review-change

Review a change systematically

The reviewer judges a change, not the whole tree. The auditor productizes this project's F1–F13 whole-tree regression detectors; the reviewer is orthogonal — it reads this diff vs HEAD against the things you only learn by looking at what the edit did: the invariants and risk map the explorer recorded, the callers that must still resolve, and the scope the plan approved. Never re-source the auditor's lib at runtime (vendoring rule); the release gate aggregates both.

Two layers:

1. Deterministic checks (scripts/lib-review-checks.sh, pure shell/awk, no python) — the floor. Diff source is git diff HEAD. Each check is individually callable and SILENT on a clean / in-spec change, so a finding is trustworthy.
2. Review sub-agents (reviewer-scout breadth / reviewer-critical depth) — judgment on top: does the change violate a recorded invariant, mishandle a risk-mapped area, or drift a producer/consumer contract? Things static analysis cannot see.

Finding line format (BOTH layers emit this)

One finding per line, TAB-separated:

<SEVERITY>\t<check>\t<file:line-or-path>\t<message>

SEVERITY ∈ BLOCKING | CONCERN | NOTE (see the check-invariants skill). verify-review.sh tallies BLOCKING/CONCERN (NOTE excluded), writes .claude/reviewer/REVIEW.md, and records bd_status_write reviewer review <state> "" blocking=$B concern=$C. The pipeline release gate (verify-release.sh) reads reviewer blocking and requires it to be 0 to release.

The deterministic check map (what each reviews, and why)

BLOCKING — the change breaks something concrete (these feed the 0-blocking gate):

• R1 caller-integrity — a shell function REMOVED or RENAMED in the diff that still has surviving call-sites in the tree. A rename surfaces as the old name removed; if the name is no longer defined anywhere but is still called, that caller is broken. (A function merely moved / edited in place — still defined somewhere — is not flagged.)
• R2 convention-regression — a changed .sh that, versus its HEAD blob, drops set -uo pipefail, stops sourcing ../lib/common.sh, or (re)introduces an errexit set -e. Each regresses the house safety preamble that keeps the gates from failing open.

CONCERN — the change warrants a human look but does not hard-block:

• R3 risk-touch — a changed file whose path is named in the explorer MEMORY.md Risk map (- <area> — <risk> — <severity> — <evidence>). Touching recorded-risk surface → re-read the risk.
• R4 scope-discipline — a changed file NOT listed in the approved .claude/builder/PLAN.md Scope (same parser/membership test as the builder scope guard). Possible scope creep.

NOTE — informational only, NEVER gates, EXCLUDED from the tally (a stylistic observation, or an agent's context note that should travel with the review without affecting the verdict).

Sub-agent method (scout + critical)

• Recall, don't re-scan. Read .claude/explorer/MEMORY.md (architecture, invariants, risk map) for ground truth, then read the diff under review (git diff HEAD, or the base ref the orchestrator passed).
• Reason about: broken invariants the explorer recorded, risk-mapped areas the change touches, producer/consumer contract drift (STATUS keys, gate inputs), dangling callers the static check could not resolve, and scope creep.
• Cite path:line. Separate evidence from inference (a hypothesis is labelled, never a BLOCKING).
• Append each finding to .claude/reviewer/findings/<agent>.tsv in the line format above.
• Do not duplicate a deterministic check's finding; add what it cannot see.

Keep findings dense and evidence-first. A BLOCKING must be a real, defensible breakage — the gate trusts it.