diff --git a/CHANGELOG.md b/CHANGELOG.md index 63195a86..dab2e96a 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -6,6 +6,14 @@ adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html). ## [Unreleased] +### Added + +- **NeMo Relay native plugin** — a dynamically loaded integration that runs + libsy's weighted-random and LLM-classifier algorithms in process while + Switchyard owns provider HTTP dispatch, credentials, translation, retries, + and fallback. Managed calls require NeMo Relay 0.7 or newer and do not depend + on `switchyard-server`. + ### Removed - **Latency-aware router** — the `latency_service` route type and its diff --git a/Cargo.lock b/Cargo.lock index 93cdbf64..8d9f415f 100644 --- a/Cargo.lock +++ b/Cargo.lock @@ -86,6 +86,18 @@ dependencies = [ "serde_json", ] +[[package]] +name = "async-channel" +version = "2.5.0" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "924ed96dd52d1b75e9c1a3e6275715fd320f5f9439fb5a4a11fa51f4221158d2" +dependencies = [ + "concurrent-queue", + "event-listener-strategy", + "futures-core", + "pin-project-lite", +] + [[package]] name = "async-stream" version = "0.3.6" @@ -239,6 +251,9 @@ name = "bitflags" version = "2.13.1" source = "registry+https://github.com/rust-lang/crates.io-index" checksum = "b588b76d00fde79687d7646a9b5bdf3cc0f655e0bbd080335a95d7e96f3587da" +dependencies = [ + "serde_core", +] [[package]] name = "bumpalo" @@ -287,6 +302,16 @@ dependencies = [ "rand_core 0.10.1", ] +[[package]] +name = "chrono" +version = "0.4.45" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "1aa79e62e7697b8e29b513a68abacf485adcd1fe8284a4316c5ae868e6633327" +dependencies = [ + "num-traits", + "serde", +] + [[package]] name = "clap" version = "4.6.2" @@ -352,6 +377,15 @@ dependencies = [ "memchr", ] +[[package]] +name = "concurrent-queue" +version = "2.5.0" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "4ca0197aee26d1ae37445ee532fefce43251d24cc7c166799f4d46817f1d3973" +dependencies = [ + "crossbeam-utils", +] + [[package]] name = "core-foundation" version = "0.10.1" @@ -377,6 +411,12 @@ dependencies = [ "libc", ] +[[package]] +name = "crossbeam-utils" +version = "0.8.22" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "61803da095bee82a81bb1a452ecc25d3b2f1416d1897eb86430c6159ef717c17" + [[package]] name = "deadpool" version = "0.12.3" @@ -440,6 +480,26 @@ dependencies = [ "windows-sys 0.61.2", ] +[[package]] +name = "event-listener" +version = "5.4.2" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "5a23add41df1562121a9393cb065eab5146a1242410f23a644851e90cfd669d2" +dependencies = [ + "parking", + "pin-project-lite", +] + +[[package]] +name = "event-listener-strategy" +version = "0.5.4" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "8be9f3dfaaffdae2972880079a491a1a8bb7cbed0b8dd7a347f668b4150a3b93" +dependencies = [ + "event-listener", + "pin-project-lite", +] + [[package]] name = "fastrand" version = "2.5.0" @@ -1051,6 +1111,31 @@ dependencies = [ "windows-sys 0.61.2", ] +[[package]] +name = "nemo-relay-plugin" +version = "0.7.0-rc.5" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "ae9f5d2e2f16f0db4e9b666068fb3c470eb655b6067ac7f5fa4c66ce8051c220" +dependencies = [ + "nemo-relay-types", + "serde", + "serde_json", +] + +[[package]] +name = "nemo-relay-types" +version = "0.7.0-rc.5" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "70b2358889074b00eba89ec42bc8114f3515541f928306ed770ddf995cfa416b" +dependencies = [ + "bitflags", + "chrono", + "serde", + "serde_json", + "typed-builder", + "uuid", +] + [[package]] name = "nu-ansi-term" version = "0.50.3" @@ -1060,6 +1145,15 @@ dependencies = [ "windows-sys 0.61.2", ] +[[package]] +name = "num-traits" +version = "0.2.19" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "071dfc062690e90b734c0b2273ce72ad0ffa95f0c74596bc250dcfd960262841" +dependencies = [ + "autocfg", +] + [[package]] name = "num_cpus" version = "1.17.0" @@ -1172,6 +1266,12 @@ dependencies = [ "tokio", ] +[[package]] +name = "parking" +version = "2.2.1" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "f38d5652c16fde515bb1ecef450ab0f6a219d619a7274976324d5e377f7dceba" + [[package]] name = "parking_lot" version = "0.12.5" @@ -2007,6 +2107,24 @@ dependencies = [ "wiremock", ] +[[package]] +name = "switchyard-nemo-relay-plugin" +version = "0.1.0" +dependencies = [ + "async-channel", + "async-trait", + "futures-util", + "http", + "nemo-relay-plugin", + "serde", + "serde_json", + "switchyard-libsy", + "switchyard-llm-client", + "switchyard-protocol", + "switchyard-translation", + "tokio", +] + [[package]] name = "switchyard-protocol" version = "0.2.0" @@ -2452,6 +2570,26 @@ version = "0.2.5" source = "registry+https://github.com/rust-lang/crates.io-index" checksum = "e421abadd41a4225275504ea4d6566923418b7f05506fbc9c0fe86ba7396114b" +[[package]] +name = "typed-builder" +version = "0.23.2" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "31aa81521b70f94402501d848ccc0ecaa8f93c8eb6999eb9747e72287757ffda" +dependencies = [ + "typed-builder-macro", +] + +[[package]] +name = "typed-builder-macro" +version = "0.23.2" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "076a02dc54dd46795c2e9c8282ed40bcfb1e22747e955de9389a1de28190fb26" +dependencies = [ + "proc-macro2", + "quote", + "syn", +] + [[package]] name = "unicode-ident" version = "1.0.24" @@ -2488,6 +2626,18 @@ version = "0.2.2" source = "registry+https://github.com/rust-lang/crates.io-index" checksum = "06abde3611657adf66d383f00b093d7faecc7fa57071cce2578660c9f1010821" +[[package]] +name = "uuid" +version = "1.18.1" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "2f87b8aa10b915a06587d0dec516c282ff295b475d94abf425d62b57710070a2" +dependencies = [ + "getrandom 0.3.4", + "js-sys", + "serde", + "wasm-bindgen", +] + [[package]] name = "valuable" version = "0.1.1" diff --git a/Cargo.toml b/Cargo.toml index 2325e1ac..0718ca3d 100644 --- a/Cargo.toml +++ b/Cargo.toml @@ -12,6 +12,7 @@ members = [ "crates/switchyard-server", "crates/switchyard-skill-distillation", "crates/switchyard-translation", + "crates/switchyard-nemo-relay-plugin", ] [workspace.package] @@ -23,6 +24,7 @@ repository = "https://github.com/NVIDIA-NeMo/Switchyard" rust-version = "1.96.1" [workspace.dependencies] +async-channel = "2" async-stream = "0.3" async-trait = "0.1" futures = "0.3" diff --git a/README.md b/README.md index 751ad11c..1a420c80 100644 --- a/README.md +++ b/README.md @@ -21,6 +21,7 @@ algorithm you write yourself. - **Protocol Translation**: convert between OpenAI Chat, Anthropic Messages, and OpenAI Responses formats - **Multi-Backend Routing**: random routing, LLM-as-classifier routing, signal-driven stage-router, or your own algorithm - **Operational Metrics**: Prometheus metrics cover requests, errors, latency, tokens, and routing overhead +- **NeMo Relay Plugin**: run random or LLM-classifier routing in Relay while Switchyard owns provider HTTP dispatch ## Quick Start @@ -119,6 +120,7 @@ configured LLM client selects one upstream format. - **[`switchyard-libsy`](crates/libsy/README.md)**: embed routing algorithms in a Rust application - **[`switchyard-protocol`](crates/protocol/README.md)**: provider-neutral request, response, and streaming types - **[`switchyard-translation`](crates/switchyard-translation/README.md)**: request, response, and stream translation +- **[`switchyard-nemo-relay-plugin`](crates/switchyard-nemo-relay-plugin/README.md)**: install Switchyard as a native NeMo Relay plugin ## Community diff --git a/crates/switchyard-nemo-relay-plugin/Cargo.toml b/crates/switchyard-nemo-relay-plugin/Cargo.toml new file mode 100644 index 00000000..4806da86 --- /dev/null +++ b/crates/switchyard-nemo-relay-plugin/Cargo.toml @@ -0,0 +1,32 @@ +# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# SPDX-License-Identifier: Apache-2.0 + +[package] +name = "switchyard-nemo-relay-plugin" +version = "0.1.0" +description = "Switchyard-owned HTTP routing plugin for NeMo Relay" +authors.workspace = true +edition.workspace = true +license.workspace = true +repository.workspace = true +rust-version.workspace = true +publish = false + +[lib] +crate-type = ["cdylib"] + +[dependencies] +async-channel.workspace = true +async-trait.workspace = true +futures-util.workspace = true +http = "1" +# Use the published release candidate during development. This becomes `0.7.0` +# before the integration lands, after the stable crate is published. +nemo-relay-plugin = "=0.7.0-rc.5" +serde.workspace = true +serde_json.workspace = true +switchyard-libsy.workspace = true +switchyard-llm-client.workspace = true +switchyard-protocol.workspace = true +switchyard-translation.workspace = true +tokio.workspace = true diff --git a/crates/switchyard-nemo-relay-plugin/README.md b/crates/switchyard-nemo-relay-plugin/README.md new file mode 100644 index 00000000..53ee58e3 --- /dev/null +++ b/crates/switchyard-nemo-relay-plugin/README.md @@ -0,0 +1,282 @@ + + +# Switchyard NeMo Relay Dynamic Plugin + +This crate builds the external `nvidia.switchyard` native plugin. It embeds +`switchyard-libsy`, drives `Algorithm::run_stream`, and uses +`switchyard-llm-client` for provider HTTP calls. Managed calls use Relay's +public Rust LLM execution-intercept SDK and do not require a targeted provider +continuation from Relay. + +The plugin uses NeMo Relay native API v1. It depends on the small +`nemo-relay-plugin` authoring SDK, not the Relay runtime, and does not start +`switchyard-server`. Managed provider calls do not use Relay's provider +continuation. + +## Ownership boundary + +For a managed LLM call: + +1. Relay invokes the native LLM execution intercept. +2. The plugin decodes the caller JSON through `switchyard-translation`. +3. The plugin drives the configured libsy algorithm with `run_stream` and + records each genuine decision. +4. For every `CallLlm`, the plugin's `switchyard-llm-client` instance translates + the neutral request, applies the selected target's URL and credentials, and + performs the HTTP request. +5. The plugin passes the real buffered response or response stream to + `CallLlmRequest::respond` and continues until `ReturnToAgent`. +6. The plugin encodes the final neutral response into the caller's protocol. + +Relay still owns the outer LLM lifecycle, dynamic-plugin loading, plugin +configuration, and event substrate. Relay's downstream LLM continuation is +used only for calls whose inbound protocol is not managed by this plugin. + +```mermaid +flowchart LR + A["Caller JSON"] --> B["Relay LLM execution intercept"] + B --> C["Switchyard decode"] + C --> D["libsy run_stream"] + D --> E["CallLlm"] + E --> F["switchyard-llm-client"] + F --> G["Provider HTTP endpoint"] + G --> H["Switchyard response or event decode"] + H --> I["CallLlmRequest.respond"] + I --> J["ReturnToAgent"] + J --> K["Switchyard encode"] + K --> A + + U["Unmanaged profile"] -.-> V["Relay v1 continuation"] +``` + +This boundary has two important consequences: + +- Managed provider calls do not traverse Relay middleware registered after the + Switchyard intercept and do not use the host's provider callback. Provider + transport activity is therefore not represented as nested Relay LLM + lifecycle events. Relay records the outer managed call and the plugin emits + Switchyard routing marks; bridging Switchyard transport spans into Relay is + future work. Routing marks are delivered through the public SDK while Relay + is polling the active execution callback or stream, so they remain attached + to the current Relay scope stack. +- Switchyard owns provider URLs, credentials, HTTP retry behavior, and + translation for managed calls. Relay neither validates nor transports those + target details. + +## Native API v1 and the public Rust SDK + +The manifest uses `compat.native_api = "1"`. The implementation registers with +`PluginContext::register_llm_execution_intercept` and +`PluginContext::register_llm_stream_execution_intercept`, receives typed +`LlmRequest`, `LlmNext`, and `LlmStreamNext` values, and returns the SDK's JSON +result or `LlmJsonStream`. The `nemo-relay-plugin` SDK owns the C callback +trampolines, host strings, continuation handles, panic containment, and native +stream lifecycle. Switchyard contains no raw C callback or host-table adapter. + +Switchyard performs libsy and provider I/O on a plugin-owned Tokio runtime. A +buffered SDK callback waits for that executor to finish. A managed streaming +callback returns a pull-based Rust iterator backed by a bounded 32-message +channel; the async producer waits when the consumer is slow. Dropping that +iterator closes the channel and aborts its in-flight routing task. + +The public native API v1 callback shapes do not provide full in-flight caller +cancellation. A buffered callback has no cancellation token, so a provider +request already in progress continues until it responds or reaches the client +timeout after the caller disconnects. Relay can cancel a streaming iterator +between pulls, but its synchronous `Iterator::next` call cannot be interrupted +while it is waiting for the next provider event. Supporting prompt disconnect +propagation would require an asynchronous public SDK callback or stream-polling +contract; the plugin does not bypass the SDK to recover that behavior. + +### Relay runtime capacity + +In Relay 0.7, the safe native LLM callbacks run directly in Relay's asynchronous +middleware future; Relay does not move them to Tokio's separate blocking pool. +Consequently, each active buffered Switchyard call occupies a normal Relay +Tokio worker while it waits for the plugin executor, and a streaming call +occupies a worker while its synchronous `Iterator::next` waits for the next +provider event. Exhausting those workers can delay unrelated Relay work. + +The Relay CLI constructs a Tokio multi-thread runtime without setting an +explicit worker count. Tokio therefore defaults to the number of CPU cores +available to the process and honors its `TOKIO_WORKER_THREADS` environment +variable. Operators can provide additional capacity while using this native API +v1 integration, for example: + +```bash +TOKIO_WORKER_THREADS=32 nemo-relay ... +``` + +This adjusts the normal asynchronous worker pool, not Tokio's blocking pool; +increasing `max_blocking_threads` does not address this integration. There is no +universal recommended value. Size the pool with headroom above the expected +number of concurrent managed calls and Relay's other work, then validate it +under representative provider latency and streaming concurrency. Increasing +the worker count mitigates starvation but does not restore cancellation or make +the synchronous boundary non-blocking. + +Embedded Relay hosts own their Tokio runtime and should configure the same +capacity explicitly: + +```rust +let runtime = tokio::runtime::Builder::new_multi_thread() + .worker_threads(32) + .enable_all() + .build()?; +``` + +The durable resolution is a public Relay SDK callback and stream-polling +contract that can yield while provider I/O is pending and receive caller +cancellation. Until that exists, provision worker capacity and load-test the +intended concurrency rather than relying on thread-pool growth alone. + +Unmanaged profiles call the typed `LlmNext` or `LlmStreamNext` continuation and +return its result directly. Managed streams use the bounded Switchyard channel; +unmanaged streams use Relay's SDK-owned pull iterator without an additional +bridge. None of the managed HTTP or routing operations requires a targeted LLM +continuation or direct access to Relay's C ABI. + +## Supported routers + +This initial plugin supports exactly two libsy algorithms: + +- seeded, weighted `random` routing; and +- capability-based `llm_classifier` routing, where a judge selects the weak or + strong target before the final provider call. + +`stage_router` and response-judging escalation are intentionally deferred. +Unsupported algorithm kinds are rejected instead of being approximated. + +The plugin owns the outer routing retry loop. Each retry starts a fresh libsy +run. Random routing draws again; an algorithm configured with persistent state, +such as classifier session affinity, may intentionally retain its assignment. +Each target's built-in HTTP retry count is set to zero to avoid retrying a +failed target invisibly before reselection. A random target with `weight = 0` +is fallback-only and is not considered by the algorithm. Trusted fallback is +attempted at most once and, for streaming responses, only before the first +caller event is emitted. Outer routing retries use exponential backoff starting +at 250 milliseconds and capped at 2 seconds. They do not currently honor +provider `Retry-After` headers because the client error contract does not expose +that metadata to the routing loop. + +## Translation and stream fidelity + +`switchyard-translation` is the only request, response, and event translation +layer. It decodes caller JSON into Switchyard's neutral protocol, encodes each +selected call for the target protocol, decodes provider results, and encodes +`ReturnToAgent` back to the caller protocol. Relay codecs are not used. + +The streaming contract carries each parsed provider JSON event in a preservation +envelope alongside its normalized `LlmResponseChunk` representation. +Same-protocol routes replay the preserved JSON unchanged, including +provider-specific fields; this preserves parsed events, not raw SSE bytes or +framing. Cross-protocol routes encode only normalized chunks, and the streaming +helpers still do not expose the buffered translation engine's reject-lossy +diagnostics, so unsupported fields may be normalized or omitted. Replacing +normalized stream content or folding a stream into an aggregate drops the +per-event preservation envelope. + +## Configuration + +During release-candidate validation the manifest declares +`compat.native_api = "1"` and Relay `>=0.7.0-rc.5,<1.0`, and the Rust SDK uses +the exact published `0.7.0-rc.5` crate. Before release, move both lower bounds +to stable `0.7.0`. The manifest API value selects Relay's released native +plugin contract; plugin authors use its safe Rust SDK rather than the underlying +C table directly. Rebuild the bundle when changing SDK versions rather than +assuming Rust dynamic-library compatibility from the manifest value alone. + +A Relay project can configure a seeded weighted-random router as follows: + +```toml +version = 1 + +[[plugins.dynamic]] +manifest = "/opt/switchyard-relay-plugin/relay-plugin.toml" + +[plugins.dynamic.config] +version = 2 +priority = 0 +max_retries = 3 + +[plugins.dynamic.config.algorithm] +kind = "random" +seed = 42 + +[plugins.dynamic.config.default_targets] +openai_chat = "fast" + +[plugins.dynamic.config.targets.fast] +model = "provider/model" +protocol = "openai_chat" +endpoint = "/v1/chat/completions" +base_url = "https://provider.example.com" +weight = 1 + +[plugins.dynamic.config.targets.fast.header_env] +authorization = "PROVIDER_AUTHORIZATION" +``` + +Target map keys such as `fast` are stable semantic names visible to libsy. The +target binding is authoritative for the provider model, protocol, endpoint, +base URL, weight, and environment-backed headers. Each `default_targets` key +both enables that inbound protocol and names its trusted fallback. + +`header_env` is the only custom provider-header source. It resolves values in +the plugin process at registration time so literal header values never appear +in configuration. Environment values must not appear in errors, routing marks, +spans, or debug output. The plugin does not inherit caller credentials for +managed calls. Each variable supplies the complete header value, so an +`authorization` value must include its scheme, such as `Bearer`. Literal +`headers` configuration is rejected; non-secret routing or tenancy headers must +also use `header_env`. + +For `kind = "llm_classifier"`, the classifier target must use `openai_chat` or +`openai_responses`; libsy's judge request uses a JSON-schema response format +that cannot be represented losslessly by Anthropic Messages. The remaining +classifier fields map directly to libsy's capability classifier configuration. + +Version-1 service configuration, decision-only execution, and observe-only +mode are rejected. + +## Build and bundle + +The crate is a source/build unit with `publish = false`. Operators install a +binary bundle rather than a Rust crate: + +```bash +cargo build --release -p switchyard-nemo-relay-plugin +python3 crates/switchyard-nemo-relay-plugin/scripts/package_bundle.py \ + --library target/release/libswitchyard_nemo_relay_plugin.so \ + --output dist/switchyard-nemo-relay-plugin-linux-x86_64 +``` + +On macOS the library suffix is `.dylib`; Windows builds use `.dll`. The bundle +builder creates the minimal Relay package: the shared library, a materialized +manifest with Relay's inline SHA-256 integrity digest, and the JSON schema. + +Install the materialized bundle with Relay's normal lifecycle commands: + +```bash +nemo-relay plugins validate /opt/switchyard-relay-plugin/relay-plugin.toml +nemo-relay plugins add /opt/switchyard-relay-plugin/relay-plugin.toml +nemo-relay plugins enable nvidia.switchyard +nemo-relay plugins inspect nvidia.switchyard +``` + +## Validation expectations + +Before release, validate both routers against buffered and streaming OpenAI +Chat, OpenAI Responses, and Anthropic Messages providers. The acceptance suite +must cover same- and supported cross-protocol routes, deterministic weighted +routing, independent runs, classifier weak and strong selections, retry +reselection, exhaustion, exactly-once fallback, stream commitment, empty +streams, late errors, cancellation, credential privacy, and unmanaged +pass-through. + +The tests must also prove that managed target traffic reaches the provider +through `switchyard-llm-client`, never through Relay's provider continuation, +and that no Switchyard service or health endpoint is involved. diff --git a/crates/switchyard-nemo-relay-plugin/config.schema.json b/crates/switchyard-nemo-relay-plugin/config.schema.json new file mode 100644 index 00000000..6305755a --- /dev/null +++ b/crates/switchyard-nemo-relay-plugin/config.schema.json @@ -0,0 +1,119 @@ +{ + "$schema": "https://json-schema.org/draft/2020-12/schema", + "title": "Switchyard NeMo Relay Plugin", + "description": "In-process Switchyard routing with Switchyard-owned provider HTTP dispatch.", + "type": "object", + "additionalProperties": false, + "required": ["version", "algorithm", "targets", "default_targets"], + "properties": { + "version": { + "const": 2, + "description": "Library-only Switchyard configuration version." + }, + "priority": { + "type": "integer", + "default": 0 + }, + "max_retries": { + "type": "integer", + "minimum": 0, + "maximum": 10, + "default": 3, + "description": "Routing retries after the initial libsy run. Every retry starts a fresh run." + }, + "algorithm": { + "description": "Router configuration. This release supports random and capability-based llm_classifier only.", + "oneOf": [ + { + "type": "object", + "additionalProperties": false, + "required": ["kind"], + "properties": { + "kind": { "const": "random" }, + "seed": { "type": ["integer", "null"], "minimum": 0 } + } + }, + { + "type": "object", + "additionalProperties": false, + "required": [ + "kind", + "classifier_target", + "weak_target", + "strong_target", + "base_threshold" + ], + "properties": { + "kind": { "const": "llm_classifier" }, + "classifier_target": { + "type": "string", + "minLength": 1, + "description": "Semantic target name for the judge. The target must use openai_chat or openai_responses because the judge requires a JSON-schema response format." + }, + "weak_target": { "type": "string", "minLength": 1 }, + "strong_target": { "type": "string", "minLength": 1 }, + "base_threshold": { "type": "number", "minimum": 0, "maximum": 1 }, + "min_confidence": { "type": "number", "minimum": 0, "maximum": 1 }, + "capability_elevated_floor": { + "type": ["number", "null"], + "minimum": 0, + "maximum": 1 + }, + "recent_turn_window": { + "type": ["integer", "null"], + "minimum": 0 + }, + "max_output_tokens": { + "type": "integer", + "minimum": 1, + "default": 4096 + }, + "session_affinity": { "type": "boolean", "default": false }, + "message_hash_fallback": { "type": "boolean", "default": false } + } + } + ] + }, + "targets": { + "type": "object", + "minProperties": 1, + "additionalProperties": { + "type": "object", + "additionalProperties": false, + "required": ["model", "protocol", "base_url"], + "properties": { + "model": { "type": "string", "minLength": 1 }, + "protocol": { + "enum": ["openai_chat", "openai_responses", "anthropic_messages"] + }, + "endpoint": { + "type": "string", + "pattern": "^$|^/", + "description": "Optional provider endpoint override. The resolved URL must end in the canonical route for the selected protocol." + }, + "base_url": { + "type": "string", + "pattern": "^https?://" + }, + "weight": { "type": "number", "minimum": 0, "default": 1 }, + "header_env": { + "type": "object", + "description": "Sole custom provider-header source. Maps header names to environment-variable names resolved by the plugin process so literal values are never stored in configuration.", + "additionalProperties": { "type": "string", "minLength": 1 } + } + } + } + }, + "default_targets": { + "type": "object", + "description": "Maps each managed inbound protocol to its trusted fallback target.", + "minProperties": 1, + "additionalProperties": false, + "properties": { + "openai_chat": { "type": "string", "minLength": 1 }, + "openai_responses": { "type": "string", "minLength": 1 }, + "anthropic_messages": { "type": "string", "minLength": 1 } + } + } + } +} diff --git a/crates/switchyard-nemo-relay-plugin/relay-plugin.toml b/crates/switchyard-nemo-relay-plugin/relay-plugin.toml new file mode 100644 index 00000000..0d7a75f1 --- /dev/null +++ b/crates/switchyard-nemo-relay-plugin/relay-plugin.toml @@ -0,0 +1,33 @@ +# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# SPDX-License-Identifier: Apache-2.0 + +manifest_version = 1 + +[plugin] +id = "nvidia.switchyard" +kind = "rust_dynamic" + +[compat] +# Development lower bound for validation against the published Relay RC. Move +# this to `>=0.7.0,<1.0` together with the SDK dependency before release. +relay = ">=0.7.0-rc.5,<1.0" +native_api = "1" + +[defaults] +enabled = false + +[capabilities] +items = ["plugin_native", "config_schema"] + +[config_schema] +path = "config.schema.json" + +[source] +artifact = "" + +[integrity] +sha256 = "sha256:" + +[load] +library = "" +symbol = "nemo_relay_register_plugin" diff --git a/crates/switchyard-nemo-relay-plugin/scripts/package_bundle.py b/crates/switchyard-nemo-relay-plugin/scripts/package_bundle.py new file mode 100644 index 00000000..550f836e --- /dev/null +++ b/crates/switchyard-nemo-relay-plugin/scripts/package_bundle.py @@ -0,0 +1,62 @@ +# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# SPDX-License-Identifier: Apache-2.0 + +"""Materialize the minimal Relay plugin bundle from a compiled cdylib.""" + +from __future__ import annotations + +import argparse +import hashlib +import shutil +from pathlib import Path + +CRATE_ROOT = Path(__file__).resolve().parents[1] + + +def digest(path: Path) -> str: + """Return the lowercase SHA-256 digest for a file.""" + value = hashlib.sha256() + with path.open("rb") as stream: + for chunk in iter(lambda: stream.read(1024 * 1024), b""): + value.update(chunk) + return value.hexdigest() + + +def main() -> None: + """Materialize a Relay-loadable plugin bundle in an empty directory.""" + parser = argparse.ArgumentParser() + parser.add_argument("--library", required=True, type=Path) + parser.add_argument("--output", required=True, type=Path) + args = parser.parse_args() + + library = args.library.resolve() + if not library.is_file(): + parser.error(f"compiled plugin library does not exist: {library}") + + manifest = (CRATE_ROOT / "relay-plugin.toml").read_text(encoding="utf-8") + placeholders = ("", "") + missing = [placeholder for placeholder in placeholders if placeholder not in manifest] + if missing: + parser.error(f"plugin manifest is missing placeholders: {', '.join(missing)}") + + output = args.output.resolve() + if output.exists() and not output.is_dir(): + parser.error(f"bundle output exists and is not a directory: {output}") + if output.is_dir() and any(output.iterdir()): + parser.error(f"bundle output directory must be empty: {output}") + output.mkdir(parents=True, exist_ok=True) + + artifact = output / library.name + shutil.copy2(library, artifact) + shutil.copy2(CRATE_ROOT / "config.schema.json", output / "config.schema.json") + + artifact_digest = digest(artifact) + manifest = manifest.replace("", artifact.name) + manifest = manifest.replace("", artifact_digest) + (output / "relay-plugin.toml").write_text(manifest, encoding="utf-8") + + print(output) + + +if __name__ == "__main__": + main() diff --git a/crates/switchyard-nemo-relay-plugin/src/client.rs b/crates/switchyard-nemo-relay-plugin/src/client.rs new file mode 100644 index 00000000..316a658d --- /dev/null +++ b/crates/switchyard-nemo-relay-plugin/src/client.rs @@ -0,0 +1,165 @@ +// SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. +// SPDX-License-Identifier: Apache-2.0 + +//! Switchyard-owned HTTP clients bound to one semantic routing target. + +use std::collections::BTreeMap; +use std::sync::Arc; + +use async_trait::async_trait; +use switchyard_llm_client::{Backend, HttpBackendConfig, ModelConfig, TranslatingLlmClient}; +use switchyard_protocol::{ + Context, Decision, LlmClientError, Request, Response, RoutedLlmClient, WireFormat, +}; +use switchyard_translation::TranslationEngine; + +use crate::translation; + +/// A provider client bound to one configured Switchyard target. +/// +/// libsy routes with a stable semantic name (for example `fast`). The provider +/// still expects its own model id (for example `meta/llama-3.1-8b-instruct`). +/// Keeping that mapping here prevents an algorithm's semantic labels from +/// leaking into provider requests. +pub(crate) struct TargetClient { + provider_model: String, + target_format: WireFormat, + inner: TranslatingLlmClient, + translation: TranslationEngine, +} + +impl TargetClient { + pub(crate) fn new( + provider_model: String, + target_format: WireFormat, + dispatch_url: String, + headers: BTreeMap, + ) -> Result { + let backend_config = HttpBackendConfig { + // `dispatch_url` is already resolved by configuration. Backend URL + // joining accepts a complete canonical endpoint as well as a base + // URL/prefix. + base_url: dispatch_url, + api_key: None, + extra_headers: headers, + extra_body: BTreeMap::new(), + // Routing retries belong to the plugin: every retry must start a + // fresh libsy run and obtain a fresh decision. + max_retries: 0, + }; + let backend = match target_format { + WireFormat::OpenAiChat => Backend::OpenAiChat(backend_config), + WireFormat::OpenAiResponses => Backend::OpenAiResponses(backend_config), + WireFormat::AnthropicMessages => Backend::Anthropic(backend_config), + }; + let model = ModelConfig::new(provider_model.clone(), backend, None); + let inner = TranslatingLlmClient::new(&[model])?; + Ok(Self { + provider_model, + target_format, + inner, + translation: TranslationEngine::default(), + }) + } + + /// Retargets only the provider-facing transport metadata. + /// + /// Correlation and agent identity remain available to libsy, while inbound + /// HTTP headers are deliberately removed. Provider credentials come solely + /// from this target's `header_env` configuration. + fn prepare_request(&self, mut request: Request) -> Request { + let metadata = request.metadata.get_or_insert_default(); + metadata.wire_format = Some(self.target_format); + metadata.http_headers = None; + request + } +} + +#[async_trait] +impl RoutedLlmClient for TargetClient { + async fn call( + &self, + ctx: Context, + request: Request, + _decision: Arc, + ) -> Result { + let request = self.prepare_request(request); + translation::validate_target_request( + &self.translation, + self.target_format, + &request.llm_request, + ) + .map_err(LlmClientError::RequestEncoding)?; + self.inner + .call_rewrite_model(ctx, request, Some(&self.provider_model)) + .await + } + + fn supports_count_tokens(&self) -> bool { + self.target_format == WireFormat::AnthropicMessages + } + + async fn count_tokens(&self, request: Request) -> Result { + self.inner.count_tokens(self.prepare_request(request)).await + } +} + +#[cfg(test)] +mod tests { + use super::*; + use switchyard_protocol::Metadata; + + fn client(format: WireFormat) -> TargetClient { + TargetClient::new( + "provider/model".into(), + format, + match format { + WireFormat::OpenAiChat => "https://provider.example/v1/chat/completions".into(), + WireFormat::OpenAiResponses => "https://provider.example/v1/responses".into(), + WireFormat::AnthropicMessages => "https://provider.example/v1/messages".into(), + }, + BTreeMap::new(), + ) + .unwrap() + } + + #[test] + fn target_preparation_forces_format_and_removes_inbound_headers() { + let client = client(WireFormat::AnthropicMessages); + let request = Request { + metadata: Some(Metadata { + correlation_id: Some("request-123".into()), + wire_format: Some(WireFormat::OpenAiChat), + http_headers: Some(BTreeMap::from([ + ("authorization".into(), "Bearer caller-secret".into()), + ("x-caller-only".into(), "must-not-forward".into()), + ])), + ..Metadata::default() + }), + ..Request::default() + }; + + let prepared = client.prepare_request(request); + let metadata = prepared.metadata.unwrap(); + assert_eq!(metadata.wire_format, Some(WireFormat::AnthropicMessages)); + assert_eq!(metadata.correlation_id.as_deref(), Some("request-123")); + assert!(metadata.http_headers.is_none()); + } + + #[test] + fn missing_metadata_is_created_for_the_target_format() { + let client = client(WireFormat::OpenAiResponses); + let prepared = client.prepare_request(Request::default()); + assert_eq!( + prepared.metadata.and_then(|metadata| metadata.wire_format), + Some(WireFormat::OpenAiResponses) + ); + } + + #[test] + fn only_anthropic_targets_advertise_count_tokens() { + assert!(client(WireFormat::AnthropicMessages).supports_count_tokens()); + assert!(!client(WireFormat::OpenAiChat).supports_count_tokens()); + assert!(!client(WireFormat::OpenAiResponses).supports_count_tokens()); + } +} diff --git a/crates/switchyard-nemo-relay-plugin/src/config.rs b/crates/switchyard-nemo-relay-plugin/src/config.rs new file mode 100644 index 00000000..fa1ee4ec --- /dev/null +++ b/crates/switchyard-nemo-relay-plugin/src/config.rs @@ -0,0 +1,740 @@ +// SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. +// SPDX-License-Identifier: Apache-2.0 + +use std::collections::{BTreeMap, BTreeSet}; +use std::sync::Arc; + +use http::Uri; +use http::header::{HeaderName, HeaderValue}; +use serde::Deserialize; +use switchyard_libsy::{ + Algorithm, LlmTarget, LlmTargetSet, LlmTaskClassifier, Random, TaskClassifierConfig, +}; +use switchyard_protocol::{RoutedLlmClient, WireFormat}; + +use crate::client::TargetClient; + +pub(crate) fn protocol_from_call(name: &str) -> Option { + match name { + "openai.chat_completions" => Some(WireFormat::OpenAiChat), + "openai.responses" => Some(WireFormat::OpenAiResponses), + "anthropic.messages" => Some(WireFormat::AnthropicMessages), + _ => None, + } +} + +const fn default_endpoint(protocol: WireFormat) -> &'static str { + match protocol { + WireFormat::OpenAiChat => "/v1/chat/completions", + WireFormat::OpenAiResponses => "/v1/responses", + WireFormat::AnthropicMessages => "/v1/messages", + } +} + +#[derive(Deserialize)] +#[serde(deny_unknown_fields)] +struct TargetBinding { + model: String, + protocol: WireFormat, + #[serde(default)] + endpoint: String, + base_url: String, + #[serde(default = "default_weight")] + weight: f64, + #[serde(default)] + header_env: BTreeMap, +} + +impl TargetBinding { + fn dispatch_url(&self) -> String { + let base = self.base_url.trim_end_matches('/'); + let default = default_endpoint(self.protocol); + if self.endpoint.is_empty() && base.ends_with(default) { + return base.to_string(); + } + let endpoint = if self.endpoint.is_empty() { + default + } else { + &self.endpoint + }; + let endpoint = if base.ends_with("/v1") && endpoint.starts_with("/v1/") { + &endpoint[3..] + } else { + endpoint + }; + format!("{base}{endpoint}") + } + + fn validate(&self, name: &str) -> Result<(), String> { + if self.model.trim().is_empty() { + return Err(format!("target {name:?} model must be non-empty")); + } + if !self.endpoint.is_empty() && !self.endpoint.starts_with('/') { + return Err(format!( + "target {name:?} endpoint must be empty or begin with '/'" + )); + } + if !self.weight.is_finite() || self.weight < 0.0 { + return Err(format!( + "target {name:?} weight must be finite and nonnegative" + )); + } + validate_dispatch_url(name, self.protocol, &self.dispatch_url())?; + self.validate_headers(name) + } + + fn validate_headers(&self, target_name: &str) -> Result<(), String> { + let mut normalized = BTreeSet::new(); + for (name, variable) in &self.header_env { + let canonical = validate_header_name(name)?; + if !normalized.insert(canonical) { + return Err(format!( + "target {target_name:?} configures header {name:?} more than once (header names are case-insensitive)" + )); + } + if variable.trim().is_empty() { + return Err(format!( + "environment variable name for target header {name:?} must not be empty" + )); + } + if variable.as_bytes().contains(&b'=') || variable.as_bytes().contains(&b'\0') { + return Err(format!( + "environment variable name for target header {name:?} must not contain '=' or NUL" + )); + } + } + Ok(()) + } + + fn prepare(&self) -> Result { + let mut headers = BTreeMap::new(); + for (name, variable) in &self.header_env { + let value = std::env::var(variable) + .map_err(|_| format!("environment variable {variable:?} is not set"))?; + validate_header(name, &value)?; + headers.insert(name.clone(), value); + } + let dispatch_url = self.dispatch_url(); + let client = TargetClient::new(self.model.clone(), self.protocol, dispatch_url, headers) + .map_err(|error| format!("failed to create target HTTP client: {error}"))?; + Ok(PreparedTargetBinding { + client: Arc::new(client), + }) + } +} + +pub(crate) struct PreparedTargetBinding { + pub(crate) client: Arc, +} + +impl PreparedTargetBinding { + fn as_llm_target(&self, semantic_name: &str) -> LlmTarget { + LlmTarget { + semantic_name: semantic_name.to_string(), + llm_client: Some(self.client.clone()), + } + } +} + +#[derive(Deserialize)] +#[serde(tag = "kind", rename_all = "snake_case", deny_unknown_fields)] +enum AlgorithmConfig { + Random { + #[serde(default)] + seed: Option, + }, + LlmClassifier { + classifier_target: String, + weak_target: String, + strong_target: String, + #[serde(flatten)] + config: TaskClassifierConfig, + }, +} + +#[derive(Deserialize)] +#[serde(deny_unknown_fields)] +pub(crate) struct SwitchyardConfig { + version: u32, + #[serde(default)] + pub(crate) priority: i32, + #[serde(default = "default_max_retries")] + max_retries: u32, + algorithm: AlgorithmConfig, + targets: BTreeMap, + default_targets: BTreeMap, +} + +pub(crate) struct PreparedConfig { + pub(crate) max_retries: u32, + pub(crate) algorithm: Arc, + pub(crate) targets: BTreeMap, + pub(crate) default_targets: BTreeMap, +} + +impl SwitchyardConfig { + pub(crate) fn validate(&self) -> Result<(), String> { + self.validate_structure()?; + self.build_algorithm(None).map(drop) + } + + fn validate_structure(&self) -> Result<(), String> { + if self.version != 2 { + return Err(format!( + "unsupported Switchyard config version {}; version 1 used switchyard-server; migrate to version = 2", + self.version + )); + } + if self.max_retries > 10 { + return Err("max_retries must not exceed 10".into()); + } + if self.targets.is_empty() { + return Err("targets must not be empty".into()); + } + if self.default_targets.is_empty() { + return Err("default_targets must not be empty".into()); + } + for (name, target) in &self.targets { + if name.trim().is_empty() { + return Err("target names must be non-empty".into()); + } + target.validate(name)?; + } + for (protocol, fallback) in &self.default_targets { + let target = self + .targets + .get(fallback) + .ok_or_else(|| format!("default target {fallback:?} is not configured"))?; + if target.protocol != *protocol { + return Err(format!( + "default target {fallback:?} must use protocol {}", + protocol.as_str() + )); + } + } + Ok(()) + } + + pub(crate) fn prepare(self) -> Result { + self.validate_structure()?; + let targets = self + .targets + .iter() + .map(|(name, target)| target.prepare().map(|prepared| (name.clone(), prepared))) + .collect::, _>>()?; + let algorithm = self.build_algorithm(Some(&targets))?; + Ok(PreparedConfig { + max_retries: self.max_retries, + algorithm, + targets, + default_targets: self.default_targets, + }) + } + + fn build_algorithm( + &self, + prepared: Option<&BTreeMap>, + ) -> Result, String> { + let target = |name: &str| { + if !self.targets.contains_key(name) { + return Err(format!("algorithm target {name:?} is not configured")); + } + Ok(match prepared { + Some(targets) => targets + .get(name) + .ok_or_else(|| format!("algorithm target {name:?} was not prepared"))? + .as_llm_target(name), + None => LlmTarget { + semantic_name: name.to_string(), + llm_client: None, + }, + }) + }; + + match &self.algorithm { + AlgorithmConfig::Random { seed } => { + let routable = self + .targets + .iter() + .filter(|(_, binding)| binding.weight > 0.0) + .collect::>(); + if routable.is_empty() { + return Err( + "random routing requires at least one positive target weight".into(), + ); + } + let targets = routable + .iter() + .map(|(name, _)| target(name)) + .collect::, _>>()?; + let weights = routable + .iter() + .map(|(_, binding)| binding.weight) + .collect::>(); + Random::new(LlmTargetSet::new(targets), Some(weights), *seed) + .map(|algorithm| Arc::new(algorithm) as Arc) + .map_err(|error| error.to_string()) + } + AlgorithmConfig::LlmClassifier { + classifier_target, + weak_target, + strong_target, + config, + } => { + let classifier_binding = self.targets.get(classifier_target).ok_or_else(|| { + format!("algorithm target {classifier_target:?} is not configured") + })?; + if classifier_binding.protocol == WireFormat::AnthropicMessages { + return Err(format!( + "classifier target {classifier_target:?} uses anthropic_messages, which cannot encode the required JSON-schema response format without loss; use an openai_chat or openai_responses target" + )); + } + LlmTaskClassifier::new( + target(classifier_target)?, + target(weak_target)?, + target(strong_target)?, + config.clone(), + ) + .map(|algorithm| Arc::new(algorithm) as Arc) + .map_err(|error| error.to_string()) + } + } + } +} + +fn validate_dispatch_url( + target_name: &str, + protocol: WireFormat, + dispatch_url: &str, +) -> Result<(), String> { + let uri = dispatch_url + .parse::() + .map_err(|error| format!("target {target_name:?} has invalid URL: {error}"))?; + if !matches!(uri.scheme_str(), Some("http" | "https")) { + return Err(format!( + "target {target_name:?} base_url must use http or https" + )); + } + let authority = uri + .authority() + .ok_or_else(|| format!("target {target_name:?} URL must include a host"))?; + if authority.host().is_empty() { + return Err(format!("target {target_name:?} URL must include a host")); + } + if authority.as_str().contains('@') { + return Err(format!( + "target {target_name:?} URL must not contain embedded credentials" + )); + } + if uri.query().is_some() { + return Err(format!( + "target {target_name:?} URL query parameters are not supported" + )); + } + + // The current switchyard-llm-client accepts provider base URLs and complete + // canonical endpoints. Reject a custom terminal route to avoid + // allowing Backend::url() to append another provider suffix silently. + let expected_suffix = match protocol { + WireFormat::OpenAiChat => "/chat/completions", + WireFormat::OpenAiResponses => "/responses", + WireFormat::AnthropicMessages => "/v1/messages", + }; + if !uri.path().ends_with(expected_suffix) { + return Err(format!( + "target {target_name:?} endpoint must resolve to a canonical {protocol} route ending in {expected_suffix:?}" + )); + } + Ok(()) +} + +fn validate_header_name(name: &str) -> Result { + let parsed = HeaderName::from_bytes(name.as_bytes()) + .map_err(|error| format!("invalid target header name {name:?}: {error}"))?; + let canonical = parsed.as_str().to_ascii_lowercase(); + if is_forbidden_target_header(&canonical) { + return Err(format!( + "target header {name:?} is controlled by the HTTP transport and cannot be configured" + )); + } + Ok(canonical) +} + +fn validate_header(name: &str, value: &str) -> Result { + let canonical = validate_header_name(name)?; + HeaderValue::from_str(value) + .map_err(|error| format!("invalid target header value for {name:?}: {error}"))?; + Ok(canonical) +} + +fn is_forbidden_target_header(name: &str) -> bool { + matches!( + name, + "connection" + | "content-length" + | "host" + | "keep-alive" + | "proxy-connection" + | "proxy-authenticate" + | "proxy-authorization" + | "te" + | "trailer" + | "transfer-encoding" + | "upgrade" + ) || name.starts_with("x-nemo-relay-internal-") +} + +const fn default_max_retries() -> u32 { + 3 +} + +const fn default_weight() -> f64 { + 1.0 +} + +#[cfg(test)] +mod tests { + use super::*; + use serde_json::json; + + fn binding(protocol: WireFormat, model: &str) -> TargetBinding { + TargetBinding { + model: model.into(), + protocol, + endpoint: String::new(), + base_url: "https://provider.example/v1".into(), + weight: 1.0, + header_env: BTreeMap::new(), + } + } + + fn config() -> SwitchyardConfig { + SwitchyardConfig { + version: 2, + priority: 0, + max_retries: 3, + algorithm: AlgorithmConfig::Random { seed: Some(42) }, + targets: BTreeMap::from([ + ( + "chat".into(), + binding(WireFormat::OpenAiChat, "provider/chat"), + ), + ( + "responses".into(), + binding(WireFormat::OpenAiResponses, "provider/responses"), + ), + ( + "anthropic".into(), + binding(WireFormat::AnthropicMessages, "provider/anthropic"), + ), + ]), + default_targets: BTreeMap::from([ + (WireFormat::OpenAiChat, "chat".into()), + (WireFormat::OpenAiResponses, "responses".into()), + (WireFormat::AnthropicMessages, "anthropic".into()), + ]), + } + } + + #[test] + fn version_two_random_configuration_builds_clients_without_a_service() { + let config = config(); + config.validate().unwrap(); + let prepared = config.prepare().unwrap(); + assert_eq!(prepared.algorithm.name(), "random"); + assert_eq!(prepared.targets.len(), 3); + assert!( + prepared + .targets + .values() + .all(|target| Arc::strong_count(&target.client) >= 2) + ); + } + + #[test] + fn version_one_reports_the_service_to_library_migration() { + let mut config = config(); + config.version = 1; + let error = config.validate().unwrap_err(); + assert!(error.contains("version 1 used switchyard-server")); + assert!(error.contains("version = 2")); + } + + #[test] + fn target_endpoints_must_be_canonical_for_the_current_http_client() { + let mut config = config(); + config.targets.get_mut("chat").unwrap().endpoint = "/custom/chat".into(); + let error = config.validate().unwrap_err(); + assert!(error.contains("ending in \"/chat/completions\"")); + + config.targets.get_mut("chat").unwrap().endpoint = "/custom/chat/completions".into(); + config.validate().unwrap(); + assert_eq!( + config.targets["chat"].dispatch_url(), + "https://provider.example/v1/custom/chat/completions" + ); + } + + #[test] + fn complete_provider_endpoint_is_not_appended_twice() { + let mut config = config(); + let chat = config.targets.get_mut("chat").unwrap(); + chat.base_url = "https://provider.example/v1/chat/completions/".into(); + assert_eq!( + chat.dispatch_url(), + "https://provider.example/v1/chat/completions" + ); + config.validate().unwrap(); + } + + #[test] + fn absolute_urls_cannot_embed_credentials_or_query_parameters() { + let mut config = config(); + config.targets.get_mut("chat").unwrap().base_url = + "https://user:password@provider.example/v1".into(); + assert!( + config + .validate() + .unwrap_err() + .contains("embedded credentials") + ); + + config.targets.get_mut("chat").unwrap().base_url = + "https://provider.example/v1?api-version=1".into(); + assert!(config.validate().unwrap_err().contains("query parameters")); + } + + #[test] + fn transport_owned_and_case_duplicate_environment_headers_are_rejected() { + let mut host_header_config = config(); + let chat = host_header_config.targets.get_mut("chat").unwrap(); + chat.header_env.insert("Host".into(), "TARGET_HOST".into()); + assert!( + host_header_config + .validate() + .unwrap_err() + .contains("HTTP transport") + ); + + let mut duplicate_config = config(); + let chat = duplicate_config.targets.get_mut("chat").unwrap(); + chat.header_env + .insert("X-Tenant".into(), "TARGET_TENANT_A".into()); + chat.header_env + .insert("x-tenant".into(), "TARGET_TENANT_B".into()); + assert!( + duplicate_config + .validate() + .unwrap_err() + .contains("more than once") + ); + } + + #[test] + fn only_canonical_relay_execution_names_resolve_protocols() { + assert_eq!( + protocol_from_call("openai.chat_completions"), + Some(WireFormat::OpenAiChat) + ); + assert_eq!( + protocol_from_call("openai.responses"), + Some(WireFormat::OpenAiResponses) + ); + assert_eq!( + protocol_from_call("anthropic.messages"), + Some(WireFormat::AnthropicMessages) + ); + assert_eq!(protocol_from_call("openai_chat"), None); + } + + #[test] + fn schema_required_contract_fields_do_not_default_during_deserialization() { + let base = json!({ + "version": 2, + "algorithm": {"kind": "random"}, + "targets": { + "chat": { + "model": "provider/chat", + "protocol": "openai_chat", + "base_url": "https://provider.example/v1" + } + }, + "default_targets": {"openai_chat": "chat"} + }); + for field in ["version", "algorithm", "default_targets"] { + let mut value = base.clone(); + value.as_object_mut().unwrap().remove(field); + let error = serde_json::from_value::(value) + .err() + .expect("required field must not default"); + assert!(error.to_string().contains(field), "field={field}: {error}"); + } + } + + #[test] + fn unknown_target_fields_are_rejected() { + let value = json!({ + "version": 2, + "algorithm": {"kind": "random"}, + "targets": { + "chat": { + "model": "provider/chat", + "protocol": "openai_chat", + "base_url": "https://provider.example/v1", + "unexpected_setting": true + } + }, + "default_targets": {"openai_chat": "chat"} + }); + let error = serde_json::from_value::(value) + .err() + .expect("unknown target field must be rejected"); + assert!(error.to_string().contains("unexpected_setting")); + } + + #[test] + fn literal_target_headers_are_rejected() { + let value = json!({ + "version": 2, + "algorithm": {"kind": "random"}, + "targets": { + "chat": { + "model": "provider/chat", + "protocol": "openai_chat", + "base_url": "https://provider.example/v1", + "headers": {"x-provider-token": "plaintext-secret"} + } + }, + "default_targets": {"openai_chat": "chat"} + }); + let error = serde_json::from_value::(value) + .err() + .expect("literal target headers must be rejected") + .to_string(); + assert!(error.contains("unknown field `headers`")); + assert!(!error.contains("plaintext-secret")); + } + + #[test] + fn unknown_algorithm_fields_are_rejected() { + let error = serde_json::from_value::(json!({ + "kind": "random", + "seed": 42, + "unexpected_setting": true + })) + .err() + .expect("unknown algorithm field must be rejected"); + assert!(error.to_string().contains("unexpected_setting")); + } + + #[test] + fn classifier_attaches_clients_to_judge_and_routed_targets() { + let mut config = config(); + config.algorithm = serde_json::from_value(json!({ + "kind": "llm_classifier", + "classifier_target": "chat", + "weak_target": "responses", + "strong_target": "anthropic", + "base_threshold": 0.5, + "recent_turn_window": 4, + "max_output_tokens": 512 + })) + .unwrap(); + config.validate().unwrap(); + let prepared = config.prepare().unwrap(); + assert_eq!(prepared.algorithm.name(), "llm_task_classifier"); + assert!( + prepared + .targets + .values() + .all(|target| Arc::strong_count(&target.client) >= 2) + ); + } + + #[test] + fn classifier_rejects_anthropic_judge_targets_before_dispatch() { + let mut config = config(); + config.algorithm = AlgorithmConfig::LlmClassifier { + classifier_target: "anthropic".into(), + weak_target: "responses".into(), + strong_target: "chat".into(), + config: TaskClassifierConfig { + base_threshold: 0.5, + ..Default::default() + }, + }; + + let error = config.validate().unwrap_err(); + assert!(error.contains("classifier target \"anthropic\" uses anthropic_messages")); + } + + #[test] + fn validation_does_not_resolve_environment_backed_headers() { + let mut config = config(); + config.targets.get_mut("chat").unwrap().header_env = BTreeMap::from([( + "authorization".into(), + "SWITCHYARD_TEST_ENVIRONMENT_VARIABLE_THAT_IS_NOT_SET".into(), + )]); + + config.validate().unwrap(); + let error = config + .prepare() + .err() + .expect("preparation must resolve headers"); + assert!(error.contains("SWITCHYARD_TEST_ENVIRONMENT_VARIABLE_THAT_IS_NOT_SET")); + } + + #[test] + fn invalid_environment_variable_names_are_rejected_before_resolution() { + for variable in ["INVALID=VARIABLE", "INVALID\0VARIABLE"] { + let mut config = config(); + config.targets.get_mut("chat").unwrap().header_env = + BTreeMap::from([("authorization".into(), variable.into())]); + + let error = config.validate().unwrap_err(); + assert!(error.contains("must not contain '=' or NUL")); + } + } + + #[test] + fn static_validation_preserves_algorithm_constructor_checks() { + let mut random = config(); + for target in random.targets.values_mut() { + target.weight = 0.0; + } + assert!( + random + .validate() + .unwrap_err() + .contains("at least one positive target weight") + ); + + let mut classifier = config(); + classifier.algorithm = AlgorithmConfig::LlmClassifier { + classifier_target: "chat".into(), + weak_target: "responses".into(), + strong_target: "anthropic".into(), + config: TaskClassifierConfig { + base_threshold: 1.1, + ..Default::default() + }, + }; + assert!( + classifier + .validate() + .unwrap_err() + .contains("base_threshold must be between 0 and 1") + ); + } + + #[test] + fn zero_weight_random_targets_are_fallback_only() { + let mut config = config(); + config.targets.get_mut("anthropic").unwrap().weight = 0.0; + let prepared = config.prepare().unwrap(); + + assert_eq!(Arc::strong_count(&prepared.targets["anthropic"].client), 1); + assert!(Arc::strong_count(&prepared.targets["chat"].client) >= 2); + assert!(Arc::strong_count(&prepared.targets["responses"].client) >= 2); + } +} diff --git a/crates/switchyard-nemo-relay-plugin/src/executor.rs b/crates/switchyard-nemo-relay-plugin/src/executor.rs new file mode 100644 index 00000000..38b4d0f4 --- /dev/null +++ b/crates/switchyard-nemo-relay-plugin/src/executor.rs @@ -0,0 +1,152 @@ +// SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. +// SPDX-License-Identifier: Apache-2.0 + +use std::future::Future; +use std::sync::{Arc, Mutex, mpsc}; +use std::thread::{self, JoinHandle}; + +use tokio::runtime::{Builder, Handle}; +use tokio::sync::oneshot; +use tokio::task::AbortHandle; + +/// Plugin-owned async executor. +/// +/// The public native-plugin SDK uses synchronous Rust callbacks and pull-based +/// iterators at the dynamic-library boundary. Switchyard performs provider I/O +/// on this dedicated runtime rather than entering Relay's Tokio runtime from a +/// separately linked cdylib. +#[derive(Clone)] +pub(crate) struct PluginExecutor { + inner: Arc, +} + +struct ExecutorInner { + handle: Handle, + shutdown: Mutex>>, + thread: Mutex>>, +} + +impl PluginExecutor { + pub(crate) fn new() -> Result { + let (ready_tx, ready_rx) = mpsc::sync_channel(1); + let thread = thread::Builder::new() + .name("switchyard-relay-http".into()) + .spawn(move || { + let runtime = match Builder::new_multi_thread() + .worker_threads(2) + .thread_name("switchyard-relay-http-worker") + .enable_all() + .build() + { + Ok(runtime) => runtime, + Err(error) => { + let _ = ready_tx.send(Err(error.to_string())); + return; + } + }; + let (shutdown_tx, shutdown_rx) = oneshot::channel(); + if ready_tx + .send(Ok((runtime.handle().clone(), shutdown_tx))) + .is_err() + { + return; + } + runtime.block_on(async { + let _ = shutdown_rx.await; + }); + }) + .map_err(|error| format!("failed to start Switchyard HTTP runtime: {error}"))?; + let (handle, shutdown) = ready_rx + .recv() + .map_err(|_| "Switchyard HTTP runtime stopped during startup".to_string())??; + Ok(Self { + inner: Arc::new(ExecutorInner { + handle, + shutdown: Mutex::new(Some(shutdown)), + thread: Mutex::new(Some(thread)), + }), + }) + } + + pub(crate) fn spawn(&self, future: F) -> AbortHandle + where + F: Future + Send + 'static, + { + self.inner.handle.spawn(future).abort_handle() + } + + pub(crate) fn run(&self, future: F) -> Result + where + F: Future + Send + 'static, + F::Output: Send + 'static, + { + let (sender, receiver) = mpsc::sync_channel(1); + self.inner.handle.spawn(async move { + let _ = sender.send(future.await); + }); + receiver + .recv() + .map_err(|_| "Switchyard HTTP runtime stopped before completing work".to_string()) + } +} + +impl Drop for ExecutorInner { + fn drop(&mut self) { + if let Some(shutdown) = self + .shutdown + .lock() + .unwrap_or_else(std::sync::PoisonError::into_inner) + .take() + { + let _ = shutdown.send(()); + } + if let Some(thread) = self + .thread + .lock() + .unwrap_or_else(std::sync::PoisonError::into_inner) + .take() + { + if std::thread::current() + .name() + .is_some_and(|name| name.starts_with("switchyard-relay-http-worker")) + { + // The runtime owner will join this worker after the current + // task returns. Waiting here would deadlock that shutdown. + drop(thread); + } else { + let _ = thread.join(); + } + } + } +} + +#[cfg(test)] +mod tests { + use super::*; + + #[test] + fn executor_runs_buffered_and_spawned_work() { + let executor = PluginExecutor::new().unwrap(); + assert_eq!(executor.run(async { 42 }).unwrap(), 42); + let (sender, receiver) = mpsc::sync_channel(1); + executor.spawn(async move { + sender.send("done").unwrap(); + }); + assert_eq!(receiver.recv().unwrap(), "done"); + } + + #[test] + fn last_reference_can_drop_on_a_worker() { + let executor = PluginExecutor::new().unwrap(); + let worker_reference = executor.clone(); + let (sender, receiver) = mpsc::sync_channel(1); + executor.spawn(async move { + drop(worker_reference); + sender.send(()).unwrap(); + }); + drop(executor); + receiver + .recv_timeout(std::time::Duration::from_secs(5)) + .expect("dropping the executor on its own worker must not deadlock"); + } +} diff --git a/crates/switchyard-nemo-relay-plugin/src/lib.rs b/crates/switchyard-nemo-relay-plugin/src/lib.rs new file mode 100644 index 00000000..85fd634a --- /dev/null +++ b/crates/switchyard-nemo-relay-plugin/src/lib.rs @@ -0,0 +1,246 @@ +// SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. +// SPDX-License-Identifier: Apache-2.0 + +mod client; +mod config; +mod executor; +mod runtime; +mod translation; + +use std::panic::AssertUnwindSafe; +use std::sync::Arc; + +use futures_util::FutureExt; +use nemo_relay_plugin::{ + ConfigDiagnostic, DiagnosticLevel, Json, LlmJsonStream, LlmNext, LlmRequest, LlmStreamNext, + NativePlugin, PluginContext, PluginRuntime, +}; +use serde_json::Map; +use tokio::task::AbortHandle; + +use crate::config::SwitchyardConfig; +use crate::executor::PluginExecutor; +use crate::runtime::{RoutingMark, StreamMessage, SwitchyardRuntime}; + +struct CallbackState { + runtime: Arc, + executor: PluginExecutor, + relay: PluginRuntime, +} + +#[derive(Default)] +struct SwitchyardPlugin; + +impl NativePlugin for SwitchyardPlugin { + fn plugin_kind(&self) -> &str { + "nvidia.switchyard" + } + + fn allows_multiple_components(&self) -> bool { + false + } + + fn validate(&self, plugin_config: &Map) -> Vec { + match parse_config(plugin_config).and_then(|config| config.validate()) { + Ok(()) => Vec::new(), + Err(message) => vec![ConfigDiagnostic { + level: DiagnosticLevel::Error, + code: "switchyard.invalid_config".into(), + component: Some("nvidia.switchyard".into()), + field: Some("config".into()), + message, + }], + } + } + + fn register( + &mut self, + plugin_config: &Map, + ctx: &mut PluginContext<'_>, + ) -> nemo_relay_plugin::Result<()> { + let config = parse_config(plugin_config)?; + let priority = config.priority; + let state = Arc::new(CallbackState { + runtime: Arc::new(SwitchyardRuntime::new(config)?), + executor: PluginExecutor::new()?, + relay: ctx.runtime(), + }); + + let buffered = Arc::clone(&state); + ctx.register_llm_execution_intercept( + "switchyard.run_stream.buffered", + priority, + move |name, request, next| execute_buffered(&buffered, name, request, next), + )?; + ctx.register_llm_stream_execution_intercept( + "switchyard.run_stream.streaming", + priority, + move |name, request, next| execute_stream(&state, name, request, next), + )?; + Ok(()) + } +} + +fn parse_config(plugin_config: &Map) -> Result { + match plugin_config.get("version").and_then(Json::as_u64) { + Some(2) => {} + Some(version) => { + return Err(format!( + "unsupported Switchyard config version {version}; version 1 used switchyard-server; migrate to version = 2" + )); + } + None => { + return Err("invalid Switchyard configuration: version must be the integer 2".into()); + } + } + serde_json::from_value(Json::Object(plugin_config.clone())) + .map_err(|error| format!("invalid Switchyard configuration: {error}")) +} + +fn execute_buffered( + state: &CallbackState, + name: &str, + request: LlmRequest, + next: LlmNext<'_>, +) -> nemo_relay_plugin::Result { + let Some(inbound) = state.runtime.managed_protocol(name) else { + return next.call(request); + }; + let request = state.runtime.decode_request(inbound, &request, false)?; + let runtime = Arc::clone(&state.runtime); + let (result, marks) = state.executor.run(async move { + let mut marks = Vec::new(); + let result = AssertUnwindSafe(runtime.execute_buffered(inbound, request, &mut marks)) + .catch_unwind() + .await + .unwrap_or_else(|_| Err("Switchyard buffered execution panicked".into())); + (result, marks) + })?; + emit_marks(&state.relay, marks); + result +} + +fn execute_stream( + state: &CallbackState, + name: &str, + request: LlmRequest, + next: LlmStreamNext<'_>, +) -> nemo_relay_plugin::Result { + let Some(inbound) = state.runtime.managed_protocol(name) else { + return Ok(Box::new(next.call(request)?)); + }; + let request = state.runtime.decode_request(inbound, &request, true)?; + let (sender, receiver) = async_channel::bounded(32); + let runtime = Arc::clone(&state.runtime); + let task = state.executor.spawn(async move { + let result = AssertUnwindSafe(runtime.execute_stream(inbound, request, &sender)) + .catch_unwind() + .await + .unwrap_or_else(|_| Err("Switchyard streaming execution panicked".into())); + if let Err(error) = result { + let _ = sender.send(StreamMessage::Error(error)).await; + } + }); + Ok(Box::new(SwitchyardStream { + receiver, + task: Some(task), + relay: state.relay.clone(), + finished: false, + })) +} + +fn emit_marks(relay: &PluginRuntime, marks: Vec) { + for mark in marks { + emit_mark(relay, mark); + } +} + +fn emit_mark(relay: &PluginRuntime, mark: RoutingMark) { + if let Err(error) = relay.emit_mark(&mark.name, Some(&mark.data), Some(&mark.metadata)) { + eprintln!( + "Switchyard could not emit routing mark {:?}: {error}", + mark.name + ); + } +} + +struct SwitchyardStream { + receiver: async_channel::Receiver, + task: Option, + relay: PluginRuntime, + finished: bool, +} + +impl Iterator for SwitchyardStream { + type Item = nemo_relay_plugin::Result; + + fn next(&mut self) -> Option { + loop { + match self.receiver.recv_blocking() { + Ok(StreamMessage::Mark(mark)) => emit_mark(&self.relay, mark), + Ok(StreamMessage::Event(event)) => return Some(Ok(event)), + Ok(StreamMessage::Error(error)) => { + self.finished = true; + self.task.take(); + return Some(Err(error)); + } + Err(_) => { + self.finished = true; + self.task.take(); + return None; + } + } + } + } +} + +impl Drop for SwitchyardStream { + fn drop(&mut self) { + self.receiver.close(); + if !self.finished + && let Some(task) = self.task.take() + { + task.abort(); + } + } +} + +nemo_relay_plugin::nemo_relay_plugin!(nemo_relay_register_plugin, SwitchyardPlugin::default); + +#[cfg(test)] +mod tests { + use serde_json::json; + + use super::*; + + #[test] + fn version_one_service_config_gets_a_migration_error_before_v2_deserialization() { + let value = json!({ + "version": 1, + "service_url": "http://127.0.0.1:8080", + "health_endpoint": "/healthz" + }); + let plugin_config = value.as_object().unwrap(); + + let error = parse_config(plugin_config) + .err() + .expect("version one must be rejected"); + assert!(error.contains("version 1 used switchyard-server")); + assert!(error.contains("migrate to version = 2")); + assert!(!error.contains("unknown field")); + } + + #[test] + fn version_must_be_an_integer() { + let value = json!({"version": "2"}); + let plugin_config = value.as_object().unwrap(); + + let error = parse_config(plugin_config) + .err() + .expect("non-integer versions must be rejected"); + assert_eq!( + error, + "invalid Switchyard configuration: version must be the integer 2" + ); + } +} diff --git a/crates/switchyard-nemo-relay-plugin/src/runtime.rs b/crates/switchyard-nemo-relay-plugin/src/runtime.rs new file mode 100644 index 00000000..8898ef6d --- /dev/null +++ b/crates/switchyard-nemo-relay-plugin/src/runtime.rs @@ -0,0 +1,941 @@ +// SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. +// SPDX-License-Identifier: Apache-2.0 + +use std::collections::BTreeMap; +use std::sync::Arc; +use std::time::Duration; + +use futures_util::{StreamExt, stream}; +use nemo_relay_plugin::{Json, LlmRequest as RelayRequest}; +use serde_json::{Map, json}; +use switchyard_libsy::{Algorithm, CallLlmRequest, LibsyError, Step}; +use switchyard_protocol::{ + Context, Decision, LlmClientError, LlmResponse, Metadata, Request, Response, SimpleDecision, + WireFormat, +}; +use switchyard_translation::{TranslationEngine, encode_stream}; + +use crate::config::{PreparedTargetBinding, SwitchyardConfig, protocol_from_call}; +use crate::translation; + +const INITIAL_RETRY_BACKOFF: Duration = Duration::from_millis(250); +const MAX_RETRY_BACKOFF: Duration = Duration::from_secs(2); + +#[derive(Debug)] +pub(crate) struct RoutingMark { + pub(crate) name: String, + pub(crate) data: Json, + pub(crate) metadata: Json, +} + +#[derive(Debug)] +pub(crate) enum StreamMessage { + Mark(RoutingMark), + Event(Json), + Error(String), +} + +pub(crate) struct SwitchyardRuntime { + max_retries: u32, + algorithm: Arc, + targets: BTreeMap, + default_targets: BTreeMap, + translation: TranslationEngine, +} + +impl SwitchyardRuntime { + pub(crate) fn new(config: SwitchyardConfig) -> Result { + let prepared = config.prepare()?; + Ok(Self { + max_retries: prepared.max_retries, + algorithm: prepared.algorithm, + targets: prepared.targets, + default_targets: prepared.default_targets, + translation: TranslationEngine::default(), + }) + } + + pub(crate) fn managed_protocol(&self, name: &str) -> Option { + protocol_from_call(name).filter(|protocol| self.default_targets.contains_key(protocol)) + } + + pub(crate) fn decode_request( + &self, + inbound: WireFormat, + request: &RelayRequest, + streaming: bool, + ) -> Result { + let mut llm_request = translation::decode_request(&self.translation, inbound, request)?; + llm_request.stream = streaming; + let headers = string_headers(&request.headers); + let mut metadata = Metadata::from_headers(&headers); + // Keep identity/routing metadata, but target clients deliberately clear + // these caller headers before HTTP dispatch. + metadata.http_headers = Some(headers); + metadata.wire_format = Some(inbound); + Ok(Request { + llm_request, + raw_request: Some(request.content.clone()), + metadata: Some(metadata), + }) + } + + pub(crate) async fn execute_buffered( + &self, + inbound: WireFormat, + request: Request, + marks: &mut Vec, + ) -> Result { + let metadata = identity_metadata(request.metadata.as_ref()); + let max_attempts = self.max_retries + 1; + let mut attempt = 1; + loop { + self.mark( + marks, + "switchyard.routing.requested", + json!({"algorithm": self.algorithm.name(), "attempt": attempt}), + &metadata, + ); + let result = self + .drive(request.clone(), attempt, marks, &metadata) + .await + .and_then(|response| { + finalize_buffered_response(&self.translation, inbound, response) + .map_err(|source| LibsyError::client_call("return_to_agent", source)) + }); + match result { + Ok(response) => return Ok(response), + Err(failure) if libsy_error_retryable(&failure) && attempt < max_attempts => { + self.mark( + marks, + "switchyard.routing.retry", + failure_mark_data(attempt, &failure), + &metadata, + ); + sleep_before_retry(attempt).await; + attempt += 1; + } + Err(failure) => { + self.mark( + marks, + "switchyard.routing.error", + failure_mark_data(attempt, &failure), + &metadata, + ); + let response = self + .fallback_response(inbound, request, marks, &metadata) + .await?; + return finalize_buffered_response(&self.translation, inbound, response) + .map_err(|error| { + public_response_failure("trusted fallback response", &error) + }); + } + } + } + } + + pub(crate) async fn execute_stream( + &self, + inbound: WireFormat, + request: Request, + output: &async_channel::Sender, + ) -> Result<(), String> { + let metadata = identity_metadata(request.metadata.as_ref()); + let max_attempts = self.max_retries + 1; + let mut attempt = 1; + let mut marks = Vec::new(); + 'attempts: loop { + self.mark( + &mut marks, + "switchyard.routing.requested", + json!({"algorithm": self.algorithm.name(), "attempt": attempt}), + &metadata, + ); + let (response, mut fallback_used) = match self + .drive(request.clone(), attempt, &mut marks, &metadata) + .await + { + Ok(response) => (response, false), + Err(failure) if libsy_error_retryable(&failure) && attempt < max_attempts => { + self.mark( + &mut marks, + "switchyard.routing.retry", + failure_mark_data(attempt, &failure), + &metadata, + ); + send_marks(output, &mut marks).await?; + sleep_before_retry(attempt).await; + attempt += 1; + continue; + } + Err(failure) => { + self.mark( + &mut marks, + "switchyard.routing.error", + failure_mark_data(attempt, &failure), + &metadata, + ); + let fallback = self + .fallback_response(inbound, request.clone(), &mut marks, &metadata) + .await; + send_marks(output, &mut marks).await?; + (fallback?, true) + } + }; + send_marks(output, &mut marks).await?; + + let mut events = match returned_events(response, inbound).await { + Ok(events) => events, + Err(failure) + if !fallback_used + && libsy_error_retryable(&failure) + && attempt < max_attempts => + { + self.mark( + &mut marks, + "switchyard.routing.retry", + failure_mark_data(attempt, &failure), + &metadata, + ); + send_marks(output, &mut marks).await?; + sleep_before_retry(attempt).await; + attempt += 1; + continue; + } + Err(failure) if !fallback_used => { + self.mark( + &mut marks, + "switchyard.routing.error", + failure_mark_data(attempt, &failure), + &metadata, + ); + fallback_used = true; + let fallback = self + .fallback_response(inbound, request.clone(), &mut marks, &metadata) + .await; + send_marks(output, &mut marks).await?; + let fallback = fallback?; + returned_events(fallback, inbound) + .await + .map_err(|error| public_libsy_failure("trusted fallback stream", &error))? + } + Err(failure) => { + return Err(public_libsy_failure("trusted fallback stream", &failure)); + } + }; + + let mut committed = false; + while let Some(item) = events.next().await { + match item { + Ok(event) => { + send_event(output, event).await?; + committed = true; + } + Err(failure) + if !fallback_used + && !committed + && libsy_error_retryable(&failure) + && attempt < max_attempts => + { + self.mark( + &mut marks, + "switchyard.routing.retry", + failure_mark_data(attempt, &failure), + &metadata, + ); + send_marks(output, &mut marks).await?; + sleep_before_retry(attempt).await; + attempt += 1; + continue 'attempts; + } + Err(failure) if !fallback_used && !committed => { + self.mark( + &mut marks, + "switchyard.routing.error", + failure_mark_data(attempt, &failure), + &metadata, + ); + let fallback = self + .fallback_response(inbound, request.clone(), &mut marks, &metadata) + .await; + send_marks(output, &mut marks).await?; + let fallback = fallback?; + let mut fallback = + returned_events(fallback, inbound).await.map_err(|error| { + public_libsy_failure("trusted fallback stream", &error) + })?; + while let Some(item) = fallback.next().await { + let event = item.map_err(|error| { + public_libsy_failure("trusted fallback stream", &error) + })?; + send_event(output, event).await?; + } + return Ok(()); + } + Err(failure) if !committed => { + return Err(public_libsy_failure("trusted fallback stream", &failure)); + } + Err(failure) => { + self.mark( + &mut marks, + "switchyard.routing.error", + failure_mark_data(attempt, &failure), + &metadata, + ); + send_marks(output, &mut marks).await?; + return Err(public_libsy_failure( + "Switchyard stream failed after response commitment", + &failure, + )); + } + } + } + if committed { + return Ok(()); + } + return Err("Switchyard response stream produced no caller events".into()); + } + } + + async fn drive( + &self, + request: Request, + attempt: u32, + marks: &mut Vec, + mark_metadata: &Json, + ) -> Result { + let context = context_from_metadata(request.metadata.as_ref()); + let mut steps = self.algorithm.clone().run_stream(context, request, None); + while let Some(step) = steps.next().await { + match step { + Ok(Step::Decision(decision)) => { + self.emit_decision(marks, decision.as_ref(), attempt, mark_metadata); + } + Ok(Step::CallLlm(call)) => self.serve_call(*call).await?, + Ok(Step::ReturnToAgent(response)) => return Ok(*response), + Err(error) => return Err(error), + } + } + Err(LibsyError::MissingFinalResponse) + } + + async fn serve_call(&self, call: CallLlmRequest) -> switchyard_libsy::Result<()> { + let routed = call.get_routed().clone(); + let target_name = routed.decision.selected_model().to_string(); + let result = match routed.default_client { + Some(client) => client + .call(routed.ctx, routed.request, routed.decision) + .await + .map_err(|source| LibsyError::client_call(target_name, source)), + None => Err(LibsyError::client_call( + target_name, + LlmClientError::Configuration { + message: "libsy CallLlm step has no Switchyard HTTP client".into(), + }, + )), + }; + call.respond(result) + } + + async fn fallback_response( + &self, + inbound: WireFormat, + request: Request, + marks: &mut Vec, + metadata: &Json, + ) -> Result { + let target_name = self.default_target(inbound)?; + let target = self.target(target_name)?; + self.mark( + marks, + "switchyard.routing.fallback", + json!({"selected_target": target_name}), + metadata, + ); + let decision: Arc = Arc::new(SimpleDecision { + selected_model: target_name.to_string(), + reasoning: Some("trusted fallback target".into()), + }); + let context = context_from_metadata(request.metadata.as_ref()); + target + .client + .call(context, request, decision) + .await + .map_err(|error| public_client_failure("trusted fallback", &error)) + } + + fn target(&self, name: &str) -> Result<&PreparedTargetBinding, String> { + self.targets + .get(name) + .ok_or_else(|| format!("libsy selected unknown target {name:?}")) + } + + fn default_target(&self, protocol: WireFormat) -> Result<&str, String> { + self.default_targets + .get(&protocol) + .map(String::as_str) + .ok_or_else(|| format!("managed protocol {protocol} has no default target")) + } + + fn mark(&self, marks: &mut Vec, name: &str, data: Json, metadata: &Json) { + marks.push(RoutingMark { + name: name.to_string(), + data, + metadata: metadata.clone(), + }); + } + + fn emit_decision( + &self, + marks: &mut Vec, + decision: &dyn Decision, + attempt: u32, + metadata: &Json, + ) { + self.mark( + marks, + "switchyard.routing.decision", + json!({ + "algorithm": self.algorithm.name(), + "attempt": attempt, + "selected_target": decision.selected_model(), + "reasoning": decision.reasoning(), + "routing_tier": decision.routing_tier(), + "is_routed_call": decision.is_routed_call(), + }), + metadata, + ); + } +} + +async fn send_marks( + output: &async_channel::Sender, + marks: &mut Vec, +) -> Result<(), String> { + for mark in marks.drain(..) { + output + .send(StreamMessage::Mark(mark)) + .await + .map_err(|_| "Relay cancelled the Switchyard response stream".to_string())?; + } + Ok(()) +} + +async fn send_event( + output: &async_channel::Sender, + event: Json, +) -> Result<(), String> { + output + .send(StreamMessage::Event(event)) + .await + .map_err(|_| "Relay cancelled the Switchyard response stream".to_string()) +} + +type ReturnedEventStream = + std::pin::Pin> + Send>>; + +fn finalize_buffered_response( + translation_engine: &TranslationEngine, + inbound: WireFormat, + response: Response, +) -> Result { + let LlmResponse::Agg(response) = response.llm_response else { + return Err(LlmClientError::InvalidResponse { + source: Box::new(std::io::Error::other( + "libsy returned a stream for a buffered request", + )), + }); + }; + translation::encode_response(translation_engine, inbound, &response) + .map_err(LlmClientError::ResponseTranslation) +} + +async fn returned_events( + response: Response, + inbound: WireFormat, +) -> Result { + let chunks = match response.llm_response { + LlmResponse::Agg(response) => response.into_stream(), + LlmResponse::Stream(mut chunks) => { + let Some(first) = chunks.next().await else { + return Err(LibsyError::client_call( + "return_to_agent", + LlmClientError::InvalidResponse { + source: Box::new(std::io::Error::new( + std::io::ErrorKind::UnexpectedEof, + "provider returned an empty stream", + )), + }, + )); + }; + Box::pin(stream::once(async move { first }).chain(chunks)) + } + }; + let events = encode_stream(chunks, inbound, None) + .map_err(|error| LibsyError::client_call("return_to_agent", error))?; + Ok(Box::pin(events.map(|item| { + item.map_err(|source| match source.downcast::() { + Ok(source) => LibsyError::client_call("return_to_agent", *source), + Err(source) => LibsyError::client_call( + "return_to_agent", + LlmClientError::ResponseTranslation(source.to_string()), + ), + }) + }))) +} + +fn libsy_error_retryable(error: &LibsyError) -> bool { + let LibsyError::ClientCall { source, .. } = error else { + return false; + }; + match source { + LlmClientError::UpstreamHttp { status, .. } => { + matches!(*status, 408 | 425 | 429 | 500 | 502 | 503 | 504) + } + LlmClientError::Transport { .. } | LlmClientError::Timeout { .. } => true, + _ => false, + } +} + +fn retry_backoff(attempt: u32) -> Duration { + let exponent = attempt.saturating_sub(1).min(3); + INITIAL_RETRY_BACKOFF + .saturating_mul(1_u32 << exponent) + .min(MAX_RETRY_BACKOFF) +} + +async fn sleep_before_retry(attempt: u32) { + tokio::time::sleep(retry_backoff(attempt)).await; +} + +fn failure_mark_data(attempt: u32, failure: &LibsyError) -> Json { + let mut data = Map::from_iter([ + ("attempt".into(), Json::from(attempt)), + ( + "retryable".into(), + Json::from(libsy_error_retryable(failure)), + ), + ]); + match failure { + LibsyError::ClientCall { + source: LlmClientError::UpstreamHttp { status, .. }, + .. + } => { + data.insert("failure_kind".into(), Json::from("http")); + data.insert("http_status".into(), Json::from(*status)); + } + LibsyError::ClientCall { source, .. } => { + data.insert("failure_kind".into(), Json::from("non_http")); + data.insert( + "non_http_kind".into(), + Json::from(client_error_label(source)), + ); + } + _ => { + data.insert("failure_kind".into(), Json::from("algorithm")); + } + } + Json::Object(data) +} + +fn client_error_label(error: &LlmClientError) -> &'static str { + match error { + LlmClientError::InvalidRequest { .. } => "invalid_request", + LlmClientError::RequestTranslation(_) => "request_translation", + LlmClientError::RequestEncoding(_) => "request_encoding", + LlmClientError::ResponseTranslation(_) => "response_translation", + LlmClientError::Configuration { .. } => "configuration", + LlmClientError::Transport { .. } => "transport", + LlmClientError::Timeout { .. } => "timeout", + LlmClientError::ContextWindowExceeded { .. } => "context_window_exceeded", + LlmClientError::UpstreamHttp { .. } => "http", + LlmClientError::InvalidResponse { .. } => "invalid_response", + LlmClientError::Ffi { .. } => "ffi", + LlmClientError::General(_) => "general", + _ => "unknown", + } +} + +fn public_libsy_failure(prefix: &str, error: &LibsyError) -> String { + match error { + LibsyError::ClientCall { source, .. } => public_client_failure(prefix, source), + _ => format!("{prefix}: Switchyard algorithm failure"), + } +} + +fn public_response_failure(prefix: &str, error: &LlmClientError) -> String { + match error { + LlmClientError::InvalidResponse { .. } => format!("{prefix}: invalid response"), + LlmClientError::ResponseTranslation(_) => { + format!("{prefix}: response translation failure") + } + _ => format!("{prefix}: response finalization failure"), + } +} + +fn public_client_failure(prefix: &str, error: &LlmClientError) -> String { + match error { + LlmClientError::UpstreamHttp { status, .. } => { + format!("{prefix}: provider returned HTTP {status}") + } + _ => format!("{prefix}: provider {} failure", client_error_label(error)), + } +} + +fn string_headers(headers: &Map) -> BTreeMap { + headers + .iter() + .filter_map(|(name, value)| value.as_str().map(|value| (name.clone(), value.into()))) + .collect() +} + +fn identity_metadata(metadata: Option<&Metadata>) -> Json { + json!({ + "session_id": metadata.and_then(|value| value.session_id.as_deref()), + "agent_id": metadata.and_then(|value| value.agent_id.as_deref()), + "parent_agent_id": metadata.and_then(|value| value.parent_agent_id.as_deref()), + "task_id": metadata.and_then(|value| value.task_id.as_deref()), + "turn_id": metadata.and_then(|value| value.turn_id.as_deref()), + "correlation_id": metadata.and_then(|value| value.correlation_id.as_deref()), + }) +} + +fn context_from_metadata(metadata: Option<&Metadata>) -> Context { + let Some(metadata) = metadata else { + return Context::default(); + }; + let mut values = std::collections::HashMap::new(); + for (name, value) in [ + ("session_id", metadata.session_id.as_deref()), + ("agent_id", metadata.agent_id.as_deref()), + ("parent_agent_id", metadata.parent_agent_id.as_deref()), + ("agent_kind", metadata.agent_kind.as_deref()), + ("agent_role", metadata.agent_role.as_deref()), + ("task_id", metadata.task_id.as_deref()), + ("task_kind", metadata.task_kind.as_deref()), + ("turn_id", metadata.turn_id.as_deref()), + ("correlation_id", metadata.correlation_id.as_deref()), + ] { + if let Some(value) = value { + values.insert(name.to_string(), value.to_string()); + } + } + values.insert("is_subagent".into(), metadata.is_subagent.to_string()); + values.insert( + "is_delegated_work".into(), + metadata.is_delegated_work.to_string(), + ); + if let Some(session_final) = metadata.session_final { + values.insert("session_final".into(), session_final.to_string()); + } + if let Some(extra) = &metadata.extra_metadata { + for (name, value) in extra { + values.entry(name.clone()).or_insert_with(|| value.clone()); + } + } + let mut context = Context::default(); + context.values = values; + context +} + +#[cfg(test)] +mod tests { + use std::sync::atomic::{AtomicUsize, Ordering}; + + use switchyard_libsy::{LlmTarget, Passthrough}; + use switchyard_protocol::{LlmResponseStream, RoutedLlmClient}; + + use super::*; + + #[derive(Clone, Copy)] + enum StreamBehavior { + Empty, + Failing, + CallFailure, + } + + struct StreamClient { + behavior: StreamBehavior, + calls: AtomicUsize, + } + + struct BufferedClient { + calls: AtomicUsize, + } + + #[async_trait::async_trait] + impl RoutedLlmClient for StreamClient { + async fn call( + &self, + _ctx: Context, + _request: Request, + _decision: Arc, + ) -> Result { + self.calls.fetch_add(1, Ordering::Relaxed); + let stream: LlmResponseStream = match self.behavior { + StreamBehavior::Empty => Box::pin(stream::empty()), + StreamBehavior::Failing => Box::pin(stream::once(async { + Err(LlmClientError::Transport { + source: Box::new(std::io::Error::other("fallback stream failed")), + }) + })), + StreamBehavior::CallFailure => { + return Err(LlmClientError::Transport { + source: Box::new(std::io::Error::other("fallback call failed")), + }); + } + }; + Ok(Response { + llm_response: LlmResponse::Stream(stream), + metadata: None, + }) + } + } + + #[async_trait::async_trait] + impl RoutedLlmClient for BufferedClient { + async fn call( + &self, + _ctx: Context, + _request: Request, + _decision: Arc, + ) -> Result { + self.calls.fetch_add(1, Ordering::Relaxed); + Ok(Response { + llm_response: LlmResponse::Agg(Default::default()), + metadata: None, + }) + } + } + + #[tokio::test] + async fn buffered_finalization_failure_uses_fallback_once() { + let selected = Arc::new(StreamClient { + behavior: StreamBehavior::Empty, + calls: AtomicUsize::new(0), + }); + let fallback = Arc::new(BufferedClient { + calls: AtomicUsize::new(0), + }); + let runtime = SwitchyardRuntime { + max_retries: 1, + algorithm: Arc::new(Passthrough::new(LlmTarget { + semantic_name: "selected".into(), + llm_client: Some(selected.clone()), + })), + targets: BTreeMap::from([( + "fallback".into(), + PreparedTargetBinding { + client: fallback.clone(), + }, + )]), + default_targets: BTreeMap::from([(WireFormat::OpenAiChat, "fallback".into())]), + translation: TranslationEngine::default(), + }; + let mut marks = Vec::new(); + + let response = runtime + .execute_buffered(WireFormat::OpenAiChat, Request::default(), &mut marks) + .await + .expect("the buffered fallback response should be encoded"); + + assert!(response.is_object()); + assert_eq!(selected.calls.load(Ordering::Relaxed), 1); + assert_eq!(fallback.calls.load(Ordering::Relaxed), 1); + assert!( + !marks + .iter() + .any(|mark| mark.name == "switchyard.routing.retry") + ); + let error = marks + .iter() + .find(|mark| mark.name == "switchyard.routing.error") + .expect("finalization failure should emit an error mark"); + assert_eq!(error.data["retryable"], false); + assert_eq!(error.data["non_http_kind"], "invalid_response"); + assert_eq!( + marks + .iter() + .filter(|mark| mark.name == "switchyard.routing.fallback") + .count(), + 1 + ); + } + + #[tokio::test] + async fn returned_events_replays_preserved_openai_chat_without_duplicate_terminal() { + let content = json!({ + "id": "chatcmpl-test", + "object": "chat.completion.chunk", + "model": "gpt-4o", + "system_fingerprint": "fp_provider_specific", + "choices": [{ + "index": 0, + "delta": {"content": "Hi"}, + "finish_reason": null + }] + }); + let terminal = json!({ + "id": "chatcmpl-test", + "object": "chat.completion.chunk", + "model": "gpt-4o", + "choices": [{ + "index": 0, + "delta": {}, + "finish_reason": "stop" + }] + }); + let body = format!("data: {content}\n\ndata: {terminal}\n\ndata: [DONE]\n\n").into_bytes(); + let stream = switchyard_translation::decode_stream( + stream::once(async move { Ok::<_, LlmClientError>(body) }), + WireFormat::OpenAiChat, + ) + .expect("provider SSE should decode"); + let response = Response { + llm_response: LlmResponse::Stream(stream), + metadata: None, + }; + + let replayed = returned_events(response, WireFormat::OpenAiChat) + .await + .expect("return stream should encode") + .collect::>() + .await + .into_iter() + .collect::, _>>() + .expect("return stream should not fail"); + + assert_eq!(replayed, vec![content, terminal]); + } + + #[tokio::test] + async fn invalid_selected_stream_does_not_invoke_failing_fallback_twice() { + let selected = Arc::new(StreamClient { + behavior: StreamBehavior::Empty, + calls: AtomicUsize::new(0), + }); + let fallback = Arc::new(StreamClient { + behavior: StreamBehavior::Failing, + calls: AtomicUsize::new(0), + }); + let runtime = SwitchyardRuntime { + max_retries: 0, + algorithm: Arc::new(Passthrough::new(LlmTarget { + semantic_name: "selected".into(), + llm_client: Some(selected.clone()), + })), + targets: BTreeMap::from([( + "fallback".into(), + PreparedTargetBinding { + client: fallback.clone(), + }, + )]), + default_targets: BTreeMap::from([(WireFormat::OpenAiChat, "fallback".into())]), + translation: TranslationEngine::default(), + }; + let (output, _messages) = async_channel::bounded(32); + + let error = runtime + .execute_stream(WireFormat::OpenAiChat, Request::default(), &output) + .await + .expect_err("the failing fallback stream must fail the request"); + + assert_eq!(error, "trusted fallback stream: provider transport failure"); + assert_eq!(selected.calls.load(Ordering::Relaxed), 1); + assert_eq!(fallback.calls.load(Ordering::Relaxed), 1); + } + + #[tokio::test] + async fn failing_fallback_call_flushes_error_and_fallback_marks() { + let selected = Arc::new(StreamClient { + behavior: StreamBehavior::Empty, + calls: AtomicUsize::new(0), + }); + let fallback = Arc::new(StreamClient { + behavior: StreamBehavior::CallFailure, + calls: AtomicUsize::new(0), + }); + let runtime = SwitchyardRuntime { + max_retries: 0, + algorithm: Arc::new(Passthrough::new(LlmTarget { + semantic_name: "selected".into(), + llm_client: Some(selected.clone()), + })), + targets: BTreeMap::from([( + "fallback".into(), + PreparedTargetBinding { + client: fallback.clone(), + }, + )]), + default_targets: BTreeMap::from([(WireFormat::OpenAiChat, "fallback".into())]), + translation: TranslationEngine::default(), + }; + let (output, messages) = async_channel::bounded(32); + + let error = runtime + .execute_stream(WireFormat::OpenAiChat, Request::default(), &output) + .await + .expect_err("the failing fallback call must fail the request"); + + assert_eq!(error, "trusted fallback: provider transport failure"); + assert_eq!(selected.calls.load(Ordering::Relaxed), 1); + assert_eq!(fallback.calls.load(Ordering::Relaxed), 1); + let mut terminal_marks = Vec::new(); + while let Ok(message) = messages.try_recv() { + if let StreamMessage::Mark(mark) = message + && matches!( + mark.name.as_str(), + "switchyard.routing.error" | "switchyard.routing.fallback" + ) + { + terminal_marks.push(mark.name); + } + } + assert_eq!( + terminal_marks, + ["switchyard.routing.error", "switchyard.routing.fallback"] + ); + } + + #[test] + fn retry_backoff_increases_exponentially_and_is_capped() { + assert_eq!(retry_backoff(1), Duration::from_millis(250)); + assert_eq!(retry_backoff(2), Duration::from_millis(500)); + assert_eq!(retry_backoff(3), Duration::from_secs(1)); + assert_eq!(retry_backoff(4), Duration::from_secs(2)); + assert_eq!(retry_backoff(u32::MAX), Duration::from_secs(2)); + } + + #[test] + fn context_carries_identity_without_http_headers() { + let context = context_from_metadata(Some(&Metadata { + session_id: Some("session-1".into()), + agent_id: Some("agent-1".into()), + is_subagent: true, + extra_metadata: Some(BTreeMap::from([("tenant".into(), "blue".into())])), + http_headers: Some(BTreeMap::from([( + "authorization".into(), + "Bearer caller-secret".into(), + )])), + ..Metadata::default() + })); + + assert_eq!( + context.values.get("session_id").map(String::as_str), + Some("session-1") + ); + assert_eq!( + context.values.get("agent_id").map(String::as_str), + Some("agent-1") + ); + assert_eq!( + context.values.get("is_subagent").map(String::as_str), + Some("true") + ); + assert_eq!( + context.values.get("tenant").map(String::as_str), + Some("blue") + ); + assert!(!context.values.contains_key("authorization")); + } +} diff --git a/crates/switchyard-nemo-relay-plugin/src/translation.rs b/crates/switchyard-nemo-relay-plugin/src/translation.rs new file mode 100644 index 00000000..6f6f965f --- /dev/null +++ b/crates/switchyard-nemo-relay-plugin/src/translation.rs @@ -0,0 +1,116 @@ +// SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. +// SPDX-License-Identifier: Apache-2.0 + +use nemo_relay_plugin::LlmRequest as RelayRequest; +use serde_json::Value as Json; +use switchyard_protocol::{AggLlmResponse, LlmRequest, WireFormat}; +use switchyard_translation::{ + DeterministicIdPolicy, DiagnosticSeverity, LossyConversionPolicy, PreservationPolicy, + TargetCapabilities, TranslationDiagnostic, TranslationEngine, TranslationPolicy, + UnknownFieldPolicy, +}; + +pub(crate) fn decode_request( + engine: &TranslationEngine, + protocol: WireFormat, + request: &RelayRequest, +) -> Result { + let output = engine + .decode_request(protocol, &request.content, &policy()) + .map_err(error)?; + safe(&output.diagnostics)?; + Ok(output.request) +} + +pub(crate) fn validate_target_request( + engine: &TranslationEngine, + protocol: WireFormat, + request: &LlmRequest, +) -> Result<(), String> { + let output = engine + .encode_request(protocol, request, &request_policy(protocol)) + .map_err(error)?; + safe(&output.diagnostics) +} + +pub(crate) fn encode_response( + engine: &TranslationEngine, + protocol: WireFormat, + response: &AggLlmResponse, +) -> Result { + let output = engine + .encode_response(protocol, response, &policy()) + .map_err(error)?; + safe(&output.diagnostics)?; + Ok(output.body) +} + +fn policy() -> TranslationPolicy { + TranslationPolicy { + unknown_field_policy: UnknownFieldPolicy::Preserve, + lossy_conversion_policy: LossyConversionPolicy::Reject, + deterministic_ids: DeterministicIdPolicy::GenerateStable { + prefix: "relay".into(), + }, + preservation: PreservationPolicy::InMemory, + target_capabilities: TargetCapabilities::default(), + } +} + +fn request_policy(protocol: WireFormat) -> TranslationPolicy { + let mut policy = policy(); + if protocol == WireFormat::AnthropicMessages { + policy + .target_capabilities + .supports_json_schema_response_format = Some(false); + } + policy +} + +fn safe(diagnostics: &[TranslationDiagnostic]) -> Result<(), String> { + let unsafe_diagnostics = diagnostics + .iter() + .filter(|diagnostic| diagnostic.severity != DiagnosticSeverity::Info) + .collect::>(); + if unsafe_diagnostics.is_empty() { + Ok(()) + } else { + Err(format!( + "Switchyard translation was not lossless: {unsafe_diagnostics:?}" + )) + } +} + +fn error(error: switchyard_translation::TranslationError) -> String { + format!("Switchyard translation failed: {error}") +} + +#[cfg(test)] +mod tests { + use serde_json::{Map, json}; + + use super::*; + + #[test] + fn same_protocol_request_preserves_unknown_fields() { + let request = RelayRequest { + headers: Map::new(), + content: json!({ + "model": "route", + "messages": [{"role": "user", "content": "hello"}], + "provider_extension": {"exact": true} + }), + }; + let engine = TranslationEngine::default(); + let decoded = decode_request(&engine, WireFormat::OpenAiChat, &request).unwrap(); + validate_target_request(&engine, WireFormat::OpenAiChat, &decoded).unwrap(); + assert_eq!( + decoded + .preservation + .requests + .get(&WireFormat::OpenAiChat.into()) + .and_then(|body| body.get("provider_extension")), + Some(&json!({"exact": true})) + ); + } +} diff --git a/docs/index.md b/docs/index.md index 1dc5056c..93814e9c 100644 --- a/docs/index.md +++ b/docs/index.md @@ -10,6 +10,7 @@ It supports OpenAI Chat Completions, OpenAI Responses, and Anthropic Messages. | Run Claude Code, Codex, or OpenClaw through Switchyard | Launcher Path | [Install and launch an agent](getting_started.md#launcher-path) | | Run Switchyard as a standalone proxy for API clients | Server Path | [Build and run the Rust server](getting_started.md#server-path) | | Add Switchyard routing to a Rust application | Library Path | [`switchyard-libsy`](../crates/libsy/README.md) | +| Add Switchyard routing to NeMo Relay | Native Plugin Path | [`switchyard-nemo-relay-plugin`](../crates/switchyard-nemo-relay-plugin/README.md) | The Launcher Path installs the `switchyard` CLI and hosts the native Rust server through its packaged PyO3 binding. The Server Path builds and runs the @@ -30,3 +31,4 @@ standalone `switchyard-server` binary. - [`switchyard-libsy`](reference/rust_api.md#switchyard-libsy): embeddable routing algorithms - [`switchyard-protocol`](reference/rust_api.md#switchyard-protocol): provider-neutral API types - [`switchyard-translation`](../crates/switchyard-translation/README.md): protocol translation +- [`switchyard-nemo-relay-plugin`](../crates/switchyard-nemo-relay-plugin/README.md): native NeMo Relay integration