[Feature]: Run ! shell commands in structured Claude chats, as the TUI does #32165

Problem or use case

In a Claude Code chat that runs as a structured chat (Claude launched with --input-format stream-json --output-format stream-json), a !-prefixed line does not run as a shell command. The composer sends !pwd as an ordinary prompt, and the model answers it in prose. Running the command takes a full model turn, and the result depends on the model choosing to call its Bash tool.

The same line does run in a terminal-backed chat, because the composer types it into the Claude TUI and the TUI's bash mode runs it (see #25956, fixed by #31204). So ! works or silently doesn't depending on which kind of chat the user opened, and the composer gives no hint which one it is.

Reproduction outside Orca (Claude Code 2.1.304):

printf '%s\n' '{"type":"user","message":{"role":"user","content":"!pwd"}}' \
  | claude -p --input-format stream-json --output-format stream-json --verbose --max-turns 1

No <bash-input> turn is recorded. The assistant replies with the working directory it read from its system context and says it did not run a shell.

On main (77f1523), classifyNativeChatSend in src/shared/native-chat-slash-commands.ts special-cases only / (and $ for Codex), and the structured send path has no ! handling either. src/shared/claude-bash-input.ts only displays <bash-input> turns that already exist in a transcript.

Orca 1.4.237, macOS.

Proposed solution

Support ! in structured Claude chats the way the TUI does:

1. When a structured Claude chat's draft starts with !, Orca runs the rest of the line itself in the chat's cwd, using the shell environment structured chats already use (feat(native-chat): one shell-environment setting for every structured chat #22387), instead of sending it to the agent.
2. The chat shows the command and its stdout/stderr as a command row.
3. Orca passes the command and its output to the agent with the next message, in the same shape the TUI records (<bash-input> plus <bash-stdout>/<bash-stderr>), so the agent sees it as context without spending a turn on it.

If this is out of scope, a smaller step would help: when a structured chat's draft starts with !, show a hint that bash mode is only available in terminal-backed chats.

Alternatives or additional context

Workarounds today: open the conversation in a terminal tab and use the TUI's !, or ask the agent to run the command and wait for a turn.