modelrig 0.3.0 → 0.4.0

package/CHANGELOG.md

@@ -10,6 +10,146 @@ published package diverge.
 This project adheres to [Semantic Versioning](https://semver.org/). Dates are
 UTC.
 
+## [0.4.0] — 2026-08-22
+
+Released with the docs/registry mirror in one R-A ceremony. NOT in this tag:
+the R3 declare-first `trial:` candidate flag (#239) and raw-lane artifact
+steps (G1) — both land in 0.5.0.
+
+CE-9 raw-lane multi-provider parity — the `reasoning` and native-grounding knobs
+generalize beyond gemini, a `responseFormat` knob joins the raw lane, and the
+provider×knob support set becomes ONE table (`RAW_KNOB_SUPPORT`) that the runtime
+guard AND the published docs matrix both read, so they can never disagree. Every
+knob stays ADDITIVE: absent ⇒ byte-identical to the 0.3.0 request shape.
+
+### Added — docs
+
+- **Routing & reliability page** (`https://modelrig.dev/routing-reliability.html`) —
+  the runtime model in one concept page: resolve -> candidate ladder -> validate ->
+  repair -> fall-through -> exhaust, what is stored per attempt, and a **code-anchored
+  boundary** (`ROUTING_BOUNDARY`, one tripwire test per entry) that lists what does not
+  trigger fall-through and what is not built — the build fails if a listed feature ships
+  without the entry being updated.
+
+### Fixed
+
+- **Failure-teaching links resolve.** `RigFailureError.docsUrl` (and the CLI/console
+  error render) now point at `routing-reliability.html#<class>` — one stable heading per
+  failure class — instead of the never-built `/docs/failures` path, which 404'd on
+  every typed failure.
+
+### Added — `rig.runRaw` / `rig.runRawStream`
+
+- **`reasoning` generalizes beyond gemini WITHOUT a per-model allowlist (R1).**
+  openai / deepinfra / fireworks forward `level` → `reasoning_effort` verbatim
+  and let the provider enforce — a model that does not accept it surfaces a
+  CLASSIFIED, teaching failure ("this model does not accept reasoning_effort —
+  remove reasoning or pick a reasoning model"), never a cryptic 400 or a silent
+  drop. grok forwards a DISCLOSED coarsening (`minimal,low → "low"`;
+  `medium,high → "high"`; xAI accepts `low|high` only). gemini stays
+  model-gated (thinking-capable models only). deepseek + anthropic FAIL CLOSED
+  pre-dispatch with a provider-specific fixHint (deepseek → select the
+  `deepseek-reasoner` model; anthropic → extended thinking unwired).
+- **Grok native grounding (R2) — WITHDRAWN before release.** The CE-9 live gate hit xAI's `410 Live search is deprecated`; the replacement `web_search` tool is Responses-API-only, which the grok adapter does not speak yet. grok grounding FAILS CLOSED (matrix `supported: false`, teaching note) until CE-9b lands that path. Original entry, kept for the record: `grounding: { mode: "native" }` on grok →
+  xAI Live Search (`search_parameters: { mode: "on" }`). Citations are not
+  surfaced on `meta` in v1 (same documented limitation as gemini).
+- **`responseFormat?: "json_object"` (R4).** MIME-shaping on the OpenAI-dialect
+  providers (openai/deepseek/grok/deepinfra/fireworks → `response_format: { type:
+  "json_object" }`) and gemini (`responseMimeType: "application/json"`, no
+  schema). This is syntactic-JSON forcing, NOT schema enforcement — F4 stands:
+  no validation, no repair, `meta.validated` stays false. anthropic has no
+  equivalent and FAILS CLOSED. DeepSeek's word-"JSON" prompt wrinkle is the
+  caller's responsibility on the raw lane (documented, never injected).
+- **DeepInfra/Fireworks prompt-cache hint (R3).** `cache.key` now maps to
+  `prompt_cache_key` on the openai-compatible adapter, closing the DeepInfra cost
+  gap. (openai and grok already mapped `cache.key → prompt_cache_key` on 0.3.0 —
+  that was a docs gap, not a code gap; it is now documented as plain fact.)
+
+### Added — exports (R5/R6)
+
+- **`RAW_KNOB_SUPPORT`** (the provider×knob matrix) + **`rawKnobSupport(provider)`**
+  (a pure lookup) — a multi-provider dispatcher ASKS which knobs a provider
+  honors instead of duplicating our lists. The guard (`assertRawKnobsSupported`)
+  and the docs generator both import this table.
+- **`GEMINI_THINKING_CAPABLE_MODELS`** — the thinking-capable set (previously
+  un-exported), so customers stop duplicating it (the §3.1 drift complaint).
+- **`coarsenGrokReasoningEffort`** + **`renderRawKnobMatrix`** — the disclosed
+  grok coarsening and the docs matrix renderer.
+
+### Docs
+
+- A generated provider×knob matrix on the modelrig.dev route-bundles page and in
+  `llms.txt`, rendered FROM `RAW_KNOB_SUPPORT` (never a hand-written twin).
+- `caching-lifecycle.md`: the cache directive is provider-neutral — a resource
+  HANDLE on gemini, a routing HINT (`prompt_cache_key`, retention on openai) on
+  the OpenAI-dialect providers.
+
+CE-8 — first-class **runRaw-only construction**. A Lane-B customer who only calls
+`rig.runRaw`/`runRawStream` can now build a rig with no routes at all, so a stray
+route file it never serves cannot down construction — retiring the empty-temp-dir
+workaround.
+
+### Added
+
+- **`RigConfig.routesDir` now accepts `null`** (was `string`). With `null`,
+  `createRig` SKIPS route loading, the `rig.yaml` manifest, and the
+  serveable-candidate gate; `runRaw`/`runRawStream` work exactly as before. The
+  `string` path is byte-identical to 0.3.0 (the existing serveability hard-error
+  on a zero-serveable route is unchanged and regression-pinned).
+- **`requireRoutesDir(config, surface)`** exported — narrows `routesDir` to a
+  string for route-required contexts, throwing the teaching failure on `null`.
+  Used by `modelrig-server` and `modelrig-oracle` (both always serve routes).
+
+### Behaviour
+
+- On a `routesDir: null` rig, `run`/`runStream`/`runBundle`/`runBundleStream`/
+  `replay`/`bakeoff`/`bakeoffFromEvalCases`/`verifySwap` throw `RigFailureError`
+  class `invariant_violation` (no new failure class) with the fixHint
+  *"this rig was constructed runRaw-only (routesDir: null); give createRig a
+  routesDir to serve routes"*. `rig.serveability` is `[]`.
+- `loadConfigFromEnv` stays string-only (no env var disables routes, by design);
+  opt out in code: `createRig({ ...loadConfigFromEnv(), routesDir: null })`.
+
+### Docs
+
+- README, `docs/migration-playbook.md` (Lane-B seam + C4 checkpoint), and the
+  generated `llms.txt` gain the runRaw-only construction snippet, the `apiKey`
+  BYOK line on every `runRaw` example, the combined-prompt → `userPrompt` note,
+  and the note that the cache-ratio-vs-baseline metric is a REAL-run measurement
+  (deferred-to-first-run at the checkpoint), not a probe.
+
+R0 — the **per-call telemetry ladder key** (routing-console spec §2.1). Every
+inference row now records which call it belongs to and its position in the
+routing ladder, so a dashboard can reconstruct the exact sequence one `rig.run`
+walked and compute fallback / repair rates.
+
+### Added — telemetry ladder key
+
+- **`RunMeta.callId`** — the `call_id` shared by every telemetry row a single
+  `run` / `runStream` / `runBundle` / `runRaw` invocation writes (primary, retry,
+  repair, grounding-search, tool-call), so a caller can correlate a `RunResult`
+  with its ladder in the console. Additive & OPTIONAL — a pre-0.4.0 caller's meta
+  omits it.
+- Telemetry rows gain **`call_id`**, **`attempt_seq`** (1-based, monotonic across
+  every row of the call), and **`candidate_index`** (position of the dispatched
+  candidate in the call's EFFECTIVE declared list — after variant narrowing and
+  per-run filters; a repair row carries the repaired candidate's index; the raw
+  lane is seq 1 / index 0). All three are OPTIONAL on the wire and nullable in
+  storage: an older SDK OMITS them and the payload is byte-identical to before
+  R0; a reader treats null as *"recorded by an SDK older than 0.4.0."*
+- OTLP spans gain `modelrig.call_id` / `modelrig.attempt_seq` /
+  `modelrig.candidate_index`, emitted only when present.
+- **`CandidateIterator.index`** (read-only getter, public type) — the index of the
+  candidate the last `next()` returned (−1 before the first call). Additive for
+  callers; an external IMPLEMENTER of the interface must add the member.
+- **`tags.rung`** on every telemetry row (R1 review, F1) — one of `primary` |
+  `retry` | `repair` | `search` | `tool_call` | `predispatch_fail`, so the routing
+  dashboard and ladder view can tell a real candidate dispatch from a tool-call
+  turn or a pre-dispatch failure without column heuristics. Additive (rides
+  `tags_json`, no schema change); a pre-0.4.0 row omits it and the rollup falls
+  back to `repaired_by` / `provider = "search"`. Public type `TelemetryRung` is
+  exported.
+
 ## [0.3.0] — 2026-08-21
 
 The `modelrig-oracle` MCP server ships to npm alongside this release (its own

package/README.md

@@ -99,27 +99,75 @@ over the route's `cache: auto` stamp, and a non-matching fallback dispatches
 without it. ModelRig carries the handle and prices the hits — it never creates,
 refreshes, or deletes the resource, so a long job needs your own TTL heartbeat
 (see the [caching lifecycle](https://modelrig.dev/caching-lifecycle.html) doc).
-The Lane-B `rig.runRaw({ provider, model, …, cache: { key } })` seam takes the
-same field.
+The Lane-B `rig.runRaw({ provider, model, apiKey, …, cache: { key } })` seam takes
+the same field.
 
 `rig.runRaw` also accepts
-three provider knobs, each additive — absent ⇒ byte-identical dispatch:
-`grounding: { mode: "native" }` (native provider web grounding, gemini-only in
-v1), `serviceTier: "standard" | "flex" | "priority"`, and
-`reasoning: { level: "minimal" | "low" | "medium" | "high" }` (gemini-only in
-v1, mapped to the thinking level). A declared directive a provider cannot honor
-FAILS CLOSED — a thrown `RigFailureError` with class `invariant_violation`,
-pre-dispatch (no meter, no provider call) — never a silent drop. A grounded step
-that also passes a `cache.key` keeps its cached-token ratio: bake the
-`googleSearch` tool into the cache resource at creation (ModelRig sends
-`cachedContent` without tools, which the provider requires).
+provider knobs, each additive — absent ⇒ byte-identical dispatch:
+`grounding: { mode: "native" }` (native provider web grounding —
+gemini; grok fail-closed until CE-9b), `serviceTier: "standard" | "flex" | "priority"`,
+`reasoning: { level: "minimal" | "low" | "medium" | "high" }` (reasoning effort¹),
+and `responseFormat: "json_object"` (syntactic-JSON forcing¹ — MIME-shaping, NOT
+schema enforcement, so `meta.validated` stays `false`). A declared directive a
+provider cannot honor FAILS CLOSED — a thrown `RigFailureError` with class
+`invariant_violation`, pre-dispatch (no meter, no provider call) — never a silent
+drop. **Which provider honors which knob is one table** — the runtime guard and
+the published
+[provider × knob matrix](https://modelrig.dev/route-bundles.html#raw-lane-provider-knob-support-rigrunraw)
+are generated from the same `RAW_KNOB_SUPPORT`; ask it in code with
+`rawKnobSupport(provider)` (and `GEMINI_THINKING_CAPABLE_MODELS` is exported so
+you don't duplicate the thinking-model set). A grounded step that also passes a
+`cache.key` keeps its cached-token ratio: bake the `googleSearch` tool into the
+cache resource at creation (ModelRig sends `cachedContent` without tools, which
+the provider requires).
+
+_¹ Multi-provider `reasoning` and `responseFormat` ship in
+0.4.0; on the published package `reasoning`/`grounding` are gemini-only._
+
+**runRaw — a BYOK passthrough (Lane B).** When you don't need a route — no
+schema, no validation, no retries — call a `provider/model` directly with your
+own provider key:
+
+```ts
+const rig = createRig(loadConfigFromEnv());
+const result = await rig.runRaw({
+  provider: "gemini",
+  model: "gemini-3.1-pro",
+  apiKey: process.env.GEMINI_API_KEY!,   // runRaw is BYOK — pass your provider key
+  systemPrompt: "",                       // one combined prompt? put it ALL in userPrompt, leave systemPrompt empty
+  userPrompt: buildPrompt(input),
+});
+```
+
+`runRaw` is BYOK — pass your provider key; ModelRig holds it for the one call and
+never stores it. It records one metered `lane=raw` telemetry row at the
+provider's list price (zero ModelRig margin) and throws a typed `RigFailureError`
+on failure, exactly like `rig.run`.
+
+**runRaw-only construction (`routesDir: null`) — 0.4.0+.** A
+Lane-B deployment that ONLY calls `runRaw`/`runRawStream` has no routes. Build it
+with `routesDir: null` and `createRig` skips route loading, the `rig.yaml`
+manifest, and the serveability gate entirely — so a stray route file it never
+calls (or a route whose provider key is absent in this deployment) can no longer
+down construction:
+
+```ts
+const rig = createRig({ ...loadConfigFromEnv(), routesDir: null });
+await rig.runRaw({ provider, model, apiKey, systemPrompt, userPrompt });
+```
+
+`loadConfigFromEnv()` always yields a string `routesDir`, so the opt-out is a
+deliberate code-level choice — override it as above; there is no env var that
+disables routes, by design. On such a rig, `rig.run` and every other
+route-dependent surface (`runStream`, `runBundle`, bake-offs, watch, replay,
+`verifySwap`) throws a terminal `invariant_violation` naming the fix.
 
 Every attempt is recorded in a local SQLite file (`.modelrig/telemetry.db`)
 tagged with your `tags`. Failures are typed —
 `RigFailureError.failure.class` is one of `content_invalid`, `capacity_shed`,
-`network`, `refusal`, `cache_invalid`, `timeout`, `budget_exhausted`,
+`network`, `refusal`, `cache_invalid`, `timeout`, `config_auth`, `budget_exhausted`,
 `invariant_violation` — with a separate retry budget per class, so a flaky
-network never spends your content-validation retries.
+network never spends your content-validation retries. What each class means, whether it retries, and how it backs off is on [Routing & reliability](https://modelrig.dev/routing-reliability.html#5-failure-classes).
 
 ## CLI
 
@@ -155,6 +203,10 @@ on the quickstart:
   stable enough to build on, not frozen.
 - **`modelrig serve` needs the `modelrig-server` package** — it is not bundled
   here.
+- **What the ladder does and does not do.** Fall-through is structural and
+  infrastructural, never a quality judgment; the full boundary (quality gate,
+  raw-lane structure, latency ordering…) is generated and tripwire-tested on
+  [Routing & reliability](https://modelrig.dev/routing-reliability.html#10-not-yet-implemented-the-boundary).
 
 ## License
 

package/dist/adapters/deepseek.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"deepseek.d.ts","sourceRoot":"","sources":["../../src/adapters/deepseek.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;GAaG;AAGH,OAAO,KAAK,EAIV,cAAc,EACd,eAAe,EAChB,MAAM,mBAAmB,CAAC;AAE3B,OAAO,KAAK,EAAE,iBAAiB,EAAE,YAAY,EAAE,MAAM,UAAU,CAAC;AAMhE,OAAO,KAAK,EAAuB,cAAc,EAAE,oBAAoB,EAAE,MAAM,gBAAgB,CAAC;AAMhG,MAAM,WAAW,qBAAqB;IACpC,QAAQ,CAAC,MAAM,EAAE,MAAM,CAAC;CACzB;AAKD,wBAAgB,gBAAgB,CAAC,OAAO,EAAE,cAAc,EAAE,KAAK,EAAE,MAAM,GAAG,OAAO,CAMhF;AAED,wBAAgB,uBAAuB,CAAC,OAAO,EAAE,MAAM,GAAG,iBAAiB,CAiB1E;AAED,UAAU,WAAW;IACnB,QAAQ,CAAC,IAAI,EAAE,QAAQ,GAAG,MAAM,CAAC;IACjC,QAAQ,CAAC,OAAO,EAAE,MAAM,CAAC;CAC1B;AAED,MAAM,WAAW,oBAAoB;IACnC,QAAQ,CAAC,KAAK,EAAE,MAAM,CAAC;IACvB,QAAQ,CAAC,UAAU,EAAE,MAAM,CAAC;IAC5B,QAAQ,CAAC,QAAQ,EAAE,SAAS,WAAW,EAAE,CAAC;IAC1C,QAAQ,CAAC,MAAM,EAAE,IAAI,CAAC;IACtB,QAAQ,CAAC,cAAc,EAAE;QAAE,QAAQ,CAAC,aAAa,EAAE,IAAI,CAAA;KAAE,CAAC;IAC1D,8EAA8E;IAC9E,QAAQ,CAAC,WAAW,CAAC,EAAE,MAAM,CAAC;IAC9B,QAAQ,CAAC,KAAK,CAAC,EAAE,MAAM,CAAC;IACxB,QAAQ,CAAC,eAAe,CAAC,EAAE;QAAE,QAAQ,CAAC,IAAI,EAAE,aAAa,CAAA;KAAE,CAAC;IAC5D,QAAQ,CAAC,KAAK,CAAC,EAAE,SAAS,cAAc,EAAE,CAAC;IAC3C,QAAQ,CAAC,WAAW,CAAC,EAAE,oBAAoB,CAAC;CAC7C;AAOD,wBAAgB,mBAAmB,CAAC,KAAK,EAAE,MAAM,EAAE,GAAG,EAAE,YAAY,GAAG,oBAAoB,CA2B1F;AAED,wBAAgB,qBAAqB,CAAC,MAAM,EAAE,qBAAqB,GAAG,eAAe,CA8JpF"}
+{"version":3,"file":"deepseek.d.ts","sourceRoot":"","sources":["../../src/adapters/deepseek.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;GAaG;AAGH,OAAO,KAAK,EAIV,cAAc,EACd,eAAe,EAChB,MAAM,mBAAmB,CAAC;AAE3B,OAAO,KAAK,EAAE,iBAAiB,EAAE,YAAY,EAAE,MAAM,UAAU,CAAC;AAMhE,OAAO,KAAK,EAAuB,cAAc,EAAE,oBAAoB,EAAE,MAAM,gBAAgB,CAAC;AAMhG,MAAM,WAAW,qBAAqB;IACpC,QAAQ,CAAC,MAAM,EAAE,MAAM,CAAC;CACzB;AAKD,wBAAgB,gBAAgB,CAAC,OAAO,EAAE,cAAc,EAAE,KAAK,EAAE,MAAM,GAAG,OAAO,CAMhF;AAED,wBAAgB,uBAAuB,CAAC,OAAO,EAAE,MAAM,GAAG,iBAAiB,CAiB1E;AAED,UAAU,WAAW;IACnB,QAAQ,CAAC,IAAI,EAAE,QAAQ,GAAG,MAAM,CAAC;IACjC,QAAQ,CAAC,OAAO,EAAE,MAAM,CAAC;CAC1B;AAED,MAAM,WAAW,oBAAoB;IACnC,QAAQ,CAAC,KAAK,EAAE,MAAM,CAAC;IACvB,QAAQ,CAAC,UAAU,EAAE,MAAM,CAAC;IAC5B,QAAQ,CAAC,QAAQ,EAAE,SAAS,WAAW,EAAE,CAAC;IAC1C,QAAQ,CAAC,MAAM,EAAE,IAAI,CAAC;IACtB,QAAQ,CAAC,cAAc,EAAE;QAAE,QAAQ,CAAC,aAAa,EAAE,IAAI,CAAA;KAAE,CAAC;IAC1D,8EAA8E;IAC9E,QAAQ,CAAC,WAAW,CAAC,EAAE,MAAM,CAAC;IAC9B,QAAQ,CAAC,KAAK,CAAC,EAAE,MAAM,CAAC;IACxB,QAAQ,CAAC,eAAe,CAAC,EAAE;QAAE,QAAQ,CAAC,IAAI,EAAE,aAAa,CAAA;KAAE,CAAC;IAC5D,QAAQ,CAAC,KAAK,CAAC,EAAE,SAAS,cAAc,EAAE,CAAC;IAC3C,QAAQ,CAAC,WAAW,CAAC,EAAE,oBAAoB,CAAC;CAC7C;AAOD,wBAAgB,mBAAmB,CAAC,KAAK,EAAE,MAAM,EAAE,GAAG,EAAE,YAAY,GAAG,oBAAoB,CAkC1F;AAED,wBAAgB,qBAAqB,CAAC,MAAM,EAAE,qBAAqB,GAAG,eAAe,CA8JpF"}

package/dist/adapters/deepseek.js

@@ -62,6 +62,13 @@ function classifyDeepSeekFailure(message) {
 const JSON_GUARD = "Respond with a single valid JSON object.";
 function buildDeepSeekParams(model, req) {
     const wantJson = req.outputSchema !== null;
+    // CE-9 R4: the raw lane's `responseFormat: "json_object"` also drives DeepSeek's
+    // json_object mode (the flag previously fired ONLY on the routed outputSchema
+    // path, which is always null on the raw lane). Absent both ⇒ no key
+    // (byte-identical). The JSON_GUARD below stays keyed on the ROUTED path only:
+    // on the raw lane the word-"JSON" wrinkle is the caller's prompt responsibility
+    // (spec R4), not silently intercepted here.
+    const jsonObject = wantJson || req.responseFormat === "json_object";
     let systemPrompt = req.systemPrompt;
     if (wantJson && !req.systemPrompt.includes("JSON") && !req.userPrompt.includes("JSON")) {
         systemPrompt = systemPrompt === "" ? JSON_GUARD : `${systemPrompt}\n\n${JSON_GUARD}`;
@@ -77,7 +84,7 @@ function buildDeepSeekParams(model, req) {
         stream: true,
         stream_options: { include_usage: true },
         ...(0, sampling_1.openAISamplingFields)(req),
-        ...(wantJson ? { response_format: { type: "json_object" } } : {}),
+        ...(jsonObject ? { response_format: { type: "json_object" } } : {}),
         ...(req.tools !== undefined && req.tools.length > 0
             ? {
                 tools: (0, tool_mapping_1.toOpenAITools)(req.tools),

package/dist/adapters/gemini.d.ts

@@ -18,6 +18,12 @@ import type { FailureClass } from "../types/failures";
 export interface GeminiAdapterConfig {
     readonly apiKey: string;
 }
+/** Models that accept thinkingConfig — passing it to 2.x models 400s.
+ * Ported from the production THINKING_CAPABLE_MODELS set. EXPORTED (CE-9 R6) so
+ * a multi-provider customer's dispatcher reads THIS set instead of duplicating
+ * it (the §3.1 drift complaint) — the same fact the raw-lane guard uses to keep
+ * gemini reasoning model-gated. */
+export declare const GEMINI_THINKING_CAPABLE_MODELS: ReadonlySet<string>;
 export declare function geminiSupports(feature: CapabilityFlag, model: string): boolean;
 export interface ClassifiedFailure {
     readonly cls: FailureClass;

package/dist/adapters/gemini.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"gemini.d.ts","sourceRoot":"","sources":["../../src/adapters/gemini.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;GAcG;AAGH,OAAO,KAAK,EACV,cAAc,EAGd,cAAc,EACd,eAAe,EAEhB,MAAM,mBAAmB,CAAC;AAC3B,OAAO,KAAK,EAAE,YAAY,EAAc,MAAM,mBAAmB,CAAC;AAKlE,MAAM,WAAW,mBAAmB;IAClC,QAAQ,CAAC,MAAM,EAAE,MAAM,CAAC;CACzB;AAkCD,wBAAgB,cAAc,CAAC,OAAO,EAAE,cAAc,EAAE,KAAK,EAAE,MAAM,GAAG,OAAO,CAG9E;AAED,MAAM,WAAW,iBAAiB;IAChC,QAAQ,CAAC,GAAG,EAAE,YAAY,CAAC;IAC3B,QAAQ,CAAC,OAAO,CAAC,EAAE,MAAM,CAAC;CAC3B;AAED;;;;;GAKG;AACH,wBAAgB,qBAAqB,CAAC,OAAO,EAAE,MAAM,GAAG,iBAAiB,CAyCxE;AAED;6EAC6E;AAC7E,MAAM,MAAM,YAAY,GAAG,IAAI,CAAC,cAAc,EAAE,WAAW,CAAC,CAAC;AAE7D;;;;;;;;GAQG;AACH,wBAAgB,0BAA0B,CAAC,KAAK,EAAE,MAAM,EAAE,GAAG,EAAE,YAAY,GAAG,iBAAiB,GAAG,IAAI,CAQrG;AAED,MAAM,WAAW,mBAAmB;IAClC,QAAQ,CAAC,KAAK,EAAE,MAAM,CAAC;IACvB,QAAQ,CAAC,QAAQ,EAAE,MAAM,CAAC;IAC1B,QAAQ,CAAC,MAAM,EAAE,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,CAAC;CAC1C;AAED,iEAAiE;AACjE,wBAAgB,kBAAkB,CAAC,KAAK,EAAE,MAAM,EAAE,GAAG,EAAE,YAAY,GAAG,mBAAmB,CAgDxF;AA0BD,wBAAgB,mBAAmB,CAAC,MAAM,EAAE,mBAAmB,GAAG,eAAe,CA+JhF"}
+{"version":3,"file":"gemini.d.ts","sourceRoot":"","sources":["../../src/adapters/gemini.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;GAcG;AAGH,OAAO,KAAK,EACV,cAAc,EAGd,cAAc,EACd,eAAe,EAEhB,MAAM,mBAAmB,CAAC;AAC3B,OAAO,KAAK,EAAE,YAAY,EAAc,MAAM,mBAAmB,CAAC;AAKlE,MAAM,WAAW,mBAAmB;IAClC,QAAQ,CAAC,MAAM,EAAE,MAAM,CAAC;CACzB;AAED;;;;mCAImC;AACnC,eAAO,MAAM,8BAA8B,EAAE,WAAW,CAAC,MAAM,CAI7D,CAAC;AA0BH,wBAAgB,cAAc,CAAC,OAAO,EAAE,cAAc,EAAE,KAAK,EAAE,MAAM,GAAG,OAAO,CAG9E;AAED,MAAM,WAAW,iBAAiB;IAChC,QAAQ,CAAC,GAAG,EAAE,YAAY,CAAC;IAC3B,QAAQ,CAAC,OAAO,CAAC,EAAE,MAAM,CAAC;CAC3B;AAED;;;;;GAKG;AACH,wBAAgB,qBAAqB,CAAC,OAAO,EAAE,MAAM,GAAG,iBAAiB,CAyCxE;AAED;6EAC6E;AAC7E,MAAM,MAAM,YAAY,GAAG,IAAI,CAAC,cAAc,EAAE,WAAW,CAAC,CAAC;AAE7D;;;;;;;;GAQG;AACH,wBAAgB,0BAA0B,CAAC,KAAK,EAAE,MAAM,EAAE,GAAG,EAAE,YAAY,GAAG,iBAAiB,GAAG,IAAI,CAQrG;AAED,MAAM,WAAW,mBAAmB;IAClC,QAAQ,CAAC,KAAK,EAAE,MAAM,CAAC;IACvB,QAAQ,CAAC,QAAQ,EAAE,MAAM,CAAC;IAC1B,QAAQ,CAAC,MAAM,EAAE,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,CAAC;CAC1C;AAED,iEAAiE;AACjE,wBAAgB,kBAAkB,CAAC,KAAK,EAAE,MAAM,EAAE,GAAG,EAAE,YAAY,GAAG,mBAAmB,CAqDxF;AA0BD,wBAAgB,mBAAmB,CAAC,MAAM,EAAE,mBAAmB,GAAG,eAAe,CA+JhF"}

package/dist/adapters/gemini.js

@@ -15,6 +15,7 @@
  * gemini transformation logic is consulted during the port.
  */
 Object.defineProperty(exports, "__esModule", { value: true });
+exports.GEMINI_THINKING_CAPABLE_MODELS = void 0;
 exports.geminiSupports = geminiSupports;
 exports.classifyGeminiFailure = classifyGeminiFailure;
 exports.geminiReasoningUnsupported = geminiReasoningUnsupported;
@@ -24,8 +25,11 @@ const genai_1 = require("@google/genai");
 const tool_mapping_1 = require("./tool-mapping");
 const sampling_1 = require("./sampling");
 /** Models that accept thinkingConfig — passing it to 2.x models 400s.
- * Ported from the production THINKING_CAPABLE_MODELS set. */
-const THINKING_CAPABLE_MODELS = new Set([
+ * Ported from the production THINKING_CAPABLE_MODELS set. EXPORTED (CE-9 R6) so
+ * a multi-provider customer's dispatcher reads THIS set instead of duplicating
+ * it (the §3.1 drift complaint) — the same fact the raw-lane guard uses to keep
+ * gemini reasoning model-gated. */
+exports.GEMINI_THINKING_CAPABLE_MODELS = new Set([
     "gemini-3.1-pro-preview",
     "gemini-3.0-pro-preview",
     "gemini-3.1-flash-preview",
@@ -110,7 +114,7 @@ function classifyGeminiFailure(message) {
  * reasoning declared, or a thinking-capable model). Pure — unit-tested directly.
  */
 function geminiReasoningUnsupported(model, req) {
-    if (req.reasoning !== undefined && !THINKING_CAPABLE_MODELS.has(model)) {
+    if (req.reasoning !== undefined && !exports.GEMINI_THINKING_CAPABLE_MODELS.has(model)) {
         return {
             cls: "invariant_violation",
             fixHint: `reasoning control requires a thinking-capable Gemini model; "${model}" is not one — remove reasoning or choose a thinking-capable model`,
@@ -134,12 +138,18 @@ function buildGeminiRequest(model, req) {
         maxOutputTokens: (0, sampling_1.effectiveMaxOutputTokens)(req, DEFAULT_MAX_OUTPUT_TOKENS),
         topP: req.sampling?.topP ?? TOP_P,
         topK: TOP_K,
-        ...(THINKING_CAPABLE_MODELS.has(model) ? { thinkingConfig: { thinkingLevel } } : {}),
+        ...(exports.GEMINI_THINKING_CAPABLE_MODELS.has(model) ? { thinkingConfig: { thinkingLevel } } : {}),
     };
     if (req.outputSchema !== null) {
         config.responseMimeType = "application/json";
         config.responseJsonSchema = req.outputSchema;
     }
+    else if (req.responseFormat === "json_object") {
+        // CE-9 R4: mime-shaping ONLY — force a JSON mime type WITHOUT a schema. This
+        // is the raw lane, so F4 stands: no responseJsonSchema, no validation, no
+        // repair, `validated` stays false. Absent ⇒ no key (byte-identical).
+        config.responseMimeType = "application/json";
+    }
     // cachedContent excludes systemInstruction and tools — the API rejects
     // requests setting both (they are baked into the cache at creation time,
     // including the googleSearch tool for cached grounded steps).

package/dist/adapters/grok.d.ts

@@ -36,7 +36,14 @@ export interface GrokShapedParams {
     /** Prompt-cache routing hint (WS-C) — the stable-prefix fingerprint the run
      * loop stamps; xAI keeps a warm cache keyed to it (writes free today). */
     readonly prompt_cache_key?: string;
+    /** CE-9 R1 — xAI accepts `reasoning_effort: "low" | "high"` ONLY, so the
+     * neutral four-level `level` is coarsened by `coarsenGrokReasoningEffort`
+     * (a DISCLOSED, lossy mapping: `minimal,low → "low"`; `medium,high → "high"`).
+     * Present only when the caller declared `reasoning`. */
+    readonly reasoning_effort?: "low" | "high";
     readonly response_format?: {
+        readonly type: "json_object";
+    } | {
         readonly type: "json_schema";
         readonly json_schema: {
             readonly name: string;

package/dist/adapters/grok.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"grok.d.ts","sourceRoot":"","sources":["../../src/adapters/grok.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;GAWG;AAGH,OAAO,KAAK,EAIV,cAAc,EACd,eAAe,EAChB,MAAM,mBAAmB,CAAC;AAE3B,OAAO,KAAK,EAAE,iBAAiB,EAAE,YAAY,EAAE,MAAM,UAAU,CAAC;AAMhE,OAAO,KAAK,EAAuB,cAAc,EAAE,oBAAoB,EAAE,MAAM,gBAAgB,CAAC;AAMhG,MAAM,WAAW,iBAAiB;IAChC,QAAQ,CAAC,MAAM,EAAE,MAAM,CAAC;CACzB;AAKD,wBAAgB,YAAY,CAAC,OAAO,EAAE,cAAc,EAAE,MAAM,EAAE,MAAM,GAAG,OAAO,CAc7E;AAED,wBAAgB,mBAAmB,CAAC,OAAO,EAAE,MAAM,GAAG,iBAAiB,CAsBtE;AAED,UAAU,WAAW;IACnB,QAAQ,CAAC,IAAI,EAAE,QAAQ,GAAG,MAAM,CAAC;IACjC,QAAQ,CAAC,OAAO,EAAE,MAAM,CAAC;CAC1B;AAED,MAAM,WAAW,gBAAgB;IAC/B,QAAQ,CAAC,KAAK,EAAE,MAAM,CAAC;IACvB,QAAQ,CAAC,UAAU,EAAE,MAAM,CAAC;IAC5B,QAAQ,CAAC,QAAQ,EAAE,SAAS,WAAW,EAAE,CAAC;IAC1C,QAAQ,CAAC,MAAM,EAAE,IAAI,CAAC;IACtB,QAAQ,CAAC,cAAc,EAAE;QAAE,QAAQ,CAAC,aAAa,EAAE,IAAI,CAAA;KAAE,CAAC;IAC1D,8EAA8E;IAC9E,QAAQ,CAAC,WAAW,CAAC,EAAE,MAAM,CAAC;IAC9B,QAAQ,CAAC,KAAK,CAAC,EAAE,MAAM,CAAC;IACxB;8EAC0E;IAC1E,QAAQ,CAAC,gBAAgB,CAAC,EAAE,MAAM,CAAC;IACnC,QAAQ,CAAC,eAAe,CAAC,EAAE;QACzB,QAAQ,CAAC,IAAI,EAAE,aAAa,CAAC;QAC7B,QAAQ,CAAC,WAAW,EAAE;YAAE,QAAQ,CAAC,IAAI,EAAE,MAAM,CAAC;YAAC,QAAQ,CAAC,MAAM,EAAE,MAAM,CAAC;YAAC,QAAQ,CAAC,MAAM,EAAE,IAAI,CAAA;SAAE,CAAC;KACjG,CAAC;IACF,QAAQ,CAAC,KAAK,CAAC,EAAE,SAAS,cAAc,EAAE,CAAC;IAC3C,QAAQ,CAAC,WAAW,CAAC,EAAE,oBAAoB,CAAC;CAC7C;AAED,wBAAgB,eAAe,CAAC,KAAK,EAAE,MAAM,EAAE,GAAG,EAAE,YAAY,GAAG,gBAAgB,CAsClF;AAED,wBAAgB,iBAAiB,CAAC,MAAM,EAAE,iBAAiB,GAAG,eAAe,CAgK5E"}
+{"version":3,"file":"grok.d.ts","sourceRoot":"","sources":["../../src/adapters/grok.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;GAWG;AAGH,OAAO,KAAK,EAIV,cAAc,EACd,eAAe,EAChB,MAAM,mBAAmB,CAAC;AAE3B,OAAO,KAAK,EAAE,iBAAiB,EAAE,YAAY,EAAE,MAAM,UAAU,CAAC;AAMhE,OAAO,KAAK,EAAuB,cAAc,EAAE,oBAAoB,EAAE,MAAM,gBAAgB,CAAC;AAOhG,MAAM,WAAW,iBAAiB;IAChC,QAAQ,CAAC,MAAM,EAAE,MAAM,CAAC;CACzB;AAKD,wBAAgB,YAAY,CAAC,OAAO,EAAE,cAAc,EAAE,MAAM,EAAE,MAAM,GAAG,OAAO,CAc7E;AAED,wBAAgB,mBAAmB,CAAC,OAAO,EAAE,MAAM,GAAG,iBAAiB,CA+BtE;AAED,UAAU,WAAW;IACnB,QAAQ,CAAC,IAAI,EAAE,QAAQ,GAAG,MAAM,CAAC;IACjC,QAAQ,CAAC,OAAO,EAAE,MAAM,CAAC;CAC1B;AAED,MAAM,WAAW,gBAAgB;IAC/B,QAAQ,CAAC,KAAK,EAAE,MAAM,CAAC;IACvB,QAAQ,CAAC,UAAU,EAAE,MAAM,CAAC;IAC5B,QAAQ,CAAC,QAAQ,EAAE,SAAS,WAAW,EAAE,CAAC;IAC1C,QAAQ,CAAC,MAAM,EAAE,IAAI,CAAC;IACtB,QAAQ,CAAC,cAAc,EAAE;QAAE,QAAQ,CAAC,aAAa,EAAE,IAAI,CAAA;KAAE,CAAC;IAC1D,8EAA8E;IAC9E,QAAQ,CAAC,WAAW,CAAC,EAAE,MAAM,CAAC;IAC9B,QAAQ,CAAC,KAAK,CAAC,EAAE,MAAM,CAAC;IACxB;8EAC0E;IAC1E,QAAQ,CAAC,gBAAgB,CAAC,EAAE,MAAM,CAAC;IACnC;;;4DAGwD;IACxD,QAAQ,CAAC,gBAAgB,CAAC,EAAE,KAAK,GAAG,MAAM,CAAC;IAC3C,QAAQ,CAAC,eAAe,CAAC,EACrB;QAAE,QAAQ,CAAC,IAAI,EAAE,aAAa,CAAA;KAAE,GAChC;QACE,QAAQ,CAAC,IAAI,EAAE,aAAa,CAAC;QAC7B,QAAQ,CAAC,WAAW,EAAE;YAAE,QAAQ,CAAC,IAAI,EAAE,MAAM,CAAC;YAAC,QAAQ,CAAC,MAAM,EAAE,MAAM,CAAC;YAAC,QAAQ,CAAC,MAAM,EAAE,IAAI,CAAA;SAAE,CAAC;KACjG,CAAC;IACN,QAAQ,CAAC,KAAK,CAAC,EAAE,SAAS,cAAc,EAAE,CAAC;IAC3C,QAAQ,CAAC,WAAW,CAAC,EAAE,oBAAoB,CAAC;CAC7C;AAED,wBAAgB,eAAe,CAAC,KAAK,EAAE,MAAM,EAAE,GAAG,EAAE,YAAY,GAAG,gBAAgB,CAmDlF;AAED,wBAAgB,iBAAiB,CAAC,MAAM,EAAE,iBAAiB,GAAG,eAAe,CAgK5E"}

package/dist/adapters/grok.js

@@ -23,6 +23,7 @@ const openai_1 = __importDefault(require("openai"));
 const tool_mapping_1 = require("./tool-mapping");
 const openai_stream_1 = require("./openai-stream");
 const sampling_1 = require("./sampling");
+const raw_support_1 = require("../run/raw-support");
 const GROK_BASE_URL = "https://api.x.ai/v1";
 const DEFAULT_MAX_TOKENS = 32000;
 function grokSupports(feature, _model) {
@@ -65,6 +66,14 @@ function classifyGrokFailure(message) {
             fixHint: "the candidate model was not found — fix the model id in the route's candidate list",
         };
     }
+    // CE-9 R1: a non-reasoning grok variant rejects `reasoning_effort` — teach the
+    // fix (pick a reasoning variant) rather than surface a cryptic 400.
+    if (/reasoning_effort/.test(msg)) {
+        return {
+            cls: "network",
+            fixHint: "this Grok model does not accept reasoning_effort — remove reasoning or pick a Grok reasoning variant",
+        };
+    }
     return { cls: "network" };
 }
 function buildGrokParams(model, req) {
@@ -84,6 +93,17 @@ function buildGrokParams(model, req) {
         ...(req.cache?.key !== undefined && req.cache.key !== ""
             ? { prompt_cache_key: req.cache.key }
             : {}),
+        // CE-9 R1: coarsen the neutral level to xAI's low|high (disclosed mapping).
+        // Absent ⇒ no key. A non-reasoning grok model's rejection is taught below.
+        ...(req.reasoning !== undefined
+            ? { reasoning_effort: (0, raw_support_1.coarsenGrokReasoningEffort)(req.reasoning.level) }
+            : {}),
+        // CE-9b: grounding is NOT mapped here. xAI retired Live Search
+        // (`search_parameters` → HTTP 410 on the CE-9 live gate); the replacement
+        // `web_search` tool is Responses-API-only. The raw-lane guard fails grok
+        // grounding closed (matrix: supported=false) until CE-9b adds that path.
+        // response_format: a native strict json_schema (routed lane) wins; else the
+        // RAW lane's mime-shaping json_object (CE-9 R4). Absent both ⇒ no key.
         ...(req.outputSchema !== null
             ? {
                 response_format: {
@@ -95,7 +115,9 @@ function buildGrokParams(model, req) {
                     },
                 },
             }
-            : {}),
+            : req.responseFormat === "json_object"
+                ? { response_format: { type: "json_object" } }
+                : {}),
         ...(req.tools !== undefined && req.tools.length > 0
             ? {
                 tools: (0, tool_mapping_1.toOpenAITools)(req.tools),

package/dist/adapters/openai-compat.d.ts

@@ -49,6 +49,13 @@ export interface OpenAICompatShapedParams {
     /** CE-6 declared sampling — present only when the route/raw asks for them. */
     readonly temperature?: number;
     readonly top_p?: number;
+    /** CE-9 R1 — reasoning effort forwarded verbatim (1:1). Present only when the
+     * caller declared `reasoning`; a host/model that does not accept it rejects at
+     * the provider and `classifyCompatFailure` teaches (no per-model allowlist). */
+    readonly reasoning_effort?: "minimal" | "low" | "medium" | "high";
+    /** CE-9 R3 — the DeepInfra prompt_cache_key gap: `cache.key` → the routing
+     * hint. Present only when the caller supplies `cache.key`. */
+    readonly prompt_cache_key?: string;
     readonly response_format?: {
         readonly type: "json_object";
     } | {

package/dist/adapters/openai-compat.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"openai-compat.d.ts","sourceRoot":"","sources":["../../src/adapters/openai-compat.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;GAkBG;AAGH,OAAO,KAAK,EAIV,cAAc,EACd,eAAe,EACf,UAAU,EACX,MAAM,mBAAmB,CAAC;AAE3B,OAAO,KAAK,EAAE,iBAAiB,EAAE,YAAY,EAAE,MAAM,UAAU,CAAC;AAMhE,OAAO,KAAK,EAAuB,cAAc,EAAE,oBAAoB,EAAE,MAAM,gBAAgB,CAAC;AAQhG;;8CAE8C;AAC9C,wBAAgB,qBAAqB,CAAC,KAAK,EAAE,MAAM,GAAG,MAAM,CAE3D;AAED,MAAM,WAAW,gBAAgB;IAC/B,QAAQ,CAAC,OAAO,EAAE,MAAM,CAAC;IACzB,QAAQ,CAAC,YAAY,CAAC,EAAE,CAAC,KAAK,EAAE,MAAM,KAAK,MAAM,CAAC;CACnD;AAED;;;0EAG0E;AAC1E,eAAO,MAAM,mBAAmB,EAAE,QAAQ,CAAC,MAAM,CAAC,WAAW,GAAG,WAAW,EAAE,gBAAgB,CAAC,CAG7F,CAAC;AAEF,wBAAgB,qBAAqB,CAAC,OAAO,EAAE,MAAM,EAAE,IAAI,EAAE,MAAM,GAAG,iBAAiB,CAiBtF;AAED,UAAU,WAAW;IACnB,QAAQ,CAAC,IAAI,EAAE,QAAQ,GAAG,MAAM,CAAC;IACjC,QAAQ,CAAC,OAAO,EAAE,MAAM,CAAC;CAC1B;AAED,MAAM,WAAW,wBAAwB;IACvC,QAAQ,CAAC,KAAK,EAAE,MAAM,CAAC;IACvB,QAAQ,CAAC,UAAU,EAAE,MAAM,CAAC;IAC5B,QAAQ,CAAC,QAAQ,EAAE,SAAS,WAAW,EAAE,CAAC;IAC1C,QAAQ,CAAC,MAAM,EAAE,IAAI,CAAC;IACtB,QAAQ,CAAC,cAAc,EAAE;QAAE,QAAQ,CAAC,aAAa,EAAE,IAAI,CAAA;KAAE,CAAC;IAC1D,8EAA8E;IAC9E,QAAQ,CAAC,WAAW,CAAC,EAAE,MAAM,CAAC;IAC9B,QAAQ,CAAC,KAAK,CAAC,EAAE,MAAM,CAAC;IACxB,QAAQ,CAAC,eAAe,CAAC,EACrB;QAAE,QAAQ,CAAC,IAAI,EAAE,aAAa,CAAA;KAAE,GAChC;QACE,QAAQ,CAAC,IAAI,EAAE,aAAa,CAAC;QAC7B,QAAQ,CAAC,WAAW,EAAE;YAAE,QAAQ,CAAC,IAAI,EAAE,MAAM,CAAC;YAAC,QAAQ,CAAC,MAAM,EAAE,IAAI,CAAC;YAAC,QAAQ,CAAC,MAAM,EAAE,MAAM,CAAA;SAAE,CAAC;KACjG,CAAC;IACN,QAAQ,CAAC,KAAK,CAAC,EAAE,SAAS,cAAc,EAAE,CAAC;IAC3C,QAAQ,CAAC,WAAW,CAAC,EAAE,oBAAoB,CAAC;CAC7C;AAOD;;2DAE2D;AAC3D,wBAAgB,uBAAuB,CACrC,KAAK,EAAE,MAAM,EACb,GAAG,EAAE,YAAY,EACjB,YAAY,EAAE,WAAW,CAAC,cAAc,CAAC,GACxC,wBAAwB,CAqC1B;AAED,MAAM,WAAW,yBAAyB;IACxC,QAAQ,CAAC,UAAU,EAAE,UAAU,CAAC;IAChC,QAAQ,CAAC,OAAO,EAAE,MAAM,CAAC;IACzB,QAAQ,CAAC,MAAM,EAAE,MAAM,CAAC;IACxB,QAAQ,CAAC,YAAY,CAAC,EAAE,CAAC,KAAK,EAAE,MAAM,KAAK,MAAM,CAAC;IAClD;+DAC2D;IAC3D,QAAQ,CAAC,iBAAiB,EAAE,CAAC,KAAK,EAAE,MAAM,KAAK,WAAW,CAAC,cAAc,CAAC,CAAC;CAC5E;AAED,wBAAgB,yBAAyB,CAAC,MAAM,EAAE,yBAAyB,GAAG,eAAe,CAwK5F"}
+{"version":3,"file":"openai-compat.d.ts","sourceRoot":"","sources":["../../src/adapters/openai-compat.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;GAkBG;AAGH,OAAO,KAAK,EAIV,cAAc,EACd,eAAe,EACf,UAAU,EACX,MAAM,mBAAmB,CAAC;AAE3B,OAAO,KAAK,EAAE,iBAAiB,EAAE,YAAY,EAAE,MAAM,UAAU,CAAC;AAMhE,OAAO,KAAK,EAAuB,cAAc,EAAE,oBAAoB,EAAE,MAAM,gBAAgB,CAAC;AAQhG;;8CAE8C;AAC9C,wBAAgB,qBAAqB,CAAC,KAAK,EAAE,MAAM,GAAG,MAAM,CAE3D;AAED,MAAM,WAAW,gBAAgB;IAC/B,QAAQ,CAAC,OAAO,EAAE,MAAM,CAAC;IACzB,QAAQ,CAAC,YAAY,CAAC,EAAE,CAAC,KAAK,EAAE,MAAM,KAAK,MAAM,CAAC;CACnD;AAED;;;0EAG0E;AAC1E,eAAO,MAAM,mBAAmB,EAAE,QAAQ,CAAC,MAAM,CAAC,WAAW,GAAG,WAAW,EAAE,gBAAgB,CAAC,CAG7F,CAAC;AAEF,wBAAgB,qBAAqB,CAAC,OAAO,EAAE,MAAM,EAAE,IAAI,EAAE,MAAM,GAAG,iBAAiB,CA2BtF;AAED,UAAU,WAAW;IACnB,QAAQ,CAAC,IAAI,EAAE,QAAQ,GAAG,MAAM,CAAC;IACjC,QAAQ,CAAC,OAAO,EAAE,MAAM,CAAC;CAC1B;AAED,MAAM,WAAW,wBAAwB;IACvC,QAAQ,CAAC,KAAK,EAAE,MAAM,CAAC;IACvB,QAAQ,CAAC,UAAU,EAAE,MAAM,CAAC;IAC5B,QAAQ,CAAC,QAAQ,EAAE,SAAS,WAAW,EAAE,CAAC;IAC1C,QAAQ,CAAC,MAAM,EAAE,IAAI,CAAC;IACtB,QAAQ,CAAC,cAAc,EAAE;QAAE,QAAQ,CAAC,aAAa,EAAE,IAAI,CAAA;KAAE,CAAC;IAC1D,8EAA8E;IAC9E,QAAQ,CAAC,WAAW,CAAC,EAAE,MAAM,CAAC;IAC9B,QAAQ,CAAC,KAAK,CAAC,EAAE,MAAM,CAAC;IACxB;;oFAEgF;IAChF,QAAQ,CAAC,gBAAgB,CAAC,EAAE,SAAS,GAAG,KAAK,GAAG,QAAQ,GAAG,MAAM,CAAC;IAClE;kEAC8D;IAC9D,QAAQ,CAAC,gBAAgB,CAAC,EAAE,MAAM,CAAC;IACnC,QAAQ,CAAC,eAAe,CAAC,EACrB;QAAE,QAAQ,CAAC,IAAI,EAAE,aAAa,CAAA;KAAE,GAChC;QACE,QAAQ,CAAC,IAAI,EAAE,aAAa,CAAC;QAC7B,QAAQ,CAAC,WAAW,EAAE;YAAE,QAAQ,CAAC,IAAI,EAAE,MAAM,CAAC;YAAC,QAAQ,CAAC,MAAM,EAAE,IAAI,CAAC;YAAC,QAAQ,CAAC,MAAM,EAAE,MAAM,CAAA;SAAE,CAAC;KACjG,CAAC;IACN,QAAQ,CAAC,KAAK,CAAC,EAAE,SAAS,cAAc,EAAE,CAAC;IAC3C,QAAQ,CAAC,WAAW,CAAC,EAAE,oBAAoB,CAAC;CAC7C;AAOD;;2DAE2D;AAC3D,wBAAgB,uBAAuB,CACrC,KAAK,EAAE,MAAM,EACb,GAAG,EAAE,YAAY,EACjB,YAAY,EAAE,WAAW,CAAC,cAAc,CAAC,GACxC,wBAAwB,CAmD1B;AAED,MAAM,WAAW,yBAAyB;IACxC,QAAQ,CAAC,UAAU,EAAE,UAAU,CAAC;IAChC,QAAQ,CAAC,OAAO,EAAE,MAAM,CAAC;IACzB,QAAQ,CAAC,MAAM,EAAE,MAAM,CAAC;IACxB,QAAQ,CAAC,YAAY,CAAC,EAAE,CAAC,KAAK,EAAE,MAAM,KAAK,MAAM,CAAC;IAClD;+DAC2D;IAC3D,QAAQ,CAAC,iBAAiB,EAAE,CAAC,KAAK,EAAE,MAAM,KAAK,WAAW,CAAC,cAAc,CAAC,CAAC;CAC5E;AAED,wBAAgB,yBAAyB,CAAC,MAAM,EAAE,yBAAyB,GAAG,eAAe,CAwK5F"}

package/dist/adapters/openai-compat.js

@@ -66,6 +66,15 @@ function classifyCompatFailure(message, host) {
             fixHint: "the candidate model was not found — fix the model id in the route's candidate list",
         };
     }
+    // CE-9 R1: a model that does not accept `reasoning_effort` rejects the request
+    // — teach it rather than surfacing a cryptic generic 400 (the reason we forward
+    // verbatim instead of keeping a per-model allowlist).
+    if (/reasoning_effort/.test(msg)) {
+        return {
+            cls: "network",
+            fixHint: `this ${host} model does not accept reasoning_effort — remove reasoning or pick a reasoning model`,
+        };
+    }
     return { cls: "network" };
 }
 /** json_object mode 400s on several compat backends when the literal word
@@ -93,6 +102,18 @@ function buildOpenAICompatParams(model, req, profileFlags) {
         stream: true,
         stream_options: { include_usage: true },
         ...(0, sampling_1.openAISamplingFields)(req),
+        // CE-9 R1: forward reasoning effort verbatim (1:1). Absent ⇒ no key.
+        ...(req.reasoning !== undefined ? { reasoning_effort: req.reasoning.level } : {}),
+        // CE-9 R3: the DeepInfra prompt_cache_key gap — map cache.key → the routing
+        // hint (absent ⇒ no key, byte-identical). Additive; unrecognized on a host
+        // that doesn't cache, the hint is simply ignored.
+        ...(req.cache?.key !== undefined && req.cache.key !== ""
+            ? { prompt_cache_key: req.cache.key }
+            : {}),
+        // response_format: a routed strict json_schema (structured_native profile)
+        // wins; else routed json_object; else the RAW lane's mime-shaping json_object
+        // (CE-9 R4). The raw path carries NO JSON_GUARD injection — the word-"JSON"
+        // wrinkle is the caller's prompt responsibility on the raw lane.
         ...(wantJson
             ? strict
                 ? {
@@ -102,7 +123,9 @@ function buildOpenAICompatParams(model, req, profileFlags) {
                     },
                 }
                 : { response_format: { type: "json_object" } }
-            : {}),
+            : req.responseFormat === "json_object"
+                ? { response_format: { type: "json_object" } }
+                : {}),
         ...(req.tools !== undefined && req.tools.length > 0
             ? {
                 tools: (0, tool_mapping_1.toOpenAITools)(req.tools),

package/dist/adapters/openai.d.ts

@@ -38,7 +38,14 @@ export interface OpenAIShapedParams {
     readonly service_tier?: "flex" | "priority";
     readonly prompt_cache_key?: string;
     readonly prompt_cache_retention?: "24h";
+    /** CE-9 R1 — reasoning effort forwarded verbatim (1:1 from the neutral
+     * `level`). Present only when the caller declared `reasoning`. A model that
+     * does not accept it rejects at the provider; `classifyOpenAIFailure` turns
+     * that 400 into a teaching failure (no per-model allowlist). */
+    readonly reasoning_effort?: "minimal" | "low" | "medium" | "high";
     readonly response_format?: {
+        readonly type: "json_object";
+    } | {
         readonly type: "json_schema";
         readonly json_schema: {
             readonly name: string;

package/dist/adapters/openai.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"openai.d.ts","sourceRoot":"","sources":["../../src/adapters/openai.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;GAYG;AAGH,OAAO,KAAK,EAIV,cAAc,EACd,eAAe,EAEhB,MAAM,mBAAmB,CAAC;AAE3B,OAAO,KAAK,EAAE,iBAAiB,EAAE,YAAY,EAAE,MAAM,UAAU,CAAC;AAMhE,OAAO,KAAK,EAAuB,cAAc,EAAE,oBAAoB,EAAE,MAAM,gBAAgB,CAAC;AAMhG,MAAM,WAAW,mBAAmB;IAClC,QAAQ,CAAC,MAAM,EAAE,MAAM,CAAC;CACzB;AAMD,wBAAgB,mBAAmB,CAAC,SAAS,EAAE,MAAM,GAAG,OAAO,CAI9D;AAED,wBAAgB,cAAc,CAAC,OAAO,EAAE,cAAc,EAAE,KAAK,EAAE,MAAM,GAAG,OAAO,CAe9E;AAED,wBAAgB,qBAAqB,CAAC,OAAO,EAAE,MAAM,GAAG,iBAAiB,CAiBxE;AAED,UAAU,WAAW;IACnB,QAAQ,CAAC,IAAI,EAAE,QAAQ,GAAG,MAAM,CAAC;IACjC,QAAQ,CAAC,OAAO,EAAE,MAAM,CAAC;CAC1B;AAED,MAAM,WAAW,kBAAkB;IACjC,QAAQ,CAAC,KAAK,EAAE,MAAM,CAAC;IACvB,QAAQ,CAAC,qBAAqB,EAAE,MAAM,CAAC;IACvC,QAAQ,CAAC,QAAQ,EAAE,SAAS,WAAW,EAAE,CAAC;IAC1C,QAAQ,CAAC,MAAM,EAAE,IAAI,CAAC;IACtB,QAAQ,CAAC,cAAc,EAAE;QAAE,QAAQ,CAAC,aAAa,EAAE,IAAI,CAAA;KAAE,CAAC;IAC1D,8EAA8E;IAC9E,QAAQ,CAAC,WAAW,CAAC,EAAE,MAAM,CAAC;IAC9B,QAAQ,CAAC,KAAK,CAAC,EAAE,MAAM,CAAC;IACxB,QAAQ,CAAC,YAAY,CAAC,EAAE,MAAM,GAAG,UAAU,CAAC;IAC5C,QAAQ,CAAC,gBAAgB,CAAC,EAAE,MAAM,CAAC;IACnC,QAAQ,CAAC,sBAAsB,CAAC,EAAE,KAAK,CAAC;IACxC,QAAQ,CAAC,eAAe,CAAC,EAAE;QACzB,QAAQ,CAAC,IAAI,EAAE,aAAa,CAAC;QAC7B,QAAQ,CAAC,WAAW,EAAE;YAAE,QAAQ,CAAC,IAAI,EAAE,MAAM,CAAC;YAAC,QAAQ,CAAC,MAAM,EAAE,MAAM,CAAC;YAAC,QAAQ,CAAC,MAAM,EAAE,IAAI,CAAA;SAAE,CAAC;KACjG,CAAC;IACF,QAAQ,CAAC,KAAK,CAAC,EAAE,SAAS,cAAc,EAAE,CAAC;IAC3C,QAAQ,CAAC,WAAW,CAAC,EAAE,oBAAoB,CAAC;CAC7C;AAID,wBAAgB,iBAAiB,CAAC,KAAK,EAAE,MAAM,EAAE,GAAG,EAAE,YAAY,GAAG,kBAAkB,CA6CtF;AAcD,wBAAgB,mBAAmB,CAAC,MAAM,EAAE,mBAAmB,GAAG,eAAe,CAwMhF"}
+{"version":3,"file":"openai.d.ts","sourceRoot":"","sources":["../../src/adapters/openai.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;GAYG;AAGH,OAAO,KAAK,EAIV,cAAc,EACd,eAAe,EAEhB,MAAM,mBAAmB,CAAC;AAE3B,OAAO,KAAK,EAAE,iBAAiB,EAAE,YAAY,EAAE,MAAM,UAAU,CAAC;AAMhE,OAAO,KAAK,EAAuB,cAAc,EAAE,oBAAoB,EAAE,MAAM,gBAAgB,CAAC;AAMhG,MAAM,WAAW,mBAAmB;IAClC,QAAQ,CAAC,MAAM,EAAE,MAAM,CAAC;CACzB;AAMD,wBAAgB,mBAAmB,CAAC,SAAS,EAAE,MAAM,GAAG,OAAO,CAI9D;AAED,wBAAgB,cAAc,CAAC,OAAO,EAAE,cAAc,EAAE,KAAK,EAAE,MAAM,GAAG,OAAO,CAe9E;AAED,wBAAgB,qBAAqB,CAAC,OAAO,EAAE,MAAM,GAAG,iBAAiB,CA8BxE;AAED,UAAU,WAAW;IACnB,QAAQ,CAAC,IAAI,EAAE,QAAQ,GAAG,MAAM,CAAC;IACjC,QAAQ,CAAC,OAAO,EAAE,MAAM,CAAC;CAC1B;AAED,MAAM,WAAW,kBAAkB;IACjC,QAAQ,CAAC,KAAK,EAAE,MAAM,CAAC;IACvB,QAAQ,CAAC,qBAAqB,EAAE,MAAM,CAAC;IACvC,QAAQ,CAAC,QAAQ,EAAE,SAAS,WAAW,EAAE,CAAC;IAC1C,QAAQ,CAAC,MAAM,EAAE,IAAI,CAAC;IACtB,QAAQ,CAAC,cAAc,EAAE;QAAE,QAAQ,CAAC,aAAa,EAAE,IAAI,CAAA;KAAE,CAAC;IAC1D,8EAA8E;IAC9E,QAAQ,CAAC,WAAW,CAAC,EAAE,MAAM,CAAC;IAC9B,QAAQ,CAAC,KAAK,CAAC,EAAE,MAAM,CAAC;IACxB,QAAQ,CAAC,YAAY,CAAC,EAAE,MAAM,GAAG,UAAU,CAAC;IAC5C,QAAQ,CAAC,gBAAgB,CAAC,EAAE,MAAM,CAAC;IACnC,QAAQ,CAAC,sBAAsB,CAAC,EAAE,KAAK,CAAC;IACxC;;;oEAGgE;IAChE,QAAQ,CAAC,gBAAgB,CAAC,EAAE,SAAS,GAAG,KAAK,GAAG,QAAQ,GAAG,MAAM,CAAC;IAClE,QAAQ,CAAC,eAAe,CAAC,EACrB;QAAE,QAAQ,CAAC,IAAI,EAAE,aAAa,CAAA;KAAE,GAChC;QACE,QAAQ,CAAC,IAAI,EAAE,aAAa,CAAC;QAC7B,QAAQ,CAAC,WAAW,EAAE;YAAE,QAAQ,CAAC,IAAI,EAAE,MAAM,CAAC;YAAC,QAAQ,CAAC,MAAM,EAAE,MAAM,CAAC;YAAC,QAAQ,CAAC,MAAM,EAAE,IAAI,CAAA;SAAE,CAAC;KACjG,CAAC;IACN,QAAQ,CAAC,KAAK,CAAC,EAAE,SAAS,cAAc,EAAE,CAAC;IAC3C,QAAQ,CAAC,WAAW,CAAC,EAAE,oBAAoB,CAAC;CAC7C;AAID,wBAAgB,iBAAiB,CAAC,KAAK,EAAE,MAAM,EAAE,GAAG,EAAE,YAAY,GAAG,kBAAkB,CAuDtF;AAcD,wBAAgB,mBAAmB,CAAC,MAAM,EAAE,mBAAmB,GAAG,eAAe,CAwMhF"}

package/dist/adapters/openai.js

@@ -71,6 +71,18 @@ function classifyOpenAIFailure(message) {
             fixHint: "the candidate model was not found — fix the model id in the route's candidate list",
         };
     }
+    // CE-9 R1: a model that does not accept `reasoning_effort` rejects the request
+    // (400 "unsupported/unknown parameter reasoning_effort"). Teach it instead of
+    // surfacing a cryptic generic 400 — the whole point of forwarding verbatim
+    // rather than keeping a per-model allowlist. Mirrors gemini's invalid_argument
+    // → network + fixHint (a request-shape rejection, retryable to the next
+    // candidate, never terminal).
+    if (/reasoning_effort/.test(msg)) {
+        return {
+            cls: "network",
+            fixHint: "this model does not accept reasoning_effort — remove reasoning or pick a reasoning model (e.g. an o-series / gpt-5 reasoning model)",
+        };
+    }
     return { cls: "network" };
 }
 const DEFAULT_MAX_COMPLETION_TOKENS = 32000;
@@ -98,6 +110,14 @@ function buildOpenAIParams(model, req) {
                 ...((req.cache.ttlSeconds ?? 0) >= 86400 ? { prompt_cache_retention: "24h" } : {}),
             }
             : {}),
+        // CE-9 R1: forward reasoning effort verbatim (1:1). Absent ⇒ no key
+        // (byte-identical). The provider — not an allowlist here — decides whether
+        // the model accepts it; a rejection is classified teachably below.
+        ...(req.reasoning !== undefined ? { reasoning_effort: req.reasoning.level } : {}),
+        // response_format precedence: a native/strict json_schema (routed lane,
+        // outputSchema set) WINS; else the raw lane's mime-shaping json_object
+        // (CE-9 R4). The two never co-occur — the routed lane sets outputSchema, the
+        // raw lane responseFormat — but the precedence keeps resolution deterministic.
         ...(req.outputSchema !== null
             ? {
                 response_format: {
@@ -109,7 +129,9 @@ function buildOpenAIParams(model, req) {
                     },
                 },
             }
-            : {}),
+            : req.responseFormat === "json_object"
+                ? { response_format: { type: "json_object" } }
+                : {}),
         ...(req.tools !== undefined && req.tools.length > 0
             ? {
                 tools: (0, tool_mapping_1.toOpenAITools)(req.tools),

package/dist/cli.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"cli.d.ts","sourceRoot":"","sources":["../src/cli.ts"],"names":[],"mappings":";AACA;;;;;;;;;;;;;;;GAeG;AAiCH,MAAM,WAAW,KAAK;IACpB,QAAQ,CAAC,UAAU,EAAE,SAAS,MAAM,EAAE,CAAC;IACvC,QAAQ,CAAC,KAAK,EAAE,WAAW,CAAC,MAAM,EAAE,MAAM,CAAC,CAAC;IAC5C,QAAQ,CAAC,IAAI,EAAE,WAAW,CAAC,MAAM,CAAC,CAAC;CACpC;AAeD,wBAAgB,SAAS,CAAC,IAAI,EAAE,SAAS,MAAM,EAAE,GAAG,KAAK,CAuBxD;AA0gBD,wBAAsB,IAAI,CAAC,IAAI,EAAE,SAAS,MAAM,EAAE,GAAG,OAAO,CAAC,MAAM,CAAC,CAqCnE"}
+{"version":3,"file":"cli.d.ts","sourceRoot":"","sources":["../src/cli.ts"],"names":[],"mappings":";AACA;;;;;;;;;;;;;;;GAeG;AAkCH,MAAM,WAAW,KAAK;IACpB,QAAQ,CAAC,UAAU,EAAE,SAAS,MAAM,EAAE,CAAC;IACvC,QAAQ,CAAC,KAAK,EAAE,WAAW,CAAC,MAAM,EAAE,MAAM,CAAC,CAAC;IAC5C,QAAQ,CAAC,IAAI,EAAE,WAAW,CAAC,MAAM,CAAC,CAAC;CACpC;AAeD,wBAAgB,SAAS,CAAC,IAAI,EAAE,SAAS,MAAM,EAAE,GAAG,KAAK,CAuBxD;AA0gBD,wBAAsB,IAAI,CAAC,IAAI,EAAE,SAAS,MAAM,EAAE,GAAG,OAAO,CAAC,MAAM,CAAC,CAqCnE"}

package/dist/cli.js

@@ -35,6 +35,7 @@ const samples_1 = require("./bakeoff/samples");
 const report_1 = require("./bakeoff/report");
 const cli_run_1 = require("./verify/cli-run");
 const loader_1 = require("./routes/loader");
+const rawonly_1 = require("./run/rawonly");
 const validate_cli_1 = require("./validate-cli");
 const store_1 = require("./telemetry/store");
 const budget_1 = require("./watch/budget");
@@ -156,7 +157,7 @@ async function cmdBakeoff(flags) {
             throw new Error(`proposal ${proposalId} is ${proposal.status} — evidence runs apply to ` +
                 "suggested/evidenced proposals only");
         }
-        const routes = (0, loader_1.loadRoutes)(config.routesDir);
+        const routes = (0, loader_1.loadRoutes)((0, rawonly_1.requireRoutesDir)(config, "modelrig bakeoff"));
         const bundle = routes.get(proposal.route);
         const variants = ["default", ...bundle.variants.map((v) => v.name)];
         const lastRaw = flags.named.get("replay-last");
@@ -243,7 +244,7 @@ async function cmdWatch(flags) {
     }
     const config = (0, config_1.loadConfigFromEnv)();
     const client = proposalsClientFrom(config);
-    const routes = (0, loader_1.loadRoutes)(config.routesDir);
+    const routes = (0, loader_1.loadRoutes)((0, rawonly_1.requireRoutesDir)(config, "modelrig watch"));
     const routeEconomics = routes.list().map((b) => ({
         route: b.route,
         candidates: b.candidates.map((c) => ({ provider: c.provider, model: c.model })),
@@ -318,7 +319,7 @@ async function cmdSwapExecute(flags) {
     const config = (0, config_1.loadConfigFromEnv)();
     const proposals = proposalsClientFrom(config);
     const swaps = swapsClientFrom(config);
-    const routes = (0, loader_1.loadRoutes)(config.routesDir);
+    const routes = (0, loader_1.loadRoutes)((0, rawonly_1.requireRoutesDir)(config, "modelrig swap execute"));
     // The orchestration (guards, ordering, baseline snapshot) lives in the
     // hermetically-tested executeSwap; the CLI only injects the real config
     // deps and prints. The --approve gate is enforced here (fast-fail before
@@ -359,7 +360,7 @@ async function cmdSwapRevert(flags) {
     const incumbents = swap.baseline.incumbent_candidates ?? [];
     console.log(`swap ${swapId} marked reverted — the odometer stops accruing it at ${revertedAt}\n`);
     console.log("APPLY THIS CHANGE (git is the actuator, as always):\n");
-    const routes = (0, loader_1.loadRoutes)(config.routesDir);
+    const routes = (0, loader_1.loadRoutes)((0, rawonly_1.requireRoutesDir)(config, "modelrig swap revert"));
     let version = null;
     try {
         version = routes.get(swap.route).version + 1;