pr-review.yml: upgrade review model to GLM-5.2 (MiniMax M3 fallback) and disable focus_only_on_problems

[uuta]

Background

On uuta/resorn PR #143, the reusable pr-review.yml workflow ran successfully but reported "No major issues detected" and "No code suggestions found", while Gemini Code Assist (reviewing the same commit fed341e1) caught a real defect: return (await lookup(id)).job; crashes on a malformed 200 response. The author fixed it in a follow-up commit (84e53a54).

Root-cause analysis of the run logs (run 29729720651) showed the miss was configuration, not tool failure:

1. pr_code_suggestions.focus_only_on_problems was effectively true, so the improve tool is prompted to report only critical bugs and suppress defensive-programming findings.
2. CONFIG.MODEL is a free-tier model (openrouter/nvidia/nemotron-3-ultra-550b-a55b:free), which has weaker recall than paid coding models. The model genuinely returned code_suggestions: [] — no rate limit or fallback occurred.

Gemini consumer code review is sunset as of 2026-07-17, so this workflow is becoming the only automated reviewer. Model research (OpenRouter pricing + mid-2026 coding benchmarks) selected z-ai/glm-5.2 as primary (same price band as kimi-k2.7-code, and the model behind the top Code-Review-Bench result) with minimax/minimax-m3 as fallback (frontier-adjacent coding at ~1/3 the price).

Implementation contract

In .github/workflows/pr-review.yml, in the "Review pull request with PR-Agent" step env block:

1. Change CONFIG.MODEL to openrouter/z-ai/glm-5.2
2. Change CONFIG.FALLBACK_MODELS to '["openrouter/minimax/minimax-m3"]'
3. Add pr_code_suggestions.focus_only_on_problems: "false"
4. Check the latest release of The-PR-Agent/pr-agent (gh api repos/The-PR-Agent/pr-agent/releases/latest). If it is newer than the currently pinned release, update the uses: line to pin the new release's tag commit SHA and update the trailing # vX.Y.Z comment to match. As of 2026-07-21 the latest release is v0.39.0 = 8e4d32e5497defd43c023a404f73560c62728961, which is already the pinned SHA — in that case leave the line untouched.
5. Do NOT add pr_code_suggestions.extra_instructions (deferred; initial phase)
6. Leave everything else unchanged: CONFIG.CUSTOM_MODEL_MAX_TOKENS, CONFIG.MAX_MODEL_TOKENS, auto_review/auto_describe/auto_improve flags, pr_actions, commitable_code_suggestions, the review-start timestamp step, and the "Verify PR-Agent published a review" step

Done when

• pr-review.yml on main contains exactly the config changes above (three env changes, plus the action pin bump only if a newer release exists) and no other diff
• actionlint (or equivalent YAML/workflow validation) passes on the modified file
• Blackbox: a test PR in a consumer repo (e.g. uuta/resorn) triggers the workflow; the run succeeds; the PR receives a "PR Reviewer Guide" comment; run logs show z-ai/glm-5.2 as the model actually called (not the fallback)

Not done if

• The pr-agent action is unpinned (tag or branch reference instead of a full commit SHA), or pinned to an untagged commit from main
• extra_instructions keys are added anywhere
• The verification step ("Verify PR-Agent published a review") is modified or removed
• Model IDs are written without the openrouter/ prefix (pr-agent's LiteLLM routing requires it)

[uuta]

Blackbox verification complete: after merging #3, the reusable workflow ran on five consumer PRs (uuta/resorn#145, uuta/SpeakRush#24, uuta/uuter#16, uuta/astro-blog#20, uuta/dotfiles#27). All five AI Review runs succeeded; resorn run 29825001965 logs confirm openrouter/z-ai/glm-5.2 was the model actually called (not the fallback), and the PR Reviewer Guide comment was posted.