Enum LlmError 

pub enum LlmError {
    RateLimit {
        retry_after_ms: Option<u64>,
        status: u16,
    },
    PaymentRequired {
        status: u16,
    },
    ServerError {
        status: u16,
    },
    ContextOverflow {
        tokens: u32,
        limit: u32,
    },
    BadRequest {
        status: u16,
        body: String,
    },
    Api {
        status: u16,
        body: String,
    },
    Transport(Box<dyn Error + Send + Sync + 'static>),
    Parse(Box<dyn Error + Send + Sync + 'static>),
    Other(Box<dyn Error + Send + Sync + 'static>),
}

Typed error produced at the AiModel boundary.

Variants

RateLimit

Fields

retry_after_ms: Option<u64>

status: u16

PaymentRequired

Fields

status: u16

ServerError

Fields

status: u16

ContextOverflow

vLLM/OpenAI-compatible context-window-exceeded error after reactive shrink retries are exhausted. Carries the server- reported limit (model max context tokens) and tokens (the request’s input-token count) so dashboards can chart how often agents are crossing which model’s window without re-parsing the error message.

Fields

tokens: u32

limit: u32

BadRequest

HTTP 400 that isn’t a recognized context-overflow. The provider’s response body carries the real reason (bad tool schema, token math, unsupported param). Display stays terse — the body can echo a fragment of the prompt — so logs/dumps don’t leak it; the body is exposed only via LlmError::detail, which operator-local surfaces (e.g. quorum smoke-test) opt into showing.

Fields

status: u16

body: String

Api

A non-2xx HTTP status that isn’t one of the specific categories above (401 auth, 403, 404 model-unavailable, 422, …). Preserves status structurally so telemetry charts it and operators can tell a dead model id (404) from an auth failure (401) — the previous Other catch-all flattened all of these to an opaque “other” with no status. body is withheld from Display like BadRequest; recover it via LlmError::detail.

Fields

status: u16

body: String

Transport(Box<dyn Error + Send + Sync + 'static>)

Parse(Box<dyn Error + Send + Sync + 'static>)

Other(Box<dyn Error + Send + Sync + 'static>)

Implementations

impl LlmError

pub fn display_chain(&self) -> String

Render the full error chain (this error + every wrapped source) as a multi-line string, root cause last:

transport
  caused by: hyper error
  caused by: connection reset by peer (os error 104)

LlmError only wraps a source on Transport, Parse, and Other; the structured variants (RateLimit, PaymentRequired, ServerError, ContextOverflow) render as a single line because they carry their detail inline in the thiserror format string.

Operators see the chain in logs (tracing::error!(error.chain = %err.display_chain(), ...)) and dashboards (the JSON shape exposes one string field) without losing the root cause to a flattened one-liner.

pub fn detail(&self) -> Option<&str>

Provider-supplied detail that Display deliberately withholds. Currently the captured HTTP 400 response body. Operator-local surfaces show this; logs/dumps keep using Display/display_chain so the body never leaks server-side.

pub fn classify(&self) -> (LlmErrorClass, Option<u16>)

Map a typed error to the telemetry taxonomy.

Trait Implementations

impl Debug for LlmError

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.

impl Display for LlmError

fn fmt(&self, __formatter: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.

impl Error for LlmError

fn source(&self) -> Option<&(dyn Error + 'static)>

Returns the lower-level source of this error, if any.

fn description(&self) -> &str

👎Deprecated since 1.42.0:

use the Display impl or to_string()

fn cause(&self) -> Option<&dyn Error>

👎Deprecated since 1.33.0:

replaced by Error::source, which can support downcasting

fn provide<'a>(&'a self, request: &mut Request<'a>)

🔬This is a nightly-only experimental API. (error_generic_member_access)

Provides type-based access to context intended for error reports.