Module models 

Re-exports

pub use tool_call::FunctionCall;

pub use tool_call::ToolCall;

Modules

catalog

The data-driven model-capability catalog: one FIRST-MATCH-WINS const table replacing the model-name string gates that were scattered across the adapters (anthropic thinking/temperature/effort tiers, gemini thinking dispatch, ollama’s gpt-oss think-string, openai-compat reasoning-model and vision markers) and the domain’s static context windows.

tool_call

Wire-format tool call type.

Structs

BackendConfig

Backend connection configuration

ChatMessage

Represents a chat message

ModelCapabilities

Capability flags advertised by a model adapter.

ModelConfig

Unified model configuration

ModelResponse

Response from a model

OllamaOptions

Ollama-specific options (extracted from backend_options)

ProviderProfile

Static description of one OpenAI-compatible provider.

ReasoningChunk

A typed chunk of reasoning content emitted during streaming.

ResponseDebugContext

Correlation ids captured from a provider’s HTTP response headers. Appended (one plain-text line) to the user-facing error message so bug reports to the provider can quote them; deliberately NOT part of Display, which feeds logs and try_extract_error_message.

TokenUsage

Token usage statistics normalized across providers.

UserFacingError

User-facing error information with actionable suggestions

Enums

BackendError

Backend-specific errors

ChatMessageKind

CompatStyle

User-friendly string form for compat = "..." in config.toml when a fully custom provider needs to declare which profile shape to follow.

ConfigError

Configuration errors

ErrorCategory

Error categories for visual differentiation

FinishReason

Why a model stopped generating, normalized across providers.

MaxTokensParam

Provider-specific spelling for the completion-token budget.

MessageRole

MetaResponseItem

One Meta Responses output item saved for stateless replay. Reasoning items split their encrypted payload from the remaining JSON so the ciphertext can be serialized as base64 bytes. This keeps generic persistence redaction from mistaking a ciphertext for a credential and corrupting the replay state.

ModelError

Top-level error type for all model operations

ProviderContinuation

Opaque provider-owned state that must be replayed with a committed assistant turn. The reducer carries this as inert data; only the matching provider interprets it.

ReasoningCapability

Describes the reasoning controls a model exposes.

ReasoningExtraction

Where reasoning content shows up in a streaming response delta.

ReasoningLevel

Reasoning depth requested for a model call.

ReasoningStrategy

How to put ReasoningLevel onto the wire for a given provider.

StreamEvent

A single event emitted during a streaming model call.

TokenUsageSource

Where a token count came from. Provider-reported counts are the billing/request truth; estimates are only for preflight context diagnostics.

Constants

PROVIDER_REGISTRY

Built-in provider registry. Lookups are case-insensitive on name. Add a provider here when its quirks fit the existing strategies; add a new ReasoningStrategy variant when a provider needs something the existing ones can’t express.

Traits

Model

Core trait that all model adapters implement.

Functions

lookup_provider

Look up a built-in provider by name. Case-insensitive.

nearest_effort

Map a requested reasoning level onto the closest level the model actually supports.

Type Aliases

Result

Result type alias for model operations

StreamCallback

Callback invoked once per StreamEvent during a chat request.