Downgrade review agent model from Opus to Sonnet

Open .fullsend/rhdh/harness/review.yaml and compare its settings with .fullsend/rhdh/harness/e2e-triage.yaml, especially the existing model override. Set the review harness model to Sonnet, then verify the configuration and monitor the next workspace-update reviews against the listed cost and quality criteria.

What happened

The review agent on PR #3490 used Opus (claude-opus-4-6) to review a 26-line mechanical diff (6 patch-level version bumps in metadata YAML files + a commit SHA update in source.json). The review workflow took 5 minutes and cost $0.72. The human reviewer (gabemontero) approved at 13:37:00 UTC — 3.5 minutes before the agent completed at 13:40:28 UTC — meaning the agent's review was not used as a decision input.

The review output was accurate and thorough (correct patch-level classification, OCI tag format verification, 12-item checklist), but every item it checked is mechanically verifiable. The real validation came from /publish (OCI image build) and smoke tests (plugin loading), both of which passed independently.

This repo generates 50+ workspace-update PRs per ~2 weeks via the daily update-plugins-repo-refs.yaml workflow. The vast majority are routine patch-level bumps like this one. At $0.72 per review, Opus reviews on workspace updates alone cost an estimated $36+/week.

What could go better

The review agent model (Opus) is disproportionate to the review complexity for this repo's dominant PR type. The review harness at .fullsend/rhdh/harness/review.yaml inherits the default model from the upstream base harness (fullsend-ai/agents) without overriding it, resulting in Opus for all reviews.

Sonnet is well-suited for workspace update verification: checking version bump levels (semver comparison), OCI tag format consistency (pattern matching), and file scope validation (path enumeration). These tasks do not require Opus-level reasoning. For the rare complex PR (major version bumps, new workspace additions, patch/overlay changes), Sonnet is still capable, and the human reviewer provides a safety net.

Repo-level model overrides are an established pattern — the e2e-triage agent in this repo already overrides to model: opus in its local harness config.

Confidence: High that Sonnet can handle routine workspace updates. Moderate confidence that Sonnet handles complex workspace updates adequately — the human reviewer backstops either way. The model can be switched back if review quality degrades on complex PRs.

Proposed change

Add model: sonnet to .fullsend/rhdh/harness/review.yaml:

base: https://raw.githubusercontent.com/fullsend-ai/agents/4bbe4f50ed8e33c60539eaa30ddc320edf8bcda0/harness/review.yaml#sha256=08456e534c7d1251b3a3b2bc1236643403809ecab27832c273764769f219a45b
image: ghcr.io/redhat-developer/rhdh-fullsend-code:latest
model: sonnet
policy: rhdh/policies/review.yaml
skills:
  - skills/e2e-failure-analysis
  - skills/playwright-trace

This follows the same override pattern used by the e2e-triage harness (model: opus in .fullsend/rhdh/harness/e2e-triage.yaml). The change reduces cost for all reviews while maintaining review quality for the predominantly mechanical workspace-update PR mix.

Validation criteria

1. The next 5 workspace-update PR reviews should cost under $0.20 each (vs $0.72 with Opus), representing at least a 3x cost reduction.
2. Review verdicts on workspace-update PRs should remain accurate — no false approvals on PRs that contain major version bumps, overlay changes, or scope creep beyond the target workspace.
3. If a complex PR (major version bump, new workspace, patch modification) is reviewed in the first 30 days, manually compare the Sonnet review quality against the Opus baseline from PR #3490 to confirm no material quality regression.
4. Monitor for human reviewers requesting changes that the agent missed — if the rate increases after the switch, re-evaluate the model choice.

Generated by retro agent from https://github.com/redhat-developer/rhdh-plugin-export-overlays/pull/3490