Frequently Asked Questions

No. agentic-go bundles an exact, tested gopls v0.21.0 companion named agentic-go-gopls. It uses gopls as a private stdio sidecar for type analysis, definition lookups, and symbol references, while layering immutable snapshot lineage, change continuity contracts, guarded deterministic refactoring, and executed whole-package verification around it.

Open models (DeepSeek, Llama, Qwen, Mistral) excel at code generation, but often lack closed-lab tool harnesses. agentic-go gives them deterministic AST lookups, change contracts to prevent drift, and whole-package test verification - producing results on par with proprietary frontier setups.

agentic-go will never mutate Git repository state (it never executes git add, git commit, git branch, or git stash). The only mutating tool is go_refactor, which applies deterministic AST edits only to existing contained non-generated files after checking file preimages and writing a recovery journal.

Execution tools compile and run Go test suites using the local go toolchain with the caller's privileges (same as running go test ./...). All workspace paths are contained and symlink-resolved, but this is containment, not a sandbox. Do not run verification on untrusted code.

We explicitly support Go 1.25, Go 1.26, and Go 1.27 on macOS (Darwin) and Linux for both amd64 and arm64 architectures. Server startup verifies the host toolchain against the workspace's minimum Go requirement using GOTOOLCHAIN=local.

No. agentic-go is 100% deterministic local software. It does not embed any language models, does not require an API key, and makes zero outbound network calls. The bundled gopls sidecar has telemetry permanently disabled.

In calibration across 10 pinned Go repositories and 467 reviewed findings, our static analyzers produced zero false positives. Our analyzers are conservative: when source patterns cannot be definitively proven problematic, the engine reports an explicit uncertainty lens rather than spamming low-confidence warnings.