Enhancement/136 Migrate the review flow to Opus 5 / Sonnet 5 - #135
Open
s100ian wants to merge 3 commits into
Open
Enhancement/136 Migrate the review flow to Opus 5 / Sonnet 5#135s100ian wants to merge 3 commits into
s100ian wants to merge 3 commits into
Conversation
Move the backend reviewer subagent to Opus 5 (deep bug-finding is the highest-value step in the review) and the CI review orchestrator to Sonnet 5 (diff classification, subagent dispatch, comment posting). Both replace claude-sonnet-4-6. No other API surface changes are needed: neither call site sets budget_tokens, sampling parameters, or assistant prefills, and effort: high remains valid on Opus 5. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01NkKxkLYPzNDTWUak4Wsb8D
Follow-up to the model migration, covering prompt-side and flow-side adjustments the model IDs alone did not address. - Reviewer: cap finding length at one to two sentences. Opus 5 writes longer user-facing text than Sonnet 4.6, and this output is posted verbatim as a PR comment. Verbosity responds to prompting, not effort. - Reviewer: effort high -> medium. Opus 5 holds review accuracy at lower effort; revert to high if recall drops on real PRs. - Skill/workflow: give the skill a real CI override and set CI=true on the job. Step 2 previously said "never skip this step" while the workflow prompt told the orchestrator to skip it — a direct contradiction that matters more now that both models follow instructions more literally. - Workflow: state the review-only guardrail (no commits, no edits, no new PRs), matching the uniplan-web workflow. - Reviewer: fix stale /review reference; the skill is /review-pr. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01NkKxkLYPzNDTWUak4Wsb8D
Reverts the medium step-down from the previous commit. Review recall is worth more here than the token saving; the other tuning changes (finding length cap, CI override, review-only guardrail) are unaffected. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01NkKxkLYPzNDTWUak4Wsb8D
s100ian
changed the title
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Closes #136
Migrates this repo's Claude model references off
claude-sonnet-4-6, plus the prompt- and flow-side fixes that the model swap alone doesn't cover.Companion to uni-dev-lab/uniplan-web#105, which makes the equivalent change on the frontend side.
What changed
.claude/agents/uniplan-be-reviewer.mdclaude-sonnet-4-6→claude-opus-5; added a finding-length cap; fixed a stale/reviewreference (the skill is/review-pr).claude/skills/review-pr/SKILL.md.github/workflows/claude-code-review.ymlclaude-sonnet-4-6→claude-sonnet-5; setCI: "true"; added the review-only guardrailWhy the model split. The reviewer subagent does the actual bug-finding, where Opus 5's precision and recall gain is worth the most. The CI
--modelflag drives the orchestrator, which classifies the diff, dispatches the subagent, merges findings, and posts the comment — plumbing that Sonnet 5 handles at lower cost. Both paths stay behind thereviewers-onlyenvironment gate, so token spend is bounded either way.Finding-length cap. Opus 5 writes longer user-facing text than Sonnet 4.6, and the reviewer's output is posted verbatim as a PR comment. The cap holds review length steady. It's prompting rather than an
effortchange because on Opus 5 verbosity responds to instructions, not to effort level.Skill/workflow contradiction.
SKILL.mdStep 2 previously said "never skip this step" while the workflow prompt told the orchestrator to skip worktree creation in CI — a direct conflict that matters more now that both models follow instructions more literally. The skill now carries its own CI override tripped byCI: "true"on the job, and Step 6 printsImplementation target: N/Ain CI instead of a worktree path it never created. This mirrors what uniplan-web already had.effort: highis retained. An intermediate commit stepped it down tomediumon the basis that Opus 5 holds review accuracy at lower effort; that was reverted, so the net diff leaveseffortuntouched. Recall was judged worth more than the token saving.Verification
No API-surface migration was required: the reviewer sets no
budget_tokens, no sampling parameters, and no assistant prefills, so nothing in the 4.6 → 5 breaking-change set applies. The workflow YAML was re-parsed after editing. No behaviour was executed — this diff is configuration and prompt text only, and the flow it configures runs onworkflow_dispatchagainst a PR.Not included
claude-opus-5uses a different tokenizer than Sonnet 4.6 and thinks by default, so a review run will spend more tokens than before. Worth watching on the first few real PRs; revert either model line if the cost profile doesn't hold.🤖 Generated with Claude Code
https://claude.ai/code/session_01NkKxkLYPzNDTWUak4Wsb8D