ci: Version Packages

[github-actions]

This PR was opened by the Changesets release GitHub action. When you're ready to do a release, you can merge this and the packages will be published to npm automatically. If you're not ready to do a release yet, that's fine, whenever you add more changesets to main, this PR will be updated.

Releases

@tanstack/ai@0.20.0

Minor Changes

• feat(ai): systemPrompts accept { content, metadata } with adapter-inferred metadata typing (#575)

  chat({ systemPrompts }) now accepts either a plain string (the existing
  shape — fully backward compatible) or { content, metadata }. The metadata
  field's type is inferred from the adapter via a new
  TSystemPromptMetadata generic on TextAdapter / BaseTextAdapter:

  • @tanstack/ai-anthropic declares AnthropicSystemPromptMetadata →
    users get cache_control autocomplete and type-checking on
    systemPrompts[i].metadata for Anthropic chats.
  • Adapters with no per-prompt metadata (OpenAI, Gemini, Ollama,
    OpenRouter, openai-base) inherit the default never, which means the
    metadata field carries no meaningful value at the call site —
    TypeScript only accepts undefined there. Provider-foreign metadata
    that reaches an adapter via JS / as any is silently dropped, never
    written to the wire.

  import { chat } from '@tanstack/ai'
  import { anthropicText } from '@tanstack/ai-anthropic'
  
  // Anthropic — `cache_control` is autocompleted, no `satisfies` needed.
  chat({
    adapter: anthropicText({ apiKey }, 'claude-sonnet-4-6'),
    systemPrompts: [
      {
        content: 'Stable instructions — cache me.',
        metadata: { cache_control: { type: 'ephemeral' } },
      },
      'Volatile per-request instruction.',
    ],
  })
  
  // OpenAI — `metadata` is `never`; only `undefined` is assignable, so the
  // field is effectively unusable. The object form without `metadata` still
  // works for portability.
  chat({
    adapter: openaiText({ apiKey }, 'gpt-4o-mini'),
    systemPrompts: [
      'Plain string.',
      { content: 'Object form without metadata is allowed.' },
    ],
  })

  New exports:

  • @tanstack/ai: SystemPrompt, NormalizedSystemPrompt types and the
    normalizeSystemPrompts() helper adapters use to normalize the wide
    input shape to { content, metadata? } before consumption.
  • @tanstack/ai-anthropic: AnthropicSystemPromptMetadata interface
    (currently exposes cache_control for prompt caching).

  Internal:

  • New TSystemPromptMetadata = never generic on TextAdapter /
    BaseTextAdapter, surfaced via '~types'['systemPromptMetadata']
    for inference at the chat() call site.
  • Anthropic adapter reads metadata.cache_control and attaches it to
    the corresponding TextBlockParam.
  • All other text adapters call normalizeSystemPrompts() and join
    .content for their respective instructions / system /
    systemInstruction fields. Foreign metadata that reaches them via JS
    / as any is dropped (never written to the wire).
  • normalizeSystemPrompts() is the public API boundary and throws
    TypeError (naming the offending index) for object-form entries whose
    content isn't a string — preventing literal "undefined" from
    reaching the model on stale call sites.
  • OpenTelemetry middleware attaches per-prompt metadata as the
    tanstack.ai.system_prompt.metadata JSON span attribute when
    captureContent: true and at least one entry carries metadata, so
    observability backends can distinguish cache hit/miss for Anthropic.
  • @tanstack/ai-event-client mirrors the SystemPrompt shape locally
    (avoids a circular import) and projects metadata away on the devtools
    wire — devtools UI still receives Array<string>.

Patch Changes

• Updated dependencies [496db9c]:
  • @tanstack/ai-event-client@0.3.5

@tanstack/ai-anthropic@0.10.0

Minor Changes

• feat(ai): systemPrompts accept { content, metadata } with adapter-inferred metadata typing (#575)

  chat({ systemPrompts }) now accepts either a plain string (the existing
  shape — fully backward compatible) or { content, metadata }. The metadata
  field's type is inferred from the adapter via a new
  TSystemPromptMetadata generic on TextAdapter / BaseTextAdapter:

  • @tanstack/ai-anthropic declares AnthropicSystemPromptMetadata →
    users get cache_control autocomplete and type-checking on
    systemPrompts[i].metadata for Anthropic chats.
  • Adapters with no per-prompt metadata (OpenAI, Gemini, Ollama,
    OpenRouter, openai-base) inherit the default never, which means the
    metadata field carries no meaningful value at the call site —
    TypeScript only accepts undefined there. Provider-foreign metadata
    that reaches an adapter via JS / as any is silently dropped, never
    written to the wire.

  import { chat } from '@tanstack/ai'
  import { anthropicText } from '@tanstack/ai-anthropic'
  
  // Anthropic — `cache_control` is autocompleted, no `satisfies` needed.
  chat({
    adapter: anthropicText({ apiKey }, 'claude-sonnet-4-6'),
    systemPrompts: [
      {
        content: 'Stable instructions — cache me.',
        metadata: { cache_control: { type: 'ephemeral' } },
      },
      'Volatile per-request instruction.',
    ],
  })
  
  // OpenAI — `metadata` is `never`; only `undefined` is assignable, so the
  // field is effectively unusable. The object form without `metadata` still
  // works for portability.
  chat({
    adapter: openaiText({ apiKey }, 'gpt-4o-mini'),
    systemPrompts: [
      'Plain string.',
      { content: 'Object form without metadata is allowed.' },
    ],
  })

  New exports:

  • @tanstack/ai: SystemPrompt, NormalizedSystemPrompt types and the
    normalizeSystemPrompts() helper adapters use to normalize the wide
    input shape to { content, metadata? } before consumption.
  • @tanstack/ai-anthropic: AnthropicSystemPromptMetadata interface
    (currently exposes cache_control for prompt caching).

  Internal:

  • New TSystemPromptMetadata = never generic on TextAdapter /
    BaseTextAdapter, surfaced via '~types'['systemPromptMetadata']
    for inference at the chat() call site.
  • Anthropic adapter reads metadata.cache_control and attaches it to
    the corresponding TextBlockParam.
  • All other text adapters call normalizeSystemPrompts() and join
    .content for their respective instructions / system /
    systemInstruction fields. Foreign metadata that reaches them via JS
    / as any is dropped (never written to the wire).
  • normalizeSystemPrompts() is the public API boundary and throws
    TypeError (naming the offending index) for object-form entries whose
    content isn't a string — preventing literal "undefined" from
    reaching the model on stale call sites.
  • OpenTelemetry middleware attaches per-prompt metadata as the
    tanstack.ai.system_prompt.metadata JSON span attribute when
    captureContent: true and at least one entry carries metadata, so
    observability backends can distinguish cache hit/miss for Anthropic.
  • @tanstack/ai-event-client mirrors the SystemPrompt shape locally
    (avoids a circular import) and projects metadata away on the devtools
    wire — devtools UI still receives Array<string>.

Patch Changes

• Updated dependencies [496db9c]:
  • @tanstack/ai@0.20.0

@tanstack/ai-client@0.11.2

Patch Changes

• Updated dependencies [496db9c]:
  • @tanstack/ai@0.20.0
  • @tanstack/ai-event-client@0.3.5

@tanstack/ai-code-mode@0.1.15

Patch Changes

• Updated dependencies [496db9c]:
  • @tanstack/ai@0.20.0

@tanstack/ai-code-mode-skills@0.1.15

Patch Changes

• Updated dependencies [496db9c]:
  • @tanstack/ai@0.20.0
  • @tanstack/ai-code-mode@0.1.15

@tanstack/ai-devtools-core@0.3.32

Patch Changes

• Updated dependencies [496db9c]:
  • @tanstack/ai@0.20.0
  • @tanstack/ai-event-client@0.3.5

@tanstack/ai-elevenlabs@0.2.7

Patch Changes

• Updated dependencies [496db9c]:
  • @tanstack/ai@0.20.0
  • @tanstack/ai-client@0.11.2

@tanstack/ai-event-client@0.3.5

Patch Changes

• feat(ai): systemPrompts accept { content, metadata } with adapter-inferred metadata typing (#575)

  chat({ systemPrompts }) now accepts either a plain string (the existing
  shape — fully backward compatible) or { content, metadata }. The metadata
  field's type is inferred from the adapter via a new
  TSystemPromptMetadata generic on TextAdapter / BaseTextAdapter:

  • @tanstack/ai-anthropic declares AnthropicSystemPromptMetadata →
    users get cache_control autocomplete and type-checking on
    systemPrompts[i].metadata for Anthropic chats.
  • Adapters with no per-prompt metadata (OpenAI, Gemini, Ollama,
    OpenRouter, openai-base) inherit the default never, which means the
    metadata field carries no meaningful value at the call site —
    TypeScript only accepts undefined there. Provider-foreign metadata
    that reaches an adapter via JS / as any is silently dropped, never
    written to the wire.

  import { chat } from '@tanstack/ai'
  import { anthropicText } from '@tanstack/ai-anthropic'
  
  // Anthropic — `cache_control` is autocompleted, no `satisfies` needed.
  chat({
    adapter: anthropicText({ apiKey }, 'claude-sonnet-4-6'),
    systemPrompts: [
      {
        content: 'Stable instructions — cache me.',
        metadata: { cache_control: { type: 'ephemeral' } },
      },
      'Volatile per-request instruction.',
    ],
  })
  
  // OpenAI — `metadata` is `never`; only `undefined` is assignable, so the
  // field is effectively unusable. The object form without `metadata` still
  // works for portability.
  chat({
    adapter: openaiText({ apiKey }, 'gpt-4o-mini'),
    systemPrompts: [
      'Plain string.',
      { content: 'Object form without metadata is allowed.' },
    ],
  })

  New exports:

  • @tanstack/ai: SystemPrompt, NormalizedSystemPrompt types and the
    normalizeSystemPrompts() helper adapters use to normalize the wide
    input shape to { content, metadata? } before consumption.
  • @tanstack/ai-anthropic: AnthropicSystemPromptMetadata interface
    (currently exposes cache_control for prompt caching).

  Internal:

  • New TSystemPromptMetadata = never generic on TextAdapter /
    BaseTextAdapter, surfaced via '~types'['systemPromptMetadata']
    for inference at the chat() call site.
  • Anthropic adapter reads metadata.cache_control and attaches it to
    the corresponding TextBlockParam.
  • All other text adapters call normalizeSystemPrompts() and join
    .content for their respective instructions / system /
    systemInstruction fields. Foreign metadata that reaches them via JS
    / as any is dropped (never written to the wire).
  • normalizeSystemPrompts() is the public API boundary and throws
    TypeError (naming the offending index) for object-form entries whose
    content isn't a string — preventing literal "undefined" from
    reaching the model on stale call sites.
  • OpenTelemetry middleware attaches per-prompt metadata as the
    tanstack.ai.system_prompt.metadata JSON span attribute when
    captureContent: true and at least one entry carries metadata, so
    observability backends can distinguish cache hit/miss for Anthropic.
  • @tanstack/ai-event-client mirrors the SystemPrompt shape locally
    (avoids a circular import) and projects metadata away on the devtools
    wire — devtools UI still receives Array<string>.

• Updated dependencies [496db9c]:

  • @tanstack/ai@0.20.0

@tanstack/ai-fal@0.7.8

Patch Changes

• Updated dependencies [496db9c]:
  • @tanstack/ai@0.20.0

@tanstack/ai-gemini@0.10.7

Patch Changes

• feat(ai): systemPrompts accept { content, metadata } with adapter-inferred metadata typing (#575)

  chat({ systemPrompts }) now accepts either a plain string (the existing
  shape — fully backward compatible) or { content, metadata }. The metadata
  field's type is inferred from the adapter via a new
  TSystemPromptMetadata generic on TextAdapter / BaseTextAdapter:

  • @tanstack/ai-anthropic declares AnthropicSystemPromptMetadata →
    users get cache_control autocomplete and type-checking on
    systemPrompts[i].metadata for Anthropic chats.
  • Adapters with no per-prompt metadata (OpenAI, Gemini, Ollama,
    OpenRouter, openai-base) inherit the default never, which means the
    metadata field carries no meaningful value at the call site —
    TypeScript only accepts undefined there. Provider-foreign metadata
    that reaches an adapter via JS / as any is silently dropped, never
    written to the wire.

  import { chat } from '@tanstack/ai'
  import { anthropicText } from '@tanstack/ai-anthropic'
  
  // Anthropic — `cache_control` is autocompleted, no `satisfies` needed.
  chat({
    adapter: anthropicText({ apiKey }, 'claude-sonnet-4-6'),
    systemPrompts: [
      {
        content: 'Stable instructions — cache me.',
        metadata: { cache_control: { type: 'ephemeral' } },
      },
      'Volatile per-request instruction.',
    ],
  })
  
  // OpenAI — `metadata` is `never`; only `undefined` is assignable, so the
  // field is effectively unusable. The object form without `metadata` still
  // works for portability.
  chat({
    adapter: openaiText({ apiKey }, 'gpt-4o-mini'),
    systemPrompts: [
      'Plain string.',
      { content: 'Object form without metadata is allowed.' },
    ],
  })

  New exports:

  • @tanstack/ai: SystemPrompt, NormalizedSystemPrompt types and the
    normalizeSystemPrompts() helper adapters use to normalize the wide
    input shape to { content, metadata? } before consumption.
  • @tanstack/ai-anthropic: AnthropicSystemPromptMetadata interface
    (currently exposes cache_control for prompt caching).

  Internal:

  • New TSystemPromptMetadata = never generic on TextAdapter /
    BaseTextAdapter, surfaced via '~types'['systemPromptMetadata']
    for inference at the chat() call site.
  • Anthropic adapter reads metadata.cache_control and attaches it to
    the corresponding TextBlockParam.
  • All other text adapters call normalizeSystemPrompts() and join
    .content for their respective instructions / system /
    systemInstruction fields. Foreign metadata that reaches them via JS
    / as any is dropped (never written to the wire).
  • normalizeSystemPrompts() is the public API boundary and throws
    TypeError (naming the offending index) for object-form entries whose
    content isn't a string — preventing literal "undefined" from
    reaching the model on stale call sites.
  • OpenTelemetry middleware attaches per-prompt metadata as the
    tanstack.ai.system_prompt.metadata JSON span attribute when
    captureContent: true and at least one entry carries metadata, so
    observability backends can distinguish cache hit/miss for Anthropic.
  • @tanstack/ai-event-client mirrors the SystemPrompt shape locally
    (avoids a circular import) and projects metadata away on the devtools
    wire — devtools UI still receives Array<string>.

• Updated dependencies [496db9c]:

  • @tanstack/ai@0.20.0

@tanstack/ai-grok@0.8.4

Patch Changes

• Updated dependencies [496db9c]:
  • @tanstack/ai@0.20.0
  • @tanstack/openai-base@0.3.3

@tanstack/ai-groq@0.2.3

Patch Changes

• Updated dependencies [496db9c]:
  • @tanstack/ai@0.20.0
  • @tanstack/openai-base@0.3.3

@tanstack/ai-isolate-cloudflare@0.2.6

Patch Changes

• Updated dependencies []:
  • @tanstack/ai-code-mode@0.1.15

@tanstack/ai-isolate-node@0.1.15

Patch Changes

• Updated dependencies []:
  • @tanstack/ai-code-mode@0.1.15

@tanstack/ai-isolate-quickjs@0.1.15

Patch Changes

• Updated dependencies []:
  • @tanstack/ai-code-mode@0.1.15

@tanstack/ai-ollama@0.6.18

Patch Changes

• feat(ai): systemPrompts accept { content, metadata } with adapter-inferred metadata typing (#575)

  chat({ systemPrompts }) now accepts either a plain string (the existing
  shape — fully backward compatible) or { content, metadata }. The metadata
  field's type is inferred from the adapter via a new
  TSystemPromptMetadata generic on TextAdapter / BaseTextAdapter:

  • @tanstack/ai-anthropic declares AnthropicSystemPromptMetadata →
    users get cache_control autocomplete and type-checking on
    systemPrompts[i].metadata for Anthropic chats.
  • Adapters with no per-prompt metadata (OpenAI, Gemini, Ollama,
    OpenRouter, openai-base) inherit the default never, which means the
    metadata field carries no meaningful value at the call site —
    TypeScript only accepts undefined there. Provider-foreign metadata
    that reaches an adapter via JS / as any is silently dropped, never
    written to the wire.

  import { chat } from '@tanstack/ai'
  import { anthropicText } from '@tanstack/ai-anthropic'
  
  // Anthropic — `cache_control` is autocompleted, no `satisfies` needed.
  chat({
    adapter: anthropicText({ apiKey }, 'claude-sonnet-4-6'),
    systemPrompts: [
      {
        content: 'Stable instructions — cache me.',
        metadata: { cache_control: { type: 'ephemeral' } },
      },
      'Volatile per-request instruction.',
    ],
  })
  
  // OpenAI — `metadata` is `never`; only `undefined` is assignable, so the
  // field is effectively unusable. The object form without `metadata` still
  // works for portability.
  chat({
    adapter: openaiText({ apiKey }, 'gpt-4o-mini'),
    systemPrompts: [
      'Plain string.',
      { content: 'Object form without metadata is allowed.' },
    ],
  })

  New exports:

  • @tanstack/ai: SystemPrompt, NormalizedSystemPrompt types and the
    normalizeSystemPrompts() helper adapters use to normalize the wide
    input shape to { content, metadata? } before consumption.
  • @tanstack/ai-anthropic: AnthropicSystemPromptMetadata interface
    (currently exposes cache_control for prompt caching).

  Internal:

  • New TSystemPromptMetadata = never generic on TextAdapter /
    BaseTextAdapter, surfaced via '~types'['systemPromptMetadata']
    for inference at the chat() call site.
  • Anthropic adapter reads metadata.cache_control and attaches it to
    the corresponding TextBlockParam.
  • All other text adapters call normalizeSystemPrompts() and join
    .content for their respective instructions / system /
    systemInstruction fields. Foreign metadata that reaches them via JS
    / as any is dropped (never written to the wire).
  • normalizeSystemPrompts() is the public API boundary and throws
    TypeError (naming the offending index) for object-form entries whose
    content isn't a string — preventing literal "undefined" from
    reaching the model on stale call sites.
  • OpenTelemetry middleware attaches per-prompt metadata as the
    tanstack.ai.system_prompt.metadata JSON span attribute when
    captureContent: true and at least one entry carries metadata, so
    observability backends can distinguish cache hit/miss for Anthropic.
  • @tanstack/ai-event-client mirrors the SystemPrompt shape locally
    (avoids a circular import) and projects metadata away on the devtools
    wire — devtools UI still receives Array<string>.

• Updated dependencies [496db9c]:

  • @tanstack/ai@0.20.0

@tanstack/ai-openai@0.9.4

Patch Changes

• feat(ai): systemPrompts accept { content, metadata } with adapter-inferred metadata typing (#575)

  chat({ systemPrompts }) now accepts either a plain string (the existing
  shape — fully backward compatible) or { content, metadata }. The metadata
  field's type is inferred from the adapter via a new
  TSystemPromptMetadata generic on TextAdapter / BaseTextAdapter:

  • @tanstack/ai-anthropic declares AnthropicSystemPromptMetadata →
    users get cache_control autocomplete and type-checking on
    systemPrompts[i].metadata for Anthropic chats.
  • Adapters with no per-prompt metadata (OpenAI, Gemini, Ollama,
    OpenRouter, openai-base) inherit the default never, which means the
    metadata field carries no meaningful value at the call site —
    TypeScript only accepts undefined there. Provider-foreign metadata
    that reaches an adapter via JS / as any is silently dropped, never
    written to the wire.

  import { chat } from '@tanstack/ai'
  import { anthropicText } from '@tanstack/ai-anthropic'
  
  // Anthropic — `cache_control` is autocompleted, no `satisfies` needed.
  chat({
    adapter: anthropicText({ apiKey }, 'claude-sonnet-4-6'),
    systemPrompts: [
      {
        content: 'Stable instructions — cache me.',
        metadata: { cache_control: { type: 'ephemeral' } },
      },
      'Volatile per-request instruction.',
    ],
  })
  
  // OpenAI — `metadata` is `never`; only `undefined` is assignable, so the
  // field is effectively unusable. The object form without `metadata` still
  // works for portability.
  chat({
    adapter: openaiText({ apiKey }, 'gpt-4o-mini'),
    systemPrompts: [
      'Plain string.',
      { content: 'Object form without metadata is allowed.' },
    ],
  })

  New exports:

  • @tanstack/ai: SystemPrompt, NormalizedSystemPrompt types and the
    normalizeSystemPrompts() helper adapters use to normalize the wide
    input shape to { content, metadata? } before consumption.
  • @tanstack/ai-anthropic: AnthropicSystemPromptMetadata interface
    (currently exposes cache_control for prompt caching).

  Internal:

  • New TSystemPromptMetadata = never generic on TextAdapter /
    BaseTextAdapter, surfaced via '~types'['systemPromptMetadata']
    for inference at the chat() call site.
  • Anthropic adapter reads metadata.cache_control and attaches it to
    the corresponding TextBlockParam.
  • All other text adapters call normalizeSystemPrompts() and join
    .content for their respective instructions / system /
    systemInstruction fields. Foreign metadata that reaches them via JS
    / as any is dropped (never written to the wire).
  • normalizeSystemPrompts() is the public API boundary and throws
    TypeError (naming the offending index) for object-form entries whose
    content isn't a string — preventing literal "undefined" from
    reaching the model on stale call sites.
  • OpenTelemetry middleware attaches per-prompt metadata as the
    tanstack.ai.system_prompt.metadata JSON span attribute when
    captureContent: true and at least one entry carries metadata, so
    observability backends can distinguish cache hit/miss for Anthropic.
  • @tanstack/ai-event-client mirrors the SystemPrompt shape locally
    (avoids a circular import) and projects metadata away on the devtools
    wire — devtools UI still receives Array<string>.

• Updated dependencies [496db9c]:

  • @tanstack/ai@0.20.0
  • @tanstack/openai-base@0.3.3
  • @tanstack/ai-client@0.11.2

@tanstack/ai-openrouter@0.9.4

Patch Changes

• feat(ai): systemPrompts accept { content, metadata } with adapter-inferred metadata typing (#575)

  chat({ systemPrompts }) now accepts either a plain string (the existing
  shape — fully backward compatible) or { content, metadata }. The metadata
  field's type is inferred from the adapter via a new
  TSystemPromptMetadata generic on TextAdapter / BaseTextAdapter:

  • @tanstack/ai-anthropic declares AnthropicSystemPromptMetadata →
    users get cache_control autocomplete and type-checking on
    systemPrompts[i].metadata for Anthropic chats.
  • Adapters with no per-prompt metadata (OpenAI, Gemini, Ollama,
    OpenRouter, openai-base) inherit the default never, which means the
    metadata field carries no meaningful value at the call site —
    TypeScript only accepts undefined there. Provider-foreign metadata
    that reaches an adapter via JS / as any is silently dropped, never
    written to the wire.

  import { chat } from '@tanstack/ai'
  import { anthropicText } from '@tanstack/ai-anthropic'
  
  // Anthropic — `cache_control` is autocompleted, no `satisfies` needed.
  chat({
    adapter: anthropicText({ apiKey }, 'claude-sonnet-4-6'),
    systemPrompts: [
      {
        content: 'Stable instructions — cache me.',
        metadata: { cache_control: { type: 'ephemeral' } },
      },
      'Volatile per-request instruction.',
    ],
  })
  
  // OpenAI — `metadata` is `never`; only `undefined` is assignable, so the
  // field is effectively unusable. The object form without `metadata` still
  // works for portability.
  chat({
    adapter: openaiText({ apiKey }, 'gpt-4o-mini'),
    systemPrompts: [
      'Plain string.',
      { content: 'Object form without metadata is allowed.' },
    ],
  })

  New exports:

  • @tanstack/ai: SystemPrompt, NormalizedSystemPrompt types and the
    normalizeSystemPrompts() helper adapters use to normalize the wide
    input shape to { content, metadata? } before consumption.
  • @tanstack/ai-anthropic: AnthropicSystemPromptMetadata interface
    (currently exposes cache_control for prompt caching).

  Internal:

  • New TSystemPromptMetadata = never generic on TextAdapter /
    BaseTextAdapter, surfaced via '~types'['systemPromptMetadata']
    for inference at the chat() call site.
  • Anthropic adapter reads metadata.cache_control and attaches it to
    the corresponding TextBlockParam.
  • All other text adapters call normalizeSystemPrompts() and join
    .content for their respective instructions / system /
    systemInstruction fields. Foreign metadata that reaches them via JS
    / as any is dropped (never written to the wire).
  • normalizeSystemPrompts() is the public API boundary and throws
    TypeError (naming the offending index) for object-form entries whose
    content isn't a string — preventing literal "undefined" from
    reaching the model on stale call sites.
  • OpenTelemetry middleware attaches per-prompt metadata as the
    tanstack.ai.system_prompt.metadata JSON span attribute when
    captureContent: true and at least one entry carries metadata, so
    observability backends can distinguish cache hit/miss for Anthropic.
  • @tanstack/ai-event-client mirrors the SystemPrompt shape locally
    (avoids a circular import) and projects metadata away on the devtools
    wire — devtools UI still receives Array<string>.

• Updated dependencies [496db9c]:

  • @tanstack/ai@0.20.0

@tanstack/ai-preact@0.6.27

Patch Changes

• Updated dependencies [496db9c]:
  • @tanstack/ai@0.20.0
  • @tanstack/ai-client@0.11.2

@tanstack/ai-react@0.11.2

Patch Changes

• Updated dependencies [496db9c]:
  • @tanstack/ai@0.20.0
  • @tanstack/ai-client@0.11.2

@tanstack/ai-solid@0.10.2

Patch Changes

• Updated dependencies [496db9c]:
  • @tanstack/ai@0.20.0
  • @tanstack/ai-client@0.11.2

@tanstack/ai-svelte@0.10.2

Patch Changes

• Updated dependencies [496db9c]:
  • @tanstack/ai@0.20.0
  • @tanstack/ai-client@0.11.2

@tanstack/ai-vue@0.10.3

Patch Changes

• Updated dependencies [496db9c]:
  • @tanstack/ai@0.20.0
  • @tanstack/ai-client@0.11.2

@tanstack/ai-vue-ui@0.1.39

Patch Changes

• Updated dependencies []:
  • @tanstack/ai-vue@0.10.3

@tanstack/openai-base@0.3.3

Patch Changes

• feat(ai): systemPrompts accept { content, metadata } with adapter-inferred metadata typing (#575)

  chat({ systemPrompts }) now accepts either a plain string (the existing
  shape — fully backward compatible) or { content, metadata }. The metadata
  field's type is inferred from the adapter via a new
  TSystemPromptMetadata generic on TextAdapter / BaseTextAdapter:

  • @tanstack/ai-anthropic declares AnthropicSystemPromptMetadata →
    users get cache_control autocomplete and type-checking on
    systemPrompts[i].metadata for Anthropic chats.
  • Adapters with no per-prompt metadata (OpenAI, Gemini, Ollama,
    OpenRouter, openai-base) inherit the default never, which means the
    metadata field carries no meaningful value at the call site —
    TypeScript only accepts undefined there. Provider-foreign metadata
    that reaches an adapter via JS / as any is silently dropped, never
    written to the wire.

  import { chat } from '@tanstack/ai'
  import { anthropicText } from '@tanstack/ai-anthropic'
  
  // Anthropic — `cache_control` is autocompleted, no `satisfies` needed.
  chat({
    adapter: anthropicText({ apiKey }, 'claude-sonnet-4-6'),
    systemPrompts: [
      {
        content: 'Stable instructions — cache me.',
        metadata: { cache_control: { type: 'ephemeral' } },
      },
      'Volatile per-request instruction.',
    ],
  })
  
  // OpenAI — `metadata` is `never`; only `undefined` is assignable, so the
  // field is effectively unusable. The object form without `metadata` still
  // works for portability.
  chat({
    adapter: openaiText({ apiKey }, 'gpt-4o-mini'),
    systemPrompts: [
      'Plain string.',
      { content: 'Object form without metadata is allowed.' },
    ],
  })

  New exports:

  • @tanstack/ai: SystemPrompt, NormalizedSystemPrompt types and the
    normalizeSystemPrompts() helper adapters use to normalize the wide
    input shape to { content, metadata? } before consumption.
  • @tanstack/ai-anthropic: AnthropicSystemPromptMetadata interface
    (currently exposes cache_control for prompt caching).

  Internal:

  • New TSystemPromptMetadata = never generic on TextAdapter /
    BaseTextAdapter, surfaced via '~types'['systemPromptMetadata']
    for inference at the chat() call site.
  • Anthropic adapter reads metadata.cache_control and attaches it to
    the corresponding TextBlockParam.
  • All other text adapters call normalizeSystemPrompts() and join
    .content for their respective instructions / system /
    systemInstruction fields. Foreign metadata that reaches them via JS
    / as any is dropped (never written to the wire).
  • normalizeSystemPrompts() is the public API boundary and throws
    TypeError (naming the offending index) for object-form entries whose
    content isn't a string — preventing literal "undefined" from
    reaching the model on stale call sites.
  • OpenTelemetry middleware attaches per-prompt metadata as the
    tanstack.ai.system_prompt.metadata JSON span attribute when
    captureContent: true and at least one entry carries metadata, so
    observability backends can distinguish cache hit/miss for Anthropic.
  • @tanstack/ai-event-client mirrors the SystemPrompt shape locally
    (avoids a circular import) and projects metadata away on the devtools
    wire — devtools UI still receives Array<string>.

• Updated dependencies [496db9c]:

  • @tanstack/ai@0.20.0

@tanstack/preact-ai-devtools@0.1.36

Patch Changes

• Updated dependencies []:
  • @tanstack/ai-devtools-core@0.3.32

@tanstack/react-ai-devtools@0.2.36

Patch Changes

• Updated dependencies []:
  • @tanstack/ai-devtools-core@0.3.32

@tanstack/solid-ai-devtools@0.2.36

Patch Changes

• Updated dependencies []:
  • @tanstack/ai-devtools-core@0.3.32