fix(llm): nudge LLM when it expresses tool intent without calling tools (#653)

* fix(llm): nudge LLM when it expresses tool intent without calling tools

Non-Anthropic models (especially GLM-5 via NEAR AI) frequently output
text like "Let me search for X" without including tool_calls, creating
a frustrating loop where the user waits but nothing happens.

Add llm_signals_tool_intent() detection that matches intent phrases
("let me search", "I'll fetch") while excluding conversational phrases
("let me explain", "let me know") and content inside code blocks.
When detected, inject a nudge message telling the model to actually
call the tool. Cap at 2 consecutive nudges to avoid infinite loops.

Applied to all three agentic loops: dispatcher (interactive chat),
agent/worker (background jobs), and worker/runtime (sandbox containers).

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

* fix(nudge): address PR #653 review comments

1. Update doc comment to match implementation (code blocks only, not quotes)
2. Use match_indices() instead of find() to check all prefix occurrences
3. Add !available_tools.is_empty() guard in dispatcher nudge check
4. Reset consecutive_tool_intent_nudges on non-intent text responses
5. Add regression test for shadowed prefix detection

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

* fix(nudge): address second round of PR #653 review comments

1. Strip double-quoted strings in tool-intent detection to avoid false
   positives on quoted prose like `"Let me search the database"`.
2. Only reset consecutive_tool_intent_nudges when text does NOT signal
   intent — preserves the 2-nudge cap when intent is detected but cap
   is already reached.
3. Fix assertion message in nudge_cap test to report correct call index.
4. Add regression test for quoted strings outside code blocks.

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

---------

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

src/agent/dispatcher.rs

@@ -138,6 +138,8 @@ impl Agent {
         let force_text_at = max_tool_iterations;
         let nudge_at = max_tool_iterations.saturating_sub(1);
         let mut iteration = 0;
+        const MAX_TOOL_INTENT_NUDGES: u32 = 2;
+        let mut consecutive_tool_intent_nudges: u32 = 0;
         loop {
             iteration += 1;
             // Hard ceiling one past the forced-text iteration (should never be reached
@@ -294,6 +296,24 @@ impl Agent {
 
             match output.result {
                 RespondResult::Text(text) => {
+                    // Nudge the LLM if it expressed tool intent without calling tools.
+                    // This is common with non-Anthropic models (e.g. GLM-5 via NEAR AI)
+                    // that output "Let me search…" but don't issue tool_calls.
+                    if !force_text
+                        && !context.available_tools.is_empty()
+                        && consecutive_tool_intent_nudges < MAX_TOOL_INTENT_NUDGES
+                        && crate::llm::llm_signals_tool_intent(&text)
+                    {
+                        consecutive_tool_intent_nudges += 1;
+                        tracing::info!(
+                            iteration,
+                            "LLM expressed tool intent without calling a tool, nudging"
+                        );
+                        context_messages.push(ChatMessage::assistant(&text));
+                        context_messages.push(ChatMessage::user(crate::llm::TOOL_INTENT_NUDGE));
+                        continue;
+                    }
+
                     // Strip internal "[Called tool ...]" text that can leak when
                     // provider flattening (e.g. NEAR AI) converts tool_calls to
                     // plain text and the LLM echoes it back.
@@ -304,6 +324,7 @@ impl Agent {
                     tool_calls,
                     content,
                 } => {
+                    consecutive_tool_intent_nudges = 0;
                     // Add the assistant message with tool_calls to context.
                     // OpenAI protocol requires this before tool-result messages.
                     context_messages.push(ChatMessage::assistant_with_tool_calls(

src/agent/worker.rs

@@ -305,6 +305,8 @@ Report when the job is complete or if you encounter issues you cannot resolve."#
         let mut iteration = 0;
         const MAX_CONSECUTIVE_RATE_LIMITS: usize = 10;
         let mut consecutive_rate_limits = 0usize;
+        const MAX_TOOL_INTENT_NUDGES: u32 = 2;
+        let mut consecutive_tool_intent_nudges: u32 = 0;
 
         // Initial tool definitions for planning (will be refreshed in loop)
         reason_ctx.available_tools = self.tools().tool_definitions().await;
@@ -502,17 +504,34 @@ Report when the job is complete or if you encounter issues you cannot resolve."#
                             }),
                         );
 
-                        // Give it one more chance to select a tool
-                        if iteration > 3 && iteration % 5 == 0 {
-                            reason_ctx.messages.push(ChatMessage::user(
-                                "Are you stuck? Do you need help completing this job?",
-                            ));
+                        // Nudge the LLM if it expressed tool intent without calling tools
+                        let signals_intent = !reason_ctx.available_tools.is_empty()
+                            && crate::llm::llm_signals_tool_intent(&response);
+                        if signals_intent && consecutive_tool_intent_nudges < MAX_TOOL_INTENT_NUDGES
+                        {
+                            consecutive_tool_intent_nudges += 1;
+                            tracing::info!(
+                                job_id = %self.job_id,
+                                "LLM expressed tool intent without calling a tool, nudging"
+                            );
+                            reason_ctx
+                                .messages
+                                .push(ChatMessage::user(crate::llm::TOOL_INTENT_NUDGE));
+                        } else if !signals_intent {
+                            consecutive_tool_intent_nudges = 0;
+                            if iteration > 3 && iteration % 5 == 0 {
+                                // Generic fallback nudge
+                                reason_ctx.messages.push(ChatMessage::user(
+                                    "Are you stuck? Do you need help completing this job?",
+                                ));
+                            }
                         }
                     }
                     RespondResult::ToolCalls {
                         tool_calls,
                         content,
                     } => {
+                        consecutive_tool_intent_nudges = 0;
                         // Model returned tool calls - execute them
                         tracing::debug!(
                             "Job {} respond_with_tools returned {} tool calls",
@@ -558,6 +577,7 @@ Report when the job is complete or if you encounter issues you cannot resolve."#
                     }
                 }
             } else if selections.len() == 1 {
+                consecutive_tool_intent_nudges = 0;
                 // Single tool: execute directly
                 let selection = &selections[0];
                 tracing::debug!(