/repo-harness-cross-review

name: repo-harness-cross-review
description: Independent acceptance review through a persistent Herdr task session that runs the deep-reasoner role. Generic review owns scope, request-bound Result validation and typed Receipt.
when_to_use: "cross review, second opinion, outside voice, codex review, independent review, 让 codex 审, 找外部意见, 二审"

repo-harness-cross-review

Bot entrypoint. Canonical routing guidance for the existing generic review domain. Runtime ownership lives in `src/effects/review/generic-review.ts`, the OAR host and `src/effects/terminal/task-session.ts`; the Skill does not implement a provider.

Mode Selection

• Acceptance through a persistent Herdr task session: `references/generic-review.md`.
• Operator-attended acceptance in the Claude desktop app without Herdr:

`references/generic-review.md`, section "App terminal launch".

• Explicit harness selection belongs to `rh review round --harness`.

Boundaries

Attach the CI status of the head, the recorded evidence and the contract before review (`repo-harness-issue-flow`, `references/acceptor.md` §2.1). The task session owns the run: the addressed hc terminal or app-terminal host, the isolated reviewer checkout and the OAR Session. The reviewer takes its instructions from the `deep-reasoner` role in `agents/fleet/`. History/idle are observations; the request-bound file Result and generic-review Receipt own acceptance. Unknown model, delivery or cleanup stays unverified or pending. Never synthesize PASS, alias the retired command, replay an ambiguous request or bypass a frozen reviewer policy. Review does not merge; the owner makes that separate decision. Use only the documented admitted runtime/isolation boundary; library permission approval is not an OS sandbox.