swiftide_agents/
agent.rs

1#![allow(dead_code)]
2use crate::{
3    default_context::DefaultContext,
4    hooks::{
5        AfterCompletionFn, AfterEachFn, AfterToolFn, BeforeAllFn, BeforeCompletionFn, BeforeToolFn,
6        Hook, HookTypes, MessageHookFn, OnStartFn,
7    },
8    state,
9    system_prompt::SystemPrompt,
10    tools::{arg_preprocessor::ArgPreprocessor, control::Stop},
11};
12use std::{
13    collections::{HashMap, HashSet},
14    hash::{DefaultHasher, Hash as _, Hasher as _},
15    sync::Arc,
16};
17
18use anyhow::Result;
19use derive_builder::Builder;
20use swiftide_core::{
21    chat_completion::{
22        ChatCompletion, ChatCompletionRequest, ChatMessage, Tool, ToolCall, ToolOutput,
23    },
24    prompt::Prompt,
25    AgentContext,
26};
27use tracing::{debug, Instrument};
28
29/// Agents are the main interface for building agentic systems.
30///
31/// Construct agents by calling the builder, setting an llm, configure hooks, tools and other
32/// customizations.
33///
34/// # Important defaults
35///
36/// - The default context is the `DefaultContext`, executing tools locally with the `LocalExecutor`.
37/// - A default `stop` tool is provided for agents to explicitly stop if needed
38/// - The default `SystemPrompt` instructs the agent with chain of thought and some common safeguards, but is otherwise quite bare. In a lot of cases this can be sufficient.
39#[derive(Clone, Builder)]
40pub struct Agent {
41    /// Hooks are functions that are called at specific points in the agent's lifecycle.
42    #[builder(default, setter(into))]
43    pub(crate) hooks: Vec<Hook>,
44    /// The context in which the agent operates, by default this is the `DefaultContext`.
45    #[builder(
46        setter(custom),
47        default = Arc::new(DefaultContext::default()) as Arc<dyn AgentContext>
48    )]
49    pub(crate) context: Arc<dyn AgentContext>,
50    /// Tools the agent can use
51    #[builder(default = Agent::default_tools(), setter(custom))]
52    pub(crate) tools: HashSet<Box<dyn Tool>>,
53
54    /// The language model that the agent uses for completion.
55    #[builder(setter(custom))]
56    pub(crate) llm: Box<dyn ChatCompletion>,
57
58    /// System prompt for the agent when it starts
59    ///
60    /// Some agents profit significantly from a tailored prompt. But it is not always needed.
61    ///
62    /// See [`SystemPrompt`] for an opiniated, customizable system prompt.
63    ///
64    /// Swiftide provides a default system prompt for all agents.
65    ///
66    /// # Example
67    ///
68    /// ```no_run
69    /// # use swiftide_agents::system_prompt::SystemPrompt;
70    /// # use swiftide_agents::Agent;
71    /// Agent::builder()
72    ///     .system_prompt(
73    ///         SystemPrompt::builder().role("You are an expert engineer")
74    ///         .build().unwrap())
75    ///     .build().unwrap();
76    /// ```
77    #[builder(setter(into, strip_option), default = Some(SystemPrompt::default().into()))]
78    pub(crate) system_prompt: Option<Prompt>,
79
80    /// Initial state of the agent
81    #[builder(private, default = state::State::default())]
82    pub(crate) state: state::State,
83
84    /// Optional limit on the amount of loops the agent can run.
85    /// The counter is reset when the agent is stopped.
86    #[builder(default, setter(strip_option))]
87    pub(crate) limit: Option<usize>,
88
89    /// The maximum amount of times the failed output of a tool will be send
90    /// to an LLM before the agent stops. Defaults to 3.
91    ///
92    /// LLMs sometimes send missing arguments, or a tool might actually fail, but retrying could be
93    /// worth while. If the limit is not reached, the agent will send the formatted error back to
94    /// the LLM.
95    ///
96    /// The limit is hashed based on the tool call name and arguments, so the limit is per tool call.
97    ///
98    /// This limit is _not_ reset when the agent is stopped.
99    #[builder(default = 3)]
100    pub(crate) tool_retry_limit: usize,
101
102    /// Internally tracks the amount of times a tool has been retried. The key is a hash based on
103    /// the name and args of the tool.
104    #[builder(private, default)]
105    pub(crate) tool_retries_counter: HashMap<u64, usize>,
106}
107
108impl std::fmt::Debug for Agent {
109    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
110        f.debug_struct("Agent")
111            //display hooks as a list of type: number of hooks
112            .field(
113                "hooks",
114                &self
115                    .hooks
116                    .iter()
117                    .map(std::string::ToString::to_string)
118                    .collect::<Vec<_>>(),
119            )
120            .field(
121                "tools",
122                &self
123                    .tools
124                    .iter()
125                    .map(swiftide_core::Tool::name)
126                    .collect::<Vec<_>>(),
127            )
128            .field("llm", &"Box<dyn ChatCompletion>")
129            .field("state", &self.state)
130            .finish()
131    }
132}
133
134impl AgentBuilder {
135    /// The context in which the agent operates, by default this is the `DefaultContext`.
136    pub fn context(&mut self, context: impl AgentContext + 'static) -> &mut AgentBuilder
137    where
138        Self: Clone,
139    {
140        self.context = Some(Arc::new(context) as Arc<dyn AgentContext>);
141        self
142    }
143
144    /// Disable the system prompt.
145    pub fn no_system_prompt(&mut self) -> &mut Self {
146        self.system_prompt = Some(None);
147
148        self
149    }
150
151    /// Add a hook to the agent.
152    pub fn add_hook(&mut self, hook: Hook) -> &mut Self {
153        let hooks = self.hooks.get_or_insert_with(Vec::new);
154        hooks.push(hook);
155
156        self
157    }
158
159    /// Add a hook that runs once, before all completions. Even if the agent is paused and resumed,
160    /// before all will not trigger again.
161    pub fn before_all(&mut self, hook: impl BeforeAllFn + 'static) -> &mut Self {
162        self.add_hook(Hook::BeforeAll(Box::new(hook)))
163    }
164
165    /// Add a hook that runs once, when the agent starts. This hook also runs if the agent stopped
166    /// and then starts again. The hook runs after any `before_all` hooks and before the
167    /// `before_completion` hooks.
168    pub fn on_start(&mut self, hook: impl OnStartFn + 'static) -> &mut Self {
169        self.add_hook(Hook::OnStart(Box::new(hook)))
170    }
171
172    /// Add a hook that runs before each completion.
173    pub fn before_completion(&mut self, hook: impl BeforeCompletionFn + 'static) -> &mut Self {
174        self.add_hook(Hook::BeforeCompletion(Box::new(hook)))
175    }
176
177    /// Add a hook that runs after each tool. The `Result<ToolOutput, ToolError>` is provided
178    /// as mut, so the tool output can be fully modified.
179    ///
180    /// The `ToolOutput` also references the original `ToolCall`, allowing you to match at runtime
181    /// what tool to interact with.
182    pub fn after_tool(&mut self, hook: impl AfterToolFn + 'static) -> &mut Self {
183        self.add_hook(Hook::AfterTool(Box::new(hook)))
184    }
185
186    /// Add a hook that runs before each tool. Yields an immutable reference to the `ToolCall`.
187    pub fn before_tool(&mut self, hook: impl BeforeToolFn + 'static) -> &mut Self {
188        self.add_hook(Hook::BeforeTool(Box::new(hook)))
189    }
190
191    /// Add a hook that runs after each completion, before tool invocation and/or new messages.
192    pub fn after_completion(&mut self, hook: impl AfterCompletionFn + 'static) -> &mut Self {
193        self.add_hook(Hook::AfterCompletion(Box::new(hook)))
194    }
195
196    /// Add a hook that runs after each completion, after tool invocations, right before a new loop
197    /// might start
198    pub fn after_each(&mut self, hook: impl AfterEachFn + 'static) -> &mut Self {
199        self.add_hook(Hook::AfterEach(Box::new(hook)))
200    }
201
202    /// Add a hook that runs when a new message is added to the context. Note that each tool adds a
203    /// separate message.
204    pub fn on_new_message(&mut self, hook: impl MessageHookFn + 'static) -> &mut Self {
205        self.add_hook(Hook::OnNewMessage(Box::new(hook)))
206    }
207
208    /// Set the LLM for the agent. An LLM must implement the `ChatCompletion` trait.
209    pub fn llm<LLM: ChatCompletion + Clone + 'static>(&mut self, llm: &LLM) -> &mut Self {
210        let boxed: Box<dyn ChatCompletion> = Box::new(llm.clone()) as Box<dyn ChatCompletion>;
211
212        self.llm = Some(boxed);
213        self
214    }
215
216    /// Define the available tools for the agent. Tools must implement the `Tool` trait.
217    ///
218    /// See the [tool attribute macro](`swiftide_macros::tool`) and the [tool derive macro](`swiftide_macros::Tool`)
219    /// for easy ways to create (many) tools.
220    pub fn tools<TOOL, I: IntoIterator<Item = TOOL>>(&mut self, tools: I) -> &mut Self
221    where
222        TOOL: Into<Box<dyn Tool>>,
223    {
224        self.tools = Some(
225            tools
226                .into_iter()
227                .map(Into::into)
228                .chain(Agent::default_tools())
229                .collect(),
230        );
231        self
232    }
233}
234
235impl Agent {
236    /// Build a new agent
237    pub fn builder() -> AgentBuilder {
238        AgentBuilder::default()
239    }
240}
241
242impl Agent {
243    /// Default tools for the agent that it always includes
244    fn default_tools() -> HashSet<Box<dyn Tool>> {
245        HashSet::from([Box::new(Stop::default()) as Box<dyn Tool>])
246    }
247
248    /// Run the agent with a user message. The agent will loop completions, make tool calls, until
249    /// no new messages are available.
250    #[tracing::instrument(skip_all, name = "agent.query")]
251    pub async fn query(&mut self, query: impl Into<String> + std::fmt::Debug) -> Result<()> {
252        self.run_agent(Some(query.into()), false).await
253    }
254
255    /// Run the agent with a user message once.
256    #[tracing::instrument(skip_all, name = "agent.query_once")]
257    pub async fn query_once(&mut self, query: impl Into<String> + std::fmt::Debug) -> Result<()> {
258        self.run_agent(Some(query.into()), true).await
259    }
260
261    /// Run the agent with without user message. The agent will loop completions, make tool calls, until
262    /// no new messages are available.
263    #[tracing::instrument(skip_all, name = "agent.run")]
264    pub async fn run(&mut self) -> Result<()> {
265        self.run_agent(None, false).await
266    }
267
268    /// Run the agent with without user message. The agent will loop completions, make tool calls, until
269    #[tracing::instrument(skip_all, name = "agent.run_once")]
270    pub async fn run_once(&mut self) -> Result<()> {
271        self.run_agent(None, true).await
272    }
273
274    /// Retrieve the message history of the agent
275    pub async fn history(&self) -> Vec<ChatMessage> {
276        self.context.history().await
277    }
278
279    async fn run_agent(&mut self, maybe_query: Option<String>, just_once: bool) -> Result<()> {
280        if self.state.is_running() {
281            anyhow::bail!("Agent is already running");
282        }
283
284        if self.state.is_pending() {
285            if let Some(system_prompt) = &self.system_prompt {
286                self.context
287                    .add_messages(vec![ChatMessage::System(system_prompt.render().await?)])
288                    .await;
289            }
290            for hook in self.hooks_by_type(HookTypes::BeforeAll) {
291                if let Hook::BeforeAll(hook) = hook {
292                    let span = tracing::info_span!(
293                        "hook",
294                        "otel.name" = format!("hook.{}", HookTypes::AfterTool)
295                    );
296                    tracing::info!("Calling {} hook", HookTypes::AfterTool);
297                    hook(self).instrument(span.or_current()).await?;
298                }
299            }
300        }
301
302        for hook in self.hooks_by_type(HookTypes::OnStart) {
303            if let Hook::OnStart(hook) = hook {
304                let span = tracing::info_span!(
305                    "hook",
306                    "otel.name" = format!("hook.{}", HookTypes::OnStart)
307                );
308                tracing::info!("Calling {} hook", HookTypes::OnStart);
309                hook(self).instrument(span.or_current()).await?;
310            }
311        }
312
313        self.state = state::State::Running;
314
315        if let Some(query) = maybe_query {
316            self.context.add_message(ChatMessage::User(query)).await;
317        }
318
319        let mut loop_counter = 0;
320
321        while let Some(messages) = self.context.next_completion().await {
322            if let Some(limit) = self.limit {
323                if loop_counter >= limit {
324                    tracing::warn!("Agent loop limit reached");
325                    break;
326                }
327            }
328            let result = self.run_completions(&messages).await;
329
330            if let Err(err) = result {
331                self.stop();
332                tracing::error!(error = ?err, "Agent stopped with error {err}");
333                return Err(err);
334            }
335
336            if just_once || self.state.is_stopped() {
337                break;
338            }
339            loop_counter += 1;
340        }
341
342        // If there are no new messages, ensure we update our state
343        self.stop();
344
345        Ok(())
346    }
347
348    #[tracing::instrument(skip_all, err)]
349    async fn run_completions(&mut self, messages: &[ChatMessage]) -> Result<()> {
350        debug!(
351            "Running completion for agent with {} messages",
352            messages.len()
353        );
354
355        let mut chat_completion_request = ChatCompletionRequest::builder()
356            .messages(messages)
357            .tools_spec(
358                self.tools
359                    .iter()
360                    .map(swiftide_core::Tool::tool_spec)
361                    .collect::<HashSet<_>>(),
362            )
363            .build()?;
364
365        for hook in self.hooks_by_type(HookTypes::BeforeCompletion) {
366            if let Hook::BeforeCompletion(hook) = hook {
367                let span = tracing::info_span!(
368                    "hook",
369                    "otel.name" = format!("hook.{}", HookTypes::BeforeCompletion)
370                );
371                tracing::info!("Calling {} hook", HookTypes::BeforeCompletion);
372                hook(&*self, &mut chat_completion_request)
373                    .instrument(span.or_current())
374                    .await?;
375            }
376        }
377
378        debug!(
379            "Calling LLM with the following new messages:\n {}",
380            self.context
381                .current_new_messages()
382                .await
383                .iter()
384                .map(ToString::to_string)
385                .collect::<Vec<_>>()
386                .join(",\n")
387        );
388
389        let mut response = self.llm.complete(&chat_completion_request).await?;
390
391        for hook in self.hooks_by_type(HookTypes::AfterCompletion) {
392            if let Hook::AfterCompletion(hook) = hook {
393                let span = tracing::info_span!(
394                    "hook",
395                    "otel.name" = format!("hook.{}", HookTypes::AfterCompletion)
396                );
397                tracing::info!("Calling {} hook", HookTypes::AfterCompletion);
398                hook(&*self, &mut response)
399                    .instrument(span.or_current())
400                    .await?;
401            }
402        }
403        self.add_message(ChatMessage::Assistant(
404            response.message,
405            response.tool_calls.clone(),
406        ))
407        .await?;
408
409        if let Some(tool_calls) = response.tool_calls {
410            self.invoke_tools(tool_calls).await?;
411        };
412
413        for hook in self.hooks_by_type(HookTypes::AfterEach) {
414            if let Hook::AfterEach(hook) = hook {
415                let span = tracing::info_span!(
416                    "hook",
417                    "otel.name" = format!("hook.{}", HookTypes::AfterEach)
418                );
419                tracing::info!("Calling {} hook", HookTypes::AfterEach);
420                hook(&*self).instrument(span.or_current()).await?;
421            }
422        }
423
424        Ok(())
425    }
426
427    async fn invoke_tools(&mut self, tool_calls: Vec<ToolCall>) -> Result<()> {
428        debug!("LLM returned tool calls: {:?}", tool_calls);
429
430        let mut handles = vec![];
431        for tool_call in tool_calls {
432            let Some(tool) = self.find_tool_by_name(tool_call.name()) else {
433                tracing::warn!("Tool {} not found", tool_call.name());
434                continue;
435            };
436            tracing::info!("Calling tool `{}`", tool_call.name());
437
438            let tool_args = tool_call.args().map(String::from);
439            let context: Arc<dyn AgentContext> = Arc::clone(&self.context);
440
441            for hook in self.hooks_by_type(HookTypes::BeforeTool) {
442                if let Hook::BeforeTool(hook) = hook {
443                    let span = tracing::info_span!(
444                        "hook",
445                        "otel.name" = format!("hook.{}", HookTypes::BeforeTool)
446                    );
447                    tracing::info!("Calling {} hook", HookTypes::BeforeTool);
448                    hook(&*self, &tool_call)
449                        .instrument(span.or_current())
450                        .await?;
451                }
452            }
453
454            let tool_span =
455                tracing::info_span!("tool", "otel.name" = format!("tool.{}", tool.name()));
456
457            let handle = tokio::spawn(async move {
458                    let tool_args = ArgPreprocessor::preprocess(tool_args.as_deref());
459                    let output = tool.invoke(&*context, tool_args.as_deref()).await.map_err(|e| { tracing::error!(error = %e, "Failed tool call"); e })?;
460
461                    tracing::debug!(output = output.to_string(), args = ?tool_args, tool_name = tool.name(), "Completed tool call");
462
463                    Ok(output)
464                }.instrument(tool_span.or_current()));
465
466            handles.push((handle, tool_call));
467        }
468
469        for (handle, tool_call) in handles {
470            let mut output = handle.await?;
471
472            // Invoking hooks feels too verbose and repetitive
473            for hook in self.hooks_by_type(HookTypes::AfterTool) {
474                if let Hook::AfterTool(hook) = hook {
475                    let span = tracing::info_span!(
476                        "hook",
477                        "otel.name" = format!("hook.{}", HookTypes::AfterTool)
478                    );
479                    tracing::info!("Calling {} hook", HookTypes::AfterTool);
480                    hook(&*self, &tool_call, &mut output)
481                        .instrument(span.or_current())
482                        .await?;
483                }
484            }
485
486            if let Err(error) = output {
487                if self.tool_calls_over_limit(&tool_call) {
488                    tracing::error!(
489                        "Tool call failed, retry limit reached, stopping agent: {err}",
490                        err = error
491                    );
492                    self.stop();
493                    return Err(error.into());
494                }
495                tracing::warn!(
496                    error = error.to_string(),
497                    tool_call = ?tool_call,
498                    "Tool call failed, retrying",
499                );
500                self.add_message(ChatMessage::ToolOutput(
501                    tool_call,
502                    ToolOutput::Fail(error.to_string()),
503                ))
504                .await?;
505                continue;
506            }
507
508            let output = output?;
509            self.handle_control_tools(&output);
510            self.add_message(ChatMessage::ToolOutput(tool_call, output))
511                .await?;
512        }
513
514        Ok(())
515    }
516
517    fn hooks_by_type(&self, hook_type: HookTypes) -> Vec<&Hook> {
518        self.hooks
519            .iter()
520            .filter(|h| hook_type == (*h).into())
521            .collect()
522    }
523
524    fn find_tool_by_name(&self, tool_name: &str) -> Option<Box<dyn Tool>> {
525        self.tools
526            .iter()
527            .find(|tool| tool.name() == tool_name)
528            .cloned()
529    }
530
531    // Handle any tool specific output (e.g. stop)
532    fn handle_control_tools(&mut self, output: &ToolOutput) {
533        if let ToolOutput::Stop = output {
534            tracing::warn!("Stop tool called, stopping agent");
535            self.stop();
536        }
537    }
538
539    fn tool_calls_over_limit(&mut self, tool_call: &ToolCall) -> bool {
540        let mut s = DefaultHasher::new();
541        tool_call.hash(&mut s);
542        let hash = s.finish();
543
544        if let Some(retries) = self.tool_retries_counter.get_mut(&hash) {
545            let val = *retries >= self.tool_retry_limit;
546            *retries += 1;
547            val
548        } else {
549            self.tool_retries_counter.insert(hash, 1);
550            false
551        }
552    }
553
554    #[tracing::instrument(skip_all, fields(message = message.to_string()))]
555    async fn add_message(&self, mut message: ChatMessage) -> Result<()> {
556        for hook in self.hooks_by_type(HookTypes::OnNewMessage) {
557            if let Hook::OnNewMessage(hook) = hook {
558                let span = tracing::info_span!(
559                    "hook",
560                    "otel.name" = format!("hook.{}", HookTypes::OnNewMessage)
561                );
562                if let Err(err) = hook(self, &mut message).instrument(span.or_current()).await {
563                    tracing::error!(
564                        "Error in {hooktype} hook: {err}",
565                        hooktype = HookTypes::OnNewMessage,
566                    );
567                }
568            }
569        }
570        self.context.add_message(message).await;
571        Ok(())
572    }
573
574    /// Tell the agent to stop. It will finish it's current loop and then stop.
575    pub fn stop(&mut self) {
576        self.state = state::State::Stopped;
577    }
578
579    /// Access the agent's context
580    pub fn context(&self) -> &dyn AgentContext {
581        &self.context
582    }
583
584    /// The agent is still running
585    pub fn is_running(&self) -> bool {
586        self.state.is_running()
587    }
588
589    /// The agent stopped
590    pub fn is_stopped(&self) -> bool {
591        self.state.is_stopped()
592    }
593
594    /// The agent has not (ever) started
595    pub fn is_pending(&self) -> bool {
596        self.state.is_pending()
597    }
598}
599
600#[cfg(test)]
601mod tests {
602
603    use serde::ser::Error;
604    use swiftide_core::chat_completion::errors::ToolError;
605    use swiftide_core::chat_completion::{ChatCompletionResponse, ToolCall};
606    use swiftide_core::test_utils::MockChatCompletion;
607
608    use super::*;
609    use crate::{
610        assistant, chat_request, chat_response, summary, system, tool_failed, tool_output, user,
611    };
612
613    use crate::test_utils::{MockHook, MockTool};
614
615    #[test_log::test(tokio::test)]
616    async fn test_agent_builder_defaults() {
617        // Create a prompt
618        let mock_llm = MockChatCompletion::new();
619
620        // Build the agent
621        let agent = Agent::builder().llm(&mock_llm).build().unwrap();
622
623        // Check that the context is the default context
624
625        // Check that the default tools are added
626        assert!(agent.find_tool_by_name("stop").is_some());
627
628        // Check it does not allow duplicates
629        let agent = Agent::builder()
630            .tools([Stop::default(), Stop::default()])
631            .llm(&mock_llm)
632            .build()
633            .unwrap();
634
635        assert_eq!(agent.tools.len(), 1);
636
637        // It should include the default tool if a different tool is provided
638        let agent = Agent::builder()
639            .tools([MockTool::new("mock_tool")])
640            .llm(&mock_llm)
641            .build()
642            .unwrap();
643
644        assert_eq!(agent.tools.len(), 2);
645        assert!(agent.find_tool_by_name("mock_tool").is_some());
646        assert!(agent.find_tool_by_name("stop").is_some());
647
648        assert!(agent.context().history().await.is_empty());
649    }
650
651    #[test_log::test(tokio::test)]
652    async fn test_agent_tool_calling_loop() {
653        let prompt = "Write a poem";
654        let mock_llm = MockChatCompletion::new();
655        let mock_tool = MockTool::new("mock_tool");
656
657        let chat_request = chat_request! {
658            user!("Write a poem");
659
660            tools = [mock_tool.clone()]
661        };
662
663        let mock_tool_response = chat_response! {
664            "Roses are red";
665            tool_calls = ["mock_tool"]
666
667        };
668
669        mock_llm.expect_complete(chat_request.clone(), Ok(mock_tool_response));
670
671        let chat_request = chat_request! {
672            user!("Write a poem"),
673            assistant!("Roses are red", ["mock_tool"]),
674            tool_output!("mock_tool", "Great!");
675
676            tools = [mock_tool.clone()]
677        };
678
679        let stop_response = chat_response! {
680            "Roses are red";
681            tool_calls = ["stop"]
682        };
683
684        mock_llm.expect_complete(chat_request, Ok(stop_response));
685        mock_tool.expect_invoke_ok("Great!".into(), None);
686
687        let mut agent = Agent::builder()
688            .tools([mock_tool])
689            .llm(&mock_llm)
690            .no_system_prompt()
691            .build()
692            .unwrap();
693
694        agent.query(prompt).await.unwrap();
695    }
696
697    #[test_log::test(tokio::test)]
698    async fn test_agent_tool_run_once() {
699        let prompt = "Write a poem";
700        let mock_llm = MockChatCompletion::new();
701        let mock_tool = MockTool::default();
702
703        let chat_request = chat_request! {
704            system!("My system prompt"),
705            user!("Write a poem");
706
707            tools = [mock_tool.clone()]
708        };
709
710        let mock_tool_response = chat_response! {
711            "Roses are red";
712            tool_calls = ["mock_tool"]
713
714        };
715
716        mock_tool.expect_invoke_ok("Great!".into(), None);
717        mock_llm.expect_complete(chat_request.clone(), Ok(mock_tool_response));
718
719        let mut agent = Agent::builder()
720            .tools([mock_tool])
721            .system_prompt("My system prompt")
722            .llm(&mock_llm)
723            .build()
724            .unwrap();
725
726        agent.query_once(prompt).await.unwrap();
727    }
728
729    #[test_log::test(tokio::test(flavor = "multi_thread"))]
730    async fn test_multiple_tool_calls() {
731        let prompt = "Write a poem";
732        let mock_llm = MockChatCompletion::new();
733        let mock_tool = MockTool::new("mock_tool1");
734        let mock_tool2 = MockTool::new("mock_tool2");
735
736        let chat_request = chat_request! {
737            system!("My system prompt"),
738            user!("Write a poem");
739
740
741
742            tools = [mock_tool.clone(), mock_tool2.clone()]
743        };
744
745        let mock_tool_response = chat_response! {
746            "Roses are red";
747
748            tool_calls = ["mock_tool1", "mock_tool2"]
749
750        };
751
752        dbg!(&chat_request);
753        mock_tool.expect_invoke_ok("Great!".into(), None);
754        mock_tool2.expect_invoke_ok("Great!".into(), None);
755        mock_llm.expect_complete(chat_request.clone(), Ok(mock_tool_response));
756
757        let chat_request = chat_request! {
758            system!("My system prompt"),
759            user!("Write a poem"),
760            assistant!("Roses are red", ["mock_tool1", "mock_tool2"]),
761            tool_output!("mock_tool1", "Great!"),
762            tool_output!("mock_tool2", "Great!");
763
764            tools = [mock_tool.clone(), mock_tool2.clone()]
765        };
766
767        let mock_tool_response = chat_response! {
768            "Ok!";
769
770            tool_calls = ["stop"]
771
772        };
773
774        mock_llm.expect_complete(chat_request, Ok(mock_tool_response));
775
776        let mut agent = Agent::builder()
777            .tools([mock_tool, mock_tool2])
778            .system_prompt("My system prompt")
779            .llm(&mock_llm)
780            .build()
781            .unwrap();
782
783        agent.query(prompt).await.unwrap();
784    }
785
786    #[test_log::test(tokio::test)]
787    async fn test_agent_state_machine() {
788        let prompt = "Write a poem";
789        let mock_llm = MockChatCompletion::new();
790
791        let chat_request = chat_request! {
792            user!("Write a poem");
793            tools = []
794        };
795        let mock_tool_response = chat_response! {
796            "Roses are red";
797            tool_calls = []
798        };
799
800        mock_llm.expect_complete(chat_request.clone(), Ok(mock_tool_response));
801        let mut agent = Agent::builder()
802            .llm(&mock_llm)
803            .no_system_prompt()
804            .build()
805            .unwrap();
806
807        // Agent has never run and is pending
808        assert!(agent.state.is_pending());
809        agent.query_once(prompt).await.unwrap();
810
811        // Agent is stopped, there might be more messages
812        assert!(agent.state.is_stopped());
813    }
814
815    #[test_log::test(tokio::test)]
816    async fn test_summary() {
817        let prompt = "Write a poem";
818        let mock_llm = MockChatCompletion::new();
819
820        let mock_tool_response = chat_response! {
821            "Roses are red";
822            tool_calls = []
823
824        };
825
826        let expected_chat_request = chat_request! {
827            system!("My system prompt"),
828            user!("Write a poem");
829
830            tools = []
831        };
832
833        mock_llm.expect_complete(expected_chat_request, Ok(mock_tool_response.clone()));
834
835        let mut agent = Agent::builder()
836            .system_prompt("My system prompt")
837            .llm(&mock_llm)
838            .build()
839            .unwrap();
840
841        agent.query_once(prompt).await.unwrap();
842
843        agent
844            .context
845            .add_message(ChatMessage::new_summary("Summary"))
846            .await;
847
848        let expected_chat_request = chat_request! {
849            system!("My system prompt"),
850            summary!("Summary"),
851            user!("Write another poem");
852            tools = []
853        };
854        mock_llm.expect_complete(expected_chat_request, Ok(mock_tool_response.clone()));
855
856        agent.query_once("Write another poem").await.unwrap();
857
858        agent
859            .context
860            .add_message(ChatMessage::new_summary("Summary 2"))
861            .await;
862
863        let expected_chat_request = chat_request! {
864            system!("My system prompt"),
865            summary!("Summary 2"),
866            user!("Write a third poem");
867            tools = []
868        };
869        mock_llm.expect_complete(expected_chat_request, Ok(mock_tool_response));
870
871        agent.query_once("Write a third poem").await.unwrap();
872    }
873
874    #[test_log::test(tokio::test)]
875    async fn test_agent_hooks() {
876        let mock_before_all = MockHook::new("before_all").expect_calls(1).to_owned();
877        let mock_on_start_fn = MockHook::new("on_start").expect_calls(1).to_owned();
878        let mock_before_completion = MockHook::new("before_completion")
879            .expect_calls(2)
880            .to_owned();
881        let mock_after_completion = MockHook::new("after_completion").expect_calls(2).to_owned();
882        let mock_after_each = MockHook::new("after_each").expect_calls(2).to_owned();
883        let mock_on_message = MockHook::new("on_message").expect_calls(4).to_owned();
884
885        // Once for mock tool and once for stop
886        let mock_before_tool = MockHook::new("before_tool").expect_calls(2).to_owned();
887        let mock_after_tool = MockHook::new("after_tool").expect_calls(2).to_owned();
888
889        let prompt = "Write a poem";
890        let mock_llm = MockChatCompletion::new();
891        let mock_tool = MockTool::default();
892
893        let chat_request = chat_request! {
894            user!("Write a poem");
895
896            tools = [mock_tool.clone()]
897        };
898
899        let mock_tool_response = chat_response! {
900            "Roses are red";
901            tool_calls = ["mock_tool"]
902
903        };
904
905        mock_llm.expect_complete(chat_request.clone(), Ok(mock_tool_response));
906
907        let chat_request = chat_request! {
908            user!("Write a poem"),
909            assistant!("Roses are red", ["mock_tool"]),
910            tool_output!("mock_tool", "Great!");
911
912            tools = [mock_tool.clone()]
913        };
914
915        let stop_response = chat_response! {
916            "Roses are red";
917            tool_calls = ["stop"]
918        };
919
920        mock_llm.expect_complete(chat_request, Ok(stop_response));
921        mock_tool.expect_invoke_ok("Great!".into(), None);
922
923        let mut agent = Agent::builder()
924            .tools([mock_tool])
925            .llm(&mock_llm)
926            .no_system_prompt()
927            .before_all(mock_before_all.hook_fn())
928            .on_start(mock_on_start_fn.on_start_fn())
929            .before_completion(mock_before_completion.before_completion_fn())
930            .before_tool(mock_before_tool.before_tool_fn())
931            .after_completion(mock_after_completion.after_completion_fn())
932            .after_tool(mock_after_tool.after_tool_fn())
933            .after_each(mock_after_each.hook_fn())
934            .on_new_message(mock_on_message.message_hook_fn())
935            .build()
936            .unwrap();
937
938        agent.query(prompt).await.unwrap();
939    }
940
941    #[test_log::test(tokio::test)]
942    async fn test_agent_loop_limit() {
943        let prompt = "Generate content"; // Example prompt
944        let mock_llm = MockChatCompletion::new();
945        let mock_tool = MockTool::new("mock_tool");
946
947        let chat_request = chat_request! {
948            user!(prompt);
949            tools = [mock_tool.clone()]
950        };
951        mock_tool.expect_invoke_ok("Great!".into(), None);
952
953        let mock_tool_response = chat_response! {
954            "Some response";
955            tool_calls = ["mock_tool"]
956        };
957
958        // Set expectations for the mock LLM responses
959        mock_llm.expect_complete(chat_request.clone(), Ok(mock_tool_response.clone()));
960
961        // // Response for terminating the loop
962        let stop_response = chat_response! {
963            "Final response";
964            tool_calls = ["stop"]
965        };
966
967        mock_llm.expect_complete(chat_request, Ok(stop_response));
968
969        let mut agent = Agent::builder()
970            .tools([mock_tool])
971            .llm(&mock_llm)
972            .no_system_prompt()
973            .limit(1) // Setting the loop limit to 1
974            .build()
975            .unwrap();
976
977        // Run the agent
978        agent.query(prompt).await.unwrap();
979
980        // Assert that the remaining message is still in the queue
981        let remaining = mock_llm.expectations.lock().unwrap().pop();
982        assert!(remaining.is_some());
983
984        // Assert that the agent is stopped after reaching the loop limit
985        assert!(agent.is_stopped());
986    }
987
988    #[test_log::test(tokio::test)]
989    async fn test_tool_retry_mechanism() {
990        let prompt = "Execute my tool";
991        let mock_llm = MockChatCompletion::new();
992        let mock_tool = MockTool::new("retry_tool");
993
994        // Configure mock tool to fail twice. First time is fed back to the LLM, second time is an
995        // error
996        mock_tool.expect_invoke(
997            Err(ToolError::WrongArguments(serde_json::Error::custom(
998                "missing `query`",
999            ))),
1000            None,
1001        );
1002        mock_tool.expect_invoke(
1003            Err(ToolError::WrongArguments(serde_json::Error::custom(
1004                "missing `query`",
1005            ))),
1006            None,
1007        );
1008
1009        let chat_request = chat_request! {
1010            user!(prompt);
1011            tools = [mock_tool.clone()]
1012        };
1013        let retry_response = chat_response! {
1014            "First failing attempt";
1015            tool_calls = ["retry_tool"]
1016        };
1017        mock_llm.expect_complete(chat_request.clone(), Ok(retry_response));
1018
1019        let chat_request = chat_request! {
1020            user!(prompt),
1021            assistant!("First failing attempt", ["retry_tool"]),
1022            tool_failed!("retry_tool", "arguments for tool failed to parse: missing `query`");
1023
1024            tools = [mock_tool.clone()]
1025        };
1026        let will_fail_response = chat_response! {
1027            "Finished execution";
1028            tool_calls = ["retry_tool"]
1029        };
1030        mock_llm.expect_complete(chat_request.clone(), Ok(will_fail_response));
1031
1032        let mut agent = Agent::builder()
1033            .tools([mock_tool])
1034            .llm(&mock_llm)
1035            .no_system_prompt()
1036            .tool_retry_limit(1) // The test relies on a limit of 2 retries.
1037            .build()
1038            .unwrap();
1039
1040        // Run the agent
1041        let result = agent.query(prompt).await;
1042
1043        assert!(result.is_err());
1044        assert!(result.unwrap_err().to_string().contains("missing `query`"));
1045        assert!(agent.is_stopped());
1046    }
1047}
swiftide_agents/agent.rs

swiftide_agents/
agent.rs
