[leehack]

Summary

Fixes #1887. Refs #1874 (0.19.4 plan, lane 7).

Stacked on #1890 (base fix/1531-qwen3-coder-required-grammar at b75aa2e91, on #1885 and #1882). Review commits 933e48bee (fix) and 302757be6 (audit follow-ups); 22fe46da5 and bf8735ab8 merge the moved base (95d4986c9 -> 55c69df04 -> b75aa2e91). Both merges were clean except that the union-merged changelogs kept #1531's old required-call bullet beside its reworded one; the stale copy was removed each time, so each changelog gains only the #1887 bullet. Retarget once #1890 merges.

After a prompt that opens no thought (MiniMax M3 with thinking off, or thinking_mode: adaptive), llama.cpp cd959bae9 (common/parsers/minimax-m3.cpp) takes a thought only where <mm:think> starts the reply, with no whitespace before it, drops a </mm:think> there, and keeps every other thinking tag as content. llamadart took a <mm:think> thought from anywhere in the reply, and a </mm:think> after text ended a thought.

• minimaxM3 gets the takesThoughtOnlyAtReplyStart trait from Take a thought only at the start of a reply after a prompt that opens no thought, as llama.cpp does (#1865) #1882.
• New trait thoughtStartsReplyExactly (MiniMax M3 only): the reply-start thought skips no whitespace, and an end tag that starts the reply is dropped. withThoughtOnlyAtReplyStart(parse, exact:) carries it to extractThinking; the stream waits at the reply start for either tag.
• The stream's tool-envelope check (_isToolCallEnvelopeBuffer) applies the same rule, so a <mm:think> after leading whitespace or after a leading </mm:think>, followed by a call before any end tag, streams as content instead of being taken as a thought and suppressed (audit note 1).

Why the extra trait. The issue's proposed fix (the #1882 trait alone) fixes the mid-reply thought, but in a wider corpus it breaks 300 rows that matched: a reply starting with a bare </mm:think> would keep the tag as content, where llama.cpp drops it. The #1882 audit corpus had no such reply.

The 16 leading-thought rows from the issue. With thinking off, llama.cpp consumes a leading <mm:think>…</mm:think> and drops it. llamadart's parseRendered still returns it as reasoningContent (documented: with thinking off the engine drops reasoning chunks, while parseRendered still returns it), so the engine's output matches llama.cpp. The leading-whitespace variants ( \n<mm:think>…) were real differences and are fixed.

Production-readiness scope

• User-facing scope: MiniMax M3 replies to a prompt that opens no thought parse and stream as in llama-server: a thought after text or after leading whitespace, and a </mm:think> after text, stay content.
• Supported platforms/paths: All backends that use the Dart chat parser (native llama.cpp, WebGPU bridge). VM and Chrome suites pass.
• Unsupported or intentionally unavailable paths: Unchanged differences, listed in closed_thought_upstream.md: with thinking off, parseRendered returns a leading thought as reasoning that llama.cpp drops (the engine drops it too); a partial parse drops a trailing partial thinking tag after content, which llama.cpp keeps (final parse and stream unaffected). With thinking on and no thinking_mode kwarg, llamadart renders enabled (forced-open <mm:think>) where llama-server renders adaptive, by design since DeepSeek V3.1 and Qwen thinking parity gaps against llama.cpp (thinking flag, escapes, repeated <think>) #750 and documented in the guide.
• Out of scope / follow-ups: None.

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: parse behavior only.)
• 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. skills/llamadart-chat-streaming/SKILL.md already states the rule MiniMax M3 now follows; unchanged.)
• 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 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: Root cause: minimaxM3 lacked the reply-start rule, so extractThinking took a <mm:think> thought anywhere and ended a thought at a </mm:think> after text; the generic rule also skips leading whitespace and keeps a leading bare end tag, which MiniMax M3's upstream parser does not; and the stream's tool-envelope check stripped any leading thought. Regression tests:
  • Two MiniMax M3 entries in closed_thought_upstream.json (thinking off with default kwargs; thinking on with thinking_mode: adaptive), 21 + 26 cases plus 5 dropped_thought_cases, recorded from the cd959bae9 probe (common_chat_session, whole = byte-by-byte = finish-only for every case; partial cases from common_chat_peg_parse). Each case checks parseRendered with and without tools and the stream at whole, per-char, 3-char and every 2-split chunking; dropped-thought cases check the engine stream with thinking off. On base 95d4986c9 47 of the first set failed. The audit follow-up cases unfinished-thought-after-space-then-call and unfinished-thought-after-end-tag-then-call fail (both entries) without the envelope fix; end-tag-only pins a reply of just </mm:think>.
  • thinking_utils_test.dart: withThoughtOnlyAtReplyStart(exact: true).
  • chat_format_registry_test.dart: the reviewed-dispatch text for the final envelope check now includes exactReplyStart.

Test Plan

All at 302757be6 unless noted.

• dart run tool/prepare_workspace.dart (no tracked changes)
• dart format --output=none --set-exit-if-changed .; dart analyze; dart run tool/testing/check_platform_boundaries.dart
• dart test -p vm -j 1 --exclude-tags local-only (isolated TMPDIR): 12,728 passed, 129 skipped, 6 failed, all windows_registry_qualification_test.dart (needs pwsh, absent on this host)
• dart test -p chrome --exclude-tags local-only test/unit/core test/integration/core: 3,819 passed, 1 skipped (full Chrome run at 3cdc8948f: 4,667 passed, 1 skipped)
• coverage-lib: Test Linux VM (with Coverage) and Test Linux & Web (with Coverage) pass in run 42802975044
• ./tool/testing/run_template_parity_suites.sh: detection 74, diagnostic 103, unit template 1,876, upstream parser 10, GBNF acceptance 81, all passed

Matrix Evidence

Matrix row	Scope covered	Platform / model / backend	Result	Evidence / notes
root-vm	native-safe unit/integration	macOS arm64 VM	PASS	12,728 passed, 129 skipped, 6 pwsh-only failures
root-chrome	browser-compatible tests	Chrome	PASS	core suites at head; CI
template-parity	detection, render and parse parity; GBNF acceptance	macOS arm64	PASS	see Test Plan
structured-output-adversarial	partial/final parse and every-chunking stream vs llama.cpp; thinking on/off; tool choice none/auto/required/parallel; with and without tools	llama.cpp cd959bae9 common_chat_session + common_chat_peg_parse oracle	PASS	corpora below; 0 regressions
coverage-lib	lib/ line coverage	Linux CI	PASS	run 42802975044
high-risk-exact-head-independent-qa	independent blocking review of exact head	—	PASS at 22fe46da5; re-audit needed at 302757be6	audit notes 1–3 addressed here
Other rows	—	—	N/A	no release, LiteRT-LM or runtime routing change

Upstream comparison (llama.cpp cd959bae9), base b75aa2e91 vs head 302757be6

MiniMax M3 corpus: the upstream template x thinking on/off x kwargs (none, thinking, thinking_mode enabled/disabled, thinking_mode: adaptive) x 5 tool modes x 235 replies (#1858/#1865/#1525 shapes, the #1882 audit shapes, bare </mm:think>, namespaced calls, and unfinished thoughts after whitespace or an end tag before a call). "Engine" is the stream with the request's own enableThinking.

Rows (finished unless noted)	base	head
parse = llama.cpp, all 9,400	4,405	6,117
partial parse = llama.cpp, all 9,400	4,305	5,967
engine stream = llama.cpp, all 9,400	4,739	6,859
every chunking = parse, all 9,400	7,385	8,695
prompt opens no thought (thinking off, or adaptive): engine = llama.cpp, of 5,800 rows llama.cpp parses	3,680	5,800
same rows: every chunking = parse	4,490	5,800
regressions (row matched on base, not on head)	—	0

The remaining mismatches are the forced-open render with thinking on and no thinking_mode (see above), the documented thinking-off parseRendered reasoning, and partial trailing tag prefixes.

All-template sweeps (76 upstream models/templates + vendored test/fixtures/templates; thinking on/off x kwargs x 5 tool modes):

Corpus	parse = llama.cpp	partial = llama.cpp	every chunking = parse	regressions
#1865 bare/mid-reply, upstream	60,852 -> 61,242 / 104,160	61,569 -> 61,939	96,922 -> 97,302	0
#1525 thinking, upstream	67,921 -> 68,241 / 103,212	—	99,892 -> 100,192	0
native tool calls, upstream	3,572 -> 3,572 / 7,040	3,218 -> 3,218	6,912 -> 6,912	0
#1865 bare/mid-reply, vendored	18,116 -> 18,116 / 24,000	18,480 -> 18,480	23,592 -> 23,592	0
#1525 thinking, vendored	20,256 -> 20,256 / 24,120	—	23,696 -> 23,696	0

Only MiniMax M3 rows change, apart from 7 Gemma 4 rows whose byte, 2-char or 3-char stream already differs from their parse on base and leaks a different length of call JSON as content from run to run (the stream's time-based partial parse interval); parse results are identical and both runs mismatch, so they are unrelated to this change.

Mutation check

16 mutants over core/unit, core/engine and core/template tests. 12 killed: trait removed, exact trait removed, parse keeps the leading end tag, parse skips leading space, engine passes exact: false, stream exact rule off, stream does not wait on an end-tag prefix, stream keeps the end tag, stream skips leading space, zone rule ignores exact, B5 (rest.length <= endTag.length, killed by end-tag-only), and envelope check without the exact rule (killed by the two new tool-call cases). 4 give identical results on all 18,800 MiniMax M3 corpus rows: the tool-mode decision path without the exact rule (unreachable for MiniMax M3's stream, so that argument was dropped), the exact strip not dropping a leading end tag or not waiting on a partial tag (the envelope check only suppresses content the parse leaves empty in those shapes). The leadingThoughtOnly guard on the exact flag also survives but is not equivalent: _isToolCallEnvelopeBuffer does reach the exact strip on forced-open prompts, and removing the guard makes 88 forced-open corpus rows chunking-dependent; no test pins it yet (independent audit at 302757be, follow-up on #1861).

Review Notes

• Independent review status: PASS at 22fe46da5 (audit notes 1–3 addressed in 302757be6); re-audit of 302757be6 pending
• CI status / head SHA: 302757be6; dispatched ci.yml run 42802975044: 26 passed, 6 skipped, 0 failed

High-risk regression review

• Classification: high-risk (backendRuntime, structuredOutput)
• Implementation task: 0.19.4 lane 7 (#1874)
• Independent blocking QA task: PASS at 22fe46da5; pending at 302757be6
• Exact head / current base: 302757be6 / b75aa2e91 (fix/1531-qwen3-coder-required-grammar)
• Production-branch deletion, bypass, or miswire proof: mutation check above (12/16 killed, 4 equivalent on the corpus with reasons)
• Affected-family real-model/artifact evidence: MiniMax M3 weights unavailable (several hundred GB MoE). Evidence is the upstream template and the pinned cd959bae9 parser via the oracle, with durable fixtures in closed_thought_upstream.json.
• Explicit unavailable-family or other N/A evidence: MiniMax M3 (above)
• 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).