cruciblelisted
Install: claude install-skill Udhay-Adithya/skills
# Crucible
A loop that makes a change more correct before, or sometimes after, it becomes a pull
request. One reviewer, several rounds, a written ledger, and a hard budget.
Make the smallest changes that answer the findings. Watch for scope creep, and push
back on findings that do not deserve a fix.
## The reviewer
The reviewer must not be the model doing the work. A model reviewing its own diff
shares every blind spot that produced it, which is the whole reason this loop exists.
In rough order of preference:
1. A subagent on a different provider's flagship model, where this harness can reach
one.
2. A subagent on a different model from the same provider, started with the `Agent`
tool and a `model` override, at high effort.
3. The harness's own review command at its highest effort.
Pick the latest model unless a version was named. Never hardcode a model ID that will
go stale. Record provider, model and effort if the harness exposes them, and say "not
exposed" rather than guessing.
One reviewer only. The reviewer must not start subagents of its own, and must not edit,
commit, push or comment.
State which tier was used in the report. A round 2 reviewer on the same model as the
worker is worth saying out loud, because it is a weaker signal than it looks.
## Budget
- Default 3 rounds. Round 1 is full. Later rounds look only at changed code and open
findings.
- Extend to 5 only for security, permissions, auth, data loss, or state and
concurrency, and only whil