fix: change executor tool_choice from required to auto so reasoning respects no-action decisions

[saakshisinghal14]

Fixes #253

Summary

When CoT reasoning concludes that no action should be taken (e.g., "I choose NOT to call adopt_opinion"), the executor phase still forces a tool call because tool_choice="required". This causes the LLM to pick an arbitrary tool — often an unrelated one like move_one_step — contradicting its own reasoning.

The same issue exists in the base Reasoning.execute_tool_call() and aexecute_tool_call() methods used by ReAct and ReWOO.

Changes

• Changed tool_choice="required" to tool_choice="auto" in all four executor call sites (CoT sync/async, base Reasoning sync/async)
• Updated the executor system prompt to explicitly instruct the LLM that it may skip tool calls when the plan concludes no action is needed
• Updated existing test assertion to match the new tool_choice value

Files changed

• mesa_llm/reasoning/cot.py — plan() and aplan() executor phases
• mesa_llm/reasoning/reasoning.py — execute_tool_call() and aexecute_tool_call()
• tests/test_reasoning/test_reasoning.py — updated assertion for tool_choice

Tests

All 21 reasoning tests pass:

[coderabbitai]

Important

Review skipped

Auto reviews are disabled on this repository. Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: Organization UI

Review profile: CHILL

Plan: Pro

Run ID: 7cd1a7ed-ff1d-4bb6-8916-187362965217

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands and usage tips.}

[ZhehaoZhao423]

Hi @saakshisinghal14, fantastic fix! I actually ran into this exact framework limitation last week while building my Information Cascade benchmark (PR #251). My financial agents were hallucinating move_one_step actions purely because the CoT executor was coercing them to pick a tool, even when their reasoning explicitly concluded to "HOLD" their positions.

Following the review guidelines, this is a massive improvement for the big picture of agent autonomy. Shifting tool_choice from required to auto correctly returns the decision boundary to the LLM's cognitive phase.

I also took a quick look at the downstream impact. Since tool_choice="auto" means the LLM might return responses without tool calls entirely, I checked ToolManager.call_tools / acall_tools. Fortunately, the existing getattr(llm_response, "tool_calls", []) with the if not tool_calls: check perfectly catches None or missing payloads and safely returns []. This means your PR integrates flawlessly without triggering any downstream TypeErrors.

Incredibly elegant and necessary fix for production-ready deterministic simulations. Great work!

[saakshisinghal14]

Thanks @ZhehaoZhao423 ! Good to know the downstream call_tools path handles the None case cleanly — I checked the same code path but appreciate the independent confirmation. Glad this helps your Information Cascade work too.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 91.88%. Comparing base (b6ece59) to head (038d2d1).
⚠️ Report is 7 commits behind head on main.

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #286      +/-   ##
==========================================
+ Coverage   91.55%   91.88%   +0.33%     
==========================================
  Files          19       19              
  Lines        1645     1627      -18     
==========================================
- Hits         1506     1495      -11     
+ Misses        139      132       -7     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[wang-boyu]

Thanks for the contribution and the discussion here. I pushed some follow-up commits / fixes:

• fixed the ReWOO no-op case when llm_plan.tool_calls is None
• added a tool_calls parameter so execution-side tool_choice is configurable (default to "auto") while planning still uses tool_choice="none"
• refactored CoT to use the shared execute_tool_call() path to align with ReAct and ReWOO
• aligned memory handling across CoT, ReAct, and ReWOO to have consistent plan_execution logging
• updated some docstrings and test cases
• merged main branch to resolve conflicts

Thanks again for the contribution. I'm merging this now.

[saakshisinghal14]

Thanks @wang-boyu for the detailed follow-up changes and improvements!

I really like the alignment across CoT/ReAct/ReWOO and the configurable tool_calls design — makes the execution flow much cleaner and more consistent.

Glad I could contribute to this fix. Looking forward to building on this further!