Skip to content

ci: switch /ai-review to the shared multi-model review - #1032

Open
MauroToscano wants to merge 5 commits into
mainfrom
ci/ai-review-multi
Open

MauroToscano wants to merge 5 commits into
mainfrom
ci/ai-review-multi

Conversation

@MauroToscano

@MauroToscano MauroToscano commented Nov 13, 2026 •

Copy link
Copy Markdown
Contributor

What

/ai-review (and the ai-review* labels) now call the shared multi-model review, pr_review_multi.yml from yetanotherco/actions (yetanotherco/actions#24), pinned at that PR's latest commit, b265aff.

  • Reviewers. Opus, GPT, Kimi, GLM and DeepSeek review the PR as read-only agents.
  • Judge. A judge checks the findings against the code.
  • One comment. It holds a table of findings, collapsed sections, housekeeping notes (stale docs, missing tests) and the review's cost.
  • Partial reviews. /ai-review continue finishes a review that ran out of budget, and /ai-review budget=N raises the budget.

The review is steered by .github/ai-review/config.toml (paths, review order) and threat_model.md (what matters in a zkVM). Both are read from the PR's base branch.

Removed. The opencode swarm: .github/scripts/ai_review.py, its tests and test workflow, the matrix and prompts, and .opencode/. docs/ai-review.md is rewritten and lists the changes from the previous review.

What needs to be done

The full checklist is in yetanotherco/actions#24. For this repository:

1. Before the first run

  • The org secret OPENROUTER_API_KEY is a key from the OpenRouter account with the air-* presets, with a credit limit.
  • The ANTHROPIC_API_KEY secret here belongs to an Anthropic workspace with a monthly spend limit. Opus is billed there.
  • The OpenRouter account is topped up.

2. First run

  • Add the ai-review label to this PR. Label events run this PR's version of the workflow.
  • In the job log, check that the sandbox check passes (sandbox check: lane commands run, ...). If it fails, the review stops before any paid call, and the comment says why.
  • Check the posted comment: table of findings, judge's count, cost line.
  • main has no .github/ai-review/config.toml yet, so this run uses the defaults: every file sensitive, no threat model. It reviews only CI files, so the findings themselves don't matter; the run tests the pipeline.

3. Merge

  • After feat: multi-model PR review workflow actions#24 is merged and released, re-pin uses: in .github/workflows/pr_ai_review.yaml from the branch commit (b265aff) to the release tag's commit.
  • Merge this PR. /ai-review comments work from then on, because comment triggers run main's workflow. The config and threat model in .github/ai-review/ apply to every review after that.

Unrelated failing check

"Spec tests" fails here because spec/tooling/chip.py breaks on the newer Python on the runners: it reads __annotations__ from an instance. That workflow last ran on main in June, and this PR doesn't touch spec/.

/ai-review (and the ai-review* labels) now call pr_review_multi.yml from
yetanotherco/actions: Opus, GPT, Kimi, GLM and DeepSeek review the PR as
read-only agents, a judge checks every finding against the code, and one
comment is posted with the findings, housekeeping notes and the cost.
'/ai-review continue' finishes a budget-limited review and
'/ai-review budget=N' raises the budget.

The repository's review config (.github/ai-review/config.toml and
threat_model.md) replaces the opencode swarm's matrix and prompts, which
are removed with its runner, tests and opencode config.
Pin pr_review_multi.yml to the commit with the review fixes. The docs
now give the enforced soft budgets, say the config is read from the PR's
base branch, that Opus is billed by Anthropic outside the OpenRouter
limit, and list the changes from the previous review. The config's
priority list is narrowed to the soundness-critical code, with the VM
and syscalls no longer reviewed last.
Pins yetanotherco/actions b265aff, which adds #25 (files lost when a split area was chunked) and
the fixes from a second review: failed reviewers and changed files no reviewer saw are listed in
the comment, whitespace changes in Python, shell, YAML and Makefiles are reviewed, and odd file
names or Rust test modules can no longer hide code. docs/ai-review.md says what changed for
this repository.
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