MAUI PR Review Orchestrator

PR Review — 3-Phase Orchestrator

End-to-end PR review workflow that orchestrates phases to explore independent fix alternatives and produce a recommendation.

Trigger phrases: "review PR #XXXXX", "work on PR #XXXXX", "fix issue #XXXXX"

🚨 NEVER use gh pr review --approve or --request-changes. AI agents must NEVER post review comments. 🚨 DO NOT post any comments to the PR. This skill only produces output files in CustomAgentLogsTmp/PRState/.

Overview

Gate (pre-run)    → Already completed by Review-PR.ps1 before this skill runs
Phase 1: Pre-Flight   → Gather context, classify files, code review     → .github/pr-review/pr-preflight.md
Phase 2: Try-Fix      → ⚠️ MANDATORY multi-model exploration           → invoke try-fix skill (×2 models)
Phase 3: Report       → Write review recommendation                     → .github/pr-review/pr-report.md

Gate and Branch setup are handled by Review-PR.ps1 before this skill is invoked. The gate result is passed in the prompt. Do NOT re-run gate verification.

All phases write output to: CustomAgentLogsTmp/PRState/{PRNumber}/PRAgent/{phase}/content.md Pre-Flight also writes: CustomAgentLogsTmp/PRState/{PRNumber}/PRAgent/pre-flight/code-review.md

Critical Rules

• ❌ Never run git checkout or git switch to change branches — stay on the review branch set up by the caller
• ❌ Never stop and ask the user — use best judgment to skip blocked phases and continue
• ❌ Never mark a phase complete with pending fields
• ❌ Never skip Phase 2 multi-model exploration — it is MANDATORY for every review, no exceptions
• ❌ Never run git commands that change branch state during Phases 2-3 (scripts handle file manipulation)
• ❌ Never duplicate phase content — each phase writes ONLY to its own content.md. Do NOT copy gate results into try-fix or report content files.
• ✅ Always create CustomAgentLogsTmp/ output files for every phase
• ✅ Always include Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> in any commits
• ✅ Always use skills' scripts — don't bypass with manual commands
• ✅ Each phase's content.md must use the exact template from the phase instruction doc — no extra prose

Multi-Model Configuration

Phase 2 uses these 2 AI models (run SEQUENTIALLY — they modify the same files):

Two GPT-family models keep the try-fix phase fast (each attempt is a full build+test cycle, so every extra model adds ~15–20 min to the review) while still varying optimization profiles: gpt-5.3-codex explores the implementation first, and gpt-5.6-sol provides the higher-reasoning comparison.

🚨 MANDATORY: Use mode: "sync" for ALL try-fix task invocations. Never use mode: "background". Background mode causes the orchestrator to move on before the attempt finishes, which means try-fix/content.md is never written and try-fix results are lost from the PR comment. Each try-fix task MUST complete and return its result before you proceed to the next attempt or to the Phase 3 completion checklist.

Environment Blockers

Phase 1: Pre-Flight

Read and follow .github/pr-review/pr-preflight.md

Gather context from the issue, PR, comments, classify changed files, and perform a deep code review using the code-review skill.

Pre-Flight now has two parts:

• Part A (Steps 1–6): Context gathering — read issue, PR, comments, classify files
• Part B (Step 7): Code review — independence-first code analysis using .github/skills/code-review/SKILL.md and .github/skills/code-review/references/review-rules.md

Outputs:

• pre-flight/content.md — Context + code review summary
• pre-flight/code-review.md — Full code-review output (findings, blast radius, failure-mode probes, verdict)

Gate: None — always runs.

Why code review runs here: The code-review findings (❌ Errors, ⚠️ Warnings, failure-mode probes, blast radius) become structured hints for Phase 2 (Try-Fix). Instead of each model starting from scratch, they receive concrete code concerns to address, leading to higher-quality fix exploration.

Phase 2: Try-Fix → Invoke try-fix Skill (×2 Models)

Read and follow .github/skills/try-fix/SKILL.md

⚠️ THIS PHASE IS MANDATORY. YOU MUST NEVER SKIP IT. NO EXCEPTIONS.

Even if the PR's fix looks correct and Gate passed, you MUST still run both models to explore alternative approaches. The purpose is to find the BEST fix, not just validate one.

⏱️ HARD TIME BUDGET — Phase 2 must finish within ~90 minutes. Task 3 (this whole Copilot Review step) has a 180-minute safety cap. The pipeline preserves partial output when that cap is reached, but the review remains incomplete and can lose its final comparison — so reaching it is still unacceptable. Track wall-clock time from the moment you enter Phase 2. Order of work: (1) run each of the two models once, writing try-fix/content.md after each attempt; (2) select the best fix. In the CI split-step reviewer, skip cross-pollination entirely; direct interactive invocations may do one optional round only when comfortably under budget. The moment you approach ~90 minutes — or sooner if attempts stop making progress — STOP immediately, finalize try-fix/content.md with the results so far, and move to Phase 3. Never run open-ended "exhaustion", per-candidate deep-dive, repeated candidate repair loops, or repeated cross-pollination that can consume the whole budget.

🚨 CRITICAL: try-fix is Independent of PR's Fix

"Independent" means each model explores a different fix approach from the PR's fix — not that models are isolated from code-review context. Code-review findings are provided as advisory background to improve fix quality.

The purpose is NOT to re-test the PR's fix, but to:

1. Generate independent fix ideas — What would YOU do to fix this bug?
2. Test those ideas empirically — Actually implement and run tests
3. Compare with PR's fix — Is there a simpler/better alternative?
4. Learn from failures — Record WHY failed attempts didn't work