Parent: #398
Goal
Reject pending edit previews when replacement content appears contaminated by repeated tool-call echoes or apply markers before the user can approve corrupted output.
Implementation Notes
Target area: PendingChangesPreviewService and its new apply utility layer from the fuzzy apply issue.
Add a pure artifact guard that scans replacement/preview text for known hallucinated tool markers, using thresholds to avoid false positives in legitimate source files.
Candidate markers:
to=functionsfunctions.<|tool|>- repeated
tool_use
- repeated
apply_edits
- repeated JSON snippets like
"name":"Edit" or "name":"MultiEdit"
Use hard rejection for very strong markers and repeated-marker thresholds for weaker markers. Return a clear rejected outcome with offending markers.
Acceptance Criteria
- Replacement content with repeated tool-call echoes is rejected before preview approval.
- A single legitimate mention of a marker string in source code is not rejected.
- Rejected edits include enough reason text for the UI to display what was detected.
- Nothing is written or accepted when an edit is rejected.
Tests
Add focused tests in PendingChangesPreviewServiceTests or a new pure utility test suite for marker thresholds and false-positive cases.
Run the targeted AgentHubCore test for the affected suite.
Parent: #398
Goal
Reject pending edit previews when replacement content appears contaminated by repeated tool-call echoes or apply markers before the user can approve corrupted output.
Implementation Notes
Target area:
PendingChangesPreviewServiceand its new apply utility layer from the fuzzy apply issue.Add a pure artifact guard that scans replacement/preview text for known hallucinated tool markers, using thresholds to avoid false positives in legitimate source files.
Candidate markers:
to=functionsfunctions.<|tool|>tool_useapply_edits"name":"Edit"or"name":"MultiEdit"Use hard rejection for very strong markers and repeated-marker thresholds for weaker markers. Return a clear rejected outcome with offending markers.
Acceptance Criteria
Tests
Add focused tests in
PendingChangesPreviewServiceTestsor a new pure utility test suite for marker thresholds and false-positive cases.Run the targeted AgentHubCore test for the affected suite.