SuperAPI AI Directory

code-review

Coding & development

View repository

Updated 10-16 09:30

About this skill

Code review

One skill, clear modes. Model and orchestration policy lives in the shared AGENTS.md; runtime capability names live in ../RUNTIME_TOOLS.md.

Mode When Depth Primary model
closeout (default) Before commit/ship; "ready?"; autoreview; second opinion P0/P1 blockers first One reviewer selected by shared policy
axes User wants Standards vs Spec; "review since main"; issue/PRD fidelity Two independent reports Runtime-native agents; distinct model only when requested and available
full Large PR, multi-surface, security-critical, or /workflows-review Multi-agent P1–P3 Load workflows-review skill

How this differs from workflows-review

They are different jobs, not duplicates:

code-review workflows-review
Job Focused gate or two-axis check Exhaustive multi-agent audit
Agents 0–2 reviewers Bounded relevant specialist checks
Severity Closeout defaults to P0/P1 Always P1/P2/P3 synthesis
Outputs Summary + issues (or Standards/Spec) Full report; tracking files only when requested
Cost Cheap enough to run often Expensive; use when risk justifies
Pipeline Standalone closeout Part of workflows-* (after plan/work)

Rule of thumb: ship-readiness → closeout. "Did we build the right thing the right way?" → axes. "Throw the whole review bench at this PR" → full (workflows-review).

Models (all modes)

Follow the model roles and one lead reviewer in shared AGENTS.md. Use the active runtime's native reviewer by default. Select a different model only when the user requests a second opinion or shared policy calls for one and that model is available through its proper harness. Never describe a review as independent or model-specific unless that model was selected explicitly and the command proves it.

Mode selection

Parse the user request (and any args) in order:

  1. Explicit full / exhaustive / /workflows-review → full
  2. Explicit axes / standards / spec / review since <ref> → axes
  3. Explicit closeout / autoreview / ready to ship / before commit → closeout
  4. Empty or "review this" with a dirty tree or open PR → closeout
  5. Ambiguous large multi-package PR → ask once: closeout, axes, or full

Optional flags (if present in the message):

  • --max-priority P0|P1|P2|P3 — closeout severity ceiling (default P1)
  • --base <ref> — branch base (default origin/main or PR base)
  • --second-opinion — also run an available independent reviewer after the primary gate

Mode: closeout

Load and follow closeout.md in full:

  1. Freeze scope baseline
  2. Collect the right diff (local / branch / PR base)
  3. Run the reviewer selected by shared policy
  4. Verify → accept/reject → fix with implementer → re-test → re-gate
  5. Scope governor + two-cycle pause
  6. Final report (command, models, proof, clean/rejected)

Prose-only internal docs: skim, skip heavy gate.

Mode: axes

Load and follow axes.md in full. Standards and Spec may run as independent runtime-native subagents when available. After both return, present them side by side without cross-axis reranking.

Mode: full

Do not reimplement the multi-agent fan-out here. Load:

skill: workflows-review

Pass the same target/args. That skill owns agent discovery, skill slices, stakeholder/scenario passes, and todo/prd output. It must still obey the closeout scope governor and one lead reviewer rules (see closeout.md and AGENTS.md).

Invocation examples

/code-review                         → closeout on current changes
review before I ship                 → closeout
review since main                    → axes (fixed point = main)
code-review axes                     → axes
code-review full                     → workflows-review
codex review --uncommitted           → Codex closeout using configured model

Natural language that should hit this skill: "autoreview", "second model review", "is this ready to merge", "review my branch", "Standards and Spec".

Final rules

  • A review-only request stays read-only. When fixes are also authorized, only verified in-scope findings get implementer edits.
  • Do not push to create a review target.
  • Empty diff → report "nothing to review" and stop.
  • After material fixes, re-run the same mode until clean or scope stops you.