feat(cli): add `stats` subcommand — per-agent React Doctor leaderboard by aidenybai · Pull Request #932 · millionco/react-doctor

aidenybai · 2026-06-22T01:06:38Z

Summary

Adds react-doctor stats: a per-model/per-tool code-quality leaderboard built from your local AI agent history. It answers one question — which agent writes the cleanest React code in my repo?

It reads local agent history (Claude Code + Codex transcripts, the Cursor composer database), reconstructs the file content each model actually wrote, lints it with the existing engine, and ranks models and providers by a confidence-weighted React Doctor score.

Faithful reconstruction per provider: Claude post-edit snapshots, Cursor afterContentId full-content blobs from state.vscdb (real model attribution, not "Auto"), Codex apply_patch replay. Only actual React files (JSX/TSX, use client/use server, or a React-ecosystem import) are scored, so backend/util/config files don't dilute the result.
Confidence-weighted ranking: each group's raw score regresses toward the global mean by its evidence (files are the dominant signal, lightly discounted by sessions since many files from one chat are one correlated sample), bounded by a floor. A tiny clean sample can't top the board; the raw and weighted scores both ship in --json.
Plain-language terminal UI: ranked model table + by-tool table with color-coded tools (cursor gray, claude orange, codex cyan), a best/worst callout, and honest "skipped" notes. Adds an orange formatter to the shared highlighter (honors --no-color).
Flags: --global (all repos), --since, --limit, --provider, --json. Default scope is the current repo.

Coverage is honest about limits: Codex shell edits aren't reconstructable (surfaced as skipped), the Cursor DB needs node:sqlite (Node 22.13+) and covers GUI composer sessions (not cursor-agent CLI), and the score requires network access.

Test plan

pnpm --filter react-doctor test (32 new stats tests: adapters, reconstruct, apply-patch, aggregate/weighting, is-react-source, render)
pnpm typecheck / pnpm lint / pnpm format:check
react-doctor stats in a repo with local Cursor/Claude/Codex history renders a sane leaderboard
react-doctor stats --json emits { schemaVersion, models, providers, best, worst, … } with both score and weightedScore
react-doctor stats --global ranks across repos; --provider cursor narrows the source

Note

Medium Risk
Large new CLI surface that reads local agent/SQLite data and spawns many lint subprocesses; scoring still depends on the network score API. Logic is heavily tested and read-only on user data, but reconstruction and ranking edge cases could mis-rank in real histories.

Overview
Adds react-doctor stats, a leaderboard that ranks AI agents and models by how clean the React code they wrote is, using local chat history and the existing lint/score pipeline.

The command discovers sessions from Claude Code and Codex JSONL transcripts plus Cursor’s state.vscdb composer DB (replacing model-less transcript JSONL for Cursor). It replays edits into file snapshots (Claude tool results, Cursor afterContentId, Codex apply_patch), keeps only React sources, materializes temp trees with project config, and runs oxlint per session (bounded concurrency). Failed or skipped scans are dropped so they cannot look like “perfect” code.

Ranking groups by model and provider, calls the score API per group, and sorts on a confidence-weighted score (Bayesian shrink toward the global mean; files dominate, sessions lightly discount correlation). Terminal output shows top models, per-tool tables, score bars, and best/worst callouts; --json ships the full report. Flags: --global, --since, --limit, --provider.

Also adds highlighter.orange for Claude branding in stats output, stats.run telemetry, CLI flag stripping for stats, a changeset, and deslop-js in the publish workflow. Broad unit test coverage for adapters, reconstruction, aggregation, and rendering.

^{Reviewed by Cursor Bugbot for commit 509f229. Bugbot is set up for automated code reviews on this repo. Configure here.}

…d from agent history Adds `react-doctor stats`, which reads local AI agent history (Claude Code + Codex transcripts, the Cursor composer database), reconstructs the React code each model actually wrote, lints it with the existing engine, and ranks models and providers by a confidence-weighted React Doctor score. - Reconstructs faithful post-edit file content per provider (Claude snapshots, Cursor `afterContentId` blobs, Codex `apply_patch`), filtered to real React. - Confidence-weighted ranking: each group's raw score regresses toward the global mean by its evidence (files dominant, lightly discounted by sessions), so a tiny clean sample can't top the board. - Plain-language terminal leaderboard with color-coded tools (adds an `orange` to the shared highlighter for Claude); `--json` for the machine-readable report.

pkg-pr-new · 2026-06-22T01:07:17Z

Open in StackBlitz

npm i https://pkg.pr.new/deslop-js@932

npm i https://pkg.pr.new/eslint-plugin-react-doctor@932

npm i https://pkg.pr.new/oxlint-plugin-react-doctor@932

npm i https://pkg.pr.new/react-doctor@932

commit: 509f229

Discovery loaded each candidate session from the Cursor SQLite DB synchronously, blocking the event loop so the ora spinner appeared frozen for a few seconds. Yield to the event loop periodically and report live "(N found)" progress during the history walk.

Cap the terminal table to the top 5 with a "+ N more" pointer to --json; the full ranking still ships in the JSON report and the best/worst callout.

Consolidate the asString/asRecord/asArray/parseJson coercers (copied across the Claude/Codex/Cursor adapters) into a shared coerce.ts, extract the "most common model" tally into most-common-key.ts, and reuse statMtimeMs in findJsonlFiles. Behavior unchanged.

The static `node:sqlite` import crashed the whole adapter test file on Node 20 (where the module doesn't exist), failing the 20.19 CI matrix job. Load it via a guarded require and skip the Cursor suite when unavailable, mirroring cursor-db.ts's runtime degradation.

- closeCursorDb now closes the underlying node:sqlite database instead of only dropping the cached reference, so the fixture file is unlocked and Windows can unlink the temp dir (was EBUSY in the adapter test teardown). - The reconstruct test compared emitted absolute paths against hardcoded POSIX strings; on Windows resolveAgainstCwd normalizes to backslashes, so expectations now mirror that normalization. Production code unchanged.

- A failed apply_patch update hunk left the prior in-session buffer in place and still emitted the file as faithfully reconstructed; drop it to unreconstructable so stale content is never linted as the model's output. - Sessions touching only non-lintable files (e.g. markdown) had zero reconstructed files and zero failures but were counted as "unreconstructable"; require an actual reconstruction failure for that bucket so the skip note stays accurate.

Replace the readFileSync + split("\n") transcript reader with a streaming node:readline parser so memory stays flat on large Claude/Codex transcripts. Makes session loading async (SessionCandidate.load + the parse adapters); the Cursor composer load wraps its sync DB walk to match.

…--since) - Drop scans that error/skip/lint-fail instead of counting them as clean code, which was inflating the leaderboard. - Emit structured JSON on failure in --json mode (reuse enableJsonMode), which also silences the incidental score-API stderr warning. - Exclude unknown-timestamp candidates under --since so the filter is consistent. - Consolidate the path-inside predicate, move render magic numbers to constants, type-guard the provider flag, throw on invalid --limit, rename op -> operation.

… content A replace/Edit whose oldString isn't in the in-session buffer now marks the file unreconstructable (like a failed apply_patch hunk) rather than keeping the stale snapshot and scoring it as the model's final output.

cursor

Cursor Bugbot has reviewed your changes using default effort and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit 9a20e3d. Configure here.}

Confidence weighting now counts only sessions that contributed scanned files, so non-React/failed/skipped sessions no longer raise session reliability or effective file weight. The reported per-group session count still reflects every analyzed session.

react-doctor has a runtime `deslop-js: workspace:*` dependency, but the Continuous Releases workflow didn't publish deslop-js, so pkg.pr.new couldn't rewrite the ref and `npx https://pkg.pr.new/react-doctor@<pr>` failed with EUNSUPPORTEDPROTOCOL ("workspace:"). Add deslop-js to the publish set.