Plan OpenCode V2 plugin adapter #718

Context

OpenCode V2 is currently beta and intentionally does not load V1 plugin implementations: https://opencode.ai/v2/docs/migrate-v1

Ponytail already has a strong cross-host boundary: shared instruction/config builders, AGENTS.md, skills, and commands. Only the OpenCode runtime adapter in .opencode/plugins/ponytail.mjs is V1-specific. It currently uses an async plugin export with config, experimental.chat.system.transform, and command.execute.before.

Suggested port plan

• Keep shared instruction generation and persisted mode handling unchanged.
• Add a separate V2 adapter using a unique id + setup contract.
• Register packaged skills and commands through V2 setup transforms, or rely on native disk discovery where possible.
• Inject active Ponytail instructions through the V2 session context hook.
• Handle /ponytail <mode> through the V2 command transform.
• Keep AGENTS.md / skills as the no-plugin fallback for users who do not need runtime mode switching.
• Preserve V1 adapter until V2 stabilizes rather than forcing one file to support both contracts.

Compatibility checks

• Plugin appears in opencode2 api get /api/plugin.
• Default mode is injected once with no duplicate rules.
• /ponytail lite|full|ultra|off changes later turns and persists as documented.
• Packaged commands and skills remain discoverable.
• Missing or malformed state files fall back safely.
• Installation from npm works in standalone and managed service modes.
• Existing multi-agent tests and V1 OpenCode behavior remain unchanged.

Would a thin V2 adapter be welcome once the API settles? I can help test it against current opencode2 previews.