| name | ablation-test |
| description | Runs a disciplined ablation workflow for MixJam Electron (MJE) bugs to prove which changed files or layers are actually required for a fix. Use when an MJE import, playback, or UI bug has a known pass/fail check but several edits may be involved, or when the user asks which change actually fixed it, what can be removed, or for the smallest proven fix.
|
| argument-hint | Optionally include the repro check, candidate paths, and any restart/build constraints.
|
Ablation Test
Use this after diagnose when MJE
already has a discriminating PASS/FAIL check and the remaining question is
which candidates — the changed files or groups — are actually required.
Candidate Rules
- Preserve — save user work before slicing candidates.
- Scope — ablate only relevant candidates; do not disturb unrelated dirty state.
- Isolate — change one candidate group at a time and keep the same
validation check. A candidate group is a set of files that map to one
project seam (engine, state, bridge, UI, config, tests).
- Log — record each run in
tmp/ablation-<slug>.md; use a markdown
template. Prefer
scripts/New-AblationLog.ps1; use the
manual template in EXAMPLES.md when the script is unavailable.
- Confirm — stop once the minimal candidate set is proven, then rerun
the focused check from a clean state.
Output
Report:
- root cause
- minimal required candidate set
- candidates proven unnecessary
- confidence level and any residual uncertainty
Completion Criterion
The ablation run is complete when:
- the initial hypothesis and candidate groups are recorded,
- each candidate group has been tested with the same validation check,
- the minimal candidate set that still satisfies the pass/fail check is identified,
- unnecessary candidates are listed with evidence, and
- the result is captured in
tmp/ablation-<slug>.md.
Deep Reference
Use REFERENCE.md for the full workflow and safe Git commands.
Use EXAMPLES.md for repo-shaped ablation scenarios.
