Handle persona-explorer invocation failures as first-class partial results

.github/aw/create-agentic-workflow.md

@@ -86,6 +86,7 @@ Use this mode for exploratory testing, persona walkthroughs, and "what workflow
 - Do not create or edit workflow files.
 - Return a compact recommendation covering trigger, any scoped `paths:` filters for file-event triggers, read tools, safe outputs, permissions, and explicit `noop` criteria.
 - For recurring reports or digests, always include the report window, grouping dimensions, and deduplication key. See [triggers.md](triggers.md) for key-format examples.
+- If custom-agent invocation is unavailable in the environment, use fallback evaluation mode: provide the recommendation directly from local gh-aw guidance, record scoring/response quality as unavailable, and continue with partial-results reporting.
 - Exit ad hoc evaluation mode only when the user explicitly asks to create, implement, or write the workflow file.
 - End by offering to turn the recommendation into `.github/workflows/<workflow-id>.md` if the user wants to proceed.
 

.github/workflows/agent-persona-explorer.md

@@ -148,6 +148,7 @@ Store all scenarios in cache memory.
   - Prompt clarity
   - Completeness
 - Notable patterns or issues (be concise)
+- If invocation fails: mark the scenario as `invocation_unavailable`, set quality scoring to `N/A`, and continue.
 
 **Assessment questions:** Does the suggestion include appropriate triggers (`on:`)? Correct tools (github, web-fetch, playwright, etc.)? Proper safe-outputs? Security best practices (minimal permissions, network restrictions)? A clear, actionable prompt?
 
@@ -161,7 +162,7 @@ For each selected scenario, invoke the "agentic-workflows" custom agent tool, pr
 - You are ONLY testing the agent's responses, NOT creating actual workflows
 - **Keep responses focused and concise** - summarize findings instead of verbose descriptions
 - Aim for quality over quantity - fewer well-analyzed scenarios are better than many shallow ones
-- **If any tool call fails, record the error briefly and move on to the next scenario** - do NOT retry or get stuck
+- **If any tool call fails, record the error briefly, mark scoring as unavailable for that scenario, and move on to the next scenario** - do NOT retry or get stuck
 
 ## Phase 4: Analyze Results (4 minutes)
 
@@ -175,6 +176,7 @@ Review all captured responses and identify:
 ### Quality Insights (summarize briefly)
 - Which scenarios received the best responses (average score > 4)?
 - Which scenarios received weak responses (average score < 3)?
+- If scenario invocation failed, note that scoring is unavailable for affected scenarios and exclude them from the numeric average.
 
 ### Potential Issues (only list critical issues)
 - Does the agent ever suggest insecure configurations?
@@ -189,7 +191,7 @@ Review all captured responses and identify:
 
 **MANDATORY OUTPUT**: Regardless of how many phases completed successfully, you MUST call either the `create issue` or the `noop` safe-output tool before finishing. Failing to call a safe-output tool is the most common cause of workflow failures.
 
-Create a GitHub issue with a **concise** summary report. Use the `create issue` safe-output to publish your findings. Even if only 1-2 scenarios were tested, create the issue with partial results.
+Create a GitHub issue with a **concise** summary report. Use the `create issue` safe-output to publish your findings. Even if only 1-2 scenarios were tested, create the issue with partial results. Treat invocation failures as a standard partial-results outcome and explicitly mark scoring as unavailable where applicable.
 
 **Issue title**: "Agent Persona Exploration - [DATE]" (e.g., "Agent Persona Exploration - 2024-01-16")
 
@@ -220,7 +222,7 @@ Example:
 - **Agent**: [name]
 - **Personas This Run**: [3 persona names]
 - **Scenarios Tested**: [count - should be 3-4, selected from the 6 generated in Phase 2 (2 per persona × 3 personas)]
-- **Average Quality Score**: [X.X/5.0]
+- **Average Quality Score**: [X.X/5.0 or N/A when invocation/scoring is unavailable]
 
 ### Key Findings (3-5 bullet points max)
 [High-level insights - keep concise]