Capability matrix

What each shipped provider declares. The session reads these flags to route behavior — a capability a provider doesn't declare fails typed at the call site instead of degrading silently.

Provider capabilities

How to read it:

• no (opt-in) — the SDK cannot know what a self-hosted backend implements, so the freeform defaults are conservative; declare what your server honors via its capabilities parameter. See your own backend.
• profile-dependent — image blocks ride the wire unconditionally; the tier's model decides what it can see, and the cloud advertises the capabilities it resolved for your agent and tier in its handshake. The cloud-profile provider starts from a conservative seed and adopts those advertised values once a turn streams, so capabilities reflects the wire rather than a fixed client-side guess (you can also seed them at construction). See send images.
• — means no declared value: backends manage their own token budgets and tool limits server-side.
• OS 27+ rows are platform-conditional: on macOS 26 / iOS 26 / visionOS 26 the flag is false; on OS 27 vision reflects what the model itself reports. The on-device context window is measured from the device's model at construction; the Private Cloud Compute configuration reports its own window and requiresNetworking. See on-device with Apple.
• managed conversation: no everywhere — no shipped provider stores the conversation for you; the session sends the history it owns. See conversations.

Run limit defaults

Every turn is bounded by AgentRunLimits. These are the defaults (AgentRunLimits.default), all overridable per agent:

Behavior and the error each trips: bound the turn.