aws.BedrockClient$stream

A foundation-model id (anthropic.claude-3-5-haiku-20241022-v1:0), a cross-region inference-profile id (us.anthropic.claude-haiku-4-5-…), or a full model/inference-profile ARN. It is a URL path label, not a body field: / and : are percent-encoded so an ARN stays one path segment. There is no default — model availability is per-account and per-region, so there is no id that is right everywhere.

Explicit region. null resolves AWS_REGION, then AWS_DEFAULT_REGION, then the active profile's region; if none of those produce one, the request fails with a message naming all of them (unless endpoint_url is set, which needs no region for the URL — signing still scopes to the resolved region).

Named profile in the shared config/credentials files. null uses AWS_PROFILE, else default.

Explicit static credentials. Both access_key_id and secret_access_key must be set to short-circuit the provider chain; a lone key id falls through rather than signing with half an identity.

Session token for temporary (STS/SSO) credentials.

Replaces https://bedrock-runtime.<region>.amazonaws.com — a VPC endpoint, a LocalStack instance, or a test server. A trailing slash is tolerated. Setting it bypasses region resolution for the URL.

A literal string, or env.NAME (a late-bound baml.env.Ref read at request time) so a client Name = ... declaration — evaluated during $init, which cannot run io sysops — stays pure while a deployment still points the client somewhere else.

Maximum tokens to generate. Wire type is a 32-bit int.

toolConfig.toolChoice: "auto", "any" (alias "required"), or a tool name to force. null omits the field, which is the API default and the only shape every Bedrock model family accepts.

NOTE on replayed tool results: toolResult.status ("success"/"error") is only supported by Amazon Nova and Anthropic Claude 3/4, so it is sent only when model names one of those families. Other families still get the result text, which carries the tool's error message.

Append a cachePoint after the system content, making the system prompt reusable across calls. Per-message checkpoints come from the cache_point key of ai.PromptMessage.metadata instead.

Append a cachePoint after the tool list, making the tool definitions reusable across calls.

additionalModelRequestFields: model-specific parameters Converse has no first-class field for, e.g. {"top_k": 50} or {"reasoning_config": {"type": "enabled", "budget_tokens": 1024}}. This is the engine's additional_model_request_fields option.

Deep-merged into the serialized request as the last step before it is signed: maps merge recursively, a null value deletes that key, and everything else replaces. Use it to reach any Converse field the typed request does not model (guardrailConfig, promptVariables, …).

Extra request headers, applied BEFORE signing so they are covered by the signature. Names are lowercased; a name the client already sets is replaced, not duplicated.

Extra query parameters, appended (and percent-encoded) BEFORE signing.

Record request/response bodies on the LLMCall wire records delivered through on_event. Status, headers, and timing are always recorded; this flag only gates the bodies.