fix(ai): fixed Kimi regex to match path-prefixed router model IDs

- Changed `^kimi[-.]/i` to `(^|\/)kimi[-.]/i` so Fireworks router IDs like `accounts/fireworks/routers/kimi-k2p6-turbo` correctly trigger Kimi-specific max_tokens behavior.
- Applied the same fix in both the compat detector and buildParams.
- Added a regression test covering the canonical Fire Pass router ID format.

packages/ai/src/providers/openai-completions-compat.ts

@@ -52,7 +52,7 @@ export function detectOpenAICompat(model: Model<"openai-completions">, resolvedB
 	const isCerebras = provider === "cerebras" || baseUrl.includes("cerebras.ai");
 	const isZai = provider === "zai" || baseUrl.includes("api.z.ai");
 	const isKilo = provider === "kilo" || baseUrl.includes("api.kilo.ai");
-	const isKimiModel = model.id.includes("moonshotai/kimi") || /^kimi[-.]/i.test(model.id);
+	const isKimiModel = model.id.includes("moonshotai/kimi") || /(^|\/)kimi[-.]/i.test(model.id);
 	const isMoonshotKimi =
 		isKimiModel &&
 		(provider === "moonshot" ||

packages/ai/src/providers/openai-completions.ts

@@ -1038,13 +1038,15 @@ function buildParams(
 	maybeAddOpenRouterAnthropicCacheControl(model, messages);
 	const supportsReasoningParams = model.provider !== "github-copilot";
 
-	// Kimi (including via OpenRouter and the Fireworks `accounts/fireworks/routers/kimi-*` family)
-	// calculates TPM rate limits based on max_tokens, not actual output. The official Kimi K2 model
-	// guidance (https://docs.fireworks.ai/models/kimi-k2) also requires `max_tokens` for every call
-	// since the family can otherwise emit very long reasoning traces before the final answer.
-	// Always send max_tokens — match the same Kimi-family regex used by the compat detector.
+	// Kimi (including via OpenRouter and Fireworks router-form IDs such as
+	// `accounts/fireworks/routers/kimi-*`) calculates TPM rate limits based on
+	// max_tokens, not actual output. The official Kimi K2 model guidance
+	// (https://docs.fireworks.ai/models/kimi-k2) also requires `max_tokens` for
+	// every call since the family can otherwise emit very long reasoning traces
+	// before the final answer. Always send max_tokens — match the same
+	// Kimi-family regex used by the compat detector.
 	// Note: Direct kimi-code provider is handled by the dedicated Kimi provider in kimi.ts.
-	const isKimi = model.id.includes("moonshotai/kimi") || /^kimi[-.]/i.test(model.id);
+	const isKimi = model.id.includes("moonshotai/kimi") || /(^|\/)kimi[-.]/i.test(model.id);
 	const effectiveMaxTokens = options?.maxTokens ?? (isKimi ? model.maxTokens : undefined);
 
 	const requestModelId =

packages/ai/test/firepass.test.ts

@@ -135,4 +135,37 @@ describe("Fire Pass provider", () => {
 		const parsed = JSON.parse(captured.body ?? "{}") as { max_tokens?: unknown };
 		expect(parsed.max_tokens).toBe(model.maxTokens);
 	});
+
+	it("applies the Kimi max_tokens default to canonical Fire Pass router ids", async () => {
+		const bundled = getBundledModel<"openai-completions">("firepass", "kimi-k2.6-turbo");
+		const model: Model<"openai-completions"> = {
+			...bundled,
+			id: "accounts/fireworks/routers/kimi-k2p6-turbo",
+		};
+		const captured: { body: string | null } = { body: null };
+		global.fetch = (async (_input: unknown, init?: RequestInit) => {
+			captured.body = typeof init?.body === "string" ? init.body : null;
+			return sseResponse([
+				{ choices: [{ delta: { content: "ok" }, index: 0 }] },
+				{ choices: [{ delta: {}, finish_reason: "stop", index: 0 }] },
+				"[DONE]",
+			]);
+		}) as typeof global.fetch;
+
+		const context: Context = {
+			systemPrompt: [],
+			messages: [{ role: "user", content: "ping", timestamp: Date.now() }],
+		};
+		const stream = streamOpenAICompletions(model, context, {
+			apiKey: "fpk_test",
+		});
+		for await (const _event of stream) {
+			/* drain */
+		}
+
+		expect(captured.body).not.toBeNull();
+		const parsed = JSON.parse(captured.body ?? "{}") as { max_tokens?: unknown; model?: unknown };
+		expect(parsed.model).toBe("accounts/fireworks/routers/kimi-k2p6-turbo");
+		expect(parsed.max_tokens).toBe(model.maxTokens);
+	});
 });