ConductorScore Scanner

The open-source scanner behind conductorscore.com.

This is the code that runs on your machine, reads your local Claude Code, Codex, and Cursor transcripts, and uploads a numbers-only payload to the server for scoring. It's public, auditable, and dependency-free on purpose — the whole point of ConductorScore is that you can verify what crosses the wire before you trust the score.

~/.claude/projects/**/*.jsonl   (your transcripts)
            ↓
scanner (this repo)             (pure function)
            ↓
numeric payload                 (≈38 fields/session)
            ↓
conductorscore.com              (scoring)

What gets uploaded

Per Claude Code, Codex, or Cursor session in the last 30 days, the client emits ~50 fields (output_schema.py is the source of truth). Every field is one of:

• a number (counts, minute durations, token totals, line counts),
• a 16-char SHA-256 prefix (session id, project root — not reversible),
• a boolean — including a small number of provider-specific flags, e.g. token_data_missing (Cursor-only in practice: set when the device couldn't read a session's token usage, so the server excludes it from cost/efficiency denominators instead of treating it as zero-cost), or
• a categorical label — built-in tool names like Bash / Edit, MCP server/tool names in plaintext like mcp__github__create_issue, plugin command names in plaintext like my-plugin:deploy, model IDs like claude-opus-4-7 / gpt-5-codex / composer-2.5, slash-command names like /plan (read from Claude Code's structured <command-name> marker, never scraped from your prose), a closed provider enum (claude / codex / cursor), and plan-signal enums like EnterPlanMode. These are the names you configured — never their arguments, inputs, or outputs.

What is never uploaded

• No transcript content — not your messages, not Claude's responses.
• No code — not file contents, not diffs.
• No file paths — only a hash of the project root.
• No tool arguments or outputs — only tool names and counts. For the redundant-approvals signature, repeated Bash commands contribute their first command token (any leading NAME=value env assignment is stripped first, so a secret value can't ride along; and if the command is invoked by path — e.g. ./deploy.sh, /Users/you/clients/acme/run.sh — the token collapses to the literal path, so the path never crosses) and Edit/Write approvals contribute a one-way hash of the top-level directory — never the path itself.
• No CLAUDE.md content — only the line count.
• No prompts or planning text — only which structural signals fired.

This invariant is enforced in CI on every push and pull request. An integration test feeds synthetic transcripts — seeded with planted secret markers in every place content could leak (user prompts, file paths, tool inputs, slash-command arguments, assistant text, inline Bash environment variables, and Bash commands invoked by path) — through the real scanner, then asserts that none of those planted secrets appears anywhere in the upload payload. A companion sweep re-runs the check for every anti-pattern detector. (Like any sentinel test, it proves the absence of the secrets it plants rather than a mathematical impossibility — but the planting covers every field that crosses the wire, and the client is open source so you can audit the rest.)

For the field-by-field schema, see WIRE_FORMAT.md. A sample payload lives at wire_format_sample.json.

Verify the privacy invariant yourself

Don't take our word for it — the test that enforces the no-leak guarantee ships in this repo and runs in under a second:

git clone https://github.com/flatironconsulting/conductorscore
cd conductorscore
pip install -e ".[dev]"        # the only dev dependency is pytest; the client itself has zero runtime deps

# Run just the privacy-invariant test:
pytest tests/integration/test_extractor_integration.py -v

# Or run the whole suite (unit + integration):
pytest

The privacy test (tests/integration/test_extractor_integration.py) plants unique secret strings into a synthetic ~/.claude transcript tree, runs the actual extract() path the installed skill uses, and fails if any planted secret — or any raw prompt, path, or tool argument — survives into the serialized payload. To convince yourself it's real, edit the scanner to leak something on purpose and watch the test go red.

This is the same pytest invocation our CI runs (.github/workflows/ci.yml), so a green badge means the invariant held on the exact code you're reading.

Visualize your sessions locally

The same on-device data the scanner reads can be rendered as a timeline you can inspect yourself — no server round-trip, nothing uploaded. Ask the skill to "visualize this session", or run the viewer directly:

# the session you're in (after a `claude --resume` / `/resume`, or `codex resume`)
python3 ~/.claude/skills/conductorscore/scripts/session_viewer.py

# list your local sessions, then render one
python3 ~/.claude/skills/conductorscore/scripts/session_viewer.py --list
python3 ~/.claude/skills/conductorscore/scripts/session_viewer.py --pick 0

It writes a self-contained HTML page — the session timeline, the Interactive / Background leverage table your score is computed from, the longest unattended run, and subagent panels — and opens it in your browser.

This is the beginning of the README. Read the rest at the original. Read the full README ↗