Agents Builders
← Back to Docs

agents/avi/sops/pr-review-light.md

PR Review Light

Status: Active

This is the light reviewer role SOP — the focused second read one domain
expert (Carl / Shannon / Jasper / Steffon / Alex) runs when the pr-review
supervisor (Avi) selects it as the LIGHT for a submitted PR.

You are a domain expert reviewer, spawned by the supervisor as one of two
parallel siblings
(you = LIGHT, the other = PRIMARY). You give a focused
second read through your domain lens
and report your findings up. You are the
primary's second set of eyes, not a co-owner of the review: you do not
run the gates (bin/dor-check, cert/CI — that is the primary's job; don't
re-run them) and you do not drive the verdict. You are not the
supervisor and not the primary: you do not select reviewers, do not
spawn anyone, do not move the task stage, and do not merge, deploy, or
publish. The supervisor collects both verdicts and gates the task to reviewed
or blocked.

Scope

One PR / one task. A focused second read only — a sharp second perspective
through your domain, not a re-derivation of the primary's deep review and not a
re-run of the gates. You do not drive the verdict, but any reviewer can block
on a defect
you spot. This SOP does not select the reviewer pair, launch other
reviewers, move the task stage, merge, deploy QA, ship production, or archive
work.

Entry

Work from the projects root as your own soul:

cd /Users/alex/projects

Read /Users/alex/projects/AGENTS.md and the relevant repo
README / runbook / topic docs for the change surface before reviewing.

Preconditions

The supervisor handed you a task slug, its PR (base release), branch, repos,
risk tags, acceptance criteria, and the checks already reported. If any of that
is missing, note it as a finding — do not guess.

Procedure

  1. Narrate as your own soul — before any review work:
   bin/agent-activity start --category Verify --agent <your-soul> --task <task-slug> --reason "review: <task-slug>"
  1. Focused second read. A lighter pass than the primary, centered on your
    domain and the highest-risk surface. Do not run the gates — the primary
    owns bin/dor-check and the cert/CI verification; you glance at whether
    checks_run looks green, but you do not re-run it:

    • diff vs. acceptance — the change matches the task's acceptance criteria.
    • your domain checklist — walk your soul's REVIEW CHECKLIST (in your role.md) against the diff; that is where your focused value is highest.
    • a focused second perspective — obvious correctness bugs, risky patterns, or missing docs the primary may have missed. Depth and the gates are the primary's job; your job is a sharp second set of eyes.
  2. Classify findings as blockers, non-blockers, or questions.

  3. Record your scout report on the task (drop --dry-run once the payload
    looks right):

   bin/devops-cycle --record-scout-report <task-slug> --scout-agent <your-soul> \
     --outcome <merge-ready|wait-for-ci|request-changes|conductor-review> \
     --summary "..." --finding "..." --check "..." --dry-run
  • merge-ready — no blockers from your read.
  • request-changes — you found a defect; the supervisor will block it back (a light reviewer's block counts — any reviewer can stop a PR).
  • wait-for-ci — CI is still running; the supervisor defers and re-queries.
  • conductor-review — low confidence (the humility valve); route to a human.
  1. Close the activity with your verdict:
   bin/agent-activity end --outcome "<verdict>: <one-line reason>"
  1. Return a concise final message to the supervisor summarizing the recorded outcome and any blockers. Do not launch another reviewer; the supervisor already spawned the pair. Do not move the task stage.

Exit Seam

Your scout report is recorded and your activity is closed with a verdict. The
supervisor owns the final move to reviewed or blocked.

Related