yoagent 0.8.4

Simple, effective agent loop with tool execution and event streaming
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
//! Sub-agent tool — delegates tasks to a child agent loop.
//!
//! The `SubAgentTool` implements `AgentTool` and internally runs `agent_loop()`
//! with its own system prompt, tools, and provider. The parent LLM invokes it
//! like any other tool, passing a natural-language `task` string.
//!
//! # Design
//!
//! - **Context isolation**: each invocation starts a fresh conversation
//! - **Nesting supported**: sub-agents can contain other SubAgentTools for recursive delegation (use `with_max_turns()` to bound depth)
//! - **Cancellation propagation**: the parent's cancel token is forwarded
//! - **Event forwarding**: sub-agent events stream to the parent via `on_update`
//!
//! # Example
//!
//! ```rust,no_run
//! use yoagent::sub_agent::SubAgentTool;
//! use yoagent::provider::AnthropicProvider;
//! use std::sync::Arc;
//!
//! let researcher = SubAgentTool::new("researcher", Arc::new(AnthropicProvider))
//!     .with_description("Searches codebases and documents")
//!     .with_system_prompt("You are a research assistant.")
//!     .with_model("claude-sonnet-4-20250514")
//!     .with_api_key("sk-...");
//! ```

use crate::agent_loop::{agent_loop, AgentLoopConfig};
use crate::context::ExecutionLimits;
use crate::provider::model::ModelConfig;
use crate::provider::StreamProvider;
use crate::shared_state::SharedState;
use crate::tools::shared_state_tool::SharedStateTool;
use crate::types::*;
use std::sync::Arc;
use tokio::sync::mpsc;

/// Default max turns for sub-agents (prevents runaway execution).
const DEFAULT_MAX_TURNS: usize = 10;

/// A tool that delegates work to a child agent loop.
///
/// When the parent LLM calls this tool, it spawns a fresh `agent_loop()` with
/// its own system prompt, tools, and provider. The sub-agent runs to completion
/// and its final text output is returned as the tool result.
pub struct SubAgentTool {
    tool_name: String,
    tool_description: String,
    system_prompt: String,
    skills_prompt: String,
    model: String,
    api_key: String,
    provider: Arc<dyn StreamProvider>,
    tools: Vec<Arc<dyn AgentTool>>,
    thinking_level: ThinkingLevel,
    max_tokens: Option<u32>,
    cache_config: CacheConfig,
    tool_execution: ToolExecutionStrategy,
    retry_config: crate::retry::RetryConfig,
    max_turns: usize,
    shared_state: Option<SharedState>,
    turn_delay: Option<std::time::Duration>,
    model_config: Option<ModelConfig>,
}

impl SubAgentTool {
    /// Create a new sub-agent tool with a name and provider.
    pub fn new(name: impl Into<String>, provider: Arc<dyn StreamProvider>) -> Self {
        let name = name.into();
        Self {
            tool_description: format!("Delegate a task to the '{}' sub-agent", name),
            tool_name: name,
            system_prompt: String::new(),
            skills_prompt: String::new(),
            model: String::new(),
            api_key: String::new(),
            provider,
            tools: Vec::new(),
            thinking_level: ThinkingLevel::Off,
            max_tokens: None,
            cache_config: CacheConfig::default(),
            tool_execution: ToolExecutionStrategy::default(),
            retry_config: crate::retry::RetryConfig::default(),
            max_turns: DEFAULT_MAX_TURNS,
            shared_state: None,
            turn_delay: None,
            model_config: None,
        }
    }

    pub fn with_description(mut self, desc: impl Into<String>) -> Self {
        self.tool_description = desc.into();
        self
    }

    pub fn with_system_prompt(mut self, prompt: impl Into<String>) -> Self {
        self.system_prompt = prompt.into();
        self
    }

    /// Attach a skill set so the sub-agent sees the skills index.
    ///
    /// Mirrors [`Agent::with_skills`](crate::agent::Agent::with_skills): the skills
    /// index is formatted as XML per the [AgentSkills standard](https://agentskills.io)
    /// and appended to the sub-agent's system prompt at dispatch time. The sub-agent
    /// can then read individual SKILL.md files via the `read_file` tool when it
    /// decides a skill is relevant (make sure the sub-agent has such a tool).
    pub fn with_skills(mut self, skills: crate::skills::SkillSet) -> Self {
        self.skills_prompt = skills.format_for_prompt();
        self
    }

    pub fn with_model(mut self, model: impl Into<String>) -> Self {
        self.model = model.into();
        self
    }

    pub fn with_api_key(mut self, key: impl Into<String>) -> Self {
        self.api_key = key.into();
        self
    }

    pub fn with_tools(mut self, tools: Vec<Arc<dyn AgentTool>>) -> Self {
        self.tools = tools;
        self
    }

    pub fn with_thinking(mut self, level: ThinkingLevel) -> Self {
        self.thinking_level = level;
        self
    }

    pub fn with_max_tokens(mut self, max: u32) -> Self {
        self.max_tokens = Some(max);
        self
    }

    pub fn with_cache_config(mut self, config: CacheConfig) -> Self {
        self.cache_config = config;
        self
    }

    pub fn with_tool_execution(mut self, strategy: ToolExecutionStrategy) -> Self {
        self.tool_execution = strategy;
        self
    }

    pub fn with_retry_config(mut self, config: crate::retry::RetryConfig) -> Self {
        self.retry_config = config;
        self
    }

    pub fn with_max_turns(mut self, max: usize) -> Self {
        self.max_turns = max;
        self
    }

    /// Attach a shared key-value store. Sub-agents get a `shared_state` tool
    /// to read/write variables. The parent can also read/write programmatically
    /// via the `SharedState` handle.
    pub fn with_shared_state(mut self, state: SharedState) -> Self {
        self.shared_state = Some(state);
        self
    }

    /// Add an inter-turn delay to throttle API requests.
    /// Useful when using OAuth tokens or providers with low rate limits.
    /// The delay is applied before each turn except the first.
    pub fn with_turn_delay(mut self, delay: std::time::Duration) -> Self {
        self.turn_delay = Some(delay);
        self
    }

    /// Set the model configuration for multi-provider support.
    /// Required for non-Anthropic providers (OpenAI-compat, Google, etc.)
    /// to specify base URL, compat flags, and other provider-specific settings.
    pub fn with_model_config(mut self, config: ModelConfig) -> Self {
        self.model_config = Some(config);
        self
    }
}

/// Thin adapter: wraps `Arc<dyn AgentTool>` so it can be placed in a
/// `Vec<Box<dyn AgentTool>>` (required by `AgentContext`).
struct ArcToolWrapper(Arc<dyn AgentTool>);

#[async_trait::async_trait]
impl AgentTool for ArcToolWrapper {
    fn name(&self) -> &str {
        self.0.name()
    }
    fn label(&self) -> &str {
        self.0.label()
    }
    fn description(&self) -> &str {
        self.0.description()
    }
    fn parameters_schema(&self) -> serde_json::Value {
        self.0.parameters_schema()
    }
    async fn execute(
        &self,
        params: serde_json::Value,
        ctx: ToolContext,
    ) -> Result<ToolResult, ToolError> {
        self.0.execute(params, ctx).await
    }
}

#[async_trait::async_trait]
impl AgentTool for SubAgentTool {
    fn name(&self) -> &str {
        &self.tool_name
    }

    fn label(&self) -> &str {
        &self.tool_name
    }

    fn description(&self) -> &str {
        &self.tool_description
    }

    fn parameters_schema(&self) -> serde_json::Value {
        serde_json::json!({
            "type": "object",
            "properties": {
                "task": {
                    "type": "string",
                    "description": "The task to delegate to this sub-agent"
                }
            },
            "required": ["task"]
        })
    }

    async fn execute(
        &self,
        params: serde_json::Value,
        ctx: ToolContext,
    ) -> Result<ToolResult, ToolError> {
        let cancel = ctx.cancel;
        let on_update = ctx.on_update;
        let on_progress = ctx.on_progress;
        // Extract the task parameter
        let task = params
            .get("task")
            .and_then(|v| v.as_str())
            .ok_or_else(|| ToolError::InvalidArgs("Missing required 'task' parameter".into()))?
            .to_string();

        // Build tool list from Arc wrappers
        let mut tools: Vec<Box<dyn AgentTool>> = self
            .tools
            .iter()
            .map(|t| Box::new(ArcToolWrapper(Arc::clone(t))) as Box<dyn AgentTool>)
            .collect();

        // Append the skills index (if any) so the sub-agent can discover skills.
        let mut system_prompt = self.system_prompt.clone();
        if !self.skills_prompt.is_empty() {
            if system_prompt.is_empty() {
                system_prompt = self.skills_prompt.clone();
            } else {
                system_prompt = format!("{}\n\n{}", system_prompt, self.skills_prompt);
            }
        }

        // Inject SharedStateTool when shared state is configured
        if let Some(ref state) = self.shared_state {
            tools.push(Box::new(SharedStateTool::new(state.clone())));
            let summary = state.summary().await;
            system_prompt.push_str(&format!(
                "\n\n## Shared State\nYou have access to a shared variable store via the `shared_state` tool.\nAvailable: {}",
                summary
            ));
        }

        // Fresh context for the sub-agent
        let mut context = AgentContext {
            system_prompt,
            messages: Vec::new(),
            tools,
        };

        // Config with Arc'd provider
        let config = AgentLoopConfig {
            provider: self.provider.clone(),
            model: self.model.clone(),
            api_key: self.api_key.clone(),
            thinking_level: self.thinking_level,
            max_tokens: self.max_tokens,
            temperature: None,
            model_config: self.model_config.clone(),
            convert_to_llm: None,
            transform_context: None,
            get_steering_messages: None,
            get_follow_up_messages: None,
            context_config: None,
            compaction_strategy: None,
            execution_limits: Some(ExecutionLimits {
                max_turns: self.max_turns,
                // Generous token/duration limits — turn limit is the primary guard
                max_total_tokens: 1_000_000,
                max_duration: std::time::Duration::from_secs(300),
            }),
            cache_config: self.cache_config.clone(),
            tool_execution: self.tool_execution.clone(),
            retry_config: self.retry_config.clone(),
            before_turn: None,
            after_turn: None,
            on_error: None,
            input_filters: vec![],
            turn_delay: self.turn_delay,
        };

        // Channel for sub-agent events
        let (tx, mut rx) = mpsc::unbounded_channel();

        // Forward sub-agent events to parent via on_update and on_progress callbacks
        let forward_handle = if on_update.is_some() || on_progress.is_some() {
            let tool_name = self.tool_name.clone();
            Some(tokio::spawn(async move {
                while let Some(event) = rx.recv().await {
                    // Forward progress messages via on_progress
                    if let AgentEvent::ProgressMessage { text, .. } = &event {
                        if let Some(ref cb) = on_progress {
                            cb(text.clone());
                        }
                    }

                    // Convert interesting events to ToolResult updates for the parent
                    if let Some(ref on_update) = on_update {
                        let update_text = match &event {
                            AgentEvent::MessageUpdate {
                                delta: StreamDelta::Text { delta },
                                ..
                            } => Some(delta.clone()),
                            AgentEvent::ToolExecutionStart { tool_name, .. } => {
                                Some(format!("[sub-agent calling tool: {}]", tool_name))
                            }
                            _ => None,
                        };

                        if let Some(text) = update_text {
                            on_update(ToolResult {
                                content: vec![Content::Text { text }],
                                details: serde_json::json!({ "sub_agent": tool_name }),
                            });
                        }
                    }
                }
            }))
        } else {
            None
        };

        // Run the sub-agent loop
        let prompt = AgentMessage::Llm(Message::user(task));
        let new_messages = agent_loop(vec![prompt], &mut context, &config, tx, cancel).await;

        // Wait for event forwarding to complete
        if let Some(handle) = forward_handle {
            let _ = handle.await;
        }

        // Check if the last message was an error
        if let Some(error_msg) = extract_error(&new_messages) {
            return Err(ToolError::Failed(format!(
                "Sub-agent '{}' failed: {}",
                self.tool_name, error_msg
            )));
        }

        // Extract final assistant text from the returned messages
        let result_text = extract_final_text(&new_messages);

        // Include full sub-agent conversation in details for debugging
        let details = serde_json::json!({
            "sub_agent": self.tool_name,
            "turns": new_messages.len(),
        });

        Ok(ToolResult {
            content: vec![Content::Text { text: result_text }],
            details,
        })
    }
}

/// Check if the last assistant message was an error, return the error message.
fn extract_error(messages: &[AgentMessage]) -> Option<String> {
    for msg in messages.iter().rev() {
        if let AgentMessage::Llm(Message::Assistant {
            stop_reason,
            error_message,
            ..
        }) = msg
        {
            if *stop_reason == StopReason::Error {
                return Some(
                    error_message
                        .clone()
                        .unwrap_or_else(|| "Unknown error".into()),
                );
            }
        }
    }
    None
}

/// Extract the final assistant text from agent messages.
/// Collects text from the last assistant message, or returns a fallback.
fn extract_final_text(messages: &[AgentMessage]) -> String {
    for msg in messages.iter().rev() {
        if let AgentMessage::Llm(Message::Assistant { content, .. }) = msg {
            let texts: Vec<&str> = content
                .iter()
                .filter_map(|c| match c {
                    Content::Text { text } if !text.is_empty() => Some(text.as_str()),
                    _ => None,
                })
                .collect();
            if !texts.is_empty() {
                return texts.join("\n");
            }
        }
    }
    "(sub-agent produced no text output)".to_string()
}