CIE Integration

What CIE is

CIE — the Code Insight Engine is a separate code-graph engine by the same author. forge can consume it as its code-graph backend, served as a real MCP server over stdio — the same way Claude Code / Cursor consume tools. Division of labor:

	Responsibility
CIE	Localization + blast radius: callers, affected_by, failing_context, file_skeleton, …
forge	draft tasks → generate → test → sample → select → gate → commit

forge itself stays unchanged — CIE is another ToolBackend, exactly like the bundled local/graph backends (Architecture).

Install + one-line benchmark

pip install git+https://github.com/arunsoman/cie.git pytest
python benchmarks/cie_forge_realbugs/forge_cie_bench.py boltons_bits_offbyone

Needs a tool-calling model at http://localhost:11434/v1 (Ollama default); override with FORGE_MODEL / FORGE_BASE_URL / FORGE_API_KEY.

What it changes

• Localization & blast radius come from a real index — multi-hop callers/affected_by, failing_context, file_skeleton — instead of the built-in per-process index. Measured effect: the same agent with CIE fixed a planted bug in ~63% fewer tokens; without the graph it broke the suite and did not converge (Benchmarks).
• CIE generates regression tests grounded in real signatures — testgen.py then validates the oracle: the test must fail on the buggy code (assertion, not import error) before repair starts (Issue-to-PR).

Where it's required vs optional

• fix / fix-comment (Issue-to-PR): required.
• run / repair (Repair-Loop): optional; the built-in LocalToolBackend/GraphToolBackend work standalone.

Methodology and raw numbers: Benchmarks; harness in benchmarks/cie_forge_realbugs/.