CtrlK
BlogDocsLog inGet started
Tessl Logo
SkillPublic

review-correctness

Review one code change for correctness and rate it Low, Medium, or High. Looks for bugs, regressions in existing behavior, broken flows, concurrency and retry problems, and migrations that can fail or lose data. Use when asked to "review this PR for bugs", "check this diff for regressions", "is this change correct", or when a review loop runs the correctness lens. Read-only; it reports findings and does not fix them.

Quality

89%

Does it follow best practices?

Run evals on this skill

Adds up to 20 points to the overall score

View guide
SecuritybySnyk

Passed

No findings from the security scan

SKILL.md
Quality
Evals
Security

Review a change for correctness

You review one change and report what it breaks or can break, and how much it matters. You do not edit code, and you do not post anything.

Inputs

The caller gives you these. If the caller leaves one out, work it out and say what you used.

  • Checkout path. A checkout of the repo at the head commit. Read code there.
  • Base. The commit the change starts from. For a PR, use the merge base of the PR head and its base branch, never the current tip of the base branch.
  • Head. The last commit of the change.
  • Context file (optional). Decisions that are already settled: findings that were fixed, rejected with a reason, or deferred, and product rules from the owner. Do not raise a settled point again unless the fix is wrong or incomplete. Then say why.

Read the change with git -C <checkout> diff <base> <head>. Add -- <path> to read one file.

What to look for

Read lens.md in this skill's directory and apply it.

Rules

  • Read-only. Do not install dependencies, and do not run builds, type checks, linters, or tests. Judge tests by reading them next to the code they test.
  • Judge the final diff, not the commit history.
  • Do not charge the change for problems in code it does not touch, unless the change makes an existing problem reachable.
  • Raise a finding only with a concrete trigger you can reach. Say which input, state, or timing reaches it. Mark a finding you could not confirm as a hypothesis, and say what would confirm it.

Output format

Write the report in plain English, in exactly this shape:

# Correctness review

Base: <base sha>
Head: <head sha>

## Findings

### 1. <one-line title>
- Severity: High | Medium | Low
- Where: <file>:<line> (more locations if needed)
- What is wrong: <the wrong behavior, the trigger that reaches it, and who sees it>
- Smallest fix: <the smallest change that removes the problem>

(more findings, most severe first; write "No findings." if there are none)

## Rationale

<the two strongest reasons for the rating, as concrete behavior problems>

Rating: Low | Medium | High

The last line must be Rating: <value> with nothing after it, so a script can read it.

Repository

Agenta-AI/agenta

Path

.agents/skills/review-correctness/SKILL.md

Last updated
First committed

Is this your skill?

If you maintain this skill, you can claim it as your own. Once claimed, you can manage eval scenarios, bundle related skills, attach documentation or rules, and ensure cross-agent compatibility.