polyc_agent/handoff.rs
1//! Handoff (sub-agent transfer) primitive for the agent turn loop.
2//!
3//! # Why a reserved tool name, not a wire message
4//!
5//! Two clean designs exist for "the parent wants to spawn a child agent":
6//!
7//! 1. A first-class wire message — e.g. a new `AgentResponse` oneof variant
8//! the planner emits. The control plane sees the variant on the response
9//! stream and reacts.
10//! 2. A reserved *tool name* the model can call from inside the standard
11//! function-calling loop. The tool executor recognises the name, doesn't
12//! execute anything, and surfaces it as a structured break-out of
13//! `run_turn`.
14//!
15//! We pick **(2) reserved tool name** (`HANDOFF_TOOL_NAME`) because:
16//!
17//! * It rides the existing provider function-calling shape (every modern
18//! provider models "call function X with JSON args"). No new channel is
19//! needed, no provider integration is touched.
20//! * It keeps the agent crate the single authority on the turn loop — the
21//! handoff is "the model asked to delegate", uniformly across providers.
22//! * Convergent with prior-art agent SDKs that advertise handoffs to the
23//! model as tools (turning the choice of delegation into a function the
24//! planner can reason about with the rest of its toolbox).
25//!
26//! Wire-level surfacing happens upstream in the control plane: when
27//! `run_turn` returns a [`TurnResult`](crate::TurnResult) with a populated
28//! [`TurnResult::handoff`](crate::TurnResult), the control plane emits a
29//! signed [`polyc_proto::proto::polychrome::handoff::v1::Handoff`] event
30//! into the *parent* conversation's journal, creates a child `Conversation`
31//! resource, and suspends the parent until a
32//! `polyc_proto::proto::polychrome::handoff::v1::HandoffReturn` event
33//! lands.
34
35use polyc_llm::{Message as LlmMessage, ToolSpec};
36
37/// The reserved tool name the model emits to request a sub-agent handoff.
38///
39/// Any [`crate::ToolExecutor`] implementation that mixes user tools with the
40/// handoff primitive must avoid using this name for a real tool — the runtime
41/// short-circuits the name before [`crate::ToolExecutor::execute`] is invoked.
42pub const HANDOFF_TOOL_NAME: &str = "__handoff_to";
43
44/// JSON-schema spec for the handoff tool. Provided alongside the user's tool
45/// specs so the model knows the shape of `handoff_to(child_agent_id, ...)`.
46/// Use [`handoff_tool_spec`] to obtain it.
47#[must_use]
48pub fn handoff_tool_spec() -> ToolSpec {
49 ToolSpec {
50 name: HANDOFF_TOOL_NAME.to_owned(),
51 description:
52 "Delegate the current task to a child sub-agent. The current turn suspends; the \
53 child runs in its own isolated conversation with the carried context; the child's \
54 final message is returned to this conversation on the next turn. Use sparingly — \
55 handoffs are expensive (cold-start a sandbox). `child_agent_id` selects the \
56 planner; `reason` is recorded for operator visibility; `max_carry` bounds how many \
57 of the most recent messages the child sees (default 5)."
58 .to_owned(),
59 schema_json: serde_json::json!({
60 "type": "object",
61 "properties": {
62 "child_agent_id": {
63 "type": "string",
64 "description": "Identifier of the child agent / planner to spawn."
65 },
66 "reason": {
67 "type": "string",
68 "description": "Short rationale for the delegation."
69 },
70 "max_carry": {
71 "type": "integer",
72 "minimum": 0,
73 "description": "How many of the most recent messages from this conversation to inject as the child's initial transcript. Default 5."
74 }
75 },
76 "required": ["child_agent_id"]
77 }),
78 title: None,
79 // Handoff is a control-plane delegation handled by the harness's own
80 // suspend/resume path, not the HITL spend gate; leave it ungated here.
81 needs_approval: false,
82 }
83}
84
85/// Parsed `__handoff_to` arguments, surfaced as
86/// [`crate::TurnResult::handoff`].
87///
88/// The control plane consumes this to:
89/// 1. Slice the parent's transcript per [`HandoffRequest::carried_context`]
90/// (the agent crate has already done the slicing).
91/// 2. Sign-and-write a `Handoff` event to the parent's eventlog partition.
92/// 3. Create the child `Conversation` resource.
93#[derive(Debug, Clone)]
94pub struct HandoffRequest {
95 /// Tool-call id the provider assigned to the `__handoff_to` call. Echoed
96 /// back to the provider on the parent's *next* turn as the `tool_result`
97 /// id, so the function-calling loop on the parent's resumed turn sees a
98 /// matched call → result pair (otherwise the provider rejects the turn
99 /// for a dangling call).
100 pub call_id: String,
101 /// The model's chosen child agent identifier.
102 pub child_agent_id: String,
103 /// Optional free-form reason captured for operator visibility.
104 pub reason: String,
105 /// Sliding-window cap on the carried transcript. The runtime applies it
106 /// against the *current* turn's transcript when packaging the handoff.
107 pub max_carry: usize,
108 /// The pre-sliced carried transcript — the last `max_carry` messages of
109 /// the parent's transcript at the moment of the handoff call. Stored as
110 /// llm messages; the control plane maps them to wire `Message`s on the
111 /// way into the `Handoff` event payload.
112 pub carried_context: Vec<LlmMessage>,
113}
114
115/// Default carried-context window when the model omits `max_carry`.
116///
117/// Kept small — the child gets a fresh sandbox / new sliding window, so
118/// dragging in the parent's full context defeats the isolation a handoff is
119/// for. A future variant accepts a per-handoff `input_filter` predicate (a
120/// projection from the parent's transcript to the child's seed).
121pub const DEFAULT_MAX_CARRY: usize = 5;
122
123/// Parse the JSON arguments of a `__handoff_to` tool call into a structured
124/// [`HandoffRequest`].
125///
126/// `transcript_so_far` is the parent's *current* turn messages — the runtime
127/// slices the tail per `max_carry` (or [`DEFAULT_MAX_CARRY`] when absent)
128/// into `carried_context`.
129///
130/// Returns `None` if `args_json` doesn't parse or the required
131/// `child_agent_id` is missing — the runtime then treats the call as a
132/// no-op and lets the loop continue (so a malformed call doesn't deadlock
133/// the turn).
134#[must_use]
135pub fn parse_handoff_args(
136 call_id: &str,
137 args_json: &str,
138 transcript_so_far: &[LlmMessage],
139) -> Option<HandoffRequest> {
140 let v: serde_json::Value = serde_json::from_str(args_json).ok()?;
141 let child_agent_id = v.get("child_agent_id")?.as_str()?.to_owned();
142 if child_agent_id.is_empty() {
143 return None;
144 }
145 let reason = v
146 .get("reason")
147 .and_then(serde_json::Value::as_str)
148 .unwrap_or("")
149 .to_owned();
150 // Clamp max_carry to the actual transcript length so the slice never
151 // panics; a malicious model asking for u64::MAX still produces a sane
152 // slice (the whole transcript).
153 #[allow(clippy::cast_possible_truncation)]
154 let raw_max_carry = v
155 .get("max_carry")
156 .and_then(serde_json::Value::as_u64)
157 .map_or(DEFAULT_MAX_CARRY, |n| n as usize);
158 let max_carry = raw_max_carry.min(transcript_so_far.len());
159 let start = transcript_so_far.len().saturating_sub(max_carry);
160 let carried_context = transcript_so_far[start..].to_vec();
161 Some(HandoffRequest {
162 call_id: call_id.to_owned(),
163 child_agent_id,
164 reason,
165 max_carry,
166 carried_context,
167 })
168}
169
170#[cfg(test)]
171mod tests {
172 #![allow(clippy::pedantic, clippy::nursery, missing_docs)]
173
174 use polyc_llm::Message as LlmMessage;
175
176 use super::*;
177
178 fn transcript(n: usize) -> Vec<LlmMessage> {
179 (0..n)
180 .map(|i| {
181 if i % 2 == 0 {
182 LlmMessage::user(format!("u{i}"))
183 } else {
184 LlmMessage::assistant(format!("a{i}"))
185 }
186 })
187 .collect()
188 }
189
190 #[test]
191 fn parses_minimum_required_args() {
192 let t = transcript(10);
193 let h = parse_handoff_args("c-1", r#"{"child_agent_id":"researcher"}"#, &t).unwrap();
194 assert_eq!(h.child_agent_id, "researcher");
195 assert_eq!(h.max_carry, DEFAULT_MAX_CARRY);
196 assert_eq!(h.carried_context.len(), DEFAULT_MAX_CARRY);
197 assert_eq!(h.call_id, "c-1");
198 }
199
200 #[test]
201 fn slices_last_n_messages() {
202 let t = transcript(10);
203 let h = parse_handoff_args(
204 "c-2",
205 r#"{"child_agent_id":"x","max_carry":3,"reason":"because"}"#,
206 &t,
207 )
208 .unwrap();
209 assert_eq!(h.max_carry, 3);
210 assert_eq!(h.carried_context.len(), 3);
211 assert_eq!(h.reason, "because");
212 // The tail of the transcript.
213 let last_text = match h.carried_context.last().unwrap().content.first().unwrap() {
214 polyc_llm::Content::Text(s) => s.clone(),
215 _ => panic!("expected text"),
216 };
217 assert_eq!(last_text, "a9");
218 }
219
220 #[test]
221 fn clamps_max_carry_to_transcript_length() {
222 let t = transcript(2);
223 let h = parse_handoff_args("c", r#"{"child_agent_id":"x","max_carry":1000}"#, &t).unwrap();
224 assert_eq!(h.max_carry, 2, "clamped to len");
225 assert_eq!(h.carried_context.len(), 2);
226 }
227
228 #[test]
229 fn rejects_missing_child_agent_id() {
230 let t = transcript(2);
231 assert!(parse_handoff_args("c", r#"{"reason":"x"}"#, &t).is_none());
232 }
233
234 #[test]
235 fn rejects_empty_child_agent_id() {
236 let t = transcript(2);
237 assert!(parse_handoff_args("c", r#"{"child_agent_id":""}"#, &t).is_none());
238 }
239
240 #[test]
241 fn rejects_garbage_json() {
242 let t = transcript(2);
243 assert!(parse_handoff_args("c", "not-json", &t).is_none());
244 }
245
246 #[test]
247 fn empty_transcript_yields_empty_carry() {
248 let h = parse_handoff_args("c", r#"{"child_agent_id":"x"}"#, &[]).unwrap();
249 assert_eq!(h.max_carry, 0);
250 assert!(h.carried_context.is_empty());
251 }
252
253 #[test]
254 fn handoff_tool_spec_has_required_field() {
255 let spec = handoff_tool_spec();
256 assert_eq!(spec.name, HANDOFF_TOOL_NAME);
257 let required = spec
258 .schema_json
259 .get("required")
260 .and_then(|v| v.as_array())
261 .cloned()
262 .unwrap_or_default();
263 assert!(required.iter().any(|v| v == "child_agent_id"));
264 }
265}