switchyard_libsy/algorithms/
plan_execute.rs1use std::collections::HashSet;
7use std::sync::Arc;
8
9use parking_lot::Mutex;
10use switchyard_protocol::{Category, ContentBlock, Request};
11
12use super::util::prompts::{append_note, drop_exact_replay, prepend_system_prompt};
13use super::util::tool_signals::ToolSignals;
14use crate::core::algorithm::{Algorithm, Driver, RoutingIdentity};
15use crate::{LibsyError, Result, RoutingOutcome};
16
17pub const DEFAULT_PLANNING_PROMPT: &str =
19 include_str!("../prompts/plan-execute/planning-system-prompt.md");
20
21const MAX_EXECUTING_SESSIONS: usize = 4_096;
22
23#[derive(Clone, Debug)]
25pub struct PlanExecuteConfig {
26 pub planning_prompt: String,
28 pub handoff_prompt: Option<String>,
30 pub planner_reasoning_as_text: bool,
32}
33
34impl Default for PlanExecuteConfig {
35 fn default() -> Self {
36 Self {
37 planning_prompt: DEFAULT_PLANNING_PROMPT.trim().to_string(),
38 handoff_prompt: None,
39 planner_reasoning_as_text: false,
40 }
41 }
42}
43
44#[derive(Clone, Copy, Debug, Eq, PartialEq)]
45enum Phase {
46 Plan,
47 Handoff,
48 Execute,
49}
50
51pub struct PlanExecute {
54 config: PlanExecuteConfig,
55 executing_sessions: Mutex<HashSet<RoutingIdentity>>,
56}
57
58impl PlanExecute {
59 pub fn new(config: PlanExecuteConfig) -> Result<Self> {
63 if config.planning_prompt.trim().is_empty() {
64 return Err(LibsyError::AlgorithmError {
65 message: "planning_prompt must not be empty".to_string(),
66 });
67 }
68 if config
69 .handoff_prompt
70 .as_deref()
71 .is_some_and(|prompt| prompt.trim().is_empty())
72 {
73 return Err(LibsyError::AlgorithmError {
74 message: "handoff_prompt must not be empty".to_string(),
75 });
76 }
77 Ok(Self {
78 config,
79 executing_sessions: Mutex::new(HashSet::new()),
80 })
81 }
82
83 fn phase(&self, request: &Request) -> Phase {
84 let signals = ToolSignals::from_request(request, None);
85 let mutation_seen = signals.edit_count > 0 || signals.write_count > 0;
86 let Some(identity) = RoutingIdentity::from_request(request) else {
87 return if mutation_seen {
88 Phase::Handoff
89 } else {
90 Phase::Plan
91 };
92 };
93
94 let mut sessions = self.executing_sessions.lock();
95 let phase = if sessions.contains(&identity) {
96 Phase::Execute
97 } else if mutation_seen {
98 if sessions.len() >= MAX_EXECUTING_SESSIONS
99 && let Some(evicted) = sessions.iter().next().cloned()
100 {
101 sessions.remove(&evicted);
102 }
103 sessions.insert(identity.clone());
104 Phase::Handoff
105 } else {
106 Phase::Plan
107 };
108 if request
109 .metadata
110 .as_ref()
111 .and_then(|metadata| metadata.session_final)
112 == Some(true)
113 {
114 sessions.remove(&identity);
115 }
116 phase
117 }
118
119 fn replay_planner_reasoning_as_text(request: &mut Request) -> usize {
120 let mut converted = 0;
121 for message in &mut request.llm_request.messages {
122 message.content = std::mem::take(&mut message.content)
123 .into_iter()
124 .filter_map(|block| match block {
125 ContentBlock::Reasoning { text, .. } => {
126 converted += 1;
127 (!text.is_empty()).then_some(ContentBlock::Text { text })
128 }
129 other => Some(other),
130 })
131 .collect();
132 }
133 request
134 .llm_request
135 .messages
136 .retain(|message| !message.content.is_empty());
137 if converted > 0 {
138 drop_exact_replay(request);
139 }
140 converted
141 }
142
143 fn route_to(driver: &Driver, category: Category, request: Request) -> Result<RoutingOutcome> {
144 let models = driver.models_for(&category);
145 let Some((selected, fallbacks)) = models.split_first() else {
146 return Err(LibsyError::AlgorithmError {
147 message: format!("no models available for category {}", category.as_str()),
148 });
149 };
150 Ok(RoutingOutcome::route_to(
151 selected.clone(),
152 fallbacks.to_vec(),
153 request,
154 ))
155 }
156}
157
158#[async_trait::async_trait]
159impl Algorithm for PlanExecute {
160 fn name(&self) -> &str {
161 "plan_execute"
162 }
163
164 async fn route(
165 self: Arc<Self>,
166 driver: Driver,
167 mut request: Request,
168 ) -> Result<RoutingOutcome> {
169 match self.phase(&request) {
170 Phase::Plan => {
171 prepend_system_prompt(&mut request, &self.config.planning_prompt);
172 tracing::debug!(phase = "plan", "plan-execute selected capable tier");
173 Self::route_to(&driver, Category::Capable, request)
174 }
175 Phase::Handoff => {
176 let reasoning_converted = if self.config.planner_reasoning_as_text {
177 Self::replay_planner_reasoning_as_text(&mut request)
178 } else {
179 0
180 };
181 let prompt_applied = if let Some(prompt) = &self.config.handoff_prompt {
182 append_note(&mut request, prompt);
183 true
184 } else {
185 false
186 };
187 tracing::debug!(
188 phase = "handoff",
189 handoff_prompt_applied = prompt_applied,
190 planner_reasoning_converted = reasoning_converted,
191 "plan-execute selected efficient tier"
192 );
193 Self::route_to(&driver, Category::Efficient, request)
194 }
195 Phase::Execute => {
196 tracing::debug!(phase = "execute", "plan-execute selected efficient tier");
197 Self::route_to(&driver, Category::Efficient, request)
198 }
199 }
200 }
201}
202
203#[cfg(test)]
204mod tests {
205 use std::collections::HashMap;
206 use std::sync::Arc;
207
208 use serde_json::json;
209 use switchyard_protocol::{
210 ContentBlock, LlmRequest, Message, Metadata, ModelId, Request, Role, ToolCall,
211 };
212
213 use super::*;
214 use crate::RuntimeModels;
215 use crate::core::testing::{reply, test_drive_with_models};
216
217 const CAPABLE: &str = "model/capable";
218 const EFFICIENT: &str = "model/efficient";
219
220 fn algorithm(config: PlanExecuteConfig) -> Arc<dyn Algorithm> {
221 Arc::new(PlanExecute::new(config).expect("config should be valid"))
222 }
223
224 fn request(messages: Vec<Message>, session_id: Option<&str>) -> Request {
225 Request {
226 llm_request: LlmRequest {
227 model: Some("switchyard/plan-execute".to_string()),
228 messages,
229 ..LlmRequest::default()
230 },
231 metadata: session_id.map(|session_id| Metadata {
232 session_id: Some(session_id.to_string()),
233 ..Metadata::default()
234 }),
235 ..Request::default()
236 }
237 }
238
239 fn tool_call(name: &str, arguments: serde_json::Value) -> Message {
240 Message {
241 role: Role::Assistant,
242 content: vec![ContentBlock::ToolCall(ToolCall {
243 id: "call-1".to_string(),
244 name: name.to_string(),
245 arguments,
246 })],
247 }
248 }
249
250 fn models() -> RuntimeModels {
251 RuntimeModels::new(HashMap::from([
252 (Category::Capable, vec![ModelId::from(CAPABLE)]),
253 (Category::Efficient, vec![ModelId::from(EFFICIENT)]),
254 ]))
255 }
256
257 async fn route_and_capture(
258 algorithm: Arc<dyn Algorithm>,
259 request: Request,
260 ) -> (ModelId, Request) {
261 let captured = Arc::new(Mutex::new(None));
262 let capture = Arc::clone(&captured);
263 let (selected, _) =
264 test_drive_with_models(algorithm, request, models(), move |_target, request| {
265 let capture = Arc::clone(&capture);
266 async move {
267 *capture.lock() = Some(request);
268 Ok(reply("ok"))
269 }
270 })
271 .await
272 .expect("routing should succeed");
273 let request = captured
274 .lock()
275 .take()
276 .expect("answer request should be captured");
277 (selected, request)
278 }
279
280 #[tokio::test]
281 async fn plans_then_hands_off_and_latches_execution() {
282 const HANDOFF: &str = "Continue from the plan and repository evidence.";
283 let algorithm = algorithm(PlanExecuteConfig {
284 handoff_prompt: Some(HANDOFF.to_string()),
285 planner_reasoning_as_text: true,
286 ..PlanExecuteConfig::default()
287 });
288
289 let read_only = request(
290 vec![tool_call(
291 "exec_command",
292 json!({"cmd": "rg parser crates"}),
293 )],
294 Some("task-1"),
295 );
296 let (selected, routed) = route_and_capture(Arc::clone(&algorithm), read_only).await;
297 assert_eq!(selected, CAPABLE);
298 assert_eq!(routed.llm_request.instructions.len(), 1);
299
300 let first_edit = request(
301 vec![Message {
302 role: Role::Assistant,
303 content: vec![
304 ContentBlock::Reasoning {
305 text: "The parser needs a boundary check.".to_string(),
306 signature: Some("planner-signature".to_string()),
307 details: vec![json!({"type": "reasoning.encrypted", "data": "opaque"})],
308 },
309 ContentBlock::ToolCall(ToolCall {
310 id: "call-1".to_string(),
311 name: "apply_patch".to_string(),
312 arguments: json!({"patch": "*** Begin Patch"}),
313 }),
314 ],
315 }],
316 Some("task-1"),
317 );
318 let (selected, routed) = route_and_capture(Arc::clone(&algorithm), first_edit).await;
319 assert_eq!(selected, EFFICIENT);
320 assert_eq!(
321 routed.llm_request.messages[0].content[0],
322 ContentBlock::Text {
323 text: "The parser needs a boundary check.".to_string()
324 }
325 );
326 assert_eq!(
327 routed.llm_request.messages.last(),
328 Some(&Message::text(Role::User, HANDOFF))
329 );
330
331 let mut final_request = request(
332 vec![Message::text(Role::User, "Continue after compaction")],
333 Some("task-1"),
334 );
335 final_request
336 .metadata
337 .as_mut()
338 .expect("session metadata should exist")
339 .session_final = Some(true);
340 let (selected, routed) = route_and_capture(Arc::clone(&algorithm), final_request).await;
341 assert_eq!(selected, EFFICIENT);
342 assert!(routed.llm_request.instructions.is_empty());
343 assert_eq!(routed.llm_request.messages.len(), 1);
344
345 let reused = request(vec![Message::text(Role::User, "New task")], Some("task-1"));
346 let (selected, _) = route_and_capture(algorithm, reused).await;
347 assert_eq!(selected, CAPABLE);
348 }
349
350 #[tokio::test]
351 async fn mutation_without_a_session_uses_the_efficient_tier() {
352 let messages = vec![tool_call(
353 "exec_command",
354 json!({"cmd": "printf 'done\\n' > task.txt"}),
355 )];
356
357 let (selected, routed) = route_and_capture(
358 algorithm(PlanExecuteConfig::default()),
359 request(messages.clone(), None),
360 )
361 .await;
362
363 assert_eq!(selected, EFFICIENT);
364 assert_eq!(routed.llm_request.messages, messages);
365 assert!(routed.llm_request.instructions.is_empty());
366 }
367
368 #[test]
369 fn rejects_blank_prompts() {
370 for config in [
371 PlanExecuteConfig {
372 planning_prompt: " ".to_string(),
373 ..PlanExecuteConfig::default()
374 },
375 PlanExecuteConfig {
376 handoff_prompt: Some(" ".to_string()),
377 ..PlanExecuteConfig::default()
378 },
379 ] {
380 assert!(matches!(
381 PlanExecute::new(config),
382 Err(LibsyError::AlgorithmError { .. })
383 ));
384 }
385 }
386}