Crate tensor_wasm_api 

HTTP serverless API gateway built on axum 0.7.

Exposes deploy/invoke/metrics/healthz endpoints with structured JSON errors. Application state is a DashMap<Uuid, FunctionRecord> shared via Arc<AppState>. The synchronous and async invoke paths both drive tensor_wasm_exec::executor::TensorWasmExecutor.

Security surface

• Body limit. Every request is capped at 64 MiB by axum::extract::DefaultBodyLimit::max; oversized bodies are rejected with 413 Payload Too Large before any handler runs.
• CORS. The gateway installs an explicit-allowlist tower_http::cors::CorsLayer; the allowlist is empty by default (every cross-origin request rejected). Widen by setting TENSOR_WASM_API_CORS_ALLOWED_ORIGINS to a comma-separated list of origins. See middleware::CorsConfig.
• Bearer auth. Reads TENSOR_WASM_API_TOKENS (comma-separated allowlist) at startup. Empty/unset means dev mode (pass-through with warning); otherwise requests must carry Authorization: Bearer <token>.
• Scoped tokens. Each entry in TENSOR_WASM_API_TOKENS may carry a :tenant= scope clause (token:tenant=1,2,3 or token:tenant=*). Routes that bind to a tenant return 403 tenant_scope_denied when the caller’s bearer token is not scoped to that tenant. Bare entries are treated as wildcard with a one-shot deprecation warning at startup (removal targeted for v1.0). See token_scope.
• Tenant scoping. The X-TensorWasm-Tenant: <u64> header is parsed and threaded through to the executor. Set TENSOR_WASM_API_REQUIRE_TENANT=1 to make the header mandatory.
• Per-token rate limiting. Configurable QPS + burst per bearer token, enforced behind bearer auth. Reads TENSOR_WASM_API_RATE_LIMIT_QPS and TENSOR_WASM_API_RATE_LIMIT_BURST; either zero or unset disables the limiter. Rejections return 429 Too Many Requests with a Retry-After header. See rate_limit.
• Audit log. Every state-mutating call (POST /functions, DELETE /functions/{id}, POST /functions/{id}/invoke[-async]) emits a structured JSON record to the sink selected by TENSOR_WASM_API_AUDIT_LOG (default: stdout). Read-only routes emit nothing. See audit and docs/AUDIT-LOG.md.
• XFCC spoofing gate via TENSOR_WASM_API_TRUSTED_XFCC_PROXIES. Comma-separated allowlist of IPv4/IPv6 addresses or CIDR ranges whose X-Forwarded-Client-Cert headers the audit middleware will honour. Empty / unset = trust nobody (safe default — every inbound XFCC is dropped). Deployments behind an mTLS-terminating proxy should set this to the proxy’s IP(s); operators not running an XFCC-stripping proxy should leave it unset. See audit::TrustedProxies.
• Snapshot HMAC key. When TENSOR_WASM_API_SNAPSHOT_HMAC_KEY is set (64-char hex, 32 bytes) the /snapshot/save route HMAC-SHA256-signs the snapshot blob it returns and /snapshot/restore verifies the signature (always requiring a signature on restore). Set TENSOR_WASM_API_SNAPSHOT_REQUIRE_SIGNATURE=true for the matching strict-restore posture. The routes are wired into the protected stack (bearer auth + tenant scope + per-tenant ownership) by build_router, which reads the key via AppConfig::from_env; when the key is unset both routes return 503 snapshot_signing_not_configured. See routes::snapshot_save / routes::snapshot_restore and config for the schema.
• HTTP request metrics. A tower middleware emits tensor_wasm_http_requests_total, tensor_wasm_http_request_duration_seconds, and tensor_wasm_http_requests_in_flight per (route, method, status), labelled with the axum route template (never the substituted id). The layer sits OUTSIDE bearer auth so 401/429 responses are counted too — required by the availability_http SLI in docs/SLO.md. See http_metrics.

See API.md for the wire-format contract.

Re-exports

pub use audit::audit_log_middleware;

pub use audit::AuditAction;

pub use audit::AuditActor;

pub use audit::AuditActorKind;

pub use audit::AuditConfig;

pub use audit::AuditOutcome;

pub use audit::AuditRecord;

pub use audit::AuditResource;

pub use audit::AuditSink;

pub use audit::FileJsonSink;

pub use audit::NoopSink;

pub use audit::StdoutJsonSink;

pub use audit::TokenScopeView;

pub use audit::TrustedProxies;

pub use audit::ENV_AUDIT_LOG;

pub use audit::ENV_TRUSTED_XFCC_PROXIES;

pub use audit::HEADER_XFCC;

pub use config::AppConfig;

pub use config::ConfigError;

pub use config::HexParseReason;

pub use config::ENV_SNAPSHOT_HMAC_KEY;

pub use config::ENV_SNAPSHOT_REQUIRE_SIGNATURE;

pub use config::SNAPSHOT_HMAC_KEY_LEN;

pub use http_metrics::http_metrics_middleware;

pub use http_metrics::HttpMetricsLayerConfig;

pub use http_metrics::RouteAllowList;

pub use http_metrics::DEFAULT_ROUTE_ALLOWLIST;

pub use http_metrics::UNKNOWN_ROUTE;

pub use middleware::normalize_method;

pub use middleware::sanitize_path;

pub use middleware::AuthConfig;

pub use middleware::CorsConfig;

pub use middleware::KernelPublishTokens;

pub use middleware::TenantConfig;

pub use middleware::TrustedHosts;

pub use middleware::ENV_API_TOKENS;

pub use middleware::ENV_CORS_ALLOWED_ORIGINS;

pub use middleware::ENV_KERNEL_PUBLISH_TOKENS;

pub use middleware::ENV_REQUIRE_TENANT;

pub use middleware::ENV_TRUSTED_HOSTS;

pub use middleware::HEADER_TENANT;

pub use middleware::MAX_AUTH_HEADER_BYTES;

pub use middleware::MAX_PATH_LEN;

pub use middleware::MAX_REQUEST_BODY_BYTES;

pub use openai::ChatCompletionsRequest;

pub use openai::ChatMessage;

pub use openai::CompletionsRequest;

pub use openai::OpenAiError;

pub use openai::OpenAiErrorBody;

pub use openai_translator::assemble_chat_prompt;

pub use openai_translator::model_map_from_env;

pub use openai_translator::parse_model_map_env;

pub use openai_translator::translate_chat_completions_request;

pub use openai_translator::translate_completions_request;

pub use openai_translator::ModelMap;

pub use openai_translator::TranslatedRequest;

pub use openai_translator::ENV_OPENAI_MODEL_MAP;

pub use rate_limit::AuthContext;

pub use rate_limit::PerTenantRateLimitConfig;

pub use rate_limit::RateLimitConfig;

pub use rate_limit::RateLimiter;

pub use rate_limit::TokenId;

pub use routes::snapshot_restore;

pub use routes::snapshot_save;

pub use routes::ApiError;

pub use routes::AppState;

pub use routes::FunctionRecord;

pub use routes::JobRecord;

pub use routes::JobStatus;

pub use routes::SnapshotRestoreRequest;

pub use routes::SnapshotRestoreResponse;

pub use routes::SnapshotSaveRequest;

pub use routes::SnapshotSaveResponse;

pub use server::build_router;

pub use server::build_router_with_audit;

pub use server::build_router_with_config;

pub use server::build_router_with_full_config;

pub use server::build_router_with_kernel_publish_tokens;

pub use server::build_router_with_trusted_proxies;

pub use server::serve;

pub use token_scope::parse_token_entry;

pub use token_scope::parse_tokens_env;

pub use token_scope::ParsedTokens;

pub use token_scope::ScopeParseError;

pub use token_scope::TenantScope;

pub use token_scope::TokenScope;

pub use trace_propagation::current_trace_id;

pub use trace_propagation::extract_parent_context;

pub use trace_propagation::inject_trace_id_header;

pub use trace_propagation::install_w3c_propagator;

pub use trace_propagation::HeaderMapExtractor;

pub use trace_propagation::HEADER_TRACE_ID;

Modules

audit

Structured audit log for state-mutating API calls.

config

Top-level application config knobs that don’t fit naturally into any single middleware-scoped config struct.

http_metrics

HTTP request metrics middleware.

kernels

/kernels HTTP endpoints (roadmap feature #3 — server-side path).

middleware

Tower middleware helpers: timeouts, concurrency limits, body limits, authentication, tenant scoping, and tracing spans.

openai

OpenAI-compatible inference gateway shim.

openai_translator

OpenAI request → internal invoke translation (T41, v0.4 wire-up).

rate_limit

Per-token QPS + burst rate limiting (token-bucket).

routes

REST route handlers (deploy, invoke, metrics, healthz).

server

Axum router builder and listener.

token_scope

Per-tenant scoped bearer tokens.

trace_propagation

W3C Trace Context propagation helpers.