---
name: reviewing
description: "Use when an Auditor or Critic must examine a plan, proposal, implementation diff, gate, or evidence package before acceptance."
---

# reviewing: evidence-backed, read-only judgment

## Modes

`reviewing` has two native modes. Select the mode before reading the candidate:

| Mode | Input | Question | Owner |
|---|---|---|---|
| Critic, pre-hoc | seed, plan, graph, lane slices, gates | Can it safely and completely run? | Plan author |
| Auditor, post-hoc | brief, result, diff, evidence, gate | Did assigned work meet the brief on this tree? | Conductor/root |

Critic challenges necessity, logic, scope, complexity, assumptions, graph coverage, and gates;
Auditor checks intent, implementation, evidence, fragile builds, duplicate helpers, and CI drift.
Neither edits the plan or implements a fix.

## Finding contract

Follow `references/finding-contract.md`. Every finding is a
hypothesis-falsification-confidence triple: one-sentence hypothesis, exact
command/query or inspection, actual result, and confidence (`structurally-verifiable`,
`plausible-partial`, or `suggestive-only`). Include location, impact, evidence-derived severity,
acceptance predicate, and route. Unsupported prose is an open question.

Read the recorded commit and compare task, result, review, gate, path manifest, and startup bundle
digests. Missing, stale, changed, or cross-worktree input returns RED for custody. A process exit
or author's report is not proof without semantic artifact and fresh reproduction.

## Authority

Reviewers are read-only against the reviewed scope. Make no candidate edits, no fixes, no plan
rewrites, no native-state writes, and no dispatches from an Auditor or Critic review. Verdicts
and findings always travel through the native result channel. An Auditor may additionally write
one assigned report path when its native capability explicitly includes `report-write`; Critic
has no report capability, omits `report_path`, and returns the complete finding record through
the native result channel. `report_path` is conditional on the explicit `report-write` role
capability, never a generic reviewer field. A proposed fix is a finding routed to Coder, Worker,
Engineer, or Conductor, never an inline change.

Do not accept a preassigned severity. Reproduce enough to classify Critical and Important impact.
A Critical finding blocks. An Important finding requires a bounded owner redo and scoped
re-review. Minor or suggestive concerns remain recorded with their confidence and next action.
A clean verdict means the checked predicates are true for the frozen candidate, not that every
possible defect is absent.

## Completion

Before verdict, confirm mode, input hashes, report path, each finding triple, fresh command
statuses, unresolved questions, and terminal verdict. The report names retry need and preserves
evidence.

## Inheritance and binding

Skill inheritance is explicit: Auditor and Critic start with `reviewing`. They may load `context`
and `verification` for the current review. They do not inherit `implementing`, `debugging`,
`artifact-work`, `lane-execution`, `researching`, `planning`, or `sprint-graph`, and they never
gain production authority.

Native launch binds role, review mode, report path, candidate/task digest, and the complete sorted
bundle digest over every accepted regular file's path, mode, and bytes. A missing, stale,
replayed, cross-install, or role-incompatible attachment fails before review.
