panel

Orchestrate a rigorous multi-agent implementation, review, and verification panel inside Herdr. Use when the user requests Panel or a Herdr-managed panel review.

SKILL.md

GitHub

Panel

Run independent reviewers in panes created for the current task, then return one deduplicated set of evidence-backed findings.

  1. Verify HERDR_ENV=1. When it is absent, tell the user to start the primary agent inside Herdr and stop; Panel does not bootstrap or attach to Herdr from an ordinary terminal.
  2. Read ../herdr/SKILL.md and follow its discovery, targeting, lifecycle, and safety rules. Use /code-review as the authority for standards and specification review.
  3. Learn the installed CLI with herdr --help and herdr agent. Treat its reported agent kinds as authoritative. Honor agent kinds requested by the user; otherwise select from installed kinds appropriate to the task. Never assume one vendor or executable.
  4. Inspect the calling pane, create the required sibling panes with herdr pane split, and preserve the current working directory and user focus. Panel creates its own panes; the user need not prepare a layout.
  5. Start each reviewer with a unique role-based name through herdr agent start <name> --kind <kind> --pane <pane-id>. Supported kinds may include Claude, Codex, Copilot, OpenCode, and others reported by the installed Herdr version.
  6. Give reviewers non-overlapping assignments. For implementation work, cover specification compliance, repository standards, and executable verification. Include the fixed comparison point, issue or specification, and repository instructions each reviewer needs.
  7. Wait for each reviewer to settle, inspect blocked or timed-out panes before sending more input, and collect the complete results. A timeout does not prove a prompt was not delivered.
  8. Validate findings against the diff and test evidence, remove duplicates and unsupported claims, and report actionable findings by severity. If no valid findings remain, say so and report the verification performed.

Panel coordinates local agents and commands only. Pushing, opening or merging a change request, answering unfamiliar approval prompts, and deleting panes or workspaces require the user's existing authorization.