SKILL.md
$2b
-
Read the plan and extract:
- Files to Modify section
- Full body text
- Plan scope (count files changed, count body lines)
-
Assemble the review panel — invoke /assemble-panel with scope = plan file. It returns the panel based on file types and keywords in the plan. Pass --include/--exclude as overrides. Log which agents were selected and why.
Codex/Cursor: read .claude/skills/assemble-panel/SKILL.md directly and apply the scope/keyword maps inline.
If /assemble-panel is unavailable, fall back to [technical-editor, code-reviewer] with gate=P2, cap=3.
-
Dispatch all selected agents IN PARALLEL using the Agent tool (single message, multiple Agent tool calls). Each agent receives:
- The plan file content
- Their specific review mandate
- Instructions to return: APPROVE / REVISE / DROP with findings tagged P0/P1/P2
-
Collect verdicts and synthesize into a summary table:
## Plan Review Summary
| Agent | Verdict | P0 | P1 | P2 | Key Finding |
|-------|---------|----|----|-----|-------------|-
Gate logic and review loop — invoke /assemble-panel for panel governance. It provides the policy algebra (gate=P2, cap=3) and handles RETAIN, EXPAND, CONVERGE, and ESCALATE_RECURRING. Do not duplicate those rules here.
After each round:
- All APPROVE with no above-gate findings → "Proceed to Build"
- Above-gate findings remain → fix, then re-review per assemble-panel's RETAIN + EXPAND rules
- Cap reached with unresolved findings → escalate to human
If /assemble-panel is unavailable, fall back: keep all reviewers with above-gate findings, exit at round 3 or when all clean.
Panel Bounds
- Minimum: technical-editor alone (1 agent, trivial plans)
- Typical: technical-editor + code-reviewer + architect-reviewer (3)
- Maximum: all agents (rare, large cross-cutting plans)