Full language reference, compiler architecture, and quick start guides.
Skills Compiler — compile agent skills into optimized runtime modules.
Agent benchmark suite — measure and compare autonomous agent performance.
Evolutionary context management — genetic algorithms for optimal context windows.
GitHub organization with all Nexa projects and repositories.
Owen — creator of the Nexa language and Harness Native methodology.
Write goals and contracts in .nx. People and models collaborate through the compiler. Programs retain their implementations, evidence and revisions as requirements change.
Python 3.10+ · Linux/macOS · Docker/OCI for execution.
python -m pip install https://github.com/Nexa-Language/Nexa/releases/download/v3.0.0/nexa_lang-3.0.0-py3-none-any.whlInstall this GitHub wheel directly. The PyPI name nexa-lang belongs to another project. Installation guide →
language "3.0"
program Answer {
goal answer() -> Int {
intent "Return the number 42"
ensure checks.answer
implementation open
}
}
nexa check answer.nx --json checks the contract offline. A configured provider and frozen check can resolve the open implementation. Start with a complete example →
Open Programs extends the Harness Native foundation: the language owns what a program must achieve, the conditions under which it is accepted, and how its next revision becomes active.
Typed open goals can be implemented by a person or resolved by a model. The compiler enforces a shared budget and executes candidates and checks in OCI.
Activities stay pinned to immutable revisions. Checkpoints admit new versions; durable events and process leases support restart and conflict detection.
Deterministic checks, model reviews and human decisions remain separate records. Changed inputs, policies or dependencies create new verification obligations.
Maintenance produces a reviewable patch. A grant binds its baseline, paths and evidence; a journal and retained originals protect edits during recovery.
Local implementation capsules retain provenance, license, dependencies and evidence. A receiving task checks compatibility and revalidates under its current inputs.
The author protocol supports clarification and diagnostics. A diagnostic advice pass is selected after frozen tests, locked bootstrap and explicit approval.
Check a typed program offline. Build it with a configured execution environment, then replay its locked artifact without asking a model to generate it again.
language "3.0"
program Answer {
goal answer() -> Int {
intent "Return 42"
implementation value 42
}
}
nexa check answer.nx --json validates the source without Docker or credentials. The explicit value is the implementation; open goals let a configured model propose one.
nexa build answer.nx \ --environment examples/open_programs/offline.json \ --state .nexa/demo.db --json nexa build answer.nx \ --environment examples/open_programs/offline.json \ --state .nexa/demo.db --locked --json
Start Docker and pull the pinned image listed in the guide. Keep both the lock file and state database. Locked builds revalidate current evidence and make zero model requests. Execution and maintenance guide →
The deterministic kernel owns identity, policy and lifecycle. Models propose implementations and reviews; isolated execution supplies evidence. Python is the 3.0 reference implementation.
Types + goals
Clarify + resolve
Fixed checks
Context + policy
Admit + checkpoint
Validation covers current evidence, authorization, concurrent edits, crash recovery and locked replay. CI runs Python 3.10, 3.12 and 3.14 on Linux with real OCI execution. Five unconfigured PostgreSQL tests are skipped; timing benchmarks run separately.
The six-stage Aliyun demonstration completed with 11 model requests. In fixed synthetic tasks, Open Programs accepted 73/90 stages and the adapted 2.4 baseline 70/90. The report includes failures, token totals and the baseline-schema correction. Read the evidence →
language "3.0" selects Open Programs. Existing Agent and flow sources retain their 2.4 entry point. Legacy state exports into an inert archive with an explicit conversion report.
Unknown non-idempotent effects require reconciliation. Patch application preserves original files in recovery directories. Cross-file updates use a journal and require storage for retained originals.
Open Programs candidates run as non-root in offline, resource-limited OCI containers without host mounts or credentials. Legacy Python tools keep their documented host semantics; Rust retains its 2.4 subset.
Download 3.0.0, start with a typed goal, and keep the evidence behind every accepted revision.