switchyard-libsy 0.3.0

Provider-neutral multi-LLM routing and orchestration for Switchyard
Documentation
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
// SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
// SPDX-License-Identifier: Apache-2.0

//! Escalation routing that judges an efficient model's answer before selecting a serving tier.

use std::sync::Arc;

use async_trait::async_trait;
use switchyard_protocol::{
    AggLlmResponse, Category, LlmClientError, Message, Request, Response, Role,
};

use super::util::buffered_response::buffer_response;
use super::util::classifier_contract::ClassifierContractConfig;
use super::util::decisive;
use super::util::escalation::{self, EscalationJudge, EscalationJudgeConfig, EscalationPolicy};
use super::util::llm_judge::JudgeClassifier;
use crate::core::algorithm::Driver;
use crate::core::classifier::{Classification, Classifier};
use crate::core::state::{State, StateValue};
use crate::{LibsyError, Result};

/// Session-state key holding the consecutive-escalate streak.
const STREAK_KEY: &str = "escalation_streak";

fn streak(state: &State) -> u32 {
    match state.extra.get(STREAK_KEY) {
        Some(StateValue::Count(n)) => *n,
        _ => 0,
    }
}

fn assistant_message(response: &AggLlmResponse) -> Message {
    Message {
        role: Role::Assistant,
        content: response
            .first_output()
            .map(|output| output.content.clone())
            .unwrap_or_default(),
    }
}

/// Calls the efficient model, judges its response, and latches to capable once the streak
/// confirms. Returns the efficient response directly when not escalating so the caller does
/// not pay for a second model call.
struct EscalationClassifier {
    judge: JudgeClassifier<EscalationJudge, EscalationPolicy>,
    /// Consecutive escalate verdicts required to latch.
    confirmations: u32,
}

/// Builds the escalation classifier used by the shared LLM classifier route shell.
pub(super) fn build_classifier(
    contract_config: ClassifierContractConfig,
    config: EscalationJudgeConfig,
    max_output_tokens: u64,
) -> Result<Arc<dyn Classifier<State>>> {
    let confirmations = config.confirmations;
    let classifier: Arc<dyn Classifier<State>> = Arc::new(EscalationClassifier {
        judge: escalation::build_judge(&contract_config, config, max_output_tokens)?,
        confirmations,
    });
    Ok(classifier)
}

#[async_trait]
impl Classifier<State> for EscalationClassifier {
    async fn score(
        &self,
        state: &mut State,
        request: &mut Request,
        driver: &Driver,
    ) -> Result<(Classification, Option<Response>)> {
        let capable = driver.first_model_for(&Category::Capable)?.clone();
        let efficient = driver.first_model_for(&Category::Efficient)?.clone();

        // A confirmed session stays capable without a judge call.
        if streak(state) >= self.confirmations {
            driver.set_evidence(serde_json::json!({
                "source": "escalation",
                "verdict": "latched",
            }));
            return Ok((decisive(&capable), None));
        }

        // Call efficient model and buffer the response so the judge can read it.
        //
        // If the efficient model exceeds its context window, fall through to capable. This call
        // deliberately has one candidate so the classifier sees the efficient model's error.
        tracing::info!(
            target = %efficient,
            "escalation classifier selected efficient tier"
        );
        let efficient_response = match driver
            .call_model(request.clone(), vec![efficient.clone()])
            .await
        {
            Ok(r) => r,
            Err(LibsyError::ClientCall {
                source: LlmClientError::ContextWindowExceeded { .. },
                ..
            }) => {
                driver.set_evidence(serde_json::json!({
                    "source": "fallback",
                    "reason_code": "context_window",
                }));
                return Ok((decisive(&capable), None));
            }
            Err(e) => return Err(e),
        };
        // The call resolves when its stream handle arrives; context and transport failures can
        // still occur while buffering.
        let efficient_response = match buffer_response(efficient.as_str(), efficient_response).await
        {
            Ok(response) => response,
            Err(LibsyError::ClientCall {
                source: LlmClientError::ContextWindowExceeded { .. },
                ..
            }) => {
                driver.set_evidence(serde_json::json!({
                    "source": "fallback",
                    "reason_code": "context_window",
                }));
                return Ok((decisive(&capable), None));
            }
            Err(LibsyError::ClientCall {
                source: LlmClientError::Transport { .. },
                ..
            }) => {
                driver.set_evidence(serde_json::json!({
                    "source": "fallback",
                    "reason_code": "transport",
                }));
                return Ok((decisive(&capable), None));
            }
            Err(error) => return Err(error),
        };
        // Append the efficient reply so the judge reads this turn's completed trajectory.
        let mut judge_request = request.clone();
        judge_request
            .llm_request
            .messages
            .push(assistant_message(&efficient_response.agg));

        let (classification, _) = self.judge.score(state, &mut judge_request, driver).await?;

        let held = streak(state);
        let best = classification.argmax(false)?;
        let (escalate, pending) = match &best {
            Some(score) if score.target == capable => (true, held + 1),
            Some(_) => (false, 0),
            None => (false, held),
        };
        state
            .extra
            .insert(STREAK_KEY.to_string(), StateValue::Count(pending));

        if escalate && pending >= self.confirmations {
            // Streak confirmed: drop the efficient response, caller will serve capable.
            driver.set_evidence(serde_json::json!({
                "source": "escalation",
                "verdict": "escalate",
            }));
            return Ok((decisive(&capable), None));
        }

        if escalate {
            driver.set_evidence(serde_json::json!({
                "source": "escalation",
                "verdict": "pending",
            }));
        }

        Ok((
            decisive(&efficient),
            Some(efficient_response.into_response()),
        ))
    }
}

#[cfg(test)]
mod tests {
    use std::collections::{HashMap, VecDeque};
    use std::sync::Arc;

    use parking_lot::Mutex;
    use switchyard_protocol::{
        ContentBlock, LlmClientError, LlmResponse, LlmResponseChunk, Metadata, ModelId, Request,
        Response, completion_text, text_request, text_response,
    };

    use super::*;
    use crate::algorithms::llm_class::{LlmClassifierConfig, LlmTaskClassifier};
    use crate::algorithms::util::DEFAULT_JUDGE_MAX_OUTPUT_TOKENS;
    use crate::core::testing::{Serve, reply, test_drive_with_models};

    /// A queue of replies, drained in order.
    struct Queue(Mutex<VecDeque<String>>);

    impl Queue {
        fn new(replies: impl IntoIterator<Item = &'static str>) -> Arc<Self> {
            Arc::new(Self(Mutex::new(
                replies.into_iter().map(String::from).collect(),
            )))
        }

        fn take(&self) -> String {
            self.0
                .lock()
                .pop_front()
                .unwrap_or_else(|| "unexpected call".to_string())
        }
    }

    /// Serves the judge and model targets from separate reply queues.
    fn queued(model: Arc<Queue>, judge: Arc<Queue>) -> impl Serve {
        move |target: ModelId, request: Request| {
            let queue = if target == "judge" {
                Arc::clone(&judge)
            } else {
                Arc::clone(&model)
            };
            async move {
                Ok(Response {
                    llm_response: LlmResponse::Agg(text_response(None, queue.take())),
                    metadata: request.metadata,
                    upstream_headers: http::HeaderMap::new(),
                })
            }
        }
    }

    fn classify_request() -> Request {
        Request {
            llm_request: text_request(Some("auto".to_string()), "classify this task"),
            raw_request: None,
            metadata: None,
        }
    }

    fn classify_session_request() -> Request {
        Request {
            metadata: Some(Metadata {
                session_id: Some("session-1".to_string()),
                ..Metadata::default()
            }),
            ..classify_request()
        }
    }

    fn runtime_models() -> HashMap<Category, Vec<ModelId>> {
        [
            (Category::Judge, vec![ModelId::from("judge")]),
            (Category::Efficient, vec![ModelId::from("efficient")]),
            (Category::Capable, vec![ModelId::from("capable")]),
            (
                Category::Any,
                vec![ModelId::from("capable"), ModelId::from("efficient")],
            ),
        ]
        .into()
    }

    /// Returns a stream that emits partial content before failing during aggregation.
    fn streamed_then_error(error: LlmClientError) -> Response {
        Response {
            llm_response: LlmResponse::Stream(Box::pin(futures::stream::iter([
                Ok(LlmResponseChunk::TextDelta {
                    index: 0,
                    text: "partial".to_string(),
                }
                .into()),
                Err(error),
            ]))),
            metadata: None,
            upstream_headers: http::HeaderMap::new(),
        }
    }

    /// Builds a router with escalation enabled (`confirmations=1` latches immediately).
    fn escalation_router() -> Result<Arc<LlmTaskClassifier>> {
        Ok(Arc::new(LlmTaskClassifier::new(
            LlmClassifierConfig::Escalation {
                contract: ClassifierContractConfig::default(),
                config: EscalationJudgeConfig {
                    confirmations: 1,
                    ..EscalationJudgeConfig::default()
                },
                max_output_tokens: DEFAULT_JUDGE_MAX_OUTPUT_TOKENS,
            },
        )?))
    }

    #[tokio::test]
    async fn serves_efficient_when_judge_declines() -> Result<()> {
        let judge = Queue::new([r#"{"escalate":false,"reason":"progressing"}"#]);
        let model = Queue::new(["efficient answer"]);

        let (selected_model, response) = test_drive_with_models(
            escalation_router()?,
            classify_request(),
            runtime_models(),
            queued(model, judge),
        )
        .await?;

        assert_eq!(selected_model, "efficient");
        assert_eq!(
            response.llm_response.as_agg().map(completion_text),
            Some("efficient answer".to_string())
        );
        Ok(())
    }

    #[tokio::test]
    async fn config_overrides_the_packaged_prompt() -> Result<()> {
        let prompts = Arc::new(Mutex::new(Vec::new()));
        let recorded = Arc::clone(&prompts);
        let serve = move |target: ModelId, request: Request| {
            if target == "judge" {
                let prompt = request
                    .llm_request
                    .instructions
                    .first()
                    .and_then(|instruction| {
                        instruction.content.iter().find_map(|block| match block {
                            ContentBlock::Text { text } => Some(text.clone()),
                            _ => None,
                        })
                    });
                recorded.lock().extend(prompt);
                std::future::ready(Ok(reply(r#"{"escalate":false,"reason":"progressing"}"#)))
            } else {
                std::future::ready(Ok(reply("efficient answer")))
            }
        };
        let router = Arc::new(LlmTaskClassifier::new(LlmClassifierConfig::Escalation {
            contract: ClassifierContractConfig::default().with_prompt("Custom trajectory rubric."),
            config: EscalationJudgeConfig {
                confirmations: 1,
                ..EscalationJudgeConfig::default()
            },
            max_output_tokens: DEFAULT_JUDGE_MAX_OUTPUT_TOKENS,
        })?);

        test_drive_with_models(router, classify_request(), runtime_models(), serve).await?;

        assert_eq!(&*prompts.lock(), &["Custom trajectory rubric."]);
        Ok(())
    }

    #[tokio::test]
    async fn upgrades_to_capable_when_judge_escalates() -> Result<()> {
        let judge = Queue::new([r#"{"escalate":true,"reason":"stuck in a loop"}"#]);
        let model = Queue::new(["efficient draft", "capable answer"]);

        let (selected_model, response) = test_drive_with_models(
            escalation_router()?,
            classify_request(),
            runtime_models(),
            queued(model, judge),
        )
        .await?;

        assert_eq!(selected_model, "capable");
        assert_eq!(
            response.llm_response.as_agg().map(completion_text),
            Some("capable answer".to_string())
        );
        Ok(())
    }

    #[tokio::test]
    async fn stays_capable_after_latch() -> Result<()> {
        let judge = Queue::new([r#"{"escalate":true,"reason":"stuck"}"#]);
        let model = Queue::new(["efficient draft", "capable t1", "capable t2"]);
        let router = escalation_router()?;
        let request = classify_session_request();

        test_drive_with_models(
            router.clone(),
            request.clone(),
            runtime_models(),
            queued(Arc::clone(&model), Arc::clone(&judge)),
        )
        .await?;
        let (selected_model, _) =
            test_drive_with_models(router, request, runtime_models(), queued(model, judge)).await?;

        assert_eq!(selected_model, "capable");
        Ok(())
    }

    #[tokio::test]
    async fn falls_back_to_capable_when_efficient_overflows() -> Result<()> {
        for streamed in [false, true] {
            let serve = move |target: ModelId, _request: Request| async move {
                match target.as_str() {
                    "efficient" if streamed => {
                        Ok(streamed_then_error(LlmClientError::ContextWindowExceeded {
                            model: target,
                            message: "prompt is too long".to_string(),
                        }))
                    }
                    "efficient" => Err(LlmClientError::ContextWindowExceeded {
                        model: target,
                        message: "prompt is too long".to_string(),
                    }),
                    "judge" => {
                        panic!("the judge must not be consulted when efficient overflows")
                    }
                    _ => Ok(reply("capable answer")),
                }
            };
            let mut request = classify_request();
            request.llm_request.stream = streamed;

            let (selected_model, response) =
                test_drive_with_models(escalation_router()?, request, runtime_models(), serve)
                    .await?;

            assert_eq!(selected_model, "capable");
            assert_eq!(
                response.llm_response.as_agg().map(completion_text),
                Some("capable answer".to_string())
            );
        }
        Ok(())
    }

    /// A transport failure while buffering efficient must bypass the judge and serve capable.
    #[tokio::test]
    async fn falls_back_when_efficient_stream_transport_fails() -> Result<()> {
        let calls = Arc::new(Mutex::new(Vec::new()));
        let serve = {
            let calls = Arc::clone(&calls);
            move |model: ModelId, _request: Request| {
                let calls = Arc::clone(&calls);
                async move {
                    let model = model.to_string();
                    calls.lock().push(model.clone());
                    match model.as_str() {
                        "efficient" => Ok(streamed_then_error(LlmClientError::Transport {
                            source: Box::new(std::io::Error::other("stream disconnected")),
                        })),
                        "judge" => {
                            panic!("the judge must not be consulted after a transport failure")
                        }
                        _ => Ok(reply("capable answer")),
                    }
                }
            }
        };
        let mut request = classify_request();
        request.llm_request.stream = true;

        let result =
            test_drive_with_models(escalation_router()?, request, runtime_models(), serve).await;

        assert_eq!(&*calls.lock(), &["efficient", "capable"]);
        let (_, response) = result?;
        assert_eq!(
            response.llm_response.as_agg().map(completion_text),
            Some("capable answer".to_string())
        );
        Ok(())
    }

    /// Non-transport aggregation failures remain typed and do not silently change targets.
    #[tokio::test]
    async fn preserves_non_transport_stream_errors() -> Result<()> {
        let serve = |target: ModelId, _request: Request| async move {
            match target.as_str() {
                "efficient" => Ok(streamed_then_error(LlmClientError::InvalidResponse {
                    source: Box::new(std::io::Error::other("invalid stream event")),
                })),
                other => panic!("unexpected call to {other}"),
            }
        };
        let mut request = classify_request();
        request.llm_request.stream = true;

        match test_drive_with_models(escalation_router()?, request, runtime_models(), serve).await {
            Err(LibsyError::ClientCall {
                target,
                source: LlmClientError::InvalidResponse { .. },
            }) => {
                assert_eq!(target, "efficient");
                Ok(())
            }
            Err(other) => panic!("expected InvalidResponse client error, got {other:?}"),
            Ok(_) => panic!("expected stream aggregation to fail"),
        }
    }
}