Skip to content

refactor: simpler review config rules - #26

Merged
MauroToscano merged 1 commit into
feat/multi-model-reviewfrom
refactor/simple-review-config
Nov 17, 2026
Merged

MauroToscano merged 1 commit into
feat/multi-model-reviewfrom
refactor/simple-review-config

Conversation

@MauroToscano

Copy link
Copy Markdown
Contributor

Simplifies the repository config of the multi-model review (#24). This targets #24's branch, like #25.

Rules

Each changed file is handled by the first rule that matches:

  1. always_review: reviewed as code, whatever the rules below say (binary files and whitespace-only changes excepted). For text that drives behaviour, such as model prompts.
  2. ignore: not reviewed.
  3. Built in, not reviewed: lockfiles, binary files, generated and vendored code, whitespace-only changes (except in Python, shell, YAML and Makefiles), docs.
  4. Tests (built-in patterns plus test_globs): reviewers get the test file names and the test functions they change, not the diff.
  5. Everything else gets all five reviewers.

A part made only of low_risk code gets low_risk_lanes (default gpt, kimi, glm), in any PR, split or not, unless the cheap triage call finds that it touches keys, auth, funds, cryptography or untrusted input. Large PRs are reviewed in path order, low-risk parts last. If the budget runs out, the comment lists the rest and <command> continue budget=N reviews them.

Key Required Default
threat_model No threat_model.md
ignore No none
test_globs No none
always_review No none
low_risk No none
low_risk_lanes No gpt, kimi, glm

Removed

  • sensitive and default_sensitive: with the default, sensitive only set the review order. Its one other effect, reviewing listed Markdown files as code, is now always_review.
  • priority and the double per-reviewer budget for priority parts: the order only matters when a PR can't finish within its budget, and then continue reviews the rest. This also removes the mismatch where Opus was told 80 commands on a priority part while its limit allowed 160.
  • nonsensitive_lanes: renamed to low_risk_lanes. The default changes from claude, gpt, kimi to gpt, kimi, glm: Opus alone costs more than those three together (Nov 13 runs: Opus $1.60, GPT $0.36, Kimi $0.34, GLM $0.40 per PR).
  • unit_depth: fixed at 2.

The runner refuses a config that still uses a removed key and says what replaces it, so an old config can't silently stop doing what it says.

Other changes

  • Low-risk parts used to get fewer reviewers only in split PRs. They now do in any PR, so a PR that only changes demo apps gets three reviewers. Triage runs for those PRs too.
  • When a part's diff doesn't fit the prompt, files are ranked by size, low-risk files last, instead of by sensitive and priority.
  • This repository's config lists ai-review/prompts/**, .github/prompts/** and .github/ai-review/** in always_review.

Callers

yetanotherco/lambda_vm#1032 pins b265aff, which still accepts the old keys. Its config needs the new keys before it is re-pinned to a commit with this change; so does yetanotherco/waas#198.

Tests

24 offline tests pass (7 new):

  • always_review beats ignore, lockfile, generated, docs and test rules, but not binary or whitespace-only files;
  • low-risk parts are packed apart and get low_risk_lanes whether the PR is split or not;
  • removed keys are refused with their replacement;
  • run order is path order with low-risk parts last.

I also rendered the report's coverage and "who reviewed what" tables for a split PR with a low-risk part and a triage promotion. Nothing ran against a model.

Each changed file is now handled by the first rule that matches: always_review
(reviewed as code), ignore, the built-in skips (lockfiles, binaries, generated
and vendored code, whitespace-only changes, docs), tests, and everything else
gets every reviewer. A part made only of low_risk code gets low_risk_lanes
(default gpt, kimi, glm) in any PR, split or not, unless triage promotes it.
Large PRs are reviewed in path order, low-risk parts last; `continue
budget=N` reviews what the budget did not reach.

Removed: sensitive, default_sensitive, priority (and its double budget),
nonsensitive_lanes (now low_risk_lanes) and unit_depth. A config that still
uses one of them is refused with what replaces it.
@MauroToscano

Copy link
Copy Markdown
Contributor Author

/review

@MauroToscano
MauroToscano merged commit e60b8b1 into feat/multi-model-review Nov 17, 2026
1 check passed
@MauroToscano
MauroToscano deleted the refactor/simple-review-config branch November 17, 2026 18:22
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant