The review pipeline

A review is five ordered passes. Each one is priced before it runs, and the verdict is computed mechanically at the end rather than asked of a model.

The passes

1. Triage

One call on the triage model over the whole diff. It ranks the changed hunks by risk, assigns review lenses per file cluster, and may already report findings it can anchor. Content your configuration ignores never reaches this pass, and neither do generated files or whitespace-only churn.

2. Deep dives

One call per file cluster per active lens, on the deep model. A cluster that triage assigned security and correctness costs two calls. This is the pass that dominates the bill, which is why two separate dials bound it: the limits.deep_calls ceiling bounds the call count, and the budget bounds the money. Whichever binds first is disclosed in the review.

3. Cross-examination

Deep profile only. One call over the whole change that tries to defeat it as a whole rather than file by file: adversarial inputs, rollback safety, concurrency hazards, migration safety, and breaking interface changes. Findings from it are held to the same contract as every other pass.

4. Verdict synthesis

Findings are deduplicated and force ranked by the harm they argue. Anything missing a location or arguing no concrete harm is dropped here. The comment budget then decides how many are published inline.

The verdict itself is not a model output. It is computed from the severities that survive synthesis against your block_on list. A model cannot approve a pull request by saying so, and prompt injection cannot change the verdict because nothing reads a verdict out of model text.

The verdict model is asked for one thing only: a two sentence summary of the strongest case against merging. If that call fails, the review publishes without the paragraph. It never costs you the findings the deep dives paid for.

5. Publication

One review event with inline comments and suggestion blocks where a fix can be expressed, plus one check run whose conclusion is the verdict. Publication reports the verdict and never recomputes it, so the review and the check can never disagree.

Publication is atomic. Nothing is posted until synthesis completes, so a cancelled job costs tokens but never a partial verdict.

Profiles

Profile	Triage	Deep dives	Cross-examination
quick	yes	no	no
standard (default)	yes	yes	no
deep	yes	yes	yes

The finding contract

Every published finding carries:

• a file path and a line range that exist in the diff,
• exactly one severity: blocker, warning, or note,
• the concrete harm merging would cause,
• optionally a suggestion, rendered as a GitHub suggestion block.

Findings that fail any of the first three are dropped at synthesis rather than published with a hedge. This is deliberate: a review padded with unanchored concerns trains people to skim it.

Lenses

Lens	What it argues
correctness	The change does not do what it claims, or breaks something that worked.
security	The change exposes data, credentials, or an attack surface.
performance	The change costs time, memory, or queries that matter at real volume.
style	Off by default. Naming and form, only when you ask for it.

Lenses multiply deep dive calls. Enabling all four roughly quadruples the deep pass cost of a wide pull request, which is what limits.deep_calls is for.