# Bug: Tool-call instruction template is read aloud as part of the agent's speech (Now you are invoking function tool ...) **URL:** \ **Category:** Support Help\ **Tags:** bug\ **Created:** [April 29, 2026, 9:22pm UTC](https://community.retellai.com/t/bug-tool-call-instruction-template-is-read-aloud-as-part-of-the-agents-speech-now-you-are-invoking-function-tool/2537 "2026-04-29T21:22:58Z")\ **Posts on this page:** 4\ **Page:** 1
**Author:** ![andrew3](https://yyz2.discourse-cdn.com/flex008/user_avatar/community.retellai.com/andrew3/32/1713_2.png) [@andrew3](https://community.retellai.com/u/andrew3)\ **Post date:** [April 29, 2026, 9:22pm UTC](https://community.retellai.com/t/bug-tool-call-instruction-template-is-read-aloud-as-part-of-the-agents-speech-now-you-are-invoking-function-tool/2537/1 "2026-04-29T21:22:58Z")
Severity: High — caller hears English system-prompt text in the middle of a Japanese conversation, including the literal tool ID. Summary When an agent invokes a custom tool with speak\_during\_execution: true, the instruction template that tells the LLM what filler to say is being concatenated into the agent’s spoken output instead of being consumed. The English system text plays out loud through TTS, sometimes twice in the same turn. Reproduction - Agent ID: agent\_2c260cc96dc5143125a5664790 - Conversation flow: conversation\_flow\_963b78c1fdbd (single custom tool: rag\_lookup\_\) - Voice: Cartesia (Aiko) - LLM: Retell-managed (GPT-5.1) - Language: multi - speak\_during\_execution: true on the tool node - Call type in our reproduction: web call - Caller utterances: Japanese - Frequency: at least 2 of ~8 test calls in one test session Reference calls Example 1 — short leak call\_69cf7c1215dfefe7eb4005e9ab1 [https://dashboard.retellai.com/call-history?history=call\_69cf7c1215dfefe7eb4005e9ab1](https://dashboard.retellai.com/call-history?history=call_69cf7c1215dfefe7eb4005e9ab1) Excerpt of what the caller heard (one continuous agent utterance): ▎ お子様が０歳で、プールをご利用いただけるかどうかでございますね。こちらを確認いたしますので、少々お待ちくださいませ。 ▎ Now you are invoking function. ▎ お子様が０歳でプールをご利用いただけるかについてでございますね。こちらを確認いたしますので、少々お待ちくださいませ。 The English string Now you are invoking function. is spoken aloud, sandwiched between two paraphrases of the same Japanese filler line. Example 2 — full instruction-template leak call\_7dec8e72f3a888115a1aaef8ac6 [https://dashboard.retellai.com/call-history?history=call\_7dec8e72f3a888115a1aaef8ac6](https://dashboard.retellai.com/call-history?history=call_7dec8e72f3a888115a1aaef8ac6) Excerpt: ▎ レストランでの喫煙についてでございますね。館内のルールをお調べいたしますので、少々お待ちくださいませ。 ▎ Now you are invoking function tool rag\_lookup\_b28fdb84-b7e8-5f97-bab4-324df5d6e43f, and you should come up with what to say to user during execution given this instruction: Acknowledge the caller’s question by briefly paraphrasing what ▎ they asked about, then let them know you’re looking it up. Keep it natural and conversational. Example: “Let me check on [topic] for you, one moment please.” ▎ レストランでの喫煙についてでございますね。館内のルールをお調べいたしますので、少々お待ちくださいませ。 ▎ Now you are invoking function tool rag\_lookup\_b28fdb84-b7e8-5f97-bab4-324df5d6e43f, and you should come up with what to say to user during execution given this instruction: … In this call the leak is the complete speak\_during\_execution instruction template — including the literal tool ID and the prompt’s own example sentence (“Let me check on [topic] for you, one moment please.”). It plays through twice in the same utterance. What we’d expect The speak\_during\_execution instruction is a system-level directive telling the LLM what filler to produce; only the LLM’s response to that instruction (the filler itself) should reach the TTS. The instruction text and the tool ID should never be spoken. What we’d like investigated 1. Why is the speak\_during\_execution instruction text being forwarded to TTS along with the LLM’s filler response? 2. Is this consistent across LLM versions (GPT-5.1 vs. GPT-4.1)? We have a hunch this regressed with GPT-5.1. 3. Is there a way to suppress the instruction echo — e.g. a flag on the tool, or a known-good prompt template — until the upstream fix lands? ---
**Author:** ![mark1](https://yyz2.discourse-cdn.com/flex008/user_avatar/community.retellai.com/mark1/32/2524_2.png) [@mark1](https://community.retellai.com/u/mark1)\ **Post date:** [April 29, 2026, 9:25pm UTC](https://community.retellai.com/t/bug-tool-call-instruction-template-is-read-aloud-as-part-of-the-agents-speech-now-you-are-invoking-function-tool/2537/2 "2026-04-29T21:25:26Z")
Hey @andrew3 Thank you for the details. I’ve forwarded them to our team for review. We’ll get back to you as soon as we have an update. Best regards ---
**Author:** ![mark1](https://yyz2.discourse-cdn.com/flex008/user_avatar/community.retellai.com/mark1/32/2524_2.png) [@mark1](https://community.retellai.com/u/mark1)\ **Post date:** [April 30, 2026, 1:03pm UTC](https://community.retellai.com/t/bug-tool-call-instruction-template-is-read-aloud-as-part-of-the-agents-speech-now-you-are-invoking-function-tool/2537/3 "2026-04-30T13:03:05Z")
Hey @andrew3 GPT‑5.1 is meaningfully more prone to echoing the system directive verbatim than GPT‑4.1, especially in non-English conversations, so we believe your hunch is correct. Team checked the logs for both of these calls, and they appear to be the LLM hallucination leaking the prompts rather than a system bug. We’d recommend going back to GPT-4.1 for now. Team can see that our prompt for tool calls is a bit unclear, so they will work on updating them. The fix should be done by today. Thank You ---
**Author:** ![andrew3](https://yyz2.discourse-cdn.com/flex008/user_avatar/community.retellai.com/andrew3/32/1713_2.png) [@andrew3](https://community.retellai.com/u/andrew3)\ **Post date:** [April 30, 2026, 6:47pm UTC](https://community.retellai.com/t/bug-tool-call-instruction-template-is-read-aloud-as-part-of-the-agents-speech-now-you-are-invoking-function-tool/2537/4 "2026-04-30T18:47:30Z")
Thanks for the quick update! We’ll use 4.1 for production use and keep trying 5.1 to see if your fix improves things.