[leehack]

Summary

When a thinking budget runs out, llamadart forced a bare </think>. llama.cpp (cd959bae9) forces \n</think> for the Qwen3-Coder XML family, and without the line break Qwen3.5 keeps thinking or leaks the raw call as content. A budget on ChatFormat.qwen3CoderXml (Qwen3.5, Nemotron 3 Nano, StepFun 3.5 Flash) now appends \n to the forced message. Fixes #1914; lane 1 of #1918.

• New trait ChatFormatTraits.budgetThoughtClosePrefix (default '', '\n' for qwen3CoderXml), read through a top-level budgetThoughtClosePrefix(formatIndex); public thinkingTagsFor is unchanged.
• The planner appends it to forcedMessage only when the caller set no endTag and the plan renders the prompt (not native-chat, not schema backend): LiteRT-LM throws on any forced message.
• Budget token counting is unchanged (maintainer decision).
• Docs: changelog (both files), generation-and-streaming, runtime-parameters, template-engine-internals, ThinkingBudget.forcedMessage Dartdoc. Also fixed the stale claim that WebGPU and LiteRT-LM reject a budget.

Production-readiness scope

• User-facing scope: a thinking budget on Qwen3-Coder XML templates closes the thought as llama.cpp does.
• Supported platforms/paths: rendered-prompt generation (native llama.cpp, WebGPU).
• Unsupported or intentionally unavailable paths: native LiteRT-LM chat is untouched and still rejects forcedMessage.
• Out of scope / follow-ups: upstream's other end sequences (<tool_call>, DSML) as a single native end sequence; budget tags for gpt-oss, Apriel and the Gemma 4 start tag.

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.
• Public API docs, README/website docs, examples, support matrices, and changelog entries are updated where relevant.
• 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 is tracked above.

PR type guidance

• Bugfix PR: root cause above; regression tests below; affected paths: rendered-prompt backends.

Test Plan

• dart run tool/prepare_workspace.dart
• dart format --output=none --set-exit-if-changed .
• dart analyze and check_platform_boundaries.dart
• dart test -p vm on the changed tests and test/unit/tooling/ (all pass except 6 windows_registry_qualification_test cases that need pwsh, absent locally)
• dart test -p chrome test/unit/core/engine/chat_completion_request_planner_test.dart
• coverage-lib: 93.79% lines (threshold 70%); the only failures in the full VM run were 6 windows_registry_qualification_test cases needing pwsh
• ./tool/testing/run_template_parity_suites.sh
• run_local_e2e.dart --scenario gguf-chat-features-smoke with Qwen3.5-2B-Q4_K_M
• ./tool/docs/build_site.sh: not run locally (port 8080 in use by another lane); left to CI

Matrix Evidence

Matrix row	Scope covered	Platform / model / backend	Result	Evidence / notes
static-format-analyze	format, analyzer, import boundaries	macOS, Flutter 3.47.1	PASS	clean
root-vm	planner and parity tests, tooling suite	VM	PASS	6 pwsh-less Windows registry tests excluded; full root VM run left to CI
root-chrome	planner test	Chrome	PASS
coverage-lib	lib coverage	VM	93.79%
template-parity	vendored template parity	VM	PASS	run_template_parity_suites.sh
structured-output-adversarial	upstream emission parity of the forced close	VM; Qwen3.5/Nemotron 3 Nano/StepFun vs QwQ/MiniMax-M2/DeepSeek-V3.1 controls	PASS	thinking_budget_close_parity_test.dart against test/fixtures/thinking_budget_close_upstream.json (llama.cpp cd959bae9); fails first on the three Qwen3-Coder XML templates
gguf-chat-features-smoke	real GGUF thinking budget with tool call	Metal, Qwen3.5-2B-Q4_K_M	PASS	toolCallWithThinkingBudget finish tool_calls, 1 call, no content leak
docs-site	site build		N/A	port conflict locally; CI

Review Notes

• Independent review status: pending auditor
• CI status / head SHA: pending; head 388678c575bfce28139b1b3c3deda78e0e6c5c2a

High-risk regression review

• Classification: high-risk (backendRuntime, structuredOutput)
• Implementation task: lane 1914 worker
• Independent blocking QA task: pending
• Exact head / current base: 388678c575bfce28139b1b3c3deda78e0e6c5c2a / edaf8b5629
• Production-branch deletion, bypass, or miswire proof: mutants all fail a test: M1 drop trait (parity + planner), M2 render-always (planner native-chat case), M3 ignore endTag (planner), M4 replace instead of append (planner), M5 prefix first (planner)
• Affected-family real-model/artifact evidence: Qwen3.5-2B-Q4_K_M smoke above; upstream llama-server emissions recorded in the fixture; Nemotron 3 Nano and StepFun weights unavailable, covered by upstream thinking-tag fixtures for their vendored templates
• Explicit unavailable-family or other N/A evidence: Nemotron 3 Nano, StepFun 3.5 Flash weights
• Known PR-caused P1 regressions: 0
• Unresolved review threads: 0

[leehack]

Landed via batch integration PR #1926 (rebase merge). Integration audit PASS at 6439022; this PR's audited head is recorded in #1926's constituent table.

[github-actions]

Chat app preview removed for leehack/llamadart-chat-pr-1924.