feat(ai): add reranker touchpoint to LiteLLM proxy recipe (takeover of #2455)

LiteLLM normalizes Cohere/Voyage/Jina rerank backends to the wire shape
gateway.rerank() already speaks, so a reranker touchpoint on the litellm
recipe makes any proxied rerank model reachable via
`search.reranker.model litellm:<model>` with no adapter.

Repairs from the original PR:
- path is the LEAF '/rerank' (not '/v1/rerank'): LiteLLM serves both
  /rerank and /v1/rerank, and the recipe's setup_hint allows
  LITELLM_BASE_URL with or without the /v1 suffix — pinning '/v1/rerank'
  doubled to /v1/v1/rerank (404) on /v1-suffixed bases.
- setup_hint appends the rerank guidance to master's current line instead
  of replacing it with a stale pre-/v1-suffix version.
- cost_per_1m_tokens_usd stays undefined (pricing-unknown), matching the
  recipe's embedding/chat touchpoints and budget-tracker's deliberate
  litellm exclusion from the free-provider sets (a proxy can front a paid
  provider; the touchpoint field isn't consumed by rerank pricing anyway).

Test drives gateway.rerank()'s real URL builder via the stubbed transport
for both base-URL forms; the /v1-suffixed case fails with the original
PR's path.

Co-authored-by: ozp <ozp@users.noreply.github.com>
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

docs/architecture/KEY_FILES.md

@@ -79,7 +79,7 @@ per-release `**vX.Y.Z:**` narration — CI enforces this
 - `docs/architecture/RETRIEVAL.md` + `docs/architecture/RETRIEVAL_MAXPOOL_INCIDENT.md` — retrieval-pipeline architecture reference + the named-thing-miss incident write-up (root cause, the five-layer fix, the eval that pins it).
 - `src/core/types.ts` extension + `src/core/operations.ts:search` + `src/core/import-file.ts` + `src/cli.ts` + `src/core/search/telemetry.ts` — the wiring layer for the retrieval cathedral. `SearchResult` gains `evidence`, `create_safety`, `title_match_boost`, `alias_hit` (all optional; evidence/create_safety reference the union types in `evidence.ts`). The `search` MCP op uses a cheap-hybrid path by default and accepts a per-call `mode` (conservative|balanced|tokenmax) honored ONLY for trusted/local callers (`resolvePerCallMode(ctx, ...)` — remote callers use the configured mode so a remote provider can't force tokenmax spend); every search path stamps evidence fail-soft. `importFromContent` projects frontmatter `aliases:` into `page_aliases` via `normalizeAliasList` + `engine.setPageAliases` so new + changed pages register aliases at ingest. `src/cli.ts` adds the `gbrain search diagnose` dispatch (lazy import) and reconciles the `search` CLI path with the cheap-hybrid op. `src/core/search/telemetry.ts` extends the rollup with the rank-1 base_score drift signal (sum/count + 3 coarse buckets, aggregate not per-query), surfaced via `gbrain search stats`, backed by migration v111's `search_telemetry` columns. Tests: `test/cli-search-dispatch.test.ts`, `test/search/per-call-mode.test.ts`, `test/search/telemetry-rank1.test.ts`, `test/search/title-boost-stage.test.ts`, `test/search/alias-hop.test.ts`, `test/search/evidence.test.ts`, `test/search/searchvector-maxpool.test.ts`, `test/search/pre-migration-failopen.test.ts`.
 - `src/commands/eval.ts` — `gbrain eval` command: single-run table + A/B config comparison. Sub-subcommand dispatch on `args[0]` routes `gbrain eval export` + `gbrain eval prune` + `gbrain eval replay` into session-capture handlers; bare `gbrain eval --qrels …` fall-through preserves the legacy IR-metrics flow. `gbrain eval cross-modal` is in the dispatch (the user-facing path is the cli.ts no-DB branch — `src/commands/eval.ts:cross-modal` only fires when callers re-enter with an existing engine).
-- `src/commands/eval-cross-modal.ts` — multi-model quality gate. Three different-provider frontier models score the OUTPUT against the TASK on a 5-dim list. Verdict `pass` (exit 0) / `fail` (exit 1) / `inconclusive` (exit 2; <2/3 model successes). Reuses `src/core/ai/gateway.ts:chat()` so config/auth/aliasing comes from the gateway recipe registry — no parallel provider stack. Self-configures the gateway via `buildGatewayConfig(loadConfig() ?? {})` since the cli.ts dispatch bypasses `connectEngine()`, so file-plane keys, env base URLs, and provider_chat_options follow the same adapter path as runtime. Default cycles 3 in TTY, 1 in non-TTY (partial cost guardrail) via the shared `resolveCycleDefault(explicit, isTty)` in `src/core/eval/cycle-default.ts`; the cost-estimate banner appends `cycleDefaultSuffix(...)` (`for 1 cycle(s) (non-interactive default; --cycles N for more)`) when the value is the silent non-TTY fallback, so the 1-vs-3 difference isn't hidden. Receipts land at `gbrainPath('eval-receipts')/<slug>-<sha8-of-output>.json`. `--batch <jsonl> [--limit N] [--concurrent N] [--max-usd FLOAT] [--yes]` fans out cross-modal scoring across a LongMemEval-shape JSONL; mutually exclusive with `--task` (fail-fast usage error if both set); filters `kind: "by_type_summary"` rows; pre-flight cost estimate refuses if `> --max-usd` without `--yes` (default cap 5.00 USD). Semaphore-bounded fan-out via inline `runWithLimit<T>(items, limit, fn)` (exported for unit tests): max N questions in-flight × 3 model slots = ceiling of 3N parallel API calls (default `--concurrent 3` → 9). Per-question receipts land in a per-batch tempdir and are deleted at end of run; the summary receipt inlines per-question verdicts as JSON, not file paths. Exit precedence (batch-level policy, NOT inherited from aggregate.ts): ERROR > FAIL > INCONCLUSIVE > PASS. DI seam: `runEvalCrossModal(args, opts?: {runEval?: typeof runEval})` mirrors `runEvalLongMemEval(args, {client?})`; tests pass `opts.runEval` to bypass real LLM calls AND the gateway availability check. Pinned by `test/eval-cross-modal-batch.test.ts`.
+- `src/commands/eval-cross-modal.ts` — multi-model quality gate. Three different-provider frontier models score the OUTPUT against the TASK on a 5-dim list. Verdict `pass` (exit 0) / `fail` (exit 1) / `inconclusive` (exit 2; <2/3 model successes). Reuses `src/core/ai/gateway.ts:chat()` so config/auth/aliasing comes from the gateway recipe registry — no parallel provider stack. Self-configures the gateway (`configureGateway(loadConfig() + process.env)`) since the cli.ts dispatch bypasses `connectEngine()`. Default cycles 3 in TTY, 1 in non-TTY (partial cost guardrail) via the shared `resolveCycleDefault(explicit, isTty)` in `src/core/eval/cycle-default.ts`; the cost-estimate banner appends `cycleDefaultSuffix(...)` (`for 1 cycle(s) (non-interactive default; --cycles N for more)`) when the value is the silent non-TTY fallback, so the 1-vs-3 difference isn't hidden. Receipts land at `gbrainPath('eval-receipts')/<slug>-<sha8-of-output>.json`. `--batch <jsonl> [--limit N] [--concurrent N] [--max-usd FLOAT] [--yes]` fans out cross-modal scoring across a LongMemEval-shape JSONL; mutually exclusive with `--task` (fail-fast usage error if both set); filters `kind: "by_type_summary"` rows; pre-flight cost estimate refuses if `> --max-usd` without `--yes` (default cap 5.00 USD). Semaphore-bounded fan-out via inline `runWithLimit<T>(items, limit, fn)` (exported for unit tests): max N questions in-flight × 3 model slots = ceiling of 3N parallel API calls (default `--concurrent 3` → 9). Per-question receipts land in a per-batch tempdir and are deleted at end of run; the summary receipt inlines per-question verdicts as JSON, not file paths. Exit precedence (batch-level policy, NOT inherited from aggregate.ts): ERROR > FAIL > INCONCLUSIVE > PASS. DI seam: `runEvalCrossModal(args, opts?: {runEval?: typeof runEval})` mirrors `runEvalLongMemEval(args, {client?})`; tests pass `opts.runEval` to bypass real LLM calls AND the gateway availability check. Pinned by `test/eval-cross-modal-batch.test.ts`.
 - `src/core/eval/cycle-default.ts` — single source of truth for the eval cycle-count default. Exports `DEFAULT_CYCLES_TTY = 3`, `DEFAULT_CYCLES_NONTTY = 1`, `resolveCycleDefault(explicit, isTty): {cycles, usedNonTtyDefault}`, and `cycleDefaultSuffix(r)` (returns ` (non-interactive default; --cycles N for more)` only when the non-TTY default was applied, else `''`). Consumed by `eval-cross-modal.ts`, `eval-takes-quality.ts` (run + regress), and `takes-quality-eval/runner.ts` (core uses only the constant — library stays TTY-agnostic; the CLI owns the TTY=3 upgrade + banner annotation). `eval-suspected-contradictions.ts` applies the same transparency to its `$5`/`$1` budget default via a `budgetUsdExplicit` flag (the budget is overwritten in-place so explicitness can't be inferred post-hoc). Not shared with `resolveWorkersWithClamp` (different domain, no engine, no dedup). Pinned by `test/eval/cycle-default.test.ts`, `test/eval-suspected-contradictions-budget-default.test.ts`.
 - `src/core/cross-modal-eval/json-repair.ts` — `parseModelJSON(raw)` named export with a 4-strategy fallback chain (direct parse → fence-strip → trailing-comma + single-quote + embedded-newline repair → regex nuclear option). Adversarial input throws rather than fabricating scores — the aggregator treats a throw as "this model contributed nothing this cycle" so the gate stays correct at >=2/3 successes.
 - `src/core/cross-modal-eval/aggregate.ts` — pure verdict logic. Pass criterion: `(successes >= 2) AND (every dim mean >= 7) AND (every dim min across models >= 5)`. Inconclusive when <2/3 models returned parseable scores (regression guard for the v1 `Object.values({}).every(...) === true` empty-array PASS bug).
@@ -138,7 +138,7 @@ per-release `**vX.Y.Z:**` narration — CI enforces this
 - `src/core/minions/queue.ts` extension — `MinionQueue.add()` rejects `subagent` jobs whose `data.model` resolves via `isAnthropicProvider()` to a non-Anthropic provider. Lazy-imports `model-config.ts` to avoid pulling engine types into queue's eager-load surface. Layer 1 of the three-layer subagent provider enforcement (layers 2+3: `model-config.ts:enforceSubagentAnthropic` runtime fallback + `src/commands/doctor.ts` `subagent_provider` check). Pinned by `test/agent-cli.test.ts`.
 - `src/commands/models.ts` — `gbrain models [--json]` read-only routing dashboard: prints tier defaults (`utility`/`reasoning`/`deep`/`subagent`), the resolved value for each (re-walking the resolution chain), every per-task override (11 `PER_TASK_KEYS`: `models.dream.synthesize`, `models.dream.patterns`, `models.drift`, `models.auto_think`, `models.think`, `models.subagent`, `facts.extraction_model`, `models.eval.longmemeval`, `models.expansion`, `models.chat`, `models.dream.synthesize_verdict`), the alias map, and a source-of-truth column (`default` / `config: <key>` / `env: <VAR>`). `gbrain models doctor [--skip=<provider>] [--json]` fires a 1-token `gateway.chat()` probe against each configured chat + expansion model and classifies failures into `{model_not_found, auth, rate_limit, network, unknown}`. Wired into `cli.ts` dispatch + `CLI_ONLY` set. A zero-token `embedding_config` probe runs FIRST, before any chat/expansion probes spend money: `probeEmbeddingConfig()` reads `getEmbeddingModel()` + `getEmbeddingDimensions()` and (for Voyage flexible-dim models) checks `isValidVoyageOutputDim(dims)` against `VOYAGE_VALID_OUTPUT_DIMS`. `ProbeStatus` variant `'config'` + optional `fix?: string` on `ProbeResult` surface a paste-ready `gbrain config set ...` line in human + JSON output; touchpoint label `'embedding_config'` joins `'chat'` and `'expansion'`.
 - `src/core/init-embed-check.ts` — embedding-key validation at `gbrain init`. `runInitEmbedCheck(opts)` runs a config-only `diagnoseEmbedding` (catches a missing key for ANY provider) plus a best-effort `liveTestEmbed` (1-token `gateway.embed(['probe'], {inputType:'query', abortSignal})`, 5s `AbortController` timeout, never throws — catches an invalid/expired key). Loud warning to stderr; init still exits 0 (`--no-embedding` is the deferred-setup escape; `--skip-embed-check` / `GBRAIN_INIT_SKIP_EMBED_CHECK=1` skip the check). Builds the effective env (`process.env` + file-plane `openai/anthropic/zeroentropy_api_key` from `loadConfigFileOnly()` + `opts.apiKey`) and configures the gateway via `buildGatewayConfig` before diagnose/probe, so the check sees the same keys AND provider base URLs runtime will (no false "missing key" for config.json-keyed users; the probe hits the right endpoint). Init-specific warning text names `--no-embedding` / `--skip-embed-check`, not the sync-flavored `--no-embed`. Wired into `initPGLite` + `initPostgres` in `src/commands/init.ts`, with the result added to the `--json` envelope as `embedding_check {ok, reason?, live_ok?}`. Pinned by `test/init-embed-check.test.ts` (hermetic via the gateway embed-transport seam + `withEnv`).
-- `src/core/ai/build-gateway-config.ts` — `buildGatewayConfig(c: GBrainConfig): AIGatewayConfig`, extracted from `src/cli.ts` (which re-exports it for back-compat). Single owner of translating stored config into gateway config — consumed by CLI runtime, init (`gbrain init`'s three configureGateway sites), `init-embed-check.ts`, the eval commands (cross-modal, takes-quality), provider diagnostics, and the in-process migration path. Folds file-plane API keys (openai/anthropic/zeroentropy/openrouter) into the gateway env and threads local-server `*_BASE_URL` env vars into base_urls; caller-provided `provider_base_urls` config wins over env base URLs. `process.env` wins EXCEPT empty-string / undefined values are dropped before the merge, so an injected empty `ANTHROPIC_API_KEY=''` (Claude Code neuters subprocess LLM calls this way) can't clobber a valid config-plane key; `'0'` / `'false'` are preserved. Pinned by `test/ai/build-gateway-config.test.ts`.
+- `src/core/ai/build-gateway-config.ts` — `buildGatewayConfig(c: GBrainConfig): AIGatewayConfig`, extracted from `src/cli.ts` (which re-exports it for back-compat). Lets core modules (`init-embed-check.ts`) reuse it without importing the CLI entrypoint. Single owner of folding file-plane API keys (openai/anthropic/zeroentropy) into the gateway env and threading local-server `*_BASE_URL` env vars into base_urls. `process.env` wins EXCEPT empty-string / undefined values are dropped before the merge, so an injected empty `ANTHROPIC_API_KEY=''` (Claude Code neuters subprocess LLM calls this way) can't clobber a valid config-plane key; `'0'` / `'false'` are preserved. Pinned by `test/ai/build-gateway-config.test.ts`.
 - `src/commands/doctor.ts` extension — `subagent_provider` check (layer 3 of 3). Warns when `models.tier.subagent` is explicitly set non-Anthropic (message names the bad value + paste-ready fix `gbrain config set models.tier.subagent anthropic:claude-sonnet-4-6`); also warns when `models.default` would sneak `subagent` into a non-Anthropic provider via tier inheritance. OK when subagent tier resolves to Anthropic. Tests in `test/doctor.test.ts`.
 - `src/core/skill-trigger-index.ts` — Shared loader that unions per-skill SKILL.md frontmatter `triggers:` with curated RESOLVER.md / AGENTS.md rows from `skillsDir` AND the parent dir (preserves the OpenClaw workspace-root layout). UNION semantics: explicit RESOLVER.md rows ADD to frontmatter triggers (don't replace). Dedup keyed on `(skillPath, trigger.trim().toLowerCase())`. Three consumers fold through this primitive — `checkResolvable`, `runRoutingEvalCli`, `mounts-cache.composeResolvers` — so fixing frontmatter reaches all of them. Exports `loadSkillTriggerIndex(skillsDir): SkillTriggerEntry[]`, `entriesToResolverContent(entries): string` (synthesizes a markdown-table resolver string for `runRoutingEval`'s string-content API), `findPrimaryResolverPath(skillsDir): string | null`, the `FRONTMATTER_SECTION` constant, and `_resetWarnedSkillsForTests`. Skip rules: non-directory entries, `_*`/`.*` prefixes, `conventions/`+`migrations/` subdirs, skills with no `SKILL.md` (deprecated `install/` graceful-skipped), no `triggers:` array, or malformed YAML (warn-once + skip). Reuses `parseSkillFrontmatter` from `src/core/skill-frontmatter.ts` (regex-based, not full YAML). Pinned by `test/skill-trigger-index.test.ts` (18 hermetic cases). CI gate `bun run check:resolver` (= `bun src/cli.ts check-resolvable --strict --skills-dir skills/`) wired into `bun run verify`.
 - `src/core/skill-catalog.ts` — host-repo skill catalog backing the MCP `list_skills` / `get_skill` ops. Lets a thin MCP client (Codex desktop, Claude Code, Claude Cowork, Perplexity) DISCOVER + FOLLOW the agent repo's fat-markdown skills over `gbrain serve` — a skill is prose, so "using" one = fetching its body then calling the gbrain MCP tools the server already exposes. Read-scope, NOT localOnly (defensible only via the full mitigation stack): (1) **publish gate** — `assertPublishEnabled(ctx, publishSkills)`; remote callers require `mcp.publish_skills === true`, default-OFF so an upgrade never silently grants existing read tokens host-skill read; local callers (`ctx.remote === false`) always pass. (2) **path confinement** — `assertSkillNameShape` rejects separators/`..`/null/space before any FS access; the client `name` is a manifest LOOKUP KEY (via `loadOrDeriveManifest`), never a raw path segment; `confineManifestPath` does realpath + relative-containment + `SKILL.md`-regular-file check on EVERY entry (defeats poisoned manifest.json `path`, symlink/`..` escape). (3) **frontmatter allowlist** — `GetSkillResult.frontmatter` projects a safe subset; private `writes_to` + `sources` dropped. (4) **prose-only + 256KB cap** (`MAX_SKILL_MD_BYTES`, env `GBRAIN_MAX_SKILL_MD_BYTES`), size-checked twice (statSync + UTF-8 byte length). (5) **no install_path serve for remote** — remote callers use `autoDetectSkillsDir` (no install-path tier) so a hosted gbrain with no agent repo returns `storage_error`; local callers use `autoDetectSkillsDirReadOnly`. (6) MCP rate-limiter caps call rate. Config reads honor BOTH planes: `readMcpPublishSkills` / `readMcpSkillsDir` prefer the DB plane (`engine.getConfig`) over the file plane (`ctx.config.mcp`). Tool-honesty: `crossReferenceTools(declared, ctx)` splits a skill's declared `tools:` into `usable_tools` vs `unavailable_tools`; `buildSkillCatalog`'s `instructions` envelope (`SKILL_CATALOG_INSTRUCTIONS`) carries the "these are prose, follow-then-call-tools" protocol. Skills are host-filesystem repo-global — `sourceScopeOpts(ctx)` / `ctx.brainId` deliberately do NOT apply. `buildSkillCatalog` is resilient (one malformed/escaping skill is skipped, never throws). Config keys in `src/core/config.ts`: `GBrainConfig.mcp?: { publish_skills?, skills_dir? }` + `KNOWN_CONFIG_KEYS` entries `mcp.publish_skills`/`mcp.publish_skills_prompted`/`mcp.skills_dir` + `mcp.` prefix in `KNOWN_CONFIG_KEY_PREFIXES`. `src/commands/init.ts` writes `config.mcp = { publish_skills: true, ... }` for new installs (existing config wins on re-init). `src/commands/upgrade.ts:runPostUpgrade` adds a one-time consent prompt (gated by `mcp.publish_skills_prompted`; existing installs stay OFF until owner opts in). Two ops register in `src/core/operations.ts` (`list_skills` with optional `section` filter + `cliHints:{name:'skills'}`; `get_skill` taking `name` + `cliHints:{name:'skill', positional:['name']}`) and dynamically import this module to avoid the import cycle (skill-catalog statically imports the `operations` array). Descriptions in `src/core/operations-descriptions.ts` (`LIST_SKILLS_DESCRIPTION`, `GET_SKILL_DESCRIPTION`, `SKILL_CATALOG_INSTRUCTIONS`, `SKILL_CLIENT_GUIDANCE`), pinned by `test/operations-descriptions.test.ts`. CLI: `gbrain skills` / `gbrain skill <name>`. Pinned by `test/skill-catalog.test.ts`, `test/skill-catalog-security.test.ts` (path-confinement / poisoned-manifest / symlink-escape), `test/skill-catalog-transports.test.ts` (publish-gate + remote-vs-local) over `test/fixtures/skill-catalog/`.

src/commands/eval-cross-modal.ts

@@ -21,8 +21,7 @@ import { join } from 'path';
 import { tmpdir } from 'os';
 import { createHash } from 'crypto';
 
-import { gbrainPath, loadConfig, type GBrainConfig } from '../core/config.ts';
-import { buildGatewayConfig } from '../core/ai/build-gateway-config.ts';
+import { gbrainPath, loadConfig } from '../core/config.ts';
 import { configureGateway, isAvailable } from '../core/ai/gateway.ts';
 import { runWithLimit } from '../core/worker-pool.ts';
 import { resolveCycleDefault, cycleDefaultSuffix } from '../core/eval/cycle-default.ts';
@@ -265,11 +264,32 @@ function isTTY(): boolean {
  * Returns true on success; false (and prints a hint) when no config is found.
  */
 function configureGatewayForCli(): boolean {
-  // Route through buildGatewayConfig (the single adapter seam) so file-plane
-  // API keys, env base URLs, and provider_chat_options follow the same
-  // precedence as the runtime path. No config file is fine — env alone serves.
   const config = loadConfig();
-  configureGateway(buildGatewayConfig(config ?? ({} as GBrainConfig)));
+  if (!config) {
+    // No config file is fine for the eval command — env vars alone may serve.
+    // We still call configureGateway so gateway recipes can read the env map.
+    configureGateway({
+      embedding_model: undefined,
+      embedding_dimensions: undefined,
+      expansion_model: undefined,
+      chat_model: undefined,
+      chat_fallback_chain: undefined,
+      base_urls: undefined,
+      provider_chat_options: undefined,
+      env: { ...process.env },
+    });
+    return true;
+  }
+  configureGateway({
+    embedding_model: config.embedding_model,
+    embedding_dimensions: config.embedding_dimensions,
+    expansion_model: config.expansion_model,
+    chat_model: config.chat_model,
+    chat_fallback_chain: config.chat_fallback_chain,
+    base_urls: config.provider_base_urls,
+    provider_chat_options: config.provider_chat_options,
+    env: { ...process.env },
+  });
   return true;
 }
 

src/commands/eval-takes-quality.ts

@@ -21,8 +21,7 @@
  */
 import type { BrainEngine } from '../core/engine.ts';
 import { configureGateway } from '../core/ai/gateway.ts';
-import { buildGatewayConfig } from '../core/ai/build-gateway-config.ts';
-import { loadConfig, type GBrainConfig } from '../core/config.ts';
+import { loadConfig } from '../core/config.ts';
 import { runEval, DEFAULT_MODEL_PANEL } from '../core/takes-quality-eval/runner.ts';
 import { resolveCycleDefault, cycleDefaultSuffix } from '../core/eval/cycle-default.ts';
 import { writeReceipt } from '../core/takes-quality-eval/receipt-write.ts';
@@ -128,12 +127,8 @@ export async function runReplayNoBrain(argv: string[]): Promise<number> {
 export async function runEvalTakesQuality(engine: BrainEngine, args: string[]): Promise<void> {
   // Self-configure the AI gateway (mirrors eval-cross-modal pattern). The
   // gateway needs config.ai_gateway + env vars; configureGateway reads both.
-  // Route through buildGatewayConfig: the old `{ ...cfg, ...process.env }`
-  // spread never populated the gateway's `env` field (the gateway NEVER reads
-  // process.env at call time), so availability checks saw no keys at all and
-  // file-plane API keys / provider base URLs were dropped.
   const cfg = loadConfig();
-  configureGateway(buildGatewayConfig(cfg ?? ({} as GBrainConfig)));
+  configureGateway({ ...cfg, ...(process.env as Record<string, string>) } as any);
 
   const { subcmd, argv, json } = parseSubcmd(args);
 

src/commands/init.ts

@@ -7,7 +7,6 @@ import { homedir } from 'os';
 const __filename = fileURLToPath(import.meta.url);
 const __dirname = dirname(__filename);
 import { saveConfig, loadConfig, loadConfigFileOnly, toEngineConfig, gbrainPath, configPath, isThinClient, effectiveEnvDatabaseUrl, type GBrainConfig } from '../core/config.ts';
-import { buildGatewayConfig } from '../core/ai/build-gateway-config.ts';
 import { createEngine } from '../core/engine-factory.ts';
 import { discoverOAuth, mintClientCredentialsToken, smokeTestMcp } from '../core/remote-mcp-probe.ts';
 import { runInitEmbedCheck } from '../core/init-embed-check.ts';
@@ -723,8 +722,7 @@ async function configureGatewayWithMergedPrecedence(
   // pollutes config.json.
   const envOverlay = loadConfig() ?? ({} as GBrainConfig);
 
-  const merged: GBrainConfig = {
-    ...envOverlay,
+  const merged = {
     embedding_model: aiOpts?.embedding_model ?? envOverlay.embedding_model ?? existingFile.embedding_model,
     embedding_dimensions: aiOpts?.embedding_dimensions ?? envOverlay.embedding_dimensions ?? existingFile.embedding_dimensions,
     expansion_model: aiOpts?.expansion_model ?? envOverlay.expansion_model ?? existingFile.expansion_model,
@@ -732,9 +730,13 @@ async function configureGatewayWithMergedPrecedence(
   };
 
   const { configureGateway, getEmbeddingModel, getEmbeddingDimensions, getExpansionModel, getChatModel } = await import('../core/ai/gateway.ts');
-  // buildGatewayConfig (the single adapter seam) so file-plane API keys and
-  // env base URLs reach the gateway — a hand-rolled config here dropped them.
-  configureGateway(buildGatewayConfig(merged));
+  configureGateway({
+    embedding_model: merged.embedding_model,
+    embedding_dimensions: merged.embedding_dimensions,
+    expansion_model: merged.expansion_model,
+    chat_model: merged.chat_model,
+    env: { ...process.env },
+  });
 
   // Read back resolved values — gateway applies internal defaults for unset
   // fields, so these are the values that actually shaped the schema.
@@ -829,13 +831,13 @@ async function initPGLite(opts: {
   // resolveAIOptions above: CLI flags > env vars > existing file > gateway
   // defaults.
   const { configureGateway } = await import('../core/ai/gateway.ts');
-  configureGateway(buildGatewayConfig({
-    ...(loadConfig() ?? ({} as GBrainConfig)),
+  configureGateway({
     embedding_model: resolvedModel ?? opts.aiOpts?.embedding_model,
     embedding_dimensions: resolvedDim ?? opts.aiOpts?.embedding_dimensions,
     expansion_model: opts.aiOpts?.expansion_model,
     chat_model: opts.aiOpts?.chat_model,
-  } as GBrainConfig));
+    env: { ...process.env },
+  });
   if (resolvedModel) console.log(`  Embedding: ${resolvedModel} (${resolvedDim}d)`);
   if (opts.aiOpts?.expansion_model) console.log(`  Expansion: ${opts.aiOpts.expansion_model}`);
   if (opts.aiOpts?.chat_model) console.log(`  Chat: ${opts.aiOpts.chat_model}`);
@@ -1044,13 +1046,13 @@ async function initPostgres(opts: {
 
   // T6: unconditional configureGateway BEFORE initSchema.
   const { configureGateway } = await import('../core/ai/gateway.ts');
-  configureGateway(buildGatewayConfig({
-    ...(loadConfig() ?? ({} as GBrainConfig)),
+  configureGateway({
     embedding_model: resolvedModel ?? opts.aiOpts?.embedding_model,
     embedding_dimensions: resolvedDim ?? opts.aiOpts?.embedding_dimensions,
     expansion_model: opts.aiOpts?.expansion_model,
     chat_model: opts.aiOpts?.chat_model,
-  } as GBrainConfig));
+    env: { ...process.env },
+  });
   if (resolvedModel) console.log(`  Embedding: ${resolvedModel} (${resolvedDim}d)`);
   if (opts.aiOpts?.expansion_model) console.log(`  Expansion: ${opts.aiOpts.expansion_model}`);
   if (opts.aiOpts?.chat_model) console.log(`  Chat: ${opts.aiOpts.chat_model}`);

src/commands/migrations/in-process.ts

@@ -66,11 +66,15 @@ export async function runMigrateOnlyCore(opts?: { timeoutMs?: number }): Promise
 
   // configureGateway BEFORE initSchema (init.ts B.3): a schema bump on a brain
   // whose file config is missing embedding fields must not fall through to
-  // stale hardcoded fallbacks. Route through buildGatewayConfig so file-plane
-  // API keys and provider base URLs follow the same precedence as runtime.
+  // stale hardcoded fallbacks. loadConfig already merged env; propagate it.
   const { configureGateway } = await import('../../core/ai/gateway.ts');
-  const { buildGatewayConfig } = await import('../../core/ai/build-gateway-config.ts');
-  configureGateway(buildGatewayConfig(config));
+  configureGateway({
+    embedding_model: config.embedding_model,
+    embedding_dimensions: config.embedding_dimensions,
+    expansion_model: config.expansion_model,
+    chat_model: config.chat_model,
+    env: { ...process.env },
+  });
 
   const timeoutMs = opts?.timeoutMs ?? MIGRATE_ONLY_TIMEOUT_MS;
   const engine = await createEngine(toEngineConfig(config));

src/core/ai/recipes/litellm-proxy.ts

@@ -56,6 +56,36 @@ export const litellmProxy: Recipe = {
       cost_per_1m_output_usd: undefined,
       price_last_verified: '2026-06-14',
     },
+    // LiteLLM normalizes Cohere / Voyage / Jina / etc. rerank backends to the
+    // same wire shape gbrain's gateway.rerank() already speaks (the
+    // ZeroEntropy/llama.cpp contract):
+    //   { model, query, documents, top_n } → { results: [{ index, relevance_score }] }
+    // So any rerank model the user registers in their LiteLLM config is
+    // reachable via `gbrain config set search.reranker.model litellm:<model>`
+    // with no request/response adapter — same as embeddings ride the proxy.
+    reranker: {
+      models: [], // user-provided; whatever rerank models the proxy serves
+      // No canonical default — the proxy defines its own model ids. The user
+      // sets search.reranker.model explicitly (mirrors the embedding
+      // touchpoint's user_provided_models contract).
+      default_model: '',
+      // The proxied backend bills (Cohere/Voyage/…); pricing-unknown is the
+      // honest state — same stance as this recipe's embedding/chat
+      // touchpoints and budget-tracker's deliberate litellm exclusion from
+      // the free-provider sets.
+      cost_per_1m_tokens_usd: undefined,
+      price_last_verified: '2026-06-27',
+      max_payload_bytes: 5_000_000,
+      // LEAF path only (matches llama-server-reranker's convention). LiteLLM
+      // serves both `/rerank` and `/v1/rerank`, and LITELLM_BASE_URL may be
+      // set with or without the `/v1` suffix (the setup_hint allows both), so
+      // the leaf form yields a valid route either way:
+      //   http://localhost:4000    + /rerank → /rerank        ✓
+      //   http://localhost:4000/v1 + /rerank → /v1/rerank     ✓
+      // Pinning '/v1/rerank' here would double to /v1/v1/rerank → 404 on
+      // /v1-suffixed bases.
+      path: '/rerank',
+    },
   },
-  setup_hint: 'Run LiteLLM (https://docs.litellm.ai) in front of any provider; set LITELLM_BASE_URL (include the /v1 suffix if your proxy serves the OpenAI route there, e.g. http://localhost:4000/v1) + pass --embedding-model litellm:<model> and --embedding-dimensions <N>.',
+  setup_hint: 'Run LiteLLM (https://docs.litellm.ai) in front of any provider; set LITELLM_BASE_URL (include the /v1 suffix if your proxy serves the OpenAI route there, e.g. http://localhost:4000/v1) + pass --embedding-model litellm:<model> and --embedding-dimensions <N>. For rerank: register a rerank model in LiteLLM and set search.reranker.model litellm:<model-name>.',
 };

src/core/brain-score-recommendations.ts

@@ -12,8 +12,8 @@ import { parseModelId } from './ai/model-resolver.ts';
  * `resolveKey` closure without re-parsing recipes.
  *
  * Only OPENAI_API_KEY and ZEROENTROPY_API_KEY appear here because those are the
- * only embedding keys `buildGatewayConfig` (src/core/ai/build-gateway-config.ts)
- * folds from config into the gateway env. VOYAGE_API_KEY / GOOGLE_GENERATIVE_AI_API_KEY are deliberately
+ * only embedding keys `buildGatewayConfig` (src/cli.ts) folds from config into
+ * the gateway env. VOYAGE_API_KEY / GOOGLE_GENERATIVE_AI_API_KEY are deliberately
  * absent: their config fields are NOT threaded to the gateway today, so the
  * producer closures fall through to checking `process.env` ONLY for them. That
  * matches what the gateway can actually use (the recipes read those keys from

test/ai/build-gateway-config.test.ts

@@ -25,7 +25,6 @@ import { withEnv } from '../helpers/with-env.ts';
 
 const PASSTHROUGHS: Array<{ envVar: string; recipeId: string }> = [
   { envVar: 'LLAMA_SERVER_BASE_URL', recipeId: 'llama-server' },
-  { envVar: 'LLAMA_SERVER_RERANKER_BASE_URL', recipeId: 'llama-server-reranker' },
   { envVar: 'OLLAMA_BASE_URL', recipeId: 'ollama' },
   { envVar: 'LMSTUDIO_BASE_URL', recipeId: 'lmstudio' },
   { envVar: 'LITELLM_BASE_URL', recipeId: 'litellm' },

test/ai/recipe-litellm-reranker.test.ts

@@ -0,0 +1,88 @@
+/**
+ * litellm-proxy reranker touchpoint smoke.
+ *
+ * Sibling of recipe-llama-server-reranker.test.ts. Pins the reranker
+ * touchpoint on the LiteLLM proxy recipe so:
+ *  - the touchpoint exists with the LEAF '/rerank' path (LiteLLM serves both
+ *    /rerank and /v1/rerank, so the leaf form is valid whether or not the
+ *    user's LITELLM_BASE_URL carries the /v1 suffix the setup_hint allows)
+ *  - a /v1-suffixed base URL does NOT produce /v1/v1/rerank (the original
+ *    community PR pinned '/v1/rerank' which 404s on /v1-suffixed bases)
+ *  - models: [] (user-provided; proxy defines the model ids)
+ *  - pricing stays undefined (proxy can front a paid provider — same honest
+ *    pricing-unknown stance as the embedding/chat touchpoints)
+ *
+ * The gateway.rerank() URL tests drive the real URL builder via the stubbed
+ * transport (same seam as test/ai/rerank.test.ts).
+ */
+
+import { describe, expect, test, afterEach } from 'bun:test';
+import { getRecipe } from '../../src/core/ai/recipes/index.ts';
+import {
+  configureGateway,
+  resetGateway,
+  rerank,
+  __setRerankTransportForTests,
+} from '../../src/core/ai/gateway.ts';
+
+afterEach(() => {
+  __setRerankTransportForTests(null);
+  resetGateway();
+});
+
+describe('recipe: litellm reranker touchpoint', () => {
+  test('declares reranker touchpoint with leaf /rerank path', () => {
+    const r = getRecipe('litellm')!;
+    const tp = r.touchpoints.reranker;
+    expect(tp).toBeDefined();
+    expect(tp!.path).toBe('/rerank');
+    expect(tp!.max_payload_bytes).toBe(5_000_000);
+  });
+
+  test('reranker touchpoint uses empty models[] for user-provided model ids', () => {
+    const r = getRecipe('litellm')!;
+    expect(r.touchpoints.reranker!.models).toEqual([]);
+  });
+
+  test('pricing stays undefined — proxy can front a paid provider', () => {
+    const r = getRecipe('litellm')!;
+    expect(r.touchpoints.reranker!.cost_per_1m_tokens_usd).toBeUndefined();
+  });
+
+  test('setup_hint keeps the /v1-suffix guidance AND mentions rerank', () => {
+    const r = getRecipe('litellm')!;
+    expect(r.setup_hint).toMatch(/\/v1 suffix/);
+    expect(r.setup_hint).toMatch(/search\.reranker\.model litellm:/);
+  });
+});
+
+describe('gateway.rerank() URL via litellm recipe', () => {
+  async function capturedRerankUrl(baseUrl?: string): Promise<string> {
+    configureGateway({
+      reranker_model: 'litellm:my-reranker',
+      env: {},
+      ...(baseUrl ? { base_urls: { litellm: baseUrl } } : {}),
+    });
+    let capturedUrl = '';
+    __setRerankTransportForTests(async (url) => {
+      capturedUrl = url;
+      return new Response(
+        JSON.stringify({ results: [{ index: 0, relevance_score: 0.9 }] }),
+        { status: 200, headers: { 'content-type': 'application/json' } },
+      );
+    });
+    await rerank({ query: 'q', documents: ['d'] });
+    return capturedUrl;
+  }
+
+  test('default base (no /v1 suffix) → /rerank', async () => {
+    const url = await capturedRerankUrl();
+    expect(url).toBe('http://localhost:4000/rerank');
+  });
+
+  test('/v1-suffixed base → /v1/rerank, NOT /v1/v1/rerank', async () => {
+    const url = await capturedRerankUrl('http://localhost:4000/v1');
+    expect(url).toBe('http://localhost:4000/v1/rerank');
+    expect(url).not.toContain('/v1/v1/');
+  });
+});

test/cli-multimodal-integration.test.ts

@@ -11,13 +11,32 @@
 import { afterAll, beforeAll, beforeEach, describe, expect, test } from 'bun:test';
 import { PGLiteEngine } from '../src/core/pglite-engine.ts';
 import { loadConfigWithEngine, type GBrainConfig } from '../src/core/config.ts';
-import { buildGatewayConfig } from '../src/core/ai/build-gateway-config.ts';
 import {
   configureGateway,
   getEmbeddingModel,
   getMultimodalModel,
   resetGateway,
 } from '../src/core/ai/gateway.ts';
+import type { AIGatewayConfig } from '../src/core/ai/types.ts';
+
+// Mirror the cli.ts buildGatewayConfig helper exactly. Keeping a copy here
+// (instead of exporting from cli.ts) is intentional: the test asserts the
+// shape of the contract, not the helper's identity. If cli.ts drifts, the
+// e2e behavior these tests care about (DB-set value lands in gateway) still
+// holds, but a helper-shape test would also catch the drift in PR review.
+function buildGatewayConfig(c: GBrainConfig): AIGatewayConfig {
+  return {
+    embedding_model: c.embedding_model,
+    embedding_dimensions: c.embedding_dimensions,
+    embedding_multimodal_model: c.embedding_multimodal_model,
+    expansion_model: c.expansion_model,
+    chat_model: c.chat_model,
+    chat_fallback_chain: c.chat_fallback_chain,
+    base_urls: c.provider_base_urls,
+    provider_chat_options: c.provider_chat_options,
+    env: { ...process.env },
+  };
+}
 
 let engine: PGLiteEngine;
 
@@ -28,7 +47,6 @@ beforeAll(async () => {
 });
 
 afterAll(async () => {
-  resetGateway(); // don't leak this file's gateway config into shard siblings
   await engine.disconnect();
 });
 

test/eval-takes-quality-gateway.test.ts

@@ -1,57 +0,0 @@
-/**
- * eval-takes-quality gateway self-config — adapter-boundary regression
- * (takeover of PR #2430).
- *
- * The old callsite spread `{ ...cfg, ...process.env }` straight into
- * configureGateway. The gateway NEVER reads process.env at call time — it
- * reads `_config.env` — and that spread never populated an `env` field at
- * all, so every availability/diagnose check dereferenced `undefined.env[k]`
- * and file-plane API keys (config.json `openai_api_key` etc.) were dropped.
- * Routing through buildGatewayConfig fixes both. This test fails (throws)
- * on the old code path.
- */
-import { afterAll, describe, expect, test } from 'bun:test';
-import { mkdirSync, mkdtempSync, rmSync, writeFileSync } from 'node:fs';
-import { tmpdir } from 'node:os';
-import { join } from 'node:path';
-import { runEvalTakesQuality } from '../src/commands/eval-takes-quality.ts';
-import { isAvailable, resetGateway } from '../src/core/ai/gateway.ts';
-import { withEnv } from './helpers/with-env.ts';
-import type { BrainEngine } from '../src/core/engine.ts';
-
-afterAll(() => {
-  resetGateway();
-});
-
-describe('runEvalTakesQuality — gateway self-config routes through buildGatewayConfig', () => {
-  test('file-plane openai_api_key reaches the gateway env (help path, engine untouched)', async () => {
-    const home = mkdtempSync(join(tmpdir(), 'gbrain-etq-gw-'));
-    try {
-      mkdirSync(join(home, '.gbrain'), { recursive: true });
-      writeFileSync(
-        join(home, '.gbrain', 'config.json'),
-        JSON.stringify({
-          engine: 'pglite',
-          database_path: join(home, '.gbrain', 'brain'),
-          openai_api_key: 'sk-file-plane-test',
-        }),
-      );
-      await withEnv(
-        {
-          GBRAIN_HOME: home,
-          OPENAI_API_KEY: undefined,
-          DATABASE_URL: undefined,
-          GBRAIN_DATABASE_URL: undefined,
-        },
-        async () => {
-          // 'help' returns before touching the engine, but the gateway is
-          // configured first — exactly the seam under test.
-          await runEvalTakesQuality({} as BrainEngine, ['--help']);
-          expect(isAvailable('embedding', 'openai:text-embedding-3-small')).toBe(true);
-        },
-      );
-    } finally {
-      rmSync(home, { recursive: true, force: true });
-    }
-  });
-});