[mark-yong]

Fixes #30. Supersedes #63.

Request

With tools in the request, any <tool_call> or <function= in reasoning or answer text started a call. When a model quotes the tag in prose, for example while discussing a parser, the whole response aborts at EOS with tool call requires a function tag or incomplete function-tag tool call (#30). Both the reasoning variant and the answer-text variant reproduce on Qwen3.8-Flash-Next under Hermes.
GLM's argument-tag parser has the same defect in answer text: a quoted <tool_call> aborts with incomplete GLM tool call, and a quoted tag followed by a real call cuts the answer short and returns a call whose function name is the intervening prose. GLM reasoning is unaffected because its <tool_call> does not end reasoning.

Change

• ToolDialect::starts_call (default: always a call) lets a dialect decide whether text at an opener begins a call. ThinkReasoning (<tool_call> ending reasoning) and content-mode scanning in TaggedOutput skip mentions and continue to the next opener instead of stopping at the first match. An unresolved prefix is held until the next tokens decide.
• FunctionTags confirms an opener only when it is followed by a <function=NAME> header and then a <parameter or </function> tag. Confirmation is stricter than parse_tool's own header grammar on surrounding whitespace: <function=NAME > and </function > parse as calls without the check and are mentions with it.
• ArgumentTags (GLM) confirms <tool_call> only when it is followed by a function name and then <arg_key> or </tool_call>, matching the GLM 5.3 chat template.
• At finish an undecided opener is still parsed as a call, so a truncated call keeps failing at EOS and is dropped at the output limit as before.
• docs/api.md documents the rule, its remaining limitation, and the dialects that keep their current behavior: Gemma 4, MiniMax, DeepSeek DSML and XTML.
  Confirmation is structural rather than requiring the opener at line start, because Step 5 and MiMo chat templates place calls directly after content text (content<tool_call>…, and inline <tool_call><function=…> for MiMo).

Scope across recipes

The check is a dialect method. The function-tag rule applies to every recipe using the Qwen3-Coder format: Qwen 3.8, Qwen 3.8 dense, Nemotron-H, Step 5 (wrapped) and MiMo V2 (compact). The GLM rule applies to GLM 5.3.
No Protobuf or package change. Existing packages pick up the fix with the new Flare build.

Validation

• Helios lib suite (--features runtime-package,server, release, 32-core x86_64 build host): 1213 passed, 21 ignored, 5 failed. Unmodified bce032c0 fails the same five (1210 passed).
  • Four are engram::reader staging allocations the build host cannot register (ENOMEM).
  • The fifth, package::cache::selection_tests::failed_default_clear_keeps_the_selected_entry, expects a 0o555 refs directory to reject a write, which cannot fail when the suite runs as root.
• The three new tests pass individually: output::chatml::tests::prose_tool_tag_mentions_remain_text_with_registered_tools, output::chatml::tests::compact_calls_directly_after_text_still_parse and output::argument_tags::tests::prose_tool_tag_mentions_remain_answer_text. Together they cover answer and reasoning mentions, unconfirmed and partial inline tags, a mention followed by a real call, and a compact call directly after text, at every two-way split point.
• The at-EOS contract stays pinned by the existing terminal_cause_preserves_incomplete_call_errors_and_never_fabricates_calls.
• Formatting: rustfmt --check (edition 2024) reports no formatting difference this change introduces; the differences it still reports in chatml.rs and server/output.rs are the ones it reports on main. CI does not gate cargo fmt.
• The four captured Qwen3.8-Flash-Next streams from Qwen function-tag parser: <tool_call> mentioned in reasoning prose aborts the stream ("tool call requires a function tag") #30 (one token per chunk, 253 KB prompt, 33 tools) were replayed through OutputPipeline. On main, all four fail with tool call requires a function tag. On this branch, all four parse, reasoning equals the raw reasoning with every quoted mention kept (2, 2, 5 and 3 mentions), and all six real terminal calls are recovered.
• A probe on the unchanged GLM parser reproduced the abort and the misnamed call for the inputs in the new GLM test.

Live Flare service, Qwen3.8 27B

The qwen3.8-27b recipe uses the same Qwen function-tag parser as Qwen3.8 Flash. One package exported at bce032c0 on an RTX PRO 6000 (SM120), served on the same GPU by Flare built from bce032c0 and from this branch. 36 requests per build with one tool declared: tags quoted in answer text (<tool_call>, <function=get_weather>), reasoning that quotes the tags, ordinary tool calls, and a quoted tag followed by a real call; streaming and non-streaming /v1/chat/completions and streaming /v1/messages, thinking on and off, three seeds.

• bce032c0: 24 of 24 quoted-tag requests fail. Answer-text quotes stop at the backtick with incomplete function-tag tool call; quotes in reasoning or before a call fail with tool call requires a function tag; non-streaming returns HTTP 500. All 12 ordinary tool calls succeed.
• This branch: 0 of 36 fail. All 12 answer-text requests return the quoted tag and the closing END OF ANSWER (for example The parser chokes on a literal `<tool_call>` tag in normal text. END OF ANSWER.). All 12 ordinary calls return get_weather {"city":"Paris"}. 5 of 6 mention-then-call requests return the quoted sentence as content and the call; in the sixth the model omitted the <function=...> header (see Limitations).
• 20 further requests with logprobs: every generated token is accounted for in reasoning, content, or tool calls, so held mention prefixes are not dropped.

Live Flare service, Qwen3.8 Flash

One package exported at bce032c0 from nvidia/Qwen3.8-Flash-Next-FP4-Local@b9ce18a6 on an RTX PRO 6000 (SM120), served on the same GPU by the same two builds with the same 36 requests.

• bce032c0: 22 of 36 fail, every request in which the model wrote a tag: incomplete function-tag tool call for answer-text quotes and tool call requires a function tag for quotes in reasoning or before a call; non-streaming returns HTTP 500. All 12 ordinary tool calls succeed, as do 2 mention-then-call requests in which the model skipped the sentence and only called the tool.
• This branch: 0 of 36 fail. All 12 answer-text requests return the quoted tag and the closing END OF ANSWER. All 12 ordinary calls return get_weather {"city":"Paris"}. All 6 mention-then-call requests return the call; the 4 in which the model wrote the sentence return it with the quoted tag as content.
• Reasoning that quotes the tags completes without error in 6 of 6 requests. Two /v1/messages answers end mid-sentence before a literal </tool_call>, and one returns the model's complete example call NAME {"param_name":"value"} (see Limitations).
• The mid-sentence endings are the model's own EOS, not parser loss. In 20 further streaming chat requests with logprobs, completion usage equals the logprob entries plus the terminating EOS in every request; 7 end where a literal </tool_call> would follow. A single-token continuation of one of them with ignore_eos gives <|im_end|> about 0.97 and </tool_call> about 0.02, with or without tools declared.
  Not run: a live Flare service for Step 5, MiMo, Nemotron-H, or GLM; Helios integration tests; Heliopause, Python, and CUDA tests.

Limitations

• Step 5, MiMo, Nemotron-H and GLM are covered only by unit tests built from their chat templates, not by captured model output.
• A complete call example in prose, fenced or not, still parses as a call. Written inline in reasoning, it ends reasoning and becomes a call. vLLM and SGLang also treat these as calls.
• A real call whose <tool_call> is not followed by a <function=NAME> header is now returned as text instead of failing with tool call requires a function tag. The live run produced one (<tool_call>\n<parameter=city>…), returned with finish_reason: stop. Under this rule it is indistinguishable from a mention.
• When asked to discuss the tag format, the model often wrote a complete example call in its reasoning: 11 of 32 reasoning-mention or mention-then-call requests (including the logprob runs) were affected. Ten returned an extra call; one, whose example elided the body with ..., failed with tool function requires parameter or closing tag. Complete example calls parse as calls, as noted above.
• Gemma 4 (<|tool_call>) and MiniMax (namespace-prefixed <tool_call>) keep their current behavior and would need the same treatment if a quoted opener is observed in prose. DeepSeek DSML and XTML parse through their own state machines.
• Qwen3.8 Flash sometimes ends its turn where it would write a literal </tool_call> (7 of the 20 logprob requests above), so a response that discusses the tags can stop mid-sentence with finish_reason: stop. A parser cannot recover text the model never generated; on bce032c0 the same requests fail earlier with the parser error.
• Think-tag mentions (Literal <think>/</think> text in model output is parsed as reasoning delimiters #48) are a separate fix.
• PR CI has not run: the Native serving workflow sits at action_required for fork PRs. main's own push run on bce032c0 fails at Test native CLI, and that failure reproduces locally in flare/tests/unverified_payload.rs, so a red CLI step on this branch is not caused by it.

Written with AI assistance; reviewed before opening.

[mark-yong]

#59 (GLM argument tags) touches the same two files as this PR, argument_tags.rs and tagged.rs, but different code: #59 changes argument-key parsing and adds array coercion in parameter_value, while this PR adds the opener check (ToolDialect::starts_call). git merge-tree of the two heads (a6317630, 1f38c71d) reports no conflicts. They fix different defects, so merging one does not replace the other.