Mindmap Mcts

MindMap-MCTS

Use this skill to keep complex work on a visible, evidence-backed reasoning tree. The bundled CLI owns the JSON truth source, UCB selection, backpropagation, markdown rendering, static HTML rendering, and interactive Markmap HTML rendering; the agent owns expansion, real probes, and judgment.

Command Prefix

Choose the command form that matches the current shell. First set the skill directory for the host that loaded this skill.

Codex on Unix shells:

export MINDMAP_MCTS_SKILL_DIR="${CODEX_HOME:-$HOME/.codex}/skills/mindmap-mcts"

Claude Code on Unix shells:

export MINDMAP_MCTS_SKILL_DIR="${CLAUDE_HOME:-$HOME/.claude}/skills/mindmap-mcts"

Default install locations are $HOME/.codex/skills/mindmap-mcts for Codex and $HOME/.claude/skills/mindmap-mcts for Claude Code.

Linux, macOS, WSL, or Git Bash:

"$MINDMAP_MCTS_SKILL_DIR/scripts/mindmap" --help

Cross-platform Python launcher:

python "$MINDMAP_MCTS_SKILL_DIR/scripts/mindmap.py" --help

Windows PowerShell:

# Claude Code:
$env:MINDMAP_MCTS_SKILL_DIR = "$env:USERPROFILE\.claude\skills\mindmap-mcts"

# Codex:
# $env:MINDMAP_MCTS_SKILL_DIR = "$env:USERPROFILE\.codex\skills\mindmap-mcts"

& "$env:MINDMAP_MCTS_SKILL_DIR\scripts\mindmap.ps1" --help

If PowerShell execution policy blocks .ps1 scripts, run the Python module directly after setting PYTHONPATH:

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

For repeated commands in one shell session, define:

alias mindmap='"$MINDMAP_MCTS_SKILL_DIR/scripts/mindmap"'

If aliases are not preserved by the current shell call, use the explicit launcher form for the current shell.

On Windows, set UTF-8 output before rendering or showing trees that contain Chinese text:

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

Never hand-edit rendered markdown as the truth source. The .tree.json file is the truth source; .tree.md is only a view.

Startup

1. Decide whether this skill applies. Skip it for one-step commands, obvious one-file edits, or direct factual lookups.
2. If no tree exists, create one:

"$MINDMAP_MCTS_SKILL_DIR/scripts/mindmap" init --title "<task title>" --out <task-name>.tree.json

The root node created by init is n1; use --parent n1 for the first child. Do not use root unless a specific tree file already contains a node with that id.

3. If a tree exists, inspect it:

"$MINDMAP_MCTS_SKILL_DIR/scripts/mindmap" show <task-name>.tree.json

4. Render the readable view:

"$MINDMAP_MCTS_SKILL_DIR/scripts/mindmap" render <task-name>.tree.json --out <task-name>.tree.md
"$MINDMAP_MCTS_SKILL_DIR/scripts/mindmap" render-markmap <task-name>.tree.json --out <task-name>.markmap.html

Use render-markmap when the user wants a browser-based visual tree. It produces interactive HTML through Markmap's browser autoloader. The map starts with an Exploration status branch that shows 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 only when a static offline fallback is preferable.

Per-Round Loop

1. Ask the CLI for the next recommended action:

"$MINDMAP_MCTS_SKILL_DIR/scripts/mindmap" next <task-name>.tree.json

Use this to orient before making new changes to the tree.

2. Select the next frontier directly when you need only the node id:

"$MINDMAP_MCTS_SKILL_DIR/scripts/mindmap" select <task-name>.tree.json

3. Expand the selected node with 2 or 3 child nodes. Children must be mutually exclusive, concrete, and verifiable.

"$MINDMAP_MCTS_SKILL_DIR/scripts/mindmap" add <task-name>.tree.json --parent <node-id> --type hypothesis --content "<specific hypothesis>"

4. Evaluate each new child using the cheapest real probe available: read code, grep a symbol, run a focused test, inspect a log, or check a config.

5. Record value and evidence:

"$MINDMAP_MCTS_SKILL_DIR/scripts/mindmap" eval <task-name>.tree.json --id <node-id> --value <0-to-1> --evidence "<short evidence>" --probe-type <test|grep|log|paper|code-read|user-input> --source "<source pointer>" --confidence <low|medium|high>

Use --probe-type, --source, and --confidence when a score is backed by a concrete probe. Omit them only when there is no useful structured metadata.

6. Prune disproven branches instead of deleting them:

"$MINDMAP_MCTS_SKILL_DIR/scripts/mindmap" prune <task-name>.tree.json --id <node-id> --evidence "<why this branch is ruled out>" --probe-type <test|grep|log|paper|code-read|user-input> --source "<source pointer>" --confidence <low|medium|high>

7. Backpropagate from evaluated nodes:

"$MINDMAP_MCTS_SKILL_DIR/scripts/mindmap" backprop <task-name>.tree.json --from <node-id>

8. Render and show the updated state:

"$MINDMAP_MCTS_SKILL_DIR/scripts/mindmap" render <task-name>.tree.json --out <task-name>.tree.md
"$MINDMAP_MCTS_SKILL_DIR/scripts/mindmap" render-html <task-name>.tree.json --out <task-name>.tree.html
"$MINDMAP_MCTS_SKILL_DIR/scripts/mindmap" render-markmap <task-name>.tree.json --out <task-name>.markmap.html
"$MINDMAP_MCTS_SKILL_DIR/scripts/mindmap" show <task-name>.tree.json
"$MINDMAP_MCTS_SKILL_DIR/scripts/mindmap" path <task-name>.tree.json
"$MINDMAP_MCTS_SKILL_DIR/scripts/mindmap" doctor <task-name>.tree.json