Skip to main content

Post-review and light review

The review system audits both findings and fixes. Before acting, it tries to refute candidate findings and dedupe the finding set; after acting, it reviews the fix diff itself.

k-light-review uses the same core review engine with a smaller routing envelope for low-risk self-authored changes.

Mental model

PassRuns whenLooks atPurpose
Candidate refutation ladderbefore keeping any findingeach candidatetry to kill the claim before acting
Findings-set auditbefore fixing or draftingsurviving findings and proposed fixescollapse duplicate, verbose, unactionable, or overengineered feedback
Post-review stageafter a change-producing flow edits the working treethe fix diff, not the original diffcatch hygiene problems introduced by the fix
Light reviewbefore choosing full review for low-risk self-authored worklocal diff / branch-delta scoperun proportional depth without PR machinery

Using it

Post-review stage (verifying the review's own fixes)

Every change-producing flow ends with a post-review stage over the fix diff, not the original diff.

Applies to:

  • local-changes verify-and-fix.
  • PR-fix self-fixes.
  • self-review.
  • k-light-review.

The stage applies the canonical four dimensions defined verbatim in judging_core.md, never renamed:

DimensionMeaning
Redundancythe fix repeats something already present: re-implements a helper, re-states a rule, or adds an already-present path
Verbositythe fix is bloated beyond what the change needs: narration comments, ceremony, or over-explanation
Semantic + logical duplicationtwo places now express the same meaning/behavior via different text: parallel branches that should be one, or divergent-but-equivalent logic; this is the subtle axis literal-clone detectors miss
Gapsthe fix is incomplete: own stranded dead code, an unupdated co-edit-set member like a doc/diagram/census, a half-applied rename, or a referenced-but-missing file

Post-review behavior:

ContextBehavior
own work / self-reviewfix hygiene findings in the working tree and re-gate
reviewing otherssurface hygiene findings
read-only subagentsurface hygiene findings
trivial candidate setcontroller audits inline
non-trivial candidate setpost-review / findings-auditor runs the lens

Refuting and auditing findings (before acting)

Two engine passes run before fixing or drafting, distinct from the post-review stage that runs after.

Candidate refutation ladder runs before keeping any finding. The deciding agent tries to kill the finding in order: claim truth, reachability, severity, proposed fix, already-covered.

A candidate survives only when refutation fails with evidence. Every kept finding states reachability, and an unreachable path loses its severity.

Direct k-review/k-light-review run this as single-model self-refutation. In /k-agent-review, the cross-family adversarial lane owns it and read-only finder lanes only return candidates plus reachability.

Findings-set audit runs before acting on the survivors. The same four dimensions apply to the finding list and its proposed fixes, not the fix diff: collapse same-root-cause duplicates, trim verbose findings, and drop unanchored, unactionable, or overengineered items.

In /k-agent-review, this is the findings-auditor's job.

Reference: light review

k-light-review is a separate skill for proportional-depth, in-place audits of low-risk self-authored changes.

KeepsDrops
judging_core.md coverage checklist + trigger-based gatesmandatory SCSI/base-context preflight
four-dimension post-review lensGitHub machinery
candidate refutation ladder + findings-set auditmulti-agent fan-out + cross-family verification
opt-in base contextPR-thread/CI-specific rules

A light-eligibility predicate is evaluated first. It is the single source both the k-review router and change-auditor reference, replacing any subjective "is this low-risk?" call.

The change is light-eligible only when none of these escalate:

  • a PR exists for the branch.
  • authorship is not self, verified rather than assumed from a local checkout.
  • the diff touches security/auth/crypto/secret/migration/persisted-data/public-API paths.
  • the diff deletes or replaces/migrates code.
  • the diff is state-machine-like.
  • correctness needs base context beyond direct local reads.

Any trigger escalates to full k-review. The router applies the same predicate in reverse, offering k-light-review for a self-authored, no-PR, trigger-free diff.

change-auditor (Claude + Pi) is the read-only delegated form.

Both k-light-review and k-review's local-changes mode run the shared Verify-and-Fix Loop in judging_core.md: build queue → refutation ladder → findings-set audit → fix → quality gates → post-review stage.

Each mode only adds its own base-context stance and scaffolding on top.