fix(llm): strip inline <think> reasoning from Dahl/MiniMax streams; dedupe identical parallel tool calls; hermetic config tests

internal/agent/agent.go

@@ -178,11 +178,20 @@ func (a *Agent) Run(parent context.Context, sessionID, userText string, emit fun
 		if len(resp.ToolCalls) == 0 {
 			return resp.Content, nil
 		}
+		// Some models (seen with GLM/DeepSeek on Dahl) emit the same call several times in
+		// parallel; identical name+arguments run once and share the result.
+		seen := map[string]string{}
 		for _, tc := range resp.ToolCalls {
 			emit(Event{Type: "tool_call", Tool: tc.Name, Args: tc.Arguments, CallID: tc.ID})
-			out := "error: cancelled by user"
-			if !cancelled() {
+			key := tc.Name + "\x00" + tc.Arguments
+			out, dup := seen[key]
+			switch {
+			case dup:
+			case cancelled():
+				out = "error: cancelled by user"
+			default:
 				out = a.Tools.Call(ctx, tc.Name, tc.Arguments)
+				seen[key] = out
 			}
 			emit(Event{Type: "tool_result", Tool: tc.Name, Text: clip(out, 2000), CallID: tc.ID})
 			// Every tool call gets a result so the history stays valid for the provider.