fix: propagate real tool_call_id instead of hardcoded placeholder (#73)

The worker (both agent/worker.rs and worker/runtime.rs) was passing the
literal string "tool_call_id" to ChatMessage::tool_result instead of
the actual tool call ID from the LLM response. This breaks
OpenAI-compatible providers that match tool results to their
corresponding calls by ID.

- Add tool_call_id field to ToolSelection struct
- Propagate ToolCall.id through select_tools() into ToolSelection
- Replace all hardcoded "tool_call_id" usages with selection.tool_call_id
- Generate unique IDs for plan-based synthetic selections
- Add test verifying tool_call_id is preserved

Co-authored-by: Yi LIU <[email protected]>
Co-authored-by: Illia Polosukhin <[email protected]>

src/llm/reasoning.rs

@@ -113,6 +113,12 @@ pub struct ToolSelection {
     pub reasoning: String,
     /// Alternative tools considered.
     pub alternatives: Vec<String>,
+    /// The tool call ID from the LLM response.
+    ///
+    /// OpenAI-compatible providers assign each tool call a unique ID that must
+    /// be echoed back in the corresponding tool result message. Without this,
+    /// the provider cannot match results to their originating calls.
+    pub tool_call_id: String,
 }
 
 /// Token usage from a single LLM call.
@@ -244,6 +250,7 @@ impl Reasoning {
                 parameters: tool_call.arguments,
                 reasoning: reasoning.clone(),
                 alternatives: vec![],
+                tool_call_id: tool_call.id,
             })
             .collect();