Skip to content

t1k:review

FieldValue
Modulet1k-base
Version5.15.36
Efforthigh
ToolsAgent, AskUserQuestion, Bash, Glob, Grep, Read, Skill, Task

Keywords: adversarial, audit, coverage, pr, quality, red-team, review

/t1k:review
[#PR | COMMIT | --pending | codebase [parallel] | adversarial] [--single|--lanes] [--advice|--no-advice]

Adversarial code review with technical rigor, evidence-based claims, and verification over performative responses. Every review includes red-team analysis that actively tries to break the code.

Follow protocol: skills/t1k-cook/references/routing-protocol.md This command uses role: reviewer

InputModeWhat Gets Reviewed
#123 or PR URLPRFull PR diff fetched via gh pr diff
abc1234 (7+ hex chars)CommitSingle commit diff via git show
--pendingPendingStaged + unstaged changes via git diff
(no args, recent changes)DefaultRecent changes in context
codebaseCodebaseFull codebase scan
codebase parallelCodebase+Parallel multi-reviewer audit

If invoked WITHOUT arguments and no recent changes, use AskUserQuestion — details: references/input-mode-resolution.md

YAGNI, KISS, DRY always. Technical correctness over social comfort. Verify before implementing. Ask before assuming. Evidence before claims.

Supervision runs at advisorySupervision.level (default forks) without being asked for. Opt-in even at forks — this skill’s default path already runs the t1k-review-synthesizer (sonnet) (and --single runs t1k-code-reviewer, routed Sonnet 5.5), so supervision here is doubled review. --advice raises this run to full; --no-advice turns it off and forward-carries the opt-out. Contract: skills/t1k-cook/references/advisory-supervision.md.

Full contract: skills/t1k-cook/references/advisory-supervision.md — invocation, forward-carry, PR gate, empty-counsel fallback, and the “never bypasses a gate” clause. Read it before firing the first checkpoint.

Composes with every input mode. In PR / Codebase modes the checkpoints fire per PR / per review target, not once per run.

Review-specific checkpoints, on top of the universal ones (the after-each-phase one fires only at level full):

  • After the initial review completes — pass the target ref, the diff summary, the findings list with severities, and the tentative verdict; ask for a go/no-go on the verdict and for findings that were missed.
  • After red-team analysis — pass the attack surface examined and what survived; ask which failure mode was not tried.
  • Before the review lands anywhere durable (a PR comment, a report file) — pass the final review body and ask for a sanity check on tone, evidence, and severity assignment. Revise the body if counsel flags a Critical/Important problem with it.

The verdict stays this skill’s. Counsel informs the write-up and the severities; it is not a veto in either direction — it can neither downgrade a Critical finding nor manufacture one. Evidence rules under Practices remain authoritative.

Follow protocol: skills/t1k-cook/references/activation-protocol.md

PracticeWhenReference
Spec complianceAfter implementing from plan/spec, BEFORE quality reviewreferences/spec-compliance-review.md
Adversarial reviewAlways-on Stage 3 (except --lanes: synthesizer-owned) — actively tries to break the codereferences/adversarial-review.md
Receiving feedbackUnclear feedback, external reviewers, needs prioritizationreferences/code-review-reception.md
Requesting reviewAfter tasks, before merge, stuck on problemreferences/requesting-code-review.md
Verification gatesBefore any completion claim, commit, PRreferences/verification-before-completion.md
Edge case scoutingAfter implementation, before reviewreferences/edge-case-scouting.md
Checklist reviewPre-landing, /t1k:ship pipeline, security auditreferences/checklist-workflow.md
Task-managed reviewsMulti-file features (3+ files), parallel reviewers, fix cyclesreferences/task-management-reviews.md
Skill review (auto)Diff includes .claude/skills/*/SKILL.md or .claude/skills/*/references/*.mdinvoke t1k-skill-creator (owns Skillmark + decision-tree + line-cap + body-tightness §K)
Agent review (auto)Diff includes .claude/agents/*.mdinvoke t1k-agent-creator (owns canonical agent frontmatter + model tier)

Skill-body tightness check (auto, when SKILL.md is in diff): flag any added line matching incident-marker patterns (dates, PR refs, Originating incident:, Verified failure, commit hashes, Real-world miss). Recommend moving to references/. Rationale + full pattern list: skills/t1k-skill-creator/references/architecture-rules.md §K. CI gate is validate-skill-body-tightness.cjs — surface gate output in review summary.

Review Depth — the Model Decides (none / single / lanes)

Section titled “Review Depth — the Model Decides (none / single / lanes)”

Depth is a model decision made before every review, not a fixed default. Before spawning anything, the orchestrator gathers facts about the change — git diff --numstat (files touched, +/- lines), which paths are touched (hooks, settings*.json, anything security-sensitive, .github/workflows/**, module.json, or a file shared across modules), and whether a PR/CI run already exists for it — then picks exactly one depth and records it as a single line: review: <none|single|lanes> (<reason>).

  • none — nothing worth reviewing (a trivial docs/comment fix, a version-bump-only diff). Only when /t1k:review runs as a step inside a larger workflow (cook, ship, a checkpoint) — never when the user explicitly asked for a review; an explicit ask always gets at least single.
  • single — the ordinary posture for a contained, single-concern change: the three-stage protocol below, one reviewer agent.
  • lanes — consider it when the change is substantial or carries above-average risk: touches several unrelated areas at once, crosses a security/trust boundary, changes a hook/settings/CI/module contract, or is large/cross-cutting enough that one reviewer plausibly misses something a narrower specialist would catch. Never a fixed line-count or file-count threshold — judge the change, not a number.

An explicit --single or --lanes on the invocation always overrides the model’s choice regardless of what the facts suggest. Full resolution recipe (project scope wins if it names either flag as a standing override, else global, else the model decides): references/lane-review-workflow.md § 0.

codebase and codebase parallel never compose with lanes — see the Input Modes table; those stay on the non-lanes path always (moving to P11’s structural-audit agents is a separate phase, not this decision).

Three-Stage Review Protocol (single depth)

Section titled “Three-Stage Review Protocol (single depth)”

Runs whenever single depth is chosen — by explicit --single, by the model’s own decision, or by a lanes-preconditions fallback (saying so):

Stage 1 — Spec Compliance → references/spec-compliance-review.md Stage 2 — Code Quality (registry-routed reviewer agent) — runs AFTER Stage 1 passes. The spawn’s description or prompt MUST include the literal line [t1k:review --single] (every route into single depth converges on this same stage) — it is the stable marker generic-agent-detector.cjs’s [t1k:review-lanes-default] nudge reads to tell “this IS the single-reviewer path /t1k:review chose or fell back to” apart from “this bypassed /t1k:review entirely”, so it stays silent here instead of nudging back toward a direct spawn. Stage 3 — Adversarial Review → references/adversarial-review.md — ALWAYS-ON under single depth (skipped under lanes: synthesizer-owned)

Full decision tree and workflows: references/review-workflows.md

lanes depth (explicit --lanes, or the model’s own choice per the guidance above) runs from the MAIN session only — in-process teammates cannot spawn background agents; on that condition (or a missing lane agent), fall back to single depth and say so. With lanes chosen in PR mode: assert the PR number is all digits, then git fetch origin +pull/N/head:review/N (leading + so a re-review after a force-push cannot resolve stale SHAs; non-zero fetch exit = stop) so hops need no gh auth; preconditions (the four perspective t1k-review-* agent files plus t1k-review-synthesizer resolve, else fall back to single depth and say so) and ref resolution: references/lane-review-workflow.md § 0–§ 1. Then the four perspective lanes — t1k-review-correctness|security|tests|contracts — plus a 5th domain lane when a reviewChecklist routing-fragment key resolves (t1k-list-classifier, haiku, rubric = the checklist file; skew guard when the reviewer role resolves to a kit agent with no reviewChecklist: rubric = that kit reviewer’s own agent body, so an old kit never loses its domain checks) — are spawned in ONE message with run_in_background UNSET (never pass false); results arrive as task notifications, and the wait for all report files (4 or 5) is unconditional (§ 4 of the reference). Each brief is ≤4,000 chars (repo path + BASE/HEAD SHAs + report id + lane contract, never the diff). Then ONE t1k-review-synthesizer spawn merges them. The main session reads only VERDICT.md, then hands the synthesizer’s unpushed commit to t1k-git-manager (§ 6 push hand-off). Stage 3 is skipped under lanes depth: the synthesizer owns the adversarial pass. The kit’s opus reviewer agent (if any) is NOT spawned on the lanes path — it stays reserved for design review and single depth.

Fact-gathering, decision recording, brief template, collect step, domain lane, failure handling, rollback: references/lane-review-workflow.md

PR mode is for ANOTHER person’s PR (a PR the user or their sessions open gets no post-open review; arm --auto at open). Lane choice comes from data and starts in parallel before CI, with scoped re-review: references/review-at-pr-open.md (table: references/review-at-pr-open.json).

Comment-only — no state transition. Full contract: skills/t1k-plane/references/workflow-enforcement.md (../t1k-plane/references/workflow-enforcement.md) § Stage 3. Governing rule: modules/t1k-plane/rules/plane-workitem-workflow.md.

After Stage 3 completes, post the verdict and finding count to every work item bound to this session. Critical findings do NOT move the item backwards. If the project defines an extra started-group state named In Review / Review / QA, move there instead of commenting alone; never create such a state to satisfy this. Plane enforcement now blocks, not warns, on an unbound repo or absent plane server — see the cited Stage 0 contract. Escapes: T1K_PLANE_MODE=advisory|off, --no-plane, recorded optout.

  • No hardcoded values
  • Error handling present
  • No unnecessary complexity (YAGNI/KISS)
  • No duplication (DRY)
  • Security: no secrets, credentials, or sensitive data
  • Tests present for new functionality

Project-type checklists: references/checklists/base.md, references/checklists/api.md, references/checklists/web-app.md

Follow protocol: skills/t1k-cook/references/subagent-injection-protocol.md

Sub-agent forking: see skills/t1k-architecture/references/fork-hygiene.md.