Fix OpenRouter strict parameter routing for runner sessions

runner/MODEL-VALIDATION.md

@@ -14,6 +14,8 @@ Primary provider documentation checked on 2026-09-13:
 
 The record pins the Groq upstream through OpenRouter with `only: ["groq"]`, `allow_fallbacks: false` and `require_parameters: true`. The endpoint metadata reports 131,072 context tokens, 117,964 maximum completion tokens and tool support for that upstream. Other upstreams have different limits; changing upstream requires a new configuration ID. Authentication and billing go through OpenRouter; no Groq key is required. The original direct-Groq record remains available unchanged.
 
+The pinned SDK's OpenRouter serializer inherits OpenAI-specific request fields. The runner adapts its serialized body to send `max_tokens` and omit `store` and `prompt_cache_key`, which the selected endpoint does not advertise. Strict parameter matching remains enabled. The HTTP transport test checks the actual outgoing body for initial turns, tool continuations and local compaction.
+
 Provider documentation establishes advertised support. It does not prove the pinned native transport works with the account or model. The live check is still a release prerequisite.
 
 ## Limits and compaction

runner/src/ModelConfiguration.ts

@@ -8,7 +8,7 @@
 // catalog or fixture metadata.
 import { Duration, Effect, Layer, Redacted, Schema, Stream } from "effect"
 import { HttpClient, HttpClientError } from "effect/unstable/http"
-import { Auth } from "@opencode/ai/route"
+import { Auth, Route } from "@opencode/ai/route"
 import { OpenAIChat } from "@opencode/ai/protocols"
 import * as OpenRouter from "@opencode/ai/providers/openrouter"
 import { SessionRunnerModel } from "@opencode/core/session/runner/model"
@@ -126,9 +126,34 @@ const credentialFor = (record: ModelConfigurationRecord, secrets: SecretReader)
     }),
   )
 
+// The pinned SDK inherits OpenAI-only fields in its OpenRouter serializer.
+// Strict OpenRouter routing rejects those fields for the selected upstream.
+const openRouterRoute = Route.make({
+  id: "openrouter",
+  provider: OpenRouter.id,
+  providerMetadataKey: "openrouter",
+  endpoint: OpenRouter.route.endpoint,
+  framing: OpenAIChat.framing,
+  protocol: {
+    ...OpenRouter.protocol,
+    body: {
+      ...OpenRouter.protocol.body,
+      from: (request): ReturnType<typeof OpenRouter.protocol.body.from> =>
+        OpenRouter.protocol.body.from(request).pipe(
+          Effect.map(
+            ({ store: _store, prompt_cache_key: _cacheKey, max_completion_tokens, ...body }) => ({
+              ...body,
+              ...(max_completion_tokens === undefined ? {} : { max_tokens: max_completion_tokens }),
+            }),
+          ),
+        ),
+    },
+  },
+})
+
 /** Builds the native route model for a record. Secrets resolve lazily per request. */
 export const languageModelFor = (record: ModelConfigurationRecord, secrets: SecretReader) => {
-  const route = (record.route === "openrouter" ? OpenRouter.route : OpenAIChat.route).with({
+  const route = (record.route === "openrouter" ? openRouterRoute : OpenAIChat.route).with({
     provider: record.provider,
     endpoint: { baseURL: record.endpoint },
     auth: Auth.bearer(credentialFor(record, secrets)),

runner/test/ModelProvider.live.test.ts

@@ -40,6 +40,11 @@ const validate = async (live: boolean) => {
       outboundService: async (request) => {
         const body = await request.text()
         const parsed = JSON.parse(body)
+        // OpenRouter strict routing must not exclude Groq with OpenAI-only options.
+        expect(parsed).not.toHaveProperty("store")
+        expect(parsed).not.toHaveProperty("prompt_cache_key")
+        expect(parsed).not.toHaveProperty("max_completion_tokens")
+        expect(parsed.max_tokens).toBe(2048)
         // Every retry and compaction request spends this same allowance. No body rewrite.
         if (
           ++requests > 8 ||