[leehack]

Fixes #1958

Summary

• On native llama.cpp, a ThinkingBudget with ToolChoice.required on Seed-OSS no longer leaves the call unconstrained after the budget closes a thought the model opened.
• Cause: the required root seed-thought? text-before-calls (...) accepts <seed:think>, then the budget sampler paused it (its root is not think-body), so it resumed inside seed-thought-text and the forced close and call were unconstrained.
• Fix: ToolCallGrammarUtils.keepsAppliedThroughThought (the Keep the response-format grammar applied through a budgeted forced-open thought #1956 think-body predicate plus a seed-thought root) replaces takesForcedOpenThought in the pause decision in llama_cpp_service.dart. Chosen over teaching the paused grammar to accept the forced close: it is one predicate, and it matches upstream, which only pauses lazy grammars (grammar_should_apply in common/sampling.cpp).

Production-readiness scope

• User-facing scope: Seed-OSS required tool calls stay grammar-constrained after a spent thinking budget.
• Supported platforms/paths: native llama.cpp. WebGPU still pauses every grammar (#1954).
• Unsupported or intentionally unavailable paths: N/A.
• Out of scope / follow-ups: None. The broader pause-every-eager-grammar divergence from upstream is not widened here.

Completeness checklist

• Declared scope is fully implemented.
• Unsupported platform/option combinations: unchanged.
• Dartdoc, website guide and both changelogs updated.
• Regression coverage covers the original issue plus the unchanged paths.
• Validation coverage registry: N/A, no new public capability.
• Security/privacy review: no secrets, tokens or paths touched.
• Follow-ups: None.

PR type guidance

• Bugfix PR: root cause above; regression tests below; affected path is native llama.cpp only.

Test Plan

• dart run tool/prepare_workspace.dart
• 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 (macOS, isolated TMPDIR; 14237 passed) with coverage-lib: 93.87% lines
• dart test -p chrome --exclude-tags local-only: left to CI (no Web code changed)
• tool/testing/run_template_parity_suites.sh, dart test -p vm test/unit/tooling/

Matrix Evidence

Matrix row	Scope covered	Platform / model / backend	Result	Evidence / notes
coverage-lib	lib/ line coverage	macOS VM	93.87%	full VM suite passed
template-parity	vendored template detection/render/parse, grammar acceptance	macOS VM	pass	run_template_parity_suites.sh
structured-output-adversarial	required grammar under a budgeted thought	stories15M (mechanism), Seed-OSS template fixture (grammar acceptance)	pass	see below

Failing-first and mutations (stories15M integration test test/integration/thinking_budget_grammar_test.dart, plus unit tests):

Mutation	Result
M0: service uses takesForcedOpenThought (pre-fix)	integration test fails
M1: pause every grammar	both existing and new integration tests fail
M2: predicate matches any root	predicate unit test and the caller-grammar integration test fail
M3: predicate requires seed-thought without ?	predicate unit test and integration test fail

Cross-format probe (460 template x mode combinations from every .jinja fixture, before vs after pause decision): exactly 2 change, both Seed-OSS required (think on and off). Ling3, Kimi K3, MiniMax M3, DeepSeek R1/V32/V4, Laguna, Qwen3-Coder XML and every other format are unchanged.

Upstream comparison: upstream llama-server never pauses an eager grammar under a reasoning budget (common/sampling.cpp grammar_should_apply: lazy grammars only), so the Seed-OSS grammar stays applied there. A live run was not possible: the Seed-OSS weights are too large to run locally, and llama-server b12110 refuses to initialise the Seed-OSS required grammar on stories15M (common_sampler_init: error initializing grammar sampler). The evidence is the durable template fixture (ByteDance-Seed-OSS.jinja), the grammar-acceptance unit test (forced budget close accepted, free text after it rejected) and the stories15M test with a seed-thought-rooted grammar. The test uses <s> as the start tag because the model's greedy tokenization of a grammar-forced <seed:think> never equals the tag's own tokens.

Review Notes

• Independent review status: PASS (fresh Opus audit at ee6384e, delta re-audit at 69a7e21)
• CI status / head SHA: green, run 43266650445 (32 pass, 9 skipped) / 69a7e21

High-risk regression review

• Classification: high-risk (surfaces: backendRuntime, structuredOutput)
• Implementation task: this PR, lane 3 of plan 0.19.8 release plan: thinking-budget grammar fixes (native, WebGPU, Seed-OSS) and #1950 sweep #1959
• Independent blocking QA task: fresh auditor-deep delta re-audit, PASS, 0 P1/P2; P3s logged in 0.19.8 audit follow-ups #1950
• Exact head / current base: 69a7e217af221d7950c9327e028bf2aace8baad6 / 083166ef; readiness evaluator exit 2 (unverifiedPrerequisites)
• Production-branch deletion, bypass, or miswire proof: 11/13 mutants killed; M7/M8 equivalent
• Affected-family real-model/artifact evidence: Seed-OSS weights unavailable; durable template fixture, grammar-acceptance test and upstream grammar_should_apply comparison used
• Explicit unavailable-family or other N/A evidence: Seed-OSS model not run locally; unrelated stories15M is mechanism-only evidence
• Known PR-caused P1 regressions: 0
• Unresolved review threads: 0

[github-actions]

Chat app preview removed for leehack/llamadart-chat-pr-1962.