[leehack]

Summary

• genkit_llamadart builds its ThinkingBudget with ChatTemplateEngine.resolveThinkingBudget(template.format, budget) instead of hand-built tags, so Qwen3.5 / Qwen3-Coder XML closes a spent budget with \n</think> rather than a bare </think>.
• Raises the package floor to llamadart: ^0.19.7 and the matching lower-bound pins (ci.yml, publish_genkit_pubdev.yml, genkit-package.md, ci_workflow_selection_test.dart); adds an Unreleased package changelog entry. No genkit version bump (separate release-prep PR).

Fixes #1929

Production-readiness scope

• User-facing scope: A spent thinking budget on Qwen3-Coder XML templates (Qwen3.5) closes the thought with the template's \n</think> instead of a bare </think>. This does not make Qwen3.5 constrained-JSON output with a budget work end to end: a separate pre-existing bug, Qwen3.5: constrained JSON with a thinking budget ignores the grammar after the forced </think> #1953, still breaks that.
• Supported platforms/paths: Genkit structured-output path (_runStructuredModel), where Genkit renders the prompt itself (rendersPrompt: true).
• Unsupported or intentionally unavailable paths: None added. A caller-set endTag gets no prefix.
• Out of scope / follow-ups: Genkit release (needs separate maintainer approval; next version: minor, 3.9.0, since the llamadart floor moves, as in 3.6.0 to 3.8.0).

Completeness checklist

• Declared scope is fully implemented, or explicitly reduced above.
• Unsupported platform/option combinations fail loudly or are documented as unavailable (N/A: no new options).
• Changelog updated (package CHANGELOG.md). The genkit docs usage.md, migration.md and tool-input-schemas.md are updated for the \n close prefix and the ^0.19.7 floor. No public API change, so no skills change.
• Regression coverage covers the original issue plus the caller-endTag path.
• Validation coverage registry: N/A, no public capability change.
• Security/privacy review: no secrets or paths touched.
• Follow-ups: None.

PR type guidance

• Bugfix PR: root cause is the hand-built budget in model_action.dart duplicating, and missing, core's \n close prefix for Qwen3-Coder XML. Regression test: inference_controls_test.dart, "Qwen3-Coder XML thinking budget". Platforms: all that run the structured path.

Test Plan

• dart run tool/prepare_workspace.dart (N/A: only the independently resolved adapter changed)
• dart format --output=none --set-exit-if-changed . (root: 0 changed)
• dart analyze (adapter: dart analyze --fatal-infos, no issues)
• dart test -p vm -j 1 --exclude-tags local-only N/A: no root code changed; root tests skip packages/
• dart test -p chrome --exclude-tags local-only N/A: same
• coverage-lib N/A: no lib/ change
• Other: adapter dart format, python3 .github/check_docs.py, dart test -x real-model (474 passed)

Matrix Evidence

Matrix row	Scope covered	Platform / model / backend	Result	Evidence / notes
genkit-contracts	adapter analyze, docs, tests against published llamadart 0.19.7	macOS, fake runtime	PASS	local run; hosted CI adds lower-bound and local-core legs
structured-output-adversarial	thinking-budget close prefix on Qwen3-Coder XML	fake runtime capturing params.thinkingBudget	PASS	forced message + \n, \n alone, none with caller endTag; no real model
genkit-native-cpu, genkit-release	not run	N/A	N/A	no native or release-routing change

Mutation check: reverting to the old construction fails the two prefix tests; rendersPrompt: false fails the same two.

Review Notes

• Independent review status: deep audit PASS at f46824cc; delta re-audit PASS at d8b49b28
• CI status / head SHA: all checks green at d8b49b2

High-risk regression review

• Classification: high-risk (surfaces: artifactConsumer, backendRuntime, regressionPolicy, releaseAutomation, structuredOutput; the workflow pin touches the genkit publisher)
• Implementation task: this PR (lane-1929-genkit)
• Independent blocking QA task: fresh auditor sessions (not the author). Genkit and core agreed in all 576 parity cases; zero PR-caused P1s and zero unresolved threads. P3s go to 0.19.8 audit follow-ups #1950.
• Exact head / current base: d8b49b2 / 4b96fde
• Production-branch deletion, bypass, or miswire proof: mutation check above (old construction and rendersPrompt: false both fail the test)
• Affected-family real-model/artifact evidence: N/A: Qwen3.5 weights not run; template-level fake-runtime evidence only (core Add ChatTemplateEngine.resolveThinkingBudget (Refs #1929) #1943 carries the parity evidence)
• Explicit unavailable-family or other N/A evidence: the Qwen3 and Qwen3.5 real-model tests skip in hosted CI. The audit's local real Qwen3.5-0.8B probe showed genkit and core giving byte-identical output (llamadart-evidence/pr-1952/prior-f46824cc/qwen35-raw-probe.log). End-to-end constrained JSON with a budget still fails until Qwen3.5: constrained JSON with a thinking budget ignores the grammar after the forced </think> #1953 is fixed.
• Known PR-caused P1 regressions: 0
• Unresolved review threads: 0

[github-actions]

Chat app preview removed for leehack/llamadart-chat-pr-1952.