LLMProvider
Interface: LLMProvider
Defined in: src/provider-contract/types.ts:646
Properties
carriesForcedToolChoice?
readonlyoptionalcarriesForcedToolChoice?:boolean
Defined in: src/provider-contract/types.ts:730
v7.26 — whether this adapter puts LLMRequest.toolChoice on its wire as a forced choice of one named tool.
Absence means NO, not "probably". That is the opposite of
carriesInMessages, whose absence means the floor every wire supports,
and the difference is what the two capabilities are for. A role that
quietly vanishes costs a message; a tool choice that quietly vanishes
costs the guarantee the consumer selected the strategy FOR — the model
would answer in whatever shape it liked while the config said the shape
was constrained. So an agent using strategy: 'tool-forced' on a
provider that has not declared this refuses at run start, by name.
Declare it only where it is true of the endpoint, not of the SDK: the
OpenAI adapter declares it for real OpenAI and Azure and NOT behind a
custom baseURL (Ollama, vLLM, Together, …), because what an
OpenAI-compatible server does with tool_choice is that server's
business and this library does not get to promise it.
A WRAPPER must forward it; withFallback publishes the AND of the two
providers it holds. Model-specific restrictions are declared separately
by supportsForcedToolChoice; this existing wire flag remains boolean.
carriesInMessages?
readonlyoptionalcarriesInMessages?: readonlyWireRole[]
Defined in: src/provider-contract/types.ts:706
v7.21 — which roles this provider carries INSIDE the messages array.
The wires disagree, and the disagreement is invisible from the outside:
the Anthropic-family adapters (Anthropic, Bedrock, Browser Anthropic) DROP
a role: 'system' message inside messages because system rides a
separate top-level field, while the OpenAI-family adapters carry it. So a
slot: 'messages' injection with role: 'system' would arrive on one
provider and vanish on another — and nothing in the recording would tell
the two apart. Declaring the capability is what lets the engine refuse at
run start instead, naming the provider and the roles it does carry.
Consulted at DELIVERY time by the agent's Deliver stage. A role that is
not listed is REFUSED, never silently re-roled: changing who appears to
speak is a meaning change the app must make, not the library.
Optional, and absence is not "carries everything" — a provider that
omits it is treated as ['user', 'assistant']
(DEFAULT_CARRIES_IN_MESSAGES), the floor every known wire
supports. Declare it if your adapter carries more.
A WRAPPER must forward it (the three src/resilience/ decorators do);
withFallback publishes the INTERSECTION of the two providers it holds,
because a role only one of them carries is a role the call might drop.
carriesNativeOutputSchema?
readonlyoptionalcarriesNativeOutputSchema?:boolean
Defined in: src/provider-contract/types.ts:750
Whether this adapter puts LLMRequest.outputFormat on its wire —
the provider's own structured output, strategy: 'native'. Absence
means NO, for the reason carriesForcedToolChoice gives: a format that
quietly vanishes costs the guarantee the strategy was chosen for. Declared
by anthropic(), browserAnthropic() and the mock; read with
nativeOutputSchemaFor(provider, model). A WRAPPER must forward it;
withFallback publishes the AND of its pair.
name
readonlyname:string
Defined in: src/provider-contract/types.ts:647
promptCaching?
readonlyoptionalpromptCaching?:PromptCaching| ((model) => PromptCaching | undefined)
Defined in: src/provider-contract/types.ts:657
How this provider's wire caches a repeated prompt prefix (PromptCaching,
from agentfootprint/cache). The agent picks its cache strategy from
this, never from name. Absent = no caching the agent drives or meters.
A wrapper must forward it, like carriesInMessages.
A model-aware declaration returns undefined for an unknown/unsupported
model. cacheStrategyFor(provider, model) resolves either form; adapters
must consult the same rule before putting cache markers on the wire.
supportsForcedToolChoice?
readonlyoptionalsupportsForcedToolChoice?: (model) =>boolean
Defined in: src/provider-contract/types.ts:740
Optional model-specific restriction on the forced-choice wire capability.
Read with forcedToolChoiceFor(provider, model): the wire flag must be
true AND this predicate (when present) must answer true. Absence preserves
the wire flag's existing behaviour. The adapter resolves its shorthand.
Like thinkingMode, this function must not depend on this; wrappers
forward it as a value. withFallback intersects both sides per model.
Parameters
model
string
Returns
boolean
supportsNativeOutputSchema?
readonlyoptionalsupportsNativeOutputSchema?: (model) =>boolean
Defined in: src/provider-contract/types.ts:756
Optional per-model restriction on carriesNativeOutputSchema, the
supportsForcedToolChoice twin: both must answer true. A plain function
of the model id (no this); the adapter resolves its own shorthand.
Parameters
model
string
Returns
boolean
thinkingHandler?
readonlyoptionalthinkingHandler?:ThinkingHandler
Defined in: src/provider-contract/types.ts:665
The handler that normalizes this wire's thinking (rawThinking). The
agent uses it unless .thinkingHandler() overrides it; absent = no
thinking stage. Never chosen by name — a wrapper must forward it.
thinkingMode?
readonlyoptionalthinkingMode?: (model) =>ThinkingMode
Defined in: src/provider-contract/types.ts:680
Which thinking request each model takes on this wire — 'budget',
'adaptive' or 'none' (ThinkingMode, from agentfootprint/providers),
asked with the model id a request names (the adapter resolves its own
shorthand, e.g. 'anthropic'). The adapter maps LLMRequest.thinking to
the right wire shape for that model; the agent reads this at build to
refuse .thinking() for a model that cannot think and to say, in dev
mode, when 'adaptive' will not send the budget.
Absent = no per-model claim: requests carry thinking: { budget } and the
adapter decides. A plain function of the model id — it must not use
this, because a wrapper forwards it as a value, like carriesInMessages;
withFallback answers for its pair.
Parameters
model
string
Returns
ThinkingMode
Methods
complete()
complete(
req,hooks?):Promise<LLMResponse>
Defined in: src/provider-contract/types.ts:764
hooks (v7.8) is optional and additive — implementations may declare
complete(req) with no second parameter and stay assignable. A LEAF
provider (one that talks to a vendor) may ignore it. A WRAPPER must
forward it, or everything it wraps goes silently dark — see the
LLMCallHooks docs above.
Parameters
req
hooks?
Returns
Promise<LLMResponse>
stream()?
optionalstream(req,hooks?):AsyncIterable<LLMChunk>
Defined in: src/provider-contract/types.ts:765
Parameters
req
hooks?
Returns
AsyncIterable<LLMChunk>