fix(config): prefer context_length from API in model discovery

Updated model discovery to use `context_length` reported by the API when available, falling back to bundled reference data and default. Added `context_length` field to parsed model response and modified context window assignment logic.

packages/agent/CHANGELOG.md

@@ -16,6 +16,9 @@
 ### Fixed
 
 - Fixed dynamic forced tool choices from queue hooks being filtered against the active per-turn tool set before provider dispatch. ([#1701](https://github.com/can1357/oh-my-pi/issues/1701))
+### Fixed
+- Honor `context_length` reported by OpenAI-compatible proxy `/v1/models` discovery (`discovery: { type: "proxy" }` or `discovery: { type: "openai-models-list" }`), so aggregator-reported windows override the bundled reference when they differ.
+
 
 ## [15.12.4] - 2026-06-13
 ### Fixed

packages/coding-agent/src/config/model-discovery.ts

@@ -393,7 +393,7 @@ export async function discoverOpenAIModelsList(
 	const response = apiKey
 		? await withAuth(apiKey, key => attempt({ ...baseHeaders, Authorization: `Bearer ${key}` }))
 		: await attempt(baseHeaders);
-	const payload = (await response.json()) as { data?: Array<{ id: string }> };
+	const payload = (await response.json()) as { data?: Array<{ id: string; context_length?: number }> };
 	const models = payload.data ?? [];
 	const discovered: Model<Api>[] = [];
 	for (const item of models) {
@@ -409,7 +409,7 @@ export async function discoverOpenAIModelsList(
 				reasoning: false,
 				input: ["text"],
 				cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 },
-				contextWindow: 128000,
+				contextWindow: toPositiveNumberOrUndefined(item.context_length) ?? 128000,
 				maxTokens: discoveryDefaultMaxTokens(providerConfig.api),
 				headers,
 				compat: {
@@ -463,7 +463,7 @@ export async function discoverProxyModels(
 		? await withAuth(apiKey, key => attempt({ ...baseHeaders, Authorization: `Bearer ${key}` }))
 		: await attempt(baseHeaders);
 	const payload = (await response.json()) as {
-		data?: Array<{ id?: string; name?: string; supported_endpoint_types?: string[] }>;
+		data?: Array<{ id?: string; name?: string; supported_endpoint_types?: string[]; context_length?: number }>;
 	};
 	const items = payload.data ?? [];
 	const discovered: Model<Api>[] = [];
@@ -499,7 +499,12 @@ export async function discoverProxyModels(
 				// upstream bundled catalogs, so keep costs local-unknown even when
 				// we successfully recover the upstream model identity.
 				cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 },
-				contextWindow: reference?.contextWindow ?? 128000,
+				// Prefer the context_length the API reports for this model; fall
+				// back to the bundled reference, then a sane default.
+				contextWindow: 
+					toPositiveNumberOrUndefined(item.context_length) ??
+					reference?.contextWindow ??
+					128000,
 				maxTokens: reference?.maxTokens ?? discoveryDefaultMaxTokens(api),
 				headers,
 				// OpenAI-compat fields are no-ops on anthropic models; the

packages/coding-agent/test/model-discovery.test.ts

@@ -607,4 +607,68 @@ describe("ModelRegistry runtime discovery", () => {
 		expect(llama?.maxTokens).toBe(32_768);
 		expect(llama?.input).toEqual(["text", "image"]);
 	});
+	test("openai-models-list discovery honors API-reported context_length over fallback", async () => {
+		writeRawModelsJson({
+			"openai-test": {
+				baseUrl: "http://127.0.0.1:9999",
+				api: "openai-completions",
+				auth: "none",
+				discovery: { type: "openai-models-list" },
+			},
+		});
+		const fetchMock: FetchImpl = async input => {
+			const url = String(input);
+			if (url === "http://127.0.0.1:9999/v1/models") {
+				return new Response(
+					JSON.stringify({
+						data: [
+							{ id: "openai-test/contextual-model", context_length: 16385 },
+							{ id: "openai-test/no-context-model" },
+						],
+					}),
+					{ status: 200, headers: { "Content-Type": "application/json" } },
+				);
+			}
+			throw new Error(`Unexpected URL: ${url}`);
+		};
+		const registry = new ModelRegistry(authStorage, modelsJsonPath, { fetch: fetchMock });
+		await registry.refresh();
+		const contextual = registry.getAll().find(m => m.provider === "openai-test" && m.id === "openai-test/contextual-model");
+		expect(contextual?.contextWindow).toBe(16385);
+		const fallback = registry.getAll().find(m => m.provider === "openai-test" && m.id === "openai-test/no-context-model");
+		expect(fallback?.contextWindow).toBe(128000);
+	});
+
+	test("proxy discovery honors API-reported context_length and endpoint routing", async () => {
+		writeRawModelsJson({
+			"proxy-test": {
+				baseUrl: "http://127.0.0.1:9998",
+				auth: "none",
+				discovery: { type: "proxy" },
+			},
+		});
+		const fetchMock: FetchImpl = async input => {
+			const url = String(input);
+			if (url === "http://127.0.0.1:9998/v1/models") {
+				return new Response(
+					JSON.stringify({
+						data: [
+							{ id: "anthropic-model", supported_endpoint_types: ["anthropic"], context_length: 200000 },
+							{ id: "openai-model", supported_endpoint_types: ["openai"], context_length: 65536 },
+						],
+					}),
+					{ status: 200, headers: { "Content-Type": "application/json" } },
+				);
+			}
+			throw new Error(`Unexpected URL: ${url}`);
+		};
+		const registry = new ModelRegistry(authStorage, modelsJsonPath, { fetch: fetchMock });
+		await registry.refresh();
+		const anthropic = registry.getAll().find(m => m.provider === "proxy-test" && m.id === "anthropic-model");
+		expect(anthropic?.api).toBe("anthropic-messages");
+		expect(anthropic?.contextWindow).toBe(200000);
+		const openai = registry.getAll().find(m => m.provider === "proxy-test" && m.id === "openai-model");
+		expect(openai?.api).toBe("openai-completions");
+		expect(openai?.contextWindow).toBe(65536);
+	});
 });