fix: preserve text before tool-call XML in forced-text responses (#852)

* fix: preserve text before tool-call XML in forced-text responses (#789)

Local models (Qwen3, DeepSeek, GLM) emit <tool_call> XML even when no
tools are available (force_text mode). The existing strip_xml_tag()
discards everything from an unclosed opening tag onward, producing an
empty string that triggers the "I'm not sure how to respond" fallback.

Add truncate_at_tool_tags() — a code-region-aware pre-processing step
that truncates at the first tool-call XML tag BEFORE clean_response()
runs, preserving all useful text before the tag. Protect all 7
clean_response() call sites. Case-insensitive matching handles models
that emit <TOOL_CALL> or <Tool_Call> variants.

Secondary fix: add has_native_thinking() model detection to skip
<think>/<final> system prompt injection for models with built-in
reasoning (Qwen3, QwQ, DeepSeek-R1, GLM-Z1, etc.), preventing
thinking-only responses that clean to empty.

Wire with_model_name(active_model_name()) at all 9 production sites
that construct Reasoning, so the runtime model name (not static config)
drives system prompt generation.

126 new/updated tests covering truncation edge cases, code-block
awareness, Unicode, case-insensitivity, StubLlm integration for
complete/plan/evaluate_success/respond_with_tools paths, model
detection, and conditional system prompt generation.

Closes #789

Co-Authored-By: Claude Opus 4.6 <[email protected]>

* fix: address Copilot review — unclosed-only truncation, ASCII case folding

- truncate_at_tool_tags() now only truncates at UNCLOSED tool tags;
  properly closed tags (e.g. <tool_call>...</tool_call>) are left intact
  for clean_response() to strip normally, preserving any text after them
- Switch from to_lowercase() to to_ascii_lowercase() to prevent byte
  offset misalignment with non-ASCII characters whose lowercase form
  has different byte length (e.g. Kelvin sign U+212A)
- Add closing_tag_for() helper to derive closing tags from open patterns
- Fix doc comment: "fenced markdown code blocks or inline code spans"
  (not "indented", which find_code_regions() doesn't detect)
- Add regression tests: closed vs unclosed for each tag variant,
  Unicode + case-insensitive offset safety, and mixed closed/unclosed

Co-Authored-By: Claude Opus 4.6 <[email protected]>

* fix: minor review items — consistent ascii_lowercase, closing_tag_for tests

- Switch has_native_thinking() from to_lowercase() to to_ascii_lowercase()
  for consistency with truncate_at_tool_tags() approach
- Add unit tests for closing_tag_for(): standard tags, space-suffixed
  patterns, pipe-delimited tags, and exhaustive coverage of all
  TOOL_TAG_PATTERNS entries
- Add test for mixed closed+unclosed tags of different types

Co-Authored-By: Claude Opus 4.6 <[email protected]>

---------

Co-authored-by: Claude Opus 4.6 <[email protected]>

src/agent/commands.rs

@@ -405,7 +405,8 @@ impl Agent {
             .with_max_tokens(512)
             .with_temperature(0.3);
 
-        let reasoning = Reasoning::new(self.llm().clone());
+        let reasoning = Reasoning::new(self.llm().clone())
+            .with_model_name(self.llm().active_model_name());
         match reasoning.complete(request).await {
             Ok((text, _usage)) => Ok(SubmissionResult::response(format!(
                 "Thread Summary:\n\n{}",
@@ -453,7 +454,8 @@ impl Agent {
             .with_max_tokens(512)
             .with_temperature(0.5);
 
-        let reasoning = Reasoning::new(self.llm().clone());
+        let reasoning = Reasoning::new(self.llm().clone())
+            .with_model_name(self.llm().active_model_name());
         match reasoning.complete(request).await {
             Ok((text, _usage)) => Ok(SubmissionResult::response(format!(
                 "Suggested Next Steps:\n\n{}",

src/agent/compaction.rs

@@ -227,7 +227,8 @@ Be brief but capture all important details. Use bullet points."#,
             .with_max_tokens(1024)
             .with_temperature(0.3);
 
-        let reasoning = Reasoning::new(self.llm.clone());
+        let reasoning = Reasoning::new(self.llm.clone())
+            .with_model_name(self.llm.active_model_name());
         let (text, _) = reasoning.complete(request).await?;
         Ok(text)
     }

src/agent/heartbeat.rs

@@ -303,7 +303,8 @@ impl HeartbeatRunner {
             .with_max_tokens(max_tokens)
             .with_temperature(0.3);
 
-        let reasoning = Reasoning::new(self.llm.clone());
+        let reasoning = Reasoning::new(self.llm.clone())
+            .with_model_name(self.llm.active_model_name());
         let (content, _usage) = match reasoning.complete(request).await {
             Ok(r) => r,
             Err(e) => return HeartbeatResult::Failed(format!("LLM call failed: {}", e)),

src/agent/worker.rs

@@ -212,7 +212,8 @@ impl Worker {
         let job_ctx = self.context_manager().get_context(self.job_id).await?;
 
         // Create reasoning engine
-        let reasoning = Reasoning::new(self.llm().clone());
+        let reasoning = Reasoning::new(self.llm().clone())
+            .with_model_name(self.llm().active_model_name());
 
         // Build initial reasoning context (tool definitions refreshed each iteration in execution_loop)
         let mut reason_ctx = ReasoningContext::new().with_job(&job_ctx.description);