Add Fix with Copilot for failed local agent launches

[!IMPORTANT]
This experience is currently blocked by a missing terminal canvas capability. The existing API only provides send_terminal_input and read_terminal_output; it does not expose command identity, process state, completion, cancellation, or exit codes.

Problem

When a user selects Inspect locally, Microsoft Foundry Canvas starts the agent with azd ai agent run --no-client in the integrated terminal. If startup fails—for example, because the Python version does not satisfy a dependency requirement—the terminal displays the actionable error, but Canvas cannot reliably determine that the launch command has exited.

Skylight PR microsoft/Skylight#5862 improves the immediate experience by keeping the terminal focused until the agent becomes reachable. However, Canvas cannot reliably return with a failure state or offer Fix with Copilot because the required process-level terminal API does not exist today.

Why the current API is insufficient

• send_terminal_input confirms only that text was written to the shell; it does not identify or track the resulting process.
• read_terminal_output returns rendered terminal text; it does not indicate whether the command is still running or what exit code it produced.
• Port readiness distinguishes a successfully started agent from one that is not ready, but it cannot distinguish slow startup from an exited command.
• Shell wrappers and marker parsing would alter the visible command, add noise to troubleshooting context, and require shell-specific behavior.

Proposed experience

1. Keep the terminal visible while the local agent starts.
2. When the launch command exits unsuccessfully, return to Foundry Canvas with a concise failure state.
3. Offer Fix with Copilot, Open terminal, and Retry actions.
4. Send a bounded terminal-output tail to Copilot only after the user selects Fix with Copilot.

Required new API capability

Add a terminal canvas API that executes the original visible command and exposes structured lifecycle information, for example:

run_terminal_command -> { commandId }
get_terminal_command_status -> {
  state: "running" | "completed" | "failed" | "cancelled",
  exitCode?: number,
  outputTail?: string
}

An event-driven completion notification would also work. The API must correlate status with the specific command, rather than with the terminal instance alone.

How this API solves the issue

• Reliable failure detection: a nonzero exit code immediately distinguishes failure from slow startup.
• No command pollution: the terminal shows the original azd ai agent run --no-client command without wrappers or markers.
• Accurate Copilot context: Canvas can attach the output associated with the failed command rather than unrelated terminal scrollback.
• Safe user consent: output is sent to Copilot only after the user selects Fix with Copilot.
• Correct state transitions: Canvas can separately handle successful startup, command failure, user cancellation, terminal closure, and timeout.

Acceptance criteria

• Canvas can execute the original command and receive a stable command identifier.
• Canvas can determine whether that command is running, completed, failed, or cancelled and obtain its exit code.
• Failure diagnostics can include a bounded, ANSI-resolved output tail correlated to that command.
• Terminal cancellation and closure are represented distinctly from command failure.
• Foundry Canvas can use the result to present a reliable, user-initiated Fix with Copilot action.