fix(llm-client): drop replayed web_search_call items under responses_reasoning = "drop"

crates/libsy-llm-client/src/client.rs

@@ -2652,7 +2652,9 @@ mod tests {
             {"type": "reasoning", "encrypted_content": ""},
             {"type": "reasoning"},
             {"type": "function_call", "call_id": "call_1", "name": "shell", "arguments": "{}"},
-            {"type": "function_call_output", "call_id": "call_1", "output": "ok"}
+            {"type": "function_call_output", "call_id": "call_1", "output": "ok"},
+            {"type": "web_search_call", "id": "ws_1", "status": "completed", "action": {"type": "search", "query": "switchyard"}},
+            {"type": "message", "role": "assistant", "content": [{"type": "output_text", "text": "found it"}]}
         ]);
         for store in [Some(true), Some(false), None] {
             for is_drop in [false, true] {
@@ -2711,6 +2713,10 @@ mod tests {
                         }
                     }
                     expected.extend([history[6].clone(), history[7].clone()]);
+                    if !is_drop {
+                        expected.push(history[8].clone());
+                    }
+                    expected.push(history[9].clone());
                     assert_eq!(
                         body["input"],
                         json!(expected),

crates/libsy-llm-client/src/responses_reasoning.rs

@@ -15,7 +15,8 @@ pub enum ResponsesReasoningPolicy {
     /// This is the default for every Responses model, regardless of its URL.
     #[default]
     PreserveEncrypted,
-    /// Drop all reasoning items while preserving messages and tool-call history.
+    /// Drop all reasoning items and replayed hosted web searches while preserving
+    /// messages and tool-call history.
     ///
     /// Use this for local Responses-compatible servers that cannot consume
     /// another provider's encrypted reasoning representation.
@@ -37,6 +38,14 @@ impl ResponsesReasoningPolicy {
         let Some(object) = item.as_object_mut() else {
             return true;
         };
+        // A hosted provider's server-side web search is replayed as a `web_search_call` item,
+        // which a local server cannot run or type (llama.cpp answers 400 "Cannot determine type
+        // of 'item'"). Its findings are in the assistant message that follows it.
+        if self == Self::Drop
+            && object.get("type").and_then(Value::as_str) == Some("web_search_call")
+        {
+            return false;
+        }
         if object.get("type").and_then(Value::as_str) != Some("reasoning") {
             return true;
         }

docs/reference/toml_schema.md

@@ -66,7 +66,11 @@ non-empty `encrypted_content` or a stored reasoning `id` are retained with empty
 `content`. ID-only reasoning is removed when the request sets `store: false`.
 Switchyard's translated Anthropic reasoning is always removed.
 Reasoning without either is removed. Set `responses_reasoning = "drop"` explicitly for a backend
-that cannot consume provider reasoning. Both modes keep messages, tool
+that cannot consume provider reasoning. `"drop"` also removes replayed
+`web_search_call` items: a conversation that ran a hosted web search replays
+them, and a local server such as llama.cpp rejects them with 400
+`Cannot determine type of 'item'`. The search results stay in the assistant
+message that follows. Both modes keep messages, tool
 calls and tool results. The setting is rejected on other client formats, even
 when no target uses the client. It is not inferred from the model name or URL.