Skillget

cognitive-project-layer

mcp
Guvenlik Denetimi
Uyari
Health Uyari
  • License — License: NOASSERTION
  • Description — Repository has a description
  • Active repo — Last push 0 days ago
  • Low visibility — Only 5 GitHub stars
Code Uyari
  • network request — Outbound network request in assets/dashboard.js
Permissions Gecti
  • Permissions — No dangerous permissions requested

Bu listing icin henuz AI raporu yok.

SUMMARY

Local-first context engine for coding agents: project skeleton, symbols, references, graph, retrieval, confidence, MCP, CLI, and HTTP API.

README.md

Cognitive Project Layer

CI
License
Rust

Local-first context engine for coding agents.

Cognitive Project Layer (CPL) gives an agent a structured, inspectable view of a
codebase: project skeleton, symbols, references, graph relations, local retrieval,
confidence scoring, and fallback tools exposed through CLI, MCP stdio, and an
optional local HTTP API.

The goal is not "RAG instead of understanding". The goal is a predictable context
pipeline for agents:

scan -> skeleton -> adaptive budget -> symbols/references -> grep -> vector search
     -> graph expansion -> confidence -> managed context -> fallback plan

Features

  • Fast project scanner with ignored build/cache/vendor folders.
  • Root .gitignore and .cplignore support for large/generated trees.
  • Always-on Skeleton prompt: entry points, modules, configs, public API, recent changes.
  • Tree-sitter-first symbol index for Rust, TypeScript/TSX, JavaScript, Python, C++, Go.
  • Regex fallback for niche/unsupported syntax, including ArkTS/HarmonyOS components.
  • References/usages index.
  • Impact/affected analysis for symbol blast radius, git diff/status, likely changed-file test targets, and verification plans.
  • Code-aware chunks with stable IDs and inferred line ranges.
  • Local TF-IDF vector search with no network dependency.
  • Persistent structural SQLite index in .cpl/index.sqlite.
  • Warm-start from fresh SQLite structural indexes.
  • Incremental SQLite index refresh with rebuild fallback.
  • Index freshness diagnostics, MCP/HTTP/watch auto-refresh, cpl doctor, and safe cpl heal.
  • Persistent embedding DB in .cpl/vectors.sqlite with legacy .cpl/vector_db.json fallback.
  • Lazy SQLite vector loading with DB-backed streaming embed-search.
  • Incremental embedding refresh by changed chunk path.
  • Embedding backends:
    • local-hash offline default;
    • Ollama local neural embeddings;
    • OpenAI via OPENAI_API_KEY;
    • OpenAI-compatible endpoints.
  • Qdrant external vector backend adapter.
  • Structural graph: files, modules, configs, imports, call-ish references, tests.
  • Confidence engine with RagOnly, Hybrid, and FallbackExplore strategies.
  • Adaptive retrieval/explore budget, context budget manager, and text transparency panel.
  • CLI cpl, MCP stdio server cpl-mcp, Python MCP wrapper, and local HTTP API.
  • Built-in local dashboard UI for health, retrieval, refresh actions, and eval history.
  • File watcher and background refresh worker.
  • Self-healing local maintenance for stale SQLite indexes and embedding DBs.
  • Fixture-based retrieval evals and CLI latency benchmarks.
  • Optional ArkTS/HarmonyOS profile; see docs/PROFILES.md.

Quick install

Windows PowerShell:

irm https://raw.githubusercontent.com/kharkilirov1/cognitive-project-layer/main/install.ps1 | iex

Linux/macOS:

curl -fsSL https://raw.githubusercontent.com/kharkilirov1/cognitive-project-layer/main/install.sh | sh

Verify:

cpl --version
cpl scan --root .

More options: docs/INSTALL.md.

Install prebuilt binaries

Download the latest archive for your OS from
releases/latest:

  • linux-x86_64
  • windows-x86_64
  • macos-x86_64
  • macos-aarch64

Each release also includes SHA256SUMS. The archives contain:

  • cpl
  • cpl-mcp
  • README.md
  • LICENSE
  • NOTICE
  • CHANGELOG.md
  • docs/
  • install.sh
  • install.ps1

Put cpl and cpl-mcp on your PATH, or run them from the extracted folder.

Install from source

Requirements:

  • Rust 1.85+.
  • Optional: Ollama if you want local neural embeddings.
git clone https://github.com/kharkilirov1/cognitive-project-layer.git
cd cognitive-project-layer
cargo build --bins

Install directly with Cargo:

cargo install --git https://github.com/kharkilirov1/cognitive-project-layer.git

The binaries are:

  • target/debug/cpl
  • target/debug/cpl-mcp

Quick start

cargo run -- scan --root .
cargo run -- skeleton --root .
cargo run -- symbols --root . retrieve
cargo run -- retrieve --root . "Where is retrieve implemented?"
cargo run -- context --root . --max-tokens 64000 "Why does the build fail around hilog?"
cargo run -- panel --root . "architecture retrieval"
cargo run -- index-build --root .
cargo run -- index-refresh --root .
cargo run -- index-db --root .
cargo run -- doctor --root .
cargo run -- heal --root .

After cargo install --git, use the installed binary:

cpl scan --root .
cpl retrieve --root . "Where is retrieve implemented?"
cpl heal --root .
cpl init --root . --server native

Self-heal local project state:

cpl heal --root .
cpl doctor --root . --fix
cpl heal --root . --embeddings ensure --embedding-backend local-hash --embedding-dimensions 1536

heal is intentionally local and safe: it refreshes/rebuilds .cpl/index.sqlite,
refreshes an existing .cpl/vectors.sqlite by default, and only creates missing
embeddings when --embeddings ensure is requested. Embedding refresh skips
potentially external backends unless you explicitly pass an embedding backend.

Build local embeddings:

cargo run -- embed-index --root . --backend local-hash --dimensions 1536
cargo run -- embed-refresh --root . --backend local-hash --dimensions 1536
cargo run -- vector-db --root .
cargo run -- embed-search --root . "project graph retrieval" --limit 10

Build the structural SQLite index:

cargo run -- index-build --root .
cargo run -- index-refresh --root .
cargo run -- index-search --root . "validate token"
cargo run -- index-db --root .

Persistence details: docs/PERSISTENCE.md.

Use Ollama embeddings:

ollama pull nomic-embed-text
cargo run -- embed-index --root . --backend ollama --model nomic-embed-text --dimensions 768

Use OpenAI embeddings:

$env:OPENAI_API_KEY="..."
cargo run -- embed-index --root . --backend openai --model text-embedding-3-small

MCP / OpenCode

Generate an OpenCode MCP config for a project:

cargo run -- init --root . --server native

This writes a local opencode.json and ensures local/private state is ignored:

  • /.cpl
  • .env
  • .env.*
  • opencode.json

A portable example is in examples/opencode.native.json.

Details and examples: docs/OPENCODE.md.

HTTP API

cargo run -- serve --root . --host 127.0.0.1 --port 3878

Endpoints:

GET  http://127.0.0.1:3878/health
GET  http://127.0.0.1:3878/ui
GET  http://127.0.0.1:3878/scan
GET  http://127.0.0.1:3878/skeleton
GET  http://127.0.0.1:3878/graph
GET  http://127.0.0.1:3878/retrieve?query=symbol_lookup
GET  http://127.0.0.1:3878/context?query=auth%20login&max_tokens=64000
GET  http://127.0.0.1:3878/symbols?query=retrieve
GET  http://127.0.0.1:3878/references?symbol=retrieve
GET  http://127.0.0.1:3878/impact?symbol=retrieve&depth=2
GET  http://127.0.0.1:3878/affected?files=src/lib.rs,tests/foo.rs&filter=*_test.rs
GET  http://127.0.0.1:3878/affected?git=true&include_untracked=true&filter=*_test.rs
GET  http://127.0.0.1:3878/verify-plan?git=true&full=false
GET  http://127.0.0.1:3878/embed-search?query=opencode%20mcp&limit=5
POST http://127.0.0.1:3878/embeddings/rebuild
POST http://127.0.0.1:3878/embeddings/refresh
POST http://127.0.0.1:3878/heal
POST http://127.0.0.1:3878/index/rebuild
POST http://127.0.0.1:3878/index/refresh
GET  http://127.0.0.1:3878/index-db
GET  http://127.0.0.1:3878/index/freshness
GET  http://127.0.0.1:3878/index/search?query=validate%20token
GET  http://127.0.0.1:3878/doctor
GET  http://127.0.0.1:3878/benchmarks
GET  http://127.0.0.1:3878/tree?depth=3
GET  http://127.0.0.1:3878/grep?pattern=EntryAbility

Security note: keep the HTTP API bound to 127.0.0.1 unless you intentionally
want another process or machine to access your project context.

Dashboard details: docs/UI.md. The dashboard includes an
interactive graph explorer with search, node/edge highlighting, minimap,
force-layout pass, local maintenance actions, retrieval panels, and benchmark
history from .cpl/eval-results/*.json.

Optional project config lives at .cpl/config.toml; see
docs/CONFIG.md and examples/cpl.config.toml.

Public retrieval benchmark

CPL can run a public CodeSearchNet retrieval eval using
mteb/CodeSearchNetRetrieval:

cargo build --release --bins
python scripts/eval_public_codesearchnet.py `
  --cpl .\target\release\cpl.exe `
  --language python `
  --limit 100 `
  --mode http `
  --min-recall10 0.90 `
  --min-ndcg10 0.70 `
  --json-out .cpl\eval-results\public-codesearchnet-python.json

Recent local release run on the first 100 Python queries:

  • Recall@1: 0.53
  • Recall@3: 0.71
  • Recall@5: 0.79
  • Recall@10: 0.90
  • MRR: 0.645
  • NDCG@10: 0.705

Local latency still depends on hardware; these quality metrics are the portable
part of the benchmark.

CLI overview

cpl scan
cpl skeleton
cpl symbols [query]
cpl retrieve <query...> [--disable-stage adaptive-budget]
cpl context [--max-tokens N] <query...>
cpl index
cpl index-build
cpl index-db
cpl index-freshness
cpl index-search <query...>
cpl index-refresh [--max-incremental-files N]
cpl doctor
cpl doctor --fix
cpl heal [--embeddings off|existing|ensure]
cpl graph
cpl impact <symbol> [--depth N] [--limit N]
cpl affected [file...] [--stdin|--git] [--base REV] [--no-untracked] [--depth N] [--filter PATTERN]
cpl verify-plan [file...] [--stdin|--git] [--base REV] [--no-untracked] [--full]
cpl chunks [query]
cpl embed-index
cpl embed-refresh [--max-incremental-paths N]
cpl embed-search <query...>
cpl vector-db
cpl qdrant-upsert
cpl qdrant-search <query...>
cpl references <symbol>
cpl panel [query...]
cpl nav <section> [filter]
cpl git-status
cpl git-diff [range]
cpl tree --depth 3
cpl grep <pattern>
cpl watch
cpl serve
cpl init

Architecture

User query
  -> Query analyzer
  -> Skeleton always included
  -> Adaptive budget based on repo size and query intent
  -> Symbol lookup
  -> References/usages index
  -> Lexical search / grep
  -> Code-aware vector search
  -> Persistent embedding DB search
  -> Structural graph expansion
  -> Merge ranking
  -> Confidence engine
  -> Context budget manager
  -> Agent context / fallback plan / transparency panel
  -> Working memory update

Repository layout

src/
  lib.rs             CognitiveProjectLayer facade
  main.rs            CLI
  scanner.rs         project scan, mode selection, entry/config detection
  skeleton.rs        skeleton data model and prompt renderer
  symbols.rs         Tree-sitter-first symbol index + regex fallback
  ast.rs             Tree-sitter parsing helpers
  references.rs      references/usages index
  graph.rs           structural project graph
  impact.rs          symbol blast-radius, git changed-file collection, and affected analysis
  verify_plan.rs     verification command planner from affected/git state
  retrieval.rs       adaptive-budgeted hybrid retrieval pipeline
  confidence.rs      confidence engine
  doctor.rs          local diagnostics
  tools.rs           fallback tools
  memory.rs          working memory
  budget.rs          context budget manager
  chunk.rs           rich code-aware chunks
  vector.rs          local TF-IDF vector store
  embedding.rs       embedding providers
  persistent_index.rs SQLite structural index
  persistent_vector.rs SQLite persistent vector DB
  qdrant.rs          Qdrant adapter
  dashboard.rs       embedded local dashboard UI and benchmark summaries
  mcp_server.rs      MCP stdio server
  http_server.rs     local HTTP API
  watcher.rs         file watcher daemon
  background.rs      background refresh worker
scripts/
  eval_retrieval.py    fixture retrieval eval runner
  bench_cli.py         CLI latency benchmark runner
  bench_mcp.py         MCP stdio warm benchmark runner
  bench_large_repo.py  synthetic large-repository benchmark
  check_bench_thresholds.py benchmark regression gate
evals/
  retrieval.json       retrieval eval cases
  fixtures/            small Rust, TypeScript, and ArkTS projects
docs/
  INSTALL.md
  PERSISTENCE.md
  SCALE.md
  UI.md
  EVALS.md
  PROFILES.md
  OPENCODE.md
  SPEC.md
  ROADMAP.md

Development

cargo fmt --check
cargo test
cargo clippy --all-targets -- -D warnings
cargo build --bins

Retrieval evals and CLI benchmarks:

python scripts/eval_retrieval.py --cpl .\target\debug\cpl.exe --top-k 3
python scripts/bench_cli.py --cpl .\target\debug\cpl.exe --iterations 3
python scripts/bench_mcp.py --mcp .\target\debug\cpl-mcp.exe --iterations 3

Details: docs/EVALS.md.

Do not commit local project cognition state:

  • .cpl/
  • target/
  • .env*
  • logs and caches

License

Apache-2.0. See LICENSE and NOTICE.

Yorumlar (0)

Sonuc bulunamadi