crune

Claude Code Rune — Traces linger, skills emerge

Decipher the traces etched in past sessions and resurrect them as reusable skills. crune is a static web dashboard and CLI tool that analyzes Claude Code session logs, providing session playback, analytics, a semantic knowledge graph, and skill synthesis.

Features

• Session Playback --- Turn-by-turn conversation replay with minimap navigation, tool call inspection, and subagent branch expansion
• Overview Dashboard --- Activity heatmap, project distribution, tool usage trends, duration distribution, model usage, and top edited files
• Semantic Knowledge Graph --- TF-IDF + Tool-IDF + structural features, Truncated SVD, agglomerative clustering, Louvain community detection, Brandes centrality (algorithm details)
• Tacit Knowledge --- Extracted workflow patterns, tool sequences, and pain points (long sessions, hot files)
• Session Summarization --- Automatic session summary and classification (no LLM required), enriched with /insights facets data when available
• Skill Synthesis --- Synthesize reusable skill definitions from the knowledge graph (algorithm details)

Overview Dashboard

Session Playback

Semantic Knowledge Graph (Rune)

CLI — Generate Skills via npx

Generate reusable Claude Code skill definitions directly from your session logs. No clone required.

npx @chigichan24/crune --dry-run                    # Preview skill candidates
npx @chigichan24/crune --skip-synthesis              # Generate heuristic skills (no LLM)
npx @chigichan24/crune --count 3 --model haiku       # LLM-synthesized skills (requires claude CLI)
npx @chigichan24/crune --output-dir ~/.claude/skills  # Install skills directly

Output follows the Claude Code skill format (<name>/SKILL.md), ready to use as /skill-name commands.

Web Dashboard

Quick Start

npm install

# Analyze Claude Code session logs (~/.claude/projects/)
npm run analyze-sessions

# Start dev server
npm run dev

Data Pipeline

npm run analyze-sessions reads JSONL session files from ~/.claude/projects/ and outputs structured JSON to public/data/sessions/.

~/.claude/projects/**/*.jsonl
  -> parse & build turns
  -> extract metadata, subagents, linked plans
  -> session summarization (centrality-based representative prompt, workType classification)
  -> TF-IDF + Tool-IDF + structural features -> Truncated SVD -> agglomerative clustering
  -> /insights facets integration (narrow cluster merging, goal-based labeling)
  -> Louvain community detection
  -> skill synthesis (reusability score top-N -> claude -p, enriched with facets insights)
  -> output:
       public/data/sessions/index.json      (session list)
       public/data/sessions/overview.json   (cross-session analytics + knowledge graph)
       public/data/sessions/detail/*.json   (individual session playback data)

Custom paths:

npm run analyze-sessions -- --sessions-dir /path/to/sessions --output-dir /path/to/output

/insights Integration

Before analysis, analyze-sessions automatically runs /insights to refresh facets data. This enriches the pipeline with:

• Better topic labels using LLM-generated underlying_goal instead of TF-IDF keywords
• Narrow cluster merging based on shared goal categories
• Quality-aware reusability scoring using session outcome and helpfulness
• Richer synthesis prompts with friction details and success rates
• Session summaries using brief_summary from facets

Skip with --skip-facets or customize the facets path with --facets-dir <path>.

Session Summarization

The session list displays auto-generated summaries. When /insights facets data is available, the LLM-generated brief_summary is used. Otherwise, summaries are processed locally without LLM.

• Representative prompt selection: Selects the most representative prompt from plan mode turns using Jaccard centrality with position weighting
• workType classification: Automatically classifies each session into one of four types:
  • investigation --- Exploratory, read-heavy sessions
  • implementation --- Edit/write-heavy coding sessions
  • debugging --- Bash-heavy sessions with edits
  • planning --- Plan mode sessions with few tool calls
• Keyword extraction: Extracts top keywords from session prompts
• Scope estimation: Infers the session scope from the common directory prefix of edited files

Skill Synthesis

Synthesizes reusable skill definitions from knowledge graph analysis. Follows the anthropics/skills format.

• Pre-synthesis: Automatically synthesizes the top 5 skill candidates via claude -p during analyze-sessions
• Instant display: Synthesized skills are immediately viewable and copyable in the UI
• On-demand re-synthesis: The re-synthesis button generates a full version enriched with graph context (connected topics, community, centrality)
• Local server: npm run skill-server starts the synthesis API server (localhost:3456)

Synthesis options:

# Specify model (e.g. haiku for speed)
npm run analyze-sessions -- --synthesis-model haiku

# Change number of candidates to synthesize (default: 5)
npm run analyze-sessions -- --synthesis-count 10

# Skip LLM synthesis (graph construction only)
npm run analyze-sessions -- --skip-synthesis

Scripts

Options for analyze-sessions

Tech Stack

• React 19 + TypeScript 5.9
• Vite 8
• Chart.js + react-chartjs-2
• react-force-graph-2d (d3-force)
• Plain CSS (no CSS-in-JS, no Tailwind)

Project Structure

src/
  components/
    overview/     # Dashboard cards, session list, charts
    playback/     # Session replay, tool call blocks, subagent branches
    knowledge/    # Force graph, node detail, tacit knowledge, skill synthesis
  hooks/          # Data fetching (useSessionIndex, useSessionDetail, useSessionOverview)
  types/          # TypeScript type definitions
scripts/
  cli.ts                     # npx CLI entry point
  session-parser.ts          # JSONL parsing, turn building, metadata extraction
  analyze-sessions.ts        # Dashboard JSON pipeline
  session-summarizer.ts      # Session summarization (local NLP)
  skill-synthesizer.ts       # Skill synthesis (claude -p)
  skill-server.ts            # Synthesis HTTP server
  knowledge-graph-builder.ts # Semantic embedding + graph construction
  knowledge-graph/
    facets-reader.ts         # /insights facets data reader + normalization
public/
  data/sessions/             # Generated JSON (gitignored)

Prerequisites

• Node.js >= 18
• Claude Code session logs at ~/.claude/projects/

License

Apache-2.0 license