Skip to content

CONTRIBUTING.md

Latest commit

 

History

History
76 lines (57 loc) · 2.81 KB

CONTRIBUTING.md

File metadata and controls

76 lines (57 loc) · 2.81 KB
Raw

Contributing

Keep the capability small: agent guidance, human approval, native skills. Read AGENTS.md and docs/research.md before changing the runtime. Do not add independent persistence, autonomous writes, or cleanup infrastructure.

Development setup

Requires Node.js 24+ and npm. OpenCode 1.18.30 is required for live CLI tests.

git clone https://github.com/Neverdecel/opencode-guided-learning.git
cd opencode-guided-learning
npm ci
npm run check
npm test

GitHub Actions runs the same deterministic checks. They cover hook registration, isolation, options, and bundled skill format. They do not certify model consent or secret detection.

Pull requests

  1. Branch from main. Direct pushes, force-pushes, and deleting main are blocked.
  2. Keep the change focused. Update docs and examples with options or behavior.
  3. Inspect pinned @opencode-ai/plugin and SDK types before changing hooks.
  4. Open a PR targeting main. The check CI job must pass.
  5. Squash or rebase merge only (linear history).

Ship product skills under skills/ and load them via skills.paths. Keep examples outside auto-discovery directories. Do not copy product skills into another project’s .opencode/skills/ in docs or tests unless the point is a fork.

Live CLI and model checks

Linux, Docker, and a native OpenCode binary. From the checkout root:

# Actual OpenCode system request and persisted-session isolation; fake provider.
EVAL_SCRIPT=context sh test/live.sh

# Real model, native file tools, disposable skill libraries.
sh test/live.sh

EVAL_CASE=4 sh test/live.sh
EVAL_MODEL=opencode/mimo-v2.5-free EVAL_CASE=4 sh test/live.sh

The scripts mount the checkout and CLI read-only in a disposable container. They do not mount a real home or credentials. Default live model: opencode/big-pickle (provider-dependent). The noninteractive runner uses text replies and disables the question tool. These tests are opt-in, not CI.

Review tool traces and generated prose, not only exit status. No-write cases fail on attempted skill/config writes, including blocked tool calls. Keyword checks are smoke only. Acceptance criteria: test/SCENARIOS.md. Record intermittent failures; a later pass does not erase them. Never run evaluation writes against a real skill library.

Releases

Package, repo, and plugin id are opencode-guided-learning. Publish with a v* tag or workflow_dispatch on .github/workflows/publish.yml; it needs the NPM_TOKEN secret. Do not npm publish from a laptop. Bump package.json version before tagging.

Issues

Include OpenCode and model versions and a minimal, sanitized reproduction. Do not upload configuration secrets, credentials, or private session exports. Security reports go through SECURITY.md, not public issues.