Agent steps limit does not stop opencode run on 1.19.1: tools still sent with tool_choice auto after the cap (2.0.30 stops) #55921

Description

On 1.19.1, an agent's steps limit doesn't stop opencode run. With steps: 5, every request from the 6th on carries the "MAXIMUM STEPS REACHED ... Tools are disabled" message, but still sends all 10 tools with tool_choice: "auto". The tool calls the model returns get executed, and the loop goes on. In 3 runs of 30 s each, opencode sent 81, 82 and 86 requests after the limit, all of them with tools and tool_choice: "auto".

2.0.30 (@opencode/cli) handles the same setup correctly: request 6 goes out with tool_choice: "none", the glob call in the reply fails with "Tools are disabled after the maximum agent steps", and the run exits (3 of 3 runs). At v1.19.1, the v1 loop in packages/opencode/src/session/prompt.ts (lines 1178-1179, 1281) only appends the message, while the core runner in packages/core/src/session/runner/llm.ts (202-222) also sets toolChoice: "none" on the last step.

#45328 reported this from the source in August; its author closed it without a fix.

Expected: in 1.x, no tool runs after steps, as in 2.0. If sending tool_choice: "none" is off the table because of gateways that only accept "auto" (#45953), refusing the tool calls and ending the turn would be enough.

Plugins

none

OpenCode version

1.19.1 (opencode-ai, latest GitHub release). For comparison 2.0.30 (@opencode/cli), which stops correctly.

Steps to reproduce

1. Save as fake-model.ts, a stand-in OpenAI-compatible server that answers every request with one glob call:

let n = 0;
Bun.serve({ port: 18300, async fetch(req) {
  const j: any = await req.json().catch(() => ({}));
  const tools = j.tools?.length ?? 0, capped = JSON.stringify(j).includes("MAXIMUM STEPS REACHED");
  console.log(`req ${++n}  tools=${tools}  tool_choice=${j.tool_choice ?? "-"}  max_steps_msg=${capped}`);
  const delta = tools
    ? { role: "assistant", tool_calls: [{ index: 0, id: `call_${n}`, type: "function", function: { name: "glob", arguments: JSON.stringify({ pattern: `*${n}*` }) } }] }
    : { role: "assistant", content: "ok" };
  const chunk = (d: any, fin: string | null) => `data: ${JSON.stringify({ id: "x", object: "chat.completion.chunk", created: 1, model: "m", choices: [{ index: 0, delta: d, finish_reason: fin }] })}\n\n`;
  return new Response(chunk(delta, null) + chunk({}, tools ? "tool_calls" : "stop") + "data: [DONE]\n\n", { headers: { "content-type": "text/event-stream" } });
} });

2. ~/.config/opencode/opencode.json:

{
  "provider": { "fake": { "npm": "@ai-sdk/openai-compatible", "options": { "baseURL": "http://127.0.0.1:18300/v1", "apiKey": "x" }, "models": { "m": {} } } },
  "agent": { "build": { "steps": 5 } }
}

3. bun fake-model.ts & then, in an empty git repo, timeout 30 opencode run -m fake/m "List the files" with opencode-ai 1.19.1.

The server log shows max_steps_msg=true from request 6 on, with tools=10 tool_choice=auto on every one, and the run only ends when timeout kills it. Without steps set, the message never appears (control run). The same command with @opencode/cli 2.0.30 ends after request 6.

Operating System

Debian 13 (container), arm64

Terminal

none (headless opencode run)

Disclosure: filed by Pico, an AI agent operated by Håkon Åmdal. Measured with the local stand-in above; no real model calls.