Improve try-fix skill: add eval.yaml and fix prompt issues

.github/skills/try-fix/SKILL.md

@@ -14,7 +14,7 @@ Attempts ONE fix for a given problem. Receives all context upfront, tries a sing
 2. **Single-shot** - Each invocation = ONE fix idea, tested, reported
 3. **Alternative-focused** - Always propose something DIFFERENT from existing fixes (review PR changes first)
 4. **Empirical** - Actually implement and test, don't just theorize
-5. **Context-driven** - All information provided upfront; don't search for additional context
+5. **Context-driven** - Work with what's provided and git history; don't search external sources
 
 **Every invocation:** Review existing fixes → Think of DIFFERENT approach → Implement and test → Report results
 
@@ -23,7 +23,7 @@ Attempts ONE fix for a given problem. Receives all context upfront, tries a sing
 🚨 **Try-fix runs MUST be executed ONE AT A TIME - NEVER in parallel.**
 
 **Why:** Each try-fix run:
-- Modifies the same source files (SafeAreaExtensions.cs, etc.)
+- Modifies the same target source files
 - Uses the same device/emulator for testing
 - Runs EstablishBrokenBaseline.ps1 which reverts files to a known state
 
@@ -147,6 +147,8 @@ The skill is complete when:
 
 **Never stop due to:** Compile errors (fix them), infrastructure blame (debug your code), giving up too early.
 
+> **Session limits:** Each try-fix *invocation* allows up to 3 compile/test iterations. The *calling orchestrator* controls how many invocations (attempts) to run per session (typically 4-5 as part of pr-review Phase 3).
+
 ---
 
 ## Workflow
@@ -182,7 +184,7 @@ The skill is complete when:
 - What files should be investigated?
 - Are there hints about what to try or avoid?
 
-**Do NOT search for additional context.** Work with what's provided.
+**Do NOT search for external context.** Work with what's provided and the git history.
 
 ### Step 2: Establish Baseline (MANDATORY)
 
@@ -317,7 +319,7 @@ git diff | Set-Content "$OUTPUT_DIR/fix.diff"
 pwsh .github/scripts/EstablishBrokenBaseline.ps1 -Restore
 ```
 
-🚨 Do NOT use `git checkout HEAD -- .` or `git clean` to restore — use the script.
+🚨 Use `EstablishBrokenBaseline.ps1 -Restore` — not `git checkout`, `git restore`, or `git reset` (see Step 2 for why).
 
 ### Step 9: Report Results
 

.github/skills/try-fix/tests/eval.yaml

@@ -0,0 +1,142 @@
+scenarios:
+  - name: "Happy path: propose alternative fix with different approach"
+    prompt: |
+      Use the try-fix skill to attempt a fix for the following problem:
+
+      Problem: CollectionView throws ObjectDisposedException when navigating back from a page on Android.
+
+      The existing PR fix already adds a null check on the adapter in OnMeasure(). We need a DIFFERENT approach.
+
+      Test command: pwsh .github/scripts/BuildAndRunHostApp.ps1 -Platform android -TestFilter "Issue54321"
+
+      Target files:
+        - src/Controls/src/Core/Handlers/Items/ItemsViewHandler.Android.cs
+
+      Platform: android
+
+      Hints: Focus on lifecycle/disposal timing rather than null checks.
+    assertions:
+      - type: output_contains
+        value: "Approach"
+      - type: output_contains
+        value: "Different from existing fix"
+      - type: output_not_contains
+        value: "null check on the adapter"
+      - type: output_contains
+        value: "result.txt"
+      - type: output_contains
+        value: "analysis"
+    rubric:
+      - "The agent proposes a fix approach that is clearly distinct from the null-check-on-adapter approach described as the existing fix"
+      - "The agent creates an approach.md document explaining the difference from the existing fix"
+      - "The agent saves artifacts (result.txt, fix.diff, analysis.md) to an output directory"
+      - "The agent reverts changes after testing using EstablishBrokenBaseline.ps1 rather than manual git commands"
+    timeout: 120
+
+  - name: "Negative trigger: code review request should not invoke try-fix workflow"
+    prompt: |
+      Please review the code changes in PR #34567 and tell me if the fix looks correct.
+      The PR modifies SafeAreaExtensions.cs to add padding adjustments.
+    assertions:
+      - type: output_not_contains
+        value: "EstablishBrokenBaseline"
+      - type: output_not_contains
+        value: "attempt-"
+      - type: output_not_contains
+        value: "OUTPUT_DIR"
+      - type: output_not_contains
+        value: "result.txt"
+    rubric:
+      - "The agent provides a code review rather than attempting to run a fix workflow"
+      - "The agent does not create output directories or artifact files associated with try-fix execution"
+    timeout: 60
+
+  - name: "Regression: agent must not claim Pass without running test command"
+    prompt: |
+      Use the try-fix skill for this problem:
+
+      Problem: Button tap does not fire Clicked event on iOS when inside a ScrollView.
+
+      Existing fix: Adds a gesture recognizer passthrough in ScrollViewHandler.
+
+      Test command: pwsh .github/scripts/BuildAndRunHostApp.ps1 -Platform ios -TestFilter "Issue99001"
+
+      Target files:
+        - src/Controls/src/Core/Handlers/Compatibility/ScrollViewRenderer.iOS.cs
+
+      Platform: ios
+
+      Note: No iOS simulator is currently available.
+    assertions:
+      - type: output_not_contains
+        value: "Pass"
+      - type: output_contains
+        value: "Blocked"
+      - type: output_not_contains
+        value: "code compiles successfully"
+      - type: output_not_contains
+        value: "logic looks correct"
+    rubric:
+      - "The agent reports 'Blocked' (not 'Pass') when no device/simulator is available to run the test command"
+      - "The agent does not claim success based on code review or compilation alone"
+      - "The agent explains why the result is Blocked rather than Pass or Fail"
+    timeout: 120
+
+  - name: "Edge case: second attempt reads and avoids prior failed approach"
+    prompt: |
+      Use the try-fix skill to attempt a fix. This is attempt #2 — attempt #1 already failed.
+
+      Problem: Shell navigation causes a NullReferenceException in ShellItemHandler on Android when popping to root.
+
+      Prior attempt (FAILED): Modified OnPageSelected to reset cached state after navigation.
+      Failure reason: OnPageSelected fires after layout measurement, so cached value was already consumed.
+
+      Test command: pwsh .github/scripts/BuildAndRunHostApp.ps1 -Platform android -TestFilter "Issue77123"
+
+      Target files:
+        - src/Controls/src/Core/Handlers/Shell/ShellItemHandler.Android.cs
+
+      Platform: android
+
+      Hints: Try intercepting at a different lifecycle point, before measurement occurs.
+    assertions:
+      - type: output_not_contains
+        value: "OnPageSelected"
+      - type: output_contains
+        value: "Different from existing fix"
+      - type: output_contains
+        value: "approach"
+    rubric:
+      - "The agent acknowledges the prior failed attempt and explicitly avoids repeating the OnPageSelected approach"
+      - "The agent proposes a lifecycle-based alternative that fires before layout measurement"
+      - "The approach.md clearly explains how this attempt differs from the prior failed attempt"
+    timeout: 120
+
+  - name: "Regression: agent must use EstablishBrokenBaseline script, not manual git commands"
+    prompt: |
+      Use the try-fix skill for this problem:
+
+      Problem: Entry text is not preserved on orientation change on Android.
+
+      Existing fix: Saves/restores text in OnSaveInstanceState override.
+
+      Test command: pwsh .github/scripts/BuildAndRunHostApp.ps1 -Platform android -TestFilter "Issue88200"
+
+      Target files:
+        - src/Core/src/Platform/Android/EntryHandler.Android.cs
+
+      Platform: android
+    assertions:
+      - type: output_contains
+        value: "EstablishBrokenBaseline"
+      - type: output_not_contains
+        value: "git checkout HEAD"
+      - type: output_not_contains
+        value: "git restore"
+      - type: output_not_contains
+        value: "git reset"
+    rubric:
+      - "The agent uses EstablishBrokenBaseline.ps1 to establish the baseline and restore working directory, not raw git commands"
+      - "The agent does not use 'git checkout HEAD', 'git restore', or 'git reset' to revert fix files"
+      - "The restore step is explicitly called after testing (whether pass or fail)"
+    timeout: 120