Testing And Parity Reference

Use this when designing tests or completion evidence.

Minimum Evidence

By completion, provide evidence for:

• Package entrypoint imports successfully.
• Plugin factory returns expected handlers/tools/commands.
• Shared .agents/ assets are readable or package-local copies exist when needed.
• Every source and package-local SKILL.md parses as YAML frontmatter and includes name and description.
• Package-local skill copies, when used, are tested for content equality with the declared source of truth or explicitly documented as intentionally OpenCode-specific.
• OpenCode command and agent wrappers have frontmatter when generated.
• Migrated MCP config has valid mcp entries in opencode.jsonc / opencode.json when Claude MCP servers are present.
• Every ported hook has an OpenCode event fixture test.
• State writes avoid Claude config directories and use safe OpenCode/XDG paths.
• Existing Claude tests still pass when they exist.
• Any local .opencode/ harness is intentional and documented.

Hook Test Matrix

Package Tests

Useful package tests:

• Import opencode-plugin/src/plugin.ts or opencode-plugin/plugin.js.
• Instantiate the exported plugin with a fake context.
• Assert returned object contains only expected handlers/tools.
• Exercise each handler with a minimal OpenCode-like input/output fixture.
• Verify opencode-plugin/package.json main and exports point to existing files.
• Parse every packaged SKILL.md frontmatter and assert required name and description fields exist.
• If opencode-plugin/skills/ contains copies, assert they match the declared shared source or document why they intentionally differ.

MCP Config Tests

For migrated MCP servers:

• Validate opencode.jsonc or opencode.json syntax for the chosen destination format.
• Confirm each MCP entry has type.
• Confirm local entries have command as an array.
• Confirm remote entries have url.
• Confirm Claude env keys were renamed to OpenCode environment.
• Confirm command paths exist or are documented as external dependencies.
• Confirm secrets were not inlined.
• Confirm name conflicts with existing mcp entries were resolved intentionally.

Manual Smoke Tests

If OpenCode is available and the user wants local dogfooding, create a small .opencode/plugins/<plugin>.ts shim and run OpenCode in the repo. Make clear this is a development harness, not the package distribution path.

Discovery checks for migrated surfaces:

Skills can be discoverable even when a UI view does not surface them. Prefer opencode debug skill for skill discovery evidence.

Experimental APIs

For experimental.* OpenCode APIs:

• Record the OpenCode version tested.
• Keep the feature optional where possible.
• Add a caveat in docs and completion summary.
• Do not claim exact Claude parity without a passing test or manual smoke result.