ampcode-connector 0.1.7 → 0.1.8

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "ampcode-connector",
-  "version": "0.1.7",
+  "version": "0.1.8",
   "description": "Proxy AmpCode through local OAuth subscriptions (Claude Code, Codex, Gemini CLI, Antigravity)",
   "license": "MIT",
   "repository": {

package/src/providers/base.ts

@@ -28,30 +28,59 @@ interface ForwardOptions {
   rewrite?: (data: string) => string;
 }
 
+const RETRYABLE_STATUS = new Set([408, 500, 502, 503, 504]);
+const MAX_RETRIES = 3;
+const RETRY_DELAY_MS = 500;
+
 export async function forward(opts: ForwardOptions): Promise<Response> {
-  const response = await fetch(opts.url, {
-    method: "POST",
-    headers: opts.headers,
-    body: opts.body,
-  });
-
-  const contentType = response.headers.get("Content-Type") ?? "application/json";
-
-  if (!response.ok) {
-    const text = await response.text();
-    logger.error(`${opts.providerName} API error (${response.status})`, { error: text.slice(0, 200) });
-    return new Response(text, { status: response.status, headers: { "Content-Type": contentType } });
-  }
+  for (let attempt = 0; attempt <= MAX_RETRIES; attempt++) {
+    let response: Response;
+    try {
+      response = await fetch(opts.url, {
+        method: "POST",
+        headers: opts.headers,
+        body: opts.body,
+      });
+    } catch (err) {
+      if (attempt < MAX_RETRIES) {
+        logger.debug(`${opts.providerName} fetch error, retry ${attempt + 1}/${MAX_RETRIES}`, {
+          error: String(err),
+        });
+        await Bun.sleep(RETRY_DELAY_MS * (attempt + 1));
+        continue;
+      }
+      throw err;
+    }
+
+    // Retry on server errors (429 handled at routing layer)
+    if (RETRYABLE_STATUS.has(response.status) && attempt < MAX_RETRIES) {
+      await response.text(); // consume body
+      logger.debug(`${opts.providerName} returned ${response.status}, retry ${attempt + 1}/${MAX_RETRIES}`);
+      await Bun.sleep(RETRY_DELAY_MS * (attempt + 1));
+      continue;
+    }
+
+    const contentType = response.headers.get("Content-Type") ?? "application/json";
+
+    if (!response.ok) {
+      const text = await response.text();
+      logger.error(`${opts.providerName} API error (${response.status})`, { error: text.slice(0, 200) });
+      return new Response(text, { status: response.status, headers: { "Content-Type": contentType } });
+    }
+
+    const isSSE = contentType.includes("text/event-stream") || opts.streaming;
+    if (isSSE) return sse.proxy(response, opts.rewrite);
 
-  const isSSE = contentType.includes("text/event-stream") || opts.streaming;
-  if (isSSE) return sse.proxy(response, opts.rewrite);
+    if (opts.rewrite) {
+      const text = await response.text();
+      return new Response(opts.rewrite(text), { status: response.status, headers: { "Content-Type": contentType } });
+    }
 
-  if (opts.rewrite) {
-    const text = await response.text();
-    return new Response(opts.rewrite(text), { status: response.status, headers: { "Content-Type": contentType } });
+    return new Response(response.body, { status: response.status, headers: { "Content-Type": contentType } });
   }
 
-  return new Response(response.body, { status: response.status, headers: { "Content-Type": contentType } });
+  // Unreachable, but TypeScript needs it
+  throw new Error(`${opts.providerName}: all retries exhausted`);
 }
 
 export function denied(providerName: string): Response {

package/src/providers/codex.ts

@@ -24,13 +24,14 @@ export const provider: Provider = {
 
   accountCount: () => oauth.accountCount(config),
 
-  async forward(sub, body, _originalHeaders, rewrite, account = 0) {
+  async forward(sub, body, originalHeaders, rewrite, account = 0) {
     const accessToken = await oauth.token(config, account);
     if (!accessToken) return denied("OpenAI Codex");
 
     const accountId = getAccountId(accessToken, account);
     const codexPath = codexPathMap[sub] ?? sub;
-    const { body: codexBody, needsResponseTransform } = transformForCodex(body.forwardBody);
+    const promptCacheKey = originalHeaders.get("x-amp-thread-id") ?? undefined;
+    const { body: codexBody, needsResponseTransform } = transformForCodex(body.forwardBody, promptCacheKey);
     const ampModel = body.ampModel ?? "gpt-5.2";
 
     const response = await forward({
@@ -50,6 +51,9 @@ export const provider: Provider = {
         "User-Agent": codexHeaderValues.USER_AGENT,
         Version: codexHeaderValues.VERSION,
         ...(accountId ? { [codexHeaders.ACCOUNT_ID]: accountId } : {}),
+        ...(promptCacheKey
+          ? { [codexHeaders.SESSION_ID]: promptCacheKey, [codexHeaders.CONVERSATION_ID]: promptCacheKey }
+          : {}),
       },
     });
 
@@ -79,7 +83,20 @@ interface ToolCallItem {
   function: { name: string; arguments: string };
 }
 
-function transformForCodex(rawBody: string): { body: string; needsResponseTransform: boolean } {
+function clampReasoningEffort(model: string, effort: string): string {
+  const modelId = model.includes("/") ? model.split("/").pop()! : model;
+  if (modelId === "gpt-5.1" && effort === "xhigh") return "high";
+  if ((modelId.startsWith("gpt-5.2") || modelId.startsWith("gpt-5.3")) && effort === "minimal") return "low";
+  if (modelId === "gpt-5.1-codex-mini") {
+    return effort === "high" || effort === "xhigh" ? "high" : "medium";
+  }
+  return effort;
+}
+
+function transformForCodex(
+  rawBody: string,
+  promptCacheKey?: string,
+): { body: string; needsResponseTransform: boolean } {
   if (!rawBody) return { body: rawBody, needsResponseTransform: false };
 
   let parsed: Record<string, unknown>;
@@ -111,6 +128,22 @@ function transformForCodex(rawBody: string): { body: string; needsResponseTransf
   // Strip id fields from input items
   if (Array.isArray(parsed.input)) {
     stripInputIds(parsed.input as Record<string, unknown>[]);
+    fixOrphanOutputs(parsed.input as Record<string, unknown>[]);
+  }
+
+  // Reasoning config — defaults match reference behavior
+  const model = (parsed.model as string) ?? "";
+  parsed.reasoning = {
+    effort: clampReasoningEffort(model, "high"),
+    summary: "auto",
+  };
+
+  parsed.text = { verbosity: "medium" };
+
+  parsed.include = ["reasoning.encrypted_content"];
+
+  if (promptCacheKey) {
+    parsed.prompt_cache_key = promptCacheKey;
   }
 
   // Remove fields the Codex backend doesn't accept
@@ -127,6 +160,21 @@ function transformForCodex(rawBody: string): { body: string; needsResponseTransf
   delete parsed.logit_bias;
   delete parsed.response_format;
 
+  // Normalize tool_choice for Responses API
+  if (parsed.tool_choice !== undefined && parsed.tool_choice !== null) {
+    if (typeof parsed.tool_choice === "string") {
+      // "auto", "none", "required" pass through as-is
+    } else if (typeof parsed.tool_choice === "object") {
+      const tc = parsed.tool_choice as Record<string, unknown>;
+      if (tc.type === "function" && tc.function) {
+        const fn = tc.function as Record<string, unknown>;
+        parsed.tool_choice = { type: "function", name: fn.name };
+      } else if (tc.type === "tool" && tc.name) {
+        parsed.tool_choice = { type: "function", name: tc.name };
+      }
+    }
+  }
+
   return { body: JSON.stringify(parsed), needsResponseTransform };
 }
 
@@ -251,6 +299,40 @@ function stripInputIds(items: Record<string, unknown>[]): void {
   }
 }
 
+/** Convert orphan function_call_output items (no matching function_call) to assistant messages. */
+function fixOrphanOutputs(items: Record<string, unknown>[]): void {
+  const callIds = new Set(
+    items.filter((i) => i.type === "function_call" && typeof i.call_id === "string").map((i) => i.call_id as string),
+  );
+  for (let i = 0; i < items.length; i++) {
+    const item = items[i]!;
+    if (item.type === "function_call_output" && typeof item.call_id === "string" && !callIds.has(item.call_id)) {
+      const toolName = typeof item.name === "string" ? (item.name as string) : "tool";
+      let text = "";
+      try {
+        text = typeof item.output === "string" ? (item.output as string) : JSON.stringify(item.output);
+      } catch {
+        text = String(item.output ?? "");
+      }
+      if (text.length > 16000) {
+        text = `${text.slice(0, 16000)}\n...[truncated]`;
+      }
+      items[i] = {
+        type: "message",
+        role: "assistant",
+        content: [
+          {
+            type: "output_text",
+            text: `[Previous ${toolName} result; call_id=${item.call_id}]: ${text}`,
+            annotations: [],
+          },
+        ],
+        status: "completed",
+      };
+    }
+  }
+}
+
 /** Extract chatgpt_account_id from JWT, falling back to stored credentials. */
 function getAccountId(accessToken: string, account: number): string | undefined {
   const creds = store.get("codex", account);

package/src/server/server.ts

@@ -17,6 +17,7 @@ export function startServer(config: ProxyConfig): ReturnType<typeof Bun.serve> {
   const server = Bun.serve({
     port: config.port,
     hostname: "localhost",
+    idleTimeout: 255, // seconds — LLM streaming responses can take minutes
 
     async fetch(req) {
       const startTime = Date.now();