atman-runtime 1.8.0

atman flow execution runtime: evaluator, tool dispatch, provider dispatch, executor, memory stores
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
use std::sync::Arc;
use std::sync::atomic::{AtomicUsize, Ordering};

use atman_dsl::parse::parse_file;
use atman_runtime::error::RuntimeError;
use atman_runtime::event::{NodeEvent, Observable, TurnId};
use atman_runtime::message::{Message, MessageOrigin, MessagePart, MessageRole};
use atman_runtime::provider::{AssistantMessage, LlmRequest, Provider, StopReason, TokenUsage};
use atman_runtime::session::Session;
use atman_runtime::tool::BoxFut;
use atman_runtime::{Executor, Value, tools};

static TEST_CFG_LOCK: tokio::sync::Mutex<()> = tokio::sync::Mutex::const_new(());

fn register_recording_models() {
    use atman_runtime::model_registry::{ModelConfig, ModelEntry};

    let models = ["recording", "recording-full-window"]
        .into_iter()
        .map(|name| {
            (
                name.to_string(),
                ModelEntry {
                    model: name.to_string(),
                    provider: Some("recording".into()),
                    context_budget: Some(200_000),
                    ..Default::default()
                },
            )
        })
        .collect();
    atman_runtime::model_registry::set_model_config(ModelConfig {
        models,
        providers: std::collections::HashMap::new(),
        aliases: std::collections::HashMap::new(),
    });
}

/// Records the messages each LLM call receives so we can assert that
/// `context: session` actually feeds session history into the provider.
struct RecordingProvider {
    calls: AtomicUsize,
    captured_messages: std::sync::Mutex<Vec<Vec<Message>>>,
    script: Vec<Vec<MessagePart>>,
}

impl RecordingProvider {
    fn new(script: Vec<Vec<MessagePart>>) -> Self {
        Self {
            calls: AtomicUsize::new(0),
            captured_messages: std::sync::Mutex::new(Vec::new()),
            script,
        }
    }

    fn captured(&self) -> Vec<Vec<Message>> {
        self.captured_messages.lock().unwrap().clone()
    }
}

impl Provider for RecordingProvider {
    fn name(&self) -> &str {
        "recording"
    }

    fn call<'a>(&'a self, req: LlmRequest) -> BoxFut<'a, Result<AssistantMessage, RuntimeError>> {
        Box::pin(async move {
            let idx = self.calls.fetch_add(1, Ordering::SeqCst);
            self.captured_messages
                .lock()
                .unwrap()
                .push(req.messages.clone());
            let parts = self.script.get(idx).cloned().unwrap_or_else(|| {
                vec![MessagePart::Text {
                    text: "[scripted: exhausted]".into(),
                }]
            });
            let turn_id = req
                .messages
                .first()
                .map(|m| m.turn_id.clone())
                .unwrap_or_else(TurnId::now);
            Ok(AssistantMessage {
                message: Message {
                    role: MessageRole::Assistant,
                    parts,
                    turn_id,
                    origin: MessageOrigin::User,
                },
                stop_reason: StopReason::End,
                token_usage: TokenUsage::default(),
                timing: atman_runtime::provider::CallTiming::default(),
                model: String::new(),
                response_id: None,
            })
        })
    }

    fn call_streaming(&self, req: LlmRequest) -> Observable<AssistantMessage> {
        use tokio::sync::broadcast;
        use tokio_util::sync::CancellationToken;
        let (tx, events) = broadcast::channel(4);
        let cancel = CancellationToken::new();
        let idx = self.calls.fetch_add(1, Ordering::SeqCst);
        self.captured_messages
            .lock()
            .unwrap()
            .push(req.messages.clone());
        let turn_id = req
            .messages
            .first()
            .map(|m| m.turn_id.clone())
            .unwrap_or_else(TurnId::now);
        let parts = self.script.get(idx).cloned().unwrap_or_else(|| {
            vec![MessagePart::Text {
                text: "[scripted: exhausted]".into(),
            }]
        });
        let msg = AssistantMessage {
            message: Message {
                role: MessageRole::Assistant,
                parts,
                turn_id,
                origin: MessageOrigin::User,
            },
            stop_reason: StopReason::End,
            token_usage: TokenUsage::default(),
            timing: atman_runtime::provider::CallTiming::default(),
            model: String::new(),
            response_id: None,
        };
        let output: BoxFut<'static, Result<AssistantMessage, RuntimeError>> =
            Box::pin(async move {
                let _ = tx.send(NodeEvent::LlmDone { total_tokens: 0 });
                Ok(msg)
            });
        Observable {
            output,
            events,
            cancel,
        }
    }
}

const AGENT_CONTEXT_SESSION: &str = r#"
flow agent(user_prompt: string) -> string {
    _prompt_lands_via_begin_turn = user_prompt
    return subflow(agent_loop, 0)
}

flow agent_loop(iteration: int) -> string {
    when iteration >= 5 {
        return "[agent: max iterations]"
    }
    reply = llm.call(
        model: "recording",
        context: "session",
        tools: ["fs.read", "session.push"],
    )
    tool_uses = extract_tool_uses(reply)
    when is_empty(tool_uses) {
        return text_concat(reply)
    }
    tool_results = dispatch_all(tool_uses)
    session.push(tool_results)
    j = iteration + 1
    return subflow(agent_loop, j)
}
"#;

#[tokio::test(flavor = "current_thread")]
async fn context_session_feeds_session_history_into_llm_call() {
    let _cfg_lock = TEST_CFG_LOCK.lock().await;
    register_recording_models();
    let dir = tempfile::tempdir().unwrap();
    let file_path = dir.path().join("data.txt");
    tokio::fs::write(&file_path, "hello from file")
        .await
        .unwrap();

    let provider = Arc::new(RecordingProvider::new(vec![
        vec![
            MessagePart::Text {
                text: "checking".into(),
            },
            MessagePart::ToolUse {
                id: "call_0".into(),
                name: "fs.read".into(),
                input: serde_json::json!({"path": file_path.display().to_string()}),
            },
        ],
        vec![MessagePart::Text {
            text: "done: read the file".into(),
        }],
    ]));

    let session = std::sync::Arc::new(Session::open_ephemeral());
    let ex = Executor::with_events(session.sink().clone());
    tools::register_tier_zero(&ex.tools);
    ex.providers.register(provider.clone());

    let file = parse_file(AGENT_CONTEXT_SESSION).unwrap();

    let turn_id = TurnId::now();
    let user_msg = Message::user_text(turn_id.clone(), "what's in the file?");
    session.begin_turn(user_msg);

    let result = ex
        .run_in_turn(
            &file,
            "agent",
            vec![(
                "user_prompt".into(),
                Value::Str("what's in the file?".into()),
            )],
            Some(turn_id),
            Some(session.clone()),
        )
        .await;
    let result = match result {
        Ok(v) => v,
        Err(e) => {
            let msgs = session.messages();
            eprintln!("error: {e}");
            eprintln!(
                "session messages: {:?}",
                msgs.iter()
                    .map(|m| (m.role, m.text_concat()))
                    .collect::<Vec<_>>()
            );
            panic!("agent flow failed: {e}");
        }
    };
    session.end_turn();

    match result {
        Value::Str(s) => assert!(s.contains("done: read the file"), "got: {s}"),
        other => panic!("expected str, got {other:?}"),
    }

    let captured = provider.captured();
    assert_eq!(captured.len(), 2, "two LLM calls expected");

    let first = &captured[0];
    assert!(
        first.iter().any(|m| {
            m.role == MessageRole::User && m.text_concat().contains("what's in the file?")
        }),
        "first call should include the user message from session, got: {:?}",
        first
            .iter()
            .map(|m| (m.role, m.text_concat()))
            .collect::<Vec<_>>()
    );

    let second = &captured[1];
    let has_assistant_with_tool_use = second.iter().any(|m| {
        m.role == MessageRole::Assistant
            && m.parts
                .iter()
                .any(|p| matches!(p, MessagePart::ToolUse { .. }))
    });
    let has_tool_result = second.iter().any(|m| {
        m.role == MessageRole::Tool
            && m.parts
                .iter()
                .any(|p| matches!(p, MessagePart::ToolResult { .. }))
    });
    assert!(
        has_assistant_with_tool_use,
        "second call should see the assistant message with tool_use (pushed via session.push)"
    );
    assert!(
        has_tool_result,
        "second call should see the tool_result (pushed via session.push)"
    );

    let final_session = session.messages();
    assert!(
        final_session.iter().any(|m| {
            m.role == MessageRole::Tool
                && m.parts.iter().any(|p| {
                    if let MessagePart::ToolResult { content, .. } = p {
                        content.contains("hello from file")
                    } else {
                        false
                    }
                })
        }),
        "session should contain the tool_result with file content"
    );
}

const SINGLE_SESSION_CALL: &str = r#"
flow one_shot() -> string {
    reply = llm.call(model: "recording-full-window", context: "session")
    return text_concat(reply)
}
"#;

#[tokio::test(flavor = "current_thread")]
async fn context_session_sends_the_full_live_window_without_request_projection() {
    let _cfg_lock = TEST_CFG_LOCK.lock().await;
    register_recording_models();
    let provider = Arc::new(RecordingProvider::new(vec![vec![MessagePart::Text {
        text: "ok".into(),
    }]]));
    let session = Arc::new(Session::open_ephemeral());
    let ex = Executor::with_events(session.sink().clone());
    tools::register_tier_zero(&ex.tools);
    ex.providers.register(provider.clone());
    let file = parse_file(SINGLE_SESSION_CALL).unwrap();

    for i in 0..7 {
        session.append_message(
            Message::user_text(TurnId::now(), format!("session turn {i}")),
            None,
        );
    }
    let turn_id = TurnId::now();
    ex.run_in_turn(
        &file,
        "one_shot",
        vec![],
        Some(turn_id),
        Some(session.clone()),
    )
    .await
    .unwrap();

    let captured = provider.captured();
    assert_eq!(captured.len(), 1);
    let user_texts = captured[0]
        .iter()
        .filter(|message| message.role == MessageRole::User)
        .map(Message::text_concat)
        .collect::<Vec<_>>();
    assert_eq!(
        user_texts.len(),
        7,
        "context:session must use the complete live message window"
    );
    assert_eq!(
        user_texts.first().map(String::as_str),
        Some("session turn 0")
    );
    assert_eq!(
        user_texts.last().map(String::as_str),
        Some("session turn 6")
    );
}

const AGENT_CONTEXT_NONE: &str = r#"
flow one_shot(user_prompt: string) -> string {
    reply = llm.call(
        model: "recording",
        prompt: user_prompt,
    )
    return text_concat(reply)
}
"#;

#[tokio::test(flavor = "current_thread")]
async fn context_none_default_does_not_read_session_history() {
    let _cfg_lock = TEST_CFG_LOCK.lock().await;
    register_recording_models();
    let provider = Arc::new(RecordingProvider::new(vec![vec![MessagePart::Text {
        text: "ok".into(),
    }]]));

    let session = std::sync::Arc::new(Session::open_ephemeral());
    let ex = Executor::with_events(session.sink().clone());
    tools::register_tier_zero(&ex.tools);
    ex.providers.register(provider.clone());

    let file = parse_file(AGENT_CONTEXT_NONE).unwrap();

    let turn_id = TurnId::now();
    session.begin_turn(Message::user_text(
        turn_id.clone(),
        "pre-existing session msg",
    ));

    let result = ex
        .run_in_turn(
            &file,
            "one_shot",
            vec![("user_prompt".into(), Value::Str("just this prompt".into()))],
            Some(turn_id),
            Some(session.clone()),
        )
        .await
        .unwrap();
    session.end_turn();

    match result {
        Value::Str(s) => assert!(s.contains("ok"), "got: {s}"),
        other => panic!("expected str, got {other:?}"),
    }

    let captured = provider.captured();
    assert_eq!(captured.len(), 1);
    let msgs = &captured[0];
    assert_eq!(msgs.len(), 1, "context:none should send exactly 1 message");
    assert_eq!(msgs[0].role, MessageRole::User);
    assert!(
        msgs[0].text_concat().contains("just this prompt"),
        "should only contain the prompt, got: {}",
        msgs[0].text_concat()
    );
    assert!(
        !msgs[0].text_concat().contains("pre-existing session msg"),
        "session history must NOT leak into context:none calls"
    );
}