A tool waits to be used. A machine is self-propelled.
Style, static errors, forbidden patterns.
Fast, isolated, run on every change.
Service boundaries and real dependencies.
Return to the agent with the failing output.
[08:48:38] TOOL CALL git push origin fix/checkout-latency [08:48:42] TOOL RESULT pushed 1 commit · CI run #4821 queued [08:48:48] POLL CI gh run watch 4821 --exit-status [08:49:30] TOOL RESULT CI failed · unit 4 failed, 205 passed · lint 1 error FAIL checkout.test.ts › clears SSO session · src/session.ts:88 [08:50:18] TOOL CALL edit src/session.ts · revoke SSO before clearing cookie [08:50:31] TOOL CALL git push origin fix/checkout-latency [08:50:41] POLL CI gh run watch 4822 --exit-status [08:51:34] TOOL RESULT CI passed · unit 1475/1475 · integration 65/65 · lint clean [08:53:12] ASSISTANT CI green · ready for review
Libra reads the change in context.
Inline comments on the pull request.
The agent pushes; Libra reviews again.
Libra can block a change that passes tests.
The app, services, and data running locally.
Ad-hoc journeys, not only scripted tests.
Every state, every viewport.
Layout, copy, console errors.
Unclear intent
The expected behavior is not specified.
Define the outcome before starting.
Drift
Repeated review cycles produce overly defensive code.
Stop after a fixed number of iterations.
Missing context
The agent cannot tell that an unreleased feature does not need a database migration.
Source of truth on GitHub (issues, PRs)
Slack bridge for notifications and assigning work
Mac mini running continuously
tmux sessions with pi (running Opus 5.5)
Skills drive the loops; scripts handle polling