Agent viewer re-renders session welcome banner, mislabeling subagents' model (v2.1.214)

Environment

• Claude Code v2.1.214, Linux (kernel 7.0.0-27-generic), terminal CLI
• Claude Max plan, session model: Fable 5 (high effort)

Description

When you switch into a running subagent's view (agent picker under the prompt box), the banner rendered at the top of the screen is the session welcome banner — "Fable 5 with high effort · Claude Max" — regardless of the model the subagent was actually launched with via the Agent tool's model parameter. This makes it look like per-agent model overrides are silently ignored and every subagent is burning session-model (Fable) tokens.

Repro steps

1. Start a session on Fable 5.
2. Have the main agent dispatch background subagents with subagent_type: "general-purpose" and model: "sonnet".
3. While they run, select one in the agent picker to view its transcript.
4. The top of the screen shows the session banner: "Fable 5 with high effort · Claude Max".

Expected

The agent view should show the subagent's actual model (or at minimum not re-render the session banner, which reads as an authoritative model label for the agent being viewed).

Actual

The session banner is re-rendered in the subagent view, implying the subagent runs the session model.

Evidence the override actually works (i.e. this is display-only)

Launched two identity-check subagents with model: "sonnet" — one with run_in_background: false, one background. Both reported their injected environment line verbatim: "Sonnet 5, model ID claude-sonnet-5". So dispatch honors the override on both paths; only the viewer's banner is wrong.

Impact

Users conclude that cheap-model fan-outs (e.g. 6 parallel Sonnet extraction agents) are secretly running the expensive session model, and may abort legitimate work or re-architect around a bug that doesn't exist.