Problem
The Ladon review for adcontextprotocol/adcp#6902 reproducibly completes its
analysis but never returns schema-validated findings. The job correctly fails
closed, but identical reruns cannot produce an actionable review.
Observed on both the original and refreshed PR heads:
Across the attempts, Claude Code reports:
error_max_structured_output_retries
- 17-24 turns
permission_denials_count: 1
- no buffered inline comments
- no top-level or inline PR review
Other Ladon reviews in the same repository and time window complete
successfully, so this appears specific to this review path rather than a
repository-wide outage. Refreshing the branch from current main did not
change the outcome.
Expected behavior
The reviewer should either return schema-valid { summary, findings } output
(including an empty findings array for a clean review) or expose a specific,
actionable failure explaining why it cannot do so.
Please preserve the current fail-closed behavior: missing findings must not
be treated as approval.
Useful diagnostic improvement
When permission_denials_count is nonzero and structured output is exhausted,
surface the denied tool/category in the job annotation without exposing
review content or secrets. That would distinguish a reviewer permission
regression from a model formatting failure.
Problem
The Ladon review for adcontextprotocol/adcp#6902 reproducibly completes its
analysis but never returns schema-validated findings. The job correctly fails
closed, but identical reruns cannot produce an actionable review.
Observed on both the original and refreshed PR heads:
Across the attempts, Claude Code reports:
error_max_structured_output_retriespermission_denials_count: 1Other Ladon reviews in the same repository and time window complete
successfully, so this appears specific to this review path rather than a
repository-wide outage. Refreshing the branch from current
maindid notchange the outcome.
Expected behavior
The reviewer should either return schema-valid
{ summary, findings }output(including an empty findings array for a clean review) or expose a specific,
actionable failure explaining why it cannot do so.
Please preserve the current fail-closed behavior: missing findings must not
be treated as approval.
Useful diagnostic improvement
When
permission_denials_countis nonzero and structured output is exhausted,surface the denied tool/category in the job annotation without exposing
review content or secrets. That would distinguish a reviewer permission
regression from a model formatting failure.