Add missing Foundation Models 26 APIs

[mattt]

This PR adds the Foundation Models 26 APIs that AnyLanguageModel was missing:

• @Generable(description:representNilExplicitlyInGeneratedContent:), GenerationSchema.init(type:description:representNilExplicitlyInGeneratedContent:properties:), and DynamicGenerationSchema.init(name:description:representNilExplicitlyInGeneratedContent:properties:)
• DynamicGenerationSchema.null
• logFeedbackAttachment(sentiment:issues:desiredResponseText:) and logFeedbackAttachment(sentiment:issues:desiredResponseContent:)
• SystemLanguageModel.supportedLanguages, supportsLocale(_:), and the five tokenCount(for:) overloads (OS 26.4)

To match Foundation Models, a @Generable type's generated content now leaves out optional properties whose value is nil, instead of writing null. Pass representNilExplicitlyInGeneratedContent: true to keep the null values. With that flag, a session also fills in null for optional properties that a model's structured response leaves out, as Foundation Models does.

This PR also fixes SystemLanguageModel.logFeedbackAttachment and PrivateCloudComputeLanguageModel.logFeedbackAttachment, which dropped desiredOutput instead of passing it to Foundation Models.

[Copilot]

Copilot review overview

🟡 Changes recommended

Explicit-null behavior is lost for streaming, union schemas, and schema serialization round trips.

Review effort: Balanced
Findings: 1 · 2

Open (3)

• Recurse into matching anyOf variant for explicit-nil traversal · New
• Preserve explicit-nil metadata during schema encoding and equality · New
• Propagate converted rawContent to the terminal response snapshot · New

What changed in this PR

Adds missing Foundation Models 26 APIs, explicit optional-nil representation, null schemas, feedback helpers, and system-model metadata/token-counting APIs.

Changes:

• Adds configurable explicit-nil generation and null schema support.
• Adds feedback convenience APIs and forwards desired output correctly.
• Exposes supported-language and token-count APIs.

File	Description
Sources/​AnyLanguageModel/​Generable.swift	Adds the explicit-nil macro overload.
Sources/​AnyLanguageModelMacros/​GenerableMacro.swift	Generates omitted or explicit-null optional properties.
Sources/​AnyLanguageModel/​GenerationSchema.swift	Adds null nodes and explicit-nil postprocessing.
Sources/​AnyLanguageModel/​DynamicGenerationSchema.swift	Adds dynamic null and explicit-nil schemas.
Sources/​AnyLanguageModel/​LanguageModelSession.swift	Applies explicit-null behavior and adds feedback helpers.
Sources/​AnyLanguageModel/​Transcript.swift	Exposes response schemas internally.
Sources/​AnyLanguageModel/​Models/​SystemLanguageModel.swift	Adds language/token APIs and feedback forwarding.
Sources/​AnyLanguageModel/​Shared/​StructuredGeneration.swift	Generates constrained null values.
Sources/​AnyLanguageModel/​Shared/​ResponseSchemaInlining.swift	Supports null nodes during inlining.
Tests/​AnyLanguageModelTests/​ExplicitNilTests.swift	Tests explicit and omitted nil behavior.
Tests/​AnyLanguageModelTests/​DynamicGenerationSchemaTests.swift	Tests null-schema coding.
Tests/​AnyLanguageModelTests/​LanguageModelFeedbackTests.swift	Tests desired feedback responses.
Tests/​AnyLanguageModelTests/​SystemLanguageModelTests.swift	Tests languages and token counting.

---

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

[Copilot]

Copilot review overview

🟡 Changes recommended

Dynamic response consistency, union handling, constrained nullable generation, and tvOS compilation issues remain.

Review effort: Balanced
Findings: 2 · 2

Open (4)

• Nullable anyOf generation never selects the null branch · New
• tvOS test calls unavailable tokenCount overloads · New
• Object variant matching ignores required keys · New
• Preserve explicit-nil metadata during schema encoding and equality

Resolved since last review (2)

• Recurse into matching anyOf variant for explicit-nil traversal
• Propagate converted rawContent to the terminal response snapshot

Previously missed (1)

In code that hasn't changed since last review

Schema overload returns untransformed response content

Sources/​AnyLanguageModel/​LanguageModelSession.swift:218

For the schema-based overloads, Content is GeneratedContent, so this keeps response.content in its original form while rawContent gains the explicit null properties. Callers reading the primary content value therefore see different data from rawContent; use the transformed value when it can be cast to Content.

[Copilot]

Copilot review overview

🟡 Changes recommended

Schema-based responses can return inconsistent normalized content and rawContent values.

Review effort: Balanced
Findings: 2 · 1

Open (3)

• Response content remains unnormalized in schema overloads · New
• Nullable anyOf generation never selects the null branch
• Preserve explicit-nil metadata during schema encoding and equality

Resolved since last review (2)

• tvOS test calls unavailable tokenCount overloads
• Object variant matching ignores required keys

[Copilot]

Copilot review overview

🔵 Needs a closer look

Private Cloud Compute still discards desired feedback responses passed through the new convenience APIs.

Review effort: Balanced
Findings: 1 · 1

Open (2)

• Nullable anyOf generation never selects the null branch
• Preserve explicit-nil metadata during schema encoding and equality

Resolved since last review (1)

• Response content remains unnormalized in schema overloads

Previously missed (1)

In code that hasn't changed since last review

Forward desiredOutput in PrivateCloudComputeLanguageModel feedback

Sources/​AnyLanguageModel/​LanguageModelSession.swift:1087

Both new convenience overloads ultimately rely on each model's desiredOutput forwarding. PrivateCloudComputeLanguageModel.logFeedbackAttachment still hard-codes desiredOutput: nil (Sources/AnyLanguageModel/Models/PrivateCloudComputeLanguageModel.swift:137), even though it creates the same Foundation Models session, so these APIs silently discard the requested text/content for PCC. Convert and forward the entry there as this PR now does for SystemLanguageModel.

[Copilot]

Copilot review overview

🔵 Needs a closer look

Nullable unions cannot select null during constrained generation, and nested unions skip explicit-nil normalization.

Review effort: Balanced
Findings: 1 · 1

Open (2)

• Nullable anyOf generation never selects the null branch
• Preserve explicit-nil metadata during schema encoding and equality

Previously missed (1)

In code that hasn't changed since last review

Support nested anyOf schemas during content matching

Sources/​AnyLanguageModel/​GenerationSchema.swift:1014

Nested anyOf schemas are never considered a match. For example, an outer union containing a referenced inner [ExplicitNilObject, .null] resolves that reference to .anyOf; this switch rejects it, so traversal stops and omitted optionals in the explicit object are not filled with null. Recurse into nested union variants when matching the content shape.

[Copilot]

Copilot review overview

🟡 Changes recommended

Nonliteral macro flags and pre-26.4 null-schema fallback can silently produce incorrect generated content.

Review effort: Balanced
Findings: 2

Open (2)

• Null schemas incorrectly fall back to unconstrained strings · New
• Nullable anyOf generation never selects the null branch

Resolved since last review (1)

• Preserve explicit-nil metadata during schema encoding and equality

Previously missed (1)

In code that hasn't changed since last review

Boolean macro expressions are incorrectly treated as false

Sources/​AnyLanguageModelMacros/​GenerableMacro.swift:99

Any valid Boolean expression other than a bare literal—such as a static constant or even a parenthesized true—is silently interpreted as false, so the generated content contradicts the value supplied to the public macro. Preserve the argument expression in generated code, or emit a macro diagnostic requiring a Boolean literal rather than changing its meaning.

[Copilot]

Copilot review overview

🔵 Needs a closer look

Typed and streamed response content can remain inconsistent with normalized explicit-null raw content.

Review effort: Balanced
Findings: 2

Open (2)

• Null schemas incorrectly fall back to unconstrained strings
• Nullable anyOf generation never selects the null branch

Previously missed (1)

In code that hasn't changed since last review

Initialize final snapshot partial content from normalized content

Sources/​AnyLanguageModel/​LanguageModelSession.swift:271

rawContent cannot be cast to a typed Content.PartiallyGenerated, so the synthetic final snapshot pairs normalized raw content with the previous, unnormalized partial value. Consumers iterating the stream can therefore see omitted keys through snapshot.content.generatedContent even though snapshot.rawContent contains explicit nulls. Initialize the partial value from the normalized content instead.

[Copilot]

Copilot review overview

🟡 Changes recommended

Nullable-union matching, property ordering, and tvOS availability handling can currently produce incorrect behavior.

Review effort: Balanced
Findings: 2 · 4

Open (6)

• Null schemas incorrectly fall back to unconstrained strings
• Nullable anyOf generation never selects the null branch
• Preserve property declaration order when appending missing keys · New
• Validate property values before selecting a union object variant · New
• Add correct tvOS availability to the language model check · New
• Apply tvOS availability to nullable union capability checks · New

[Copilot]

Copilot review overview

🔵 Needs a closer look

tvOS availability checks are incomplete and union normalization can select the wrong object variant.

Review effort: Balanced
Findings: 2 · 4

Open (6)

• Null schemas incorrectly fall back to unconstrained strings
• Nullable anyOf generation never selects the null branch
• Apply tvOS availability to nullable union capability checks
• Add correct tvOS availability to the language model check
• Validate property values before selecting a union object variant
• Preserve property declaration order when appending missing keys

[Copilot]

Copilot review overview

🟡 Changes recommended

Feedback sessions omit conversation history, and explicit-nil normalization can select or order schema fields incorrectly.

Review effort: Balanced
Findings: 2 · 3

Open (5)

• Null schemas incorrectly fall back to unconstrained strings
• Nullable anyOf generation never selects the null branch
• Include property order when comparing explicit-nil schemas · New
• Preserve declared order when inserting omitted optional fields · New
• Validate property values before selecting a union object variant

Resolved since last review (3)

• Apply tvOS availability to nullable union capability checks
• Add correct tvOS availability to the language model check
• Preserve property declaration order when appending missing keys

Previously missed (2)

In code that hasn't changed since last review

Rehydrate session transcript before attaching feedback

Sources/​AnyLanguageModel/​Models/​PrivateCloudComputeLanguageModel.swift:137

This newly created Foundation Models session does not contain session.transcript, so it has no most recent PCC response to which the feedback and desired output can be attached. Rehydrate it from the converted transcript before calling logFeedbackAttachment, matching the generation path.

Feedback logs against a fresh session without the response transcript

Sources/​AnyLanguageModel/​Models/​SystemLanguageModel.swift:296

The Foundation Models session is created only from tools and instructions, so it has none of session.transcript—including the response this feedback is supposed to describe. Passing desiredOutput therefore still logs against a fresh session rather than the actual exchange. Initialize the Foundation Models session from the converted transcript, as the response paths do.