Opt-in, restart-safe CI actions for native PR watches

[kaovilai]

Idea

Let a thread opt in to CI actions that T3's native PR watch performs for the agent, instead of waking the agent to do them. Today watch_pull_request wakes the agent on failed checks, passed checks, new comments and conflicts; any retry or nudge costs an agent turn.

This is a follow-up to #17644, which covers the detection signals (required checks that never report, behind-base). It builds on them and on #16924.

Why it might be worth it

• Retrying an infrastructure-looking failure, nudging a required job that never started, or re-requesting a validation run that timed out in a queue are mechanical. An agent turn is a lot of tokens for "post the retry comment".
• Doing them server-side lets the agent stay asleep until something needs judgment.

The counterpoint: the agent is already woken on these events and can act itself, so the saving is agent turns, not capability. I'd like to hear whether that is worth the extra surface and risk.

Sketch

Per-watch, opt-in rules, for example:

{ on: failed | missing | stale-pending, match: <check name pattern>, act: comment(<body>) | rerun, after: <duration>, maxPerHead: <n> }

• comment works on any hosting provider (for example a retry-style command comment). rerun only where the host has a rerun API.
• "Looks like an infrastructure failure" means a configurable match on the check name or summary text. The watch would still not read CI logs.

Restart safety

• Persist a log of (commit, check, rule) keys in the watch state so a restarted watch does not repeat an action.
• Before acting, also look for the same comment on the PR since the head commit, which catches actions taken by an older watch or by hand.

Safety

• Explicit opt-in per watch before anything is posted as the user.
• Per-commit and per-hour caps, and stop on a failed post.
• Respect hosts and organizations that disallow automated comments.
• Do not act on PRs the user does not own, or on forks without push rights.

Out of scope

Rebasing branches automatically, and diagnosing CI logs.

Note

Responses generated with Claude