System prompt race condition in reasoning

[DineshThumma9]

Describe the bug

Reasoning code mutates a shared system prompt on the agent LLM during planning and execution. In parallel runs, one in-flight reasoning call can overwrite another call’s prompt, causing cross-agent prompt bleed and non-deterministic behavior.

Relevant Files : cot.py ,react.py,reasoning.py

Expected behavior

System prompt should be scoped per generate call, not mutated as shared instance state during concurrent reasoning.

To Reproduce

1. Enable parallel stepping with multiple LLM agents.
2. Run simulation where multiple agents call planning in same step.
3. Observe occasional inconsistent tool choice or plan output that does not match the intended reasoning prompt context.
4. Inspect reasoning files and see shared prompt mutation before generate calls.

[raghav-567]

Progress update:

• Opened PR: fix(reasoning): preserve agent system prompt during planning and execution #283
• Scope implemented (issue-aligned, minimal):
  1. Added per-call system_prompt support in ModuleLLM.generate/agenerate.
  2. Updated cot.py, react.py, and reasoning.py to pass scoped prompts per call instead of mutating shared llm.system_prompt.
  3. Added regression tests for prompt scoping + no shared-state mutation.

Validation:

• .venv/bin/python -m pytest tests/test_module_llm.py tests/test_reasoning/test_reasoning.py tests/test_reasoning/test_cot.py tests/test_reasoning/test_react.py -q
• Result: 35 passed
• .venv/bin/ruff check ... on changed files
• Result: All checks passed!

Related working notes/artifacts:

• https://github.com/raghav-567/mesa-gsoc-2026-raghav/blob/main/proposal_optimized_approach.md
• https://github.com/raghav-567/mesa-gsoc-2026-raghav/blob/main/notes/mesa_llm_gsoc_2026_alignment.md
• https://github.com/raghav-567/mesa-gsoc-2026-raghav/blob/main/models/mixed_agent_latency/benchmark_results.csv