Skip to main content

​
Code Review Discipline

Most AI code reviewers fail not by reviewing too little, but by overreaching: an agent rewrites code it was asked only to review, counts legacy bugs as new problems, invents speculative risks to look thorough, or quietly fixes a cross-module issue that deserves a human decision. Myrm’s code-review skill makes the reviewer behave like a disciplined senior engineer:
  • It reviews. It doesn’t fix. Applying fixes is a separate step, done by you or a dedicated fix pipeline.
  • Every finding is attributed. Each issue is labeled Introduced-in-change, Pre-existing, or Unknown, backed by git diff / git blame evidence—never guessed.
  • Speculative risks are rejected. Findings must ground in the actual diff; unrealistic edge cases and hypothetical failures are dropped, not padded.
  • Out-of-scope blockers escalate to you. A real risk that crosses a module boundary or ownership line is reported and marked Needs user decision instead of being silently expanded into the change.

​
The four discipline layers

​
Multi-agent pipeline closure

When the review runs as a code-review-pipeline (analyzer → security reviewer → logic reviewer → verifier), the closure is explicit at the last mile:
  • The verifier role confirms fixes and runs regression tests, but does not auto-fix items marked Needs user decision—those are handed back to you.
  • The pipeline’s success criteria require out-of-scope blockers escalated to user decision, so a real risk can’t vanish into an agent’s fix list.
This matters because Myrm’s pipeline is agent-automated end-to-end. In competitor products the human is always in the loop at the PR interface, so they never needed to teach a verifier agent when to stop. Myrm must—and does.

​
Requirements traceability

Beyond finding defects, the skill maps every requirement from the task/PR description to a verification result: If no explicit requirements existed, it states the inferred purpose and confirms whether the code achieves it—so the review answers “did this change deliver what it promised?”, not just “is the code clean?”.

​
How to use it

Zero configuration. The skill is available as a prebuilt skill under Settings → Skills.
  • Review a change directly: enable the code-review skill and ask the agent to review a PR, a file, or a module.
  • Run a full review pipeline: use the code-review-pipeline skill (Kanban → Pipelines → Code Review) to fan out security + logic passes and end with fix verification and regression tests.

​
Verification & contract tests

The skill’s behavioral contract is guarded by contract tests that assert every discipline clause is present and the frontmatter stays valid, plus integration tests covering the full pipeline (instantiation → task graph → verifier task description) without mocks. Review pipeline parsing and the Needs user decision closure are covered end-to-end.