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
//! Script upstream client — a thin wrapper around the client-side
//! python runtime.
//!
//! The whole "upstream" is one reverse-channel RPC: assemble the FULL
//! conversation (request continuation ++ messages ++ internal
//! continuation items, exactly like mock), ship it with the agent's
//! script definition as [`server_request::Payload::Script`], and let
//! the CLI daemon run the code on its embedded runtime (the same one
//! the `python` command uses). The reply's assistant/tool-only output
//! messages become the agent's SOLE yielded chunk; usage carries no
//! token counts — only the measured wall-clock duration (plus its
//! duration charge).
use std::pin::Pin;
use futures::Stream;
use objectiveai_sdk::agent::completions::response::streaming::{
AgentCompletionChunk, AssistantResponseChunk, MessageChunk,
};
use super::super::{ContinuationItem, StreamItem};
type Message = objectiveai_sdk::agent::completions::message::Message;
/// Upstream client for Script agents. Stateless — the per-request
/// reverse-attach handle arrives per `create` call.
pub struct Client;
impl
super::super::UpstreamClient<
objectiveai_sdk::agent::script::Agent,
objectiveai_sdk::agent::script::Continuation,
> for Client
{
/// One turn's output as conversation messages — a script may
/// append several (assistant/tool) messages per turn, so the
/// state is the whole slice rather than a single assistant
/// message.
type State = Vec<Message>;
type Stream = Pin<Box<dyn Stream<Item = StreamItem<Self::State>> + Send>>;
type Error = super::Error;
fn create(
&self,
id: &str,
created: u64,
agent: &objectiveai_sdk::agent::script::Agent,
request_continuation: Option<&objectiveai_sdk::agent::script::Continuation>,
_params: &objectiveai_sdk::agent::completions::request::AgentCompletionCreateParams,
messages: &[Message],
_mcp_connection: Option<objectiveai_sdk::mcp::Connection>,
reverse_attach: Option<std::sync::Arc<crate::objectiveai_mcp::ReverseAttachHandle>>,
continuation: Option<&[ContinuationItem<Self::State>]>,
_byok: Option<&str>,
cost_multiplier: rust_decimal::Decimal,
duration_cost: rust_decimal::Decimal,
_tools_enabled: bool,
agent_instance_hierarchy: &str,
agent_id_arg: &str,
agent_full_id: &str,
agent_remote: Option<&objectiveai_sdk::RemotePath>,
) -> impl Future<
Output = Result<
Self::Stream,
Self::Error,
>,
> + Send
+ 'static {
let id = id.to_string();
let agent_instance_hierarchy = agent_instance_hierarchy.to_string();
let agent_id_for_chunks = agent_id_arg.to_string();
let agent_full_id = agent_full_id.to_string();
let agent_remote = agent_remote.cloned();
let script = agent.base.script.clone();
// Build the full conversation: request_continuation -> messages
// -> continuation (State items splice their whole turn).
let rc_len = request_continuation.map_or(0, |rc| rc.messages.len());
let cont_len = continuation.map_or(0, |c| c.len());
let mut all_messages: Vec<Message> =
Vec::with_capacity(rc_len + messages.len() + cont_len);
if let Some(rc) = request_continuation {
all_messages.extend_from_slice(&rc.messages);
}
all_messages.extend_from_slice(messages);
if let Some(cont) = continuation {
for item in cont {
match item {
ContinuationItem::State(turn) => {
all_messages.extend_from_slice(turn);
}
ContinuationItem::ToolMessage(t) => {
all_messages.push(Message::Tool(t.clone()));
}
ContinuationItem::UserMessage(u) => {
all_messages.push(Message::User(u.clone()));
}
}
}
}
// Message-chunk index base: how many conversation positions the
// internal continuation already produced (same counting rule as
// mock's assistant_index).
let index_base = continuation
.map(|c| {
c.iter()
.filter(|item| {
matches!(
item,
ContinuationItem::State(_) | ContinuationItem::ToolMessage(_)
)
})
.count() as u64
})
.unwrap_or(0);
// `byok` is meaningless for scripts; carried onto the usage
// trailer as `is_byok: false` via the flag below.
let is_byok = false;
async move {
let Some(handle) = reverse_attach else {
return Err(super::Error::ReverseChannelUnavailable);
};
use objectiveai_sdk::client_objectiveai_mcp::{server_request, server_response};
let start = std::time::Instant::now();
let rc = handle.channel();
let request = server_request::Request {
id: uuid::Uuid::new_v4().to_string(),
headers: indexmap::IndexMap::new(),
payload: server_request::Payload::Script(server_request::ScriptRequest {
script,
messages: all_messages,
// TYPED identity (the non-MCP convention): the CLI
// applies these to the script's execution context,
// so anything the script runs via
// `objectiveai.execute` uses THIS agent's identity
// (its response id, agent full id, ...). The
// sibling response-id group isn't threaded into
// upstream clients — `None`.
agent_instance_hierarchy: agent_instance_hierarchy.clone(),
agent_id: agent_id_for_chunks.clone(),
agent_full_id: agent_full_id.clone(),
agent_remote: agent_remote
.as_ref()
.map(|r| serde_json::to_string(r).unwrap()),
response_id: id.clone(),
response_ids: None,
}),
};
// No timeout — like every other API-direct reverse-channel
// RPC, the await resolves when the reply arrives or errors
// when the WS drops (the recv loop's teardown drops the
// pending oneshot).
let rx = crate::objectiveai_mcp::send_server_request(&rc.sink, &rc.pending, request)
.await
.map_err(|()| super::Error::ChannelClosed)?;
let response = rx.await.map_err(|_| super::Error::ChannelDropped)?;
let output = match response.payload {
server_response::Payload::Script(server_response::JsonRpcResult::Ok {
result,
}) => result.messages,
server_response::Payload::Script(server_response::JsonRpcResult::Err {
code,
message,
..
}) => return Err(super::Error::Script { code, message }),
_ => return Err(super::Error::WrongVariant),
};
if output.is_empty() {
return Err(super::Error::EmptyOutput);
}
let elapsed_ms = start.elapsed().as_millis() as u64;
// Duration is the script agent's WHOLE usage: no token
// counts, just the measured wall time and its charge —
// added raw to both `cost` and `total_cost` (see
// `crate::duration`).
let duration_charge = crate::duration::duration_charge(elapsed_ms, duration_cost);
let mut upstream_duration_ms =
objectiveai_sdk::agent::completions::response::UpstreamDurationMs::default();
upstream_duration_ms.script = Some(elapsed_ms);
let usage = objectiveai_sdk::agent::completions::response::UpstreamUsage {
completion_tokens: 0,
prompt_tokens: 0,
total_tokens: 0,
completion_tokens_details: None,
prompt_tokens_details: None,
cost: duration_charge,
cost_details: None,
total_cost: duration_charge,
upstream_duration_ms,
cost_multiplier,
is_byok,
};
let mut usage = Some(usage);
// The turn's state: the output as conversation messages.
let state: Vec<Message> =
output.iter().cloned().map(Message::from).collect();
// The SOLE chunk: every output message as its own
// MessageChunk. The LAST assistant carries the usage
// trailer; if the output has no assistant at all, an empty
// usage-only assistant chunk is appended (the mock trailer
// pattern) so the run loop still folds the usage.
let last_assistant = output
.iter()
.rposition(|m| {
matches!(
m,
objectiveai_sdk::agent::script::OutputMessage::Assistant(_)
)
});
let mut message_chunks: Vec<MessageChunk> = Vec::with_capacity(output.len() + 1);
for (i, message) in output.into_iter().enumerate() {
let index = index_base + i as u64;
match message {
objectiveai_sdk::agent::script::OutputMessage::Assistant(asst) => {
let tool_calls = asst.tool_calls.map(|tcs| {
tcs.into_iter()
.enumerate()
.map(|(k, tc)| {
let objectiveai_sdk::agent::completions::message::AssistantToolCall::Function { id, function } = tc;
objectiveai_sdk::agent::completions::message::AssistantToolCallDelta {
index: k as u64,
r#type: Some(objectiveai_sdk::agent::completions::message::AssistantToolCallType::Function),
id: Some(id),
function: Some(objectiveai_sdk::agent::completions::message::AssistantToolCallFunctionDelta {
name: Some(function.name),
arguments: Some(function.arguments),
}),
}
})
.collect::<Vec<_>>()
});
let finish_reason = if tool_calls.is_some() {
objectiveai_sdk::agent::completions::response::FinishReason::ToolCalls
} else {
objectiveai_sdk::agent::completions::response::FinishReason::Stop
};
let chunk_usage = if Some(i) == last_assistant {
usage.take()
} else {
None
};
message_chunks.push(MessageChunk::Assistant(AssistantResponseChunk {
index,
created,
model: "script".into(),
upstream_id: id.clone(),
reasoning: asst.reasoning,
tool_calls,
content: asst.content,
refusal: asst.refusal,
finish_reason: Some(finish_reason),
usage: chunk_usage,
..Default::default()
}));
}
objectiveai_sdk::agent::script::OutputMessage::Tool(tool) => {
message_chunks.push(MessageChunk::Tool(
objectiveai_sdk::agent::completions::response::ToolResponse {
role: Default::default(),
index,
inner: tool,
request_message_ids: None,
},
));
}
}
}
if let Some(usage) = usage.take() {
// No assistant message in the output — append the
// usage-only trailer chunk.
message_chunks.push(MessageChunk::Assistant(AssistantResponseChunk {
index: index_base + message_chunks.len() as u64,
created,
model: "script".into(),
upstream_id: id.clone(),
usage: Some(usage),
..Default::default()
}));
}
let chunk = AgentCompletionChunk {
id: id.clone(),
agent_instance_hierarchy: agent_instance_hierarchy.clone(),
agent_id: agent_id_for_chunks.clone(),
agent_full_id: agent_full_id.clone(),
agent_remote: agent_remote.clone(),
created,
messages: message_chunks,
object: Default::default(),
usage: None,
upstream: objectiveai_sdk::agent::Upstream::Script,
error: None,
continuation: None,
messages_queued: None,
};
let stream = async_stream::stream! {
yield StreamItem::Chunk(chunk);
yield StreamItem::State(state);
};
let boxed: Pin<Box<dyn Stream<Item = StreamItem<Self::State>> + Send>> =
Box::pin(stream);
Ok(boxed)
}
}
fn response_continuation(
&self,
request_continuation: Option<&objectiveai_sdk::agent::script::Continuation>,
messages: &[Message],
continuation: Option<&[ContinuationItem<Self::State>]>,
agent_instance_hierarchy: &str,
) -> objectiveai_sdk::agent::script::Continuation {
let rc_len = request_continuation.map_or(0, |rc| rc.messages.len());
let cont_len = continuation.map_or(0, |c| c.len());
let mut all_messages = Vec::with_capacity(rc_len + messages.len() + cont_len);
if let Some(rc) = request_continuation {
all_messages.extend_from_slice(&rc.messages);
}
all_messages.extend_from_slice(messages);
if let Some(cont) = continuation {
for item in cont {
match item {
ContinuationItem::State(turn) => {
all_messages.extend_from_slice(turn);
}
ContinuationItem::ToolMessage(t) => {
all_messages.push(Message::Tool(t.clone()));
}
ContinuationItem::UserMessage(u) => {
all_messages.push(Message::User(u.clone()));
}
}
}
}
objectiveai_sdk::agent::script::Continuation {
upstream: objectiveai_sdk::agent::script::Upstream::default(),
agent_instance_hierarchy: agent_instance_hierarchy.to_string(),
messages: all_messages,
}
}
}