Skip to main content

Model & Provider Layer

Architectural Layer: L3 Intelligence
Source Directory: src/model
File Count: 16 source files

NVIDIA NIM provider, SSE streaming, model catalog, dynamic discovery, and token usage.

Module Files & Detailed Reference​

Source FileLinesKey TypesResponsibility
catalog.rs657ModelCatalog, CatalogRefreshState, CatalogSourceDynamic Model Catalog (PRD §01, MDL-01, D-05). Provides the canonical...
mod.rs22Functions/TraitsModel subsystem. Owns model provider abstractions, wire protocol seri...
invocation.rs434ModelInvocationRecord, SqliteModelInvocationRepositoryDurable SQLite repository for ModelInvocation records and token usage ...
mod.rs5Functions/TraitsPersistence layer for model subsystem records and telemetry (MDL-05).
pricing.rs150ModelPricecanonical pricing table and cost calculator for model invocations. co...
protocol.rs384ToolCallProtocol, ToolCallValidationOutcomeLayered tool-call protocol, argument extraction, truncation detection,...
endpoint.rs399ValidatedProviderEndpoint, EndpointTrustError, EndpointTrustSourceProvider endpoint trust boundary (P0-01). Invariant: NO CREDENTIAL → ...
mod.rs146Functions/TraitsProvider-neutral abstractions for LLM reasoning and streaming. Core p...
nvidia.rs1778NvidiaProviderProduction NvidiaProvider targeting NVIDIA Build / NIM OpenAI-compatib...
nvidia_metadata.rs1215FeaturedModelEntry, FeaturedModelsResponse, NvidiaModelMetadataResolverNVIDIA Model Metadata Enrichment Pipeline (Section 1–15). Provides au...
sse.rs1012AccumulatedToolCall, StreamAccumulatorAsynchronous SSE stream accumulator, OpenAI-compatible JSON event pars...
health.rs554CircuitBreakerRegistry, ModelCircuitBreaker, ModelHealthMetricsRuntime-local circuit breaker with semantic failure classification (D-...
mod.rs11Functions/TraitsModel routing, circuit breaker health tracking, and recovery subsystem...
recovery.rs523ModelRecoveryCoordinator, StepRecoveryDecisionTwo-level bounded model recovery coordinator (D-07, MDL-04). Level 1:...
resolver.rs1088ContextLimits, ModelCandidate, ModelResolutionDetailsTwo-stage deterministic model router with eligibility filtering and ti...
types.rs544ModelInvocationOptions, ModelToolCall, TokenUsageCanonical domain types for model communication, proposals, streaming, ...

src/model/catalog.rs​

Primary Responsibility​

Dynamic Model Catalog (PRD §01, MDL-01, D-05). Provides the canonical in-memory and cached model catalog for the M31A runtime. Represents dynamically discovered provider model capabilities, context limits, availability states, and deterministic fallback selection without hardcoded model lists.

Types & Trait Declarations​

  • Structs (1): ModelCatalog
  • Enums (3): CatalogRefreshState, CatalogSource, ModelCacheStatus

Key Functions & Signatures (15 shown)​

// private
fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result;
// private
fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result;
// private
fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result;
// private
fn default_catalog_schema_version() -> u32;
// pub
fn new(provider: impl Into<String>) -> Self;
// pub
fn from_discovered(provider: impl Into<String>,
        models: Vec<ModelCandidate>,
        timestamp: u64,) -> Self;
// pub
fn from_cache(provider: impl Into<String>,
        models: Vec<ModelCandidate>,
        timestamp: u64,
        is_stale;
// pub
fn failed_no_cache(provider: impl Into<String>) -> Self;
// pub
fn update_from_provider(&mut self,
        provider: impl Into<String>,
        models: Vec<ModelCandidate>,);
// pub
fn is_empty(&self) -> bool;
// pub
fn len(&self) -> usize;
// pub
fn models_for_tier(&self, tier: ModelTier) -> Vec<&ModelCandidate>;
// pub
fn find_model(&self, model_id: &str) -> Option<&ModelCandidate>;
// pub
fn contains_model(&self, model_id: &str) -> bool;
// pub
fn is_inventory_stale(&self, max_age_secs: u64) -> bool;

Unit Verification (2 inline tests)​

Tests defined within module: test_catalog_exact_context_capacity_roundtrip, test_catalog_legacy_unversioned_cache_is_stale

src/model/mod.rs​

Primary Responsibility​

Model subsystem. Owns model provider abstractions, wire protocol serialization, SSE streaming, model routing, health management, and token telemetry (MDL-01 to MDL-05). Core principle: The model proposes. The runtime decides.

src/model/persistence/invocation.rs​

Primary Responsibility​

Durable SQLite repository for ModelInvocation records and token usage telemetry (D-08, MDL-05). Maps to the model_invocations table created in Migration 005.

Types & Trait Declarations​

  • Structs (2): ModelInvocationRecord, SqliteModelInvocationRepository

Key Functions & Signatures (12 shown)​

// pub
fn new(mission_id: MissionId,
        task_id: TaskId,
        agent_id: AgentId,
        step_number: u32,
        prov;
// pub
fn with_prompt_provenance(mut self, provenance: impl Into<String>) -> Self;
// pub
fn with_cost(mut self, cost_usd: Option<f64>, provenance: CostProvenance) -> Self;
// pub
fn new(pool: SqlitePool) -> Self;
// pub
async fn insert_invocation(&self,
        record: &ModelInvocationRecord,) -> Result<(), sqlx::Error>;
// pub
async fn get_invocations_for_task(&self,
        task_id: &TaskId,) -> Result<Vec<ModelInvocationRecord>, sqlx::Error>;
// pub
async fn get_invocations_for_mission(&self,
        mission_id: &MissionId,) -> Result<Vec<ModelInvocationRecord>, sqlx:;
// pub
async fn get_total_token_usage_for_mission(&self,
        mission_id: &MissionId,) -> Result<TokenUsage, sqlx::Error>;
// pub
async fn get_all_invocations(&self) -> Result<Vec<ModelInvocationRecord>, sqlx::Error>;
// private
fn map_row_to_invocation(row: SqliteRow) -> Result<ModelInvocationRecord, sqlx::Error>;
// private
async fn seed_test_parents(pool: &SqlitePool,
        mission_id: &MissionId,
        task_id: &TaskId,
        agent_id;
// private
async fn test_sqlite_model_invocation_crud();

Unit Verification (1 inline tests)​

Tests defined within module: test_sqlite_model_invocation_crud

src/model/persistence/mod.rs​

Primary Responsibility​

Persistence layer for model subsystem records and telemetry (MDL-05).

src/model/pricing.rs​

Primary Responsibility​

canonical pricing table and cost calculator for model invocations. computes estimated cost when the provider does not supply authoritative cost. never returns synthetic 0.00 for unknown commercial models; unknown models yield (None, CostProvenance::Unknown).

Types & Trait Declarations​

  • Structs (1): ModelPrice

Key Functions & Signatures (6 shown)​

// pub
fn calculate_cost_usd(&self, prompt_tokens: usize, completion_tokens: usize) -> f64;
// pub
fn lookup_price(provider: &str, model: &str) -> Option<ModelPrice>;
// pub
fn calculate_cost(provider: &str,
    model: &str,
    prompt_tokens: usize,
    completion_tokens: usize,) -> (Option<f;
// pub
fn calculate_cost_from_usage(provider: &str,
    model: &str,
    usage: &crate::model::types::TokenUsage,) -> (Option<f;
// private
fn test_unknown_models_return_none_and_unknown_provenance();
// private
fn test_known_models_and_local_models_pricing();

Unit Verification (2 inline tests)​

Tests defined within module: test_unknown_models_return_none_and_unknown_provenance, test_known_models_and_local_models_pricing

src/model/protocol.rs​

Primary Responsibility​

Layered tool-call protocol, argument extraction, truncation detection, and constrained safe recovery (P3-B, MDL-01, MDL-03).

Types & Trait Declarations​

  • Structs (1): ToolCallProtocol
  • Enums (1): ToolCallValidationOutcome

Key Functions & Signatures (10 shown)​

// pub
fn process_arguments(raw_args: &str,
        finish_reason: Option<&str>,) -> ToolCallValidationOutcome;
// pub
fn detect_unclosed_delimiters(text: &str) -> bool;
// pub
fn attempt_safe_recovery(raw: &str) -> Option<(Value, String)>;
// private
fn extract_from_markdown_or_quotes(text: &str) -> Option<String>;
// private
fn test_valid_json();
// private
fn test_empty_or_null_json();
// private
fn test_truncated_json_recoveries();
// private
fn test_markdown_codeblock_extraction();
// private
fn test_syntax_error_unrecoverable();
// private
fn test_token_limit_length_truncation();

Unit Verification (6 inline tests)​

Tests defined within module: test_valid_json, test_empty_or_null_json, test_truncated_json_recoveries, test_markdown_codeblock_extraction, test_syntax_error_unrecoverable, test_token_limit_length_truncation

src/model/provider/endpoint.rs​

Primary Responsibility​

Provider endpoint trust boundary (P0-01). Invariant: NO CREDENTIAL → BEFORE ENDPOINT TRUST. A repository-controlled workspace configuration must never be able to redirect a credential-bearing production provider request to an attacker-controlled URL. Endpoint authorization happens before any credential is loaded or attached, and every credential-bearing HTTP request site must go through a [ValidatedProviderEndpoint]. Trust tiers: - [EndpointTrustSource::BuiltinDefault] — immutable canonical default. - [EndpointTrustSource::System] — /etc/m31a/config.toml (trusted admin). - [EndpointTrustSource::User] — ~/.config/m31a/config.toml (trusted user). - [EndpointTrustSource::ExplicitCli] — explicit --config &lt;path&gt; (trusted operator intent). - [EndpointTrustSource::Workspace] — &lt;ws&gt;/.m31a/config.toml (UNTRUSTED repository content). - [EndpointTrustSource::SessionOverride] — generic session overlay (untrusted by default). Policy (production): - Canonical https://integrate.api.nvidia.com/v1 is always trusted. - Any custom endpoint carrying a real credential MUST come from a trusted tier (System/User/ExplicitCli/BuiltinDefault). Workspace/Session custom endpoints are rejected when a real credential would be attached. - Custom endpoints must be https:// (fail closed on http://), must parse, and must not target blocked destinations (loopback/RFC1918/ link-local/metadata/host aliases). Full DNS validation happens in [ValidatedProviderEndpoint::authorize_for_credential] before use.

Types & Trait Declarations​

  • Structs (1): ValidatedProviderEndpoint
  • Enums (2): EndpointTrustError, EndpointTrustSource

Key Functions & Signatures (15 shown)​

// pub
fn url(&self) -> &str;
// pub
fn source(&self) -> EndpointTrustSource;
// pub
fn is_canonical(&self) -> bool;
// pub
fn credential_permitted(&self) -> bool;
// pub
fn require_credential_permitted(&self) -> Result<&str, EndpointTrustError>;
// pub
fn is_test_credential(key: &str) -> bool;
// private
fn normalize(url: &str) -> String;
// pub
fn is_canonical_endpoint(url: &str) -> bool;
// pub
fn validate_nvidia_endpoint(base_url: Option<String>,
    source: EndpointTrustSource,) -> Result<ValidatedProviderEndpo;
// pub
async fn authorize_for_credential(mut self) -> Result<Self, EndpointTrustError>;
// private
fn canonical_endpoint_always_trusted_from_any_tier();
// private
fn workspace_custom_endpoint_rejected_for_credentials();
// private
fn trusted_tier_custom_https_endpoint_shape_accepted_pending_dns();
// private
fn http_custom_endpoint_rejected_even_from_trusted_tier();
// private
fn loopback_and_private_literals_rejected();

Unit Verification (7 inline tests)​

Tests defined within module: canonical_endpoint_always_trusted_from_any_tier, workspace_custom_endpoint_rejected_for_credentials, trusted_tier_custom_https_endpoint_shape_accepted_pending_dns, http_custom_endpoint_rejected_even_from_trusted_tier, loopback_and_private_literals_rejected, destination_authorization_binds_dns_before_credential, blocked_destination_fails_closed_on_authorize

src/model/provider/mod.rs​

Primary Responsibility​

Provider-neutral abstractions for LLM reasoning and streaming. Core principle: The model proposes. The runtime decides. (MDL-01, Law 1)

Types & Trait Declarations​

  • Traits (1): ModelProvider

Key Functions & Signatures (10 shown)​

// private
async fn call_model(&self,
        model_name: &str,
        system_prompt: &str,
        tools: Vec<serde_json::Value>,;
// private
async fn call_model_with_messages(&self,
        _model_name: &str,
        _messages: &[crate::model::types::ChatMessag;
// private
async fn call_model_with_options(&self,
        model_name: &str,
        system_prompt: &str,
        tools: Vec<serde_;
// private
async fn call_model_with_messages_streaming(&self,
        _model_name: &str,
        _messages: &[crate::model::types::;
// private
async fn call_model_with_messages_streaming_options(&self,
        model_name: &str,
        messages: &[crate::model::t;
// private
async fn stream_model(&self,
        model_name: &str,
        system_prompt: &str,
        tools: Vec<serde_json::Value;
// private
async fn stream_model_with_messages(&self,
        _model_name: &str,
        _messages: &[crate::model::types::ChatMess;
// private
async fn discover_models(&self,) -> Result<Vec<crate::model::router::resolver::ModelCandidate>, ModelError>;
// private
fn is_test_double(&self) -> bool;
// private
fn provider_name(&self) -> &'static str;

src/model/provider/nvidia.rs​

Primary Responsibility​

Production NvidiaProvider targeting NVIDIA Build / NIM OpenAI-compatible API surface (D-01, MDL-01, MDL-03). Connects to NVIDIA NIM endpoints via async SSE streaming. Enforces side-effect prohibition: returns validated ModelProposal for runtime decision.

Types & Trait Declarations​

  • Structs (1): NvidiaProvider

Key Functions & Signatures (15 shown)​

// private
fn record_live_provider_event(event: &str);
// private
fn record_live_provider_event(_event: &str);
// private
fn validate_chat_completion_body(body: &Value) -> Result<(), ModelError>;
// private
fn explicit_tier(item: &serde_json::Value,) -> Option<(crate::model::router::resolver::ModelTier, &'static str)>;
// pub
fn extract_context_capacity(item: &serde_json::Value) -> (usize, &'static str);
// private
fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result;
// pub
fn parse_model_candidate(item: &Value,
        now_secs: u64,) -> Option<crate::model::router::resolver::ModelCandidate>;
// pub
fn new(base_url: Option<String>, api_key: Option<String>) -> Result<Self, ModelError>;
// pub
fn new_governed(base_url: Option<String>,
        api_key: Option<String>,
        source: EndpointTrustSource,) -> Resu;
// pub
fn new_test_fixture(base_url: String, api_key: String) -> Result<Self, ModelError>;
// private
fn construct(base_url: Option<String>,
        api_key: Option<String>,
        source: EndpointTrustSource,) -> Result<;
// pub
fn with_request_tracer(mut self, tracer: RequestTraceHook) -> Self;
// pub
fn with_metadata_resolver(mut self,
        resolver: Arc<crate::model::provider::nvidia_metadata::NvidiaModelMetadataRe;
// pub
fn metadata_resolver(&self,) -> &Arc<crate::model::provider::nvidia_metadata::NvidiaModelMetadataResolver>;
// pub
fn with_client(client: Client,
        base_url: Option<String>,
        api_key: Option<String>,) -> Result<Self, Model;

Unit Verification (23 inline tests)​

Tests defined within module: test_validate_chat_completion_accepts_nonempty_text, test_validate_chat_completion_accepts_tool_calls, test_validate_chat_completion_rejects_missing_or_empty_choices, test_validate_chat_completion_rejects_missing_message, test_validate_chat_completion_rejects_empty_content_without_tool_calls, test_provider_instantiation_with_defaults, test_provider_instantiation_trims_trailing_slash, test_provider_instantiation_fails_without_key, test_build_chat_request_payload_without_tools, test_build_chat_request_payload_with_tools...

src/model/provider/nvidia_metadata.rs​

Primary Responsibility​

NVIDIA Model Metadata Enrichment Pipeline (Section 1–15). Provides authoritative per-model capability normalization, official NVIDIA model reference metadata enrichment, and capability provenance tracking. Architecture: text /v1/models (Inventory) ↓ canonical model inventory ↓ metadata enrichment (NvidiaModelMetadataResolver) ↓ capability normalization ↓ ModelCandidate (with explicit ContextLimits &amp; CapabilitySupport) ↓ ModelCatalog ↓ Router / Wizard / TUI

Types & Trait Declarations​

  • Structs (4): FeaturedModelEntry, FeaturedModelsResponse, NvidiaModelMetadataResolver, NvidiaModelReferenceRecord
  • Traits (1): ProviderModelMetadataSource

Key Functions & Signatures (15 shown)​

// private
async fn enrich_candidate(&self, candidate: &mut ModelCandidate) -> Result<(), ModelError>;
// private
async fn enrich_candidates(&self, candidates: &mut [ModelCandidate]) -> Result<(), ModelError>;
// pub
fn new(model_id: impl Into<String>,
        context_length: usize,
        model_kind: ModelKind,
        tool_calling: ;
// pub
fn with_display_name(mut self, name: impl Into<String>) -> Self;
// pub
fn with_max_output(mut self, max_tokens: usize) -> Self;
// pub
fn with_modalities(mut self, modalities: Vec<String>) -> Self;
// pub
fn new() -> Self;
// pub
fn new_offline() -> Self;
// pub
fn with_featured_models_url(mut self, url: impl Into<String>) -> Self;
// pub
fn with_remote_enrichment(mut self, enabled: bool) -> Self;
// pub
fn register_reference(&mut self, record: NvidiaModelReferenceRecord);
// pub
fn reference_record(&self, model_id: &str) -> Option<&NvidiaModelReferenceRecord>;
// pub
fn lookup_reference(&self, model_id: &str) -> Option<ModelCandidate>;
// private
fn load_official_nvidia_references(&mut self);
// private
async fn get_remote_featured_models(&self) -> HashMap<String, FeaturedModelEntry>;

src/model/provider/sse.rs​

Primary Responsibility​

Asynchronous SSE stream accumulator, OpenAI-compatible JSON event parser, and HTTP error normalizer (D-02, D-03, D-04).

Types & Trait Declarations​

  • Structs (2): AccumulatedToolCall, StreamAccumulator

Key Functions & Signatures (15 shown)​

// pub
fn new() -> Self;
// pub
fn reasoning_text(&self) -> &str;
// pub
fn process_event(&mut self, event_data: &str) -> Result<Vec<StreamChunk>, ModelError>;
// pub
fn finalize(self) -> Result<(ModelProposal, TokenUsage), ModelError>;
// private
fn try_parse_action_envelope(val: &Value) -> Option<ModelProposal>;
// private
fn extract_json_codeblock(text: &str) -> Option<String>;
// private
fn extract_embedded_action(text: &str) -> Option<ModelProposal>;
// pub
fn normalize_http_error(status: u16, body: &str) -> ModelError;
// private
fn sanitize_credentials(text: &str) -> String;
// private
fn extract_cooldown_seconds(body: &str) -> Option<u64>;
// private
fn extract_context_window_numbers(lower_body: &str) -> (usize, usize);
// private
fn test_single_and_multi_chunk_text_streaming();
// private
fn test_multi_chunk_fragmented_tool_call_accumulation();
// private
fn test_malformed_tool_call_json_arguments();
// private
fn test_empty_stream_done_handling();

Unit Verification (10 inline tests)​

Tests defined within module: test_single_and_multi_chunk_text_streaming, test_multi_chunk_fragmented_tool_call_accumulation, test_malformed_tool_call_json_arguments, test_empty_stream_done_handling, test_fallback_envelope_parsing, test_reasoning_content_tracked_without_polluting_answer, test_reasoning_only_stream_rejected_as_empty, test_whitespace_only_stream_rejected_as_empty, test_normalize_http_errors, test_normalize_http_error_credential_redaction

src/model/router/health.rs​

Primary Responsibility​

Runtime-local circuit breaker with semantic failure classification (D-06, MDL-04). Tracks model/endpoint health across Healthy, Degraded, Open, and HalfOpen states. Strictly enforces that client-side bugs (400 Bad Request, 401 Unauthorized, etc.) never degrade provider health or trip circuit breakers.

Types & Trait Declarations​

  • Structs (3): CircuitBreakerRegistry, ModelCircuitBreaker, ModelHealthMetrics
  • Enums (2): CircuitState, FailureKind

Key Functions & Signatures (15 shown)​

// private
fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result;
// pub
fn from_model_error(error: &ModelError) -> Self;
// pub
fn new(endpoint: impl Into<String>, model_name: impl Into<String>) -> Self;
// pub
fn with_failure_threshold(mut self, threshold: u32) -> Self;
// pub
fn check_admission(&mut self) -> Result<(), ModelError>;
// pub
fn check_admission_at(&mut self, now: DateTime<Utc>) -> Result<(), ModelError>;
// pub
fn record_success(&mut self);
// pub
fn record_failure(&mut self, error: &ModelError);
// pub
fn record_failure_at(&mut self, error: &ModelError, now: DateTime<Utc>);
// pub
fn state(&self) -> CircuitState;
// pub
fn consecutive_failures(&self) -> u32;
// pub
fn cooldown_until(&self) -> Option<DateTime<Utc>>;
// pub
fn last_failure_reason(&self) -> Option<&str>;
// private
fn default() -> Self;
// pub
fn new() -> Self;

Unit Verification (5 inline tests)​

Tests defined within module: test_bad_request_does_not_trip_circuit, test_consecutive_503_transitions_healthy_to_degraded_to_open, test_429_immediately_opens_circuit_with_cooldown, test_cooldown_expiry_enables_half_open_probe_and_recovery, test_circuit_breaker_registry_inspection_and_updates

src/model/router/mod.rs​

Primary Responsibility​

Model routing, circuit breaker health tracking, and recovery subsystem (MDL-02, MDL-04). Enforces deterministic two-stage model resolution and runtime-local circuit health.

src/model/router/recovery.rs​

Primary Responsibility​

Two-level bounded model recovery coordinator (D-07, MDL-04). Level 1: In-step transient retry loop strictly bounded by configured budget (default 2 retries). Level 2: Persistent failure or retry exhaustion escalates to Model Fallback via ModelRouter. Fallback re-evaluates hard capability requirements and NEVER weakens task or role constraints (T-07-06).

Types & Trait Declarations​

  • Structs (1): ModelRecoveryCoordinator
  • Enums (1): StepRecoveryDecision

Key Functions & Signatures (14 shown)​

// private
fn default() -> Self;
// pub
fn new() -> Self;
// pub
fn with_bounds(mut self, max_in_step_retries: u32, max_model_fallbacks: u32) -> Self;
// pub
fn max_in_step_retries(&self) -> u32;
// pub
fn max_model_fallbacks(&self) -> u32;
// pub
fn handle_step_failure(&self,
        current_attempt: u32,
        fallback_count: u32,
        error: &ModelError,
   ;
// pub
async fn record_telemetry_attempt(telemetry: &SqliteModelInvocationRepository,
        mission_id: MissionId,
        ta;
// private
fn test_candidates() -> Vec<ModelCandidate>;
// private
fn test_level_1_transient_failure_retries_in_step();
// private
fn test_level_1_retry_exhaustion_escalates_to_level_2_fallback();
// private
fn test_circuit_open_immediately_escalates_to_level_2_fallback();
// private
fn test_fallback_fails_closed_when_hard_constraints_not_met();
// private
async fn seed_test_parents(pool: &sqlx::SqlitePool,
        mission_id: &MissionId,
        task_id: &TaskId,
        ag;
// private
async fn test_telemetry_recording_helper();

Unit Verification (5 inline tests)​

Tests defined within module: test_level_1_transient_failure_retries_in_step, test_level_1_retry_exhaustion_escalates_to_level_2_fallback, test_circuit_open_immediately_escalates_to_level_2_fallback, test_fallback_fails_closed_when_hard_constraints_not_met, test_telemetry_recording_helper

src/model/router/resolver.rs​

Primary Responsibility​

Two-stage deterministic model router with eligibility filtering and tier ranking (D-05, MDL-02). Stage 1: Derives model requirements from AgentProfile/RoutingRequest and applies hard capability filters (tool calling, context capacity, circuit health, budget bounds). Stage 2: Deterministically ranks eligible candidates by tier suitability, operational health, cost preference, and stable model ID tie-breaking.

Types & Trait Declarations​

  • Structs (6): ContextLimits, ModelCandidate, ModelResolutionDetails, ModelRouter, ResolvedModelSelection, RoutingRequest
  • Enums (5): CapabilitySupport, ModelKind, ModelResolutionError, ModelRole, ModelTier

Key Functions & Signatures (15 shown)​

// private
fn default_candidate_availability() -> ProviderCapabilityStatus;
// private
fn default_candidate_source() -> String;
// private
fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result;
// pub
fn is_supported(&self) -> bool;
// pub
fn is_unsupported(&self) -> bool;
// pub
fn is_unknown(&self) -> bool;
// private
fn from(b: bool) -> Self;
// private
fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result;
// pub
fn is_embedding(&self) -> bool;
// pub
fn is_image_generation(&self) -> bool;
// pub
fn is_safety_guard(&self) -> bool;
// pub
fn is_text_or_code_generation(&self) -> bool;
// private
fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result;
// pub
fn new(model_id: impl Into<String>,
        provider: impl Into<String>,
        tier: ModelTier,
        context_capaci;
// pub
fn with_display_name(mut self, name: impl Into<String>) -> Self;

Unit Verification (8 inline tests)​

Tests defined within module: test_hard_filter_eliminates_insufficient_context, test_hard_filter_eliminates_missing_tool_support, test_circuit_breaker_open_rejected_in_stage_1, test_expired_cooldown_readmits_model_in_stage_1, test_role_temperature_follows_profile_registry, test_tier_matching_ranks_preferred_tier_first, test_tie_breaking_is_completely_deterministic, test_fails_closed_when_all_candidates_disqualified

src/model/types.rs​

Primary Responsibility​

Canonical domain types for model communication, proposals, streaming, and error classification. Core principle: The model proposes. The runtime decides. (MDL-01, MDL-03)

Types & Trait Declarations​

  • Structs (4): ModelInvocationOptions, ModelToolCall, TokenUsage, UserOption
  • Enums (7): ChatMessage, CostProvenance, ModelError, ModelProposal, ProviderCapabilityStatus, StreamChunk, UsageSource

Key Functions & Signatures (15 shown)​

// pub
fn is_completion(&self) -> bool;
// pub
fn tool_calls(&self) -> Vec<ModelToolCall>;
// pub
fn text_content(&self) -> Option<&str>;
// pub
fn kind_name(&self) -> &'static str;
// pub
fn tool_names(&self) -> Vec<String>;
// pub
fn tool_calls_count(&self) -> usize;
// pub
fn new(name: impl Into<String>, arguments: serde_json::Value) -> Self;
// pub
fn with_id(id: impl Into<String>,
        name: impl Into<String>,
        arguments: serde_json::Value,) -> Self;
// pub
fn new(prompt_tokens: usize,
        completion_tokens: usize,
        total_tokens: usize,
        reasoning_tokens: us;
// pub
fn accumulate(&mut self, other: &Self);
// pub
fn new() -> Self;
// pub
fn with_max_output_tokens(mut self, max_tokens: usize) -> Self;
// pub
fn with_temperature(mut self, temp: f32) -> Self;
// pub
fn with_top_p(mut self, top_p: f32) -> Self;
// pub
fn as_str(&self) -> &'static str;

Unit Verification (9 inline tests)​

Tests defined within module: test_model_proposal_serialization_tool_calls_single, test_model_proposal_serialization_tool_calls_multiple, test_model_proposal_serialization_assistant_text, test_model_proposal_serialization_ask_user, test_model_proposal_serialization_handoff, test_model_proposal_serialization_complete, test_stream_chunk_serialization, test_token_usage_defaults_and_construction, test_model_error_formatting