[calvin-archastro]

Why

Agents compute risk in two places: the sealed risk assessment on plan, task and PR events, and the per-hunk risk annotations on PRs. Today both are computed once and posted as labels. The assessment finds medium or high risks, and nobody acts on them before posting.

Change

references/monitor.md, sealed assessment: a new step 4, "Mitigate, then recompute".

• For each medium or high risk driver, reduce the risk in the subject itself: fix the code, add a test or guard, split the unverifiable step, or tighten the plan.
• Then re-author the assessment and finalize it again.
• Never lower a grade by editing the assessment alone.
• Run at most two rounds. Fix only within the existing scope. If a fix would expand scope, don't make it: record the risk as remaining and move forward.
• Name what was mitigated, and what remains, in --message.

references/monitor.md, PR hunk annotations: a new step 3.

• Fix real medium or high risk ranges with a test that would have caught them, and push.
• Annotate the new head. A new head gets a new row, so --force is never needed.
• At most two rounds, fixing only within the PR's scope. A risk whose fix would expand scope stays annotated, stated plainly, and the agent moves on.

SKILL.md: one sentence that points to this loop (within scope, at most twice).

Factory sessions are unchanged: they still skip hunk annotations, because the host writes those.

🤖 Generated with Claude Code