⏩ thunderkit

Opinionated multi-model delegation for very large repos.

skill

tk-review

Use to review and verify completed big-repo work: fans a diff to two-plus model families for cross-family review, consolidates findings by severity, and runs each lane's verification command so done means evidence, not intent.

Install: npx skills add thunderock/thunderkit -s tk-review -g

---

tk-review — cross-family review + evidence gate

Two modes

• tk-review (default) — review a completed diff (post-execute). Both jobs below.
• tk-review --plan — review the *plan* before execution (the plan-check gate, lifecycle

stage 8). Fan PLAN.md/plan.json to the reviewer families and check: are lanes truly disjoint, does every lane have a runnable verify, are the dependency layers acyclic, do lanes cite real symbols (not hallucinated names)? Output .thunderkit/PLAN-REVIEW.md. tk-execute refuses to start when review_families_min ≥ 2 and no PLAN-REVIEW.md exists.

The two jobs (default mode)

tk-review does two inseparable jobs (merged by design):

1. Cross-family review — fan the change to ≥2 model families and consolidate. A model

family reviewing its own output is not review; the author's family cannot be the only reviewer.

1. Evidence gate — run each lane's verify command. A lane without a passing verification is

not done — it's blocked. Done means evidence, never intent.

Preferred reviewers: Sol + Opus 5 (at least one different from whoever authored the lane). Verification runs on Fable 5.1 (running commands is cheap). See ../references/model-roster.md.

Cross-family review procedure

1. Identify the author family per lane (from .thunderkit/runs/<lane-id>). Choose reviewers

from *other* families — if Opus authored, review with Sol (+ Opus 5 as the strong same-lineage second, but never Opus alone).

1. Fan the diff to each reviewer via portable dispatch (roster dispatch table). Send the lane

diff, its acceptance criteria, and the goal. Ask each for findings with severity (blocker / major / minor / nit) and a file:line anchor.

1. Consolidate — merge reviewer outputs, dedupe overlapping findings, keep the highest

severity when they disagree, and record *which reviewer* raised each (families disagree — that disagreement is signal, preserve it).

1. Show each reviewer's model inline: (Sol), (Opus 5). Best-effort reviewers (Sol is

credit-capped) that fail are dropped with a note, not silently omitted.

Evidence gate procedure

For every lane in the plan:

1. Run its verify command from plan.json.
2. Record pass / fail / blocked with the actual command output (truncated), not a summary.
3. A lane is done only if: verify passes and it has no unresolved blocker-severity review

finding. Otherwise it's blocked — name what's needed.

Output contract — .thunderkit/REVIEW.md

## Lane L0-auth-token-refresh
- Author: Opus 4.8 | Reviewers: Sol, Opus 5
- Verify: `cargo test -p auth token::` → PASS (12 passed)
- Findings:
  - [major] (Sol) src/auth/token.rs:88 — backoff not jittered; thundering herd on mass expiry
  - [nit] (Opus 5) src/auth/token.rs:40 — name `t` → `token`
- Status: BLOCKED (1 major unresolved)

Plus a roll-up: N lanes, X done, Y blocked, and the consolidated blocker list that must clear before the change is shippable.

Opinions this skill enforces

• ≥2 families or it's not a review. If only one family is available/authed, say the review is

single-family (reduced confidence) and name what to install for a real cross-family pass — don't quietly downgrade.

• No verify, not done. A lane whose verify can't run is blocked, full stop.
• Preserve disagreement. When families split on a finding, record both positions; don't

average them into mush.

Degrade honestly

Sol credit-capped (429) and no other second family authed? Report the review as best-effort single-family, list the specific blocker findings you *could* get, and recommend the login that restores a cross-family gate. Never present a single-family pass as a full review.