ACLA Agentic Context Runtime v4

ACLA answers one question for an AI agent: what should the model know for this mission?

Persistent Context -> Context Compiler -> Mission Working Set (minimum relevant, current, authoritative) -> Model

An AI agent should remember far more than it processes.

Quick start

pip install "acla-agentic-context-runtime[mcp]"
acla init
acla project add MYPROJECT --workspace /path/to/workspace
acla provider add local --type ollama --model qwen3:8b      # optional; SHADOW mode needs no provider
acla start
acla status
acla mission run "What is the current blocker?" --project MYPROJECT
acla stop

Everything runtime-specific lives under ACLA_HOME (default ~/.acla); nothing machine-specific is embedded in the source. See docs/GETTING-STARTED.md, SECURITY.md, RELEASE.md and CHANGELOG.md.

Runtime engine background (V3 foundation)

A governed runtime extension of the ACLA Context Compiler research artifact.

V3 preserves the deterministic Context Compiler kernel and V2 runtime contracts, while adding provider health/circuit breakers, measured latency/cost routing, bounded adaptive routing, adapter conformance, heterogeneous read-only source intake, queue/workers, MCR capability binding, EAF authorization/outcome assurance, governed autonomous tool execution and a read-only operations console.

Core authority model

• ACLA Context Compiler decides what information enters model-active context.
• Inference Routing selects among already policy-eligible provider/model routes.
• Adaptive Routing may adjust bounded preference weights; it cannot rewrite governance gates.
• MCR Binding binds requested semantic capabilities to eligible tool/provider routes.
• EAF Authorization authorizes actions. Mutating tools require an explicit single-use mission/action grant.
• Outcome Assurance requires authoritative action evidence when configured.
• UI is read-only and never creates execution authority.

Engines

See docs/ENGINE-MATRIX.md and docs/architecture.md.

Run a local preview

python -m acla_context_compiler.cli run examples/v3-runtime-mission.json

Queue/worker reference

python -m acla_context_compiler.cli queue-enqueue runtime.db examples/v3-runtime-mission.json --job-id DEMO-1
python -m acla_context_compiler.cli worker-once runtime.db --worker-id worker-1

The SQLite backend is a reproducible reference queue, not a production distributed broker certification.

Read-only operations UI

python -m acla_context_compiler.cli ui --host 127.0.0.1 --port 8789

Rich source formats

Core/stdlib formats: .md, .txt, .json, .csv, .html, .htm, .xml.

Optional-extra formats: .yaml, .yml, .docx, .pdf.

Install the optional readers with the package extra rich-sources in an environment allowed to resolve dependencies.

Safety / qualification boundary

• No live commercial provider calls are included in acceptance.
• No SAP reads or writes are performed.
• No enterprise write adapter is shipped.
• No model output can create write authority.
• No unsigned PowerShell/CMD/BAT/shell operational launch scripts are shipped.
• Trusted digital signing remains a public promotion gate because the author's signing key/certificate is not present in this build environment.

Paper relationship

V3 is an additive runtime artifact. It does not retroactively alter historical paper measurements, and local/synthetic qualification is not a claim of universal cost or latency improvement.