Local CLI · Python 3.12
A local, CLI-based autonomous coding agent. The model reasons; deterministic software searches, executes, constrains, verifies, persists, and observes.
End-to-end agent mechanics with explicit safety boundaries and verifiable completion.
Native provider tool calls with iteration, repeat, and budget guards.
Table-driven allow, ask, and deny before execution; single-use approvals.
Git worktree by default, checkpoints, rollback, and optional Docker sandbox.
Checks pipeline and completion gate; no verified finish without evidence.
SQLite state, resume after interrupt or quota wait, full event trace.
40-task fixture suite with scripted replay in CI and optional live slices.
Terminal recording from a representative session on the payment service example.
Dependencies point downward: the loop orchestrates; tools do not call back into it.
flowchart TB
CLI[CLI_Typer_Rich]
Loop[AgentLoop]
Provider[Provider]
Tools[ToolRegistry]
Perm[PermissionEngine]
State[StateStore]
Iso[Isolation]
Verify[Verification]
Obs[Observability]
CLI --> Loop
Loop --> Provider
Loop --> Tools
Loop --> Perm
Loop --> State
Loop --> Iso
Loop --> Verify
Loop --> Obs
Tools --> Perm
Full specification: DESIGN.md
Python 3.12, strict mypy, Ruff, pytest on Ubuntu and macOS. 589 offline pytest cases in CI (live excluded), including a dedicated security suite.
Live results on a 30-task SWE-bench Lite subset, graded by hidden failing tests. Scripted replay guards the loop in CI.
About $0.02 per solved task. Identical runs flipped 6 of 30 outcomes.
Not significantly above Haiku at n = 30 (McNemar p = 0.13–0.34), at 12 times the cost per solve.
A plumbing regression gate through the real loop, tools, and permissions via FakeProvider. Not a capability measure.
Local environments, not the official SWE-bench images, so the numbers are not comparable to leaderboard results. Full report: SWE-bench subset results.
Shipped through phase 8, tuned for small repositories and Groq free-tier context and quota limits.