RetryScope

Find out where your AI coding agent got stuck.

Local-first CLI that analyzes Claude Code session JSONL for evidence-linked repeated tool failures. No LLM API. No cloud upload. Offline HTML + JSON + Markdown reports.

Status: 0.1.0-rc.1 on GitHub (Releases). Demo data is synthetic. Detects potential repeated failures only — not token waste. npm Registry: not published.

Demo (synthetic)

git clone https://github.com/flower1201/retryscope.git
cd retryscope
npm install
npm run demo

Then open examples/out-demo/index.html in a browser (file:// works). You should see finding RF-001 (Bash) with a clickable execution timeline.

node ./dist/cli/index.js analyze ./examples/synthetic-session.jsonl --out ./examples/out-demo

Install from GitHub Pre-release (no npm Registry)

curl -fsSL -o retryscope-0.1.0-rc.1.tgz \
  https://github.com/flower1201/retryscope/releases/download/v0.1.0-rc.1/retryscope-0.1.0-rc.1.tgz
npm install ./retryscope-0.1.0-rc.1.tgz
npx retryscope --help

Release page: https://github.com/flower1201/retryscope/releases/tag/v0.1.0-rc.1
Still offline. npm Registry is not published.

Quick start (from source)

npm install
npm run build
node ./dist/cli/index.js --help
node ./dist/cli/index.js analyze /path/to/session.jsonl --out ./retryscope-report

Writes:

retryscope-report/
├── index.html
├── findings.json
└── summary.md

Supported log shape (Claude Code JSONL)

Best-effort subset (format may change across Claude Code versions):

• JSONL lines with optional type, timestamp, message
• message.content[] blocks:
  • tool_use: { id, name, input }
  • tool_result: { tool_use_id, content?, is_error? }
• Results link by tool_use_id ↔ id (never by array position alone)
• Missing result → unknown (never invented failure)

Privacy

• Only reads the JSONL path you pass (no home-directory scan)
• Reports omit prompts, commands, raw tool inputs/outputs, and internal inputGroupId
• Custom / MCP tool names render as Custom tool
• Warning messages use fixed codes — not raw log lines
• Never upload private agent sessions for demos or issues

Finding limitations

• Findings are potential repeated-failure patterns, not proof of wasted loops
• Tool success means the tool reported no error — not that the user task was solved
• Root causes are not inferred

Feedback

Looking for Claude Code users to try the synthetic demo and say whether the report helps. Use Discussions or Issues with synthetic reproductions only. See SECURITY.md.

License

MIT