Repetitive tool-call loops with alibaba-coding-plan-cn/qwen3.6-plus

[Astro-Han]

Description

When using OpenCode with alibaba-coding-plan-cn/qwen3.6-plus, some sessions fall into repetitive tool-call loops.

What I observe:

• One user turn turns into many assistant turns.
• Adjacent assistant turns often contain nearly identical reasoning.
• Each assistant turn often performs only one tool call, then re-enters the loop.
• The session usually eventually finishes, so this does not look like a storage or transport hang.
• This seems more like a harness-level loop amplification issue that is exposed more often by qwen3.6-plus.

Local evidence from my machine:

• Session ses_2845d8980ffemjHrBrf2XpIrY7 had 18 user messages and 114 assistant messages.
• Many adjacent assistant turns repeated almost the same reasoning, especially around checking or updating the same section again after each tool result.
• In a local sample, qwen3.6-plus sessions showed a higher assistant-to-user ratio than recent opencode-go/glm-5.1 sessions.
• Across stored assistant messages, Qwen also produced many more single-tool turns before reaching a final stop.

This does not look like a simple “session got stuck forever” bug. It looks more like the loop keeps accepting repeated tool-calls from this model without enough guardrails against repetition.

Relevant code paths that may be involved:

• packages/opencode/src/session/prompt.ts
• packages/opencode/src/agent/agent.ts

My current suspicion:

• OpenCode keeps looping while the assistant finishes with tool-calls.
• The default build agent appears to have no practical default step limit.
• There does not seem to be a repetition guard for “same reasoning / same tool pattern again”.
• qwen3.6-plus seems more likely than some other models to produce serial one-tool-at-a-time continuations, which makes the issue much more visible.

Plugins

No custom plugins involved, as far as I can tell.

OpenCode version

Mostly observed on 1.4.3.

Affected local sessions I checked:

• 1.4.3: ses_2845d8980ffemjHrBrf2XpIrY7, ses_289442f7affexh49ZQ2S655J74, ses_28a631ed9ffedBp4DospJDEqoE
• 1.4.0: ses_28f1e807cffe55PnIHB9b7fK2C

I also have local qwen3.6-plus sessions on 1.3.13 and 1.4.1, but the strongest examples I inspected were on 1.4.3.

Steps to reproduce

1. Configure OpenCode to use alibaba-coding-plan-cn/qwen3.6-plus.
2. Use the default build agent on a task that requires multiple reads, edits, or repeated verification inside a repo.
3. Watch the session timeline.
4. In some cases, the assistant will produce many short turns with near-duplicate reasoning and one tool call per turn before finally stopping.

A representative pattern is:

1. Assistant reasons briefly.
2. Assistant performs one tool call.
3. Assistant re-enters with nearly the same reasoning.
4. Assistant performs another single tool call.
5. This repeats much longer than expected.

Screenshot and/or share link

I do not have a share link attached yet.

I can provide sanitized session details if needed. Example local session IDs:

• ses_2845d8980ffemjHrBrf2XpIrY7
• ses_289442f7affexh49ZQ2S655J74
• ses_28a631ed9ffedBp4DospJDEqoE

Operating System

macOS 26.3.1

Terminal

Not verified from the current environment.

[github-actions]

This issue might be a duplicate of existing issues. Please check:

• GLM-5 model gets stuck in infinite loop, repeating same actions #18415: GLM-5 (Alibaba Coding Plan) model gets stuck in infinite loop repeating same actions — very similar provider family and identical symptom of repetitive tool calls
• Bug Model enters infinite loop due to hallucinated oldString in edit tool calls #21850: Model enters infinite loop due to hallucinated tool calls — related loop behavior, different root cause
• Some Model stuck in looping/rampancy, need detection and stop. #18912: Some model stuck in looping/rampancy — general request for loop detection/stopping

If your issue is distinct (e.g., specific to the harness-level loop amplification without hallucination), please note that in the issue so maintainers can track it separately.

[Astro-Han]

I checked the issues the bot linked. #18415 looks like the closest match.

Both reports describe the same broad pattern with Alibaba Coding Plan models: the session runs through many assistant turns, the model makes one small tool call at a time, and the agent repeats very similar actions instead of converging.

What I am seeing in #22142 does not look like the same problem as #21850, where the model keeps retrying a bad edit call with hallucinated input. It also does not look like the older class of bugs where the harness fails to stop after finish=stop. In the sessions I checked, the model keeps returning tool-calls, and the loop keeps re-entering normally.

Because of that, a small harness-side repetition guard may help both #22142 and #18415. Detecting a short run of near-duplicate tool-call turns, then forcing the model to change course or stop, seems like the most likely minimal fix.

My read is:

• #18415 is probably the closest existing issue
• #22142 adds more concrete session-level evidence for the same or a very similar failure mode
• one minimal loop-level guard could plausibly mitigate both