v1.0.0 — the right files, measured on repositories it never saw

Ship the right code, not the whole repo. The first release of code-context-engine as its own project: a local-first MCP context engine that answers "which files does this task need?" from a per-project code graph, and hands your agent a folded evidence packet instead of whole files.

The numbers

Every figure below is reproducible with one command (bash scripts/benchmark-holdout.sh) on pinned third-party checkouts, against gold files written by reading the source before the engine ran. Full table and method: docs/measured.md.

Repositories the engine was never tuned on — every gold file reaches the packet:

Holdout	Stack	Recall	Precision	Forbidden
gin + torchvision	Go, Python	1.000	0.700	0
libuv + fmt	C, C++	1.000	0.589	0
os-lib + cli + Flux.jl	Scala, R, Julia	1.000	0.589	1
peft + keras-hub	Hugging Face, Keras 3	1.000	0.632	0

Against the engine this was forked from (same 112 tasks, same machine, same day): recall 0.735 → 0.938, precision 0.206 → 0.633.

In a real Claude Code session (cost from the CLI's own usage report): on Django, 3,516 files — 45% fewer context tokens, 15% less money, 8 turns down to 5.

With a real model answering and a separate judge grading against the real source, patch tasks verified by applying the patch and running the tests: the packet scored 1.000 task success on a Go holdout (whole gold files: 0.900) and 0.800 on a vision holdout (0.700), at roughly half the tokens.

What is in it

• Project isolation — deterministic ProjectId, single-project invariant on the live graph, a CI gate that indexes two projects sharing a file and asserts zero leakage.
• Evidence, not similarity — seeds resolve from the task as written (identifiers, quoted literals, routes, config keys, path words, body words); a file enters the packet because something points at it, never because a score was tuned.
• Config → code — YAML / JSON / argparse / Hydra keys are graph nodes; "the masks flag" reaches both the file that defines it and the code that reads it.
• ML and notebook graph — Dataset · Model · Checkpoint · Metric · TrainLoop from a PyTorch/Keras overlay; .ipynb parsed cell by cell.
• 12 tree-sitter languages plus YAML, JSON, SQL, shell, PowerShell and notebooks.
• Deterministic packets — every tie has a total order; the same question on the same graph gives the same packet on a loaded CI runner and an idle laptop.
• Folding — bodies collapse to signatures until the agent asks for one, so a 12k-token file costs a few hundred.

Install

Pre-built binaries for Linux x86_64, macOS (Apple Silicon) and Windows x86_64 are attached below.

curl -fsSL https://raw.githubusercontent.com/ParsaVictor/code-context-engine/main/install.sh | sh

irm https://raw.githubusercontent.com/ParsaVictor/code-context-engine/main/install.ps1 | iex

Then, from the root of the project you want to index:

neuromesh connect --global --agent-rules   # write MCP config for Cursor / Claude Code / VS Code / Zed
neuromesh index                            # build the graph — seconds on most repositories

Credit

An independent derivative of NeuroMesh v0.9.0 by yoosef alipour (MIT), baseline tag baseline-v0.9.0. Fixes that apply to the original are contributed back upstream. See NOTICE and ATTRIBUTION.md.

Full changelog: docs/CHANGELOG.md · commits since the baseline