Struct ReplaySanitizingAgentLlmClient 

pub struct ReplaySanitizingAgentLlmClient { /* private fields */ }

Agent-layer companion to ReplaySanitizingLlmClient.

Meerkat session services can also receive already-adapted AgentLlmClients through live replacement and hot-swap APIs. Sanitize at that boundary too so provider-emitted server tool telemetry is never replayed into the next stateless model request just because the client entered below the raw LlmClient adapter seam.

Implementations

impl ReplaySanitizingAgentLlmClient

pub fn new(inner: Arc<dyn AgentLlmClient>) -> Self

pub fn wrap(inner: Arc<dyn AgentLlmClient>) -> Arc<dyn AgentLlmClient>

Trait Implementations

impl AgentLlmClient for ReplaySanitizingAgentLlmClient

fn stream_response<'life0, 'life1, 'life2, 'life3, 'async_trait>( &'life0 self, messages: &'life1 [Message], tools: &'life2 [Arc<ToolDef>], max_tokens: u32, temperature: Option<f32>, provider_params: Option<&'life3 ProviderParamsOverride>, ) -> Pin<Box<dyn Future<Output = Result<LlmStreamResult, AgentError>> + Send + 'async_trait>>

where Self: 'async_trait, 'life0: 'async_trait, 'life1: 'async_trait, 'life2: 'async_trait, 'life3: 'async_trait,

Stream a response from the LLM

fn provider(&self) -> Provider

Get the typed catalog provider identity for this client.

fn model(&self) -> &str

Get the current effective model identifier.

fn compile_schema( &self, output_schema: &OutputSchema, ) -> Result<CompiledSchema, SchemaError>

Compile an output schema for this provider.

fn prepare_model_fallback( &self, _failure: &AgentError, ) -> Option<AgentLlmFallbackSwitch>

Prepare the next prebuilt fallback model after the generated turn authority has classified the LLM failure as recoverable.

fn commit_model_fallback(&self, _identity: &SessionLlmIdentity)

Commit a prepared fallback after the agent loop has applied the paired request policy, identity, and tool-visibility state.

fn active_capability_base_filter(&self) -> ToolFilter

Capability base filter for the current active model.

fn active_max_output_tokens(&self) -> Option<u32>

Output-token ceiling for the current active model, if the catalog knows one. Used on every provider call so a sticky fallback keeps respecting the backup model’s smaller output limit after the failover turn.

fn begin_stream_output_observation(&self)

Reset per-call observation of user-visible streaming output.

fn stream_output_observed(&self) -> bool

Whether the current LLM call has emitted user-visible streaming output.