D-Engine

The AI thinks, the gate decides.

A deterministic harness for LLM code edits. The model proposes SEARCH/REPLACE blocks; a local runtime applies them in a shadow git worktree, checks the result with tsc --noEmit, and merges only if compilation passes. The LLM never writes to your real branch.

📄 Read the full story: The AI thinks, the gate decides — how I made LLM code edits deterministic (42× fewer tokens) 🇪🇸 README en español

Why

Agentic coding loops pay their token bill twice: every turn re-sends the whole trajectory (files, tool results, reasoning) to the model, and "thinking" modes multiply both the tokens per turn and the number of turns. D-Engine removes the loop for the bounded-edit case: one prompt, one patch set, one deterministic gate.

Same task, same model (V4.1-Flash), same literal prompt:

The toolbox is not the cost — the agentic loop is.

How it works

1. You describe the change; an automatic selector (P9) picks the target files by token budget.
2. The LLM returns SEARCH/REPLACE blocks — nothing else.
3. A 4-strategy cascade applies them locally: exact → normalized newlines → ignore trailing whitespace → fuzzy (0.85 threshold).
4. Everything happens in a shadow git worktree. tsc --noEmit is the only source of truth.
5. Green compile → merge. Anything else → nothing touches your branch.

Benchmark (headline)

10 frozen tasks, same model family, one attempt each, full methodology disclosed:

Tied quality, 14–42× fewer tokens, a predictable bill, zero broken merges. The full benchmark — frozen prompts, voided rows, hallucination incident and all — ships in docs/benchmark.en.md (español).

Install & usage

Requirements: Node.js, git, TypeScript in the target repo (tsc --noEmit).

git clone https://github.com/corruchaga/D-Engine
cd D-Engine
npm install
cp .env.example .env

Fill in LLM_BASE_URL, LLM_API_KEY and LLM_MODEL in .env.

Development (no build):

npm run dev

Local production:

npm run build
npm start

The TUI asks for: change description, target file (empty Enter = automatic selector), safety mode (Fast / Verify / Shadow) and confirmation. Fast merges after applying; Verify adds a semantic audit; Shadow never merges without permission.

npm run typecheck is equivalent to tsc --noEmit.

Known limitations

• Cannot create new files (yet — see roadmap).
• Fuzzy matching (0.85 threshold) is the weak link.
• The P9 file selector is not deterministic.

Roadmap v0.3

• New-file creation
• Retry loop with compiler feedback
• Mandatory Verify when a patch applied via fuzzy
• Time and tokens in the final summary

License: MIT