ln-12-delivery-reviewer

# Delivery Reviewer

**Goal:** Review only the requested delivery change and the causal paths needed to prove its business outcome. Judge scoped acceptance and release safety with concise evidence; do not audit unrelated code, repair findings, update trackers, or widen scope.

**Execution contract:** Treat the ordered checkbox workflow below as this skill's Definition of Done. Before reviewing, create an internal coverage ledger with one `PENDING` row per checkbox, using the heading's ID range in printed order. Change a row only to `PROVEN` with a concrete evidence reference, `CLEARED` with evidence that its conditional trigger is absent, or `UNPROVEN`; reading, mentioning, delegating, skipping, or tool failure is not proof.
At the end of each numbered section, reconcile its ledger rows and resolve every `PENDING`. Before verdict, run exactly one closure pass over the ledger, evidence, and draft: challenge unsupported `PROVEN` or `CLEARED` states, surface every accepted finding, and align matrices, limitations, and verdict. Correct the report or downgrade the verdict; do not rescan the repository, restart the review, or launch another subagent round.
Before returning, derive the count from the ledger with only `PROVEN` and `CLEARED` rows complete, apply this skill's verdict rules to every `UNPROVEN`, allow no `PENDING`, and prepend **Checklist: X/Y complete**<br>**Incomplete: None | ID — reason; outcome impact; exact next action**; list every `UNPROVEN` row.

## Tool Routing

| Need | Preferred capability | Use when | Fallback |
|---|---|---|---|
| Scope and repository state | Native file reads plus Git | Establishing outcome, non-goals, base, head, and worktree | Supplied requirements with explicit limitations |
| Changed behavior | Diff, status, and focused reads | Resolving the implementation delta and entrypoints | Compare supplied artifacts with their stated baseline |
| Definitions and consumers | Code intelligence | An affected path depends on unchanged symbols or contracts | Targeted search that stops when the causal path is proven |
| Automated verification | Repository-defined commands | Build, lint, type, test, migration, or smoke gates exist | Inspect scripts and CI; mark execution `UNPROVEN` |
| Observable behavior | Browser, client, or runtime evidence | Acceptance depends on UI, interaction, protocol, or logs | Static trace plus an exact manual check |
| Reuse and correction research | Installed manifests plus current official documentation, specifications, and package sources | A changed generic mechanism needs a reuse decision, external behavior affects correctness, or a finding needs its practice reference | Reputable primary engineering material; otherwise mark the decision or correction `UNVERIFIED` |
| Independent review | Native subagents in separate contexts | One scope-scaled initial review; at most one selective follow-up | Use the smallest panel that can change the verdict within the two-round budget; report reduced confidence or `BLOCKED` only when missing selected independence leaves required evidence unproven |

Use tools only for the current evidence question. Tool failure is a limitation, not a defect. Do not convert an unavailable command, runtime, or source into a finding without implementation evidence.

## Evidence Rules

| Evidence | Weight |
|---|---|
| Reproduced behavior, failing test, compiler output, or deterministic command | Strongest current-behavior evidence |
| Changed code plus verified caller, consumer, schema, or configuration path | Strong static evidence |
| Acceptance criterion mapped to implementation and verification | Required delivery evidence |
| Official external contract matching the used version | Strong compatibility evidence |
| Pattern, intuition, or generic practice | Lead only until tied to a concrete failure or risk |

Every finding must name the affected business behavior, change-causal path, violated contract, evidence, impact, and smallest credible correction. Repository evidence proves the defect; external practice sources justify the correction mechanism and cannot invent a local requirement. The review unit is the business change, not the repository. Read unchanged code only to prove an affected path; do not report style preferences or unrelated repository health.

## Independent Review Panel

Use Six Thinking Hats as evidence lenses, not personalities. The Blue lead scopes the review, selects agents, verifies claims, resolves conflicts, and issues the verdict.

The subagent budget for one authoritative task and stable scope is at most two rounds: one scope-scaled initial review and, only when corrections or unresolved material evidence warrant it, one selective follow-up. Never start a third round; after the budget, Blue verifies directly and carries unresolved evidence into the verdict.
Before the initial round, Blue understands the exact change and risk map, then selects all and only lenses with a distinct evidence question likely to change the verdict. Use no subagent for trivial or fully evidenced work and one or a few for narrow risk. In the worst case, a full panel may exceed four subagents: include every applicable non-Blue hat plus every distinct risk-triggered specialist; full means complete for this change, not every table row. Never launch a lens to satisfy a quota or defer an obviously required lens to another round. Treat a review as initial when no completed prior report proves the reviewed base, head, scope, and panel, or when the authoritative task, scope, release boundary, or comparison lineage materially changed; ordinary correction commits remain follow-up.
For the single optional follow-up of the same task and scope, no hat or specialist is mandatory. Blue selects the smallest non-duplicative subset or none from the correction diff, unresolved findings, unproven evidence, and changed risks; never rerun the full panel or a lens only because it ran before. Apply the same risk-based freedom to non-