ace-lite-engine 0.3.84

Local-first Active Context Engine for AI coding

ACE-Lite Engine

Local-first, MCP-native Active Context Engine for AI coding.

ACE-Lite avoids "dump the whole repository into the model" workflows. Instead, it runs a deterministic pipeline to distill structure, retrieve candidates with hybrid rankers, and produce a source plan (steps + constraints + selected files/chunks) you can execute and review.

If you want full-repo packing with minimal setup, consider tools like Repomix. ACE-Lite is optimized for top-k relevant context with benchmarkable retrieval quality.

Core capabilities

• Deterministic pipeline: memory -> index -> repomap -> augment -> skills -> source_plan
• Hybrid retrieval defaults: multiple rankers + fusion (lexical / semantic / graph / worktree / feedback), with fail-open behavior
• Multi-language indexing (tree-sitter): Python / TypeScript+TSX / JavaScript / Go / Solidity / Rust / Java / C / C++ / C# / Ruby / PHP
• Skill lazy-loading: route Markdown skills by query intent, load section-level content only
• Dual-channel memory: MCP primary with REST fallback
• RepoMap generation: token-budgeted repository map (json + md)
• Plugin hooks: before/after stage hooks via in-process and MCP runtime
• Benchmark automation: quality scoring, baseline comparison, and regression gates
• Layered config: home -> repo -> cwd -> CLI

When to use ACE-Lite

• Use ACE-Lite when you want a plan + curated context (top-k files/chunks) with deterministic, regression-tested selection.
• Use repo packers when you want full-repo context quickly and accept lower precision.
• Use DB-backed vector search when you need always-on semantic search at scale and accept backend complexity.

Positioning

ACE-Lite is designed for the middle ground between "pack the whole repo into the prompt" and "deploy a heavyweight managed context platform."

Approach	Common shape	Tradeoff	ACE-Lite position
Repo packers	Single-file repository export or large prompt stuffing	Fast setup, but token cost and context precision can degrade on larger repos	ACE-Lite retrieves and ranks only the most relevant files/chunks, then emits a source plan
Managed context engines	Hosted or platform-oriented code search / semantic retrieval systems	Powerful at scale, but usually heavier to operate or adopt	ACE-Lite stays local-first and file-based, with no required external vector DB
Memory layers	Persistent user/project memory across sessions	Good for recall, but memory alone does not produce an execution plan	ACE-Lite includes memory, but centers the workflow on actionable code retrieval and planning

This is also why ACE-Lite is not trying to be a full IDE or an all-in-one coding agent. Its job is narrower: give any capable model or agent a better, smaller, and more testable working set.

Why It Stands Out

• MCP-native: ACE-Lite exposes its core capabilities through MCP, which makes it usable from modern MCP-compatible clients and agent runtimes.
• Benchmarkable retrieval: ACE-Lite treats retrieval quality as something you can measure in CI, not just judge by feel.
• Local-first: the default workflow does not require a hosted backend or a separate search cluster.
• Action-oriented output: the end product is a source plan with selected files, chunks, and constraints, not just raw retrieved text.

Competitive Notes

The high-level competitive framing is directionally right, with a few important caveats:

• Repo packing is a real comparison point, but not every local coding tool is "full dump into prompt only." Some tools also maintain repo maps or lightweight retrieval layers.
• Managed products such as Augment Code and Sourcegraph do far more than simple vector search; they combine search, code intelligence, and platform integrations.
• Memory products such as Mem0 address a different layer of the stack. They are complementary more often than they are direct substitutes.

The stronger open-source claim for ACE-Lite is not "it beats every alternative everywhere." The stronger claim is that it offers a rare combination of MCP integration, deterministic retrieval, local-first operation, and built-in regression testing for context quality.

Installation

python -m venv .venv
.venv\Scripts\activate
pip install -e .[dev]

Quickstart

See docs/README.md and docs/guides/GETTING_STARTED.md for the fastest path.

One-command demo (after install):

ace-lite demo

Build index:

ace-lite index --root . --output context-map/index.json

30-second checks:

ace-lite doctor

Build RepoMap:

ace-lite repomap build --root . --ranking-profile graph --output-json context-map/repo_map.json --output-md context-map/repo_map.md

Run planner:

ace-lite plan --query "trace how benchmark thresholds are applied" --repo ace-lite-engine --root . --skills-dir skills

Local profile and memory workflows:

# Manage local profile facts
ace-lite profile add-fact "Prefer deterministic patches" --confidence 0.9
ace-lite profile show

# Store/search/wipe local memory notes with namespace isolation
ace-lite memory store "Fix auth fallback behavior" --namespace repo:ace-lite --tag kind=bugfix
ace-lite memory search "fix auth" --namespace repo:ace-lite --limit 5
ace-lite memory wipe --namespace repo:ace-lite

Service-mode runtime utilities:

# Watch layered config changes and validate reloads
ace-lite runtime watch-config --root . --max-polls 10 --poll-interval-seconds 1

# Run heartbeat + cron scheduler from `.ace-lite.yml` runtime section
ace-lite runtime run-scheduler --root . --max-ticks 5 --tick-interval-seconds 60 --simulate-clock

Run benchmark with regression gate:

ace-lite benchmark run --cases benchmark/cases/default.yaml --repo ace-lite-engine --root . --warmup-runs 1 --output artifacts/benchmark/latest --baseline artifacts/benchmark/baseline.json --benchmark-threshold-profile strict --dependency-recall-floor 0.8 --chunk-hit-tolerance 0.02 --chunk-budget-growth-factor 1.15 --validation-test-growth-factor 1.5 --fail-on-regression
ace-lite benchmark run --cases benchmark/cases/default.yaml --repo ace-lite-engine --root . --warmup-runs 1 --no-include-plans --no-include-case-details --output artifacts/benchmark/ci-fast
ace-lite benchmark report --input artifacts/benchmark/latest/results.json

PowerShell helpers:

./scripts/run_smoke.ps1
./scripts/run_benchmark.ps1
./scripts/run_benchmark_tabi_v3.ps1

Release validation

Multi-repository matrix run (3+ repos):

python scripts/run_benchmark_matrix.py --matrix-config benchmark/matrix/repos.yaml --output-dir artifacts/benchmark/matrix/latest --fail-on-thresholds

Skill routing validation across supported CLI hosts:

python scripts/run_skill_validation.py --repo-url https://github.com/blockscout/frontend.git --repo-ref main --repo-name blockscout-frontend --repo-dir artifacts/repos-workdir/skill-validation --skills-dir skills --output-path artifacts/skill-eval/blockscout_skill_validation_matrix.json --index-cache-path artifacts/skill-eval/blockscout-index.json --languages typescript,javascript --apps codex,opencode,claude-code --min-pass-rate 1.0 --fail-on-miss

Release freeze bundle:

python scripts/run_release_freeze_regression.py --matrix-config benchmark/matrix/repos.yaml --output-dir artifacts/release-freeze/latest --fail-on-thresholds

Repository layout

• src/ace_lite/ runtime engine and CLI
• tests/ pytest test suite
• skills/ Markdown skill manifests and sections
• benchmark/ cases, matrix config, scoring contracts
• scripts/ smoke/benchmark helper scripts
• docs/ architecture and operational docs

Index filtering (.aceignore)

Create a .aceignore file at the repository root to exclude low-signal paths from indexing and retrieval.

More examples: docs/reference/ACEIGNORE_COOKBOOK.md.

Example for Go repos with generated code:

# Protobuf output
*.pb.go

# Common generated naming
*_generated.go
*_gen.go

# Go mocks
mock_*.go
*_mock.go

# Vendor
vendor/

# Fixtures
testdata/

# Common contract binding locations (abigen output)
pkg/contract/

Documentation

• Start here: docs/README.md
• Getting started: docs/guides/GETTING_STARTED.md
• MCP setup: docs/guides/MCP_SETUP.md
• Architecture overview: docs/design/ARCHITECTURE_OVERVIEW.md
• Maintainer docs: docs/maintainers/README.md
• Reference docs: docs/reference/ACEIGNORE_COOKBOOK.md

Known limitations

• Language symbol extraction is best-effort for some grammars (especially C/C++); ranking still falls back to path/module/content signals.
• Embeddings are optional, but when enabled you must keep the embedding model + dimension consistent with your vector store/collection.
• Design and maintainer notes live under docs/design/ and docs/maintainers/.

Governance

• License: LICENSE
• Contributing guide: CONTRIBUTING.md
• Code of conduct: CODE_OF_CONDUCT.md
• Security policy: SECURITY.md
• Changelog: CHANGELOG.md

Support

• Security reports: see SECURITY.md
• General issues/feature requests: open a GitHub issue with reproduction details