net-mesh 0.36.0

High-performance, schema-agnostic, backend-agnostic event bus
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
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
# AI Tool Calling on net

`net-mesh` exposes a typed AI-tool calling surface: agents discover
tools advertised by mesh nodes, invoke them by name with JSON request
payloads, and receive either a typed JSON response (unary) or a stream
of `ToolEvent` envelopes (streaming). Tools are first-class
`nrpc:<tool_id>` services with an `ai-tool:<tool_id>` capability tag
on top.

This doc covers the **Rust SDK** surface that landed in Wave 2 of
the AI tool-calling work. Bindings (Node / Python / Go) and format
translators (OpenAI / Anthropic / Gemini / MCP) are tracked in
`docs/internal/plans/NRPC_AI_TOOL_CALLING_AND_AGENT_DX.md`.

## Quickstart

```rust
use net_sdk::mesh::MeshBuilder;
use net_sdk::tool::metadata_for;
use schemars::JsonSchema;
use serde::{Deserialize, Serialize};

#[derive(JsonSchema, Deserialize, Serialize)]
struct WebSearchReq {
    /// Free-text query string.
    query: String,
}

#[derive(JsonSchema, Deserialize, Serialize)]
struct WebSearchResp {
    results: Vec<String>,
}

#[tokio::main]
async fn main() -> anyhow::Result<()> {
    let mesh = MeshBuilder::new("127.0.0.1:0", &[0u8; 32])?.build().await?;

    // Build a descriptor: name + Rust types → JSON Schemas via
    // `schemars`, plus the fluent setters for the fields not
    // derivable from the type signature.
    let descriptor = metadata_for::<WebSearchReq, WebSearchResp>("web_search")
        .description("Search the web for relevant pages.")
        .stateless(true)
        .estimated_time_ms(500)
        .tag("web")
        .tag("research")
        .build();

    // Atomically register: handler + tool registry + capability tag.
    // Drop the handle to deregister everything.
    let _handle = mesh.serve_tool::<WebSearchReq, WebSearchResp, _, _>(
        descriptor,
        |req| async move {
            Ok(WebSearchResp {
                results: vec![format!("hit for {}", req.query)],
            })
        },
    )?;

    // Announce so peers' capability folds pick up the tool.
    mesh.announce_capabilities(Default::default()).await?;

    // … keep the mesh alive …
    Ok(())
}
```

On the caller side:

```rust
use net_sdk::tool::{TagMatcher, ToolDescriptor};

// One in-memory walk of the local capability fold; no network.
let tools: Vec<ToolDescriptor> = mesh.list_tools(None);

// Capability-routed call. JSON codec under the hood.
let resp: WebSearchResp = mesh
    .call_tool("web_search", &WebSearchReq { query: "mesh".into() })
    .await?;
```

## Concepts

### One identifier — three roles

A tool's `tool_id` is simultaneously:

1. The nRPC service name (`nrpc:<tool_id>.requests` channel),
2. The AI-tool capability tag (`ai-tool:<tool_id>`),
3. The unary-call service the underlying `serve_rpc_typed` is keyed on.

No mapping table. A tool registered as `web_search` IS the nRPC service
at `nrpc:web_search` IS the announcement carrying `ai-tool:web_search`.
This is locked decision #1 of the plan.

### JSON codec everywhere

Every tool request and response is JSON-encoded. Every `ToolEvent`
envelope is one JSON chunk on the streaming path. This is locked
decision #3: every LLM provider (OpenAI, Anthropic, Gemini, MCP)
consumes JSON for tool input/output, so the substrate enforces one
codec for the whole tool surface.

### Schema format = JSON Schema draft 2020-12

The `metadata_for::<Req, Resp>(name)` helper derives input/output
schemas via the `schemars` crate. Both schemas land on the
`ToolDescriptor` as JSON-encoded strings. Format translators
(`formats/openai`, `formats/anthropic`, …) consume the schemas
verbatim or lower them to per-provider shapes.

### `ToolEvent` streaming envelope

Streaming tools emit one `ToolEvent` per chunk:

| Variant       | Direction | Purpose                                         |
|---------------|-----------|-------------------------------------------------|
| `Start`       | first     | `tool_id`, optional `call_id`, optional metadata |
| `Progress`    | mid       | Optional `pct` + `message` for spinners         |
| `Delta`       | mid       | Partial output (token, file chunk, log line)    |
| `Result`      | terminal  | Final result payload                            |
| `Error`       | terminal  | Structured failure: `code`, `message`, `details`|

Every stream ends with **exactly one** terminal event (`Result` or
`Error`). If a handler ends its stream without one, the SDK's
streaming wrapper synthesizes `ToolEvent::Error { code: "missing_terminal", ... }`
so callers can rely on the contract.

Unary tools (`serve_tool` / `call_tool`) bypass the envelope: the wire
shape is just the typed `Resp` bytes. Adapter packages synthesize a
single `Result` envelope locally when lowering a unary call into a
provider's streaming protocol.

## Server side

### Unary: `Mesh::serve_tool`

```rust
let handle = mesh.serve_tool::<Req, Resp, _, _>(descriptor, handler)?;
```

Atomically performs four steps:

1. Insert the descriptor into the local `tool_registry` (so the next
   `announce_capabilities` auto-emits the `ai-tool:<id>` tag + the
   typed `ToolCapability` + description/streaming/tags metadata keys).
2. Register the handler via `serve_rpc_typed` at `tool_id` with JSON codec.
3. On the FIRST `serve_tool` call to this `Mesh`, lazy-install the
   `tool.metadata.fetch` nRPC service handler so agents can pull full
   descriptors for tools whose schemas were too large for the fold.
4. Return a `ToolServeHandle`. Drop reverses steps 1 + 2.

Duplicate registrations of the same `tool_id` return
`ServeError::AlreadyServing(tool_id)`. The prior handler's registry
entry is preserved.

### Streaming: `Mesh::serve_tool_streaming`

```rust
let handle = mesh.serve_tool_streaming::<Req, _, _, _>(descriptor, |req| async move {
    futures::stream::iter(vec![
        ToolEvent::Start { tool_id: "web_search".into(), call_id: None, metadata: None },
        ToolEvent::Progress { pct: Some(50.0), message: Some("indexing".into()) },
        ToolEvent::Delta { data: serde_json::json!({ "token": "result " }) },
        ToolEvent::Result { data: serde_json::json!({ "results": ["hit"] }) },
    ])
})?;
```

Contract:

- The handler returns `impl Future<Output = impl Stream<Item = ToolEvent>>`.
- The SDK serializes each item as one JSON chunk on the underlying
  `serve_rpc_streaming_typed` path with `Resp = ToolEvent`.
- Stop on the first terminal event — the SDK does not drain past it.
- Missing-terminal synthesis ensures every stream closes cleanly.
- `descriptor.streaming` is forced to `true` on register so announce
  metadata reflects reality.

## Discovery

### `Mesh::list_tools`

```rust
// All tools the local fold has seen.
let tools = mesh.list_tools(None);

// Region-scoped: only EU hosts.
let matcher = TagMatcher::Prefix { value: "region.eu".into() };
let eu_tools = mesh.list_tools(Some(&matcher));
```

One in-memory walk of the capability fold. For each
`(class, NodeId) → CapabilityMembership` entry:

1. Pre-filter by the optional `TagMatcher` — entry must have ANY tag
   matching.
2. Decode `software.tool.<i>.*` tags into `Vec<ToolCapability>`.
3. Hydrate input/output schemas from
   `tool::<id>::input_schema` / `output_schema` metadata keys.
4. Build `ToolDescriptor::from_capability(cap, metadata)`.
5. Dedupe by `(tool_id, version)`; accumulate `node_count`.

Returns a `Vec<ToolDescriptor>` sorted by `(tool_id, version)` for
stable snapshots.

### `Mesh::watch_tools`

```rust
let mut changes = mesh.watch_tools(None, Some(Duration::from_millis(250)));
while let Some(change) = changes.next().await {
    match change {
        ToolListChange::Added(desc) => println!("+ {}", desc.tool_id),
        ToolListChange::Removed(desc) => println!("- {}", desc.tool_id),
        ToolListChange::NodeCountChanged { descriptor, prev_node_count } => {
            println!("~ {}: {} -> {}", descriptor.tool_id, prev_node_count, descriptor.node_count);
        }
    }
}
```

Backed by a polling task: every `interval` (default `1s`), the task
re-runs `list_tools(&matcher)` and diffs against the prior snapshot.
The first event fires AFTER the initial baseline — call `list_tools`
first if you need the starting shape.

The initial baseline is taken **synchronously** before the task
spawns so a subscribe-then-publish call sequence never loses the
`Added` event to a race.

Dropping the `ToolListWatch` ends the polling task on its next tick.

### `tool.metadata.fetch` — oversized schemas

The capability fold has a per-entry payload budget. Tools with large
JSON schemas (multi-KB Pydantic-derived shapes, deep nested Zod
output) may have their `input_schema` / `output_schema` dropped on
the fold path; the resulting descriptor lands with `input_schema: None`.

Agents that need the full schema for strict-mode adapters
(OpenAI's `strict: true`, Anthropic's tool blocks, MCP's
`inputSchema`) call the auto-installed `tool.metadata.fetch` nRPC
service against the publishing host:

```rust
let resp: ToolMetadataResponse = mesh.call_typed(
    host_node_id,
    TOOL_METADATA_FETCH_SERVICE,
    &ToolMetadataRequest { name: "web_search".into() },
    CallOptionsTyped { codec: Codec::Json, raw: Default::default() },
).await?;
match resp {
    ToolMetadataResponse::Found { descriptor } => { /* use full schema */ }
    ToolMetadataResponse::NotFound { name } => { /* host doesn't serve this tool */ }
}
```

The host's `tool.metadata.fetch` handler is lazy-installed on the
first `serve_tool` call and stays alive for the `Mesh` lifetime;
empty registries just return `NotFound` for every request.

## Client side

### Unary: `Mesh::call_tool`

```rust
let resp: WebSearchResp = mesh
    .call_tool("web_search", &WebSearchReq { query: "mesh".into() })
    .await?;
```

Capability-routed: consults the local fold for nodes advertising
`nrpc:web_search`, picks one per the default routing policy, calls.
Returns `RpcError::NoRoute` if no host currently serves the tool.
Bubbles handler errors as `RpcError::ServerError` with status
`NRPC_TYPED_HANDLER_ERROR` carrying the handler's error message.

### Streaming: `Mesh::call_tool_streaming`

```rust
let stream = mesh
    .call_tool_streaming("web_search_stream", &WebSearchReq { query: "mesh".into() })
    .await?;
let events: Vec<ToolEvent> = stream.map(|item| item.unwrap()).collect().await;
```

Returns `RpcStreamTyped<ToolEvent>`. Implements `futures::Stream`.
Caller surfaces wire events verbatim; adapter packages
(`formats/anthropic`, `formats/openai`, etc.) own the contract
enforcement (e.g. lowering `Delta` envelopes to provider streaming
protocols, accumulating partial JSON on Anthropic
`tool_use_block_delta`, etc.).

Dropping the stream emits CANCEL to the server (substrate cancel-token
contract).

## Capability scoping

Tool discovery rides the existing capability fold, so the same
filters and scope mechanisms apply:

- **Region scoping**: `TagMatcher::Prefix { value: "region.eu" }`  see Discovery above.
- **Subnet visibility**: callers in a subnet only discover tools
  advertised within their subnet (same as any other capability).
- **Capability auth**: a tool's host can populate `allowed_nodes` /
  `allowed_subnets` / `allowed_groups` on its `CapabilitySet` to gate
  invocation. Off-mesh callers fail the auth check before the handler
  runs.
- **Predicate-pushdown**: `CallOptions::with_where(&pred)` rides the
  `net-where` request header to push per-call predicate filtering.

## Atomicity

`Mesh::serve_tool` is atomic with respect to observable mesh state:
either all of (handler registration, capability-fold tag publish,
`nrpc:<id>` + `ai-tool:<id>` tags, descriptor in `tool_registry`)
succeed, or none do.

- Step 1 failure (duplicate `tool_id`): nothing changed.
- Step 2 failure (`serve_rpc_typed` returned `Err`): step 1's
  registry insert is paired-removed before the error returns.
- Drop reverses steps 1 + 2 in order. The lazy
  `tool.metadata.fetch` handler from step 3 stays installed for the
  `Mesh` lifetime (idempotent; harmless when the registry is empty).

## Wire compatibility

No wire-ABI bump for unary tool calls — they ride the existing
`call_service` + `serve_rpc_typed` path. Streaming tools use the new
substrate primitive `MeshNode::call_service_streaming` (S-1); the
wire shape of an individual stream is unchanged from `call_streaming`.
`ToolEvent` envelopes are JSON-encoded chunks on existing streams.

Receiving binaries WITHOUT the `tool` Cargo feature still apply
inbound tool announcements correctly: the `software.tool.<i>.*` tags
and the `tool::<id>::*` metadata keys ride through the standard
`CapabilitySet` shape. They just don't have a `list_tools` /
`call_tool` surface to consume them with.

## Cargo features

- `net-mesh/tool` enables substrate-side `ToolDescriptor`,
  `ToolMetadataRegistry`, `MeshNode::list_tools` / `watch_tools`, the
  announce-time merge from the registry into `CapabilitySet`, and the
  `ToolEvent` wire type.
- `net-mesh-sdk/tool` enables the SDK-facing wrappers:
  `Mesh::serve_tool` / `serve_tool_streaming` / `list_tools` /
  `watch_tools` / `call_tool` / `call_tool_streaming` plus the
  `ToolMetadataBuilder` + `metadata_for` schema helpers (pulls in
  `schemars`).

Both gated off by default. Bindings that consume `net-mesh-sdk`
through prebuilt artifacts (Node / Python / Go) enable the feature in
their build pipeline.

## Bindings

Each language binding ships a tool layer mirroring the Rust SDK
surface. Wire byte-equality across the four implementations is
mechanically enforced by golden-vector fixtures (T-1, T-2).

### Node TS (`@net-mesh/sdk`)

```ts
import {
  callTool,
  fetchToolMetadata,
  listTools,
  serveTool,
  watchTools,
  openai,
} from '@net-mesh/sdk/tool'

// Server: register a tool.
const handle = serveTool(
  rpc,
  {
    name: 'web_search',
    description: 'Search the web for relevant pages.',
    inputSchema: { type: 'object', properties: { query: { type: 'string' } } },
    tags: ['web', 'research'],
    estimatedTimeMs: 500,
  },
  async (req: { query: string }) => ({ results: [`hit for ${req.query}`] }),
)
// Tear down: handle.close()

// Discovery.
const tools = listTools(mesh)               // one-shot snapshot
for await (const change of watchTools(mesh, { intervalMs: 250 })) {
  // 'added' | 'removed' | 'node_count_changed'
  console.log(change.type, change.descriptor.toolId)
}

// Provider lowering + invocation.
const apiTools = tools.map(openai.toOpenaiTool)
// ... POST to /v1/chat/completions with `tools: apiTools` ...
const spec = openai.lowerOpenaiToolCall(replyToolCall)
const resp = await callTool(rpc, spec.name, JSON.parse(spec.argumentsJson))

// On-demand full schema (when fold dropped it under size budget):
const meta = await fetchToolMetadata(rpc, hostNodeId, 'web_search')
if (meta.type === 'found') { /* meta.descriptor.inputSchema is parseable */ }

// Streaming tool invocation — capability-routed, drains ToolEvents.
// The wrapper synthesizes a `missing_terminal` error envelope if the
// host's handler exits without a result/error (matches T-2).
for await (const event of callToolStreaming(rpc, 'web_search', { query: '...' })) {
  if (event.type === 'delta')    { /* event.data is the partial result */ }
  if (event.type === 'progress') { /* event.pct, event.message */ }
  if (event.type === 'result' || event.type === 'error') break
}
```

Streaming server-side (`serveToolStreaming`) is deferred — see the
plan's "Deferred follow-ups" section.

### Python (`net-mesh`)

```python
from net.tool import (
    call_tool,
    fetch_tool_metadata,
    list_tools,
    serve_tool,
    watch_tools,
    openai,
)

# Server: register a tool (sync handler shown; async also supported
# via `AsyncTypedMeshRpc.serve`).
handle = serve_tool(
    rpc,
    {
        "name": "web_search",
        "description": "Search the web for relevant pages.",
        "input_schema": {"type": "object", "properties": {"query": {"type": "string"}}},
        "tags": ["web", "research"],
        "estimated_time_ms": 500,
    },
    lambda req: {"results": [f"hit for {req['query']}"]},
)
# handle.close() to deregister.

# Discovery.
tools = list_tools(mesh)
async for change in watch_tools(mesh, interval=0.25):
    print(change.type, change.descriptor.tool_id)

# Provider lowering + invocation.
api_tools = [openai.to_openai_tool(t) for t in tools]
spec = openai.lower_openai_tool_call(reply_tool_call)
resp = call_tool(rpc, spec.name, json.loads(spec.arguments_json))

# On-demand metadata fetch.
meta = fetch_tool_metadata(rpc, host_node_id, "web_search")
if meta["type"] == "found": ...  # meta["descriptor"]

# Streaming tool invocation — sync (returns an iterator of dicts).
from net.tool import call_tool_streaming
for event in call_tool_streaming(rpc, "web_search", {"query": "..."}):
    if event["type"] == "delta": ...
    if event["type"] in ("result", "error"): break

# Async equivalent for AsyncTypedMeshRpc callers.
from net.tool import call_tool_streaming_async
async for event in call_tool_streaming_async(arpc, "web_search", {"query": "..."}):
    if event["type"] == "delta": ...
    if event["type"] in ("result", "error"): break
```

Pydantic models work seamlessly — pass `MyModel.model_json_schema()`
as the `input_schema` and the handler receives a dict that you
deserialize with `MyModel.model_validate(req)`. Asyncio callers
should use `serve_tool_async` + `call_tool_async` from
`net.tool` (they accept `AsyncTypedMeshRpc` and route async
handlers through the substrate's tokio runtime via the
async-bridge).

### Go (`github.com/ai-2070/net/go`)

```go
import (
    "context"
    "github.com/ai-2070/net/go"
)

// Server: register a tool (typed handler, generics).
desc, _ := net.DescriptorFor(net.ToolOptions{
    Name:        "web_search",
    Description: "Search the web for relevant pages.",
    InputSchema: map[string]any{"type": "object", "properties": map[string]any{
        "query": map[string]any{"type": "string"},
    }},
    Tags:            []string{"web", "research"},
    EstimatedTimeMs: 500,
})
handle, _ := net.RegisterTool[WebSearchReq, WebSearchResp](
    rpc, desc, func(req WebSearchReq) (WebSearchResp, error) {
        return WebSearchResp{Results: []string{"hit for " + req.Query}}, nil
    },
)
// defer handle.Close()

// Provider lowering + invocation.
apiTool := net.ToOpenAITool(desc)
spec, _ := net.LowerOpenAIToolCall(replyMap)
var req WebSearchReq
json.Unmarshal([]byte(spec.ArgumentsJSON), &req)
resp, _ := net.CallTool[WebSearchReq, WebSearchResp](
    context.Background(), rpc, spec.Name, req,
)

// On-demand metadata fetch.
meta, _ := net.FetchToolMetadata(context.Background(), rpc, hostNodeID, "web_search")
if meta.Type == "found" { /* meta.Descriptor */ }

// Discovery (local fold).
tools, _ := net.ListTools(rpc)
for _, t := range tools { fmt.Println(t.ToolID, t.NodeCount) }

// Dynamic watch via polling (1s default; ctx-cancel ends the loop).
ctx, cancel := context.WithCancel(context.Background())
defer cancel()
changes, errs, baseline, _ := net.WatchTools(ctx, rpc, net.WatchOptions{})
go func() {
    for change := range changes {
        switch change.Type {
        case "added":              fmt.Println("+", change.Descriptor.ToolID, change.Descriptor.Version)
        case "removed":            fmt.Println("-", change.Descriptor.ToolID, change.Descriptor.Version)
        case "node_count_changed": fmt.Println("~", change.Descriptor.ToolID, change.Descriptor.NodeCount)
        }
    }
}()
// drain errs separately if you care about transient poll failures
_ = baseline

// Streaming tool invocation — capability-routed; iterates ToolEvents.
stream, _ := net.CallToolStreaming[WebSearchReq](
    context.Background(), rpc, "web_search", WebSearchReq{Query: "..."},
)
defer stream.Close()
for {
    event, ok, err := stream.Recv()
    if err != nil  { /* handle */ break }
    if !ok          { break }
    if event.IsTerminal() { break }
    // event.Type == "delta" / "progress" / "start" — render incrementally
}
```

Streaming server-side ergonomics (`RegisterStreamingTool`) are
deferred — see the plan's "Deferred follow-ups" section. Until
that lands, handlers that need to emit a stream should drive the
substrate's `serve_rpc_streaming` directly and encode `ToolEvent`
envelopes manually (the JSON shape is documented in the Streaming
section above).

### Provider format coverage

All three bindings ship the same four provider translators with
identical wire shapes (pinned by T-1 golden vectors):

| Provider  | Descriptor → tool definition          | Provider reply → `ToolCallSpec`                |
|-----------|---------------------------------------|------------------------------------------------|
| OpenAI    | `openai.to_openai_tool(desc)`         | `openai.lower_openai_tool_call(call)`          |
| Anthropic | `anthropic.to_anthropic_tool(desc)`   | `anthropic.lower_anthropic_tool_use(block)`    |
| MCP       | `mcp.to_mcp_tool(desc)`               | `mcp.lower_mcp_tools_call(params)`             |
| Gemini    | `gemini.to_gemini_function_declaration(desc)` | `gemini.lower_gemini_function_call(call)` |

Same names + signatures across all four languages (snake_case in
Python/Go, camelCase in Node TS). Empty input-schema fallback to
`{"type": "object", "properties": {}}` is uniform.

## Plan reference

See `docs/internal/plans/NRPC_AI_TOOL_CALLING_AND_AGENT_DX.md` for the full
slice-by-slice plan, locked design decisions, and Wave 3 / 4 follow-
ups (bindings + format translators).