Closes archie-core-h019.5. The shipped contract already made only a confirmed
error-level finding block, and kept zero-findings distinct from not-run; what
was missing was the other half of calibration, and that is what a reviewer that
cries wolf needs.
- ReviewReport gains Checked []ReviewCheck {Property, Evidence}: the properties
the reviewer verified clean. Evidence is required, for the same reason a
finding requires a failure scenario -- a check without evidence is an
assertion, not a verification.
- The reviewer records checks through a typed record_checked tool, mirroring
record_finding, so it cannot end its turn having silently skipped the
checklist.
- Style/formatting/naming nits are stated out of contract, in the ReviewFinding
type doc and the reviewer prompt: they are the padding that trains an
operator to skim.
- The PR-body findings section renders the cleared properties, so a
zero-finding review arrives backed by what was checked rather than reading
the same as a review that never looked.
Design: docs/prds/review-calibration.md.
Implements archie-core-h019.3 (reviewer isolation) and h019.4 (wiring),
per docs/prds/adversarial-self-review.md.
- internal/domain/workflow: new Reviewer interface and StageReview,
inserted between StageDiffCap and StageOpenPR in the implement
workflow. Fail-closed: a review that didn't run, or a confirmed
error-level finding, parks the task (store.StatusParked) with
findings in Detail; warn/plausible findings never block. Inert
unless Repo.ReviewEnabled.
- internal/worktree: Manager.Snapshot exports HEAD's tracked files into
a fresh directory with no .git, for the reviewer's isolated
workspace. Trees interface gains Snapshot; hybridTrees delegates to
it.
- internal/app/agentworker: subagentReviewer implements Reviewer using
ai-sdk's agent.Subagent (nested, in-process, non-streaming
generation) -- conversation isolation is structural (Subagent.Run
never carries message history), workspace isolation is the snapshot
above. The reviewer's toolset is read-only (read/grep/find) rooted
at the snapshot, never the worker's own toolset, so it cannot reach
the worktreerpc publication grant. Findings are captured via a typed
record_finding tool appending to a slice the stage owns, read even
when Subagent.Run errors (a truncated run's captured findings
survive).
- internal/config: Repo.ReviewEnabled (default false) and
Models["reviewer"] (falls back to Models["builder"], no schema
change).
Judgment calls: Reviewer.Review never returns a Go error -- every
failure folds into ReviewReport per the PRD's fail-closed contract,
which the stage can act on uniformly. A Subagent.Run error with zero
captured findings resolves to ReviewStatusNotRun (park); a
Subagent.Run error with >=1 captured findings resolves to
ReviewStatusCompleted with those findings, since the PRD's own example
("three defects survive a step-limit truncation") requires them to
still be evaluated for blocking rather than discarded.