Skip to content

Latest commit

 

History

History
55 lines (40 loc) · 2.56 KB

File metadata and controls

55 lines (40 loc) · 2.56 KB

github-self-reviewed

Ships a change as a GitHub pull request that you review yourself. It's the same as github-peer-reviewed but without the separate review gate. There's no peer to wait on: you review the change during or after implementation and approve it by advancing the task to merging.

PLANNING → ITERATING → MERGING → COMPLETE

(plus DROPPED, reachable from any state.)

When to use: GitHub-hosted changes you review yourself, with no peer-review gate.

This workflow is opt-in: enable it for a repo before it appears in the task-creation picker.

Lifecycle

State What happens Who advances
PLANNING The agent collects requirements and writes a plan.md artifact (read it from the dashboard: highlight the task and press a) plus a token estimate. You, by invoking advance with your harness's syntax (see the workflow guide).
ITERATING The agent implements the plan, commits and pushes, opens a draft PR, and gets CI green. You self-review the change here. You: advancing to MERGING is your approval.
MERGING The agent adds the PR to the merge queue and re-adds it if it falls out. The agent, which advances itself once the PR is merged.
COMPLETE Terminal. The change has landed. n/a

There's no REVIEW state: with self-review, "tell the agent to proceed to merging" is the approval. You can also send the task back to ITERATING from MERGING at any time if you spot something.

Your part and the agent's part

  • You: approve the plan, review the change yourself, and advance out of ITERATING when it's good to merge. Nothing merges until you do.
  • The agent: does the planning, coding, testing, PR authoring, CI-watching, and merge shepherding. In MERGING it drives to a merge on its own.

Skills

The agent has these forge skills in its container (it runs gh against GitHub):

  • open-pr pushes the branch and opens a draft PR, then records the PR URL on the task (so the dashboard's p hotkey opens it).
  • babysit-ci watches the PR's CI and fixes failures (and base conflicts) until it's green, without tying up a turn while CI runs.
  • babysit-merge shepherds the PR through the merge queue, re-queuing if it's kicked out, until it merges.

Related