daimon 0.19.0

A Rust-native AI agent framework
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
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
# Daimon Architecture

This document describes the deep architectural design of the Daimon Rust AI agent framework. It is written for senior Rust developers who want to understand the design before contributing.

---

## 1. Design Philosophy

### Why Rust for Agents

Daimon is built in Rust to deliver:

- **Zero-cost abstractions** — Trait-based polymorphism compiles to static dispatch when the concrete type is known. The `Erased*` pattern adds dynamic dispatch only when needed; users choose per call site.
- **Trait-based plugin system** — New LLM providers, vector stores, and task brokers are added by implementing traits. No reflection, no runtime registration. Compile-time composition.
- **Async-first** — All I/O is async. The ReAct loop, tool execution, and streaming are built on `async fn` and `Future`. No blocking calls in the hot path.
- **No runtime overhead from unused features** — Every provider, storage backend, and protocol is behind a Cargo feature flag. The core framework compiles with zero optional dependencies. Users pay only for what they enable.
- **Memory safety without GC** — Long-running agents with tool loops and streaming avoid GC pauses. Ownership and borrowing prevent data races and leaks.
- **Single binary deployment** — Compile one binary with the features you need. No Python runtime, no Node.js, no interpreter startup.

### Core Principles

1. **daimon-core is the plugin boundary** — Provider and plugin crates depend only on `daimon-core`. The main `daimon` crate re-exports everything. This keeps the interface stable and avoids circular dependencies.
2. **Traits over concrete types** — The agent accepts `SharedModel`, `SharedMemory`, `Arc<dyn ErasedTool>`. Implementors are decoupled from the orchestration layer.
3. **Builder pattern for complex types**`Agent::builder()`, `ChainBuilder`, `PgVectorStoreBuilder`. Validation at `.build()` time; invalid config fails to compile or returns `Result`.
4. **Structured errors** — Single `DaimonError` enum. Provider crates map transport errors to `DaimonError::Model(String)` or `DaimonError::Other(String)`.

---

## 2. Workspace Layout

Daimon is a Cargo workspace (mono-repo). The root `Cargo.toml` defines the main `daimon` package and workspace members.

```
lexmata-agent-framework/
├── Cargo.toml              # Workspace root + daimon package
├── daimon-core/            # Plugin interface traits + shared types
├── daimon-macros/          # #[tool_fn] proc macro
├── daimon-provider-bedrock/
├── daimon-provider-gemini/
├── daimon-provider-azure/
├── daimon-plugin-pgvector/
├── daimon-plugin-opensearch/
├── src/                    # Main daimon crate source
│   ├── agent/              # Agent builder, ReAct loop, multi-agent patterns
│   ├── model/              # Model trait re-exports + OpenAI, Anthropic, Ollama
│   ├── tool/               # Tool trait, registry, retry
│   ├── memory/             # Memory trait + SQLite, Redis, sliding window
│   ├── stream/             # ResponseStream, StreamEvent
│   ├── hooks.rs            # AgentHook lifecycle callbacks
│   ├── middleware/         # Request/response/tool_call interception
│   ├── guardrails/         # InputGuardrail, OutputGuardrail
│   ├── retriever/          # Retriever, KnowledgeBase, RAG tool
│   ├── checkpoint/         # Checkpoint trait + file, Redis, NATS KV
│   ├── cost/               # CostModel, CostTracker, budget enforcement
│   ├── orchestration/       # Chain, Graph, DAG, Workflow
│   ├── distributed/        # TaskBroker implementations (Redis, NATS, AMQP, gRPC)
│   ├── mcp/                # Model Context Protocol client & server
│   ├── a2a/                # Google Agent-to-Agent protocol
│   ├── server/             # HTTP server (axum) for REST API
│   ├── eval/               # EvalScenario, EvalRunner, Scorer
│   └── ...
├── examples/
├── tests/
└── benches/
```

### Crate Responsibilities

| Crate | Purpose |
|-------|---------|
| **daimon** | Main crate. Agent runtime, orchestration, RAG, memory, hooks, tools, MCP, A2A, eval, server. Re-exports from daimon-core and provider/plugin crates. Users depend on `daimon` with feature flags. |
| **daimon-core** | Plugin interface. Defines `Model`, `EmbeddingModel`, `VectorStore`, `TaskBroker` traits and shared types: `Message`, `ChatRequest`, `ChatResponse`, `Document`, `ToolCall`, `StreamEvent`, `DaimonError`. Zero optional deps. |
| **daimon-macros** | `#[tool_fn]` proc macro. Derives `Tool` from async functions with JSON Schema from parameter types. |
| **daimon-provider-*** | LLM provider crates. Each implements `Model` and `EmbeddingModel` from daimon-core. Depend only on daimon-core + HTTP/SDK. |
| **daimon-plugin-*** | Storage plugins. Each implements `VectorStore` from daimon-core. Depend only on daimon-core + DB client. |

---

## 3. The Plugin Boundary

```
┌─────────────────────────────────────────────────────────────────────────┐
│                         daimon (main crate)                               │
│  Agent, Chain, Graph, Memory, Hooks, MCP, Server, Eval, ...              │
│  Re-exports: daimon_core::*, daimon_provider_*::*, daimon_plugin_*::*     │
└─────────────────────────────────────────────────────────────────────────┘
                                    │ depends on
┌─────────────────────────────────────────────────────────────────────────┐
│                         daimon-core                                       │
│  Traits: Model, EmbeddingModel, VectorStore, TaskBroker                  │
│  Types: Message, ChatRequest, ChatResponse, Document, ToolCall, ...      │
│  Error: DaimonError, Result                                              │
└─────────────────────────────────────────────────────────────────────────┘
         ▲                    ▲                    ▲
         │                    │                    │
         │ implements        │ implements        │ implements
         │                    │                    │
┌────────┴────────┐  ┌────────┴────────┐  ┌───────┴────────┐
│ daimon-provider-│  │ daimon-provider-│  │ daimon-plugin- │
│ bedrock         │  │ gemini          │  │ pgvector       │
│ azure           │  │                 │  │ opensearch     │
└─────────────────┘  └─────────────────┘  └────────────────┘
```

### Dependency Rules

1. **Provider crates** (`daimon-provider-*`) depend only on `daimon-core`. They implement `Model` and `EmbeddingModel`. They do not depend on `daimon`.
2. **Plugin crates** (`daimon-plugin-*`) depend only on `daimon-core`. They implement `VectorStore`. They do not depend on `daimon`.
3. **Main crate** (`daimon`) depends on `daimon-core` and optionally on provider/plugin crates via feature flags. It re-exports their types.
4. **Users** depend on `daimon` with features like `gemini`, `pgvector`. They never depend on `daimon-core` directly unless implementing a new provider.

### Adding New Components

| Component | Action |
|-----------|--------|
| New LLM provider | Create `daimon-provider-xyz`, implement `Model` (+ `EmbeddingModel` if supported), add `xyz` feature to daimon. |
| New vector store | Create `daimon-plugin-xyz`, implement `VectorStore`, add `xyz` feature to daimon. |
| New task broker | Implement `TaskBroker` in daimon (or a provider crate if cloud-specific), add feature flag. |
| New memory backend | Implement `Memory` in daimon, add feature flag. |

No circular dependencies. Clean separation.

---

## 4. The ReAct Loop

The central execution model is the ReAct (Reason-Act-Observe) loop. The agent receives a prompt, builds messages, calls the model, and either returns a final response or executes tools and loops.

### Flow Diagram

```
                    ┌──────────────────────────────────────────────────┐
                    │                  User Prompt                      │
                    └──────────────────────────────────────────────────┘
                    ┌──────────────────────────────────────────────────┐
                    │  Input Guardrails (optional transform/block)        │
                    └──────────────────────────────────────────────────┘
                    ┌──────────────────────────────────────────────────┐
                    │  Build messages: system + history + user           │
                    │  Add user message to memory                       │
                    └──────────────────────────────────────────────────┘
    ┌─────────────────────────────────────────────────────────────────────┐
    │                         ReAct Loop (max_iterations)                   │
    │                                                                      │
    │   ┌─────────────────────────────────────────────────────────────┐   │
    │   │  Check: CancellationToken cancelled? → Err(Cancelled)         │   │
    │   │  Check: Budget exceeded? → Err(BudgetExceeded)               │   │
    │   │  Check: iteration > max_iterations? → Err(MaxIterations)      │   │
    │   └─────────────────────────────────────────────────────────────┘   │
    │                              │                                       │
    │                              ▼                                       │
    │   ┌─────────────────────────────────────────────────────────────┐   │
    │   │  Hook: on_iteration_start                                    │   │
    │   └─────────────────────────────────────────────────────────────┘   │
    │                              │                                       │
    │                              ▼                                       │
    │   ┌─────────────────────────────────────────────────────────────┐   │
    │   │  Middleware: on_request (may short-circuit)                  │   │
    │   └─────────────────────────────────────────────────────────────┘   │
    │                              │                                       │
    │                              ▼                                       │
    │   ┌─────────────────────────────────────────────────────────────┐   │
    │   │  model.generate_erased(&ChatRequest)                           │   │
    │   └─────────────────────────────────────────────────────────────┘   │
    │                              │                                       │
    │                              ▼                                       │
    │   ┌─────────────────────────────────────────────────────────────┐   │
    │   │  Middleware: on_response (may short-circuit)                 │   │
    │   └─────────────────────────────────────────────────────────────┘   │
    │                              │                                       │
    │                              ▼                                       │
    │   ┌─────────────────────────────────────────────────────────────┐   │
    │   │  Hook: on_model_response                                     │   │
    │   └─────────────────────────────────────────────────────────────┘   │
    │                              │                                       │
    │              ┌────────────────┴────────────────┐                       │
    │              │                               │                       │
    │              ▼                               ▼                       │
    │   ┌─────────────────────┐         ┌─────────────────────────────────┐│
    │   │ has_tool_calls?     │   NO   │  Add assistant message to memory ││
    │   │                    │────────▶│  Output guardrails (optional)    ││
    │   │                    │        │  Return AgentResponse             ││
    │   └─────────────────────┘        └─────────────────────────────────┘│
    │              │ YES                                                      │
    │              ▼                                                         │
    │   ┌─────────────────────────────────────────────────────────────┐   │
    │   │  Add assistant message (with tool_calls) to memory            │   │
    │   │  execute_tools_parallel(tool_calls) — middleware on_tool_call   │   │
    │   │  Add tool result messages to memory                           │   │
    │   │  Hook: on_iteration_end                                        │   │
    │   │  iteration += 1, continue loop                                │   │
    │   └─────────────────────────────────────────────────────────────┘   │
    │                                                                      │
    └─────────────────────────────────────────────────────────────────────┘
```

### Exit Conditions

- **No tool calls** — Model returns text; agent adds message to memory, runs output guardrails, returns `AgentResponse`.
- **Tool calls** — Agent executes tools (with optional schema validation), appends results to messages, loops. Tools run in parallel via `tokio::task::JoinSet`.
- **Max iterations** — After `max_iterations` model invocations, returns `Err(DaimonError::MaxIterations(n))`.
- **Cancellation** — Before each iteration, checks `CancellationToken`. If cancelled, returns `Err(DaimonError::Cancelled)`.
- **Budget exceeded** — If `max_budget` is set and `CostTracker` cumulative cost ≥ limit, returns `Err(DaimonError::BudgetExceeded { spent, limit })`.
- **Middleware short-circuit** — Middleware can return `ShortCircuit(ChatResponse)` to skip the model call or inject a response.

### Key Implementation Details

- **Zero-copy message handling**`std::mem::take` moves `messages` and `tool_specs` in/out of `ChatRequest` to avoid cloning.
- **Tool execution** — Parallel via `JoinSet`. Each tool call is validated (if `validate_tool_inputs`), executed with optional retry, and results are collected in order.
- **Streaming**`prompt_stream` runs the same loop but uses `generate_stream_erased`. Tool call deltas are accumulated; when complete, tools run and the model is re-invoked. Events are yielded via `async_stream::try_stream!`.

---

## 5. Trait Hierarchy

### Core Traits (daimon-core)

```rust
/// LLM provider. Implement to add a new model backend.
pub trait Model: Send + Sync {
    fn generate(&self, request: &ChatRequest) -> impl Future<Output = Result<ChatResponse>> + Send;
    fn generate_stream(&self, request: &ChatRequest) -> impl Future<Output = Result<ResponseStream>> + Send;
}

/// Embedding model for RAG and vector search.
pub trait EmbeddingModel: Send + Sync {
    fn embed(&self, texts: &[&str]) -> impl Future<Output = Result<Vec<Vec<f32>>>> + Send;
    fn dimensions(&self) -> usize;
}

/// Low-level vector storage. Pre-computed embeddings only.
pub trait VectorStore: Send + Sync {
    fn upsert(&self, id: &str, embedding: Vec<f32>, document: Document) -> impl Future<Output = Result<()>> + Send;
    fn query(&self, embedding: Vec<f32>, top_k: usize) -> impl Future<Output = Result<Vec<ScoredDocument>>> + Send;
    fn delete(&self, id: &str) -> impl Future<Output = Result<bool>> + Send;
    fn count(&self) -> impl Future<Output = Result<usize>> + Send;
}

/// Distributed task broker for multi-process agent execution.
pub trait TaskBroker: Send + Sync {
    fn submit(&self, task: AgentTask) -> impl Future<Output = Result<String>> + Send;
    fn status(&self, task_id: &str) -> impl Future<Output = Result<TaskStatus>> + Send;
    fn receive(&self) -> impl Future<Output = Result<Option<AgentTask>>> + Send;
    fn complete(&self, task_id: &str, result: TaskResult) -> impl Future<Output = Result<()>> + Send;
    fn fail(&self, task_id: &str, error: String) -> impl Future<Output = Result<()>> + Send;
}
```

### Main Crate Traits

```rust
/// Tool the agent can invoke. JSON Schema for parameters.
pub trait Tool: Send + Sync {
    fn name(&self) -> &str;
    fn description(&self) -> &str;
    fn parameters_schema(&self) -> serde_json::Value;
    fn execute(&self, input: &serde_json::Value) -> impl Future<Output = Result<ToolOutput>> + Send;
    fn retry_policy(&self) -> Option<ToolRetryPolicy> { None }
}

/// Conversation memory. Stores and retrieves messages.
pub trait Memory: Send + Sync {
    fn add_message(&self, message: Message) -> impl Future<Output = Result<()>> + Send;
    fn get_messages(&self) -> impl Future<Output = Result<Vec<Message>>> + Send;
    fn clear(&self) -> impl Future<Output = Result<()>> + Send;
}

/// Lifecycle callbacks during agent execution.
pub trait AgentHook: Send + Sync {
    fn on_iteration_start(&self, state: &AgentState) -> impl Future<Output = Result<()>> + Send;
    fn on_model_response(&self, response: &ChatResponse) -> impl Future<Output = Result<()>> + Send;
    fn on_tool_call(&self, call: &ToolCall) -> impl Future<Output = Result<()>> + Send;
    fn on_tool_result(&self, call: &ToolCall, result: &ToolOutput) -> impl Future<Output = Result<()>> + Send;
    fn on_iteration_end(&self, state: &AgentState) -> impl Future<Output = Result<()>> + Send;
    fn on_error(&self, error: &DaimonError) -> impl Future<Output = Result<()>> + Send;
}

/// Request/response/tool_call interception. Can short-circuit.
pub trait Middleware: Send + Sync {
    fn on_request(&self, request: &mut ChatRequest) -> impl Future<Output = Result<MiddlewareAction>> + Send;
    fn on_response(&self, response: &mut ChatResponse) -> impl Future<Output = Result<MiddlewareAction>> + Send;
    fn on_tool_call(&self, call: &mut ToolCall) -> impl Future<Output = Result<MiddlewareAction>> + Send;
}

/// Input validation before the model sees it.
pub trait InputGuardrail: Send + Sync {
    fn check(&self, input: &str, messages: &[Message]) -> impl Future<Output = Result<GuardrailResult>> + Send;
}

/// Output validation before returning to the caller.
pub trait OutputGuardrail: Send + Sync {
    fn check(&self, response: &ChatResponse) -> impl Future<Output = Result<GuardrailResult>> + Send;
}

/// Document retrieval (vector store, search engine, etc.).
pub trait Retriever: Send + Sync {
    fn retrieve(&self, query: &str, top_k: usize) -> impl Future<Output = Result<Vec<Document>>> + Send;
}

/// High-level RAG: ingest, search, remove, count.
pub trait KnowledgeBase: Send + Sync {
    fn ingest(&self, documents: Vec<Document>) -> impl Future<Output = Result<Vec<String>>> + Send;
    fn search(&self, query: &str, top_k: usize) -> impl Future<Output = Result<Vec<Document>>> + Send;
    fn remove(&self, id: &str) -> impl Future<Output = Result<bool>> + Send;
    fn count(&self) -> impl Future<Output = Result<usize>> + Send;
}

/// Checkpoint persistence for resumable runs.
pub trait Checkpoint: Send + Sync {
    fn save(&self, state: &CheckpointState) -> impl Future<Output = Result<()>> + Send;
    fn load(&self, run_id: &str) -> impl Future<Output = Result<Option<CheckpointState>>> + Send;
    fn list_runs(&self) -> impl Future<Output = Result<Vec<String>>> + Send;
    fn delete(&self, run_id: &str) -> impl Future<Output = Result<()>> + Send;
}

/// Token cost model for budget tracking.
pub trait CostModel: Send + Sync {
    fn cost_per_token(&self, model_id: &str, direction: TokenDirection) -> f64;
}
```

### Trait Relationships

```
Model ──────────────────────────────────────────────────────────────┐
  │                                                                   │
  └──► ChatRequest, ChatResponse, Message, ToolSpec, Usage             │
EmbeddingModel ───────────────────────────────────────────────────────┤
  │                                                                   │
  └──► embed(), dimensions()                                          │
VectorStore ◄─── SimpleKnowledgeBase ◄─── Retriever ◄─── RetrieverTool │
  │                     │                    │                         │
  └──► Document, ScoredDocument               └──► Tool                 │
TaskBroker ───► AgentTask, TaskResult, TaskStatus                     │
Tool ───► ToolRegistry ───► Agent                                      │
Memory ───► Agent                                                       │
AgentHook ───► Agent                                                   │
Middleware ───► MiddlewareStack ───► Agent                              │
InputGuardrail, OutputGuardrail ───► Agent                              │
CostModel ───► CostTracker ───► Agent                                  │
Checkpoint ───► ResumableAgent                                         │
```

---

## 6. Object-Safety Pattern

Rust traits with `impl Future` return types are not object-safe (the compiler cannot know the future size). Daimon uses the **Erased* + Shared*** pattern for every async trait.

### Pattern

1. **Static trait** — Uses `impl Future<Output = T> + Send` for zero-cost static dispatch.
2. **Erased trait** — Object-safe wrapper with `Pin<Box<dyn Future<Output = T> + Send + 'a>>`.
3. **Blanket impl**`impl<T: Trait> ErasedTrait for T` forwards to the static trait.
4. **Shared type alias**`type SharedTrait = Arc<dyn ErasedTrait>`.

### Example: Model

```rust
pub trait Model: Send + Sync {
    fn generate(&self, request: &ChatRequest) -> impl Future<Output = Result<ChatResponse>> + Send;
    fn generate_stream(&self, request: &ChatRequest) -> impl Future<Output = Result<ResponseStream>> + Send;
}

pub trait ErasedModel: Send + Sync {
    fn generate_erased<'a>(&'a self, request: &'a ChatRequest)
        -> Pin<Box<dyn Future<Output = Result<ChatResponse>> + Send + 'a>>;
    fn generate_stream_erased<'a>(&'a self, request: &'a ChatRequest)
        -> Pin<Box<dyn Future<Output = Result<ResponseStream>> + Send + 'a>>;
}

impl<T: Model> ErasedModel for T {
    fn generate_erased<'a>(&'a self, request: &'a ChatRequest)
        -> Pin<Box<dyn Future<Output = Result<ChatResponse>> + Send + 'a>>
    {
        Box::pin(self.generate(request))
    }
    fn generate_stream_erased<'a>(&'a self, request: &'a ChatRequest)
        -> Pin<Box<dyn Future<Output = Result<ResponseStream>> + Send + 'a>>
    {
        Box::pin(self.generate_stream(request))
    }
}

pub type SharedModel = Arc<dyn ErasedModel>;
```

### Traits Using This Pattern

| Trait | Erased* | Shared* |
|-------|---------|---------|
| Model | ErasedModel | SharedModel |
| EmbeddingModel | ErasedEmbeddingModel | SharedEmbeddingModel |
| VectorStore | ErasedVectorStore | SharedVectorStore |
| TaskBroker | ErasedTaskBroker | (no alias; used directly) |
| Tool | ErasedTool | SharedTool |
| Memory | ErasedMemory | SharedMemory |
| AgentHook | ErasedAgentHook | (Arc<dyn ErasedAgentHook>) |
| Middleware | ErasedMiddleware | SharedMiddleware |
| InputGuardrail | ErasedInputGuardrail | (Vec<Arc<dyn ErasedInputGuardrail>>) |
| OutputGuardrail | ErasedOutputGuardrail | (Vec<Arc<dyn ErasedOutputGuardrail>>) |
| Retriever | ErasedRetriever | SharedRetriever |
| KnowledgeBase | ErasedKnowledgeBase | SharedKnowledgeBase |
| Checkpoint | ErasedCheckpoint | (Arc<dyn ErasedCheckpoint>) |

### Usage

- **Static dispatch** — When the concrete type is known at compile time, use `impl Model` or `M: Model`. No heap allocation, no vtable.
- **Dynamic dispatch** — When you need heterogeneous collections or runtime plugin loading, use `SharedModel` (`Arc<dyn ErasedModel>`). The agent stores `SharedModel` so users can pass any provider.

---

## 7. Feature Flags

Every provider, storage backend, and protocol is behind a feature flag. The core framework compiles with no features (or with `default`).

### Complete Feature Matrix

| Feature | Pulls In | Description |
|---------|----------|-------------|
| **default** | openai, anthropic, macros, ollama | Common providers + tool macro + local Ollama |
| **macros** | daimon-macros | `#[tool_fn]` proc macro |
| **openai** | reqwest | OpenAI API provider (GPT-4, etc.) |
| **anthropic** | reqwest | Anthropic Claude API |
| **gemini** | daimon-provider-gemini | Google Gemini / Vertex AI |
| **azure** | daimon-provider-azure | Azure OpenAI Service |
| **bedrock** | daimon-provider-bedrock | AWS Bedrock |
| **ollama** | reqwest | Ollama local models |
| **sqs** | bedrock, daimon-provider-bedrock/sqs | AWS SQS task broker |
| **pubsub** | gemini, daimon-provider-gemini/pubsub | Google Cloud Pub/Sub task broker |
| **servicebus** | azure, daimon-provider-azure/servicebus | Azure Service Bus task broker |
| **sqlite** | rusqlite | SQLite memory backend |
| **redis** | redis | Redis memory + task broker |
| **nats** | async-nats | NATS JetStream task broker |
| **amqp** | lapin | RabbitMQ (AMQP) task broker |
| **mcp** | reqwest, tokio-tungstenite | Model Context Protocol client & server |
| **otel** | opentelemetry, opentelemetry_sdk, opentelemetry-otlp, tracing-opentelemetry, tracing-subscriber | OpenTelemetry OTLP span export |
| **http-server** | axum, tower-http | REST API server (POST /prompt, SSE) |
| **qdrant** | qdrant-client | Qdrant vector store retriever |
| **pgvector** | daimon-plugin-pgvector | pgvector-backed VectorStore |
| **opensearch** | daimon-plugin-opensearch | OpenSearch k-NN VectorStore |
| **grpc** | tonic, prost | gRPC transport for distributed execution |
| **eval** | (none) | EvalScenario, EvalRunner, Scorer |
| **full** | All of the above | Everything enabled |

### Dependency Graph (Simplified)

```
daimon
├── daimon-core (always)
├── daimon-macros (macros)
├── daimon-provider-bedrock (bedrock)
├── daimon-provider-gemini (gemini)
├── daimon-provider-azure (azure)
├── daimon-plugin-pgvector (pgvector)
├── daimon-plugin-opensearch (opensearch)
├── reqwest (openai | anthropic | ollama | mcp)
├── rusqlite (sqlite)
├── redis (redis)
├── async-nats (nats)
├── lapin (amqp)
├── tokio-tungstenite (mcp)
├── opentelemetry* (otel)
├── axum, tower-http (http-server)
├── qdrant-client (qdrant)
├── tonic, prost (grpc)
└── ...
```

### Compiling Without Optional Deps

```bash
cargo build --no-default-features   # Core only; no providers
cargo build --no-default-features -F gemini   # Core + Gemini only
cargo build -F full   # Everything
```

---

## 8. Error Model

All fallible operations return `Result<T>`, an alias for `std::result::Result<T, DaimonError>`.

### DaimonError Variants

```rust
#[derive(Error, Debug)]
pub enum DaimonError {
    #[error("model error: {0}")]
    Model(String),

    #[error("tool execution failed for '{tool}': {message}")]
    ToolExecution { tool: String, message: String },

    #[error("tool '{0}' not found in registry")]
    ToolNotFound(String),

    #[error("duplicate tool '{0}' in registry")]
    DuplicateTool(String),

    #[error("agent builder validation failed: {0}")]
    Builder(String),

    #[error("max iterations ({0}) exceeded")]
    MaxIterations(usize),

    #[error("serialization error: {0}")]
    Serialization(#[from] serde_json::Error),

    #[error("schema validation failed for tool '{tool}': {errors}")]
    SchemaValidation { tool: String, errors: String },

    #[error("stream closed unexpectedly")]
    StreamClosed,

    #[error("request timed out after {0:?}")]
    Timeout(std::time::Duration),

    #[error("operation cancelled")]
    Cancelled,

    #[error("orchestration error: {0}")]
    Orchestration(String),

    #[error("MCP error: {0}")]
    Mcp(String),

    #[error("budget exceeded: ${spent:.6} spent, limit was ${limit:.6}")]
    BudgetExceeded { spent: f64, limit: f64 },

    #[error("guardrail blocked: {0}")]
    GuardrailBlocked(String),

    #[error("{0}")]
    Other(String),
}
```

### Provider Error Mapping

Provider crates (HTTP, gRPC, SDK) map their errors to `DaimonError`:

```rust
// Example: in daimon-provider-gemini
.map_err(|e| DaimonError::Model(e.to_string()))
```

Use `DaimonError::Model(String)` for API errors, rate limits, bad responses. Use `DaimonError::Other(String)` for unexpected or uncategorized failures.

---

## 9. Async Runtime

### Tokio

Daimon uses **tokio** as the async runtime. All traits assume a tokio executor. The main crate depends on `tokio` with `features = ["full"]`; daimon-core uses `features = ["rt"]` for minimal footprint.

### Send + Sync

Every trait is `Send + Sync`. Types used across async boundaries (agent, model, tools, memory) must be `Send` so they can be moved between tasks. `Sync` allows shared references across threads (e.g., `Arc<dyn ErasedModel>`).

### Cancellation

Cooperative cancellation via `tokio_util::sync::CancellationToken`:

```rust
pub async fn prompt_with_cancellation(&self, input: &str, cancel: &CancellationToken) -> Result<AgentResponse>
```

The ReAct loop checks `cancel.is_cancelled()` before each iteration. If cancelled, returns `Err(DaimonError::Cancelled)`.

### Streaming

Streaming uses `async-stream` and `futures::Stream`:

```rust
pub type ResponseStream = Pin<Box<dyn Stream<Item = Result<StreamEvent>> + Send>>;
```

Events: `TextDelta`, `ToolCallStart`, `ToolCallDelta`, `ToolCallEnd`, `ToolResult`, `Usage`, `Error`, `Done`. The stream runs the full ReAct loop internally; tool calls are accumulated and executed between model invocations.

---

## 10. Orchestration Layer

Beyond the single-agent ReAct loop, Daimon provides multi-step orchestration:

### Chain

Linear pipeline of steps. Each step transforms input to output; context flows sequentially.

```rust
pub trait ChainStep: Send + Sync {
    fn execute(&self, ctx: &mut ChainContext) -> impl Future<Output = Result<()>> + Send;
}
```

`Chain` runs steps in order. `AgentStep` wraps an agent; `TransformStep` wraps a closure. Use for simple pipelines: prompt → agent → post-process.

### Graph

Directed graph with conditional nodes. Supports cycles, fan-out, fan-in. Pregel-style sequential walker.

```rust
pub trait GraphNode: Send + Sync {
    fn execute(&self, ctx: &mut GraphContext) -> impl Future<Output = Result<NodeOutcome>> + Send;
}
```

`NodeOutcome` can route to another node by name. Use for branching workflows: router → specialist A or B → aggregator.

### DAG

DagBuilder builds a DAG at construction time. Independent nodes run in parallel; dependencies are enforced. AllPredecessor model: a node runs when all predecessors complete.

```rust
pub struct Dag { /* nodes, edges */ }
pub struct DagBuilder { /* ... */ }
```

Use for parallelizable workflows: fetch data from N sources → merge → agent.

### Workflow

Extends DAG with typed field-level data mapping. Nodes read/write named fields; the engine wires outputs to inputs by field name. Inspired by Eino's typed-field workflow engine.

```rust
pub trait WorkflowNode: Send + Sync {
    fn execute(&self, ctx: &mut WorkflowContext) -> impl Future<Output = Result<()>> + Send;
}
```

---

## 11. Distributed Execution

For multi-process agent execution, Daimon uses the `TaskBroker` trait from daimon-core.

### Flow

1. **Submit** — Client submits `AgentTask` (input, optional run_id, metadata). Broker returns task_id.
2. **Receive** — Worker blocks on `receive()` until a task is available.
3. **Execute** — Worker runs the agent loop with the task input.
4. **Complete/Fail** — Worker calls `complete(task_id, result)` or `fail(task_id, error)`.
5. **Status** — Client polls `status(task_id)` for Pending/Running/Completed/Failed.

### Broker Implementations

| Feature | Broker | Backend |
|---------|--------|---------|
| sqs | SqsBroker | AWS SQS (via daimon-provider-bedrock) |
| pubsub | PubSubBroker | Google Cloud Pub/Sub (via daimon-provider-gemini) |
| servicebus | ServiceBusBroker | Azure Service Bus (via daimon-provider-azure) |
| redis | RedisBroker | Redis |
| nats | NatsBroker | NATS JetStream |
| amqp | AmqpBroker | RabbitMQ (AMQP) |
| grpc | gRPC transport | Tonic + Prost |

### Worker Pattern

```rust
let broker = RedisBroker::new(redis_url)?;
loop {
    let task = broker.receive().await?.ok_or(...)?;
    let result = agent.prompt(&task.input).await?;
    broker.complete(&task.task_id, TaskResult { ... }).await?;
}
```

---

## 12. MCP (Model Context Protocol)

The `mcp` feature enables the Model Context Protocol for tool discovery and execution across process boundaries.

### Client

Connects to MCP servers (stdio, HTTP, WebSocket). Discovers tools from the server and registers them with the agent. Tool calls are forwarded to the remote server.

### Server

Exposes Daimon tools to external MCP clients. Tools registered in the agent are advertised; clients invoke them via the protocol.

### Transports

- **stdio** — For local subprocess servers.
- **HTTP/SSE** — For remote HTTP-based servers.
- **WebSocket** — For bidirectional streaming.

---

## 13. A2A (Agent-to-Agent)

The `a2a` module provides Google's Agent-to-Agent protocol support for inter-agent communication. Agents can send structured messages to other agents.

---

## 14. Checkpointing and Resumable Runs

The `Checkpoint` trait enables resumable agent execution. When a run is interrupted, the agent can load the last checkpoint and continue.

### CheckpointState

Stores messages, iteration count, and any custom state. Implementations persist to file, Redis, or NATS KV.

### Implementations

| Backend | Feature | Use Case |
|---------|---------|----------|
| FileCheckpoint | (built-in) | Local development |
| RedisCheckpoint | redis | Production |
| NatsKVCheckpoint | nats | Distributed |

### ResumableAgent

Wraps an agent with `Checkpoint`. When `prompt` is called with a `run_id`, it loads the checkpoint (if any), restores the message history, and continues from the last state.

---

## 15. Cost Tracking and Budget

Attach a `CostModel` to an agent to track token spend. Set `max_budget` to abort when a dollar limit is reached.

### CostModel

```rust
pub trait CostModel: Send + Sync {
    fn cost_per_token(&self, model_id: &str, direction: TokenDirection) -> f64;
}
```

`TokenDirection::Input` or `Output`. Built-in: `OpenAiCostModel`, `AnthropicCostModel`.

### CostTracker

Wraps a `CostModel`. On each model response, `record(model_id, usage)` computes cost and accumulates. `cumulative_cost()` returns total spent. The ReAct loop checks `cumulative_cost() >= max_budget` before each iteration.

---

## 16. Prompt System

The `prompt` module provides:

- **PromptTemplate** — String templates with `{{variable}}` placeholders.
- **Builder** — Fluent API for building prompts with variables.
- **Few-shot** — Inject example messages for in-context learning.
- **Dynamic** — Runtime prompt construction from context.

---

## 17. Eval (Evaluation)

The `eval` feature provides a testing harness for agent behavior:

- **EvalScenario** — Input + expected outcome (e.g., `expect_contains("4")`).
- **EvalRunner** — Runs scenarios through an agent, collects results.
- **EvalResult** — Pass/fail, latency, cost, score.
- **Scorer** — Custom scoring logic.

---

## 18. Shared Types (daimon-core)

Types used across the plugin boundary. All defined in `daimon-core` and re-exported by `daimon`:

| Type | Description |
|------|-------------|
| `Message` | Conversation message: role (System/User/Assistant/Tool), content, tool_calls, tool_call_id |
| `ChatRequest` | messages, tools (ToolSpec), temperature, max_tokens |
| `ChatResponse` | message, stop_reason (EndTurn/ToolUse/MaxTokens), usage |
| `ToolSpec` | name, description, parameters (JSON Schema) |
| `ToolCall` | id, name, arguments (JSON) |
| `Usage` | input_tokens, output_tokens, cached_tokens |
| `Document` | content, metadata, score |
| `ScoredDocument` | document, score |
| `StreamEvent` | TextDelta, ToolCallStart, ToolCallDelta, ToolCallEnd, ToolResult, Usage, Error, Done |
| `AgentTask` | task_id, input, run_id, metadata |
| `TaskResult` | task_id, output, iterations, cost, error |
| `TaskStatus` | Pending, Running, Completed(TaskResult), Failed(String) |

---

## 19. Memory Implementations

### SlidingWindowMemory (default)

In-memory store that keeps only the last N messages. Configurable window size (default 50). No persistence; suitable for single-session conversations.

### SQLiteMemory (sqlite feature)

Persists messages to SQLite. Survives process restarts. Use for development or low-concurrency production.

### RedisMemory (redis feature)

Stores messages in Redis. Supports multi-process or multi-instance deployments. Keyed by session ID.

### TokenWindowMemory

Wraps another memory backend. Truncates history to fit within a token budget (estimated by character count). Ensures the context never exceeds model limits.

---

## 20. Tool Registry and Validation

### ToolRegistry

Stores tools as `HashMap<String, Arc<dyn ErasedTool>>`. `register(tool)` inserts by `tool.name()`; duplicate names return `Err(DuplicateTool)`. `tool_specs()` builds `Vec<ToolSpec>` for the model (name, description, parameters_schema).

### Schema Validation

When `validate_tool_inputs` is true, each tool call's `arguments` are validated against the tool's `parameters_schema` via `jsonschema`. Invalid inputs produce an error message returned to the model (so it can correct on the next iteration) rather than executing the tool.

### Retry Policy

Tools can declare `retry_policy()` returning `Option<ToolRetryPolicy>`. The policy specifies max_retries, backoff (exponential), and `is_retryable(&str)` to decide if an error message warrants a retry. Agent-level `tool_retry_policy` applies when the tool does not override.

---

## 21. Multi-Agent Patterns

### Agent-as-Tool (AgentTool)

Wraps an agent as a `Tool`. The tool's `execute` runs `agent.prompt(input)`. Use when one agent needs to delegate to another (e.g., a router agent calling specialist agents).

### Supervisor

One agent delegates to specialized sub-agents. The supervisor receives a task, routes it to a sub-agent by name, and returns the result. Sub-agents are `HashMap<String, Arc<Agent>>`.

### HandoffNetwork

Agents transfer control to each other mid-conversation. A handoff action specifies the target agent and context. Use for multi-agent conversations where control passes between participants.

### StructuredOutput

Extracts typed data from LLM responses. Uses a schema (e.g., JSON Schema) to parse the model's text into a struct. Fails if the output does not match.

---

## 22. Middleware Stack

`MiddlewareStack` holds a list of `Arc<dyn ErasedMiddleware>`. On each request, response, and tool call, the stack runs middleware in order. Any middleware can return `ShortCircuit(ChatResponse)` to skip the rest of the pipeline and return immediately.

Use cases: caching, logging, rate limiting, request/response transformation, A/B testing (inject different prompts).

---

## 23. Proc Macro: tool_fn

The `#[tool_fn]` macro derives a `Tool` implementation from an async function:

```rust
/// Adds two numbers together.
#[tool_fn]
async fn add(
    /// The first number.
    a: f64,
    /// The second number.
    b: f64,
) -> daimon::Result<ToolOutput> {
    Ok(ToolOutput::text(format!("{}", a + b)))
}
```

- Function name → tool name (or override via `name = "..."`).
- Doc comments → tool description (or override via `description = "..."`).
- Parameters → JSON Schema properties. `Option<T>` marks optional.
- Supported types: `String`, `i8``i128`, `u8``u128`, `f32`, `f64`, `bool`, `Option<T>`, `Vec<T>`, `serde_json::Value`.

---

## 24. Summary

Daimon is a Rust-native AI agent framework with:

- **daimon-core** as the stable plugin boundary (Model, EmbeddingModel, VectorStore, TaskBroker + shared types).
- **Provider/plugin crates** that depend only on daimon-core.
- **ReAct loop** as the central execution model: prompt → model → tool calls (optional) → loop until done.
- **Erased* + Shared*** pattern for object-safe async traits.
- **Feature flags** for every optional component; zero unused deps.
- **Single DaimonError** enum; providers map to `Model` or `Other`.
- **Tokio** async runtime; `CancellationToken` for cancellation; `async-stream` for streaming.

To contribute: implement a trait from daimon-core, add a feature flag, and wire it in the main crate.