expect(json.data[1].index).toBe(1);
expect(json.usage.prompt_tokens).toBe(9);
expect(out.headers.get('content-length')).toBeNull();
});
test('non-JSON body falls back to the original bytes (rebuilt, still readable)', async () => {
const wireBody = 'plain text, not json';
globalThis.fetch = (async () =>
new Response(wireBody, {
status: 200,
headers: { 'content-type': 'application/json' },
})) as unknown as typeof fetch;
const out = await __zeroEntropyCompatFetchForTests('https://api.zeroentropy.dev/v1/embeddings', {
method: 'POST',
body: JSON.stringify({ input: ['hello'] }),
});
// Body was consumed by the shim's single read; the fallback must
// rebuild a readable Response rather than return the drained original.
expect(await out.text()).toBe(wireBody);
});
});
+4 -6
View File
@@ -98,18 +98,16 @@ describe('zeroEntropyCompatFetch — OOM caps', () => {
expect(src).toMatch(/MAX_ZEROENTROPY_RESPONSE_BYTES\s*=\s*256\s*\*\s*1024\s*\*\s*1024/);
});
test('Layer 1: Content-Length pre-check before the body is read', async () => {
test('Layer 1: Content-Length pre-check before resp.clone().json()', async () => {
const src = await Bun.file(GATEWAY_PATH).text();
// Find the zeroEntropyCompatFetch block bounds, then assert ordering
// within it (mirroring the voyage cap test pattern). #1610 moved the
// body read from `resp.clone().json()` to a single `resp.text()` (bun
// < 1.1.27 truncates clone()d bodies, oven-sh/bun#6348).
// within it (mirroring the voyage cap test pattern).
const zeFetchStart = src.indexOf('const zeroEntropyCompatFetch');
expect(zeFetchStart).toBeGreaterThan(0);
const block = src.slice(zeFetchStart, zeFetchStart + 9000);
const block = src.slice(zeFetchStart, zeFetchStart + 8000);
const preCheckIdx = block.indexOf("resp.headers.get('content-length')");
const jsonParseIdx = block.indexOf('const bodyText = await resp.text()');
const jsonParseIdx = block.indexOf('await resp.clone().json()');
expect(preCheckIdx).toBeGreaterThan(0);
expect(jsonParseIdx).toBeGreaterThan(0);
// The pre-check MUST appear before the JSON parse — Voyage's lesson
-61
View File
@@ -1,61 +0,0 @@
/**
* #1484 — invisible-miss hint. A bare `gbrain query` resolves to a single
* source (usually 'default'); on a multi-source brain a zero-hit run gave no
* signal that the answer might live in another source. sourceScopeHint
* returns the stderr hint exactly when: query/search op + zero results +
* no explicit scoping param + >1 registered source.
*/
import { describe, expect, test } from 'bun:test';
import { sourceScopeHint } from '../src/cli.ts';
import type { BrainEngine } from '../src/core/engine.ts';
function fakeEngine(sourceCount: number, fail = false): BrainEngine {
return {
executeRaw: async () => {
if (fail) throw new Error('sources table missing');
return [{ n: sourceCount }];
},
} as unknown as BrainEngine;
}
describe('sourceScopeHint (#1484)', () => {
test('fires on a bare zero-hit query against a multi-source brain', async () => {
const hint = await sourceScopeHint('query', {}, 'default', fakeEngine(3), []);
expect(hint).toContain('3 sources');
expect(hint).toContain('"default"');
expect(hint).toContain('--source-id __all__');
});
test('fires for search too', async () => {
const hint = await sourceScopeHint('search', {}, 'wiki', fakeEngine(2), []);
expect(hint).toContain('"wiki"');
});
test('silent when results were found', async () => {
expect(await sourceScopeHint('query', {}, 'default', fakeEngine(3), [{ slug: 'a' }])).toBeNull();
});
test('silent when the caller scoped explicitly', async () => {
expect(await sourceScopeHint('query', { source_id: 'wiki' }, 'wiki', fakeEngine(3), [])).toBeNull();
expect(await sourceScopeHint('query', { source: 'wiki' }, 'wiki', fakeEngine(3), [])).toBeNull();
expect(await sourceScopeHint('query', { all_sources: true }, '__all__', fakeEngine(3), [])).toBeNull();
});
test('silent when the resolved scope is already __all__', async () => {
expect(await sourceScopeHint('query', {}, '__all__', fakeEngine(3), [])).toBeNull();
});
test('silent on a single-source brain', async () => {
expect(await sourceScopeHint('query', {}, 'default', fakeEngine(1), [])).toBeNull();
});
test('silent for non-search ops and non-array results', async () => {
expect(await sourceScopeHint('get_stats', {}, 'default', fakeEngine(3), [])).toBeNull();
expect(await sourceScopeHint('query', {}, 'default', fakeEngine(3), { rows: [] })).toBeNull();
});
test('best-effort: sources lookup failure returns null, never throws', async () => {
expect(await sourceScopeHint('query', {}, 'default', fakeEngine(3, true), [])).toBeNull();
});
});
@@ -1,77 +0,0 @@
/**
* #1626 hybridSearch's text-vector arm must not fail DARK.
*
* The arm only runs when the embedding provider probed available, so a throw
* inside it (embed timeout, transient pooler error on searchVector) is a real
* failure. Pre-fix, a bare `catch {}` swallowed it and the run silently
* collapsed to keyword-only under `--source __all__` on a strained pooler
* that read as a non-deterministic "No results". The fix logs the swallowed
* reason via warnOncePerProcess while keeping the keyword fallback.
*/
import { afterAll, beforeAll, describe, expect, test } from 'bun:test';
import { PGLiteEngine } from '../src/core/pglite-engine.ts';
import { hybridSearch } from '../src/core/search/hybrid.ts';
import {
__setEmbedTransportForTests,
configureGateway,
resetGateway,
} from '../src/core/ai/gateway.ts';
import { _resetWarnOnceForTests } from '../src/core/utils.ts';
let engine: PGLiteEngine;
const origWarn = console.warn;
beforeAll(async () => {
// Pin the gateway to OpenAI with a stub key (put-page-provenance pattern):
// embed() runs instantiateEmbedding — which requires OPENAI_API_KEY — BEFORE
// the stubbed transport is reached. Without this, a keyless CI environment
// throws the config error instead of the transport's, and the assertion on
// the swallowed reason fails. The key never leaves the process.
configureGateway({
embedding_model: 'openai:text-embedding-3-large',
embedding_dimensions: 1536,
env: { ...process.env, OPENAI_API_KEY: process.env.OPENAI_API_KEY || 'sk-test-stub' },
});
engine = new PGLiteEngine();
await engine.connect({});
await engine.initSchema();
await engine.putPage('people/alice-example', {
type: 'person',
title: 'Alice Example',
compiled_truth: 'Alice Example is a test person for the vector-arm warn test.',
});
});
afterAll(async () => {
console.warn = origWarn;
__setEmbedTransportForTests(null);
resetGateway();
await engine.disconnect();
});
describe('hybridSearch vector-arm failure telemetry (#1626)', () => {
test('embed failure logs the swallowed reason and falls back to keyword', async () => {
_resetWarnOnceForTests();
// Installing a transport makes isAvailable('embedding') true (test-seam
// fast path), so the vector arm RUNS — and then throws.
__setEmbedTransportForTests(() => {
throw new Error('pooler exploded mid-fanout');
});
const warnings: string[] = [];
console.warn = (...args: unknown[]) => {
warnings.push(args.map(String).join(' '));
};
try {
const results = await hybridSearch(engine, 'alice');
// Keyword fallback still returns results — fail-open preserved.
expect(results.some((r) => r.slug === 'people/alice-example')).toBe(true);
} finally {
console.warn = origWarn;
__setEmbedTransportForTests(null);
}
const armWarnings = warnings.filter((w) => w.includes('vector arm failed'));
expect(armWarnings).toHaveLength(1);
expect(armWarnings[0]).toContain('pooler exploded mid-fanout');
});
});
+154
View File
@@ -0,0 +1,154 @@
/**
* #2561 sources.config.federated participates in UNQUALIFIED local CLI
* search/query.
*
* Pre-fix: the local CLI always emitted a scalar `{sourceId}` scope (required
* field, auto-filled 'default'), so a source registered with
* `gbrain sources add --federated` was invisible to an unqualified
* `gbrain search "X"` contradicting docs/guides/multi-source-brains.md
* ("Source participates in unqualified `gbrain search` results").
*
* Fix: the CLI context builder computes `ctx.localFederatedSourceIds`
* (resolved source + every other federated source) whenever the source
* resolved via a NON-explicit tier; `federatedSearchScope` widens the scalar
* scope to that set for the `search` / `query` ops trusted-local only
* (`ctx.remote === false`), never for remote callers, never when a per-call
* `source_id` or an explicit --source/env/dotfile was given.
*/
import { describe, test, expect, beforeAll, afterAll } from 'bun:test';
import { PGLiteEngine } from '../src/core/pglite-engine.ts';
import { localFederatedSourceIds } from '../src/core/source-resolver.ts';
import {
federatedSearchScope,
operations,
type OperationContext,
} from '../src/core/operations.ts';
let engine: PGLiteEngine;
const search = operations.find((o) => o.name === 'search')!;
function ctxOf(overrides: Partial<OperationContext> = {}): OperationContext {
return {
engine: engine as any,
config: {} as any,
logger: console as any,
dryRun: false,
remote: false,
sourceId: 'default',
...overrides,
};
}
beforeAll(async () => {
engine = new PGLiteEngine();
await engine.connect({});
await engine.initSchema();
// Seeded 'default' source is federated=true. Add:
// wiki — federated (must join unqualified search)
// private — NOT federated (must stay invisible unless explicitly named)
// oldnews — federated but archived (must stay excluded)
await engine.executeRaw(
`INSERT INTO sources (id, name, local_path, config) VALUES ('wiki', 'wiki', '/tmp/wiki', '{"federated": true}'::jsonb)`,
);
await engine.executeRaw(
`INSERT INTO sources (id, name, local_path, config) VALUES ('private', 'private', '/tmp/private', '{}'::jsonb)`,
);
await engine.executeRaw(
`INSERT INTO sources (id, name, local_path, config, archived) VALUES ('oldnews', 'oldnews', '/tmp/oldnews', '{"federated": true}'::jsonb, true)`,
);
const pages: Array<[slug: string, sourceId: string, where: string]> = [
['notes/home', 'default', 'default'],
['wiki/topic', 'wiki', 'wiki'],
['private/topic', 'private', 'private'],
['old/topic', 'oldnews', 'oldnews'],
];
for (const [slug, sourceId, where] of pages) {
await engine.putPage(slug, {
type: 'note', title: `Topic in ${where}`, compiled_truth: `the zebra telescope in ${where}`, frontmatter: {},
}, { sourceId });
await engine.upsertChunks(slug, [
{ chunk_index: 0, chunk_text: `the zebra telescope in ${where}`, chunk_source: 'compiled_truth' },
], { sourceId });
}
// Keyword-only search path: no embedding provider needed in tests.
await engine.setConfig('search.mcp_keyword_only', 'true');
}, 60_000);
afterAll(async () => {
if (engine) await engine.disconnect();
}, 60_000);
describe('localFederatedSourceIds — CLI-side scope computation', () => {
test('non-explicit tier: resolved source first, then other federated, archived excluded', async () => {
expect(await localFederatedSourceIds(engine, 'default', 'seed_default')).toEqual(['default', 'wiki']);
});
test('non-federated resolved source still joins its own scope', async () => {
expect(await localFederatedSourceIds(engine, 'private', 'brain_default')).toEqual(['private', 'default', 'wiki']);
});
test('explicit tiers (--source / env / dotfile) never expand', async () => {
expect(await localFederatedSourceIds(engine, 'default', 'flag')).toBeUndefined();
expect(await localFederatedSourceIds(engine, 'default', 'env')).toBeUndefined();
expect(await localFederatedSourceIds(engine, 'default', 'dotfile')).toBeUndefined();
});
test('single federated source (the resolved one) keeps the scalar fast path', async () => {
const solo = { executeRaw: async () => [{ id: 'default' }] } as any;
expect(await localFederatedSourceIds(solo, 'default', 'seed_default')).toBeUndefined();
});
});
describe('federatedSearchScope — trust + explicitness matrix', () => {
test('trusted local + unqualified widens to the federated set', () => {
const ctx = ctxOf({ localFederatedSourceIds: ['default', 'wiki'] });
expect(federatedSearchScope(ctx)).toEqual({ sourceIds: ['default', 'wiki'] });
});
test('remote caller NEVER widens (fail-closed), even if the field is set', () => {
const ctx = ctxOf({ remote: true, localFederatedSourceIds: ['default', 'wiki'] });
expect(federatedSearchScope(ctx)).toEqual({ sourceId: 'default' });
});
test('per-call source_id wins over the federated set', () => {
const ctx = ctxOf({ localFederatedSourceIds: ['default', 'wiki'] });
expect(federatedSearchScope(ctx, 'wiki')).toEqual({ sourceId: 'wiki' });
});
test('per-call __all__ keeps the whole-brain semantics for trusted local', () => {
const ctx = ctxOf({ localFederatedSourceIds: ['default', 'wiki'] });
expect(federatedSearchScope(ctx, '__all__')).toEqual({});
});
test('a federated OAuth grant wins over the local set', () => {
const ctx = ctxOf({
localFederatedSourceIds: ['default', 'wiki'],
auth: { allowedSources: ['a', 'b'] } as OperationContext['auth'],
});
expect(federatedSearchScope(ctx)).toEqual({ sourceIds: ['a', 'b'] });
});
test('no local federated set → unchanged scalar scope', () => {
expect(federatedSearchScope(ctxOf())).toEqual({ sourceId: 'default' });
});
});
describe('search op — unqualified local search spans federated sources', () => {
test('federated source results appear; non-federated + archived stay invisible', async () => {
const ctx = ctxOf({
localFederatedSourceIds: await localFederatedSourceIds(engine, 'default', 'seed_default'),
});
const results = (await search.handler(ctx, { query: 'zebra telescope' })) as Array<{ slug: string }>;
const slugs = results.map((r) => r.slug);
expect(slugs).toContain('notes/home');
expect(slugs).toContain('wiki/topic'); // pre-#2561 this was missing
expect(slugs).not.toContain('private/topic');
expect(slugs).not.toContain('old/topic');
});
test('explicit source resolution (no federated set on ctx) stays single-source', async () => {
const results = (await search.handler(ctxOf(), { query: 'zebra telescope' })) as Array<{ slug: string }>;
const slugs = results.map((r) => r.slug);
expect(slugs).toEqual(['notes/home']);
});
});
-87
View File
@@ -1,87 +0,0 @@
/**
* #1484 follow-up the `search` op must honor per-call `source_id` /
* `all_sources` through the canonical fail-closed resolver
* (resolveRequestedScope), exactly like `query` does.
*
* Pre-fix, `search` had no source_id param at all: the zero-hit CLI hint
* advised "retry with --source-id __all__", the flag parsed into params,
* NOTHING consumed it, and the retry silently re-ran the same single-source
* search an invisible false negative (and the retry's params.source_id
* suppressed the hint, so the user got no second warning).
*/
import { describe, expect, test } from 'bun:test';
import { operationsByName } from '../src/core/operations.ts';
import type { OperationContext } from '../src/core/operations.ts';
import type { BrainEngine } from '../src/core/engine.ts';
const searchOp = operationsByName['search'];
/** Fake engine: keyword-only config so the handler's scope goes straight to
* searchKeyword, where we capture the opts it was called with. */
function makeCtx(remote: boolean, allowedSources?: string[]) {
const captured: { opts?: Record<string, unknown> } = {};
const engine = {
getConfig: async (key: string) => (key === 'search.mcp_keyword_only' ? 'true' : null),
searchKeyword: async (_q: string, opts: Record<string, unknown>) => {
captured.opts = opts;
return [];
},
} as unknown as BrainEngine;
const ctx = {
engine,
config: { engine: 'pglite' },
logger: { info: () => {}, warn: () => {}, error: () => {} },
dryRun: false,
remote,
sourceId: 'default',
...(allowedSources ? { auth: { allowedSources } } : {}),
} as unknown as OperationContext;
return { ctx, captured };
}
describe('search op per-call source scope (#1484 follow-up)', () => {
test('op declares source_id + all_sources params (the CLI hint advises them)', () => {
expect(searchOp.params.source_id).toBeDefined();
expect(searchOp.params.all_sources).toBeDefined();
});
test('default: scopes to ctx.sourceId', async () => {
const { ctx, captured } = makeCtx(false);
await searchOp.handler(ctx, { query: 'x' });
expect(captured.opts?.sourceId).toBe('default');
});
test("local + source_id '__all__' spans the whole brain (no source filter)", async () => {
const { ctx, captured } = makeCtx(false);
await searchOp.handler(ctx, { query: 'x', source_id: '__all__' });
expect(captured.opts?.sourceId).toBeUndefined();
expect(captured.opts?.sourceIds).toBeUndefined();
});
test('local + all_sources=true spans the whole brain', async () => {
const { ctx, captured } = makeCtx(false);
await searchOp.handler(ctx, { query: 'x', all_sources: true });
expect(captured.opts?.sourceId).toBeUndefined();
expect(captured.opts?.sourceIds).toBeUndefined();
});
test('explicit source_id wins over ctx.sourceId', async () => {
const { ctx, captured } = makeCtx(false);
await searchOp.handler(ctx, { query: 'x', source_id: 'wiki' });
expect(captured.opts?.sourceId).toBe('wiki');
});
test("remote + '__all__' collapses to the caller's grant (fail-closed)", async () => {
const { ctx, captured } = makeCtx(true, ['wiki', 'essays']);
await searchOp.handler(ctx, { query: 'x', source_id: '__all__' });
expect(captured.opts?.sourceIds).toEqual(['wiki', 'essays']);
});
test('remote + out-of-grant source_id is denied', async () => {
const { ctx } = makeCtx(true, ['wiki']);
await expect(searchOp.handler(ctx, { query: 'x', source_id: 'secrets' })).rejects.toThrow(
/outside your granted sources/,
);
});
});
+3 -5
View File
@@ -34,7 +34,7 @@ describe('v0.31.8 — voyage Content-Length pre-check + per-item cap', () => {
expect(source).toMatch(/MAX_VOYAGE_RESPONSE_BYTES\s*=\s*256\s*\*\s*1024\s*\*\s*1024/);
});
test('Layer 1: Content-Length pre-check fires BEFORE the body is read (D10 OOM defense)', async () => {
test('Layer 1: Content-Length pre-check fires BEFORE resp.clone().json() (D10 OOM defense)', async () => {
const source = await Bun.file(new URL('../src/core/ai/gateway.ts', import.meta.url)).text();
// Anchor relative to the post-fetch handler block. The function declaration
// contains an OUTBOUND request body section earlier; we want to verify
@@ -47,10 +47,8 @@ describe('v0.31.8 — voyage Content-Length pre-check + per-item cap', () => {
// doesn't pin to comment text.
const preCheckIdx = inboundBlock.indexOf("resp.headers.get('content-length')");
// Use the full lvalue assignment so the match doesn't accidentally hit
// comment text that mentions the body read for context. (#1610 moved the
// read from `resp.clone().json()` to a single `resp.text()` — bun <
// 1.1.27 truncates clone()d bodies, oven-sh/bun#6348.)
const jsonParseIdx = inboundBlock.indexOf('const bodyText = await resp.text()');
// comment text that mentions `await resp.clone().json()` for context.
const jsonParseIdx = inboundBlock.indexOf('const json: any = await resp.clone().json()');
expect(preCheckIdx).toBeGreaterThan(0);
expect(jsonParseIdx).toBeGreaterThan(0);
// The pre-check MUST appear before the JSON parse — otherwise the OOM