decision-topology（决策拓扑）

Decision Topology

Records how ideas branch and evolve during conversations, producing a browsable tree the user can review at any time. Like git log for thinking — the structure is always there when you want to inspect it.

Privacy note: This skill is installed and enabled by the user. All data stays on disk in the configured trees directory — nothing is sent externally. The user can view, delete, or relocate their trees at any time.

Security Properties

• Zero network access — no HTTP calls, no sockets, no DNS lookups. Works fully offline.
• Zero external dependencies — uses only Node.js built-in fs and path modules.
• No conversation content stored — the script enforces length limits on all persisted text fields (summary: 200 chars, reasoning: 300 chars, topic: 120 chars, concept: 50 chars). Text exceeding limits is truncated. This is a code-level guardrail, not just a policy.
• No process spawning — no child_process, no exec, no eval, no Function().
• Stdin-only input — all user-derived content is piped via stdin as JSON to prevent shell injection. See SECURITY.md for details.
• Path containment enforced — all file arguments are stripped to basename and resolved inside the canonical trees directory. Absolute paths and .. traversal are rejected at runtime.
• User-controlled storage — trees are local JSON files the user can inspect, move, or delete at any time.
• ID generation — uses Math.random() for 6-char hex node IDs. Cryptographic randomness is not needed — IDs only require tree-local uniqueness across 5-30 nodes.

Activation

Active by default (user can set always: false in metadata to require explicit invocation). When a conversation involves brainstorming, problem-solving, or exploring options, record the structure as a tree.

Skip pure Q&A ("what time is it"), greetings, and small talk.

Output Style

Do not insert status messages about tree operations into the conversation. The goal is a clean conversational experience — like how git commits happen without the developer seeing each one.

• Do not say "logging node," "branch created," "adding to tree" — this adds noise without value.
• Do not change your conversational behavior because of the skill. The user gets the same conversation they'd get without it.
• The user can ask to see the topology at any time ("show me what we explored", "what did we kill?").
• Think: git commits in the background. Low-noise, not hidden — the user knows it's installed and can inspect it whenever they want.

When to Record a Node

Create a node when:

• You propose a distinct idea, direction, solution, or option
• The user introduces a new angle or topic
• The user rejects, pushes back, or corrects — this is a branch kill, always record the reason
• The conversation pivots direction because of something said
• An insight combines elements from earlier dead branches (merge node)
• An analogy, metaphor, or reframe changes how the problem is understood
• A question redirects the exploration (pivot node)

Skip when:

• Same idea continues without meaningful evolution
• Minor refinement within the same direction
• Trivial "no" to something small that doesn't change direction
• Factual answers to factual questions

Depth calibration: A good tree has 5-30 nodes recording the shape of an exploration. Not 200 nodes transcribing every sentence. Only create nodes when the direction of thinking meaningfully shifts. Heuristic: would this rejection or pivot change what comes next? If yes, record it. If no, skip it.

Auto-Initialization

Do NOT create a tree when a conversation starts. Wait until the conversation actually branches — until there is a rejection, a pivot, or a second distinct direction. Only then initialize.

Most conversations won't need a tree. That's fine.

When initializing, auto-generate the filename from the date and a 2-4 word topic slug:

• 2026-02-24-business-model-exploration.json
• 2026-02-24-vacation-planning.json

Auto-Association

When a conversation starts and ideas begin branching, check if this connects to an existing tree before creating a new one.

How to check: Run the associate command with the core topic:

echo '{"query": "short description of current topic"}' | node {baseDir}/scripts/topology.js associate

This scans existing trees and returns the best match with a relevance score.

• Score >= 0.4 — continue that tree (load it, add nodes to it)
• Score 0.25-0.4 — ambiguous. Ask the user naturally: "This feels related to [topic] we explored on [date]. Continuing that thread, or fresh start?"
• Score < 0.25 — new tree

Never ask the user to pick a tree by ID. If you need to disambiguate, ask naturally in conversation.

Setup

Script: {baseDir}/scripts/topology.js

Storage: Trees are stored in {baseDir}/trees/ by default. Override with the TOPOLOGY_TREES_DIR environment variable if you want trees stored elsewhere (e.g. in a memory directory for semantic search indexing).

Path containment: All file arguments are resolved to basenames inside the trees directory. Absolute paths and .. traversal are rejected — the script cannot read or write files outside the configured trees directory.

Invocation: Always pipe JSON args via stdin to prevent shell injection from user-derived content:

echo '<JSON args>' | node {baseDir}/scripts/topology.js <command>

Core Operations

Initialize a tree

echo '{"topic": "short topic description"}' | node {baseDir}/scripts/topology.js init

Returns the file path and root node ID. Remember both for the session.

Add a node

echo '{"file": "<path>", "parent_id": "<id>", "type": "proposal", "summary": "one-line description", "reasoning": "why", "concepts": ["keyword1", "keyword2"]}' | node {baseDir}/scripts/topology.js add-node

Types: proposal, pivot, merge. The concepts array is optional — short keyword tags extracted from the node content, used for cross-tree linking.

Kill a branch

echo '{"file": "<path>", "node_id": "<id>", "reason": "why it was rejected"}' | node {baseDir}/scripts/topology.js kill-branch

Then add the new direction as a child (pivot node linked to what was killed).

Merge branches

echo '{"file": "<path>", "source_ids": ["<id1>", "<id2>"], "summary": "merged insight", "reasoning": "combines X from A with Y from B"}' | node {baseDir}/scripts/topology.js merge

Fork from any node

echo '{"file": "<path>", "node_id": "<id>", "summary": "re-exploring from this point", "reasoning": "reason for revisiting"}' | node {baseDir}/scripts/topology.js fork

Node Types

Status values: active (still exploring), dead (rejected, has killed_by), merged (combined into a merge node).

Viewing the Topology

The user does NOT need to learn slash commands. They ask naturally:

• "Show me what we explored"
• "What did we kill?"
• "What shape is this conversation?"
• "What paths did we reject and why?"
• "Go back to that idea about X"

You understand the intent and run the appropriate commands. Present results conversationally, not as raw script output.

/tree is an optional shortcut — works if the user types it, but don't teach it or require it.

Rendering

echo '{"file": "<path>"}' | node {baseDir}/scripts/topology.js render

After the tree, append a one-line summary: {N} branches explored, {M} killed, {K} active, depth {D}

List all trees

node {baseDir}/scripts/topology.js list

Statistics

echo '{"file": "<path>"}' | node {baseDir}/scripts/topology.js stats

Export as Mermaid

echo '{"file": "<path>"}' | node {baseDir}/scripts/topology.js export

Revisiting a dead branch

When the user asks about a killed path, find the node, present:

• What was proposed
• Why it was proposed
• Why it was killed
• What came after

Cross-tree analysis

node {baseDir}/scripts/topology.js analyze

Rebuilds the concept index, scans all trees, finds concepts appearing across multiple trees, reports which ideas keep surviving vs keep getting killed, identifies cross-root connections, and regenerates all companion .md files with updated cross-tree links and weights. Shows index health stats (total concepts, cross-tree count, orphans).

Query the concept index

echo '{"name": "trust"}' | node {baseDir}/scripts/topology.js concept
echo '{"list": true}' | node {baseDir}/scripts/topology.js concept
echo '{"orphans": true}' | node {baseDir}/scripts/topology.js concept

• name — reverse-lookup: shows every node across every tree that references a concept
• list — all concepts sorted by cross-tree spread, * marks concepts spanning multiple trees
• orphans — concepts that exist in only one tree (candidates for future linking)

Rebuild concept index

node {baseDir}/scripts/topology.js rebuild-index

Full rebuild of concepts.json from all tree files. Also regenerates all companion .md files with cross-tree links and updated weights. Use as a recovery tool or after manual edits to tree JSON files.

Concept Index

A reverse-index at {trees_dir}/concepts.json that maps every concept keyword to all nodes and trees that reference it.

• Automatic: Updated incrementally on every tree save (add-node, kill-branch, merge, fork, init). No manual intervention needed.
• Cross-tree links: Companion .md files include a ## Related trees section with [[wikilinks]] to other trees that share concepts. Useful for semantic search indexing.
• Weight field: Node weight is auto-set to the number of distinct trees sharing that node's concepts. weight: 1 = single-tree concept. weight: 2+ = concept spans multiple trees.
• Lazy discovery: Links form organically as nodes are added to any tree. A new node with concepts: ["trust"] will immediately link its tree to every other tree that also references "trust" — no need to wait for analyze.

Rules

1. Clean output. Do not insert tree-operation status messages into the conversation. The user can inspect trees whenever they want.
2. Judgment over completeness. Record the shape, not the transcript. 5-30 nodes per tree. Summaries only — never store verbatim conversation content.
3. Causal links. Show WHY the conversation evolved, not just WHAT was said. Link rejections to pivots.
4. Persist. Trees are JSON files that survive sessions. They can be searched if stored in an indexed directory.
5. Continue, don't duplicate. If a conversation continues a previous topic, load and extend that tree.
6. Graceful failures. Missing or corrupted tree — re-initialize. Missing node ID — say so clearly. Never crash.
7. Natural interface. The user asks in plain language. You translate to the right operation. Slash commands are optional shortcuts, not the primary interface.
8. Local only. All data stays on disk. No network calls, no external APIs, no telemetry. The user owns their data.