fix(flows): preserve literal tool-call markup in tool-less completions (#4750)

src/openhuman/tinyagents/model.rs

@@ -61,11 +61,11 @@ fn build_chat_inputs(
 ///
 /// The text-mode fallback parse needs each tool's positional parameter layout
 /// to reconstruct named JSON arguments from a P-Format `name[a|b]` body. The
-/// harness always populates `request.tools` (schemas are rendered into the
-/// prompt for prompt-guided providers, or advertised natively otherwise), so
-/// the registry is available in both modes. An empty registry (no tools
-/// advertised) makes the P-Format-aware parser short-circuit to the canonical
-/// grammar, so this is behaviour-neutral when there are no tools.
+/// harness populates `request.tools` when tools are available (schemas are
+/// rendered into the prompt for prompt-guided providers, or advertised natively
+/// otherwise), so the registry is available in both modes. Tool-less requests
+/// skip fallback parsing entirely; this empty registry is therefore consulted
+/// only alongside a non-empty advertised tool list.
 fn pformat_registry_from_request(
     request: &ModelRequest,
 ) -> crate::openhuman::agent::pformat::PFormatRegistry {
@@ -84,22 +84,25 @@ fn pformat_registry_from_request(
 /// Translate an openhuman [`ChatResponse`] into a harness [`ModelResponse`]
 /// (visible text + tool calls + token usage).
 ///
-/// Native `tool_calls` take precedence; when absent, the response text is parsed
-/// for prompt-guided (`<tool_call>…` / p-format) calls — matching the legacy
-/// dispatcher — so text-mode models drive the tinyagents loop too. The visible
-/// text is the prose with any tool-call markup stripped.
+/// Native `tool_calls` take precedence; when absent and the request advertised
+/// tools, the response text is parsed for prompt-guided (`<tool_call>…` /
+/// p-format) calls — matching the legacy dispatcher — so text-mode models drive
+/// the tinyagents loop too. Tool-less requests preserve response text verbatim,
+/// including literal tool-call examples. When parsing is enabled, visible text
+/// is the prose with any parsed tool-call markup stripped.
 ///
 /// `pformat_registry` carries the advertised tools' positional layouts so the
 /// text-mode fallback can recover P-Format (`name[a|b]`) calls that ~10 builtin
 /// prompts still teach — the migrated parse path had dropped that grammar and
 /// silently lost those calls (issue #4465). It is empty for the native-tool
-/// path (where `response.tool_calls` is used directly) and for tool-less turns.
+/// path, where `response.tool_calls` is used directly.
 ///
 /// Unknown-tool recovery is handled by `RunPolicy::unknown_tool`, so the model
 /// adapter preserves the provider-requested tool name.
 fn response_to_model_response(
     response: &ChatResponse,
     pformat_registry: &crate::openhuman::agent::pformat::PFormatRegistry,
+    parse_text_tool_calls: bool,
 ) -> ModelResponse {
     let (visible_text, tool_calls): (String, Vec<TaToolCall>) = if !response.tool_calls.is_empty() {
         let calls = response
@@ -113,7 +116,8 @@ fn response_to_model_response(
             })
             .collect();
         (response.text.clone().unwrap_or_default(), calls)
-    } else if let Some(text) = response.text.as_deref() {
+    } else if parse_text_tool_calls {
+        let text = response.text.as_deref().unwrap_or_default();
         let (prose, parsed) =
             crate::openhuman::agent::harness::parse_tool_calls_with_pformat(text, pformat_registry);
         if parsed.is_empty() {
@@ -134,7 +138,7 @@ fn response_to_model_response(
             (prose, calls)
         }
     } else {
-        (String::new(), Vec::new())
+        (response.text.clone().unwrap_or_default(), Vec::new())
     };
 
     let mut content = Vec::new();
@@ -632,7 +636,11 @@ impl ChatModel<()> for ProviderModel {
         // Provider usage (charged USD / context window / cache-creation-reasoning)
         // now reaches the event bridge via `UsageCarryMiddleware`, which reads it
         // off the returned `ModelResponse` (G1) — the adapter no longer carries it.
-        Ok(response_to_model_response(&response, &pformat_registry))
+        Ok(response_to_model_response(
+            &response,
+            &pformat_registry,
+            !request.tools.is_empty(),
+        ))
     }
 
     /// Stream the model response, forwarding openhuman's `ProviderDelta` events
@@ -650,6 +658,7 @@ impl ChatModel<()> for ProviderModel {
         // Positional layouts for the text-mode P-Format fallback (issue #4465);
         // built here so it can move into the `'static` producer task below.
         let pformat_registry = pformat_registry_from_request(&request);
+        let parse_text_tool_calls = !request.tools.is_empty();
         let provider = self.provider.clone();
         let model = self.model.clone();
         // Per-request temperature when set (see `invoke`), else the pinned value;
@@ -749,7 +758,11 @@ impl ChatModel<()> for ProviderModel {
                     // Provider usage rides the `Completed` response's crate `Usage`
                     // + raw (G1); `UsageCarryMiddleware` reads it off the folded
                     // response for the bridge, so the adapter no longer pushes here.
-                    ModelStreamItem::Completed(response_to_model_response(&resp, &pformat_registry))
+                    ModelStreamItem::Completed(response_to_model_response(
+                        &resp,
+                        &pformat_registry,
+                        parse_text_tool_calls,
+                    ))
                 }
                 Err(e) => {
                     // Streaming failures ride `ModelStreamItem::Failed(String)`, which
@@ -827,7 +840,7 @@ mod g1_usage_tests {
             }),
             reasoning_content: None,
         };
-        let model_response = response_to_model_response(&chat, &empty_registry());
+        let model_response = response_to_model_response(&chat, &empty_registry(), false);
 
         // Crate Usage carries every token breakdown natively.
         let usage = model_response.usage.expect("usage present");
@@ -860,7 +873,7 @@ mod g1_usage_tests {
             }),
             reasoning_content: None,
         };
-        let model_response = response_to_model_response(&chat, &empty_registry());
+        let model_response = response_to_model_response(&chat, &empty_registry(), false);
         assert!(
             model_response.raw.is_none(),
             "no charged USD / window ⇒ raw stays None"
@@ -879,9 +892,37 @@ mod g1_usage_tests {
             usage: None,
             reasoning_content: None,
         };
-        let model_response = response_to_model_response(&chat, &empty_registry());
+        let model_response = response_to_model_response(&chat, &empty_registry(), false);
         assert!(usage_info_from_response(&model_response).is_none());
     }
+
+    #[test]
+    fn tool_less_response_preserves_literal_tool_call_markup() {
+        let text = r#"Example: <tool_call>{"name":"lookup","arguments":{}}</tool_call>"#;
+        let chat = ChatResponse {
+            text: Some(text.to_string()),
+            ..Default::default()
+        };
+
+        let response = response_to_model_response(&chat, &empty_registry(), false);
+
+        assert_eq!(response.text(), text);
+        assert!(response.message.tool_calls.is_empty());
+    }
+
+    #[test]
+    fn tool_enabled_response_still_extracts_tool_call_markup() {
+        let chat = ChatResponse {
+            text: Some(r#"<tool_call>{"name":"lookup","arguments":{}}</tool_call>"#.to_string()),
+            ..Default::default()
+        };
+
+        let response = response_to_model_response(&chat, &empty_registry(), true);
+
+        assert_eq!(response.text(), "");
+        assert_eq!(response.message.tool_calls.len(), 1);
+        assert_eq!(response.message.tool_calls[0].name, "lookup");
+    }
 }
 
 #[cfg(test)]