Two ways to run agents. One place they ship.

Magic Sessions take on long-horizon tasks your team steers together. Automations cover the repository events you never want to babysit again. Every run lands as a reviewable pull request.

Get startedSee pricing

Role

You detect breaking changes in a pull request and label it accordingly.

Inputs

Resolve the PR number from /workspace/CONTEXT.json at .trigger.event.data.target.number.

Process

  1. try pr diff <number> --json — get the changed files and hunks.
  2. Look for changes that break consumers:
    • Removed or renamed exported functions, types, classes, methods.
    • Changed function signatures: required argument added, type narrowed, return type changed.
    • Removed or renamed config keys, env vars, CLI flags.
    • Database column or table drops, NOT NULL added without default, type narrowed.
    • HTTP route removed or method/path changed; response shape narrowed.
  3. Distinguish from non-breaking: pure additions, optional arguments, internal-only refactors, test-only changes.

Output

  • If breaking change found:
    • try issue label <number> --add breaking-change (the issue-labels endpoint covers PRs as well).
    • try pr comment <number> --body "<summary>" with a 1–3 bullet summary of the specific breaks (file + symbol + what changed).
  • If no breaking change: do nothing. No label, no comment.

Rules

  • Only label if you can name the specific symbol or surface that broke.
  • Internal or private symbols (underscore-prefixed, files in internal/ or _internal/) do not count as breaking.
  • Do not remove labels you did not add.
  • Never speculate (this might break X) — if you cannot point to the line, do not flag it.

Stop conditions

  • After applying the label and posting one comment, stop.
  • If nothing breaks, stop silently.
Magic SessionsEarly access

A sentence in. A pull request out.

Shared sandboxes for long-horizon work. Start a task, let it run, and anyone on the team can watch, steer, or take over — until it lands as a pull request that automations and CI review first.

Get early access

Run models where the code already is

A prompt, a trigger, a model, and a permission scope. Automations run inside the platform on real repository events, using scoped repository credentials — no third-party app to install, no token handed to someone else's cloud.

Any frontier model, one balance

Frontier models from every major provider — pick one per automation, with premium routing and ZDR options when code cannot touch a retaining provider. Usage draws from the same balance as CI.

Triggered by repository events

A pull opens, CI fails, a bot is mentioned — the run and its output land on the timeline. Six official templates install in one click; edit the prompt or start from an empty one.

Premium, fast routing runs through Vercel AI Gateway and Wafer.

Start from a template, not a blank prompt

Six automations ship with the platform, each already wired to the event it belongs on. Install one as-is, edit its prompt, or start from an empty one.

Templates for the work you repeat

Review, CI repair, triage, release notes. Each arrives with its trigger, model, and repository scope already set.

Nothing runs without a receipt

Every run is filed under the automation that started it, with the status it reached, how long it took, and what it spent.

Something you can try today

Agent execution, review, permissions, and CI in one system. No credit card and no trial time. Just free.

Start freeSee pricing