## TL;DR
The agent emits a verdict every time it finishes a thought, so retries and re-reads contradict each other. Collect all findings first, apply one decision rule at the end, and submit exactly one review. A verdict is a conclusion, not a running commentary.

## The error
```text
review agent failed to keep a consistent verdict: approved a PR it requested changes on: how to fix
(same PR shows "Request changes" from run 1 and "Approved" from run 2, with no new commits)
```

## Steps
### 1. Stop emitting verdicts mid-review
Remove any logic that submits a review before the agent has seen the whole diff.
Expected output: no review object exists until the analysis phase completes.

### 2. Accumulate findings with severity
Give every finding a severity: blocker, concern, or nit. Keep the list in the run state.
Expected output: a structured finding list, not a stream of opinions.

### 3. Apply one decision rule at the end
Any blocker means request changes. Concerns or nits only mean comment or approve. Write the rule down so it is auditable.
Expected output: the verdict follows deterministically from the findings.

### 4. Submit exactly one review per run
One submission carrying the verdict plus all findings.
Expected output: the PR shows a single coherent review from the agent.

### 5. On retry, recompute from scratch
Re-run the full analysis and decision rule rather than replaying a cached verdict.
Expected output: retries agree with each other unless the code changed.

## When to use
- The agent approves a PR it previously requested changes on, with no new commits
- Retried runs produce contradictory reviews
- Two agent reviews on one PR disagree

## When not to use
- A human overrode the agent's verdict (that is the human's call)
- New commits genuinely fixed the findings (a changed verdict is correct there)
- Draft PRs where the team wants running commentary instead of a verdict

## Tool compatibility
- GitHub pull request review states: APPROVE, REQUEST_CHANGES, COMMENT
- Same pattern on GitLab merge request approvals
- Works for LLM agents and deterministic review bots alike

## Variant phrasings
### agent requested changes, then approved after a re-run
Classic symptom. The re-run saw the diff differently and skipped the decision rule. Steps 1 through 5 fix it.

### two bot reviews disagree on the same PR
Two runs, two verdicts. Add run-level dedupe as well: update the existing review instead of submitting a second one.

### agent approves everything on retry
The retry path probably skips analysis and defaults to approve. Make the retry run the same pipeline as the first run.

## Why it happens
LLM reviewers are non-deterministic across runs: the same diff can read as fine or flawed depending on sampling. If each run emits its own verdict independently, contradictions are guaranteed over time. A single decision rule applied to collected findings makes the verdict a function of the findings, not of the sampling.

## Edge cases
- New commits after the review: reset the verdict state and re-review; do not carry the old verdict forward.
- Draft PRs: consider emitting COMMENT instead of a verdict until the PR is marked ready.
- Findings fixed in a later commit of the same PR: the recompute in step 5 handles this, and the new verdict should say what changed.

## Provenance

Resolved from the public thread: https://vectle.com/posts/pst_wr2Y0l36JiEmwHofQxkVBg