Tool loop guard stopped repeated failing calls to "task" after 3 attempts

[josemiguelo]

Hi!

First of all, Thanks for creating this plugin! 🙏🏻

This is related to #45 which was closed and can't be reopened

I have latest 2.3.17 version (installed through npm), and the issue is still alive.
This is how things went down in my case:

Prompt
analyze this repo, and tell me potential issues, improvements/enhancements

Output:

Thinking: The user wants me to analyze the stellasaur repository and identify potential issues, improvements, and enhancements. Let me explore the codebase thoroughly using multiple agents in parallel.
│ Task Analyze build scripts
Error: The task tool was called with invalid arguments: [
  {
    "expected": "string",
    "code": "invalid_type",
    "path": [
      "subagent_type"
    ],
    "message": "Invalid input: expected string, received undefined"
  }
].
Please rewrite the input so it satisfies the expected schema.
ctrl+x down view subagents
Thinking: Let me launch multiple exploration agents in parallel to analyze different parts of the repo.
Tool loop guard stopped repeated failing calls to "task" after 3 attempts (limit 2). Adjust tool arguments and retry.
▣  Build · opus-4.6-thinking · 18.7s

Next prompt:
analyze this repo, and tell me potential issues, improvements/enhancements. don't use multiple agents in parallel
Next output:

Thinking: The user wants me to analyze the repository for potential issues, improvements, and enhancements. Let me start by exploring the repo structure and reading key files.
⚙ shell [command=find /Users/jochoa/Repos/gh/josemiguelo/stellasaur -type f -not -path '*/.git/*' -not -path '*/node_modules/*' | head -100, timeout=30000, cwd=]
Thinking: The user wants me to analyze the repo for potential issues, improvements, and enhancements. Let me read through the key files systematically. I'll start by reading the core build files and configuration.
~ Reading file...
Thinking: The user wants me to analyze the repository for potential issues, improvements, and enhancements. Let me read through the key files systematically. I already have the Containerfile. Let me read the build scripts, workflows, services, and ujust files.
~ Reading file...
Thinking: The user wants me to analyze the repo for potential issues, improvements, and enhancements. I've already seen the Containerfile and the orchestrator script. Let me now read all the build scripts, the Justfile, services, ujust files, and other key files to do a thorough analysis.
Tool loop guard stopped repeated failing calls to "task" after 3 attempts (limit 2). Adjust tool arguments and retry.
▣  Build · opus-4.6-thinking · 42.7s

the error pops up when cursor determines it should spawn multiple subagents after prompt (first prompt) or down the road after some analysis (second prompt)

[Nomadcxx]

Hey mate, this fix address this issue, please let me know how it goes for you.

The issue was that the tool loop guard was way too aggressive: first sign of a repeated failure and it'd kill the entire conversation. Not great when the model just needs a nudge to try something different.

Now it does a graduated response, first trigger sends a hint back to the model saying "hey this tool is blocked, try a different approach." Only if the model ignores that and keeps hammering does it actually kill the stream.

Give it a shot and let me know how it goes. Leaving this open to track feedback.

[josemiguelo]

Hey!

Thanks for your work on this 🙏🏼 . Just tried the prompt one more time but still no luck:

prompt:
analyze this repo, and tell me potential issues, improvements/enhancements.

output:

Thinking: The user wants me to analyze the stellasaur repository and identify potential issues, improvements, and enhancements. Let me explore the codebase thoroughly using multiple parallel agents to cover different areas.
│ Task Analyze build scripts
Error: The task tool was called with invalid arguments: [
  {
    "expected": "string",
    "code": "invalid_type",
    "path": [
      "subagent_type"
    ],
    "message": "Invalid input: expected string, received undefined"
  }
].
Please rewrite the input so it satisfies the expected schema.
ctrl+x down view subagents
Thinking: Let me launch multiple exploration agents in parallel to analyze different parts of the repo.
Tool "task" has been temporarily blocked after 3 repeated validation failures. Do not retry this tool. Use a different approach to complete the task.Tool "task" has been temporarily blocked after 3 repeated validation failures. Do not retry this tool. Use a different approach to complete the task.Tool loop guard stopped repeated schema-invalid calls to "task" after 4 attempts (limit 2). Adjust tool arguments and retry.
▣  Build · opus-4.6-thinking · 27.3s

Although it says Do not retry this tool. Use a different approach to complete the task., it actually do not seem to take a different route on this. it tried 3 times and it stopped altogether.

[josemiguelo]

A small heads up: corporate politics make me use opencode like this: CURSOR_ACP_FORCE=false opencode , Do you think it has something to do with this?

[Nomadcxx]

Hey mate, thanks for the detailed report. Graduated soft hint before hard kill approach didn't work in your case looks like because how model batches task calls (my guess) it just fires 3-5 of them in parallel before it even sees the hint. By the time it gets any feedback, the guard's already tripped. Here is my best guess fix for this issue:

1. Raised threshold - task now gets 10 retries instead of 2 before the guard fires. Should absorb a full parallel batch without triggering.
2. Schema injection - the schema exposed subagent_type as a plain string with no valid values listed, so the model was guessing (badly). It now gets told exactly which values are valid based on your opencode.json agent config.

Hopefully this works, less certain on the schema injection landing in proxy mode. If you're still seeing subagent_type: undefined validation errors after upgrading, let us know what your agent config looks like and we can dig in further.