forge-lcdl

Handbook

Cheap-model operating guide

This guide describes a practical loop for using forge-lcdl with small / inexpensive models: stay deterministic where the library already is, and reserve fuzzy reasoning for tightly scoped steps.

  1. Contract first — Pick a catalog task_id and version; ensure inputs match the contract. Use CONTRACT-SPEC.md and machine-readable contract.json where present.
  2. Context pack — Call forge_lcdl.context.build_context_pack with a narrow task string and a bounded budget. See CONTEXT-PACKS.md.
  3. Plan patch units — For coding-style work, use forge_lcdl.coding.planner.create_patch_units to emit small, reviewable units. See CODING-EXECUTOR.md.
  4. Verify — Run registered verifiers or subprocess checks; capture VerificationResult objects. See VERIFICATION.md.
  5. Repair — On failure, run classify_failure then reduce_failure_to_repair for deterministic next-step hints. See REPAIR-LOOPS.md.
  6. Proof — Aggregate patch outcomes with proof_report_from_results / proof_report_to_dict. See CODING-EXECUTOR.md.

Cursor and MCP

  • Cursor integration pack: CURSOR-INTEGRATION.md — rules, skills, agents shipped under .cursor/.
  • MCP sidecar: MCP-SIDECAR.md — local-only patterns; prefer least privilege and avoid shipping secrets.

What cheap models cannot safely assume

Small models cannot reliably replace architecture decisions, security review, or broad refactors without tight guardrails. Read KNOWN-LIMITATIONS.md and keep humans in the loop for high-impact changes.

Offline health checks

  • Baseline tasks (fake chat): cheap_model_baseline — BENCHMARKS.md.
  • Library dogfood on this repo: lcdl_dogfood_alpha — DOGFOODING.md.