Evidence-guided reasoning trees for Codex and Claude Code agents.

MindMap-MCTS is an installable Codex skill and Claude Code skill pattern for complex debugging, architecture decisions, and research synthesis. It turns agent reasoning into an auditable reasoning tree: propose branches, run real probes, score evidence, backpropagate values, and choose the next step with lightweight Monte Carlo Tree Search (MCTS) and UCB.

In short: this gives AI agents a visible tree search loop instead of a hidden linear trial-and-error thread.

Codex skill, Claude Code skill, AI agents, agentic AI, agent reasoning, reasoning tree, evidence-guided reasoning, tree search, Monte Carlo Tree Search, MCTS, UCB, debugging workflow, planning workflow, architecture tradeoffs, research synthesis.

中文摘要：MindMap-MCTS 给 Codex、Claude Code 和其他智能体安装一个可见的推理树，让复杂问题可以用“分支探索、证据评分、MCTS/UCB 选择、回传更新”的方式推进，而不是线性瞎试。

• Creates a JSON reasoning tree as the truth source.
• Renders readable Markdown, static HTML, and interactive Markmap HTML mindmap views.
• Selects the next branch with lightweight MCTS/UCB.
• Records V value, N visits, state, probe metadata, and evidence per node.
• Preserves pruned branches so dead ends are not retried.
• Helps agents stop when a high-evidence path converges or when a user decision is needed.

MindMap-MCTS turns Codex problem solving into a compact search architecture:

The result is an Alpha-inspired skill implementation: the language model supplies candidate moves and real-world probes; the deterministic CLI preserves the tree, scores branches, and renders an auditable reasoning map.

This project is not affiliated with Google, DeepMind, AlphaGo, AlphaZero, AlphaFold, or any official Alpha-series project.

mindmap-mcts-skill/
  mindmap-mcts/              # Installable Codex skill folder
    SKILL.md
    agents/openai.yaml
    scripts/mindmap_mcts/    # Bundled CLI tree engine
  examples/                  # Example tree state and rendered view
  docs/                      # Design notes and implementation plan
  assets/                    # GitHub README illustrations
  tests/                     # CLI and engine tests

Clone this repository and run the installer for your agent host.

Codex:

git clone git@github.com:cheshireyang/mindmap-mcts-skill.git
cd mindmap-mcts-skill
./install.sh

Restart Codex so the new skill metadata is loaded.

Claude Code:

git clone git@github.com:cheshireyang/mindmap-mcts-skill.git
cd mindmap-mcts-skill
./install-claude.sh

This installs the same skill folder to ${CLAUDE_HOME:-$HOME/.claude}/skills/mindmap-mcts; the default Claude Code path is ~/.claude/skills/mindmap-mcts. Restart Claude Code so the new skill metadata is loaded.

Manual install is also just a directory copy:

mkdir -p "${CODEX_HOME:-$HOME/.codex}/skills"
cp -R mindmap-mcts "${CODEX_HOME:-$HOME/.codex}/skills/"

mkdir -p "${CLAUDE_HOME:-$HOME/.claude}/skills"
cp -R mindmap-mcts "${CLAUDE_HOME:-$HOME/.claude}/skills/"

Ask Codex to use the skill explicitly:

Use $mindmap-mcts to explore this debugging task: login sometimes times out under load.

or:

用 $mindmap-mcts 分析这个复杂问题：Transformer 当前有哪些缺陷？

Install with ./install-claude.sh, restart Claude Code, then ask for the same skill by name:

Use $mindmap-mcts to explore this architecture tradeoff.

The skill includes a Python CLI under mindmap-mcts/scripts.

Linux, macOS, WSL, or Git Bash:

mindmap-mcts/scripts/mindmap --help

Cross-platform Python launcher:

python mindmap-mcts/scripts/mindmap.py --help

Windows PowerShell can use the native launcher, which sets PYTHONPATH and UTF-8 output before running the Python module:

.\mindmap-mcts\scripts\mindmap.ps1 --help

If PowerShell execution policy blocks .ps1 files, run the Python module directly:

$env:PYTHONPATH = "$PWD\mindmap-mcts\scripts;$env:PYTHONPATH"
$env:PYTHONIOENCODING = "utf-8"
[Console]::OutputEncoding = [System.Text.Encoding]::UTF8
python -m mindmap_mcts.cli --help

Windows cmd.exe can use:

mindmap-mcts\scripts\mindmap.cmd --help

Generate the interactive HTML mind map from PowerShell:

.\mindmap-mcts\scripts\mindmap.ps1 render-markmap task.tree.json --out task.markmap.html

Create and inspect a tree:

mindmap-mcts/scripts/mindmap init \
  --title "Fix intermittent login timeout" \
  --out task.tree.json

# The root node created by init is n1.
mindmap-mcts/scripts/mindmap add task.tree.json \
  --parent n1 \
  --type hypothesis \
  --content "DB connection pool is exhausted"

mindmap-mcts/scripts/mindmap eval task.tree.json \
  --id n2 \
  --value 0.9 \
  --evidence "Logs contain pool timeout during failed login" \
  --probe-type log \
  --source logs/auth.log \
  --confidence high

mindmap-mcts/scripts/mindmap backprop task.tree.json --from n2
mindmap-mcts/scripts/mindmap render task.tree.json --out task.tree.md
mindmap-mcts/scripts/mindmap render-html task.tree.json --out task.tree.html
mindmap-mcts/scripts/mindmap render-markmap task.tree.json --out task.markmap.html
mindmap-mcts/scripts/mindmap show task.tree.json
mindmap-mcts/scripts/mindmap path task.tree.json
mindmap-mcts/scripts/mindmap next task.tree.json
mindmap-mcts/scripts/mindmap doctor task.tree.json

Use render-markmap for an interactive browser mind map. It writes a self-contained HTML shell that loads Markmap in the browser through the official CDN autoloader, so no local Node/npm setup is required. The generated map starts with an Exploration status branch showing best path, selected frontier, state counts, open frontier nodes, verified nodes, and pruned nodes. In the reasoning tree, completed exploration uses a green ✓, partial exploration uses a yellow ◐, and unopened frontier nodes use a gray ○; no red cross is used. Node titles are bold black, and status stays immediately after the title, for example **n11 事实性与幻觉** (V=0.90 N=1 verified). Use render-html when you want the simpler static fallback.

Available commands:

init, add, eval, prune, select, backprop, render, render-html, render-markmap, show, path, next, doctor

Structured evidence fields are optional. Use them when a score is backed by a concrete probe:

--probe-type test|grep|log|paper|code-read|user-input
--source path/to/file.py:42
--confidence low|medium|high

• The tree root created by init is n1, not root.
• Prefer mindmap.ps1, mindmap.cmd, python mindmap-mcts/scripts/mindmap.py, or python -m mindmap_mcts.cli in PowerShell. The Bash mindmap wrapper is for Unix shells, WSL, and Git Bash.
• If Chinese output looks garbled in PowerShell, set UTF-8 before running commands:

$env:PYTHONIOENCODING = "utf-8"
[Console]::OutputEncoding = [System.Text.Encoding]::UTF8

See examples/login-timeout.tree.md:

Best path: n1 -> n2
Best value: 0.90

Use this skill when a task has:

• multiple plausible hypotheses or designs
• systematic debugging needs
• repeated trial-and-error risk
• option tradeoffs that should remain visible
• evidence-backed exploration rather than pure speculation

Skip it for one-step commands, obvious edits, or direct fact lookups.

Run tests from the repository root:

PYTHONPATH=mindmap-mcts/scripts pytest tests -q

Validate the skill folder with Codex's skill validator:

python3 ~/.codex/skills/.system/skill-creator/scripts/quick_validate.py mindmap-mcts

Skill evaluation prompts live in evals/evals.json. They cover debugging, architecture tradeoffs, and research synthesis, and are intended to compare agent behavior with and without $mindmap-mcts.

MIT