API Valid, Interface Invalid: A Boundary Failure in Title Generation

In the backend of a software system, an unnoticed issue began in the backlog (#2637/brand#16). This was not a standard task name but a snippet of checklist and prompt text amalgamated into an issue title. At first, it blended seamlessly into the ticket flow, but eventually, pattern recognition identified the problem. Another identical instance (#2644/brand#17) appeared, exhibiting the same glitch.

The problem wasn't "agent drift" but a boundary-validation failure: structurally valid internal text slipped into a surface artifact without being validated for human understanding.

The problem stemmed from the internal representation, boundary transformation, and human-facing artifact. Tracking the leak back to the make_issue_title function in tasks/issue_title_builder.py revealed that the boundary transformation accepted checklist prompt scaffolding and emitted it as issue/PR titles. The system succeeded in API calls, but this did not guarantee interface success. The defect was specific: boundary logic validated the structure but not the usefulness of the title for reviewers scanning work history.

Upon inspection of the current task model and title builder path, it was discovered that the description was the available human-facing label consumed at the boundary. The failure was that boundary logic didn't use this canonical human label; instead, it allowed truncated checklist/prompt scaffolding to pass through as title text. The fix corrected this at the boundary: using description as the source for title generation and flagging checklist/prompt-shaped strings even when they were API-valid.

In tests/test_title_generation.py, the implemented assertions checked for specific leakage signals: title equaling or beginning with raw checklist/prompt tokens matching the observed leak shape (e.g., #2637/brand#16). The boundary output matching known raw/truncated checklist leakage patterns emitted a warning. If a valid description was present, the generated title used that human-facing label instead of the checklist/prompt scaffolding.

This issue was subtle and practical rather than a loud outage. The lesson learned was that internal scaffolding could be structurally valid, and boundary code could pass CI and API calls, but human-facing artifacts could still be wrong. The system complied with what the API accepted, not what people needed to read. The fix added a guardrail at the boundary separating internal representation from public artifacts, making the system legible.

Currently, this guardrail detects and surfaces checklist-prompt leakage instead of silently shipping it as titles.