tui: subagent task row does not show agent name or model #57970

Summary

Since the task/subagent progress redesign (~1.2.16, carried unchanged into V2), the TUI task row no longer shows which agent or which model is running a subagent — only the short description label. With multiple custom subagents pinned to different models (local and cloud), it is impossible to tell at a glance which model a dispatch is using. Pre-redesign V1 showed the agent type on the row.

Related: #15915 (closed automatically as "not planned" in July 2026 after no fix; behavior still present in 2.1.2), #16287 (follow-up proposing agent name + session ID suffix).

Environment

• opencode version: 2.1.2
• OS: Linux 7.0.0-38-generic (Ubuntu, x64)
• Terminal: xterm-256color, COLORTERM=truecolor
• Shell: /bin/bash
• Install/channel: latest
• Active plugins: none (no plugin files configured; a leftover V1 @opencode-ai/plugin package dependency exists in the global config dir but loads nothing)

Reproduction

1. Define two custom subagents with different models, e.g. ~/.config/opencode/agents/coder.md pinned to a cloud model and ~/.config/opencode/agents/coder-local.md pinned to a local model.
2. From a primary-agent session, dispatch work to each subagent via the subagent tool.
3. Observe the task progress row in the full-screen TUI while running and after completion.

Expected Behavior

The task row identifies the agent and/or model, e.g. Explore: <description> as in pre-redesign V1, or a suffix like coder-local · qwen3.8 [d822ff] (per #16287). In mixed local/cloud subagent setups the model identity per row is the load-bearing detail.

Actual Behavior

The row shows only [Task <short description>] plus N toolcalls · duration — no agent type, no model. Applies to in-progress and completed foreground tasks alike.

Additional Context

• Regression window: introduced by the task/subagent progress redesign around v1.2.16 per Task/subagent progress display no longer shows agent name (e.g. Explore, General) #15915; never restored; present in 2.1.2.
• Impact scales with the number of configured models: in mixed local/cloud setups the transcript gives zero indication whether a dispatch hit the local or the remote model, or what it cost.
• Current workaround: encode agent + model into the subagent call's description label (e.g. coder-local · qwen3.8 — implement fix), since that label is the only user-visible metadata on the row. Works, but relies on the parent agent's labeling discipline rather than the UI.
• Happy to test a fix; running mixed local + cloud subagent fleets.