ci: upgrade pr-reviewer-action to v1.2.0 (routing + escalation)

[joryirving]

Upgrades the AI PR review workflow to pr-reviewer-action v1.2.0 and adopts the new routing/reliability features.

Repo character: application code — routing sends routine changes to the fast model; auth/route/secret-touching changes go to the smart model.

What changed and why

• Pin bumped to the v1.2.0 release SHA (ff03870). Feature-heavy release: model routing + escalation, structured findings, required-check validation, incremental-review and CI-wait fixes, large context/perf work. All backward-compatible.
• review_routing_mode: auto — PRs are classified deterministically; low-risk ones (e.g. routine dependency bumps) go to the fast model, risk-flagged ones (auth, secrets, migrations, public routes, linked security issues) go straight to the smart model. Fast stays your current PRIMARY_MODEL. Smart is wired as vars.SMART_MODEL || vars.FALLBACK_MODEL — no new vars are required; it uses your fallback model until you optionally add a SMART_MODEL (and SMART_FORMAT if it differs) repo variable pointing at a bigger model.
• Escalation (on by default with routing) — if the fast review looks insufficient (request_changes, unaddressed required checks, suspiciously short/unsure review, blocker signals), the review automatically re-runs on the smart model and only that result is published.
• on_model_failure: notice + retries 8 → 3 — when the homelab endpoint is down, instead of hammering it (~8 attempts with backoff) and leaving a red check with no explanation, the action now tries 3 times, then posts a visible "review could not run" notice as request_changes. Faster feedback, never silently approves.
• inline_findings: "true" — findings that anchor to a diff line are posted as native line-anchored review comments (capped at 20) instead of only prose in the review body.
• verdict_policy: findings_severity_gated — the approve/request-changes verdict is now derived deterministically from the structured findings (request_changes iff any blocker-severity finding) instead of trusting the model's mood, since this repo publishes native review verdicts. Falls back to the model verdict when no findings are returned.

Now-active defaults worth knowing (no config needed)

• review_scope: auto — full review first, then incremental reviews of just the new commits on later pushes. Significant token savings on long-lived PRs.
• validate_required_checks: auto (mode warn) — risky PRs get classifier-derived required checks; reviews that never address them get a visible "Unaddressed required checks" section.

Choices for you (left out on purpose — flag if you want any)

1. SMART_MODEL repo variable — add it to route risky PRs/escalations to a stronger model than your current fallback. Without it, smart = FALLBACK_MODEL.
2. model_context_tokens — if the models behind your LiteLLM endpoint have small context windows (8k–32k local models), set this to the real window and the action right-sizes the corpus instead of using the coarse context_limit_mode: normal (~55–70k tokens). I don't know your per-model windows, so I didn't guess.
3. ai_response_format: json_object — helps small local models emit parseable JSON, but errors on backends that don't support it; only enable if you know the models behind LiteLLM accept it.
4. ci_status_check: "true" — waits (up to 5 min) for CI before reviewing so the review can see CI state. Adds latency to every review; worth it only if you want the reviewer gating on test results.
5. required_check_validation_mode: fail — make unaddressed required checks block instead of warn.

🤖 Generated with Claude Code

[joryirving]

Follow-up grounded in the actual LiteLLM config (home-ops litellm-configmap):

• Added ai_response_format: json_object in the latest commit. Safe because LiteLLM runs drop_params: true (backends that don't support it just don't see the param), and the review llama.cpp servers support grammar-constrained JSON — this directly reduces parse failures/retries.
• model_context_tokens: intentionally still unset. The review alias has a 200k input window, so the default context_limit_mode: normal (~55–70k tokens) fits with lots of headroom. Setting it to 200000 would ~3x the corpus and slow llama.cpp prompt eval for little gain. (Only needed if a repo ever points at ryzen — 8k ctx.)
• SMART_MODEL recommendation: smart currently resolves to FALLBACK_MODEL (MiniMax-M2.7, anthropic format) — already a reasonable smart tier. If you want a stronger escalation model, reasoning-coder (glm-5.1 with thinking, openai format) or opencode-go-deepseek-v4-pro are the best candidates in the proxy; set repo vars SMART_MODEL + SMART_FORMAT=openai. Note the chatgpt/* entries won't work for this action — they're mode: responses, not chat completions.

Superseded by a newer automated review for this pull request.

Superseded by a newer automated review for this pull request.

[its-saffron]

AI Automated Review

Incremental review: reviewed changes since the last managed review. This is not a full re-review of the entire PR.

Analysis engine: review@https://litellm.jory.dev/v1 (openai) (incremental delta review)

Recommendation

Approved. This PR performs a minor version bump of the pr-reviewer-action from v1.2.1 to v1.2.2. Based on the provided release notes for v1.2.0 and the incremental diff, this is a routine dependency update to a newer patch/minor version within the same feature set.

Change-by-change findings

.github/workflows/ai-pr-review.yaml

• Updated uses: misospace/pr-reviewer-action@... from v1.2.1 SHA to v1.2.2 SHA (e891235e8e3e75824b88fb696b1f7b4816122737).

Standards Compliance

• Dependency Updates: The update follows the repository's semver-based release process. The PR description notes that v1.2.0 was feature-heavy and backward-compatible, and this bump to v1.2.2 is a subsequent minor update.
• CI/Lint: No changes to the build or linting logic were detected; only the action version was updated.

Linked Issue Fit

• The PR title and body indicate this is a CI maintenance task to adopt the latest features and reliability improvements from the pr-reviewer-action release. It aligns with the goal of maintaining the AI PR review workflow.