Improve visibility of long-running / externally-blocked MCP tool calls

[aheritier]

Problem

When an MCP server (e.g. IntelliJ) blocks a tool call waiting for user confirmation on its side, docker-agent shows only an indefinite spinner (⟳ Working...). The user has no way to know:

• How long the tool call has been running
• That an external application needs attention
• What to do about it

This was observed with IntelliJ's MCP server blocking on execute_terminal_command confirmation while the docker-agent TUI appeared stuck indefinitely.

Screenshots

The TUI shows a tool call spinner with no duration or hint that IntelliJ needs attention:

Proposed Improvements

1. Elapsed time on running tool calls (low effort)

Add a StartedAt timestamp to tool call messages and render elapsed time next to the spinner:

⟳ Run tools with Javascript (2m15s)

Makes stuck calls immediately obvious.

2. Long-running tool call warning (low effort)

After a configurable threshold (e.g. 60s), display a warning in the TUI:

⚠ Tool call has been running for over 60s. The MCP server may be waiting for external input. Check the MCP server application for pending confirmations. Press Esc to cancel.

3. Configurable MCP tool call timeout (medium effort)

Wire the existing Toolset.Timeout config field (currently only used by fetch) to MCP CallTool via context.WithTimeout. Default could be ~5 minutes. On timeout, return a clear error message so the model can retry or switch tools.

Relevant Code Paths

• pkg/tools/mcp/mcp.go — callTool() has no timeout, blocks indefinitely on ts.mcpClient.CallTool(ctx, request)
• pkg/tui/types/types.go — Message struct has no StartedAt field
• pkg/tui/components/toolcommon/common.go — Icon() / RenderTool() show spinner with no duration
• pkg/config/latest/types.go — Toolset.Timeout exists but is only wired for fetch