Adjudicate-regression takes a candidate, frozen sources, any reproduction and replay packages, and a registry assignment, and decides whether a change introduced a defect. It runs in a fresh context apart from the reproducer and the replayer.
The agent inspects source, setup, PoC and actual logs, rules out incomplete fixes, latent bugs, intentional changes and unreachable scenarios, and confirms only when the same PoC failed after the change and passed before, through real interfaces. The result is a sealed decision record.
What You Get
- A sealed decision.json that cites evidence for each conclusion
- An outcome of confirmed, an evidenced exclusion or unresolved, with a concise reason
- A statement of what remains unproved when the result is unresolved
- A short report back with status, record path, result and any gap
Common Use Cases
- Deciding whether a reproduced failure was introduced by a given change
- Separating regressions from latent bugs and incomplete fixes
- Checking whether candidate findings share one root cause
What Makes It Unique
A defect that passes before and fails after is not enough: exposing an existing defect counts as a latent bug. Confirmation needs the same PoC replayed independently against live before and after states, and unresolved claims are never merged into confirmed ones.
Under The Hood
The agent reads SKILL.md plus a regression definition, report definitions and an evidence file contract in the plugin, then inspects sealed packages, logs and sources. It checks known findings, issues and pull requests for novelty when it can reach them, and writes a sealed JSON decision into the campaign directory. The skill ships no scripts of its own and runs no target code, because reproduction is another stage.