Crate polyc_llm 

Provider-agnostic LLM trait and wire types for polychrome.

This crate defines the LlmProvider trait that every concrete provider backend implements, and the language-shaped Rust types that flow through it.

LlmProvider is the seam that lets the planner swap LLM backends without touching its own code: one impl crate per provider, dispatched behind a dyn LlmProvider trait object.

Modules

• request — CompletionRequest and everything reachable from it (Message, Content, ToolSpec, ToolChoice).
• chunk — Chunk (streaming response events) and friends (Usage, StopReason).
• error — the LlmError trait bound that LlmProvider::Error must satisfy.

The trait itself lives in this crate root; the wire types live in the modules above and are re-exported here for convenience.

Re-exports

pub use chunk::Chunk;

pub use chunk::StopReason;

pub use chunk::Usage;

pub use erased::BoxError;

pub use erased::DynProvider;

pub use erased::ErasedProvider;

pub use erased::into_dyn;

pub use error::LlmError;

pub use error::LlmErrorKind;

pub use error::kind_from_http_status;

pub use error::parse_retry_after;

pub use model_info::FALLBACK_CONTEXT_WINDOW;

pub use model_info::ModelInfo;

pub use model_info::lookup_model;

pub use model_info::lookup_model_or_fallback;

pub use preflight::PreflightReport;

pub use preflight::ProbeOutcome;

pub use preflight::preflight;

pub use request::APPROVAL_STATUS_GROUND_RULE;

pub use request::CacheHint;

pub use request::CompletionRequest;

pub use request::Content;

pub use request::GATED_TOOL_APPROVAL_NOTE;

pub use request::ImageRef;

pub use request::JsonSchema;

pub use request::Message;

pub use request::Role;

pub use request::ToolCall;

pub use request::ToolChoice;

pub use request::ToolResult;

pub use request::ToolSpec;

Modules

chunk

Streaming response types: Chunk, Usage, and StopReason.

erased

Type erasure for LlmProvider so the control plane can hold a single Arc<dyn LlmProvider> regardless of which backend is configured.

error

LlmError marker trait and reference DummyError implementation.

model_info

Bundled model catalog: per-model context-window facts and the longest-prefix lookup that drives compaction.

preflight

Startup preflight: verify a backend actually delivers native tool calling and schema-constrained structured output, against the live endpoint.

request

Request-side LLM types: CompletionRequest, Message, Content, ToolSpec, ToolChoice, and JsonSchema.

sse

Shared Server-Sent Events framing helpers for streaming providers.

turn

Turn helpers: a StubProvider for wiring/tests and collect_turn, which folds a provider’s Chunk stream into a single TurnOutput.

Traits

LlmProvider

The seam between the planner and any concrete LLM backend.

Functions

init_metrics

Force-register this crate’s Prometheus call-latency histogram.