[leehack]

Summary

• Parse a MiniMax M3 reply to a forced-open thought as llama.cpp (cd959bae9) does: the thought is the text before the first </mm:think>, kept as generated, and every later <mm:think> or </mm:think> is content. Replaces the generic multi-block extraction in MinimaxM3Handler with extractForcedOpenThoughtVerbatim and sets keepsForcedThoughtVerbatim on the format, so the stream ends the thought at its first end tag.
• Audit follow-up (P2): with tools on, a reply such as </mm:think><mm:think> + call lost its <mm:think> content in a 2-chunk stream, because the final tool-envelope check read the content's <mm:think> as an unended thought. For verbatim forced-open formats that check now decides from the text after the first end tag.
• Root cause: the handler extracted every thinking block, so the whole parse took later tags as thoughts, while the stream (which kept the leading-thought rule) kept a trailing </mm:think as content. The two disagreed by chunking and both differed from llama.cpp.

Closes #1905.

Production-readiness scope

• User-facing scope: MiniMax M3 forced-open replies (thinking_mode: enabled) with repeated or stray <mm:think>/</mm:think> parse and stream identically, matching llama.cpp, with any tool choice.
• Supported platforms/paths: all (pure Dart parser and stream; VM and Web).
• Unsupported or intentionally unavailable paths: None.
• Out of scope / follow-ups: None. Content whitespace stays trimmed (existing engine behavior; llama.cpp keeps it). Reasoning whitespace also differs: the engine trims leading reasoning whitespace, llama.cpp keeps it; unchanged here.

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 options)
• Public API docs, README/website docs, examples, support matrices, and changelog entries are updated where relevant. (changelog x2, template-engine-internals MiniMax paragraph; 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. (N/A: no capability change)
• Security/privacy review completed: no secrets or paths involved.
• Follow-up work tracked: None.

PR type guidance

• Bugfix PR: root cause above; regression tests recorded from the llama.cpp oracle; affected platforms VM and Chrome.

Test Plan

• dart run tool/prepare_workspace.dart
• dart format --output=none --set-exit-if-changed .
• dart analyze
• dart test -p vm -j 1 --exclude-tags local-only (locally: test/unit/core and the touched integration test; full suite left to CI at the head)
• dart test -p chrome on test/unit/core/engine and test/unit/core/template
• coverage-lib matrix row (not run locally: the change is ~10 lines, all exercised by the new tests; main enforces the threshold)
• Other: dart run tool/testing/check_platform_boundaries.dart; mutation check below

Matrix Evidence

Matrix row	Scope covered	Platform / model / backend	Result	Evidence / notes
static-format-analyze	format, analyze, platform boundaries	local	pass	exit 0
root-vm	test/unit/core, new integration test	Dart VM	pass	CI at head covers the rest
root-chrome	test/unit/core/engine, test/unit/core/template, touched stream test	Chrome	pass	new fixture test is @TestOn('vm'); inline oracle shapes run on Chrome in minimax_thinking_stream_test.dart
coverage-lib	N/A	N/A	N/A	not run locally; see above
structured-output-adversarial	forced-open thought, auto/none tool choice, whole/char/every-split streaming, partial parse, call after thought	primary upstream emissions (llama.cpp cd959bae9 oracle) on the vendored MiniMax-M3.jinja; no MiniMax M3 weights	pass	fixture test/fixtures/minimax_m3_forced_open_upstream.json (now with end-then-start-then-call rows and partial x<mm:th rows)
docs-site	N/A	N/A	N/A	one sentence in an existing paragraph; no links added

Review Notes

• Independent review status: P2 and P3 audit findings fixed at the new head; re-audit pending (high-risk)
• CI status / head SHA: pending at 99b9c89

High-risk regression review

• Classification: high-risk (structuredOutput)
• Implementation task: lane-1905 worker session
• Independent blocking QA task: pending; PR stays draft until an uninvolved auditor reviews the exact head
• Exact head / current base: 99b9c89 / 33e00f5
• Production-branch deletion, bypass, or miswire proof: mutants (all fail the new tests): dropping the keepsForcedThoughtVerbatim trait (3 failures), reverting the handler to extractThinking (30 failures), ignoring isPartial in the handler (2 failures), passing the suffix-hidden thinkingInput to the handler (fails on the new partial rows), ignoring thinkingTagsAreContent in the final envelope check (fails end-then-start-then-call split rows). Surviving equivalent variant: lastIndexOf instead of indexOf for the end tag passes on the recorded rows
• Affected-family real-model/artifact evidence: MiniMax M3 weights unavailable; primary upstream emissions from the llama.cpp cd959bae9 probe on the vendored template plus durable fixtures
• Explicit unavailable-family or other N/A evidence: MiniMax M3 real model N/A (not available locally)
• Known PR-caused P1 regressions: 0
• Unresolved review threads: 0

[leehack]

Landed via batch integration PR #1912 (rebase merge, main 2cfb936). The integration audit passed at a9c7772, and this PR's audited head is recorded in #1912's constituent table.

[github-actions]

Chat app preview removed for leehack/llamadart-chat-pr-1910.