Embed

---
name: human-pr-review
description: Run the machine PR review — dispatch the adversarial reviewer on the open draft PR and record its verdict
argument-hint: <key> --pr=<number> --branch=<branch>
---

<!-- human:include launch-briefing -->

`$ARGUMENTS` is `<KEY> --pr=<number> --branch=<branch>` — the PM ticket key, the open (draft) pull request to review, and its branch, all supplied by the daemon's deploy loop. Parse them, then delegate to the **human-pr-reviewer** agent.

Run the reviewer at the `opus` tier: it is the adversarial gate before a merge, and a weaker model gets talked out of real objections, turning the check into a rubber stamp (never tier down an adversary). It stays at `opus` rather than joining the other adversaries at `fable` for one reason and it is not a judgement about the work: at 108 spawns this is the highest-volume dispatch in the pipeline, so moving it is a measurement someone has to take, not a call to make from the policy table.

```
Task(subagent_type="human-pr-reviewer", model="opus", prompt="Review PR <number> for ticket <KEY> --branch=<branch>", run_in_background=false)
```

The agent reviews the fixer's **local** branch commit (not the stale pushed head — the fixer never pushes in board context, so origin's head is pre-fix), records its findings and the machine verdict (`approved | changes-requested | unreviewable`) in `stage.pr-review`, and mirrors findings onto the PR as inline comments when it has a write path. The daemon's loop reads that verdict to decide the next step — another fix pass, or the merge — so you do **not** post board markers, dispatch a fixer, or merge anything yourself: run the agent and report its verdict. Human review of the PR happens out of band and never gates this run.

A reviewer that could not reach the substrate records `exit: outage` instead of a verdict: the daemon parks the card on `[human:deploy-outage]` and re-drives the review when it returns, charging no round — so you neither retry it nor report it as a failure.