AI can write a fix.
You should see exactly what changes.
A real command-line repair, from a broken workflow to a reviewable patch. Then watch what happens when the file changes before the patch is exported.
Download the demo and reproduce the checks →
Inspect the actual patch
Evidence, comparison and limits
This is a replay of the September 22, 2026 pipeline-support task. The real recorded model answer is passed through the vendored ANVIL workbench; the verification script reproduces the original failure, runs three behavior tests on the reviewed replacement, and exercises normal and stale-source patch export.
Historical model: Qwen3.8-27B via stock Splash 1.0. Both arms used the same prompt and produced identical code: 944 input tokens and 879 completion tokens each. The direct request took 6.720 seconds; ANVIL took 5.662 seconds. Fixed order and shared cache prevent a causal speed claim. The full shared trial took 45.770 seconds; integration and CI were additional.
No new inference was performed for this replay. The buttons walk through verified results; they do not run a model or edit files. A hash check is point-in-time, not authentication or a lock against concurrent edits. Reviewability is not proof of correctness.
Replay receipt · Provenance · Patch · Original public change