fix(open-sse): treat CommandCode in-stream error events as request failures

Upstream emits AI SDK v5 {"type":"error"} events inside an HTTP 200 stream.
The translator turned them into fake success content ([CommandCode error: ...]
+ finish_reason stop), so account/model fallback never fired and logs showed
Status: success.

- translator: error events now emit an OpenAI-shaped error chunk (chunk.error)
  instead of content; parseSSEToOpenAIResponse already detects chunk?.error
- executor: peek the first events before committing the response; an early
  error event returns 502 so fallback runs before any byte reaches the client

open-sse/executors/commandcode.js

@@ -1,6 +1,7 @@
 import { randomUUID } from "crypto";
 import { BaseExecutor } from "./base.js";
 import { PROVIDERS } from "../config/providers.js";
+import { HTTP_STATUS } from "../config/runtimeConfig.js";
 import { commandCodeToOpenAIResponse } from "../translator/response/commandcode-to-openai.js";
 import { SSE_DONE } from "../utils/sseConstants.js";
 
@@ -14,6 +15,12 @@ import { SSE_DONE } from "../utils/sseConstants.js";
  * We translate each event to an OpenAI chat.completion.chunk and emit it as SSE so
  * both the streaming and non-streaming (forced SSE → JSON) downstream handlers in
  * 9router can consume it without further format translation.
+ *
+ * Terminal upstream failures arrive as `{"type":"error"}` events inside the HTTP
+ * 200 stream, so a plain `response.ok` check cannot see them. We peek the first
+ * events before committing the response (see peekForUpstreamError) so a stream
+ * that starts with an error fails fast — the normal `!response.ok` path then
+ * triggers account/model fallback instead of streaming fake success content.
  */
 export class CommandCodeExecutor extends BaseExecutor {
 	constructor() {
@@ -42,11 +49,192 @@ export class CommandCodeExecutor extends BaseExecutor {
 	async execute(opts) {
 		const result = await super.execute(opts);
 		if (!result?.response?.ok || !result.response.body) return result;
-    result.response = wrapNdjsonAsOpenAISse(result.response, opts.model);
+		result.response = await peekForUpstreamError(result.response, opts.model, {
+			signal: opts.signal,
+		});
 		return result;
 	}
 }
 
+// How long to hold the response open while peeking the first upstream events.
+// An upstream error event ("Network connection lost") is emitted at stream
+// start, so the peek is fast; the bound just prevents a slow-started stream
+// from being held hostage. Env: COMMANDCODE_PEEK_TIMEOUT_MS.
+const PEEK_TIMEOUT_MS = (() => {
+	const raw = process.env.COMMANDCODE_PEEK_TIMEOUT_MS;
+	const n = raw ? parseInt(raw, 10) : NaN;
+	return Number.isFinite(n) && n > 0 ? n : 10 * 1000;
+})();
+
+// Event types that count as "the stream has started producing". Everything
+// else (start, start-step, reasoning-start, text-start, ...) is metadata and
+// does not end the peek.
+const MEANINGFUL_EVENT_TYPES = new Set([
+	"text-delta",
+	"reasoning-delta",
+	"tool-input-start",
+	"tool-input-delta",
+	"tool-input-end",
+	"tool-call",
+	"finish-step",
+	"finish",
+]);
+
+function makeAbortError(reason) {
+	const error = new Error(reason?.message || reason || "Request aborted");
+	error.name = "AbortError";
+	return error;
+}
+
+function tryParseEvent(line) {
+	const trimmed = line.trim();
+	if (!trimmed) return null;
+	const json = trimmed.startsWith("data:") ? trimmed.slice(5).trim() : trimmed;
+	if (!json || json === "[DONE]") return null;
+	try {
+		return JSON.parse(json);
+	} catch {
+		return null;
+	}
+}
+
+function formatErrorValue(errVal) {
+	const errStr =
+		typeof errVal === "string"
+			? errVal
+			: typeof errVal?.message === "string"
+				? errVal.message
+				: JSON.stringify(errVal);
+	const errType =
+		typeof errVal === "string"
+			? "upstream_error"
+			: errVal?.type || "upstream_error";
+	return { message: errStr, type: errType };
+}
+
+/**
+ * Read the first upstream events before committing the response.
+ *
+ * - `{"type":"error"}` as the first meaningful event → return a 502 Response so
+ *   chatCore's `!response.ok` path parses the error and triggers fallback.
+ * - Otherwise → re-emit the buffered bytes + the rest of the stream through the
+ *   normal NDJSON → OpenAI SSE wrapper and return it untouched in spirit.
+ *
+ * Bounded by `timeoutMs` (default PEEK_TIMEOUT_MS): if no meaningful event
+ * arrives in time, or the request signal aborts, we commit whatever we have and
+ * let the regular stream pipeline (stall detection, abort handling) take over.
+ */
+export async function peekForUpstreamError(
+	originalResponse,
+	model,
+	{ signal = null, timeoutMs = PEEK_TIMEOUT_MS } = {},
+) {
+	const reader = originalResponse.body.getReader();
+	const decoder = new TextDecoder();
+	const encoder = new TextEncoder();
+	const abortController = new AbortController();
+	const forwardAbort = () => abortController.abort(signal?.reason);
+	if (signal?.aborted) abortController.abort(signal?.reason);
+	else if (signal)
+		signal.addEventListener("abort", forwardAbort, { once: true });
+
+	let peeked = "";
+	let errorEvent = null;
+	let committed = false;
+
+	const readWithTimeout = (ms) => {
+		if (abortController.signal.aborted) {
+			return Promise.reject(makeAbortError(abortController.signal.reason));
+		}
+		const timeoutPromise = new Promise((_, reject) => {
+			const t = setTimeout(() => reject(new Error("peek timeout")), ms);
+			t.unref?.();
+		});
+		const abortPromise = new Promise((_, reject) => {
+			abortController.signal.addEventListener(
+				"abort",
+				() => reject(makeAbortError(abortController.signal.reason)),
+				{ once: true },
+			);
+		});
+		return Promise.race([reader.read(), timeoutPromise, abortPromise]);
+	};
+
+	try {
+		const deadline = Date.now() + timeoutMs;
+		while (!errorEvent && !committed && Date.now() < deadline) {
+			const { done, value } = await readWithTimeout(
+				Math.max(deadline - Date.now(), 1),
+			);
+			if (done) break;
+			peeked += decoder.decode(value, { stream: true });
+			const lines = peeked.split("\n");
+			// The last segment may be a partial line — only parse complete ones.
+			for (const line of lines.slice(0, -1)) {
+				const event = tryParseEvent(line);
+				if (!event?.type) continue;
+				if (event.type === "error") {
+					errorEvent = event;
+					break;
+				}
+				if (MEANINGFUL_EVENT_TYPES.has(event.type)) {
+					committed = true;
+					break;
+				}
+			}
+		}
+	} catch {
+		// timeout / abort / read failure during the peek → commit whatever we have;
+		// the downstream stream pipeline (stall detection, abort handling) takes over.
+	}
+
+	// Flush any partial multi-byte UTF-8 sequence held by the decoder so the
+	// re-encoded peeked bytes round-trip losslessly.
+	peeked += decoder.decode();
+
+	if (signal) signal.removeEventListener("abort", forwardAbort);
+
+	if (errorEvent) {
+		await reader.cancel("commandcode early error detected").catch(() => {});
+		const { message, type } = formatErrorValue(
+			errorEvent.error ?? errorEvent.message ?? "unknown",
+		);
+		return new Response(JSON.stringify({ error: { message, type } }), {
+			status: HTTP_STATUS.BAD_GATEWAY,
+			statusText: message.slice(0, 200),
+			headers: { "Content-Type": "application/json" },
+		});
+	}
+
+	const remaining = new ReadableStream({
+		start(controller) {
+			(async () => {
+				try {
+					if (peeked) controller.enqueue(encoder.encode(peeked));
+					while (true) {
+						const { done, value } = await reader.read();
+						if (done) break;
+						controller.enqueue(value);
+					}
+					controller.close();
+				} catch (err) {
+					controller.error(err);
+				}
+			})();
+		},
+		cancel() {
+			reader.cancel("commandcode stream cancelled").catch(() => {});
+		},
+	});
+
+	const combined = new Response(remaining, {
+		status: originalResponse.status,
+		statusText: originalResponse.statusText,
+		headers: originalResponse.headers,
+	});
+	return wrapNdjsonAsOpenAISse(combined, model);
+}
+
 function wrapNdjsonAsOpenAISse(originalResponse, model) {
 	const decoder = new TextDecoder();
 	const encoder = new TextEncoder();

open-sse/translator/response/commandcode-to-openai.js

@@ -43,7 +43,7 @@ function makeChunk(state, delta, finishReason = null) {
 	return buildChunk(
 		{ id: state.responseId, created: state.created, model: state.model },
 		delta,
-    finishReason
+		finishReason,
 	);
 }
 
@@ -53,7 +53,11 @@ export function commandCodeToOpenAIResponse(chunk, state) {
 	if (!chunk) return null;
 
 	// Already-OpenAI chunk: pass through
-  if (chunk && typeof chunk === "object" && chunk.object === "chat.completion.chunk") {
+	if (
+		chunk &&
+		typeof chunk === "object" &&
+		chunk.object === "chat.completion.chunk"
+	) {
 		return chunk;
 	}
 
@@ -81,7 +85,10 @@ export function commandCodeToOpenAIResponse(chunk, state) {
 		case "text-delta": {
 			const text = event.text || event.delta || "";
 			if (!text) break;
-      const delta = state.chunkIndex === 0 ? { role: ROLE.ASSISTANT, content: text } : { content: text };
+			const delta =
+				state.chunkIndex === 0
+					? { role: ROLE.ASSISTANT, content: text }
+					: { content: text };
 			state.chunkIndex++;
 			state.openText = true;
 			out.push(makeChunk(state, delta));
@@ -97,7 +104,8 @@ export function commandCodeToOpenAIResponse(chunk, state) {
 			break;
 		}
 		case "tool-input-start": {
-      const id = event.id || event.toolCallId || fallbackToolCallId(state.toolIndex);
+			const id =
+				event.id || event.toolCallId || fallbackToolCallId(state.toolIndex);
 			let idx = state.toolIndexById.get(id);
 			if (idx == null) {
 				idx = state.toolIndex++;
@@ -106,12 +114,14 @@ export function commandCodeToOpenAIResponse(chunk, state) {
 			state.openTools.add(id);
 			const delta = {
 				...(state.chunkIndex === 0 ? { role: ROLE.ASSISTANT } : {}),
-        tool_calls: [{
+				tool_calls: [
+					{
 						index: idx,
 						id,
 						type: OPENAI_BLOCK.FUNCTION,
 						function: { name: event.toolName || "", arguments: "" },
-        }],
+					},
+				],
 			};
 			state.chunkIndex++;
 			out.push(makeChunk(state, delta));
@@ -122,10 +132,12 @@ export function commandCodeToOpenAIResponse(chunk, state) {
 			const idx = state.toolIndexById.get(id);
 			if (idx == null) break;
 			const delta = {
-        tool_calls: [{
+				tool_calls: [
+					{
 						index: idx,
 						function: { arguments: event.delta || event.inputTextDelta || "" },
-        }],
+					},
+				],
 			};
 			out.push(makeChunk(state, delta));
 			break;
@@ -136,15 +148,20 @@ export function commandCodeToOpenAIResponse(chunk, state) {
 			if (state.toolIndexById.has(id)) break;
 			const idx = state.toolIndex++;
 			state.toolIndexById.set(id, idx);
-      const argsStr = typeof event.input === "string" ? event.input : JSON.stringify(event.input ?? {});
+			const argsStr =
+				typeof event.input === "string"
+					? event.input
+					: JSON.stringify(event.input ?? {});
 			const delta = {
 				...(state.chunkIndex === 0 ? { role: ROLE.ASSISTANT } : {}),
-        tool_calls: [{
+				tool_calls: [
+					{
 						index: idx,
 						id,
 						type: OPENAI_BLOCK.FUNCTION,
 						function: { name: event.toolName || "", arguments: argsStr },
-        }],
+					},
+				],
 			};
 			state.chunkIndex++;
 			out.push(makeChunk(state, delta));
@@ -156,7 +173,8 @@ export function commandCodeToOpenAIResponse(chunk, state) {
 			break;
 		}
 		case "finish": {
-      const finishReason = state.finishReason || mapFinishReason(event.finishReason || "stop");
+			const finishReason =
+				state.finishReason || mapFinishReason(event.finishReason || "stop");
 			const finalChunk = makeChunk(state, {}, finishReason);
 			const totalUsage = event.totalUsage || state.usage;
 			const usage = toOpenAIUsage(totalUsage, "commandcode");
@@ -165,10 +183,25 @@ export function commandCodeToOpenAIResponse(chunk, state) {
 			break;
 		}
 		case "error": {
+			// Terminal upstream failure (AI SDK v5 error event) — NOT content. Emit an
+			// OpenAI-shaped error chunk (chunk.error) so downstream — parseSSEToOpenAIResponse
+			// for non-streaming, OpenAI SDK clients for streaming — treats the request as
+			// failed instead of surfacing fake success content like "[CommandCode error: ...]".
 			state.finishReason = OPENAI_FINISH.STOP;
 			const errVal = event.error ?? event.message ?? "unknown";
-      const errStr = typeof errVal === "string" ? errVal : JSON.stringify(errVal);
+			const errStr =
-      out.push(makeChunk(state, { content: `\n\n[CommandCode error: ${errStr}]` }));
+				typeof errVal === "string"
+					? errVal
+					: typeof errVal?.message === "string"
+						? errVal.message
+						: JSON.stringify(errVal);
+			const errType =
+				typeof errVal === "string"
+					? "upstream_error"
+					: errVal?.type || "upstream_error";
+			const errChunk = makeChunk(state, {});
+			errChunk.error = { message: errStr, type: errType };
+			out.push(errChunk);
 			out.push(makeChunk(state, {}, OPENAI_FINISH.STOP));
 			break;
 		}
@@ -181,4 +214,9 @@ export function commandCodeToOpenAIResponse(chunk, state) {
 	return out.length ? out : null;
 }
 
-register(FORMATS.COMMANDCODE, FORMATS.OPENAI, null, commandCodeToOpenAIResponse);
+register(
+	FORMATS.COMMANDCODE,
+	FORMATS.OPENAI,
+	null,
+	commandCodeToOpenAIResponse,
+);

tests/unit/commandcode-executor.test.js

@@ -0,0 +1,101 @@
+/**
+ * Unit tests for the CommandCode executor early-error peek.
+ *
+ * The upstream emits AI SDK v5 NDJSON over an HTTP 200 stream, so a terminal
+ * `{"type":"error"}` event is invisible to the normal `response.ok` success
+ * check. `peekForUpstreamError` reads the first events before committing the
+ * response: an error event → non-ok Response (fallback can kick in); otherwise
+ * the buffered bytes are re-emitted and streaming proceeds as before.
+ */
+import { describe, it, expect } from "vitest";
+import { peekForUpstreamError } from "../../open-sse/executors/commandcode.js";
+
+const encoder = new TextEncoder();
+
+function ndjsonResponse(lines) {
+	const body = new ReadableStream({
+		start(controller) {
+			for (const line of lines) controller.enqueue(encoder.encode(line + "\n"));
+			controller.close();
+		},
+	});
+	return new Response(body, {
+		status: 200,
+		headers: { "content-type": "text/event-stream" },
+	});
+}
+
+describe("commandcode executor — early-error peek", () => {
+	it("returns 502 when the first meaningful event is an error", async () => {
+		const res = await peekForUpstreamError(
+			ndjsonResponse([
+				'{"type":"error","error":{"type":"server_error","message":"Network connection lost."}}',
+			]),
+			"model",
+		);
+		expect(res.status).toBe(502);
+		const body = await res.json();
+		expect(body.error.message).toBe("Network connection lost.");
+		expect(body.error.type).toBe("server_error");
+	});
+
+	it("detects an error event even when metadata events arrive first", async () => {
+		const res = await peekForUpstreamError(
+			ndjsonResponse([
+				'{"type":"start"}',
+				'{"type":"start-step"}',
+				'{"type":"error","error":{"type":"server_error","message":"Network connection lost."}}',
+			]),
+			"model",
+		);
+		expect(res.status).toBe(502);
+		const body = await res.json();
+		expect(body.error.message).toContain("Network connection lost");
+	});
+
+	it("commits and streams normally when the first meaningful event is content", async () => {
+		const res = await peekForUpstreamError(
+			ndjsonResponse([
+				'{"type":"start"}',
+				'{"type":"text-delta","text":"hi there"}',
+				'{"type":"finish"}',
+			]),
+			"model",
+		);
+		expect(res.status).toBe(200);
+		const text = await res.text();
+		expect(text).toContain('"content":"hi there"');
+		expect(text).not.toContain("[CommandCode error:");
+	});
+
+	it("commits when the stream ends without any event", async () => {
+		const res = await peekForUpstreamError(ndjsonResponse([]), "model");
+		expect(res.status).toBe(200);
+		await res.body.cancel();
+	});
+
+	it("commits (does not hang) when no event arrives before the peek timeout", async () => {
+		const stalled = new Response(new ReadableStream({ start() {} }), {
+			status: 200,
+			headers: { "content-type": "text/event-stream" },
+		});
+		const res = await peekForUpstreamError(stalled, "model", { timeoutMs: 50 });
+		expect(res.status).toBe(200);
+		await res.body.cancel();
+	});
+
+	it("does not hang when the request signal aborts during the peek", async () => {
+		const controller = new AbortController();
+		const stalled = new Response(new ReadableStream({ start() {} }), {
+			status: 200,
+			headers: { "content-type": "text/event-stream" },
+		});
+		setTimeout(() => controller.abort(new Error("client gone")), 10);
+		const res = await peekForUpstreamError(stalled, "model", {
+			signal: controller.signal,
+			timeoutMs: 2000,
+		});
+		expect(res.status).toBe(200);
+		await res.body.cancel();
+	});
+});

tests/unit/commandcode-to-openai.test.js

@@ -36,9 +36,7 @@ describe("commandcode-to-openai — text-delta", () => {
 
 describe("commandcode-to-openai — reasoning-delta", () => {
 	it("maps reasoning-delta to reasoning_content delta", () => {
-    const { chunks } = feed([
+		const { chunks } = feed([{ type: "reasoning-delta", text: "thinking..." }]);
-      { type: "reasoning-delta", text: "thinking..." },
-    ]);
 		expect(chunks[0].choices[0].delta.reasoning_content).toBe("thinking...");
 	});
 });
@@ -47,8 +45,8 @@ describe("commandcode-to-openai — tool-input-* with id field (live schema)", (
 	it("registers tool index using event.id (NOT toolCallId)", () => {
 		const { chunks } = feed([
 			{ type: "tool-input-start", id: "call_X", toolName: "Bash" },
-      { type: "tool-input-delta", id: "call_X", delta: "{\"cmd" },
+			{ type: "tool-input-delta", id: "call_X", delta: '{"cmd' },
-      { type: "tool-input-delta", id: "call_X", delta: "\":\"ls\"}" },
+			{ type: "tool-input-delta", id: "call_X", delta: '":"ls"}' },
 		]);
 
 		// First chunk emits tool_calls with id
@@ -57,8 +55,12 @@ describe("commandcode-to-openai — tool-input-* with id field (live schema)", (
 		expect(startChunk.function.name).toBe("Bash");
 
 		// Subsequent deltas accumulate arguments
-    expect(chunks[1].choices[0].delta.tool_calls[0].function.arguments).toBe("{\"cmd");
+		expect(chunks[1].choices[0].delta.tool_calls[0].function.arguments).toBe(
-    expect(chunks[2].choices[0].delta.tool_calls[0].function.arguments).toBe("\":\"ls\"}");
+			'{"cmd',
+		);
+		expect(chunks[2].choices[0].delta.tool_calls[0].function.arguments).toBe(
+			'":"ls"}',
+		);
 	});
 
 	it("ignores tool-input-delta when id is unknown (no prior start)", () => {
@@ -73,8 +75,13 @@ describe("commandcode-to-openai — final tool-call event", () => {
 	it("does NOT re-emit tool_calls when tool-input-* deltas already fired", () => {
 		const { chunks } = feed([
 			{ type: "tool-input-start", id: "call_Y", toolName: "Write" },
-      { type: "tool-input-delta", id: "call_Y", delta: "{\"file\":\"a\"}" },
+			{ type: "tool-input-delta", id: "call_Y", delta: '{"file":"a"}' },
-      { type: "tool-call", toolCallId: "call_Y", toolName: "Write", input: { file: "a" } },
+			{
+				type: "tool-call",
+				toolCallId: "call_Y",
+				toolName: "Write",
+				input: { file: "a" },
+			},
 		]);
 		// Should be exactly 2 chunks (start + delta), no duplicate from final tool-call
 		expect(chunks.length).toBe(2);
@@ -82,7 +89,12 @@ describe("commandcode-to-openai — final tool-call event", () => {
 
 	it("emits a consolidated tool_calls when only the final tool-call event arrives", () => {
 		const { chunks } = feed([
-      { type: "tool-call", toolCallId: "call_Z", toolName: "Read", input: { path: "/x" } },
+			{
+				type: "tool-call",
+				toolCallId: "call_Z",
+				toolName: "Read",
+				input: { path: "/x" },
+			},
 		]);
 		expect(chunks.length).toBe(1);
 		const tc = chunks[0].choices[0].delta.tool_calls[0];
@@ -107,21 +119,45 @@ describe("commandcode-to-openai — finish", () => {
 	it("includes usage on the final chunk when totalUsage provided", () => {
 		const { chunks } = feed([
 			{ type: "text-delta", text: "hi" },
-      { type: "finish-step", finishReason: "stop", usage: { inputTokens: 10, outputTokens: 5, totalTokens: 15 } },
+			{
-      { type: "finish", totalUsage: { inputTokens: 10, outputTokens: 5, totalTokens: 15 } },
+				type: "finish-step",
+				finishReason: "stop",
+				usage: { inputTokens: 10, outputTokens: 5, totalTokens: 15 },
+			},
+			{
+				type: "finish",
+				totalUsage: { inputTokens: 10, outputTokens: 5, totalTokens: 15 },
+			},
 		]);
 		const last = chunks[chunks.length - 1];
-    expect(last.usage).toEqual({ prompt_tokens: 10, completion_tokens: 5, total_tokens: 15 });
+		expect(last.usage).toEqual({
+			prompt_tokens: 10,
+			completion_tokens: 5,
+			total_tokens: 15,
+		});
 	});
 });
 
 describe("commandcode-to-openai — error event", () => {
-  it("stringifies object errors so client sees readable message", () => {
+	it("emits an OpenAI-shaped error chunk instead of fake success content", () => {
 		const { chunks } = feed([
 			{ type: "error", error: { type: "server_error", message: "Boom" } },
 		]);
-    const text = chunks[0].choices[0].delta.content;
+		expect(chunks[0].error).toEqual({ message: "Boom", type: "server_error" });
-    expect(text).toContain("Boom");
+		expect(chunks[0].choices[0].delta.content).toBeUndefined();
-    expect(text).not.toContain("[object Object]");
+		expect(chunks[1].choices[0].finish_reason).toBe("stop");
+		expect(JSON.stringify(chunks)).not.toContain("[CommandCode error:");
+	});
+
+	it("keeps the stream terminal so clients do not hang waiting for more", () => {
+		const { chunks } = feed([
+			{ type: "start" },
+			{
+				type: "error",
+				error: { type: "server_error", message: "Network connection lost." },
+			},
+		]);
+		expect(chunks[0].error.message).toBe("Network connection lost.");
+		expect(chunks[1].choices[0].finish_reason).toBe("stop");
 	});
 });