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
//! Items yielded by the model stream: content, reasoning, and the terminating finish reason.
use ContentBlock;
/// Why the model stopped generating.
///
/// OpenAI-compatible servers report this as a `finish_reason` string on the final streaming
/// chunk. The SDK maps the well-known values onto variants and preserves anything else
/// verbatim in [`FinishReason::Other`], so a provider-specific reason is never silently
/// flattened into a generic one.
///
/// # Why `Unspecified` exists
///
/// `finish_reason` is optional in practice. llama.cpp, vLLM, and several local gateways stream
/// content and then close the connection (or send `data: [DONE]`) with `finish_reason` still
/// null. Reporting that as [`FinishReason::Stop`] would claim the model finished cleanly when
/// the SDK has no evidence either way, so it is reported as
/// [`FinishReason::Unspecified`] instead — a distinct, checkable state.
///
/// # Examples
///
/// The distinction that matters for callers parsing structured output:
///
/// ```rust
/// use open_agent::FinishReason;
///
/// // A truncated response is worth retrying with a larger budget.
/// assert!(FinishReason::Length.is_truncated());
/// // A clean stop that produced unparseable output is a model behaviour problem.
/// assert!(!FinishReason::Stop.is_truncated());
/// ```
/// One item in the stream returned by [`query()`](crate::query).
///
/// Before 0.8.0 the stream yielded bare [`ContentBlock`]s, which left no room for anything
/// that is not content — most importantly the reason generation stopped. `StreamEvent` makes
/// that explicit: content arrives as [`StreamEvent::Block`], and the stream always ends with
/// exactly one [`StreamEvent::Finish`].
///
/// # Guarantees
///
/// - Exactly one [`StreamEvent::Finish`] is emitted per stream, and it is the final event.
/// - [`StreamEvent::Reasoning`] is emitted only when
/// [`AgentOptions::include_reasoning`](crate::AgentOptions::include_reasoning) is enabled,
/// and never carries text that also appears in a [`ContentBlock::Text`].
///
/// The enum is `#[non_exhaustive]`: future channels can be added without another breaking
/// release, so match with a `_` arm.
///
/// # Examples
///
/// ```rust,no_run
/// use futures::StreamExt;
/// use open_agent::{AgentOptions, ContentBlock, FinishReason, StreamEvent, query};
///
/// # async fn example() -> Result<(), Box<dyn std::error::Error>> {
/// let options = AgentOptions::builder()
/// .model("deepseek-reasoner")
/// .base_url("http://localhost:1234/v1")
/// .build()?;
///
/// let mut answer = String::new();
/// let mut stream = query("Reply with JSON.", &options).await?;
///
/// while let Some(event) = stream.next().await {
/// match event? {
/// StreamEvent::Block(ContentBlock::Text(text)) => answer.push_str(&text.text),
/// StreamEvent::Finish(FinishReason::Length) => {
/// // Truncated at the token cap: retry with a larger budget rather than
/// // treating the unparseable body as a refusal.
/// }
/// _ => {}
/// }
/// }
/// # Ok(())
/// # }
/// ```