Step 1 — LLM produces structured plan in JSON:
{
"task_type": "code_modification",
"constraint_class": "syntactic_preservation",
"edit_scope": "minimal_diff",
"risk_flags": [],
"proposed_patch": "..."
}Step 2 — Codex consumes only the proposed_patch.
Step 3 — Diff checker validates token preservation.
Step 4 — Unit tests enforce correctness.
Step 1 — LLM produces structured plan in JSON:
{ "task_type": "code_modification", "constraint_class": "syntactic_preservation", "edit_scope": "minimal_diff", "risk_flags": [], "proposed_patch": "..." }Step 2 — Codex consumes only the proposed_patch.
Step 3 — Diff checker validates token preservation.
Step 4 — Unit tests enforce correctness.