Code review that keeps up with the code.
Taking an open-source review extension further: follow-up reviews, token-aware prioritisation and feedback that stays useful across revisions.
async function publish(plan) { await savePlan(plan); await notifyForecast(plan);↳}The challenge.
A model can comment on a diff. The harder problem is making those comments useful inside a real review process: changes arrive in iterations, context is limited and repeated feedback creates noise.
What I built.
I extended an open-source Azure DevOps review extension with follow-up reviews across PR iterations, priority-ordered line comments, structured model output and a fixture-based local test harness. This work builds on the original extension rather than claiming its entire implementation as my own.
- ↗︎Follow-up reviews across PR iterations
- ↗︎Deduplicated, line-level comments
- ↗︎Fixture-based local test harness
The decisions behind it.
Spend context where it matters
Priority-ordered review allocates a bounded token budget to the most relevant changes. It makes context limits an explicit part of the workflow.
Treat model output as input
Structured JSON, parsing fallbacks and line-level mapping turn a model response into reviewable comments. Output still needs validation before it reaches a pull request.
Remember the review state
Hidden metadata supports idempotent comments and deduplication across revisions. A follow-up review can focus on what changed instead of repeating itself.
What matters most.
The integration around the model determines whether an AI feature is useful. Context selection, validation and state management turn a one-off generation into a repeatable engineering tool.
This case study describes my extension work. The displayed code review is a synthetic illustration, not a customer pull request or a live model response.
If notification fails after the save, how does the caller safely retry?
RECOVERY · LINE 20 · SYNTHETIC EXAMPLE