Fix null handling in tool schemas for chat templates

[atdrendel]

No description provided.

[Copilot]

🔵 Needs a closer look

The core change alters a shared chat-templating path used for all tool schemas with subtle Swift Optional/existential semantics whose downstream behavior cannot be verified here, and the added tests are on-device model tests that do not run in CI.

Pull request overview

This PR fixes rendering failures that occur when tool JSON schemas contain null values (e.g., nullable types like ["string", "null"] and enums such as ["person", "organization", null]). In this codebase such nulls are represented as NSNull() inside the any Sendable tool dictionaries. The core change makes _Tokenizer.applyChatTemplate recursively convert NSNull to Swift nil before delegating the tool schemas to the upstream tokenizer, so the chat-template engine can render them. The remainder of the PR adds a shared test-input helper and a matching per-model test across every model suite.

Changes:

• Add a recursive replacingNulls helper in TokenizerLoader.swift and apply it to tools before calling upstream.applyChatTemplate.
• Add a nullableToolSearchInput(enableThinking:) helper in Helpers.swift that builds a multi-turn tool-search conversation with a nullable contact_update schema.
• Add canContinueAfterToolSearchRevealsNullableSchema tests to 19 model test files, following each file's existing conventions (e.g., Gemma files guard on SHLLM.isSupportedDevice, Qwen/others rely on the loadModel early-return).

File summaries

File	Description
Sources/SHLLM/TokenizerLoader.swift	Core fix: recursively replace NSNull with nil in tool schemas before chat templating
Tests/SHLLMTests/Helpers.swift	New nullableToolSearchInput helper providing a nullable tool schema and multi-turn messages
Tests/SHLLMTests/Models/Qwen3-*Tests.swift, Qwen3_5-*Tests.swift, Qwen2_5-*Tests.swift	New nullable-schema tool test per Qwen suite
Tests/SHLLMTests/Models/Gemma4-*Tests.swift	New nullable-schema tool test with isSupportedDevice guard and enable_thinking: true
Tests/SHLLMTests/Models/Orchestrator-8BTests.swift, NemotronNano-30BTests.swift, Ministral-3-14BTests.swift, LFM2-8B-A1BTests.swift, GPTOSS-20BTests.swift, Devstral2Small-24BTests.swift	New nullable-schema tool test per model suite

Review details

• Files reviewed: 23/23 changed files
• Comments generated: 0
• Review effort level: Balanced

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.