Propose the fix

SkillMedia

Design the smallest correct DeepReason fix for a reproduced cause, as FIX.md. No code changes yet. Use after REPRO.md confirms the diagnosis.

Available today. Use it from your connected AI after setup.

Connect ahel once, and every AI you use reads what you have installed.

Then ask your AI: use the Propose the fix skill

What this skill tells your AI

The instructions your AI receives, as published by ahepi/deepreason in .claude/skills/dr-propose-fix/SKILL.md and read by ahel’s review.

Input: DIAGNOSIS.md + REPRO.md. Output: FIX.md. You may read code freely now, but you still change nothing.

Design rules (DeepReason-specific)

  • Smallest semantic change that makes the reproduction invert. If a comparison is wrong, fix the comparison — do not redesign the data it compares.
  • The record is law. Never change what is WRITTEN to the append-only record to fix what is READ from it. Fix readers (validators, gates, accessors) so old committed roots stay valid. A fix that invalidates existing replay-valid roots is wrong by definition.
  • Frozen surfaces need a flag, not a patch. Anything that alters state digests, event application order, manifest schemas, or qualification subjects is out of a normal tranche: FIX.md must say so and stop for operator approval.
  • Budgets and priorities are guarantees. A scheduling or budget fix must state the guarantee in one sentence ("the operator's seed question always wins rank ties"; "simulation budgets meter only simulation records") and the fix must make that sentence true in EVERY selection/gate path, not just the default one.
  • Counters count one thing. When filtering a pooled collection, filter by type (isinstance) at every consumer, and check for the mirror-image bug in the sibling capability before declaring done — note it in PARKED.md if real.

FIX.md template

# Fix: <one line>
Guarantee restored: <one sentence>
Change sites (exhaustive):
  - <file:line-range> — <what changes, one line each>
Regression artifact: <the REPRO artifact that must invert, plus
  any NEW conditions this fix must be tested against>
Existing tests at risk: <tests whose fixtures assumed the old
  behavior, from grep — name them and say whether each will be
  updated (fixture was defect-dependent) or must keep passing>
Explicitly not changed: <the tempting neighbor, and why>
Estimated diff: <n> lines across <n> files (must be <=150)

Approval gate

  • Class defect (per GOAL.md) with diff estimate <=150 lines and no frozen surface: proceed to dr-implement-fix.
  • Anything else: stop, present FIX.md, await operator direction.

Exit criteria

  • FIX.md committed and pushed. No production code changed.
  • Return to the orchestrator.

Signals

GitHub stars
142
Forks
14
Last commit
Sep 2026
Advanced
Catalog kind
skill
Gateway key
dr-propose-fix
Source
github.com/ahepi/deepreason