opencode-devtools

Verification and build-status primitives for AI agent execution — type checking, testing, and CI as control-loop gates.

An OpenCode plugin that exposes development infrastructure as AI-callable tools. Instead of the agent guessing whether its changes compile or pass tests, these tools give the execution loop hard verification constraints: type-check, lint, run tests, check CI — then decide.

Why This Exists

In a control-system architecture (intent → orchestration → execution → verification → supervision), verification is the gate between "the agent did something" and "a human should look at it." Without verification primitives, the agent runs blind and the human reviews everything. With them, the loop self-corrects before surfacing results.

This plugin implements that verification layer for software development workflows. The primitives are domain-specific (TypeScript, ESLint, test runners, CI), but the pattern is general: give the agent observable constraints so it can converge before asking for human attention.

Tools

Tool	What it does	Execution
dev-server	Start and manage Expo/Next.js dev servers	PTY (long-running)
server-logs	Read server output with filtering and pagination	PTY buffer
check-types	TypeScript type checking	Bun.$
lint	ESLint with auto-fix support	Bun.$
generate	Code scaffolding from templates	Bun.$
create-pr	Interactive PR creation (auto-answers prompts)	PTY
run-tests	Test execution with aggregated results	Bun.$ / PTY
build-status	CI pipeline monitoring (Azure DevOps)	Bun.$ / PTY

Setup

Register as a plugin in your opencode.json:

{
"plugins": {
"devtools": {
"path": "path/to/opencode-devtools"
}
}
}

Then optionally run the setup CLI:

bunx devtools init

This generates a .devtools.jsonc config file with tool toggles and per-tool configuration.

Configuration

All config lives in .devtools.jsonc (or .devtools.json) at the project root. Validated with Zod schemas. Falls back to sensible defaults when no config file exists.

Default tool enablement:

• Enabled: dev-server, server-logs, check-types, lint, run-tests
• Disabled: generate, create-pr, build-status

See ARCHITECTURE.md for schema hierarchy and implementation details.

Architecture

Two execution strategies based on tool needs:

1. Bun.$ — For commands that run to completion (type checking, linting, tests)
2. bun-pty — For long-running or interactive processes (dev servers, CI watch, PR creation)

PTY sessions are managed by a singleton SessionManager with a 50k-line rolling buffer per session, supporting paginated reads and regex filtering.

See ARCHITECTURE.md for the full technical breakdown.

Part of a Larger Stack

This repo is one layer of a control-system architecture for supervised AI execution:

Layer	Repo
Supervision	react-native-opencode-client
Verification	opencode-devtools (this repo)
Execution	OpenCode
Learned Policy	opencode-learn
Epistemic Safety	agent-memory-failure-demo

Read the full thesis at cloudpresser.com.

License

MIT