---
url: >-
  https://adk.nht.io/api/@nhtio/adk/batteries/functions/stripEnvelopeSpecialTokens.md
---

# Function: stripEnvelopeSpecialTokens()

```ts
function stripEnvelopeSpecialTokens(text: string): string;
```

Defined in: [src/batteries/llm/chat\_common/helpers.ts:109](https://github.com/NHTIO/ADK/blob/v1.20260713.1/src/src/batteries/llm/chat_common/helpers.ts#L109)

Strip the non-semantic envelope/turn-boundary special tokens (see ENVELOPE\_SPECIAL\_TOKEN\_RE)
from decoded model text before it reaches the parser layer.

## Parameters

| Parameter | Type     |
| --------- | -------- |
| `text`    | `string` |

## Returns

`string`

## Remarks

**Why this exists.** The transformers.js streaming path decodes with `skip_special_tokens:false` (it
must — the live prose-stop gate watches for tool/think markers, which ARE special tokens). That leaves
envelope tokens like Llama's `<|python_tag|>{json}<|eom_id|>` or ChatML's trailing `<|im_end|>` in the
accumulated text, so the JSON tool-call parsers see `<|python_tag|>{…}` and decline — even though the
NON-streaming path (which decodes with `skip_special_tokens:true`) parses the identical call fine.
Normalising here makes the stream and batch paths parse equivalent text. Surfaced by the deep model
matrix (Llama-3.2-1B + Qwen2.5-Coder tool calls passed on batch, failed on stream).