Add scrape-and-improve skill, custom agent, and workflow for agent instruction refinement

.github/actions/scrape-and-improve/action.yml

@@ -0,0 +1,110 @@
+name: 'Scrape and Improve'
+description: 'Collects agent interaction data and generates instruction improvement recommendations.'
+
+inputs:
+  token:
+    description: 'GitHub token for API access'
+    required: false
+    default: ${{ github.token }}
+  pr-numbers:
+    description: 'Comma-separated list of PR numbers to analyze (optional)'
+    required: false
+    default: ''
+  label:
+    description: 'GitHub label to filter PRs by'
+    required: false
+    default: 'copilot'
+  since:
+    description: 'ISO 8601 date to filter PRs from (e.g., 2026-01-01)'
+    required: false
+    default: ''
+  recent-pr-count:
+    description: 'Number of most recent PRs to scrape for suggestion analysis'
+    required: false
+    default: '20'
+  memory-context:
+    description: 'Raw memory context text from agent interactions to analyze'
+    required: false
+    default: ''
+
+outputs:
+  report-file:
+    description: 'Path to the analysis report'
+    value: ${{ steps.analyze.outputs.report-file }}
+  recommendations-file:
+    description: 'Path to the recommendations JSON'
+    value: ${{ steps.analyze.outputs.recommendations-file }}
+  recommendations-count:
+    description: 'Number of recommendations generated'
+    value: ${{ steps.analyze.outputs.recommendations-count }}
+
+runs:
+  using: 'composite'
+  steps:
+
+    - name: Collect agent data
+      id: collect
+      shell: pwsh
+      env:
+        GH_TOKEN: ${{ inputs.token }}
+      run: |
+        Write-Host "Collecting agent interaction data..."
+        $params = @{
+          RepoRoot      = "${{ github.workspace }}"
+          OutputDir     = "${{ runner.temp }}/scrape-and-improve"
+          Repository    = "${{ github.repository }}"
+          RecentPRCount = ${{ inputs.recent-pr-count }}
+        }
+        if ("${{ inputs.pr-numbers }}") {
+          $params["PRNumbers"] = "${{ inputs.pr-numbers }}"
+        }
+        if ("${{ inputs.label }}") {
+          $params["Label"] = "${{ inputs.label }}"
+        }
+        if ("${{ inputs.since }}") {
+          $params["Since"] = "${{ inputs.since }}"
+        }
+        if ("${{ inputs.memory-context }}") {
+          $params["MemoryContext"] = "${{ inputs.memory-context }}"
+        }
+        & "${{ github.workspace }}/.github/skills/scrape-and-improve/scripts/Collect-AgentData.ps1" @params
+
+    - name: Analyze and generate recommendations
+      id: analyze
+      shell: pwsh
+      run: |
+        Write-Host "Analyzing collected data..."
+        $outputDir = "${{ runner.temp }}/scrape-and-improve"
+        & "${{ github.workspace }}/.github/skills/scrape-and-improve/scripts/Analyze-And-Recommend.ps1" `
+          -InputFile "$outputDir/collected-data.json" `
+          -OutputDir $outputDir `
+          -RepoRoot "${{ github.workspace }}"
+
+        "report-file=$outputDir/analysis-report.md" >> $env:GITHUB_OUTPUT
+        "recommendations-file=$outputDir/recommendations.json" >> $env:GITHUB_OUTPUT
+
+        $recommendations = Get-Content -Path "$outputDir/recommendations.json" -Raw | ConvertFrom-Json
+        $count = if ($recommendations -is [System.Array]) { $recommendations.Count } else { if ($recommendations) { 1 } else { 0 } }
+        "recommendations-count=$count" >> $env:GITHUB_OUTPUT
+
+    - name: Upload analysis artifacts
+      uses: actions/upload-artifact@v4
+      with:
+        name: scrape-and-improve-results
+        path: |
+          ${{ runner.temp }}/scrape-and-improve/analysis-report.md
+          ${{ runner.temp }}/scrape-and-improve/recommendations.json
+          ${{ runner.temp }}/scrape-and-improve/collected-data.json
+        retention-days: 30
+
+    - name: Print summary
+      shell: pwsh
+      run: |
+        $reportFile = "${{ steps.analyze.outputs.report-file }}"
+        if (Test-Path $reportFile) {
+          Write-Host "::group::Analysis Report"
+          Get-Content -Path $reportFile
+          Write-Host "::endgroup::"
+        }
+        Write-Host ""
+        Write-Host "Recommendations: ${{ steps.analyze.outputs.recommendations-count }}"

.github/agents/scrape-and-improve.md

@@ -0,0 +1,167 @@
+---
+name: scrape-and-improve
+description: Scrapes agent PR sessions, Copilot comments, CCA session data, and repository memories to identify patterns and generate instruction file updates for improved agent success rates.
+---
+
+# Scrape and Improve Agent
+
+Autonomously collects data from multiple agent interaction sources, analyzes patterns of success and failure, and **applies** instruction file updates to improve future agent performance.
+
+## When to Invoke
+
+- "Scrape and improve agent instructions"
+- "Analyze agent patterns and apply improvements"
+- "What patterns are agents struggling with? Fix them."
+- "Run scrape-and-improve and check in improvements"
+- Periodically (weekly/monthly) to improve agent instructions
+
+## When NOT to Invoke
+
+- For a single PR analysis → Use `learn-from-pr` agent or `/learn-from-pr` skill
+- During active PR work → Use `pr` agent
+- For analysis only without applying changes → Use `/scrape-and-improve` skill directly
+
+---
+
+## Workflow
+
+### Phase 1: Collect Data
+
+Run the data collection script to gather information from all sources:
+
+```bash
+pwsh .github/skills/scrape-and-improve/scripts/Collect-AgentData.ps1 \
+  -RepoRoot "." \
+  -RecentPRCount 20
+```
+
+**Data sources collected:**
+
+| Source | Location | What It Contains |
+|--------|----------|------------------|
+| Agent PR Sessions | `.github/agent-pr-session/*.md` | Fix candidates, root causes, phase statuses |
+| Copilot Comments | PR comments via `gh` CLI | AI Summary, try-fix attempts, test verification |
+| CCA Session Logs | `CustomAgentLogsTmp/PRState/` | Session state, fix attempts, convention patterns |
+| Repository Memories | Agent memory context | Stored facts (conventions, build commands, patterns) |
+| Recent PR Reviews | Most recent 20 PRs via `gh` API | Review comments, suggestion acceptance/rejection |
+
+If the agent has access to repository memories (provided in the conversation context), pass them via `-MemoryContext`:
+
+```bash
+pwsh .github/skills/scrape-and-improve/scripts/Collect-AgentData.ps1 \
+  -RepoRoot "." \
+  -RecentPRCount 20 \
+  -MemoryContext "<paste memory block>"
+```
+
+### Phase 2: Analyze Patterns
+
+Run the analysis script to identify patterns across collected data:
+
+```bash
+pwsh .github/skills/scrape-and-improve/scripts/Analyze-And-Recommend.ps1 \
+  -RepoRoot "."
+```
+
+This produces:
+- `CustomAgentLogsTmp/scrape-and-improve/analysis-report.md` — Full report
+- `CustomAgentLogsTmp/scrape-and-improve/recommendations.json` — Structured recommendations
+
+### Phase 3: Present Results
+
+Display the analysis report to the user, including:
+
+1. **Data Summary** — What was collected (session files, comments, memories, PRs)
+2. **Pattern Analysis** — Slow discovery PRs, quick successes, root causes
+3. **Memory Analysis** — Subject frequency, convention patterns
+4. **Recent PR Suggestion Analysis** — Acceptance/rejection rates, hotspot areas
+5. **Recommendations** — Prioritized list with evidence and impact
+
+### Phase 4: Apply Improvements
+
+For each **High or Medium priority** recommendation from the analysis (priorities are determined by the `Analyze-And-Recommend.ps1` script based on evidence strength, recurrence across PRs, and expected impact on agent success rates):
+
+| Category | Action |
+|----------|--------|
+| Instruction file | Edit existing or create new `.github/instructions/*.instructions.md` |
+| Skill enhancement | Edit `.github/skills/*/SKILL.md` |
+| Architecture doc | Edit `/docs/design/*.md` or create in `.github/architecture/` |
+| General AI guidance | Edit `.github/copilot-instructions.md` |
+
+**Before each edit:**
+- Read the target file first
+- Check for existing similar content (don't duplicate)
+- Match the existing style/format
+- Find the appropriate section
+
+**Skip applying if:**
+- Content already exists
+- Recommendation is too vague
+- Would require major restructuring
+
+### Phase 5: Verify and Report
+
+After applying changes:
+
+1. Run `git diff` to review all edits
+2. Verify no syntax errors in modified files (valid markdown)
+3. Confirm style matches existing content
+
+Present a summary:
+
+```markdown
+## Scrape and Improve Results
+
+### Data Collected
+| Source | Count |
+|--------|-------|
+| [source] | [count] |
+
+### Recommendations Applied
+| File | Change |
+|------|--------|
+| [path] | [what was added/modified] |
+
+### Recommendations Not Applied
+| Recommendation | Reason |
+|----------------|--------|
+| [rec] | [why skipped] |
+
+### Key Findings
+- [finding 1]
+- [finding 2]
+```
+
+---
+
+## Error Handling
+
+| Situation | Action |
+|-----------|--------|
+| No agent PR sessions found | Report empty, suggest checking date range |
+| `gh` CLI not authenticated | Report error, proceed with local data only |
+| No Copilot comments found | Proceed with session files and memories |
+| No CCA session data | Proceed with other sources |
+| No patterns identified | Report "insufficient data" with data source summary |
+| Target file doesn't exist | Create if instruction doc, skip if code |
+| Duplicate content exists | Skip, note in report |
+
+## Constraints
+
+- **Evidence-based** — Every recommendation must cite specific PRs or data
+- **Don't duplicate** — Check existing instruction files before recommending or applying
+- **Match style** — Read file before editing
+- **Only apply High/Medium priority** — Report Low priority without applying
+- **Focus on actionable** — Skip vague observations
+- **Respect existing patterns** — Match style and structure of existing instruction files
+
+---
+
+## Difference from Skill
+
+| Aspect | `/scrape-and-improve` Skill | This Agent |
+|--------|----------------------------|------------|
+| Output | Analysis report and recommendations | Applied changes to instruction files |
+| Mode | Analysis only | Autonomous — collects, analyzes, applies |
+| Use when | Want to review without applying | Want full automation |
+| Trigger | "analyze agent patterns" | "scrape and improve, check in improvements" |

.github/copilot-instructions.md

@@ -78,6 +78,39 @@ Platform-specific files use naming conventions to control compilation:
 - `src/Essentials/samples/Essentials.Sample` - Essentials API demonstrations (non-UI MAUI APIs)
 - `src/BlazorWebView/samples/` - BlazorWebView sample applications
 
+## Common Fix Patterns (From Agent Analysis)
+
+These patterns were identified from scrape-and-improve analysis of past agent PR sessions. They help agents find fixes faster by documenting what worked and what didn't.
+
+### NavigationPage Handler Disconnection
+
+When fixing handler disconnection issues with `NavigationPage.RemovePage()`:
+- ❌ **Don't** add `DisconnectHandlers()` inside `SendHandlerUpdateAsync` callback (failed in PR #32289)
+- ❌ **Don't** add `DisconnectHandlers()` synchronously after `SendHandlerUpdateAsync` (failed in PR #32289)
+- ✅ **Do** add `DisconnectHandlers()` at the end of the Legacy `RemovePage` method path
+- **Key insight**: The fix location is in the legacy code path, not the modern handler update path
+
+### CollectionView EmptyView (Android)
+
+When fixing EmptyView display issues inside unconstrained layouts (e.g., `VerticalStackLayout`):
+- ❌ **Don't** just check for infinity before casting in `GetHeight()`/`GetWidth()` (failed in PR #33134)
+- ✅ **Do** handle `double.PositiveInfinity` → `int.MaxValue` propagation in `EmptyViewAdapter`
+- **Key insight**: In C#, `(int)double.PositiveInfinity` produces `int.MaxValue`. The fix must normalize dimensions where they're consumed (e.g., convert `int.MaxValue` back to `double.PositiveInfinity` for layout measurement), not where they're cast
+- **Root cause**: When CollectionView is inside VerticalStackLayout, height constraint is `double.PositiveInfinity` which propagates through `EmptyViewAdapter.RecyclerViewHeight`
+
+### Shell Navigation Tests (iOS)
+
+When fixing Shell back button / navigation tests on iOS:
+- Check if `OnAppearing` count expectations changed due to lifecycle changes in related PRs
+- Shell navigation involves both `DidPopItem` and `SendPop()` — ensure stacks stay in sync
+- **Key insight**: Simplify `DidPopItem` to always call `SendPop()` when stacks are out of sync
+
+### Device Test Isolation
+
+When fixing flaky device tests (e.g., Stepper tests):
+- Tests that modify control state (like `Minimum`, `Maximum`) can leak state to subsequent tests
+- **Key insight**: Check if a previous test in the same class modifies shared state without resetting it
+
 ## Development Workflow
 
 ### Testing
@@ -220,6 +253,13 @@ The repository includes specialized custom agents and reusable skills for specif
    - **Output**: Applied changes to instruction files, skills, architecture docs, code comments
    - **Do NOT use for**: Analysis only without applying changes → Use `/learn-from-pr` skill instead
 
+5. **scrape-and-improve** - Scrapes agent data from multiple sources and applies instruction improvements
+   - **Use when**: Want to analyze agent patterns across multiple PRs and apply improvements automatically
+   - **Capabilities**: Collects data from PR sessions, Copilot comments, CCA sessions, memories, and recent PRs; analyzes patterns; applies High/Medium priority instruction updates
+   - **Trigger phrases**: "scrape and improve agent instructions", "analyze agent patterns and apply improvements", "run scrape-and-improve and check in improvements"
+   - **Output**: Applied changes to instruction files based on aggregated pattern analysis
+   - **Do NOT use for**: Analysis only without applying → Use `/scrape-and-improve` skill instead; single PR analysis → Use `learn-from-pr` agent
+
 ### Reusable Skills
 
 Skills are modular capabilities that can be invoked directly or used by agents. Located in `.github/skills/`:
@@ -277,9 +317,15 @@ Skills are modular capabilities that can be invoked directly or used by agents.
    - **Categories**: Build, WindowsTemplates, macOSTemplates, Blazor, MultiProject, Samples, AOT, RunOnAndroid, RunOniOS
    - **Note**: **ALWAYS use this skill** instead of manual `dotnet test` commands for integration tests
 
+9. **scrape-and-improve** (`.github/skills/scrape-and-improve/SKILL.md`)
+   - **Purpose**: Scrapes agent PR sessions, Copilot comments, CCA sessions, and memories to generate instruction updates
+   - **Trigger phrases**: "improve agent instructions", "scrape agent data", "analyze agent patterns", "what patterns are agents struggling with?"
+   - **Scripts**: `Collect-AgentData.ps1`, `Analyze-And-Recommend.ps1`
+   - **Note**: Also available as a GitHub Action (`.github/actions/scrape-and-improve/`) and scheduled workflow
+
 #### Internal Skills (Used by Agents)
 
-9. **try-fix** (`.github/skills/try-fix/SKILL.md`)
+10. **try-fix** (`.github/skills/try-fix/SKILL.md`)
    - **Purpose**: Proposes ONE independent fix approach, applies it, tests, records result with failure analysis, then reverts
    - **Used by**: pr agent Phase 3 (Fix phase) - rarely invoked directly by users
    - **Behavior**: Reads prior attempts to learn from failures. Max 5 attempts per session.