multi-agent-reviewlisted

# Multi-Agent Review Same task, several models, no one merges anyone else's work. Every participant works in its own worktree; the orchestrator only relays and, at the end, rates. This skill is invoked two ways: **directly** by a human (interactive operating mode), or by the preserved `legacy-coder`, `legacy-debugger`, and `legacy-planner` workflow. The round mechanics (Steps 2-4) are identical either way — only how participants are picked (Step 1) and how the result is resolved (Step 5) differ. ## Mode — what "the task" actually is | Mode | Used by | Task | Artifact | |---|---|---|---| | `build` (default) | `/legacy-coder` Step 2 | Implement the ticket's acceptance criteria test-first | A code diff, gated by the ticket's `Verification-command` | | `harden` | `/legacy-debugger` Steps 1-2 | Run the four-nets audit + red-team pass against the ticket's landed diff, fix everything found, test-first | A fix diff on top of the existing branch, same gate | | `plan` | `/legacy-planner` chain steps 3-4 | Turn the already-**grilled** decisions + locked invariants into a spec and dependency-ordered tickets | A spec + ticket set per participant, no code gate | `plan` mode never runs the interactive grill itself (`grill-with-docs` stays a single human conversation — you can't parallelize asking the user a question). It starts only after `/legacy-planner`'s Step 1 (grill) and Step 2 (lock invariants) are already done; participants compete on turning those locked decisions into a spec/