cog
7e8cb4a2 - fix: silently drop unknown prediction inputs instead of rejecting them (#2943)

Commit
38 days ago
fix: silently drop unknown prediction inputs instead of rejecting them (#2943) * fix: silently drop unknown prediction inputs instead of rejecting them Coglet was injecting additionalProperties: false into the JSON Schema validator, causing 422 errors for any input field not in the model's schema. This broke backwards compatibility -- Replicate's API has historically ignored unknown inputs, and deployed models rely on that. Strip unknown fields before validation instead. Stripped field names are logged at warn level for observability. * fix: eliminate TOCTOU race in training input validation fallback Inline the predict-validator fallback instead of re-entering through strip_and_validate_input. The old code dropped the train_validator lock before acquiring the input_validator lock, allowing a schema swap between the two acquisitions. * chore: update rustls-webpki to 0.103.12 (RUSTSEC-2026-0098, RUSTSEC-2026-0099) * chore: update mise.lock
Author
Parents
Loading