github-review-looplisted
Install: claude install-skill brenpike/hivemind
# GitHub Review Loop
Watch one open PR and drive review-feedback remediation to a terminal state. This
skill owns the loop lifecycle; it does NOT read, interpret, or classify feedback
content — that is `hivemind:github-reviewer` fix-mode work. The skill's only
GitHub footprint is a thin, cheap, predefined change-detect poll that answers
"did anything change?" and "is the PR terminal?".
This skill is intent-based: it describes intent plus mechanical safety rails
(Monitor wiring, terminal conditions, dispatch contract). It is deliberately NOT
a step-by-step state machine.
Load and follow: `${CLAUDE_PLUGIN_ROOT}/governance/definitions.md`, `${CLAUDE_PLUGIN_ROOT}/governance/safety-rails.md`, `${CLAUDE_PLUGIN_ROOT}/governance/security-policy.md`.
## Why this runs in the main session
Monitor is a main-session cross-turn primitive. A subagent runs exactly one turn
and returns when it stops emitting tool calls, orphaning any Monitor armed inside
it. The loop therefore lives here, in the overlord's tool context, where Monitor
survives across the reviewer-subagent dispatches this skill makes. The overlord
spawns the reviewer (only the top-level orchestrator can — ADR-0005).
## Inputs
| Input | Default | Meaning |
|---|---|---|
| `pr` | (required) | PR number or URL to watch. |
| `working_branch` | (required) | The branch the reviewer pushes fixes to. |
| `base` | (required) | The PR's base/target branch. |
| `reviewer_filter` | `codex-only` | Which reviewer identities count as a