heartbit-core 2026.507.2

The Rust agentic framework — agents, tools, LLM providers, memory, evaluation.
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
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563

//! LLM provider abstractions — `LlmProvider` trait, Anthropic/Gemini/OpenRouter/OpenAI-compat backends, retry, cascade, and circuit-breaker wrappers.

pub mod anthropic;
pub mod circuit;

/// Maximum bytes read from an upstream LLM error body.
///
/// SECURITY (F-LLM-5, F-LLM-6): a hostile or compromised provider can stream
/// gigabytes of body in response to a 4xx/5xx and OOM the agent. We also
/// truncate to keep accidentally-included secrets / internal IPs from
/// flooding logs.
const ERROR_BODY_MAX_BYTES: usize = 8 << 10; // 8 KiB

/// SECURITY (F-LLM-4): hard cap on accumulated streaming text bytes per
/// response. A drip-fed `text_delta` event sequence would otherwise grow
/// unbounded.
pub(crate) const STREAM_MAX_TEXT_BYTES: usize = 16 << 20; // 16 MiB

/// SECURITY (F-LLM-4): hard cap on accumulated tool-call arguments JSON per
/// individual tool call.
pub(crate) const STREAM_MAX_TOOL_ARGS_BYTES: usize = 1 << 20; // 1 MiB

/// SECURITY (F-LLM-4): hard cap on the number of tool calls a single
/// streaming response may emit. Protects against a hostile `tool_calls[].index`
/// of `u32::MAX` triggering a multi-billion-entry Vec allocation.
pub(crate) const STREAM_MAX_TOOL_CALLS: usize = 256;

/// Build an `Error::Api` from a failed HTTP response, sanitizing auth errors.
///
/// For 401/403 responses, the body is NOT read to avoid leaking API key
/// fragments in logs. For all other statuses the response body is included
/// but capped at `ERROR_BODY_MAX_BYTES` and stripped of control characters
/// (newlines, ANSI escapes) so a hostile body cannot poison structured logs
/// (F-LLM-6).
pub(crate) async fn api_error_from_response(response: reqwest::Response) -> Error {
    use futures::TryStreamExt;
    let status = response.status().as_u16();
    let message = if status == 401 || status == 403 {
        format!("authentication failed (HTTP {status})")
    } else {
        let mut buf: Vec<u8> = Vec::with_capacity(2048);
        let mut stream = response.bytes_stream();
        let mut overflowed = false;
        loop {
            match stream.try_next().await {
                Ok(Some(chunk)) => {
                    let remaining = ERROR_BODY_MAX_BYTES.saturating_sub(buf.len());
                    if remaining == 0 {
                        overflowed = true;
                        break;
                    }
                    let take = chunk.len().min(remaining);
                    buf.extend_from_slice(&chunk[..take]);
                    if take < chunk.len() {
                        overflowed = true;
                        break;
                    }
                }
                Ok(None) => break,
                Err(e) => {
                    return Error::Api {
                        status,
                        message: format!("<body read error: {e}>"),
                    };
                }
            }
        }
        let mut text = String::from_utf8_lossy(&buf).to_string();
        // Strip control characters (CR/LF/ANSI ESC). Keep tabs and printable.
        text.retain(|c| c == '\t' || (!c.is_control() && c != '\u{1b}'));
        if overflowed {
            text.push_str("…[truncated]");
        }
        text
    };
    Error::Api { status, message }
}
pub mod cascade;
pub mod error_class;
pub mod gemini;
pub mod openai_compat;
pub mod openrouter;
pub mod pricing;
pub mod registry;
pub mod retry;
pub mod types;

use std::future::Future;
use std::pin::Pin;
use std::sync::Arc;

use crate::error::Error;
use crate::llm::types::{CompletionRequest, CompletionResponse};

/// Callback invoked with each text delta during streaming.
pub type OnText = dyn Fn(&str) + Send + Sync;

/// Decision returned by the `OnApproval` callback.
///
/// `Allow` and `Deny` behave like the previous `true`/`false` return.
/// `AlwaysAllow` and `AlwaysDeny` additionally persist the decision as a
/// learned permission rule so it survives across sessions.
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
pub enum ApprovalDecision {
    /// Allow this time.
    Allow,
    /// Deny this time.
    Deny,
    /// Allow and persist as a permission rule.
    AlwaysAllow,
    /// Deny and persist as a permission rule.
    AlwaysDeny,
}

impl ApprovalDecision {
    /// Returns `true` when the decision allows execution.
    pub fn is_allowed(self) -> bool {
        matches!(self, Self::Allow | Self::AlwaysAllow)
    }

    /// Returns `true` when the decision should be persisted.
    pub fn is_persistent(self) -> bool {
        matches!(self, Self::AlwaysAllow | Self::AlwaysDeny)
    }
}

impl From<bool> for ApprovalDecision {
    fn from(allowed: bool) -> Self {
        if allowed { Self::Allow } else { Self::Deny }
    }
}

/// Callback invoked before tool execution for human-in-the-loop approval.
///
/// Receives the list of tool calls the LLM wants to execute.
/// Returns an [`ApprovalDecision`] indicating whether to proceed.
/// `AlwaysAllow`/`AlwaysDeny` additionally persist the decision as a
/// learned permission rule.
pub type OnApproval = dyn Fn(&[crate::llm::types::ToolCall]) -> ApprovalDecision + Send + Sync;

/// Trait for LLM providers.
///
/// Uses RPITIT (`impl Future`) which means this trait is NOT dyn-compatible.
/// All consumers are generic over `P: LlmProvider`. This is intentional:
/// one provider per process, no need for trait objects.
///
/// For dynamic dispatch, use [`BoxedProvider`] which wraps any `LlmProvider`
/// behind [`DynLlmProvider`].
pub trait LlmProvider: Send + Sync {
    /// Send a completion request and wait for the full response.
    fn complete(
        &self,
        request: CompletionRequest,
    ) -> impl Future<Output = Result<CompletionResponse, Error>> + Send;

    /// Stream a completion, calling `on_text` for each text delta as it arrives.
    ///
    /// The returned `CompletionResponse` contains the full accumulated response
    /// (same as `complete()`), but text was emitted incrementally via the callback.
    ///
    /// Default: falls back to `complete()` (no incremental streaming).
    fn stream_complete(
        &self,
        request: CompletionRequest,
        on_text: &OnText,
    ) -> impl Future<Output = Result<CompletionResponse, Error>> + Send {
        let _ = on_text;
        self.complete(request)
    }

    /// Return the model identifier, if known.
    ///
    /// Used for audit trail events. Default returns `None`.
    fn model_name(&self) -> Option<&str> {
        None
    }
}

// ---------------------------------------------------------------------------
// DynLlmProvider — object-safe adapter for LlmProvider (RPITIT → dyn)
// ---------------------------------------------------------------------------

/// Object-safe version of [`LlmProvider`] for dynamic dispatch.
///
/// `LlmProvider` uses RPITIT (not dyn-compatible). This trait wraps it via
/// `Pin<Box<dyn Future>>` so providers can be stored as `Arc<dyn DynLlmProvider>`.
///
/// A blanket impl covers all `LlmProvider` types automatically.
///
/// Used by the Restate service layer (`AgentServiceImpl`) and by
/// [`BoxedProvider`] for type-erased standalone use.
pub trait DynLlmProvider: Send + Sync {
    /// Boxed-future version of [`LlmProvider::complete`] for object-safe dispatch.
    fn complete<'a>(
        &'a self,
        request: CompletionRequest,
    ) -> Pin<Box<dyn Future<Output = Result<CompletionResponse, Error>> + Send + 'a>>;

    /// Boxed-future version of [`LlmProvider::stream_complete`] for object-safe dispatch.
    fn stream_complete<'a>(
        &'a self,
        request: CompletionRequest,
        on_text: &'a OnText,
    ) -> Pin<Box<dyn Future<Output = Result<CompletionResponse, Error>> + Send + 'a>>;

    /// Return the model identifier, if known.
    fn model_name(&self) -> Option<&str>;
}

impl<P: LlmProvider> DynLlmProvider for P {
    fn complete<'a>(
        &'a self,
        request: CompletionRequest,
    ) -> Pin<Box<dyn Future<Output = Result<CompletionResponse, Error>> + Send + 'a>> {
        Box::pin(LlmProvider::complete(self, request))
    }

    fn stream_complete<'a>(
        &'a self,
        request: CompletionRequest,
        on_text: &'a OnText,
    ) -> Pin<Box<dyn Future<Output = Result<CompletionResponse, Error>> + Send + 'a>> {
        Box::pin(LlmProvider::stream_complete(self, request, on_text))
    }

    fn model_name(&self) -> Option<&str> {
        LlmProvider::model_name(self)
    }
}

// ---------------------------------------------------------------------------
// BoxedProvider — type-erased LlmProvider via DynLlmProvider
// ---------------------------------------------------------------------------

/// Type-erased LLM provider for use when dynamic dispatch is needed.
///
/// Wraps any [`LlmProvider`] behind `Box<dyn DynLlmProvider>`. Implements
/// `LlmProvider` itself, so it can be used with `AgentRunner<BoxedProvider>`
/// and `Orchestrator<BoxedProvider>`, eliminating the need for generic code
/// at the call site.
///
/// # Example
///
/// ```ignore
/// let provider = BoxedProvider::new(AnthropicProvider::new(key, model));
/// let runner = AgentRunner::builder(Arc::new(provider))
///     .name("agent")
///     .build()?;
/// ```
pub struct BoxedProvider(Box<dyn DynLlmProvider>);

impl BoxedProvider {
    /// Create a type-erased provider from any concrete `LlmProvider`.
    pub fn new<P: LlmProvider + 'static>(provider: P) -> Self {
        Self(Box::new(provider))
    }

    /// Create a type-erased provider from an `Arc<P>`.
    ///
    /// Useful when the provider is already behind an `Arc` (e.g., shared between
    /// the orchestrator and sub-agents) and needs to be converted to `BoxedProvider`
    /// for type erasure without consuming the original.
    pub fn from_arc<P: LlmProvider + 'static>(provider: Arc<P>) -> Self {
        /// Internal adapter: delegates to the `Arc<P>` inner provider.
        struct ArcAdapter<P>(Arc<P>);

        impl<P: LlmProvider> LlmProvider for ArcAdapter<P> {
            async fn complete(
                &self,
                request: CompletionRequest,
            ) -> Result<CompletionResponse, Error> {
                self.0.complete(request).await
            }

            async fn stream_complete(
                &self,
                request: CompletionRequest,
                on_text: &OnText,
            ) -> Result<CompletionResponse, Error> {
                self.0.stream_complete(request, on_text).await
            }

            fn model_name(&self) -> Option<&str> {
                self.0.model_name()
            }
        }

        Self(Box::new(ArcAdapter(provider)))
    }
}

impl LlmProvider for BoxedProvider {
    async fn complete(&self, request: CompletionRequest) -> Result<CompletionResponse, Error> {
        self.0.complete(request).await
    }

    async fn stream_complete(
        &self,
        request: CompletionRequest,
        on_text: &OnText,
    ) -> Result<CompletionResponse, Error> {
        self.0.stream_complete(request, on_text).await
    }

    fn model_name(&self) -> Option<&str> {
        self.0.model_name()
    }
}

#[cfg(test)]
mod tests {
    use super::*;
    use crate::llm::types::{ContentBlock, Message, StopReason, TokenUsage};
    use std::sync::{Arc, Mutex};

    struct FakeProvider;

    impl LlmProvider for FakeProvider {
        async fn complete(&self, _request: CompletionRequest) -> Result<CompletionResponse, Error> {
            Ok(CompletionResponse {
                content: vec![ContentBlock::Text {
                    text: "fake".into(),
                }],
                stop_reason: StopReason::EndTurn,
                usage: TokenUsage::default(),
                model: None,
            })
        }
    }

    struct StreamingFakeProvider;

    impl LlmProvider for StreamingFakeProvider {
        async fn complete(&self, _request: CompletionRequest) -> Result<CompletionResponse, Error> {
            panic!("should call stream_complete, not complete");
        }

        async fn stream_complete(
            &self,
            _request: CompletionRequest,
            on_text: &OnText,
        ) -> Result<CompletionResponse, Error> {
            on_text("hello");
            on_text(" world");
            Ok(CompletionResponse {
                content: vec![ContentBlock::Text {
                    text: "hello world".into(),
                }],
                stop_reason: StopReason::EndTurn,
                usage: TokenUsage::default(),
                model: None,
            })
        }
    }

    fn test_request() -> CompletionRequest {
        CompletionRequest {
            system: String::new(),
            messages: vec![Message::user("test")],
            tools: vec![],
            max_tokens: 100,
            tool_choice: None,
            reasoning_effort: None,
        }
    }

    #[test]
    fn dyn_llm_provider_wraps_provider() {
        let provider = FakeProvider;
        let dyn_provider: &dyn DynLlmProvider = &provider;
        let _ = dyn_provider;
    }

    #[tokio::test]
    async fn boxed_provider_delegates_complete() {
        let provider = BoxedProvider::new(FakeProvider);
        // Disambiguate: BoxedProvider implements both LlmProvider and DynLlmProvider
        let response = LlmProvider::complete(&provider, test_request())
            .await
            .unwrap();
        assert_eq!(response.text(), "fake");
    }

    #[tokio::test]
    async fn boxed_provider_delegates_stream_complete() {
        let provider = BoxedProvider::new(StreamingFakeProvider);
        let received = Arc::new(Mutex::new(Vec::<String>::new()));
        let received_clone = received.clone();
        let on_text: &OnText = &move |text: &str| {
            received_clone
                .lock()
                .expect("test lock")
                .push(text.to_string());
        };

        let response = LlmProvider::stream_complete(&provider, test_request(), on_text)
            .await
            .unwrap();
        assert_eq!(response.text(), "hello world");

        let texts = received.lock().expect("test lock");
        assert_eq!(*texts, vec!["hello", " world"]);
    }

    #[test]
    fn boxed_provider_is_send_sync() {
        fn assert_send_sync<T: Send + Sync>() {}
        assert_send_sync::<BoxedProvider>();
    }

    #[tokio::test]
    async fn boxed_provider_default_stream_falls_back_to_complete() {
        // FakeProvider only implements complete; stream_complete should fall back
        let provider = BoxedProvider::new(FakeProvider);
        let on_text: &OnText = &|_| {};
        let response = LlmProvider::stream_complete(&provider, test_request(), on_text)
            .await
            .unwrap();
        assert_eq!(response.text(), "fake");
    }

    #[tokio::test]
    async fn boxed_provider_from_arc_delegates_complete() {
        let provider = Arc::new(FakeProvider);
        let boxed = BoxedProvider::from_arc(provider);
        let response = LlmProvider::complete(&boxed, test_request()).await.unwrap();
        assert_eq!(response.text(), "fake");
    }

    #[tokio::test]
    async fn boxed_provider_from_arc_delegates_stream_complete() {
        let provider = Arc::new(StreamingFakeProvider);
        let boxed = BoxedProvider::from_arc(provider);
        let received = Arc::new(Mutex::new(Vec::<String>::new()));
        let received_clone = received.clone();
        let on_text: &OnText = &move |text: &str| {
            received_clone
                .lock()
                .expect("test lock")
                .push(text.to_string());
        };
        let response = LlmProvider::stream_complete(&boxed, test_request(), on_text)
            .await
            .unwrap();
        assert_eq!(response.text(), "hello world");
        let texts = received.lock().expect("test lock");
        assert_eq!(*texts, vec!["hello", " world"]);
    }

    #[test]
    fn model_name_default_is_none() {
        let provider = FakeProvider;
        assert!(LlmProvider::model_name(&provider).is_none());
    }

    #[test]
    fn boxed_provider_preserves_model_name() {
        struct NamedProvider;
        impl LlmProvider for NamedProvider {
            async fn complete(
                &self,
                _request: CompletionRequest,
            ) -> Result<CompletionResponse, Error> {
                unimplemented!()
            }
            fn model_name(&self) -> Option<&str> {
                Some("test-model")
            }
        }
        let boxed = BoxedProvider::new(NamedProvider);
        assert_eq!(LlmProvider::model_name(&boxed), Some("test-model"));
    }

    #[test]
    fn boxed_provider_from_arc_preserves_model_name() {
        struct NamedProvider;
        impl LlmProvider for NamedProvider {
            async fn complete(
                &self,
                _request: CompletionRequest,
            ) -> Result<CompletionResponse, Error> {
                unimplemented!()
            }
            fn model_name(&self) -> Option<&str> {
                Some("arc-model")
            }
        }
        let boxed = BoxedProvider::from_arc(Arc::new(NamedProvider));
        assert_eq!(LlmProvider::model_name(&boxed), Some("arc-model"));
    }

    #[tokio::test]
    async fn boxed_provider_from_arc_shares_underlying_provider() {
        // Verify from_arc shares the underlying provider via Arc (not a copy)
        let call_count = Arc::new(std::sync::atomic::AtomicUsize::new(0));
        struct CountingProvider(Arc<std::sync::atomic::AtomicUsize>);
        impl LlmProvider for CountingProvider {
            async fn complete(
                &self,
                _request: CompletionRequest,
            ) -> Result<CompletionResponse, crate::error::Error> {
                self.0.fetch_add(1, std::sync::atomic::Ordering::Relaxed);
                Ok(CompletionResponse {
                    content: vec![ContentBlock::Text {
                        text: "counted".into(),
                    }],
                    stop_reason: StopReason::EndTurn,
                    usage: TokenUsage::default(),
                    model: None,
                })
            }
        }

        let inner = Arc::new(CountingProvider(call_count.clone()));
        let boxed1 = BoxedProvider::from_arc(inner.clone());
        let boxed2 = BoxedProvider::from_arc(inner);

        LlmProvider::complete(&boxed1, test_request())
            .await
            .unwrap();
        LlmProvider::complete(&boxed2, test_request())
            .await
            .unwrap();

        assert_eq!(
            call_count.load(std::sync::atomic::Ordering::Relaxed),
            2,
            "both boxed providers should share the same underlying provider"
        );
    }

    // --- ApprovalDecision ---

    #[test]
    fn approval_decision_from_true() {
        let decision = ApprovalDecision::from(true);
        assert_eq!(decision, ApprovalDecision::Allow);
        assert!(decision.is_allowed());
        assert!(!decision.is_persistent());
    }

    #[test]
    fn approval_decision_from_false() {
        let decision = ApprovalDecision::from(false);
        assert_eq!(decision, ApprovalDecision::Deny);
        assert!(!decision.is_allowed());
        assert!(!decision.is_persistent());
    }

    #[test]
    fn approval_decision_always_allow() {
        let decision = ApprovalDecision::AlwaysAllow;
        assert!(decision.is_allowed());
        assert!(decision.is_persistent());
    }

    #[test]
    fn approval_decision_always_deny() {
        let decision = ApprovalDecision::AlwaysDeny;
        assert!(!decision.is_allowed());
        assert!(decision.is_persistent());
    }
}