the cagent-output fence may not always be at the start of the response

[derekmisler]

Summary

Fixes a bug in the cagent-output extraction logic where code fences that don't start at the beginning of a line were not being properly detected. This occurs when AI agents emit conversational text before the code fence, causing the extraction to fail and potentially leak agent thoughts into the final output.

Changes

Modified

• .github/workflows/pr-describe.yml: Updated JavaScript extraction logic to use regex matching instead of relying on line-based parsing for cagent-output blocks
• action.yml: Modified AWK script to remove line-start anchoring (^) from the cagent-output pattern matching, allowing detection of code fences that appear mid-line
• tests/test-output-extraction.sh: Added comprehensive test case (1b) to verify extraction works when code fence is preceded by agent thoughts on the same line

Technical Details

The core issue was that the extraction logic used ^```cagent-output pattern matching, which only worked when the code fence started at the beginning of a line. However, AI agents sometimes emit conversational text like "I'll analyze the PR..." immediately before the code fence without a newline, causing the pattern to fail.

Key changes:

• GitHub Workflow: Added regex-based extraction (raw.match(/```cagent-output\n([\s\S]*?)```/)) as a fallback to handle cases where the composite action's cleaning doesn't catch everything
• Composite Action: Updated AWK script to use /```cagent-output/ instead of /^```cagent-output$/ and implemented proper state-based capturing that works regardless of line position
• Test Suite: Added test case that simulates real-world scenario where agent emits "I'll analyze the PR by reading..." before the code fence, ensuring the fix prevents agent thoughts from leaking into the final output

Testing

• Added test case 1b that verifies extraction works with mid-line code fences
• Test includes validation that agent thoughts are properly excluded from clean output
• Maintains backward compatibility with existing line-start code fences
• Both AWK-based (composite action) and JavaScript-based (workflow) extraction methods updated

Breaking Changes

None - this is a backward-compatible bug fix that improves robustness.

Related Issues

None explicitly referenced - this addresses a reliability issue discovered during PR description generation workflows.

Checklist

• Code follows project style guidelines
• Self-review completed
• Comments added explaining the fix
• Comprehensive test coverage added
• No new warnings introduced
• Tests pass locally
• Security implications considered (prevents information leakage)

[derekmisler]

/describe

[docker-agent]

✅ PR description has been generated and updated!

[docker-agent]

Review Summary

Found 2 issues in the changed code that should be addressed:

1. JavaScript regex inconsistency - The regex in pr-describe.yml doesn't handle mid-line code fences like the AWK script does
2. Missing error handling - File read operation lacks try-catch wrapper

Both issues are in .github/workflows/pr-describe.yml and relate to the fallback extraction logic.

[derekmisler]

/review

[docker-agent]

Review Summary

Found 2 potential issues with the regex pattern in the JavaScript extraction code. The AWK script changes look solid.

Issues Found

• Medium severity: Regex requires newline before closing fence
• Low severity: Regex captures unwanted content when no newline after opening fence

Both issues could cause extraction failures in edge cases.