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//! LLM driver trait and types.
//!
//! Abstracts over multiple LLM providers (Anthropic, OpenAI, Ollama, etc.).
use async_trait::async_trait;
use openfang_types::message::{ContentBlock, Message, StopReason, TokenUsage};
use openfang_types::tool::{ToolCall, ToolDefinition};
use serde::{Deserialize, Serialize};
use thiserror::Error;
/// Error type for LLM driver operations.
#[derive(Error, Debug)]
pub enum LlmError {
/// HTTP request failed.
#[error("HTTP error: {0}")]
Http(String),
/// API returned an error.
#[error("API error ({status}): {message}")]
Api {
/// HTTP status code.
status: u16,
/// Error message from the API.
message: String,
},
/// Rate limited — should retry after delay.
#[error("Rate limited, retry after {retry_after_ms}ms")]
RateLimited {
/// How long to wait before retrying.
retry_after_ms: u64,
},
/// Response parsing failed.
#[error("Parse error: {0}")]
Parse(String),
/// No API key configured.
#[error("Missing API key: {0}")]
MissingApiKey(String),
/// Model overloaded.
#[error("Model overloaded, retry after {retry_after_ms}ms")]
Overloaded {
/// How long to wait before retrying.
retry_after_ms: u64,
},
}
/// A request to an LLM for completion.
#[derive(Debug, Clone)]
pub struct CompletionRequest {
/// Model identifier.
pub model: String,
/// Conversation messages.
pub messages: Vec<Message>,
/// Available tools the model can use.
pub tools: Vec<ToolDefinition>,
/// Maximum tokens to generate.
pub max_tokens: u32,
/// Sampling temperature.
pub temperature: f32,
/// System prompt (extracted from messages for APIs that need it separately).
pub system: Option<String>,
/// Extended thinking configuration (if supported by the model).
pub thinking: Option<openfang_types::config::ThinkingConfig>,
}
/// A response from an LLM completion.
#[derive(Debug, Clone)]
pub struct CompletionResponse {
/// The content blocks in the response.
pub content: Vec<ContentBlock>,
/// Why the model stopped generating.
pub stop_reason: StopReason,
/// Tool calls extracted from the response.
pub tool_calls: Vec<ToolCall>,
/// Token usage statistics.
pub usage: TokenUsage,
}
impl CompletionResponse {
/// Extract text content from the response.
pub fn text(&self) -> String {
self.content
.iter()
.filter_map(|block| match block {
ContentBlock::Text { text } => Some(text.as_str()),
ContentBlock::Thinking { .. } => None,
_ => None,
})
.collect::<Vec<_>>()
.join("")
}
}
/// Events emitted during streaming LLM completion.
#[derive(Debug, Clone)]
pub enum StreamEvent {
/// Incremental text content.
TextDelta { text: String },
/// A tool use block has started.
ToolUseStart { id: String, name: String },
/// Incremental JSON input for an in-progress tool use.
ToolInputDelta { text: String },
/// A tool use block is complete with parsed input.
ToolUseEnd {
id: String,
name: String,
input: serde_json::Value,
},
/// Incremental thinking/reasoning text.
ThinkingDelta { text: String },
/// The entire response is complete.
ContentComplete {
stop_reason: StopReason,
usage: TokenUsage,
},
/// Agent lifecycle phase change (for UX indicators).
PhaseChange {
phase: String,
detail: Option<String>,
},
/// Tool execution completed with result (emitted by agent loop, not LLM driver).
ToolExecutionResult {
name: String,
result_preview: String,
is_error: bool,
},
}
/// Trait for LLM drivers.
#[async_trait]
pub trait LlmDriver: Send + Sync {
/// Send a completion request and get a response.
async fn complete(&self, request: CompletionRequest) -> Result<CompletionResponse, LlmError>;
/// Stream a completion request, sending incremental events to the channel.
/// Returns the full response when complete. Default wraps `complete()`.
async fn stream(
&self,
request: CompletionRequest,
tx: tokio::sync::mpsc::Sender<StreamEvent>,
) -> Result<CompletionResponse, LlmError> {
let response = self.complete(request).await?;
let text = response.text();
if !text.is_empty() {
let _ = tx.send(StreamEvent::TextDelta { text }).await;
}
let _ = tx
.send(StreamEvent::ContentComplete {
stop_reason: response.stop_reason,
usage: response.usage,
})
.await;
Ok(response)
}
}
/// Configuration for creating an LLM driver.
#[derive(Clone, Serialize, Deserialize)]
pub struct DriverConfig {
/// Provider name.
pub provider: String,
/// API key.
pub api_key: Option<String>,
/// Base URL override.
pub base_url: Option<String>,
}
/// SECURITY: Custom Debug impl redacts the API key.
impl std::fmt::Debug for DriverConfig {
fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
f.debug_struct("DriverConfig")
.field("provider", &self.provider)
.field("api_key", &self.api_key.as_ref().map(|_| "<redacted>"))
.field("base_url", &self.base_url)
.finish()
}
}
#[cfg(test)]
mod tests {
use super::*;
#[test]
fn test_completion_response_text() {
let response = CompletionResponse {
content: vec![
ContentBlock::Text {
text: "Hello ".to_string(),
},
ContentBlock::Text {
text: "world!".to_string(),
},
],
stop_reason: StopReason::EndTurn,
tool_calls: vec![],
usage: TokenUsage::default(),
};
assert_eq!(response.text(), "Hello world!");
}
#[test]
fn test_stream_event_clone() {
let event = StreamEvent::TextDelta {
text: "hello".to_string(),
};
let cloned = event.clone();
assert!(matches!(cloned, StreamEvent::TextDelta { text } if text == "hello"));
}
#[test]
fn test_stream_event_variants() {
let events: Vec<StreamEvent> = vec![
StreamEvent::TextDelta {
text: "hi".to_string(),
},
StreamEvent::ToolUseStart {
id: "t1".to_string(),
name: "web_search".to_string(),
},
StreamEvent::ToolInputDelta {
text: "{\"q".to_string(),
},
StreamEvent::ToolUseEnd {
id: "t1".to_string(),
name: "web_search".to_string(),
input: serde_json::json!({"query": "rust"}),
},
StreamEvent::ContentComplete {
stop_reason: StopReason::EndTurn,
usage: TokenUsage {
input_tokens: 10,
output_tokens: 5,
},
},
];
assert_eq!(events.len(), 5);
}
#[tokio::test]
async fn test_default_stream_sends_events() {
use tokio::sync::mpsc;
struct FakeDriver;
#[async_trait]
impl LlmDriver for FakeDriver {
async fn complete(
&self,
_request: CompletionRequest,
) -> Result<CompletionResponse, LlmError> {
Ok(CompletionResponse {
content: vec![ContentBlock::Text {
text: "Hello!".to_string(),
}],
stop_reason: StopReason::EndTurn,
tool_calls: vec![],
usage: TokenUsage {
input_tokens: 5,
output_tokens: 3,
},
})
}
}
let driver = FakeDriver;
let (tx, mut rx) = mpsc::channel(16);
let request = CompletionRequest {
model: "test".to_string(),
messages: vec![],
tools: vec![],
max_tokens: 100,
temperature: 0.0,
system: None,
thinking: None,
};
let response = driver.stream(request, tx).await.unwrap();
assert_eq!(response.text(), "Hello!");
// Should receive TextDelta then ContentComplete
let ev1 = rx.recv().await.unwrap();
assert!(matches!(ev1, StreamEvent::TextDelta { text } if text == "Hello!"));
let ev2 = rx.recv().await.unwrap();
assert!(matches!(
ev2,
StreamEvent::ContentComplete {
stop_reason: StopReason::EndTurn,
..
}
));
}
}