llm-switcher 1.1.6 → 1.1.7

package/README.md

@@ -191,7 +191,11 @@ flowchart LR
 
 ---
 
-## Changes in 1.1.6
+## Changes in 1.1.7
+
+- **Codex tools.** With `publicModels` set, Codex lost its tools and ended after one answer. The model catalog copied the metadata of a real OpenAI model, which puts Codex in the "Responses Lite" form. The catalog now keeps Codex in its direct tool mode, and the gateway also reads tools that arrive as an `additional_tools` input item.
+
+### Changes in 1.1.6
 
 - **Codex over WebSocket.** The gateway keeps the turns of each WebSocket session. A turn that sends `previous_response_id` gets the earlier turns back, so Codex no longer loses the task after the first tool call. An unknown id fails the turn with `previous_response_not_found`.
 - **Codex warmup.** A `response.create` frame with `generate: false` gets a local answer. It no longer spends a model call.

package/README.vi.md

@@ -191,7 +191,11 @@ flowchart LR
 
 ---
 
-## Thay đổi trong bản 1.1.6
+## Thay đổi trong bản 1.1.7
+
+- **Tool của Codex.** Khi có `publicModels`, Codex mất hết tool và dừng sau một câu trả lời. Model catalog chép metadata của một model OpenAI thật, và metadata này đưa Codex sang dạng "Responses Lite". Giờ catalog giữ Codex ở chế độ tool trực tiếp, và gateway cũng đọc tool gửi đến dưới dạng input item `additional_tools`.
+
+### Thay đổi trong bản 1.1.6
 
 - **Codex qua WebSocket.** Gateway giữ các lượt của mỗi phiên WebSocket. Lượt nào gửi `previous_response_id` sẽ nhận lại các lượt trước, nên Codex không còn mất nhiệm vụ sau lần gọi tool đầu tiên. Id không tồn tại làm lượt đó lỗi với `previous_response_not_found`.
 - **Warmup của Codex.** Frame `response.create` có `generate: false` được trả lời ngay tại máy. Frame này không còn tốn một lần gọi model.

package/formats.mjs

@@ -612,6 +612,8 @@ function responsesToIR(payload) {
     else ir.messages.push({ role: 'assistant', toolCalls: [call] });
   };
 
+  // Responses Lite sends the tools as an input item, not in payload.tools.
+  const toolList = Array.isArray(payload.tools) ? [...payload.tools] : [];
   const input = payload.input;
   if (typeof input === 'string' && input) {
     ir.messages.push({ role: 'user', content: input });
@@ -655,14 +657,19 @@ function responsesToIR(payload) {
         pushCall({ id: item.call_id || item.id || null, name: 'local_shell', args: { command: a.command || [], workdir: a.working_directory ?? undefined, timeout_ms: a.timeout_ms ?? undefined } });
       } else if (item.type === 'function_call_output' || item.type === 'custom_tool_call_output' || item.type === 'local_shell_call_output') {
         ir.messages.push({ role: 'tool', toolCallId: item.call_id || item.id, content: responsesOutputToText(item.output) });
+      } else if (item.type === 'additional_tools' && Array.isArray(item.tools)) {
+        toolList.push(...item.tools);
       }
     }
   }
 
-  if (Array.isArray(payload.tools) && payload.tools.length) {
+  if (toolList.length) {
     ir.toolMeta = {};
     const addTool = (t, namespace) => {
       if (!t || typeof t !== 'object') return;
+      // A tool in both payload.tools and an additional_tools item is sent once: upstreams refuse duplicate names.
+      if (t.name && t.type !== 'namespace' && ir.toolMeta[responsesToolName(namespace, t.name)]) return;
+      if (t.type === 'local_shell' && ir.toolMeta.local_shell) return;
       if (t.type === 'namespace' && Array.isArray(t.tools)) {
         for (const inner of t.tools) addTool(inner, t.name);
         return;
@@ -681,7 +688,7 @@ function responsesToIR(payload) {
       }
       // web_search / file_search / tool_search / image_generation...: hosted tools, skipped.
     };
-    for (const t of payload.tools) addTool(t, null);
+    for (const t of toolList) addTool(t, null);
   }
 
   const tc = payload.tool_choice;

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "llm-switcher",
-  "version": "1.1.6",
+  "version": "1.1.7",
   "description": "Zero-dependency multi-protocol edge gateway & provider switcher for Claude Code, Codex, OpenAI and Gemini clients",
   "keywords": [
     "llm",

package/state.mjs

@@ -346,8 +346,12 @@ export function buildCodexCatalog(profile) {
 
 // One catalog entry. The catalog file and the gateway's /v1/models both use it, so they cannot drift.
 export function codexModelEntry(name, is1M) {
+  // The template copies a real OpenAI model. Its Responses Lite and code modes send the tools in a
+  // form made for OpenAI's own tools, and an upstream then gets none (LS-5).
+  const { tool_mode, ...template } = structuredClone(codexCatalogTemplate() || {});
   return {
-    ...structuredClone(codexCatalogTemplate() || {}),
+    ...template,
+    use_responses_lite: false,
     slug: name,
     display_name: name,
     ...(is1M ? { context_window: 1000000, max_context_window: 1000000 } : {})

package/tests/formats.test.mjs

@@ -792,3 +792,10 @@ test('allowed_tools that matches no declared tool gives a request without tools,
     }
   }
 });
+
+test('responsesToIR reads additional_tools items and sends a tool once', () => {
+  const tool = { type: 'function', name: 'exec_command', parameters: { type: 'object' } };
+  const ir = responsesToIR({ model: 'm', tools: [tool], input: [{ type: 'additional_tools', role: 'developer', tools: [tool, { type: 'custom', name: 'apply_patch' }] }] });
+  assert.deepEqual(ir.tools.map(t => t.name), ['exec_command', 'apply_patch']);
+  assert.equal(ir.messages.length, 0, 'the item is not a message');
+});

package/tests/gateway.e2e.test.mjs

@@ -845,6 +845,26 @@ test('Codex WS: a generate:false warmup is answered locally, never sent upstream
   ws.socket.destroy();
 });
 
+// LS-5: in the Responses Lite form Codex sends its tools as an additional_tools input item of the
+// warmup, not as a tools field. The next turn must reach upstream with those tools.
+test('Codex WS: tools sent as an additional_tools item reach upstream', async () => {
+  const ws = await rawWs();
+  ws.socket.write(clientFrame(1, JSON.stringify({ type: 'response.create', model: 'main', generate: false, input: [
+    { type: 'additional_tools', id: 'at_1', role: 'developer', tools: [
+      { type: 'function', name: 'exec_command', description: 'Run a command', parameters: { type: 'object', properties: { cmd: { type: 'string' } } } },
+      { type: 'custom', name: 'apply_patch', description: 'Patch files' }] },
+    { type: 'message', role: 'developer', content: [{ type: 'input_text', text: 'You are a coding agent.' }] }] })));
+  assert.ok(await until(() => ws.messages.some(m => m.type === 'response.completed')), 'the warmup completes');
+  const warmId = ws.messages.find(m => m.type === 'response.completed').response.id;
+  const before = received.length;
+  ws.socket.write(clientFrame(1, JSON.stringify({ type: 'response.create', model: 'main', previous_response_id: warmId,
+    input: [{ type: 'message', role: 'user', content: [{ type: 'input_text', text: 'fix calc.py' }] }] })));
+  assert.ok(await until(() => received.length > before && ws.messages.filter(m => m.type === 'response.completed' || m.type === 'response.failed').length === 2), 'turn 2 ends');
+  const sent = JSON.stringify(received.slice(before).at(-1).body);
+  assert.ok(sent.includes('exec_command') && sent.includes('apply_patch'), `the tools are lost: ${sent.slice(0, 400)}`);
+  ws.socket.destroy();
+});
+
 test('Codex WS: a mid-stream error is logged with its text', async () => {
   const ws = await rawWs();
   ws.socket.write(clientFrame(1, JSON.stringify({ type: 'response.create', model: 'main', input: 'MID_STREAM_ERROR over ws' })));

package/tests/state.test.mjs

@@ -699,3 +699,13 @@ test('an empty launch lock is held, not taken over', async (t) => {
   assert.ok(waited >= 100 && waited < 4000, `the writer waited ${waited} ms`);
   assert.equal(fs.existsSync(lock), false);
 });
+
+// LS-5: the template copies a real OpenAI model that runs Codex in Responses Lite and code mode.
+// Those modes send the tools in a form made for OpenAI's own tools, so every entry turns them off.
+test('buildCodexCatalog entries keep Codex in its direct tool mode', () => {
+  const catalog = buildCodexCatalog({ inFormat: 'responses', publicModels: ['gpt-5.6-sol'] });
+  for (const entry of catalog.models) {
+    assert.equal(entry.use_responses_lite, false);
+    assert.equal(Object.hasOwn(entry, 'tool_mode'), false);
+  }
+});