OpenCode harness

OpenCode provides the coding-agent experience: prompts, model interactions, and tool calls. This directory packages its lifecycle scripts, pinned image reference, and OpenShell profiles; it does not replace OpenCode with a custom agent.

Deploy and test OpenCode from the OpenShell single-server guide. The optional Praxis route is documented separately for administrators who select it. The active pins are covered by the 0.1.4 release review. The 0.1.3 upgrade tests remain historical evidence for model responses, streamed write and bash tools, independent file readback, and policy denials with controlled model fixtures. Qualify the intended real provider and model separately. In that route, tools do not receive direct vLLM or cloud-provider access. In standalone OpenShell mode, selected model endpoints remain governed by the harness policy.

The preinstalled harness image is digest-pinned in openshell/configs/images.env. That upstream image does not yet include ripgrep; RHEL 9 also does not package it. The bootc image and manual quickstart provide the same checksum-pinned static rg, and harness creation copies it into the sandbox's ~/.local/bin so OpenCode's built-in glob and grep tools work without changing the upstream image or policy.

The pr-review profile adds delegated PR review: review.sh sends a batch of PRs as one bounded agent turn and review-collect.sh retrieves the report artifacts. See the review service guide.

The experimental pr-queue profile keeps its queue runner, collector, and agent instructions in pr-queue/. The user-facing entry point is the PR-candidate example. The sandbox can read public GitHub issues and clone/fetch a repository, but it cannot push or create pull requests.