AGENTS.md

Guidance for coding agents working on LMSSEP Reasoning Servers.

Project Overview

The reasoning set contains independent MCP servers for recursive thinking, sequential thinking, tree of thoughts, counterfactual reasoning, verbalized sampling, and neuro-symbolic reasoning. Each server is intended to run independently or in small combinations rather than loading the whole suite.

Key paths:

• reasoning/README.md: overall reasoning entrypoint.
• recursive_thinking/: iterative answer improvement.
• sequential_thinking/: step-by-step analysis.
• tree_of_thoughts/: multiple path exploration.
• counterfactual_reasoning/: what-if analysis.
• verbalized_sampling/: diverse response sampling.
• neuro_symbolic_reasoner/: hybrid symbolic/neural reasoning.
• Each server has main.py, requirements.txt, and README.md.

Setup Commands

bashcd <project-root>/<reasoning-server>
pip install -r requirements.txt
python3 main.py

Use <reasoning-server>, <QUESTION>, <SESSION_ID>, and <MAX_STEPS> placeholders in examples.

Activation Rules

• Use recursive thinking for iterative refinement.
• Use sequential thinking for ordered reasoning traces.
• Use tree of thoughts when multiple alternatives must be explored.
• Use counterfactual reasoning for what-if or causality questions.
• Use verbalized sampling when diverse candidate responses matter.
• Use neuro-symbolic reasoning when symbolic constraints and LLM reasoning both matter.

Code Style

• Keep each server's tool schema independent.
• Avoid shared mutable state across reasoning modules unless intentionally documented.
• Store outputs under the server's own output directory.
• Keep prompt examples generic and free of private task data.

Testing

• Smoke test startup for the changed reasoning server.
• Test invalid step numbers, missing session IDs, excessive branching, and finalize-before-complete paths.
• Verify output can be resumed or inspected without loading unrelated servers.

Security

• Reasoning traces can expose private user context. Do not publish raw traces unless explicitly intended.
• Do not write sensitive questions to shared output locations without operator approval.