feat: live Cursor subagent activity in the task card

[justin-carper]

A Cursor subagent's task card now behaves like a native opencode subagent card: navigable while running, with a live ↳ <Tool> <title> activity subtitle, and the child session carries the subagent's own output.

How it works

• Child session up-front. The child session is created when the task call starts and state.metadata.sessionId is stamped on the running task part via part.update (the native task tool's execute-time metadata equivalent). The TUI reads part.state.metadata.sessionId for navigation, so the card is clickable / ctrl+x-navigable from the start of the run, not just after completion.
• Live activity from streamed events. The Cursor SDK streams a local subagent's nested activity via taskUpdate payloads on the parent task's tool-call-delta updates (text, thinking, tool-start/tool-result with id + name + input). These are normalized into subagent-events and written into the child session as real tool parts via the part.update upsert — the exact mechanism opencode's processor uses — so the TUI's task card shows live activity while the subagent runs. Tool calls complete when their tool-result arrives; any call left open is completed at finalize.
• Full transcript in the child session. The child session is seeded with the subagent's prompt and receives its rendered activity: assistant text, thinking, tool calls with args and results, conversation steps, final answer, and duration. No truncation.
• Reliable child writes. upsertToolPart treats a resolved { error } response as failure (the hey-api client does not throw on 4xx) and always sends state.metadata on completed parts, which opencode's completed-state schema requires. Child-session writes are best-effort: a failed write is logged, never breaks the parent turn.
• cursor_delegate also links a child session seeded with its transcript, discoverable via the TUI's subagent panel.

Verification

• npm run typecheck && npx vitest run && npm run build — 567/567 tests pass.
• Live opencode run exercise (with the local provider override): tool parts appear in the child session mid-run with real start/end timestamps, and the parent task part carries state.metadata.sessionId while the subagent is still running.