fix: stop model calling tools for simple chat (hello → text, not tool_call)

Gemini sees 106 tools and calls one for every message including "hello".
Added explicit "do NOT call tools for simple chat" guidance to both the
system prompt and the tool injection block. Model now responds with text
for conversational messages and only calls tools when explicitly needed.

Also removed leftover debug log line.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Author: Darth-Hidious

crates/agent/src/agent_loop.rs

@@ -432,12 +432,13 @@ pub async fn run_turn(
         }
 
         // ── 2e. Emit streamed text deltas ─────────────────────────
+        tracing::debug!(delta_count = streaming_deltas.len(), "emitting streamed text deltas");
+        for delta in &streaming_deltas {
+            emit(AgentEvent::TextDelta {
+                text: delta.clone(),
+            });
+        }
         if !streaming_deltas.is_empty() {
-            for delta in &streaming_deltas {
-                emit(AgentEvent::TextDelta {
-                    text: delta.clone(),
-                });
-            }
             // Flush text so the TUI moves it from streaming buffer to chat history
             // before cost/turn-complete events arrive.
             emit(AgentEvent::TextFlush);

crates/agent/src/prompts.rs

@@ -262,6 +262,11 @@ const INTERACTIVE_PROMPT: &str = r#"You are PRISM, an interactive agent for mate
 - Use discover_capabilities, status, and tools when you need to inspect the current environment before planning.
 - Keep local, MARC27-hosted, and BYOC boundaries explicit in your reasoning when you choose a compute or storage path.
 
+# Tool Use
+- For general conversation, greetings, explanations, and questions that do not require live data, respond with plain text. Do NOT call tools for simple chat.
+- Only call tools when the user explicitly asks to search, compute, query, deploy, ingest, run workflows, or interact with the knowledge graph/platform.
+- When in doubt, respond with text first. The user will ask you to use tools if needed.
+
 # Result Quality
 - Cite providers, data sources, and workflow boundaries when they materially affect the answer.
 - Do not hallucinate materials properties, deployment state, job state, or command outcomes.
@@ -307,6 +312,11 @@ const AUTONOMOUS_PROMPT: &str = r#"You are PRISM, an autonomous agent for materi
 - Use discover_capabilities, status, and tools when you need to inspect the current environment before planning.
 - Keep local, MARC27-hosted, and BYOC boundaries explicit in your reasoning when you choose a compute or storage path.
 
+# Tool Use
+- For general conversation, greetings, explanations, and questions that do not require live data, respond with plain text. Do NOT call tools for simple chat.
+- Only call tools when the user explicitly asks to search, compute, query, deploy, ingest, run workflows, or interact with the knowledge graph/platform.
+- When in doubt, respond with text first. The user will ask you to use tools if needed.
+
 # Result Quality
 - Cite providers, data sources, and workflow boundaries when they materially affect the answer.
 - Do not hallucinate materials properties, deployment state, job state, or command outcomes.

crates/llm/src/lib.rs

@@ -483,15 +483,11 @@ impl LlmClient {
                 }
             }
 
-            debug!("MARC27 full_text ({} chars): {:?}", full_text.len(), &full_text[..full_text.len().min(300)]);
-
             // Parse tool calls — only take the FIRST batch (before any "Results:" hallucination)
             let tool_calls = parse_text_tool_calls(&full_text);
-            debug!("MARC27 parsed {} tool calls", tool_calls.len());
             // Only unique tool calls (LLM sometimes duplicates)
             let tool_calls = dedup_tool_calls(tool_calls);
             let content_text = strip_tool_call_blocks(&full_text);
-            debug!("MARC27 content_text after strip: {:?}", &content_text[..content_text.len().min(200)]);
 
             return Ok(ChatResponse {
                 message: ChatMessage {
@@ -832,8 +828,12 @@ fn build_tool_prompt_block(tools: &[ToolDefinition]) -> String {
     }
 
     block.push_str("\n\
+        ## IMPORTANT: When NOT to call tools\n\n\
+        For greetings, casual conversation, explanations, general knowledge questions, \
+        or anything that does not need live data — respond with plain text. \
+        Do NOT call tools for simple chat like \"hello\", \"what can you do?\", or \"explain X\".\n\n\
         ## How to call tools\n\n\
-        When a task requires data retrieval, computation, or platform interaction, call a tool:\n\n\
+        ONLY when a task explicitly requires data retrieval, computation, search, or platform interaction, call a tool:\n\n\
         ```tool_call\n\
         {\"name\": \"tool_name\", \"arguments\": {\"arg1\": \"value1\"}}\n\
         ```\n\n\