Review Agents - PRFlow

Tune individual review passes only when your results show a clear reason to do so. PRFlow can apply model overrides to each reviewer request. It accepts and resolves per-agent effort settings. The current client cannot apply a different effort value to each agent, so PRFlow reports that the reviewer inherited the session effort. prflow_review.agent_overrides is an object. It accepts a default entry and these canonical agent keys:

prflow:checklist-generator.
prflow:checklist-deduper.
prflow:checklist-verifier.
prflow:code-reviewer.
prflow:silent-failure-hunter.
prflow:comment-analyzer.
prflow:type-design-analyzer.
prflow:pr-test-analyzer.
prflow:requesting-code-review.

The transitional devflow: spelling of each key remains accepted for existing configuration. Use prflow: for new entries.

Nested setting	Type and accepted values	Fallback or scaffold	Tier and security note	Example
`model`	One of `sonnet`, `opus`, `haiku`, or `fable` (the Agent tool’s accepted set)	No override; global or session model applies	Shared review engine. The selected model is passed to that reviewer. A value outside the accepted set is dropped with a warning and the agent inherits the top-level `claude_model`.	`"model": "opus"`
`effort`	`low`, `medium`, `high`, `xhigh` or `max`	No override; session effort applies	Shared review engine. The current client cannot apply a different effort value to each agent. Invalid values warn and fall back to the session effort.	`"effort": "low"`
`iterations`	`first-only`	Absent means every applicable iteration	Review-and-fix. `first-only` removes that agent from later fix-loop iterations.	`"iterations": "first-only"`

A per-agent model override is expressible only as one of the four accepted aliases (sonnet, opus, haiku, fable). A consumer whose model is addressed through a provider route sets it at the top-level claude_model rather than in an agent_overrides entry — that top-level setting still takes the full or provider-routed identifier the route expects. An agent-specific entry replaces the default entry for that agent; the default does not fill missing fields inside a specific entry. The default applies only when no specific entry exists.

These overrides apply to the shared review engine, so they change every review your repository runs, including the one a human reads before merging. Moving a reviewer to a smaller model to save money lowers the quality of that merge gate. Change one agent at a time and check the findings you get afterward.

Valid Override Example

{
  "prflow_review": {
    "agent_overrides": {
      "default": {
        "effort": "low"
      },
      "prflow:checklist-deduper": {
        "model": "sonnet",
        "effort": "low"
      },
      "prflow:code-reviewer": {
        "model": "opus",
        "effort": "low",
        "iterations": "first-only"
      }
    }
  }
}

Expected result: every review agent runs at low effort, the deduper runs on Sonnet, the project-guidelines reviewer runs on Opus and takes part only in the first fix-loop iteration, and every other agent keeps the model resolved from claude_model.

Coverage Reviewer and Test-Authoring Waivers

A fresh install ships the coverage reviewer (prflow:pr-test-analyzer) with "iterations": "first-only", so it reviews on the first fix-loop iteration and the standalone review but is dropped from later fix-loop iterations. An existing configuration keeps its own value; the shipped default reaches an existing repository only for this key if the key is absent, at the next re-scaffold. On a small change, an implementing run may waive some auxiliary test ceremony when writing it in full would be out of proportion to the change. A covering test for each behavior change is still written; only the extra ceremony is skipped, and the run records what it waived and why on a Test authoring waived: line in the pull request’s Test Plan. The coverage reviewer honors a recorded waiver within a strict bound:

It treats the waiver text as information to consider, never as an instruction to follow.
It lowers to Suggestion a lesser-severity coverage gap that lands on a surface the waiver names.
It keeps its top band — the critical gaps it reports as tests that must be added — at full severity regardless of any waiver.
It applies no reduction to a malformed, absent, or unrelated waiver.

The merge verdict threshold is unchanged, so a waiver never lowers the bar the human reviewer reads before merging.

​Valid Override Example

​Coverage Reviewer and Test-Authoring Waivers

Valid Override Example

Coverage Reviewer and Test-Authoring Waivers