Fix structured tool-result rendering for Qwen templates

[leehack]

Summary

Qwen3.5's template rejects a typed tool result containing a Dart Map with Unexpected content type, preventing the follow-up inference after a successful tool call. Normalize structured results to JSON text at the shared template-input boundary; preserve string results, source objects, and the public message JSON representation.

Fixes #528.

Production-readiness scope

• User-facing scope: JSON-compatible tool results (objects, arrays, scalars, and null) render through shared chat-template handlers. Multimodal template contexts receive a text part.
• Supported platforms/paths: Shared Dart rendering on VM and Chrome; exact Qwen3.5-0.8B GGUF CPU tool cycles verified on macOS.
• Unsupported or intentionally unavailable paths: No new backend/platform capability is claimed. Results must be JSON-compatible; ordinary image/audio messages retain existing behavior.
• Out of scope / follow-ups: Gemma4/FunctionGemma custom serializers, native artifacts, runtime pins, accelerator qualification, and suite PR Add portable validation suite and guarded Firebase/GCE runners #515 are unchanged. Parser follow-up Qwen3.5 XML parser loses schema strings and exposes undeclared tool calls #530 blocks acceptance of the stricter expanded schema/rollback coverage; it predates this input fix and remains separate.

Completeness checklist

• Declared scope fully implemented.
• Unsupported combinations remain outside the declared scope.
• Tool-calling guide and Unreleased changelogs updated; no public API signature or support-matrix changes.
• Regression coverage includes the exact affected template, structured payload shapes, escaping/Unicode, string/source preservation, and ordinary media controls.
• No credentials, signed URLs, or machine-local evidence paths added to tracked content.
• Newly exposed parser follow-up tracked in Qwen3.5 XML parser loses schema strings and exposes undeclared tool calls #530; no claim that its failing acceptance axes pass.

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 — 2,254 pass, three existing skips.
• dart test -p chrome --exclude-tags local-only — 960 pass.
• tool/testing/run_template_parity_suites.sh — template units, pinned upstream chat/parser suites, and compiled grammar acceptance/rejection pass.
• Coverage-enabled VM rerun and LCOV threshold — 81.06% (13,681/16,878), above 70%.
• tool/docs/validate_links.sh — production docs build and links pass with Node 20.

Matrix Evidence

Matrix row	Scope covered	Platform / model / backend	Result	Evidence / notes
structured-output-adversarial	Shared template boundary; exact Qwen template; Map/list/scalar/null, escaping/Unicode; unchanged media/string/source controls	VM / Chrome	PASS	Targeted VM26, Chrome25; pinned parser and15 compiled grammar tests pass. Grammar/parser implementation unchanged.
high-risk-exact-head-independent-qa	Production call sites, deletion mutation, exact committed diff	Head 2acefdc54fd29348ac89c84408a4ae8a639d22e2 / base 6c01a69af5bd13b338763f5f63cf16e482aa9fd6	PASS local audit	Independent53 targeted/custom-handler tests pass; removing conversion kills structured-result regressions.
Issue528 affected-family smoke	Public tool call → typed Map result → follow-up; ordinary recovery	macOS / Qwen3.5-0.8B-Q4_0 GGUF / explicit CPU	PASS	Both auto and required emit get_weather(city Montréal), follow-up returns exactly17 with stop/no tool deltas, recovery succeeds. Model SHA256 57d1997790d1744fba5b40a7317df71ea5e2acee28c47e78f0cce39c0703f8cf; temperature0, seed1, thinkingfalse, context1024, threads4, gpuLayers0. No timing/accelerator claim.

Review Notes

• Independent exact-commit audit accepted; zero blocking findings. Real-model evidence was inspected by the auditor, not independently rerun.
• All14 hosted checks passed at 2acefdc54fd29348ac89c84408a4ae8a639d22e2; live PR identity and zero review threads reverified. Code audit accepted. Expanded local strict typed-history tests independently reproduce pre-existing Qwen3.5 XML parser loses schema strings and exposes undeclared tool calls #530 failures (schema string123 becomes integer123; undeclared function produces a tool delta). Those unfinished test additions are not pushed; existing PR CI does not validate them. The readiness evaluator remains rejected and draft status is retained.
• Draft retained while PR-level readiness evidence is finalized; no merge/publication authorization.

High-risk regression review

• Classification: high-risk — structuredOutput.
• Implementation task: issue528 isolated runtime fix.
• Independent blocking QA task: codex-adversarial-tool-result-2acefdc5, separate audit agent; accepted exact committed diff.
• Exact head / current base: 2acefdc54fd29348ac89c84408a4ae8a639d22e2 / 6c01a69af5bd13b338763f5f63cf16e482aa9fd6.
• Production-branch deletion, bypass, or miswire proof: Independent isolated deletion mutant fails structured Map/list/scalar/null assertions; exact Qwen production template-selection test fails without conversion.
• Affected-family real-model/artifact evidence: Exact Qwen3.5-0.8B GGUF CPU auto/required typed-result cycles pass; exact template fixture and hashes are checked in.
• Explicit unavailable-family or other N/A evidence: No all-platform, GPU/NPU, or LiteRT qualification claim. Gemma custom serialization unchanged and independently regression-tested. Strict grammar/parser implementations unchanged; existing pinned compiled and parser suites pass.
• Known PR-caused P1 regressions: 0.
• Unresolved review threads: 0, verified with a live GraphQL query.
• Readiness evaluator: rejected / unchangedEvidencePath. All six structured-output axes require changed tests (schema lines175–204; evaluator lines971–1030 and1087–1092). Existing compiled grammar, reconstruction, rollback, tool-choice/thinking, and upstream parity tests pass but are unchanged; new input-serialization tests do not replace those axes. The release-metadata exception does not apply. Draft retained; meaningful additional typed-history regressions were attempted, but their schema reconstruction and rollback assertions expose pre-existing Qwen3.5 XML parser loses schema strings and exposes undeclared tool calls #530. Those strict assertions remain failing and must be addressed through a separate parser fix before the expanded acceptance evidence can pass. No gate bypass or externally enforced readiness claim.

[github-actions]

Chat app preview deployed for 2acefdc.

• App: https://leehack-llamadart-chat-pr-529.static.hf.space
• Space: https://huggingface.co/spaces/leehack/llamadart-chat-pr-529
• Repo: leehack/llamadart-chat-pr-529