Did it meet the contract?
TOML contracts declare executable checks and review clauses. A favorable model review does not erase a failing harness check.
CONTRACT-DRIVEN AGENT EXECUTION
Resume the work.
Inspect the proof.
An agent saying “done” is not an acceptance test. Truss gives AI-agent tasks explicit contracts, a durable execution journal, and checks the harness runs itself.
Development project · Rust · MIT
FROM INTENT TO EVIDENCE↘
- 01Task contract
- 02Agent execution + journal
- 03Harness checks + model review
- 04Verdict + signed record
Runtime architecture · not a live benchmark
THREE QUESTIONS THE RUNTIME ANSWERS
Make completion inspectable.
What survives an interruption?
A SQLite journal records execution. Resume reconstructs state and reuses recorded model and tool outcomes instead of treating the transcript as a plan to rerun.
Has the record changed?
Hash-chained, Ed25519-signed entries support verification against a trusted key. Record integrity is distinct from the correctness of an agent’s answer.
START WITH DETERMINISTIC TESTS
No provider key required.
Use a stable Rust toolchain, Git, a C/C++ build toolchain for bundled SQLite, and a Unix-like shell. The durable-runtime tests use controlled providers. Initial dependency downloads require network access.
Full prerequisites and test tour ↗SOURCE CHECKOUT● ● ●
git clone --depth 1 https://github.com/ControlStackAI/truss.git
cd truss
cargo test --locked -p truss-core --test durable_testCURRENT SCOPE
A runtime, not a reliability guarantee.
Truss is a development-stage standalone runtime and CLI, not an OpenClaw fork. Controlled tests do not establish production-scale reliability or superior live-model task quality. Live provider execution and model review can incur API charges. Contracts execute commands; run unfamiliar projects in an appropriately isolated environment.