pr-monitor/README.md
2026-07-29 09:24:20 +00:00

69 lines
2.8 KiB
Markdown

# pr-monitor
Single-user dashboard for tracking PRs across multiple GitHub repos: CI checks
segmented into lint / specs / e2e / other (every check visible, with per-chip
dropdowns), review state derived from review-thread resolution, and a live
unified timeline of all comments and reviews with persistent read/unread state.
## How it works
- Node/TypeScript server polls the GitHub GraphQL API every 60s using a two-tier
query: a cheap `search()` list query whose fields double as change signals
(PR `updatedAt` does **not** change when checks or thread resolution change),
then batched detail queries only for changed PRs.
- State is cached in SQLite (`data/pr-monitor.sqlite`, `node:sqlite`) so the UI
loads instantly on restart; timeline events and read/unread survive restarts.
- The browser gets a `sync` event over SSE each poll and refetches via
TanStack Query. React 19 + Vite SPA, light/dark/system theme, responsive
(two-column desktop, tabbed mobile).
## Config
`config.json` — repos, the `involves:<user>` filter flag, poll interval, port,
and per-repo check-name → category mappings (exact names or `*` globs, e.g.
`"Preview E2E *"`). Unmatched checks land in `other`; nothing is ever hidden.
Auth: `GITHUB_TOKEN` env var, falling back to `gh auth token`.
`timeline.mute` hides timeline events by actor (optionally kind / body substring)
at read time — data is kept, so unmuting is retroactive.
`failureReports` (per repo) maps failing check names to workflow artifacts that
contain structured results: `rspec` parser reads RSpec JSON-formatter output
(`rspec_results_*.json`), `playwright` reads the `data/*.md` failure attachments
inside Playwright HTML-report artifacts. Failing specs are stored (capped at
1000 per report), summarized on PR cards, and enumerated in the per-PR focus
view grouped by file. Reports re-fetch only when the workflow run id changes.
## Commands
```sh
pnpm dev # server on :4000 (tsx watch)
pnpm dev:web # vite dev server on :5173, proxies /api
pnpm sync:once # one poll, prints a PR/category table (smoke test)
pnpm build # vite build + tsc server build
pnpm start # run the production build
```
## Deploy
systemd user unit (survives reboots via linger):
```sh
pnpm build
cp deploy/pr-monitor.service ~/.config/systemd/user/
systemctl --user daemon-reload
systemctl --user enable --now pr-monitor
loginctl enable-linger $USER
```
Serves the built SPA + API on http://localhost:4000.
## Notes
- Own activity appears in the timeline but is never marked unread; the first-ever
sync marks the historical backfill as read (`BACKFILL_UNREAD=1` to override).
- Events are append-only: comments later deleted on GitHub remain in the timeline.
- Polling stretches to 5 min automatically when the GraphQL rate budget dips
below 500 points.