[leehack]

Summary

Closes #1523. Refs #1531 (lane of #922).

• Cohere2 MoE: with enableThinking: false and a reasoning: true kwarg, the prompt now ends <|START_THINKING|><|END_THINKING|>, the template's own closed form, instead of adding \n. The new upstream fixture case matches llama.cpp 620d2225 with {reasoning: false, skip_thinking: true} byte for byte.
• Seed-OSS: adds a second lazy trigger, <seed:tool_call><function=. The grammar (xml-space after the scope) and the parse already accept that form. GBNF-validated. buildGrammar now builds the non-parallel grammar that render uses by default.
• Tests for surviving mutants: X2, an upstream case with two Qwen3.5 <tool_call>s in an unended thought. X7, a content opening inside a thought the model opens streams as it parses. Z2/Z5, a Hermes, DeepSeek V3 or Qwen3-Coder template that always opens the thought keeps parsing the reply's thought after the handler closes it. M11/M17, an empty first Qwen3-Coder scope starts no later calls (upstream returns 0). Hy3 reasoning_effort none/medium throws LlamaInferenceException. cities() checks exact values again, untrimmed for Qwen3-Coder and with line breaks for Seed-OSS.
• Docs: the guide and llamadart-chat-streaming skill now cover the Kimi K3 thinking switch and the defaults llamadart sets (Hy3 high, MiniMax M3 enabled) where llama-server renders no_think/adaptive. They also note that under enableThinking: false a kwarg does not fully win: llamadart still closes a thought the kwarg opens, and Cohere2 MoE still drops earlier thoughts unless skip_thinking is set too. The changelog notes the Nemotron 3 Nano parallel-grammar fix from Parse each Qwen3-Coder XML parallel call in its own <tool_call> block (#1530) #1539 and the Cohere change.

Item status

Item	Status
#1523 1 Cohere thought ends at <|START_TEXT|>/<|START_ACTION|>	Done in #1524
#1523 2 MA11	Killed in #1526
#1523 3 WebGPU decoding of the control tokens and stops	Checked in the source only, no Web Cohere model run. Bridge 883376d (assets v0.1.76) decodes every token with special=true (src/core/generation.inc 813/875), so the preserved markers reach Dart as text, as native renders its preserved tokens. Both backends drop stop sequences that are preserved tokens (_effectiveStopSequences, webgpu_backend.dart 1972), so neither stops on <|END_RESPONSE|>/<|END_ACTION|>; generation ends at the EOG token.
#1523 4 guide precedence	#1526, qualified here
#1523 5 Hy3 values	Documented in #1526; test here
#1523 6 Cohere \n	Fixed here
#1523 7 MB6, MB10	Killed in #1526
#1523 7 MB7	Moved to #750, where #1526 recorded it. Upstream lets kwargs override messages/tools but not BOS/EOS, so a test of either order would pin a divergence.
#1523 comment: guide ~60, MiniMax M3 enabled, Hy3 throw test, Kimi K3	Done: qualified; documented as intended (same mapping as Hy3 high); test; listed
#1531 X2, X7, Z2/Z5, M11/M17, cities(), Seed buildGrammar, Nemotron changelog, Seed trigger	Done

Production-readiness scope

• User-facing scope: Cohere2 MoE prompt bytes for enableThinking: false plus an opening reasoning kwarg. Seed-OSS lazy grammar also starts on <seed:tool_call><function=. Docs.
• Supported platforms/paths: all; Dart template render shared by native and Web.
• Unsupported or intentionally unavailable paths: N/A, no new option. Hy3's existing LlamaInferenceException for other reasoning_effort values is now tested.
• Out of scope / follow-ups (deferred, other lanes or decisions): #1532 streaming holdback. Ling parsed as Hermes (<think> in call arguments). Upstream dropping text after tool calls. Parse cap with parallel calls off. Dropping a partial </<tool_ at the end with thinking off. Also still open on Reasoning parity follow-ups from #1524/#1525 audit: mutants X2/X7, Ling <think> in args, partial-tag drop #1531 and not in this lane: Seed whitespace leniency, unknown tool in a later Seed scope, several <function=> in one Qwen3-Coder scope, truncated later scope, and Qwen3-Coder/StepFun thinking-off handling. A Seed empty first scope is not pinned: llama.cpp reads <seed:tool_call></seed:tool_call> as text and returns the later call, which is part of the Seed whitespace leniency.

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. (N/A: no new option.)
• Public API docs, README/website docs, examples, support matrices, and changelog entries are updated where relevant. (Guide, skill, both changelogs. README/examples/support matrix N/A: no capability 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: no public 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. (Reasoning parity follow-ups from #1524/#1525 audit: mutants X2/X7, Ling <think> in args, partial-tag drop #1531 stays open for the deferred items.)

PR type guidance

• Bugfix PR: Root cause, Cohere: CommandR7BHandler.render, which Cohere2 MoE inherits, closed a thought the template left open with <|END_THINKING|>\n, while the template's own off form has no newline. Root cause, Seed: Parse each Seed-OSS parallel call in its own <seed:tool_call> block (#1540) #1542 swapped the lazy trigger for llama.cpp's <seed:tool_call>\n, which dropped the compact form the grammar and parse still accept. Failing-first: the new kwargs fixture case and both Seed handler tests fail on main. Platforms: all.
• Feature PR / Docs-only PR: N/A.

Test Plan

• dart run tool/prepare_workspace.dart
• dart format --output=none --set-exit-if-changed .
• dart analyze lib test tool
• dart test -p vm -j 1 --exclude-tags local-only (isolated TMPDIR): +9471 ~123, all passed
• dart test -p chrome test/unit/core/template test/unit/core/engine/chat_completion_stream_parser_test.dart: +1118, all passed
• Other targeted/local-only validation: dart run tool/testing/check_platform_boundaries.dart; ./tool/testing/run_template_parity_suites.sh (b10549 templates, GBNF validator): all five suites passed; oracle, render diff and mutation runs below.

Evidence (oracle built from llama.cpp 620d2225, harnesses from llamadart-evidence/pr-1542/audit/ and issue-1540/; logs in llamadart-evidence/pr-1550/):

Check	Result
Oracle, full corpus (94 templates, 183,672 parse/stream records)	OK on main 66,896, OK on head 66,896: 0 OK->BAD, 0 BAD->OK. The one differing record is a Gemma 4 3-byte-chunk stream (ps on both). Three sequential reruns of that template give identical main and head output, so the diff comes from the timing-based partial parse under 10-way load.
Oracle, Seed-OSS parallel corpus (13,944 records)	0 changes
Render diff, 2,256 renders (94 templates x 24 configs)	Only the 8 Seed-OSS renders with a lazy grammar change, gaining the second trigger. No prompt or grammar changes.
Kwargs fixture vs oracle	All 13 prompts byte-exact, the new Cohere case included
Mutants	X2, X7, Z2, Z5, M11/M17, plus new N1 (Cohere \n back), N2/N4 (either Seed trigger dropped), N3 (Seed buildGrammar parallel): all killed, each by the new test. MB7 survives (moved, see above).

Matrix Evidence

Matrix row	Scope covered	Platform / model / backend	Result	Evidence / notes
static-format-analyze	format, analyzer, web/native import boundaries	macOS, Flutter 3.47.1	PASS	local
root-vm	native-safe unit/integration coverage on the Dart VM	macOS VM	PASS	local, isolated TMPDIR
root-chrome	browser-compatible unit/integration coverage	Chrome	PASS	template + stream suites locally; full suite in CI
structured-output-adversarial	render/history, compiled grammar acceptance, partial/final streaming, thinking prefixes, pinned upstream parity	Dart vs llama.cpp 620d2225; GBNF validator	PASS	oracle 0 OK->BAD; parity suites; Seed compact scope accepted by the grammar
high-risk-exact-head-independent-qa	independent blocking review of exact head	fresh auditor session, exact head 5229835	PASS	CLEAR: CLEAR at 5229835; evidence /opt/UnitySrc/personal/llama/llamadart-evidence/pr-1550/audit/evidence-1550-5229835.json
release-metadata-verification			N/A	not a release PR
litert-lm-json-schema			N/A	no LiteRT-LM change
docs-site	website build		CI	guide paragraph and changelog

Review Notes

• Independent review status: CLEAR at exact head 5229835 by a fresh auditor session (0 P1, 0 P2).
• CI status / head SHA: 5229835, CI pending.

High-risk regression review

• Classification: high-risk (structuredOutput)
• Implementation task: #1523 and the Reasoning parity follow-ups from #1524/#1525 audit: mutants X2/X7, Ling <think> in args, partial-tag drop #1531 test-gap items, 0.17 lane of #922
• Independent blocking QA task: fresh auditor session, CLEAR, 0 known PR-caused P1, 0 unresolved threads
• Exact head / current base: 5229835 / d926c2f (merge-tree clean)
• Production-branch deletion, bypass, or miswire proof: N1 to N4 delete or revert each production change, and each fails a new test (mutation table above).
• Affected-family real-model/artifact evidence: Cohere2 MoE, Seed-OSS, Qwen3.5, Qwen3-Coder, Hermes, DeepSeek V3 and Hy3 weights are not used. The evidence is primary upstream emissions: the llama.cpp 620d2225 oracle renders and parses, plus durable fixtures.
• Explicit unavailable-family or other N/A evidence: weights for these families are unavailable locally; the WebGPU control-token item was checked in the source only.
• Known PR-caused P1 regressions: 0
• Unresolved review threads: 0

[github-actions]

Chat app preview removed for leehack/llamadart-chat-pr-1550.