AI Pull-Request Reviewer with Re-Review and Audit

The problem

Real code review is the difference between a senior engineer and a feature factory. But on a small team, reviews pile up: nobody wants to be the bottleneck, the queue grows, and PRs eventually get merged with a thumbs-up emoji and zero feedback.

The interesting question: can an AI reviewer be useful enough to actually catch things, cheap enough to run on every push, and honest enough to know when to escalate to a human?

The approach

A reviewer that lives in the operations bot, not in CI:

Per-push trigger — when a PR's `head_sha` changes, the reviewer wakes up. If the new SHA matches the last reviewed SHA, it skips silently. No double-spend.

Cost-managed router — Claude Code CLI is the default (flat-cost subscription). The Anthropic API path is opt-in for low-latency cases and capped by a hard daily USD budget. When the budget hits zero, the reviewer queues for the next day instead of failing or silently overspending.

Severity classifier — every finding is rated (block / nit / info). Only blocks generate inline review comments; nits roll into a single summary. The PR doesn't drown.

Re-review on new commits — when the developer pushes again, the reviewer checks only the diff since the last review, not the whole PR. Fast, cheap and stays in context.

Audit + auto-resolve — when a PR is merged or rebased, the reviewer revisits its own comments. Comments that were addressed get resolved automatically; comments that were dismissed get marked dismissed in the audit log.

Human escalation — if the reviewer is genuinely unsure (security, schema migration, irreversible op), it pings the human on Telegram with the relevant lines and waits.

The outcome

The reviewer runs on every push to every PR across the project. False positives are rare enough to be tolerable; false negatives are caught on the second push. Daily Claude spend stays within budget, every cent accounted for.

The productized version (`code-review.tyber.io`, standalone outside the bot) makes the same engine available as a service to other operators.

Why it matters

Most "AI code review" tools are either too noisy to take seriously or too expensive to run on every push. The interesting design constraints are: cost discipline, per-push idempotence, severity gating, and respecting the human as the final escalation layer. Solve those and you get a reviewer that the team actually wants reading their code.