[leehack]

Summary

Refs #1531. Refs #1874 (0.19.4 plan, lane 4).

Stacked on #1882 (base fix/1865-thinking-tags-text at a6ea0316b). Review only the commits after a6ea0316b, and retarget to main once #1882 merges.

Qwen3-Coder XML replies (Qwen3-Coder, Qwen3.5, StepFun3.5-Flash) now parse and stream as llama.cpp cd959bae9 parses them (common/parsers/qwen3-coder.cpp):

• Text after the calls is dropped. The content ends at the first <tool_call>. The calls are those of the valid scopes after it, and anything after them is dropped: trailing text, a second <function=> in one scope, an empty or invalid scope, or an unknown tool. Before, that text stayed as content, and an invalid first scope kept the whole reply as content.
• One call without parallelToolCalls. Without a grammar, llama.cpp takes only the first call when parallel tool calls are off. llamadart now does the same.
• A truncated last scope still counts. A finished reply whose last valid call is missing its </tool_call>, or has only part of it, still returns that call.
• The Qwen3-Coder template has no reasoning. It has no <think>, so llama.cpp takes no thought from its replies, and a stray </think> stays as content (lane 3's 128 Plan.</think> rows).
• StepFun3.5-Flash with thinking off now ends its thought at </think> or <tool_call>. llama.cpp keeps the thought open in its prompt (<think>\n); llamadart closes it, as it does for Seed-OSS and Cohere2 MoE. Before, the stream held the whole reply. Those three formats now share one registry trait (reopensThoughtClosedForThinkingOff) in place of two OR checks.

Contract change for review: before, an invalid Qwen call stayed in the reply as content (#531's malformed-call tests). Now it is dropped, as llama.cpp drops it. The affected tests are updated to llama.cpp's results.

Root cause. parseXmlToolCalls treated any text it could not parse as a call as content, and it kept every call. llama.cpp's PEG parser does neither: it stops at the end of the last valid call, takes min..1 call groups, and has no end().

Production-readiness scope

• User-facing scope: Qwen3-Coder, Qwen3.5 and StepFun3.5-Flash tool-call replies give the content, reasoning and calls llama.cpp gives, in the final parse and in every stream chunking.

• Supported platforms/paths: All. The change is in shared Dart template parsing (native llama.cpp and WebGPU bridge).

• Unsupported or intentionally unavailable paths: Each item below still differs from llama.cpp on purpose. They are recorded as left out in test/fixtures/qwen3_coder_xml_upstream.md.

  • A reply cut inside a call: llama.cpp emits unterminated JSON arguments; llamadart emits no call.
  • Strict newline framing (<tool_call><function=, </function></tool_call>): llamadart still reads these forms.
  • A partial parse of an unclosed scope.
  • Under required, a reply with no valid call: llama.cpp fails it; llamadart returns content.

• Out of scope / follow-ups:

  • A bare <function=NAME> call without <tool_call> for the Qwen3-Coder template.
  • The Seed-OSS leniencies in #1542: llama.cpp fails those replies, so llamadart needs a policy decision first.
  • The Qwen3.5/StepFun required + forced-open grammar.

  These are reported on #1874.

Completeness checklist

• Declared scope is fully implemented, or explicitly reduced above.
• Unsupported platform/option combinations fail loudly with actionable diagnostics or are clearly documented as unavailable. (No new unsupported path.)
• Public API docs, README/website docs, examples, support matrices, and changelog entries are updated where relevant. (Both changelogs, chat-template-and-parsing.md, template-engine-internals.md, and two skills/ SKILL.md files. No public API change.)
• Regression coverage covers the original issue plus key negative/version-skew paths where relevant.
• Public capability changes update the validation coverage registry with a pack or a reasoned, issue-linked exemption; inventory checks pass (or N/A with scope). N/A: parser parity fix, no capability change.
• Security/privacy review completed: no secrets, bearer tokens, signed URLs, or raw secret-bearing paths leak through logs, cache keys, metadata, errors, or snapshots.
• Follow-up work that is useful but not required for this PR is tracked in GitHub Issues, or explicitly marked "None" above.

PR type guidance

• Bugfix PR: The root cause and the fix are above.
  • New upstream fixture. test/fixtures/qwen3_coder_xml_upstream.json records 18 cases on 5 template configs: Qwen3-Coder (thinking on), and Qwen3.5-0.8B and StepFun3.5-Flash (thinking on and off), each with parallel off and on. Every case has llama.cpp's finished parse, and its partial parse where that is meaningful. The cases cover:
    • text after a call, and two calls with text after them
    • two functions in one scope
    • an empty first scope
    • an unknown tool between calls
    • a later scope with no end, part of an end, or cut in its name
    • a scope tag in text
    • </think> followed by a JSON or XML call
    • a thought followed by text and a call
  • New test. test/integration/core/engine/qwen3_coder_xml_upstream_test.dart (180 tests) checks the parse against llama.cpp. It checks the stream against the parse for the whole reply, byte by byte, in 3-character chunks and split in two at every point. 125 of its tests fail on base.
  • Unit tests. The endsAtLastCall group in xml_tool_call_format_test.dart and the single-call mark in template_internal_metadata_test.dart.

Test Plan

• dart run tool/prepare_workspace.dart
• dart format --output=none --set-exit-if-changed .
• dart analyze, plus dart run tool/testing/check_platform_boundaries.dart
• dart test -p vm -j 1 --exclude-tags local-only (isolated TMPDIR): 12,658 passed, 55 skipped, and 6 failed. All 6 are windows_registry_qualification_test.dart, which needs pwsh and this host has none
• dart test -p chrome --exclude-tags local-only: 4,665 passed, 1 skipped
• coverage-lib: run by the dispatched CI below (ci.yml collects coverage on manual runs)
• ./tool/testing/run_template_parity_suites.sh: detection 74, diagnostic 103, unit template 1,873, upstream parser 10, GBNF acceptance 79, all passed

Matrix Evidence

Matrix row	Scope covered	Platform / model / backend	Result	Evidence / notes
root-vm	native-safe unit/integration	macOS arm64 VM	PASS	12,658 passed, 55 skipped, and 6 failed. All 6 are windows_registry_qualification_test.dart, which needs pwsh and this host has none
root-chrome	browser-compatible tests	Chrome	PASS	4,665 passed, 1 skipped
template-parity	detection, render and parse parity; GBNF acceptance	macOS arm64	PASS	detection 74, diagnostic 103, unit template 1,873, upstream parser 10, GBNF acceptance 79
structured-output-adversarial	parse, partial and every-chunking stream against llama.cpp; auto/required/none tool choice; thinking on/off; parallel on/off	llama.cpp cd959bae9 common_chat_session oracle	PASS	See the sweep tables below. 0 records regress.
high-risk-exact-head-independent-qa	independent blocking review of the exact head	—	PENDING	Required before mark-ready
coverage-lib	lib/ line coverage	Linux CI	PASS	Dispatched ci.yml run
Other rows	—	—	N/A	No runtime, backend, grammar, release or LiteRT-LM change

Lane sweep. It covers 6 XML templates (Qwen3-Coder, Qwen3.5-0.8B and 4B, StepFun3.5-Flash, Seed-OSS and Nemotron-3-Nano), 48,264 records in all, thinking on and off, and 6 tool modes (no tools, none, auto, required, and auto and required with parallel). The corpus includes the #1539/#1542 shapes. Each record is compared with llama.cpp cd959bae9:

Measure	Base a6ea0316b	Head	Regressions
Finished parse = llama.cpp	13,430	19,608	0
Partial parse = llama.cpp	13,710	19,888	0
Every chunking = parse	23,800	23,884	0

Broad corpora (lane 3's, all formats; finished parse = llama.cpp). Only the three Qwen3-Coder XML templates change.

Corpus	Base	Head	Regressions
bare	59,640	60,604	0
full	66,990	67,658	0
native	3,286	3,572	0
vbare	18,012	18,100	0
vfull	20,152	20,256	0

Mutations: 19 of 20 killed.

Area	Mutants
Handler	Q1–Q4
Call parser	X1–X6
Engine	E1–E3
Stream	S1, S3
Registry	R1–R3 (Qwen, Seed-OSS, Cohere2 MoE)
Thought marks	T1

S2 (no single-call mark on the stream's partial parse) survived as equivalent, because partial calls only decide whether the reply has calls. That plumbing was removed (47a86a526).

Review Notes

• Independent review status: pending (a fresh auditor session, at the exact head).
• CI status / head SHA: 47a86a526, dispatched ci.yml run 42745372367 (success, coverage gate included).

High-risk regression review

• Classification: high-risk (backendRuntime, structuredOutput)
• Implementation task: lane 4 of #1874
• Independent blocking QA task: pending
• Exact head / current base: 47a86a526 / fix/1865-thinking-tags-text a6ea0316b
• Production-branch deletion, bypass, or miswire proof: the 19 killed mutants above, covering the handler, parser, engine, stream and registry call sites
• Affected-family real-model/artifact evidence: primary llama.cpp cd959bae9 emissions on the vendored Qwen3-Coder, Qwen3.5 and StepFun3.5-Flash templates (durable fixture plus sweeps). No model weights were needed: the change is in parsing only.
• Explicit unavailable-family or other N/A evidence: N/A
• Known PR-caused P1 regressions: 0
• Unresolved review threads: 0

[leehack]

Landed on main via batch integration PR #1895 (rebase merge, main ba4e0e1b; tree identical to the audited integration head 5ffd1224).