[leehack]

Summary

• Add public ChatTemplateEngine.resolveThinkingBudget(int formatIndex, ThinkingBudget budget, {bool rendersPrompt = true}), which fills a budget's omitted start/end tags and the Qwen3-Coder XML line-break close prefix, exactly as LlamaEngine does.
• ChatCompletionRequestPlanner now calls the same static, so the logic lives in one place; planner behavior is unchanged.
• Lets genkit_llamadart (which renders the prompt itself) match core. Genkit adoption follows the 0.19.7 release in a separate PR.

Refs #1929; plan #1939.

Production-readiness scope

• User-facing scope: Integrations that call ChatTemplateEngine.thinkingTagsFor and run runtime.generate themselves can get the same resolved ThinkingBudget as core, including the \n close prefix on Qwen3-Coder XML templates.
• Supported platforms/paths: Pure Dart template logic; all platforms.
• Unsupported or intentionally unavailable paths: rendersPrompt: false (backends that take chat messages and close the thought themselves) adds no prefix, as before.
• Out of scope / follow-ups: genkit_llamadart adoption (separate PR after the core 0.19.7 release, #1929).

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 platform dependence.)
• Public API docs, README/website docs, examples, support matrices, and changelog entries are updated where relevant. (/// Dartdoc, both changelogs; no README/skill/example mentions thinkingTagsFor.)
• Regression coverage covers the original issue plus key negative/version-skew paths where relevant.
• Public capability changes update the validation coverage registry: N/A, a helper on an existing public class, not a new capability; validation_coverage_test.dart passes.
• Security/privacy review completed: no secrets or paths involved.
• Follow-up work tracked: genkit adoption above.

PR type guidance

• Feature PR: public static with Dartdoc and tests for the happy path (Qwen3-Coder XML, forced-message append), no-prefix format, caller end tag and rendersPrompt: false.

Test Plan

• dart run tool/prepare_workspace.dart
• dart format --output=none --set-exit-if-changed .
• dart analyze (no issues)
• dart test -p vm -j 1 --exclude-tags local-only: macOS, 14086 passed; only the 6 windows_registry_qualification_test.dart cases failed locally (no pwsh)
• dart test -p chrome --exclude-tags local-only: left to CI; the change is platform-neutral Dart
• coverage-lib: 93.79% (threshold 70)
• Targeted: test/unit/core/template, test/unit/core/engine, test_structure, validation_coverage_test.dart, check_platform_boundaries all pass.

New test test/unit/core/template/thinking_budget_resolution_test.dart failed to load before the change. Mutations, each caught:

Mutation	Failing test
drop rendersPrompt &&	no prefix when backend does not render
drop budget.endTag == null	caller end tag left as given
condition forced false	Qwen3-Coder XML newline
replace append with closePrefix	appends to forced message
ignore caller startTag	caller end tag/start tag as given
ignore caller endTag	caller end tag/start tag as given

Matrix Evidence

Matrix row	Scope covered	Platform / model / backend	Result	Evidence / notes
coverage-lib	>=70% line coverage for lib/	macOS VM	PASS	93.79% line coverage, local
high-risk-exact-head-independent-qa	independent blocking-only review of exact head	b103a1c	PASS	evidence llamadart-evidence/pr-1943/ (evaluator exit 2); 11/11 change mutants killed
structured-output-adversarial	thinking/budget parity	N/A	N/A	Planner output unchanged; the existing thinking_budget_close_parity_test.dart (#1924 upstream fixtures) passes unchanged
release-metadata-verification, litert-lm-*, chat-app-litert-json-schema		N/A	N/A	No release metadata, LiteRT-LM or device change

Review Notes

• Independent review status: independent audit PASS at b103a1cb. Planner output matched main's inline logic in all 1176 cases tried.
• CI status / head SHA: green (32 pass, 9 skipping), b103a1c

High-risk regression review

• Classification: high-risk (backendRuntime, structuredOutput; chat_completion_request_planner.dart)
• Implementation task: lane-1929-core-api
• Independent blocking QA task: a fresh auditor session (not the author) passed it, with zero PR-caused P1s and zero unresolved threads. The Dartdoc P3s are on 0.19.7 audit follow-ups #1938.
• Exact head / current base: b103a1c / 81ba323
• Production-branch deletion, bypass, or miswire proof: planner call site now delegates to the static; mutations above and the unchanged Close a Qwen3-Coder XML thinking budget with a line break (#1914) #1924 parity test cover it
• Affected-family real-model/artifact evidence: Qwen3-Coder XML upstream fixtures (thinking_budget_close_upstream.json) via the unchanged parity test; no model run
• Explicit unavailable-family or other N/A evidence: none
• Known PR-caused P1 regressions: 0
• Unresolved review threads: 0