pmm:viz

PMM Viz — Interactive Memory Graph

Generates an interactive D3.js force-directed graph from all memory files and opens it in the browser. Includes a time slider that lets you scrub through git history to watch the graph evolve. Runs entirely as a subagent to keep the main context clean.

Scope: $ARGUMENTS (default: full if empty or not provided — shows everything)

Invocation

• /pmm:viz — full graph (all memory files)
• /pmm:viz graph — relationships from graph.md only
• /pmm:viz clusters — cluster members + similarity edges only
• /pmm:viz timeline — event nodes + decision nodes + temporal edges

Behaviour

When invoked, dispatch a general-purpose agent using the Readonly Agent Model from memory/config.md (default: haiku) with the prompt below. Replace <scope> with the subcommand (full, graph, clusters, or timeline). Replace <project-root> with the actual project root path.

Agent Prompt

Generate an interactive PMM memory graph with time-travel support. This is a READ + WRITE task. You may run git read-only commands (rev-parse, log, show). Do NOT run git write commands.

Scope: <scope>

Step 1 — Cache Check

Run: git rev-parse HEAD:memory 2>/dev/null

If the file <project-root>/pmm/viz-cache.html exists, read its first line. It should contain <!-- pmm-cache: HASH SCOPE -->. If both the hash and scope match, just run open <project-root>/pmm/viz-cache.html and return "Opened cached visualization."

Step 2 — Read All Memory Files (HEAD)

Read every .md file in <project-root>/memory/. The full list:

• graph.md, vectors.md, assets.md, timeline.md, decisions.md, processes.md
• lessons.md, memory.md, config.md, standinginstructions.md, preferences.md, voices.md
• progress.md, last.md, summaries.md, taxonomies.md

Skip files that don't exist or are empty.

Step 2.5 — Reconstruct History

Build temporal data so the time slider can show how the graph evolved over git commits.

1. Run: git log --format="%H %at %s" -- memory/ to get all commits that touched memory/.
2. Sort oldest-first. If there are more than 100 commits, sample: take every Nth commit to reduce to ~50, always keeping the first and last commits.
3. For each commit hash, extract node IDs and edge keys using a lightweight parse:
   • Run git show <hash>:memory/graph.md 2>/dev/null and extract all [[...]] names and edge lines ([[A]] → rel → [[B]])
   • Run git show <hash>:memory/vectors.md 2>/dev/null and extract [[...]] names from similarity/cluster lines
   • Run git show <hash>:memory/assets.md 2>/dev/null and extract first-column entries from People/Tools/Organisations tables
   • Run git show <hash>:memory/decisions.md 2>/dev/null and extract ## headings or list items
   • Run git show <hash>:memory/processes.md 2>/dev/null and extract ## headings
   • Run git show <hash>:memory/timeline.md 2>/dev/null and extract event entries
   • Run git show <hash>:memory/lessons.md 2>/dev/null and extract lesson entries
   • Run git show <hash>:memory/standinginstructions.md 2>/dev/null and extract instruction entries
   • Ignore errors (file may not exist at early commits)
   • You do NOT need full metadata from historical commits — just the set of node IDs and edge keys present at each commit
4. For each node: firstSeen = timestamp of earliest commit where the node ID appeared. lastSeen = timestamp of latest commit where the node ID appeared.
5. For each edge: same logic. Edge key = source|target|label for stable identity across commits.
6. Build a timeline array (oldest-first):

   [{"hash": "abc123", "timestamp": 1773650133, "message": "memory: initialise", "nodeCount": 5, "edgeCount": 3}]
   

   Where nodeCount and edgeCount are the counts at that specific commit.

Performance note: Process commits in batches. For each commit, run multiple git show commands in a single bash call where possible (e.g. separated by ; echo "---SEPARATOR---"; ). Only extract what's needed — node IDs and edge keys — not full content.

Step 3 — Parse Nodes and Edges (HEAD)

Parse each file using these rules:

graph.md:

• Lines matching [[A]] → relationship → [[B]] become directed edges
• ## headings become edge sections (store in edge section field)
• Both A and B become nodes

vectors.md:

• Lines matching [[A]] ↔ [[B]] | score: X.XX | basis: ... become similarity edges (undirected, type=similarity, weight=score)
• Lines matching Cluster: name → [[[A]], [[B]], ...] | theme: ... define clusters
• Cluster members become nodes

assets.md:

• Parse markdown tables under People, Tools, Organisations headings
• First column entries become nodes
• People → type=person, Tools → type=tool, Organisations → type=tool

timeline.md:

• Lines matching **[Date]** — [Event description] → Event nodes (type=event, label=description)

decisions.md:

• Each decision entry (## heading or list item with decision text) → Concept node (type=concept)

processes.md:

• Each ## heading → Process node (type=process)

lessons.md:

• Each lesson entry → Concept node (type=concept)

standinginstructions.md:

• Each instruction entry → Concept node (type=concept)

Step 4 — Assign Node Types (Priority Order)

If a node name appears in multiple files, assign type by priority:

1. In assets.md People table → person
2. In assets.md Tools/Organisations table → tool
3. Ends in .md → file
4. In decisions.md → concept
5. In processes.md heading → process
6. From timeline.md → event
7. Otherwise → concept

Step 5 — Deduplicate

Deduplicate nodes by canonical name (the text inside [[ ]], trimmed, case-preserved). Merge metadata from all sources.

Step 6 — Apply Scope Filter

• full → include everything
• graph → only edges from graph.md, only nodes referenced by those edges
• clusters → only cluster member nodes + similarity edges from vectors.md
• timeline → only event nodes (from timeline.md) + concept nodes (from decisions.md) + any edges connecting them

Step 7 — Build JSON

Build a JSON object matching this schema:

{
  "nodes": [{"id": "str", "type": "file|person|concept|tool|process|event", "label": "str", "metadata": {}, "clusters": [], "firstSeen": 1773650133, "lastSeen": 1773736182}],
  "edges": [{"source": "str", "target": "str", "label": "str", "section": "str", "weight": 0.0, "type": "relationship|similarity", "firstSeen": 1773650133, "lastSeen": 1773736182}],
  "clusters": [{"name": "str", "members": [], "theme": "str"}],
  "timeline": [{"hash": "str", "timestamp": 0, "message": "str", "nodeCount": 0, "edgeCount": 0}],
  "metadata": {"generated": "ISO date", "treeHash": "HASH", "scope": "<scope>", "nodeCount": 0, "edgeCount": 0}
}

Node id = canonical name. Edge weight defaults to 0.5 for relationship edges. Cluster members = array of node IDs. firstSeen/lastSeen are Unix timestamps from Step 2.5. If no history data is available (single commit or no history), omit firstSeen/lastSeen — the template handles missing values gracefully.

Step 8 — Assemble HTML

1. Read the template: <project-root>/pmm/pmm-viz-template.html
2. Read D3.js: <project-root>/pmm/d3.v7.min.js
3. In the template:
   • Replace /*D3_PLACEHOLDER*/ with the full D3.js source code
   • Replace /*PMM_DATA_PLACEHOLDER*/{"nodes":[],"edges":[],"clusters":[],"timeline":[],"metadata":{}}/*END_PLACEHOLDER*/ with the actual JSON data
   • Replace the first line <!-- pmm-cache: HASH SCOPE --> with <!-- pmm-cache: <actual-hash> <scope> -->
4. Write the assembled HTML to <project-root>/pmm/viz-cache.html

Step 9 — Open

Detect platform and open:

• macOS: open <project-root>/pmm/viz-cache.html
• Linux: xdg-open <project-root>/pmm/viz-cache.html
• WSL: wslview <project-root>/pmm/viz-cache.html or cmd.exe /c start <project-root>/pmm/viz-cache.html

Return a summary: "Generated PMM graph: X nodes, Y edges, Z commits in timeline. Opened in browser."

Notes

• The template and D3.js live in pmm/ at the project root (not in .claude/)
• pmm/viz-cache.html is gitignored — it's a generated artifact
• The agent reads ALL memory files regardless of config.md active list — the graph should be greedy
• If memory/ has no content (all files template-only), generate an empty graph with a message in the HTML
• The agent must not modify any memory files — read-only for memory/, write-only for pmm/viz-cache.html
• Git read permissions (rev-parse, log, show) are pre-approved in .claude/settings.json
• The time slider only appears when there are 2+ commits in the timeline — single-commit projects get the graph without the slider
• Clusters reflect HEAD state only — they don't change as you scrub the timeline