Merge pull request #24 from southseact-3d/codex/fix-multiple-request-counting-issue

Limit tool-run loop for Chutes-like OpenAI routes and add regression test

opencode/packages/opencode/src/session/llm.ts

@@ -73,6 +73,13 @@ export namespace LLM {
     retryState: RetryState,
   ): Promise<StreamOutput> {
     const { getLanguage, cfg, provider, auth, isCodex, system, params, options, headers, maxOutputTokens, tools, rateLimitConfig } = config
+    const activeTools = Object.keys(tools).filter((x) => x !== "invalid")
+    const singleStepTools = shouldLimitToolLoop({
+      model: input.model,
+      provider,
+      options,
+      activeTools,
+    })
 
     while (retryState.attempt <= (rateLimitConfig.maxRetries || 0)) {
       try {
@@ -115,10 +122,10 @@ export namespace LLM {
           topP: params.topP,
           topK: params.topK,
           providerOptions: ProviderTransform.providerOptions(input.model, params.options),
-          activeTools: Object.keys(tools).filter((x) => x !== "invalid"),
+          activeTools,
           tools,
           // Chutes accounts each provider round-trip as a separate request, so keep SDK orchestration to one step.
-          stopWhen: input.model.providerID === "chutes" ? stepCountIs(1) : undefined,
+          stopWhen: singleStepTools ? stepCountIs(1) : undefined,
           maxOutputTokens,
           abortSignal: input.abort,
           headers: {
@@ -467,4 +474,26 @@ export namespace LLM {
     }
     return false
   }
+
+  function shouldLimitToolLoop(input: {
+    model: Provider.Model
+    provider: Provider.Info
+    options: Record<string, any>
+    activeTools: string[]
+  }) {
+    if (input.activeTools.length === 0) return false
+    if (input.model.providerID === "chutes") return true
+
+    const url = [
+      input.model.api.url,
+      input.provider.options?.baseURL,
+      input.provider.options?.baseUrl,
+      input.options?.baseURL,
+      input.options?.baseUrl,
+    ]
+      .filter((x) => typeof x === "string")
+      .map((x) => x.toLowerCase())
+
+    return url.some((x) => x.includes("chutes.ai") || x.includes("/chutes/"))
+  }
 }

opencode/packages/opencode/test/session/llm.test.ts

@@ -704,6 +704,130 @@ describe("session.llm.stream", () => {
     })
   })
 
+  test("limits tool runs to a single SDK step when OpenAI provider targets Chutes baseURL", async () => {
+    const server = state.server
+    if (!server) throw new Error("Server not initialized")
+
+    const providerID = "openai"
+    const modelID = "gpt-4o-mini"
+    const fixture = await loadFixture(providerID, modelID)
+    const model = fixture.model
+
+    const request = waitRequest(
+      "/chat/completions",
+      createEventResponse(
+        [
+          {
+            id: "chatcmpl-1",
+            object: "chat.completion.chunk",
+            choices: [{ delta: { role: "assistant" } }],
+          },
+          {
+            id: "chatcmpl-1",
+            object: "chat.completion.chunk",
+            choices: [
+              {
+                delta: {
+                  tool_calls: [
+                    {
+                      index: 0,
+                      id: "call_1",
+                      type: "function",
+                      function: {
+                        name: "echo",
+                        arguments: '{"value":"hello"}',
+                      },
+                    },
+                  ],
+                },
+              },
+            ],
+          },
+          {
+            id: "chatcmpl-1",
+            object: "chat.completion.chunk",
+            choices: [{ delta: {}, finish_reason: "tool_calls" }],
+          },
+        ],
+        true,
+      ),
+    )
+
+    await using tmp = await tmpdir({
+      init: async (dir) => {
+        await Bun.write(
+          path.join(dir, "opencode.json"),
+          JSON.stringify({
+            $schema: "https://opencode.ai/config.json",
+            enabled_providers: [providerID],
+            provider: {
+              [providerID]: {
+                options: {
+                  apiKey: "test-openai-key",
+                  baseURL: `${server.url.origin}/chutes/v1`,
+                },
+              },
+            },
+          }),
+        )
+      },
+    })
+
+    await Instance.provide({
+      directory: tmp.path,
+      fn: async () => {
+        const resolved = await Provider.getModel(providerID, model.id)
+        const sessionID = "session-test-6"
+        const agent = {
+          name: "test",
+          mode: "primary",
+          options: {},
+          permission: [{ permission: "*", pattern: "*", action: "allow" }],
+          temperature: 0.4,
+        } satisfies Agent.Info
+
+        const user = {
+          id: "user-6",
+          sessionID,
+          role: "user",
+          time: { created: Date.now() },
+          agent: agent.name,
+          model: { providerID, modelID: resolved.id },
+        } satisfies MessageV2.User
+
+        const stream = await LLM.stream({
+          user,
+          sessionID,
+          model: resolved,
+          agent,
+          system: ["You are a helpful assistant."],
+          abort: new AbortController().signal,
+          messages: [{ role: "user", content: "Use echo" }],
+          tools: {
+            echo: tool({
+              inputSchema: jsonSchema({
+                type: "object",
+                properties: {
+                  value: { type: "string" },
+                },
+                required: ["value"],
+                additionalProperties: false,
+              }),
+              execute: async () => "ok",
+            }),
+          },
+        })
+
+        for await (const _ of stream.fullStream) {
+        }
+
+        const capture = await request
+        expect(capture.url.pathname.endsWith("/chat/completions")).toBe(true)
+        expect(state.queue.length).toBe(0)
+      },
+    })
+  })
+
   test("sends Google API payload for Gemini models", async () => {
     const server = state.server
     if (!server) {