Skip to content

AGENTS.md

Latest commit

 

History

History
48 lines (31 loc) · 1.58 KB

AGENTS.md

File metadata and controls

48 lines (31 loc) · 1.58 KB
Raw

ReasonKit-think — Agent Guide

ReasonKit-think (reasonkit-think-mcp on crates.io) is a single-crate Rust stdio MCP server for auditable, fail-closed reasoning in AI coding agents.

Philosophy: Designed, Not Dreamed. Turn prompts into protocols.

Commands

Task Command
Build release binary just build
Format, clippy, tests just check
MCP wiring + smoke test just mcp-refresh

Python smoke/compat scripts: uv / uvx. JS/TS when needed: bun / bunx. Avoid make, pip, pipx, npm, and npx by default.

Invariants

  • stderr only for logs; never write diagnostics to stdout (stdio MCP transport).
  • Fail closed on model-backed or retrieval-backed claims until real provider bindings exist.
  • Do not claim verification success without evidence tiers and matrix status.
  • Do not claim MCP health without just mcp-refresh or scripts/smoke_test.py.

Thinking modes

Default user-facing mode: Auto (run_thinking_mode). Map requests to Quick / Explore / Map / Sketch / Test when appropriate.

Full mode contract: README table and docs/resources/thinking-modes.md (resource reasoning://thinking-modes).

Scoped rules

Keep this file lean. Put detailed, scoped guidance in:

  • Cursor: .cursor/rules/*.mdc
  • Claude / other agents: .claude/rules/*.md and .agents/rules/*.md (mirrored)

Docs

  • Architecture: docs/ARCHITECTURE.md
  • Tools: docs/tools/
  • MCP resources: docs/resources/
  • Client setup: docs/clients/

Workflow

Inspect → Plan → Edit → Verify (fmt/clippy/test/smoke) → Summarize