Code Review
Owns always-on review doctrine, finding standards, and report shape.
../solution-shaping/SKILL.md owns review timing, the binding objective, valid
slice shape, and whether a finding changes scope.
../verify-work/SKILL.md owns proof selection. Reviewers may reject invalid
proof; they do not replace runtime, design, security, or test owners.
Contract
Review is a blocking gate, not encouragement. A reviewer is expected to falsify
the parent story and find real defects.
Plans, work state, summaries, previous approvals, and parent claims are
hypotheses. Code, owner docs, tests, runtime topology, data shape, diffs, proof
artifacts, and repository state are evidence.
For non-trivial work, approval is binary: every reference selected by the
caller, role prompt, or handoff must be applied deeply enough for the claim. If
required context is skipped, uncertain, or impossible from available context,
return BLOCKED or BLOCK, not approval.
Valid verdicts:
APPROVE: the selected review scope passed, with no
unresolved current-scope findings.
BLOCK: the review found an actionable current-scope issue, invalid proof,
incomplete objective coverage, or unclassified material finding.
BLOCKED: required authority, context, changed surface, proof artifacts, or
repo access is missing, so the reviewer cannot honestly judge the claim.
Diff-only approval is invalid for non-trivial work. Plan-compliance approval is
invalid. Test-pass approval is invalid.
Always-On Review Setup
Before trusting the diff or parent story, independently reconstruct from repo
evidence:
- binding objective and user-accepted reductions
- expected finished repo state that would make the objective true
- atomic claims or objective coverage units in the reviewed scope
- owner/interface, lifecycle, state authority, failure policy, and proof route
- inspected authority: owner docs, code paths, runtime topology, data/state
shape, tests, changed surface, adjacent paths, and proof artifacts
- plan/note interpretation verdict
- known findings, finding dispositions, accepted temporary debt, and residual
work
Do not inherit the finished state from the diff, branch history, backlog
wording, PR status, previous completion claims, or parent handoff. An internally
consistent plan that implements the wrong shape is BLOCK.
Use ../solution-shaping/SKILL.md and its review lenses for non-trivial review.
Unassessed current-scope lens impact blocks approval.
Convergence
After independent reconstruction, compare against the parent plan,
implementation, proof, and final claim.
Convergence verdicts:
converged: objective, target shape, atomic claims, proof route, and cleanup
disposition match repo evidence.
converged with non-material deltas: wording or sequencing differs, but no
owner, proof, scope, cleanup, or objective coverage changes.
divergent: reviewer finds a materially different objective interpretation,
simpler target shape, atomic claim boundary, proof route, current-scope
disposition, or replan trigger. Verdict must be BLOCK.
blocked by missing authority: reviewer cannot inspect enough evidence to
reconstruct the target independently. Verdict must be BLOCKED.
Reconstruction depth is role-bounded:
- planning critic: objective, expected finished state, owner model, and atomic
claim boundaries
- quality guard: assigned atomic claim and enough lifecycle context to judge
implementation quality and plan validity
- final reviewer: whole objective coverage from completed claims, proof,
residuals, repo state, and artifact disposition
Simplicity And Fix-Now Review
Working behavior is not enough. Review whether the solution is simple,
owner-correct, and unified with established peer patterns.
Block unnecessary owners, wrappers, shims, flags, fallback paths, compatibility
paths, defensive branches, stale tests/docs, proof-only entrypoints, and
speculative abstractions unless a real owner and failure mode justify them.
Block missed simplification/unification when current work exposes an
inconsistency and applying an established peer pattern would make the touched
system simpler or more unified without adding capability, public surface,
optional behavior, compatibility path, or speculative abstraction.
Queue or require a parent decision instead of blocking as fix-now when the
finding adds product/runtime capability, introduces a new abstraction or
pattern, requires broad migration outside the touched owner path, has multiple
valid target patterns, or cannot be proved within the current or directly
adjacent atomic claim.
Finding Standard
Prefer fewer high-confidence findings over many weak comments. A finding is
valid only when it is concrete, actionable, and useful to the author.
Each finding must include:
- severity:
P0, P1, P2, or P3
- confidence:
high, medium, or low
- exact location:
file/path:line
- current-scope reason: why this blocks the reviewed objective or approved
scope
- impact: the failure mode, affected path/data/user, and when it happens
- required fix: the smallest owner-correct repair, not just a symptom patch
- disposition: fixed, current-scope blocker, separate debt, accepted temporary
debt, no-change, or blocked
Severity guide:
P0: must stop release immediately; critical security, data loss,
corruption, irreversible migration/deployment breakage, or broad outage.
P1: blocks the objective or normal-use correctness, security, reliability,
data integrity, migration, or recovery path.
P2: real current-scope bug or quality regression in a plausible edge path,
secondary workflow, admin/observability path, or maintainability surface.
P3: lower-severity current-scope cleanup or structural concern that should
be fixed before approval because it will make the code harder to own.
Unclassified material findings block approval. A NON-BLOCKING observation is
valid only for separate debt or explicitly accepted temporary debt.
Reference Index
Reference files are optional review lenses selected by the caller, role prompt,
or handoff before applying this skill. They are not activation rules.
references/patch-native-bug-hunt.md: changed-surface bug finding.
references/structural-quality.md: maintainability, abstraction, ownership,
and simplification review.
references/objective-proof-closeout.md: objective coverage, proof validity,
cleanup, repo health, and closeout review.
references/finding-disposition.md: current-scope, separate debt, accepted
temporary debt, and no-change classification.
Method
- Stay read-only.
- Use fresh context.
- Inspect the diff, changed files, adjacent owner/interface paths, tests, docs,
proof artifacts, generated artifacts, and active finding ledger needed to
judge the claim.
- Apply selected references only. Do not treat the reference index itself as an
activation rule.
- If no findings remain, state which bug classes and quality classes were
inspected and why the reviewed scope was sufficient.
Output
Report:
- verdict:
APPROVE, BLOCK, or BLOCKED
- convergence verdict
- binding objective and accepted reductions
- expected finished repo state
- reconstructed atomic claims or coverage units
- authority inspected
- reviewed scope and why it is sufficient or insufficient
- plan/note interpretation verdict
- selected references and verdict for each
- findings with exact location, impact, severity, confidence, required fix, and
disposition
- remaining required work or
none
Source: Vistyy/agent-harness — distributed by TomeVault.
1---2name: code-review-1493description: Review code or repository changes ruthlessly for bugs, quality regressions, objective fit, proof validity, maintainability, changed state, and final repo health. Use when the user asks for review or before claiming non-trivial work complete. Use this skill for always-on review doctrine, findings, and verdicts only; role prompts and handoffs own reference selection. Use when this capability is needed.4---56# Code Review78Owns always-on review doctrine, finding standards, and report shape.9`../solution-shaping/SKILL.md` owns review timing, the binding objective, valid10slice shape, and whether a finding changes scope.11`../verify-work/SKILL.md` owns proof selection. Reviewers may reject invalid12proof; they do not replace runtime, design, security, or test owners.1314## Contract1516Review is a blocking gate, not encouragement. A reviewer is expected to falsify17the parent story and find real defects.1819Plans, work state, summaries, previous approvals, and parent claims are20hypotheses. Code, owner docs, tests, runtime topology, data shape, diffs, proof21artifacts, and repository state are evidence.2223For non-trivial work, approval is binary: every reference selected by the24caller, role prompt, or handoff must be applied deeply enough for the claim. If25required context is skipped, uncertain, or impossible from available context,26return `BLOCKED` or `BLOCK`, not approval.2728Valid verdicts:2930- `APPROVE`: the selected review scope passed, with no31 unresolved current-scope findings.32- `BLOCK`: the review found an actionable current-scope issue, invalid proof,33 incomplete objective coverage, or unclassified material finding.34- `BLOCKED`: required authority, context, changed surface, proof artifacts, or35 repo access is missing, so the reviewer cannot honestly judge the claim.3637Diff-only approval is invalid for non-trivial work. Plan-compliance approval is38invalid. Test-pass approval is invalid.3940## Always-On Review Setup4142Before trusting the diff or parent story, independently reconstruct from repo43evidence:4445- binding objective and user-accepted reductions46- expected finished repo state that would make the objective true47- atomic claims or objective coverage units in the reviewed scope48- owner/interface, lifecycle, state authority, failure policy, and proof route49- inspected authority: owner docs, code paths, runtime topology, data/state50 shape, tests, changed surface, adjacent paths, and proof artifacts51- plan/note interpretation verdict52- known findings, finding dispositions, accepted temporary debt, and residual53 work5455Do not inherit the finished state from the diff, branch history, backlog56wording, PR status, previous completion claims, or parent handoff. An internally57consistent plan that implements the wrong shape is `BLOCK`.5859Use `../solution-shaping/SKILL.md` and its review lenses for non-trivial review.60Unassessed current-scope lens impact blocks approval.6162## Convergence6364After independent reconstruction, compare against the parent plan,65implementation, proof, and final claim.6667Convergence verdicts:6869- `converged`: objective, target shape, atomic claims, proof route, and cleanup70 disposition match repo evidence.71- `converged with non-material deltas`: wording or sequencing differs, but no72 owner, proof, scope, cleanup, or objective coverage changes.73- `divergent`: reviewer finds a materially different objective interpretation,74 simpler target shape, atomic claim boundary, proof route, current-scope75 disposition, or replan trigger. Verdict must be `BLOCK`.76- `blocked by missing authority`: reviewer cannot inspect enough evidence to77 reconstruct the target independently. Verdict must be `BLOCKED`.7879Reconstruction depth is role-bounded:8081- planning critic: objective, expected finished state, owner model, and atomic82 claim boundaries83- quality guard: assigned atomic claim and enough lifecycle context to judge84 implementation quality and plan validity85- final reviewer: whole objective coverage from completed claims, proof,86 residuals, repo state, and artifact disposition8788## Simplicity And Fix-Now Review8990Working behavior is not enough. Review whether the solution is simple,91owner-correct, and unified with established peer patterns.9293Block unnecessary owners, wrappers, shims, flags, fallback paths, compatibility94paths, defensive branches, stale tests/docs, proof-only entrypoints, and95speculative abstractions unless a real owner and failure mode justify them.9697Block missed simplification/unification when current work exposes an98inconsistency and applying an established peer pattern would make the touched99system simpler or more unified without adding capability, public surface,100optional behavior, compatibility path, or speculative abstraction.101102Queue or require a parent decision instead of blocking as fix-now when the103finding adds product/runtime capability, introduces a new abstraction or104pattern, requires broad migration outside the touched owner path, has multiple105valid target patterns, or cannot be proved within the current or directly106adjacent atomic claim.107108## Finding Standard109110Prefer fewer high-confidence findings over many weak comments. A finding is111valid only when it is concrete, actionable, and useful to the author.112113Each finding must include:114115- severity: `P0`, `P1`, `P2`, or `P3`116- confidence: `high`, `medium`, or `low`117- exact location: `file/path:line`118- current-scope reason: why this blocks the reviewed objective or approved119 scope120- impact: the failure mode, affected path/data/user, and when it happens121- required fix: the smallest owner-correct repair, not just a symptom patch122- disposition: fixed, current-scope blocker, separate debt, accepted temporary123 debt, no-change, or blocked124125Severity guide:126127- `P0`: must stop release immediately; critical security, data loss,128 corruption, irreversible migration/deployment breakage, or broad outage.129- `P1`: blocks the objective or normal-use correctness, security, reliability,130 data integrity, migration, or recovery path.131- `P2`: real current-scope bug or quality regression in a plausible edge path,132 secondary workflow, admin/observability path, or maintainability surface.133- `P3`: lower-severity current-scope cleanup or structural concern that should134 be fixed before approval because it will make the code harder to own.135136Unclassified material findings block approval. A `NON-BLOCKING` observation is137valid only for separate debt or explicitly accepted temporary debt.138139## Reference Index140141Reference files are optional review lenses selected by the caller, role prompt,142or handoff before applying this skill. They are not activation rules.143144- `references/patch-native-bug-hunt.md`: changed-surface bug finding.145- `references/structural-quality.md`: maintainability, abstraction, ownership,146 and simplification review.147- `references/objective-proof-closeout.md`: objective coverage, proof validity,148 cleanup, repo health, and closeout review.149- `references/finding-disposition.md`: current-scope, separate debt, accepted150 temporary debt, and no-change classification.151152## Method153154- Stay read-only.155- Use fresh context.156- Inspect the diff, changed files, adjacent owner/interface paths, tests, docs,157 proof artifacts, generated artifacts, and active finding ledger needed to158 judge the claim.159- Apply selected references only. Do not treat the reference index itself as an160 activation rule.161- If no findings remain, state which bug classes and quality classes were162 inspected and why the reviewed scope was sufficient.163164## Output165166Report:167168- verdict: `APPROVE`, `BLOCK`, or `BLOCKED`169- convergence verdict170- binding objective and accepted reductions171- expected finished repo state172- reconstructed atomic claims or coverage units173- authority inspected174- reviewed scope and why it is sufficient or insufficient175- plan/note interpretation verdict176- selected references and verdict for each177- findings with exact location, impact, severity, confidence, required fix, and178 disposition179- remaining required work or `none`180181---182> Source: [Vistyy/agent-harness](https://github.com/Vistyy/agent-harness) — distributed by [TomeVault](https://tomevault.io).183<!-- tomevault:4.0:skill_md:2026-06-23 -->