In multi-agent coding, the dangerous moment is when BOTH agents say "done" and nobody verified — these six wheels are the verifier.
When you run more than one AI agent on a codebase, the failure is rarely a loud crash. It's two agents that each report success, a green checkmark on both sides, and a bug that ships anyway because no one held a clean-context view of the change. Or quieter still: one agent simply stops, and nobody notices for hours. These six skills add the missing checks: a drift mirror, an adversarial review panel, a reuse compiler, a frontend smoke gate, a delegation bound, and a viral-repo triage protocol.
/plugin marketplace add starshard-ai/reviewer-wheels
/plugin install reviewer-wheels@reviewer-wheels
Or install directly from the repo:
/plugin install reviewer-wheels --from github:starshard-ai/reviewer-wheels
-
mirror-pair — A bounded k=1 observer/executor loop. One agent executes; one clean-context mirror reviews a compact state packet for drift, broken invariants, missing validation, and loop closure. It does not redo the work — it checks it, and it isn't closed until there's a review receipt. Provider- neutral: executor and mirror can be any agents, ideally with different blind spots.
-
adversarial-reviewer-panel — Turns "ask some reviewers" into a constraint-driven loop. You design 3-5 persona reviewers, each bound to a constraint class (falsifiability, control/safety, public reality, HCI, invariant discipline...), and each must report a strongest insight, a failure mode, a concrete action implication, and a disconfirming signal — no generic praise. Built to find what a design forbids or breaks, not to produce a chorus of opinions.
-
meta-pipeline-compiler — Stops agents from rebuilding the same wheel. Before building anything non-trivial, it searches existing skills, scripts, docs, task ledger, memory, and capability registry, emits a reuse-check receipt, and promotes recurring work into the right durable substrate. Also repairs drift when memory says a capability exists but the local copy is gone.
-
frontend-smoke-gate — A gate for user-facing UI changes. Run build/lint, then a focused smoke that opens the actual route, hits the backend, and produces a screenshot the agent inspects itself before reporting success — catching stale data, overlapping panels, clipped text, and missing receipts. Includes a staged release guard for anything publicly live.
-
delegation-bound-gate — For the failure where nothing crashes: a subagent hangs, the parent waits, and hours disappear with no error anywhere. Requires every delegation to carry a hard cap plus a partial-report contract, and gives the measured procedure for judging a silent worker (check open sockets and probe the target — a hung tool call shows zero connections) instead of guessing from file mtime, blocking on a non-critical path, or hedging with a second agent. Ships an opt-in
PreToolUsehook, because the rule it enforces had already been written down in plain text and broken four times. -
viral-repo-triage — For the failure where one human CODEOWNER is buried under a flood of fork PRs, no labels exist, and duplicate fixes race each other until either duplicates ship or nothing ships. This wheel turns triage into a state machine where agents do everything upstream of the merge or close decision: label, de-duplicate, route security, arbitrate claims, first-pass review, and digest. It also ships a drop-in GitHub governance kit that a maintainer can apply to their own repo with their own credentials.
MIT. See LICENSE.