Feature request: OpenCode as a Sesame Link runtime

What

Add OpenCode (opencode CLI) as a supported Sesame Link
runtime, alongside claude_code, codex, pi, and cursor
(allowed_remote_runtimes).

Why

• OpenCode is multi-provider (85+ providers via models.dev, local models via
  Ollama/LM Studio, plus OAuth for existing ChatGPT Plus/Pro and Claude
  Pro/Max subscriptions). It is the only major agent where users bring their
  own model — including fully free local setups.
• Users who can't use Claude (subscription cost) or hit Codex limits currently
  have no Link-compatible agent with model choice.
• Your own docs describe Link as "coding agent agnostic ... with more coding
  agents on the way". OpenCode is the natural next runtime.

Suggested behavior (mirrors Codex/Claude support)

• Detect opencode on PATH (opencode --version).
• sesame-link opencode starts an OpenCode session in tmux (like Claude/Pi),
  or without tmux if preferred.
• Remote create / message / interrupt / approval answers in the web client and
  voice agent, same as Codex sessions.
• Reuse OpenCode's provider auth (opencode auth login) the way Link reuses
  codex login / Claude sign-in — no credential handling by Link needed.
• Include opencode in allowed_remote_runtimes.

Workarounds tried (why native support matters)

• allow_remote_bash_commands + !opencode run "..." works for one-shot
  tasks but gives no live steering, no streamed tool steps, and needs a host
  session open. It is a hack, not a runtime.
• A shim binary can't work: Link drives Codex via its app-server protocol,
  which OpenCode doesn't speak.

Environment

• Sesame Link 0.1.33 (release), Linux x86_64