Stop sending contextWindow as the request max_tokens

docs/IMPLEMENTATION.md

@@ -253,7 +253,7 @@ Provider and model configuration lives in JSON settings files. The global file h
 
   `models` is always an array (single- and multi-model providers are uniform). `defaultModel` (or the first entry) is used when no model is selected. With exactly one provider configured, `defaultProvider` may be omitted.
 
-  Optional `contextWindow` (positive number, tokens) overrides the models.dev / heuristic window for that provider. `loadConfig` applies it after `resolveProvider` via `setProviderContextWindowOverrides`, keyed as `<provider>:<model>` for every model on a provider that sets the field, plus the bare model id for the resolved provider so occupancy lookups that only have `source.model` still hit. It takes precedence over models.dev metadata and family heuristics. OAuth-projected Codex/xAI providers still drop the field: the synthetic `ProviderSettings` written by the projection overwrites the settings entry and does not copy `contextWindow`, so a hand-edited value on `codex/...` or `xai/...` is ignored. API-key providers are unaffected.
+  Optional `contextWindow` (positive number, tokens) overrides the models.dev / heuristic window for that provider. It sizes compaction and the status-bar meter only — it never becomes the request's `max_tokens` output budget, which stays at the shared source default on every provider branch (Codex, xAI, Go, Anthropic, Bifrost, OpenAI-compatible). `loadConfig` applies it after `resolveProvider` via `setProviderContextWindowOverrides`, keyed as `<provider>:<model>` for every model on a provider that sets the field, plus the bare model id for the resolved provider so occupancy lookups that only have `source.model` still hit. It takes precedence over models.dev metadata and family heuristics. OAuth-projected Codex/xAI providers still drop the field: the synthetic `ProviderSettings` written by the projection overwrites the settings entry and does not copy `contextWindow`, so a hand-edited value on `codex/...` or `xai/...` is ignored. API-key providers are unaffected.
 
   Optional `tools` block to arm the outer per-tool wall-clock budget (unset leaves the watchdog unarmed):
 

src/config/inference-sources.test.ts

@@ -0,0 +1,87 @@
+import { describe, test, expect, afterEach } from "bun:test";
+import type { ConversationTurn, InferenceOptions } from "@intx/types/runtime";
+import { SOURCE_MAX_TOKENS, type ProviderCatalogEntry } from "./index.js";
+import {
+  buildInferenceSourceForRef,
+  type BuildSourceContext,
+} from "./inference-sources.js";
+import type { Settings } from "./settings.js";
+import {
+  buildProviderContextWindowOverrides,
+  contextWindowFor,
+  setProviderContextWindowOverrides,
+} from "../provider/context-window.js";
+import { createOpenAICompatibleAdapter } from "../provider/openai-compatible-adapter.js";
+
+const WINDOW = 400_000;
+
+function catalog(): ProviderCatalogEntry[] {
+  return [
+    {
+      name: "fp",
+      baseURL: "https://fp.example/v1",
+      apiKey: "fp-key",
+      models: ["fp-large"],
+    },
+  ];
+}
+
+function ctx(): BuildSourceContext {
+  return { sessionId: "sess-1", catalog: catalog() };
+}
+
+function settingsWithWindow(): Settings {
+  return {
+    providers: {
+      fp: {
+        baseURL: "https://fp.example/v1",
+        apiKey: "fp-key",
+        models: ["fp-large"],
+        contextWindow: WINDOW,
+      },
+    },
+  };
+}
+
+afterEach(() => {
+  setProviderContextWindowOverrides(undefined);
+});
+
+describe("contextWindow / maxTokens split (CL-7784)", () => {
+  test("setting contextWindow does not change the source output budget", () => {
+    const source = buildInferenceSourceForRef(
+      { provider: "fp", model: "fp-large" },
+      ctx(),
+      settingsWithWindow(),
+    );
+    expect(source?.defaults?.maxTokens).toBe(SOURCE_MAX_TOKENS);
+  });
+
+  test("contextWindow 400000 does not reach the wire as max_tokens 400000", () => {
+    const source = buildInferenceSourceForRef(
+      { provider: "fp", model: "fp-large" },
+      ctx(),
+      settingsWithWindow(),
+    );
+    const adapter = createOpenAICompatibleAdapter(
+      source as unknown as Parameters<typeof createOpenAICompatibleAdapter>[0],
+    );
+    const messages = [
+      { role: "user", content: [{ type: "text", text: "hi" }] },
+    ] as unknown as ConversationTurn[];
+    const built = adapter.buildRequest(messages, "fp-large", {
+      maxTokens: source?.defaults?.maxTokens,
+    } as InferenceOptions);
+    const body = JSON.parse(built.body) as Record<string, unknown>;
+    expect(body["max_tokens"]).toBe(SOURCE_MAX_TOKENS);
+    expect(body["max_tokens"]).not.toBe(WINDOW);
+  });
+
+  test("setting contextWindow still changes contextWindowFor", () => {
+    const settings = settingsWithWindow();
+    setProviderContextWindowOverrides(
+      buildProviderContextWindowOverrides(settings.providers, "fp", "fp-large"),
+    );
+    expect(contextWindowFor("fp:fp-large")).toBe(WINDOW);
+  });
+});

src/config/inference-sources.ts

@@ -14,7 +14,6 @@ import {
   resolveSessionEffort,
   type ReasoningEffort,
 } from "../provider/reasoning-effort.js";
-import { SOURCE_MAX_TOKENS } from "./index.js";
 import { isOpenCodeGoProvider } from "../../packages/opencode-go/src/index.js";
 
 export interface BuildSourceContext {
@@ -37,16 +36,6 @@ function catalogEntry(
   return catalog.find((e) => e.name === provider);
 }
 
-function maxTokensFor(
-  settings: Settings | undefined,
-  provider: string,
-  _model: string,
-): number {
-  const cw = settings?.providers[provider]?.contextWindow;
-  if (typeof cw === "number" && cw > 0) return cw;
-  return SOURCE_MAX_TOKENS;
-}
-
 export function buildInferenceSourceForRef(
   ref: ProviderRef,
   ctx: BuildSourceContext,
@@ -57,7 +46,6 @@ export function buildInferenceSourceForRef(
   const baseURL = entry?.baseURL ?? providerSettings?.baseURL;
   if (baseURL === undefined) return null;
 
-  const maxTokens = maxTokensFor(settings, ref.provider, ref.model);
   const configured = ref.reasoningEffort ?? ctx.reasoningEffort;
   const effort =
     configured !== undefined
@@ -123,7 +111,7 @@ export function buildInferenceSourceForRef(
     });
   }
   if (entry?.bifrostVirtualKey === true) {
-    const src = buildBifrostSource({
+    return buildBifrostSource({
       id: ref.provider,
       baseURL,
       ...(entry?.apiKey !== undefined
@@ -134,13 +122,9 @@ export function buildInferenceSourceForRef(
       model: ref.model,
       ...(effort !== undefined ? { reasoningEffort: effort } : {}),
     });
-    return {
-      ...src,
-      defaults: { ...src.defaults, maxTokens },
-    };
   }
 
-  const src = buildOpenAISource({
+  return buildOpenAISource({
     id: ref.provider,
     baseURL,
     ...(entry?.apiKey !== undefined
@@ -151,10 +135,6 @@ export function buildInferenceSourceForRef(
     model: ref.model,
     ...(effort !== undefined ? { reasoningEffort: effort } : {}),
   });
-  return {
-    ...src,
-    defaults: { ...src.defaults, maxTokens },
-  };
 }
 
 function buildSourceBundle(args: {