Hive Review Loop

Execute one repository task through Hive. hive_workflow owns the loop, parallel reviewer fan-out, consensus, run artifacts, resume, and cleanup; hive_registry owns every implementor and reviewer agent, capability, provider session, authentication home, and provider process.

This skill is an orchestrator only. Never spawn, delegate to, or simulate a Codex subagent. Do not launch provider CLIs directly. Invoke only the bundled launcher, which calls hive_registry for preflight and hive_workflow for the actual run.

Input

Require a single-run YAML or JSON configuration. Read references/configuration.md for the complete contract and example. The configuration names:

• one absolute Git worktree root and one external Hive run directory;
• one registry agent with act capability as implementor;
• one or more registry agents with read_only capability as reviewers;
• the round limit, diff inclusion, and successful-run cleanup policy; and
• complete task text and/or exact absolute context sources with their purpose.

Reviewer entries form a panel. Hive runs every reviewer in a round concurrently and continues only when all return PASS. Repeating one registry agent name is allowed and creates distinct reviewer slots and sessions.

Execute

1. Inspect the configuration and confirm the requested task authorizes an implementation agent to modify the named repository.

2. Validate without launching agents:

   python3 "<skill-dir>/scripts/run_hive_loops.py" single "<config.yaml>" --validate-only

   This resolves both Hive CLIs, verifies the Git root, checks registry capabilities and doctor, confirms the bundled workflow exists, and rejects unsafe artifact placement or an accidental existing run.

3. Run the loop:

   python3 "<skill-dir>/scripts/run_hive_loops.py" single "<config.yaml>"

4. Report the Hive run directory, round reports, summary.json, launcher logs, registry roster used, consensus status, and cleanup status.

Exit 0 means every reviewer passed and requested cleanup, when enabled, succeeded. Exit 1 preserves the run and sessions for diagnosis or resume. Exit 2 is a configuration or preflight failure and launches no agents.

To resume a preserved run after correcting an external problem or increasing max_rounds in the same configuration:

python3 "<skill-dir>/scripts/run_hive_loops.py" single "<config.yaml>" --resume

Never replace a failed run directory, clean its registry sessions, or claim consensus merely because all Hive processes exited. The Hive workflow's machine-readable consensus result is authoritative. The launcher requires the final collected approved verdict in summary.json and the workflow exit code to agree; disagreement preserves the run and reports needs_attention.

Boundaries

• Reviewers are strictly static and read-only; implementation-side validation belongs to the registry implementor.
• Context sources may be outside the repository, but every configured provider sandbox must be able to read them. Host-side existence does not prove that.
• Run artifacts stay outside the reviewed repository so they cannot pollute the submitted diff.
• The launcher does not commit, merge, push, or integrate repository changes.

Example activations

• "Use these three registry reviewers in parallel until they all approve this repository task."
• "Run the Hive review loop from this YAML and preserve the run if consensus fails."
• "Resume the existing Hive run with a larger round limit; do not create subagents."