Qwen chat template whitespace normalisation of tool-call and think blocks

Facts

• Normalisation sites in the Qwen3.8-27B template: every message body goes through `render_content(...)|trim`; the system text is trimmed; `reasoning_content` is trimmed (line 116); the last-query scan trims user text before testing for a `<tool_response>` wrapper. [source]
• Fixed re-wrap on render: a historical assistant turn is written as `<|im_start|>assistant\n<think>\n` + trimmed reasoning + `\n</think>\n\n` + trimmed content. The generation prompt is `<|im_start|>assistant\n<think>\n`. A model that emitted extra blank lines or trailing spaces around the think block is therefore rewritten to the canonical form on the next request. [source]
• Tool-call wrapper: `\n\n<tool_call>\n<function=NAME>\n` when the turn has text, bare `<tool_call>\n<function=NAME>\n` when it has none, `\n<tool_call>\n...` for later calls in the same turn; each parameter is `<parameter=KEY>\n` + value + `\n</parameter>\n`; the call ends `</function>\n</tool_call>`. [source]
• Value serialisation: a string argument is written verbatim; any other value is written as `args_value | tojson | safe`. Number formatting, nested-object key order and JSON spacing in the re-render come from the template runtime's `tojson`, not from the model's raw text. [source]
• Tool results: consecutive `tool` messages are merged into one `user` turn of `<tool_response>` blocks; the body is trimmed. [source]
• Maintainer position (QwenLM/Qwen3.8 #131, 2026-04-14): the normalisation difference "lacks a perfect solution" because token-level generation and message-level templating are separate; the suggested experiment is to remove the normalisation in the chat template. [source]
• Third-party claims: froggeric states it applies "strict single newline normalization at autoregressive boundaries" and rewrites `content | replace(...)` as `content.split(...) | join('')` because the C++ minja engine silently drops the whole text if the replaced string sits at index 0. Neither claim is shown with a measured token diff. [source]
• 2026-04-09/10: #131 reporter lists two template fixes (guard empty historical think; optionally drop historical think entirely) and a reproduction on oMLX plus OpenCode and Pi, and says llama.cpp-style backends show the same drift. [source]
• After 2026-08-13: froggeric v22.3 adds a prefix test and fuzz harness (see prefix-stability-regression-testing-of-chat-temp.md) but both only test template idempotence. [source]
• A template-only test cannot catch generation-versus-render drift: the froggeric tests render message lists twice and compare strings; they never tokenise model output. [source]
• If a server's reasoning parser trims the reasoning text and the template trims again, the pair is idempotent only when the model's first and last reasoning characters are non-whitespace. [source]
• `trim` on the final content means a model reply that ends in a newline is cached as `...\n<|im_end|>` but re-rendered as `...<|im_end|>`; the divergence sits one token before the end of the turn, so the loss is the turn's tail plus every later token. [source]
• On a hybrid model the divergence point forces a checkpoint restore or a full reprocess (held in hybrid-and-sliding-window-attention-kv-cache-rewinding.md), so a one-character difference costs as much as a template reorder. [source]
• Qwen maintainer: whitespace normalisation is a residual that cannot be fully fixed and may be removed experimentally. froggeric: strict single-newline normalisation yields a 100% prefix hit rate. These claims are not compared with a shared measurement; the froggeric figure is a documentation claim. [source]
• Whether llama.cpp's Qwen tool-call parser trims parameter values or keeps the model's trailing newline, which decides whether `<parameter>` values round-trip byte-exact. [source]
• Whether `tojson` in llama.cpp's Jinja runtime emits the same separators as the model's own JSON output for non-string arguments. [source]
• A measured token-level diff between a Qwen3.8 generation and its re-render on any Mac runtime. [source]
• The Qwen3.8-27B template trims message content, system text and reasoning_content before rendering [source]
• A historical assistant turn renders as an assistant header, a think block of trimmed reasoning, then trimmed content, joined by fixed newlines [source]
• The generation prompt is the assistant header followed by "<think>\n" unless enable_thinking is false, when it is "<think>\n\n</think>\n\n" [source]
• The first tool call of a turn is preceded by "\n\n" only when the turn has non-empty content [source]
• Non-string tool-call argument values are rendered with tojson and the safe filter, string values verbatim [source]
• Consecutive tool messages are merged into a single user turn of tool_response blocks with trimmed bodies [source]
• A Qwen maintainer on #131 suggests trying removal of the template's normalisation to see whether prefix mismatches improve [source]
• The #131 reporter reproduced the drift on oMLX with OpenCode and Pi and says llama.cpp-style backends show the same behaviour [source]
• froggeric documents "strict single newline normalization at autoregressive boundaries" as part of its 100% prefix-hit claim [source]
• froggeric replaced content|replace with split-and-join because minja drops the whole text when the replaced string is found at index 0 [source]
• froggeric's template tests render message lists as strings and never tokenise model output [source]
• A one-character whitespace difference between generated and re-rendered text ends the shared prefix at that character [source]
• Trimming final content in the template means a reply that ended in a newline is re-rendered one token shorter than it was cached [source]