Workflows

After each check, four rules look at the PR. Each fires once per state (per commit, or per new set of failures or threads), not on every check.

Rule	When	Action
merge	CI green, approved, no unresolved threads, mergeable, not draft	gh pr merge --<method> --match-head-commit <sha>
review_threads	CI green, approved, unresolved threads	ask the PR's agent session to go through the threads (no pushing)
ci_failed	CI failed	send failed checks and log tail to the agent: rebase if flaky, fix if real
needs_review	CI green, no approval, not draft	Slack "review needed" and gh pr merge --auto

Each rule is off, dry-run (default) or on:

"workflows": { "merge": "dry-run", "review_threads": "dry-run", "ci_failed": "dry-run", "needs_review": "dry-run" },
"merge_method": "squash"

Dry-run builds the prompt or command and stores it as the task result without running it. Actions re-fetch the PR first and skip if it changed. Every rule that fires also sends a Slack message.

Retriggering CI

watchcat pr retrigger <ref> (or the Retrigger button) re-runs failed GitHub Actions jobs with gh run rerun --failed and checks again a minute later. Other CI systems are skipped.

Agent hand-off (opencode)

watchcat talks to opencode through opencode api, so nothing extra needs to run. Link a PR to the session working on it:

watchcat agent sessions                                   # open sessions and their worktrees
watchcat pr bind owner/repo#123 --session ses_abc123
watchcat pr prompt-ci owner/repo#123                      # send the CI-failure prompt now
watchcat pr unbind owner/repo#123

or pick the session on the PR page, or let the agent do it through the MCP server. A link is only accepted if opencode reports the session exists and works in that worktree. Prompts are queued into the session and never retried automatically, so the agent never gets the same work twice.