Local CLI · Python 3.12

CodeAgent

A local, CLI-based autonomous coding agent. The model reasons; deterministic software searches, executes, constrains, verifies, persists, and observes.

Capabilities

End-to-end agent mechanics with explicit safety boundaries and verifiable completion.

Tool-calling loop

Native provider tool calls with iteration, repeat, and budget guards.

Permission engine

Table-driven allow, ask, and deny before execution; single-use approvals.

Isolation

Git worktree by default, checkpoints, rollback, and optional Docker sandbox.

Verification

Checks pipeline and completion gate; no verified finish without evidence.

Durable runs

SQLite state, resume after interrupt or quota wait, full event trace.

Evaluation harness

40-task fixture suite with scripted replay in CI and optional live slices.

Demo

Terminal recording from a representative session on the payment service example.

Architecture

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

Engineering

Python 3.12, strict mypy, Ruff, pytest on Ubuntu and macOS. 589 offline pytest cases in CI (live excluded), including a dedicated security suite.

Evaluation

Live results on a 30-task SWE-bench Lite subset, graded by hidden failing tests. Scripted replay guards the loop in CI.

12–13 / 30 Claude Haiku 5.5, three runs

About $0.02 per solved task. Identical runs flipped 6 of 30 outcomes.

17 / 30 Claude Sonnet 5.5, one run

Not significantly above Haiku at n = 30 (McNemar p = 0.13–0.34), at 12 times the cost per solve.

42 / 42 Scripted replay (CI)

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.

Scope

Shipped through phase 8, tuned for small repositories and Groq free-tier context and quota limits.

Decision records · Build progress · Documentation index