Skip to content

AI Code Reviewer implementation

Ticket #267: AI Code Reviewer implementation
Type: Quality / Governance / Automation
Affected Component: .github/agents/, .specify/scripts/powershell/, src/pr_reviewer/, tests/pr_reviewer/, specs/009-ai-code-reviewer/

1. Context

Following my brainstorming session on an agentic SDLC, I made a key decision: to retain Speckit as the sole agentic system for the structural development of this application. In line with this, I undertook the creation of an AI agent dedicated to code review, designed according to the Speckit approach, in order to integrate this quality assurance step naturally and consistently into my current production process.

2. Objective

Set up an AI review agent that helps make a clear decision:

  • either the Pull Request can be approved,
  • or it must be corrected before merge,
  • with an explanation that is easy to read and act on.

3. What was delivered

  • A structured review flow centered on the /speckit.review agent.
  • A systematic assessment across 5 axes:
    • compliance with the project constitution,
    • security risk level,
    • architecture alignment,
    • test coverage,
    • adherence to planned scope.
  • A complete locally printed report with a clear recommendation: approval when risk is acceptable, or required corrections when a critical gap is detected.
  • No automatic GitHub review, comment, approval or change request is published; the maintainer keeps the final decision.

4. Business impact

  • Lower regression risk before production release.
  • More reliable merge decisions through one consistent quality standard.
  • Coordination time savings thanks to a pre-structured review outcome.
  • Controlled decision-making: the report informs the maintainer without automatically changing the PR.

5. Validation and status

  • Review analysis and local printing modules were tested.
  • Core behavior is validated: classification, complete report formatting and absence of GitHub writes.
  • Linked traceability issue is created: #267.

6. Lessons learned

  1. A simple and stable decision framework speeds up merge arbitration.
  2. The main value of AI here is consistency of analysis, not complexity.
  3. A complete local recommendation avoids impossible attempts to review a pull request owned by the same GitHub account.