Assess portability of OpenCode plugins to Codex and Claude Code

[OpenCodeEngineer]

Context

Evaluate whether the current OpenCode plugins (reflection-3.ts, tts.ts, telegram.ts) can be adopted for:

1. Codex
2. Claude Code

Goals

• Identify hard runtime coupling to OpenCode plugin APIs/events
• Determine what can be reused vs. rewritten
• Propose a concrete adapter architecture per target runtime
• Document risks (security/test portability/operational)

Acceptance Criteria

• Portability matrix (OpenCode -> Codex/Claude) for key capabilities:
  • session lifecycle hooks
  • message/session read/write
  • child/judge session spawning
  • tool registration and command interception
  • UI/toast notifications
• Explicit blocker list with file/line references
• Migration recommendation (native port, adapter, or not viable now) for each plugin
• Follow-up implementation issues split by plugin/runtime

Notes

• Keep this issue linked to any PR that implements portability work.

[OpenCodeEngineer]

Initial review in progress.

Findings so far:

• Plugin runtime is tightly coupled to OpenCode session lifecycle APIs (session.idle, session.error, session.updated, session.create/messages/promptAsync/delete, tui.publish, command interception).
• Claude Code looks viable via native hooks/plugins/MCP bridge with moderate adapter work.
• Codex appears viable only via an external orchestrator/App Server event stream adapter, not as a direct drop-in plugin.
• Opened follow-up security issue: Security: remove hard-coded Supabase service role key from test suite #132 (hard-coded Supabase service role key in tests + production-coupled test path).

Will post a full portability matrix and migration recommendation next.

[OpenCodeEngineer]

Portability verdict (review complete):

Summary

• Claude Code: adoptable now via native hooks/plugins with moderate adapter work.
• Codex: adoptable with an external adapter/orchestrator (App Server), not as a direct drop-in plugin.

Why

Current implementation is tightly coupled to OpenCode runtime surfaces:

• event model: session.idle, session.error, session.updated
• session API: session.create/messages/promptAsync/delete/get
• UI API: tui.publish
• command interception: command.execute.before
• OpenCode message schema (info.role, parts[], tool payload shapes)

Runtime mapping

• Claude Code has first-class hooks/plugin architecture (Stop/TaskCompleted/PreToolUse/PostToolUse, plugin packaging, MCP in plugins), so Reflection/TTS/Telegram behaviors can be recreated natively.
• Codex docs expose MCP + Skills + App Server event stream (thread/*, turn/*, item/*) and turn steering; this enables parity but requires a long-running sidecar service to subscribe to events and inject continuation turns.

Key blocker outside portability

• Security test hygiene issue tracked in Security: remove hard-coded Supabase service role key from test suite #132 (hard-coded Supabase service-role credentials and production-coupled default tests).

Recommendation

1. Extract host-agnostic core (reflection evaluator + text filtering + Telegram transport) into shared module.
2. Build runtime adapters:
   • adapter-opencode (existing)
   • adapter-claude (hooks + plugin packaging)
   • adapter-codex (App Server event consumer + turn orchestrator)
3. Fix Security: remove hard-coded Supabase service role key from test suite #132 before broad adoption work so CI/test behavior is safe and reproducible.