Bound what gets loaded
Admit required and relevant context under an explicit budget instead of feeding the entire workspace into every decision.
Local-first cognitive runtime
Make agents prove what worked.
Persistent memory, bounded context, tool/model routing and verified execution — organized as an inspectable control plane instead of a black-box chain of tool calls.
281/281 Python testsWindows · Ubuntu · macOS145/145 governance testsApache-2.0
git clone https://github.com/robertoatila/jarvis-skill-registry.git
cd jarvis-skill-registry
python jarvis.pyThe missing layer is judgment.
Most agent stacks can call a model and execute a tool. Long-lived systems also need to decide what information deserves context, what authority exists, what evidence proves success and what should survive into memory.
Verify independently
Keep execution, verification, recovery and mission outcome separate. “The command ran” is not the same as “the job is proven done.”
Remember with provenance
Treat persistent memory as governed knowledge: freshness, source and confidence matter more than retaining every transcript.
Evidence before claims.
The public release is fail-closed and evidence-bearing. Exact-tag validation runs before publication, and machine-readable artifacts are attached to the release.
281/281portable Python tests across 42 suites
3 OSWindows, Ubuntu and macOS portable runtime matrix
145/145legacy PowerShell governance tests
7,428 → 1,673 Bserialized context admitted under a 1,800-byte budget
The context number measures serialized UTF-8 bytes only — not token savings, cost, latency or answer quality. Open the v0.1.0 release and raw evidence.
Inspect the decision boundary.
J.A.R.V.I.S. is evolving toward a runtime where context, skills, tools and models are resolved through explicit contracts, authority checks, durable attempts and verification receipts.
OBSERVE → PLAN → RESOLVE → EXECUTE → VERIFY → MEASURE → LEARN
Built in public. Claims stay bounded.
v0.1.0 contains the governed skill registry, execution contracts/DAG, bounded inference primitives, verification structures, local HUD and memory/cognitive-vault foundations. Broader autonomy and empirical routing remain active development.
- Zero-dependency public Python launcher
- Portable runtime CI on Windows, Ubuntu and macOS
- Dedicated Windows compatibility job for legacy governance
- Reproducible context-budget benchmark with raw JSON evidence
- Immutable tagged public release with OCI packaging evidence
- Architecture docs that distinguish tested, partial and planned scope
“Execute less blindly. Verify more. Remember only what earned it.”
Build with it · challenge it · break assumptions
Start local. Bring evidence.
Run the three-command quickstart, inspect the release evidence, or pick a contributor-sized issue. Reproducible criticism is explicitly welcome.