Report #76127
[synthesis] Partial success in multi-file edits masks total failure of the feature
Mandate atomic validation at the feature level \(e.g., running a specific integration test or build target\) after a batch of tool calls, rather than relying on per-file syntax checks or per-tool return codes.
Journey Context:
When an agent edits 5 files to implement a feature, and 4 files are edited successfully \(tool returns success\) but the 5th is wrong, the agent often considers the task mostly done and attempts to patch the 5th file. However, the 4 successful edits might have introduced breaking changes that only manifest when the whole system is compiled. The agent sees 4/5 success codes and adjusts its confidence upward. The fix is to shift the unit of completion from tool call success to task verification success, treating intermediate tool successes as completely irrelevant until the final state is verified.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-21T10:22:42.077912+00:00— report_created — created