fix: make tool-use nudge conditional on detected tool intent

[sangwa]

The agentic loop previously nudged the LLM to use tools on every text response, making casual conversation impossible ("hi, how are you" would trigger up to 3 retry iterations with "please use tools" injected).

• Add llm_mentions_tool_intent() heuristic that checks whether the LLM's text response mentions an intent phrase ("I'll use", "let me call", etc.) near a known tool name within an 80-char window
• Only nudge when the heuristic fires, letting plain conversational responses pass through immediately
• Soften the nudge prompt to give the LLM an exit if no tool was needed
• Save the pre-nudge response and fall back to it if the post-nudge response is also text (false-positive recovery)

Summary by CodeRabbit

• Bug Fixes

  • Detects when the AI mentions intent to use a tool without calling it and performs a one-time nudge instead of looping.
  • Restores the original AI response if a nudge was a false positive; preserves tool results when tools are actually executed.

• Tests

  • Added tests covering detection of textual tool-intent, case and distance edge cases, and related behaviors.

[coderabbitai]

No actionable comments were generated in the recent review. 🎉

---

Walkthrough

The agent loop now detects textual intent to use tools and performs a single nudge when the LLM mentions tool usage but didn’t call any tools. It collects tool names before adding tool definitions to the LLM context, and saves the original LLM text in pre_nudge_response before nudging so the original can be returned if the nudge was a false positive. If tools run after a nudge, the saved pre_nudge_response is discarded and the final response reflects tool results.

Sequence Diagram

sequenceDiagram
    participant User as User
    participant LLM as LLM
    participant Agent as Agent Loop
    participant Detector as Tool Intent Detector
    participant Tools as Tool Executor

    User->>LLM: Sends prompt
    LLM-->>Agent: Returns response text
    Agent->>Agent: Check pre_nudge_response
    alt pre_nudge_response exists
        Agent-->>User: Return saved original response
    else no pre_nudge_response
        Agent->>Detector: Analyze response for intent (llm_mentions_tool_intent)
        Detector-->>Agent: Intent detected? (bool)
        Agent->>Agent: Check if any tools executed
        alt No tools executed & intent detected
            Agent->>Agent: Save response as pre_nudge_response
            Agent->>LLM: Send nudged prompt (one-time)
            LLM-->>Agent: Return post-nudge response
            Agent->>Tools: Execute tools if called
            Tools-->>Agent: Return tool results
            Agent->>Agent: Discard pre_nudge_response if tools ran
            Agent-->>User: Return final response (with tool results)
        else Tools executed or no intent
            Agent-->>User: Return current response
        end
    end

Loading

Estimated code review effort

🎯 3 (Moderate) | ⏱️ ~25 minutes

Poem

A rabbit nudged the LLM with care,
Kept the first reply tucked in its lair,
One gentle prod to call a tool true,
If nothing ran, the old answer came through—
A hop, a smile, the loop stays fair. 🐇

🚥 Pre-merge checks | ✅ 3

✅ Passed checks (3 passed)

Check name	Status	Explanation
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.
Title check	✅ Passed	The title 'fix: make tool-use nudge conditional on detected tool intent' directly and clearly describes the main change—making the tool nudging behavior conditional on detecting actual tool intent rather than applying it unconditionally.
Docstring Coverage	✅ Passed	Docstring coverage is 100.00% which is sufficient. The required threshold is 80.00%.

_{✏️ Tip: You can configure your own custom pre-merge checks in the settings.}

✨ Finishing Touches

• 📝 Generate docstrings

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Post copyable unit tests in a comment
• Commit unit tests in branch fix/conversational-messages

---

_{Comment @coderabbitai help to get the list of available commands and usage tips.}

[coderabbitai]

Actionable comments posted: 2

🤖 Prompt for all review comments with AI agents

Verify each finding against the current code and only fix it if needed.

Inline comments:
In `@src/agent/agent_loop.rs`:
- Around line 1115-1126: The current RespondResult::Text branch unconditionally
returns pre_nudge_response when present, which discards cases where the nudge
led to actual tool execution; change the logic inside the RespondResult::Text
match arm to only fall back to pre_nudge_response.take() when no tools were
executed (e.g., check a field like output.tool_calls.is_empty() or
output.tools_used == 0 or call a method like output.was_tool_called()),
otherwise let the tool-augmented result continue processing; update the
condition that wraps pre_nudge_response.take() to include this "no tools
executed" check while keeping the debug message and returning
AgenticLoopResult::Response(original) only in that case.

In `@src/util.rs`:
- Around line 71-76: The slice for the intent-matching window can panic on UTF-8
boundaries because window_end is computed as a byte index; update the
calculation to use floor_char_boundary to ensure window_end is a valid char
boundary before slicing. Locate the loop over INTENT_PHRASES where
lower.find(phrase) is used and modify the computation of window_end (and any use
of window_start/window_end for slicing) to call
lower.floor_char_boundary((window_start + 80).min(lower.len())) so the resulting
&lower[window_start..window_end] slice is safe.

[coderabbitai]

Actionable comments posted: 1

🤖 Prompt for all review comments with AI agents

Verify each finding against the current code and only fix it if needed.

Inline comments:
In `@src/util.rs`:
- Around line 71-76: The code incorrectly calls floor_char_boundary as a method
on lower; change the call to the standalone function by invoking
floor_char_boundary(&lower, (window_start + 80).min(lower.len())) (or whatever
the helper signature requires) instead of lower.floor_char_boundary(...), so in
the loop with INTENT_PHRASES use window_end = floor_char_boundary(&lower,
(window_start + 80).min(lower.len())); to get a valid char boundary.