The AI thinks, the gate decides — how I made LLM code edits deterministic (and cut token usage 42×)

Описание и применение

D‑Engine снижает токенов на 44× и предотвращает некорректные коммиты, разделяя «мышление» LLM и «действие» локального гейта.

Задача

AI coding agents consume excessive tokens and risk breaking code due to lack of memory, state drift, and atomic writes.

Результат

D‑Engine achieved token usage 44× lower than dsh, comparable quality (48/50 points), and prevented unsafe commits, demonstrating that a stricter gate outperforms smarter agents.

Рабочий процесс

• 1. Selector step: LLM picks minimal relevant files (≈300 tokens).
• 2. Proposal step: LLM generates SEARCH/REPLACE patches (≈2,000 tokens).
• 3. LocalEditor applies patches via 4‑strategy cascade.
• 4. Compiler validates patched copy.
• 5. commitAndMerge stages only touched files if gate green.
• 6. (Optional) Verify mode: send modified snippet for semantic audit (330–984 tokens).

Доказательства

D-Engine (my harness) 2,552 tokens, ~4 s vs dsh 107,000 tokens, 28 s

Источник

The model receives only those files and generates the SEARCH/REPLACE blocks.

Источник

D-Engine Fast ~2,100 tokens per task, dsh ~93,000 tokens for same set.

Источник