Skip to content

Latest commit

 

History

History

README.md

Reviews

Tyrne runs four kinds of reviews, each addressing a distinct concern. Each review type has its own folder with a master plan describing the multi-agent procedure and a log of dated review artifacts.

Reviews are event-triggered, not calendar-triggered. See each type's README for its specific triggers.

The four types

Type Folder Concern Trigger
Business business-reviews/ Milestone retrospectives and roadmap-level direction. What landed, what changed in the plan, what was learned. Milestone completion; maintainer-initiated.
Code code-reviews/ Per-change code quality: correctness, style, test coverage, documentation. Per non-trivial PR / change (once PR flow begins).
Security security-reviews/ Deep security pass on changes that touch capabilities, IPC, syscalls, memory, boot, crypto, unsafe, or security-sensitive dependencies. Per change that hits the triggers in ../../standards/security-review.md.
Performance optimization performance-optimization-reviews/ Hypothesis-driven performance cycles: baseline → hotspot → proposal → measurement → regression check. Periodic or on concern (user-reported slowness, scaling limits).

Per-type structure

Every review-type folder has the same three things:

  1. README.md — what this review type is, when to conduct one, how the output differs from a task artifact.
  2. master-plan.md — the procedure, written to be parallelizable across multiple AI agents. Enumerates agent roles, each role's checklist, and the merge step that combines their outputs.
  3. YYYY-MM-DD-<context>.md — one file per review conducted.

The master plan is a living document. It is amended as the project learns how to review itself better; the amendments are recorded inline.

Multi-agent review

Each master plan describes agent roles. For a single-maintainer solo phase, these roles can all be played by the same person (or the same AI) in sequence. The value of naming them separately is twofold:

  1. Each role is a focused checklist that is easier to execute without losing state than a monolithic "do a review" instruction.
  2. When AI tooling supports parallel agents, the roles compose directly: agent X does correctness, agent Y does style, agent Z does security-implications, and a merge step combines their sections.

The procedure does not require multiple agents; the single-agent sequential path is always valid. The multi-agent option exists for when it is efficient.

Producing a review

Use the conduct-review skill with the review type as an argument. The skill loads the type's master plan and walks through its steps.

For code and security reviews specifically, the execution skills perform-code-review and perform-security-review describe how to do the review during development; they produce the artifact in the corresponding folder as their final step.

See also