Copilot agent infrastructure, emulator reliability, and try-fix workflow improvements

.github/agents/pr/PLAN-TEMPLATE.md

@@ -57,7 +57,7 @@ See `SHARED-RULES.md` for complete details. Key points:
 
 **Round 1: Run try-fix with each model (SEQUENTIAL)**
 - [ ] claude-sonnet-4.5
-- [ ] claude-opus-4.5
+- [ ] claude-opus-4.6
 - [ ] gpt-5.2
 - [ ] gpt-5.2-codex
 - [ ] gemini-3-pro-preview
@@ -68,10 +68,10 @@ See `SHARED-RULES.md` for complete details. Key points:
 - [ ] Invoke EACH model: "Any NEW fix ideas?"
 - [ ] Record responses in Cross-Pollination table
 - [ ] Run try-fix for new ideas (SEQUENTIAL)
-- [ ] Repeat until ALL 5 say "NO NEW IDEAS" (max 3 rounds)
+- [ ] Repeat until ALL 6 say "NO NEW IDEAS" (max 3 rounds)
 
 **Completion:**
-- [ ] Cross-Pollination table has all 5 responses
+- [ ] Cross-Pollination table has all 6 responses
 - [ ] Mark Exhausted: Yes
 - [ ] Compare passing candidates with PR's fix
 - [ ] Select best fix (results → simplicity → robustness)

.github/agents/pr/SHARED-RULES.md

@@ -132,7 +132,7 @@ Phase 4 uses these 5 AI models for try-fix exploration (run SEQUENTIALLY):
 | Order | Model |
 |-------|-------|
 | 1 | `claude-sonnet-4.5` |
-| 2 | `claude-opus-4.5` |
+| 2 | `claude-opus-4.6` |
 | 3 | `gpt-5.2` |
 | 4 | `gpt-5.2-codex` |
 | 5 | `gemini-3-pro-preview` |

.github/agents/pr/post-gate.md

@@ -24,6 +24,16 @@ If Gate is not passed, go back to `.github/agents/pr.md` and complete phases 1-2
 
 If try-fix cannot run due to environment issues, **STOP and ask the user**. Do NOT mark attempts as "BLOCKED" and continue.
 
+### 🚨 CRITICAL: Stop on Environment Blockers (Applies to Phase 4)
+
+The same "Stop on Environment Blockers" rule from `pr.md` applies here. If try-fix cannot run due to:
+- Missing Appium drivers
+- Device/emulator not available
+- WinAppDriver not installed
+- Platform tools missing
+
+**STOP and ask the user** before continuing. Do NOT mark try-fix attempts as "BLOCKED" and continue. Either fix the environment issue or get explicit user permission to skip.
+
 ---
 
 ## 🔧 FIX: Explore and Select Fix (Phase 3)
@@ -52,7 +62,7 @@ Phase 4 uses a **multi-model approach** to maximize fix diversity. Each AI model
 
 #### Round 1: Run try-fix with Each Model
 
-Run the `try-fix` skill **5 times sequentially**, once with each model (see `SHARED-RULES.md` for model list).
+Run the `try-fix` skill **6 times sequentially**, once with each model (see `SHARED-RULES.md` for model list).
 
 **For each model**, invoke the try-fix skill:
 ```
@@ -70,14 +80,31 @@ Generate ONE independent fix idea. Review the PR's fix first to ensure your appr
 
 **Wait for each to complete before starting the next.**
 
+**🧹 MANDATORY: Clean up between attempts.** After each try-fix completes (pass or fail), run these commands before starting the next attempt:
+
+```bash
+# 1. Restore any baseline state from the previous attempt (safe no-op if none exists)
+pwsh .github/scripts/EstablishBrokenBaseline.ps1 -Restore
+
+# 2. Restore all tracked files to HEAD (the merged PR state)
+# This catches any files the previous attempt modified but didn't restore
+git checkout HEAD -- .
+
+# 3. Remove untracked files added by the previous attempt
+# git checkout restores tracked files but does NOT remove new untracked files
+git clean -fd --exclude=CustomAgentLogsTmp/
+```
+
+**Why this is required:** Each try-fix attempt modifies source files. If an attempt fails mid-way (build error, timeout, model error), it may not run its own cleanup step. Without explicit cleanup, the next attempt starts with a dirty working tree, which can cause missing files, corrupt state, or misleading test results. Use `HEAD` (not just `-- .`) to also restore deleted files.
+
 #### Round 2+: Cross-Pollination Loop (MANDATORY)
 
 After Round 1, invoke EACH of the 5 models to ask for new ideas. **No shortcuts allowed.**
 
 **❌ WRONG**: Using `explore`/`glob`, declaring exhaustion without invoking each model
 **✅ CORRECT**: Invoke EACH model via task agent and ask explicitly
 
-**Steps (repeat until all 5 say "NO NEW IDEAS", max 3 rounds):**
+**Steps (repeat until all 6 say "NO NEW IDEAS", max 3 rounds):**
 
 1. **Compile bounded summary** (max 3-4 bullets per attempt):
    - Attempt #, approach (1 line), result (✅/❌), key learning (1 line)
@@ -176,7 +203,7 @@ Update the state file:
   | Model | Round 2 Response |
   |-------|------------------|
   | claude-sonnet-4.5 | NO NEW IDEAS |
-  | claude-opus-4.5 | NO NEW IDEAS |
+  | claude-opus-4.6 | NO NEW IDEAS |
   | gpt-5.2 | NO NEW IDEAS |
   | gpt-5.2-codex | NO NEW IDEAS |
   | gemini-3-pro-preview | NO NEW IDEAS |
@@ -298,3 +325,44 @@ Update all phase statuses to complete.
 - ❌ **Forgetting to revert between attempts** - Each try-fix must start from broken baseline, end with PR restored
 - ❌ **Declaring exhaustion prematurely** - All 5 models must confirm "no new ideas" via actual invocation
 - ❌ **Rushing the report** - Take time to write clear justification
+- ❌ **Skipping cleanup between attempts** - ALWAYS run `-Restore` + `git checkout HEAD -- .` + `git clean -fd --exclude=CustomAgentLogsTmp/` between try-fix attempts (see Step 1)
+
+---
+
+## Common Errors and Recovery
+
+### skill(try-fix) fails with "ENOENT: no such file or directory"
+
+**Symptom:** `skill(try-fix) Failed to read skill file: Error: ENOENT: no such file or directory, open '.../.github/skills/try-fix/SKILL.md'`
+
+**Root cause:** A previous try-fix attempt failed mid-way and left the working tree in a dirty state. Files may have been modified or deleted by `EstablishBrokenBaseline.ps1` without being restored.
+
+**Fix:** Run cleanup before retrying:
+```bash
+pwsh .github/scripts/EstablishBrokenBaseline.ps1 -Restore
+git checkout HEAD -- .
+git clean -fd --exclude=CustomAgentLogsTmp/
+```
+
+Then retry the try-fix attempt. The skill file should now be accessible.
+
+**Prevention:** Always run the cleanup commands between try-fix attempts (see Step 1).
+
+### try-fix attempt starts with dirty working tree
+
+**Symptom:** `git status` shows modified files before the attempt starts, or the build fails with unexpected errors from files the attempt didn't touch.
+
+**Root cause:** Previous attempt didn't restore its changes (crashed, timed out, or model didn't follow Step 8 restore instructions).
+
+**Fix:** Same as above — run `-Restore` + `git checkout HEAD -- .` + `git clean -fd --exclude=CustomAgentLogsTmp/` to reset to the merged PR state.
+
+### Build errors unrelated to the fix being attempted
+
+**Symptom:** Build fails with errors in files the try-fix attempt didn't modify (e.g., XAML parse errors, unrelated compilation failures).
+
+**Root cause:** Often caused by dirty working tree from a previous attempt. Can also be transient environment issues.
+
+**Fix:**
+1. Run cleanup: `pwsh .github/scripts/EstablishBrokenBaseline.ps1 -Restore && git checkout HEAD -- . && git clean -fd --exclude=CustomAgentLogsTmp/`
+2. Retry the attempt
+3. If it fails again with the same unrelated error, treat this as an environment/worktree blocker: STOP the try-fix workflow, do NOT continue with the next model, and ask the user to investigate (see "Stop on Environment Blockers").

.github/scripts/BuildAndRunHostApp.ps1

@@ -262,8 +262,8 @@ if ($Platform -eq "catalyst") {
 
         Write-Success "MacCatalyst app prepared (Appium will launch with test name)"
     } else {
-        Write-Warning "MacCatalyst app not found at: $appPath"
-        Write-Warning "Test may use wrong app bundle if another version is registered"
+        Write-Warn "MacCatalyst app not found at: $appPath"
+        Write-Warn "Test may use wrong app bundle if another version is registered"
     }
 
     # Set log file path directly - app will write ILogger output here
@@ -323,7 +323,7 @@ if ($Platform -eq "android") {
     & adb -s $DeviceUdid logcat -d | Select-String "com.microsoft.maui.uitests|DOTNET" > $deviceLogFile
 
     if ((Get-Item $deviceLogFile).Length -eq 0) {
-        Write-Warning "No logs found for com.microsoft.maui.uitests, dumping entire logcat..."
+        Write-Warn "No logs found for com.microsoft.maui.uitests, dumping entire logcat..."
         & adb -s $DeviceUdid logcat -d > $deviceLogFile
     }
 
@@ -397,7 +397,7 @@ if (Test-Path $deviceLogFile) {
         Write-Host ""
         Write-Info "Full device log: $deviceLogFile"
     } else {
-        Write-Warning "Could not read device log file"
+        Write-Warn "Could not read device log file"
     }
 
     Write-Host "═══════════════════════════════════════════════════════" -ForegroundColor Cyan

.github/scripts/BuildAndRunSandbox.ps1

@@ -258,7 +258,7 @@ if ($Platform -eq "catalyst") {
             Write-Success "MacCatalyst Sandbox app launched with log capture"
         }
     } else {
-        Write-Warning "MacCatalyst Sandbox app not found at: $appPath"
+        Write-Warn "MacCatalyst Sandbox app not found at: $appPath"
     }
 }
 
@@ -379,7 +379,7 @@ try {
         # Fallback: If we couldn't get PID, dump entire logcat buffer (unfiltered)
         # This ensures we always have logs for the agent to analyze
         Write-Host ""
-        Write-Warning "Could not capture app PID from Appium test output"
+        Write-Warn "Could not capture app PID from Appium test output"
         Write-Info "Dumping entire logcat buffer (unfiltered)..."
         & adb -s $DeviceUdid logcat -d > $deviceLogFile
         Write-Info "Logcat dumped to: $deviceLogFile (UNFILTERED - contains all apps)"
@@ -469,7 +469,7 @@ try {
                 Write-Info "All logs are from Sandbox app only (Maui.Controls.Sample.Sandbox)"
             }
         } else {
-            Write-Warning "Could not read device log file"
+            Write-Warn "Could not read device log file"
         }
 
         Write-Host "═══════════════════════════════════════════════════════" -ForegroundColor Cyan