fix(search): honor per-call source_id/all_sources in the search op — the #1484 hint's advised retry was silently ignored

The zero-hit CLI hint tells users to retry with --source-id __all__, which
works for `query` (routes through resolveRequestedScope) but the `search`
op had no source_id/all_sources params: the flag parsed into params, nothing
consumed it, and the retry re-ran the same single-source search — an
invisible false negative whose params.source_id also suppressed the hint on
the retry. search now resolves per-call scope through the same canonical
fail-closed resolver as query (local __all__ spans the brain; remote
collapses to the caller's grant; out-of-grant source_id is denied).

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

src/cli.ts

@@ -466,6 +466,11 @@ 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
@@ -837,6 +842,44 @@ async function makeContext(engine: BrainEngine, params: Record<string, unknown>)
   };
 }
 
+/**
+ * #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) {

src/commands/embed.ts

@@ -1,5 +1,5 @@
 import type { BrainEngine } from '../core/engine.ts';
-import { embedBatch, currentEmbeddingSignature, resolveEmbeddingModelLabel } from '../core/embedding.ts';
+import { embedBatch, currentEmbeddingSignature } from '../core/embedding.ts';
 import type { ChunkInput } from '../core/types.ts';
 import { chunkText } from '../core/chunkers/recursive.ts';
 import { createProgress, type ProgressReporter } from '../core/progress.ts';
@@ -581,16 +581,11 @@ async function embedPage(
   for (let j = 0; j < toEmbed.length; j++) {
     embeddingMap.set(toEmbed[j].chunk_index, embeddings[j]);
   }
-  // #1717: label each (re)embedded chunk with the model that actually
-  // produced its vector. Preserved chunks (not re-embedded this pass) keep
-  // their existing model so a mixed-model page isn't relabeled wholesale.
-  const embedModelLabel = resolveEmbeddingModelLabel();
   const updated: ChunkInput[] = chunks.map(c => ({
     chunk_index: c.chunk_index,
     chunk_text: c.chunk_text,
     chunk_source: c.chunk_source,
     embedding: embeddingMap.get(c.chunk_index),
-    model: embeddingMap.has(c.chunk_index) && embedModelLabel ? embedModelLabel : c.model,
     token_count: c.token_count || Math.ceil(c.chunk_text.length / 4),
   }));
 
@@ -722,16 +717,12 @@ async function embedAll(
       for (let j = 0; j < toEmbed.length; j++) {
         embeddingMap.set(toEmbed[j].chunk_index, embeddings[j]);
       }
-      // #1717: stamp the resolved embedding model on (re)embedded chunks;
-      // preserve the existing model on chunks left untouched.
-      const embedModelLabel = resolveEmbeddingModelLabel();
       // Preserve ALL chunks, only update embeddings for stale ones
       const updated: ChunkInput[] = chunks.map(c => ({
         chunk_index: c.chunk_index,
         chunk_text: c.chunk_text,
         chunk_source: c.chunk_source,
         embedding: embeddingMap.get(c.chunk_index) ?? undefined,
-        model: embeddingMap.has(c.chunk_index) && embedModelLabel ? embedModelLabel : c.model,
         token_count: c.token_count || Math.ceil(c.chunk_text.length / 4),
       }));
       await observed(pacer, () => engine.upsertChunks(page.slug, updated, pageOpts));
@@ -1021,15 +1012,11 @@ async function embedAllStale(
           for (let j = 0; j < stale.length; j++) {
             staleIdxToEmbedding.set(stale[j].chunk_index, embeddings[j]);
           }
-          // #1717: label the re-embedded (stale) chunks with the resolved
-          // model; preserve the existing model on the non-stale chunks.
-          const embedModelLabel = resolveEmbeddingModelLabel();
           const merged: ChunkInput[] = existing.map(c => ({
             chunk_index: c.chunk_index,
             chunk_text: c.chunk_text,
             chunk_source: c.chunk_source,
             embedding: staleIdxToEmbedding.get(c.chunk_index) ?? undefined,
-            model: staleIdxToEmbedding.has(c.chunk_index) && embedModelLabel ? embedModelLabel : c.model,
             token_count: c.token_count || Math.ceil(c.chunk_text.length / 4),
           }));
           await observed(pacer, () => engine.upsertChunks(slug, merged, { sourceId: keySourceId }));

src/core/ai/gateway.ts

@@ -1000,9 +1000,24 @@ 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 = await resp.clone().json();
-    if (!json || typeof json !== 'object') return resp;
+    const json: any = JSON.parse(bodyText);
+    if (!json || typeof json !== 'object') return rebuild(bodyText);
     let modified = false;
     if (Array.isArray(json.data)) {
       for (const item of json.data) {
@@ -1037,22 +1052,19 @@ const voyageCompatFetch = (async (input: RequestInfo | URL, init?: RequestInit)
         : 0;
       modified = true;
     }
-    if (!modified) return resp;
-    return new Response(JSON.stringify(json), {
-      status: resp.status,
-      statusText: resp.statusText,
-      headers: resp.headers,
-    });
+    if (!modified) return rebuild(bodyText);
+    return rebuild(JSON.stringify(json));
   } 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 response is correct in that case. Letting the
+    // to the original body 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, fall back to the original response.
-    return resp;
+    // If parsing/transformation fails, pass the original body through
+    // (rebuilt — resp's body stream is already consumed by text()).
+    return rebuild(bodyText);
   }
 }) as unknown as typeof fetch;
 
@@ -1192,9 +1204,21 @@ 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 = await resp.clone().json();
-    if (!json || typeof json !== 'object') return resp;
+    const json: any = JSON.parse(bodyText);
+    if (!json || typeof json !== 'object') return rebuild(bodyText);
     let modified = false;
     if (Array.isArray(json.results) && !Array.isArray(json.data)) {
       // Layer 2 OOM cap — per-embedding size. ZE returns float[] arrays,
@@ -1228,20 +1252,25 @@ const zeroEntropyCompatFetch = (async (input: RequestInfo | URL, init?: RequestI
       // SDK also expects total_tokens; ZE provides it directly.
       modified = true;
     }
-    if (!modified) return resp;
-    return new Response(JSON.stringify(json), {
-      status: resp.status,
-      statusText: resp.statusText,
-      headers: resp.headers,
-    });
+    if (!modified) return rebuild(bodyText);
+    return rebuild(JSON.stringify(json));
   } 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 resp;
+    return rebuild(bodyText);
   }
 }) 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, ...).

src/core/embed-stale.ts

@@ -20,7 +20,6 @@
 import type { BrainEngine } from './engine.ts';
 import type { ChunkInput } from './types.ts';
 import { embedBatchWithBackoff } from '../commands/embed.ts';
-import { resolveEmbeddingModelLabel } from './embedding.ts';
 import { type DbPacer, createNoopPacer, observed } from './db-pacer.ts';
 import { AbortError } from './abort-check.ts';
 
@@ -201,17 +200,11 @@ export async function embedStaleForSource(
         for (let j = 0; j < stale.length; j++) {
           staleIdxToEmbedding.set(stale[j].chunk_index, embeddings[j]);
         }
-        // #1717: label re-embedded chunks with the model that produced the
-        // vector; preserved chunks keep their existing model. Without this,
-        // upsertChunks falls back to DEFAULT_EMBEDDING_MODEL for every chunk
-        // (the same mislabel the embed.ts paths fixed).
-        const embedModelLabel = resolveEmbeddingModelLabel();
         const merged: ChunkInput[] = existing.map((c) => ({
           chunk_index: c.chunk_index,
           chunk_text: c.chunk_text,
           chunk_source: c.chunk_source,
           embedding: staleIdxToEmbedding.get(c.chunk_index) ?? undefined,
-          model: staleIdxToEmbedding.has(c.chunk_index) && embedModelLabel ? embedModelLabel : c.model,
           token_count: c.token_count || Math.ceil(c.chunk_text.length / 4),
           // Carry through per-chunk metadata. upsertChunks writes these as
           // EXCLUDED.<col> (not COALESCE), so omitting them here resets image

src/core/embedding.ts

@@ -113,21 +113,6 @@ export async function embedBatch(
   return results;
 }
 
-/**
- * Resolve the embedding model label (`provider:model`) to stamp onto
- * `content_chunks.model`, so each chunk records the model that actually
- * produced its vector instead of the engine's hardcoded default (#1717).
- * Returns undefined if the gateway is unconfigured; callers then fall back
- * to the chunk's existing model rather than mislabeling it.
- */
-export function resolveEmbeddingModelLabel(): string | undefined {
-  try {
-    return gatewayGetModel();
-  } catch {
-    return undefined;
-  }
-}
-
 /** Currently-configured embedding model (short form without provider prefix). */
 export function getEmbeddingModelName(): string {
   return gatewayGetModel().split(':').slice(1).join(':') || 'text-embedding-3-large';

src/core/import-file.ts

@@ -8,7 +8,7 @@ import { chunkText } from './chunkers/recursive.ts';
 import { chunkCodeText, chunkCodeTextFull, detectCodeLanguage, CHUNKER_VERSION } from './chunkers/code.ts';
 import { findChunkForOffset } from './chunkers/edge-extractor.ts';
 import { extractCodeRefs, imageOfCandidates } from './link-extraction.ts';
-import { embedBatch, embedMultimodal, currentEmbeddingSignature, resolveEmbeddingModelLabel } from './embedding.ts';
+import { embedBatch, embedMultimodal, currentEmbeddingSignature } from './embedding.ts';
 import { slugifyPath, slugifyCodePath, isCodeFilePath } from './sync.ts';
 import type { ChunkInput, PageInput, PageType } from './types.ts';
 import { computeEffectiveDate } from './effective-date.ts';
@@ -716,12 +716,8 @@ export async function importFromContent(
       ? chunks.map((c) => wrapChunkForEmbedding(c.chunk_text, prefix, c.chunk_source))
       : chunks.map((c) => c.chunk_text);
     const embeddings = await embedBatch(wrappedTexts);
-    // #1717: label each chunk with the model that actually produced its
-    // vector, not the engine's hardcoded default.
-    const embedModelLabel = resolveEmbeddingModelLabel();
     for (let i = 0; i < chunks.length; i++) {
       chunks[i].embedding = embeddings[i];
-      if (embedModelLabel) chunks[i].model = embedModelLabel;
       // token_count tracks the wrapped string length so cost reporting
       // reflects what we actually sent to the embedder.
       chunks[i].token_count = Math.ceil(wrappedTexts[i].length / 4);
@@ -1145,10 +1141,7 @@ export async function importCodeFile(
     const matched = existingByKey.get(key);
     if (matched && matched.embedding) {
       // Reuse the existing embedding verbatim. No API call, no cost.
-      // #1717: carry the existing model label along with the reused vector
-      // so the upsert doesn't relabel it with the engine default.
       chunks[i]!.embedding = matched.embedding as Float32Array;
-      chunks[i]!.model = matched.model ?? undefined;
       chunks[i]!.token_count = matched.token_count ?? undefined;
     } else {
       needsEmbedIndexes.push(i);
@@ -1160,12 +1153,9 @@ export async function importCodeFile(
     try {
       const textsToEmbed = needsEmbedIndexes.map((i) => chunks[i]!.chunk_text);
       const embeddings = await embedBatch(textsToEmbed);
-      // #1717: stamp the model that produced these vectors.
-      const embedModelLabel = resolveEmbeddingModelLabel();
       for (let j = 0; j < needsEmbedIndexes.length; j++) {
         const i = needsEmbedIndexes[j]!;
         chunks[i]!.embedding = embeddings[j]!;
-        if (embedModelLabel) chunks[i]!.model = embedModelLabel;
         chunks[i]!.token_count = Math.ceil(chunks[i]!.chunk_text.length / 4);
       }
     } catch (e: unknown) {

src/core/operations.ts

@@ -499,7 +499,7 @@ 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, code_callers, code_callees,
+ * `source_id` / `all_sources` parameter (query, search, 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
@@ -1442,13 +1442,24 @@ 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;
-    const scope = sourceScopeOpts(ctx);
+    // #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);
 
     // 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

src/core/search/hybrid.ts

@@ -1323,8 +1323,18 @@ export async function hybridSearch(
       if (effectiveModality === 'both' && imageVectorList !== null) {
         vectorLists = [...vectorLists, imageVectorList];
       }
-    } catch {
-      // Embedding failure is non-fatal, fall back to keyword-only
+    } 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)}`,
+      );
     }
   }
 

test/ai/compat-fetch-body-read.serial.test.ts

@@ -0,0 +1,115 @@
+/**
+ * #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);
+  });
+});

test/ai/zeroentropy-compat-fetch.test.ts

@@ -98,16 +98,18 @@ 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 resp.clone().json()', async () => {
+  test('Layer 1: Content-Length pre-check before the body is read', 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).
+    // 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).
     const zeFetchStart = src.indexOf('const zeroEntropyCompatFetch');
     expect(zeFetchStart).toBeGreaterThan(0);
-    const block = src.slice(zeFetchStart, zeFetchStart + 8000);
+    const block = src.slice(zeFetchStart, zeFetchStart + 9000);
 
     const preCheckIdx = block.indexOf("resp.headers.get('content-length')");
-    const jsonParseIdx = block.indexOf('await resp.clone().json()');
+    const jsonParseIdx = block.indexOf('const bodyText = await resp.text()');
     expect(preCheckIdx).toBeGreaterThan(0);
     expect(jsonParseIdx).toBeGreaterThan(0);
     // The pre-check MUST appear before the JSON parse — Voyage's lesson

test/cli-source-scope-hint.test.ts

@@ -0,0 +1,61 @@
+/**
+ * #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();
+  });
+});

test/embed-stale.serial.test.ts

@@ -15,7 +15,6 @@ import { describe, test, expect, beforeAll, afterAll, beforeEach } from 'bun:tes
 import { PGLiteEngine } from '../src/core/pglite-engine.ts';
 import { resetPgliteState } from './helpers/reset-pglite.ts';
 import { embedStaleForSource } from '../src/core/embed-stale.ts';
-import { configureGateway, resetGateway } from '../src/core/ai/gateway.ts';
 import type { ChunkInput } from '../src/core/types.ts';
 
 let engine: PGLiteEngine;
@@ -277,49 +276,4 @@ describe('embedStaleForSource', () => {
     // The stale text row actually got its embedding.
     expect(txtRow.embedded_at).not.toBeNull();
   });
-
-  // #1717: the backfill path must label re-embedded chunks with the model
-  // that produced the vector, and preserve the existing label on chunks it
-  // did not touch (before the fix, both were reset to the engine default).
-  test('labels re-embedded chunks with the gateway model, preserves untouched labels (#1717)', async () => {
-    configureGateway({
-      embedding_model: 'openai:text-embedding-3-large',
-      env: { OPENAI_API_KEY: 'sk-test-embed-stale-1717' },
-    });
-    try {
-      await engine.putPage('notes/model-label', {
-        type: 'note',
-        title: 'model-label',
-        compiled_truth: '# model-label\n\nseeded',
-      });
-      await engine.upsertChunks('notes/model-label', [
-        {
-          chunk_index: 0,
-          chunk_text: 'already embedded elsewhere',
-          chunk_source: 'compiled_truth',
-          embedding: new Float32Array(1536).fill(0.01),
-          model: 'voyage:voyage-3',
-          token_count: 4,
-        },
-        {
-          chunk_index: 1,
-          chunk_text: 'stale chunk needing embed',
-          chunk_source: 'compiled_truth',
-          token_count: 5,
-          embedding: undefined, // stale
-        },
-      ]);
-
-      const result = await embedStaleForSource(engine, 'default', { embedFn: fakeEmbedFn });
-      expect(result.embedded).toBe(1);
-
-      const after = await engine.getChunks('notes/model-label');
-      const preserved = after.find((c) => c.chunk_index === 0)!;
-      const reembedded = after.find((c) => c.chunk_index === 1)!;
-      expect(reembedded.model).toBe('openai:text-embedding-3-large');
-      expect(preserved.model).toBe('voyage:voyage-3');
-    } finally {
-      resetGateway();
-    }
-  });
 });

test/embed.serial.test.ts

@@ -37,8 +37,6 @@ mock.module('../src/core/embedding.ts', () => ({
   // setPageEmbeddingSignature / invalidateStaleSignatureEmbeddings resolve to
   // null via the Proxy default, so the signature value is inert here.
   currentEmbeddingSignature: () => 'test:model:1536',
-  // #1717: embed paths stamp this label on (re)embedded chunks.
-  resolveEmbeddingModelLabel: () => 'openai:text-embedding-3-large',
 }));
 
 // Import AFTER mocking.
@@ -805,34 +803,3 @@ describe('embedAllStale --source threading (D7)', () => {
     expect((firstCallOpts as { sourceId?: string }).sourceId).toBe('media-corpus');
   });
 });
-
-// #1717: content_chunks.model must record the model that actually produced
-// each vector, not the gateway/engine default.
-describe('content_chunks.model labeling (#1717)', () => {
-  test('stamps the resolved embedding model on re-embedded chunks, preserves it on untouched chunks', async () => {
-    let upserted: any[] | undefined;
-    // Chunk 0 is stale (no embedded_at) → gets re-embedded this pass.
-    // Chunk 1 is already embedded with a DIFFERENT model → must be preserved,
-    // not relabeled to the current model.
-    const chunks = [
-      { chunk_index: 0, chunk_text: 'a', chunk_source: 'compiled_truth', embedded_at: null, model: 'zeroentropyai:zembed-1', token_count: 1 },
-      { chunk_index: 1, chunk_text: 'b', chunk_source: 'compiled_truth', embedded_at: '2026-01-01', embedding: new Float32Array(1536), model: 'voyage:voyage-3', token_count: 1 },
-    ];
-    const engine = mockEngine({
-      getPage: async () => ({ slug: 'notes/x', compiled_truth: 'a', timeline: '', source_id: 'default' }),
-      getChunks: async () => chunks,
-      upsertChunks: async (_slug: string, c: any[]) => { upserted = c; },
-      setPageEmbeddingSignature: async () => null,
-    });
-
-    await runEmbedCore(engine, { slugs: ['notes/x'] });
-
-    expect(upserted).toBeDefined();
-    const byIdx = Object.fromEntries(upserted!.map(c => [c.chunk_index, c]));
-    // Re-embedded chunk carries the model that produced its vector (was
-    // mislabeled with the default before the fix).
-    expect(byIdx[0].model).toBe('openai:text-embedding-3-large');
-    // Untouched chunk keeps its original model — no wholesale relabel.
-    expect(byIdx[1].model).toBe('voyage:voyage-3');
-  });
-});

test/hybrid-vector-arm-warn.serial.test.ts

@@ -0,0 +1,77 @@
+/**
+ * #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');
+  });
+});

test/import-signature-stamp.serial.test.ts

@@ -73,21 +73,4 @@ describe('importFromContent embedding_signature stamping (F1)', () => {
     await importFromContent(engine, 'concepts/unstamped', '# Unstamped\n\nbody content.', { noEmbed: true });
     expect(await signatureOf('concepts/unstamped')).toBeNull();
   });
-
-  // #1717: content_chunks.model must record the model that produced the
-  // vector (the configured gateway model), not the engine's hardcoded
-  // default. The gateway here is configured to openai:text-embedding-3-large,
-  // which differs from DEFAULT_EMBEDDING_MODEL — so this fails without the
-  // import-path model stamping.
-  test('inline embed labels content_chunks.model with the configured model (#1717)', async () => {
-    await importFromContent(engine, 'concepts/labeled', '# Labeled\n\nsome body content to chunk and embed.', {});
-    const rows = await engine.executeRaw<{ model: string }>(
-      `SELECT cc.model FROM content_chunks cc
-        JOIN pages p ON p.id = cc.page_id
-       WHERE p.slug = $1 AND p.source_id = 'default'`,
-      ['concepts/labeled'],
-    );
-    expect(rows.length).toBeGreaterThan(0);
-    for (const r of rows) expect(r.model).toBe('openai:text-embedding-3-large');
-  });
 });

test/search-op-source-scope.test.ts

@@ -0,0 +1,87 @@
+/**
+ * #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/,
+    );
+  });
+});

test/voyage-response-cap.test.ts

@@ -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 resp.clone().json() (D10 OOM defense)', async () => {
+  test('Layer 1: Content-Length pre-check fires BEFORE the body is read (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
@@ -47,8 +47,10 @@ 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 `await resp.clone().json()` for context.
-    const jsonParseIdx = inboundBlock.indexOf('const json: any = await resp.clone().json()');
+    // 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()');
     expect(preCheckIdx).toBeGreaterThan(0);
     expect(jsonParseIdx).toBeGreaterThan(0);
     // The pre-check MUST appear before the JSON parse — otherwise the OOM