feat(cmdcode): tool calling support

[evangit2]

Summary

Add full tool calling support to the Command Code provider client.

Changes

Tool passing

• _build_generate_body now forwards the tools array from the OpenAI request to the Command Code API
• Converts OpenAI tool format (parameters) to CC format (input_schema)
• Handles tool role messages (tool results) in the message conversion

Structured tool call event handling (primary)

The Command Code API returns structured SSE events for tool calls when tools are passed correctly:

• tool-input-start → creates tool call placeholder with ID and name
• tool-input-delta → accumulates argument JSON fragments
• tool-input-end → finalizes arguments
• tool-call → complete tool call (id, name, input)

These are handled in both streaming and non-streaming paths and translated to OpenAI-compatible tool_calls objects.

DSML parser (fallback)

Added a DSML <｜DSML｜tool_calls> parser as a fallback for cases where a model emits tool calls as text rather than structured events. Supports:

• <｜DSML｜invoke name="..."> with <｜DSML｜parameter> tags
• <name>/<parameters> JSON format
• ASCII pipe variant <|DSML|...>
• Multiple invokes in one block
• Partial marker detection across streaming deltas

Response builders

• _make_openai_chunk now accepts tool_calls parameter for streaming
• _make_openai_response now accepts tool_calls parameter for non-streaming
• finish_reason set to "tool_calls" when tool calls are emitted

Testing

• ✅ Streaming tool call (calculator: 2+2 → {"expression": "2+2"})
• ✅ Non-streaming tool call
• ✅ Tool result round-trip (tool result → model responds with content)
• ✅ No-tools backward compatibility
• ✅ All 27 existing tests pass
• ✅ DSML parser unit tests (7 test cases)