| 1 | //! Dependency-neutral agent-run read model. |
| 2 | //! |
| 3 | //! [`AgentRunSnapshot`] is a uniform, serializable projection of "one unit of |
| 4 | //! agent work" regardless of which subsystem owns it: a direct sub-agent |
| 5 | //! worker, a workflow run, a fleet run, a core background job, or a managed |
| 6 | //! task. It carries serialized IDs, neutral enums, and scalar summaries only — |
| 7 | //! never handles, callbacks, or owner-internal types — so any surface can |
| 8 | //! consume it without depending on the owning subsystem. |
| 9 | //! |
| 10 | //! Ownership contract: |
| 11 | //! - Each owner keeps its own richer record type and maps it here through a |
| 12 | //! pure adapter function living in the owner's crate/module. This crate |
| 13 | //! depends on nothing new; owners depend on it, never the reverse. |
| 14 | //! - Adapters must apply the durable-outranks-live rule: when both a durable |
| 15 | //! record and a live in-memory handle exist, a terminal durable state always |
| 16 | //! wins over lagging live enrichment. Live state may only refine a run that |
| 17 | //! the durable record still considers in-flight. |
| 18 | //! - Adapters never fabricate values: unknown budgets and timestamps stay |
| 19 | //! `None`, and references are logical identifiers within the owner's |
| 20 | //! namespace — never filesystem paths and never secrets. |
| 21 | //! |
| 22 | //! This is a staging slice: the model compiles and is tested, but nothing |
| 23 | //! consumes it yet. Consumers arrive with later slices. |
| 24 | |
| 25 | use serde::{Deserialize, Serialize}; |
| 26 | |
| 27 | use super::Status; |
| 28 | |
| 29 | /// Which subsystem owns the run behind a snapshot. |
| 30 | #[derive(Debug, Clone, Copy, Serialize, Deserialize, PartialEq, Eq)] |
| 31 | #[serde(rename_all = "snake_case")] |
| 32 | pub enum RunSource { |
| 33 | /// A directly spawned sub-agent worker. |
| 34 | Direct, |
| 35 | /// A workflow VM run. |
| 36 | Workflow, |
| 37 | /// A fleet run reconstructed from the durable ledger. |
| 38 | Fleet, |
| 39 | /// A core background job. |
| 40 | CoreJob, |
| 41 | /// A managed background task. |
| 42 | Task, |
| 43 | } |
| 44 | |
| 45 | /// Neutral lifecycle state of a run. |
| 46 | /// |
| 47 | /// The two wait variants beyond model/tool waits exist because real owner |
| 48 | /// state machines report them: workers can wait on user input, and jobs and |
| 49 | /// fleet runs can be explicitly paused. Collapsing those into `Running` or |
| 50 | /// `Queued` would misreport liveness, so the read model keeps them distinct. |
| 51 | #[derive(Debug, Clone, Copy, Serialize, Deserialize, PartialEq, Eq)] |
| 52 | #[serde(rename_all = "snake_case")] |
| 53 | pub enum RunState { |
| 54 | /// Accepted but not started. |
| 55 | Queued, |
| 56 | /// Spawn intent recorded; startup in progress. |
| 57 | Initializing, |
| 58 | /// Actively executing. |
| 59 | Running, |
| 60 | /// Blocked on a model response. |
| 61 | WaitingModel, |
| 62 | /// Blocked on a tool execution. |
| 63 | WaitingTool, |
| 64 | /// Blocked on user input. |
| 65 | WaitingInput, |
| 66 | /// Explicitly paused by the user or system. |
| 67 | Paused, |
| 68 | /// Cancellation or shutdown requested, not yet terminal. |
| 69 | Stopping, |
| 70 | /// Finished; see [`AgentRunSnapshot::terminal`] for the outcome. |
| 71 | Terminal, |
| 72 | } |
| 73 | |
| 74 | impl Status for RunState { |
| 75 | fn is_terminal(&self) -> bool { |
| 76 | matches!(self, Self::Terminal) |
| 77 | } |
| 78 | fn is_active(&self) -> bool { |
| 79 | matches!( |
| 80 | self, |
| 81 | Self::Queued |
| 82 | | Self::Initializing |
| 83 | | Self::Running |
| 84 | | Self::WaitingModel |
| 85 | | Self::WaitingTool |
| 86 | | Self::WaitingInput |
| 87 | | Self::Stopping |
| 88 | ) |
| 89 | } |
| 90 | fn is_paused(&self) -> bool { |
| 91 | matches!(self, Self::Paused) |
| 92 | } |
| 93 | } |
| 94 | |
| 95 | /// Scalar budget/spend facts for a run. |
| 96 | /// |
| 97 | /// Every field is optional: owners report only what they actually track, and |
| 98 | /// adapters must not invent values for the rest. |
| 99 | #[derive(Debug, Clone, Copy, Default, Serialize, Deserialize, PartialEq, Eq)] |
| 100 | pub struct BudgetSummary { |
| 101 | /// Token ceiling configured for the run, if any. |
| 102 | #[serde(default, skip_serializing_if = "Option::is_none")] |
| 103 | pub token_budget: Option<u64>, |
| 104 | /// Tokens consumed so far, when the owner tracks spend. |
| 105 | #[serde(default, skip_serializing_if = "Option::is_none")] |
| 106 | pub tokens_used: Option<u64>, |
| 107 | /// Steps taken so far, when the owner counts steps. |
| 108 | #[serde(default, skip_serializing_if = "Option::is_none")] |
| 109 | pub steps_taken: Option<u32>, |
| 110 | /// Step ceiling configured for the run, if any. |
| 111 | #[serde(default, skip_serializing_if = "Option::is_none")] |
| 112 | pub max_steps: Option<u32>, |
| 113 | /// Wall-clock duration in milliseconds, when known. |
| 114 | #[serde(default, skip_serializing_if = "Option::is_none")] |
| 115 | pub duration_ms: Option<u64>, |
| 116 | } |
| 117 | |
| 118 | /// How a terminal run ended. |
| 119 | #[derive(Debug, Clone, Copy, Serialize, Deserialize, PartialEq, Eq)] |
| 120 | #[serde(rename_all = "snake_case")] |
| 121 | pub enum TerminalOutcome { |
| 122 | /// Ended successfully. |
| 123 | Completed, |
| 124 | /// Ended with an error. |
| 125 | Failed, |
| 126 | /// Cancelled by the user or a parent. |
| 127 | Cancelled, |
| 128 | /// Interrupted (e.g. process restart) before completion. |
| 129 | Interrupted, |
| 130 | /// Stopped because it exhausted its own budget. |
| 131 | BudgetExhausted, |
| 132 | } |
| 133 | |
| 134 | /// Scalar summary of a terminal run. |
| 135 | #[derive(Debug, Clone, Serialize, Deserialize, PartialEq, Eq)] |
| 136 | pub struct TerminalSummary { |
| 137 | pub outcome: TerminalOutcome, |
| 138 | /// Epoch milliseconds when the run ended, when the owner recorded it. |
| 139 | #[serde(default, skip_serializing_if = "Option::is_none")] |
| 140 | pub ended_at_ms: Option<i64>, |
| 141 | /// Short human-readable detail (result summary or error text). |
| 142 | /// Never raw logs, reasoning text, or secrets. |
| 143 | #[serde(default, skip_serializing_if = "Option::is_none")] |
| 144 | pub detail: Option<String>, |
| 145 | } |
| 146 | |
| 147 | /// Category of a durable reference attached to a run. |
| 148 | #[derive(Debug, Clone, Copy, Serialize, Deserialize, PartialEq, Eq)] |
| 149 | #[serde(rename_all = "snake_case")] |
| 150 | pub enum ReceiptKind { |
| 151 | /// A produced artifact. |
| 152 | Artifact, |
| 153 | /// A gate/verification result. |
| 154 | Gate, |
| 155 | /// A durable completion receipt. |
| 156 | Receipt, |
| 157 | /// An approval record. |
| 158 | Approval, |
| 159 | } |
| 160 | |
| 161 | /// Reference to a durable artifact, gate, receipt, or approval. |
| 162 | /// |
| 163 | /// `reference` is a logical identifier within the owner's namespace — never |
| 164 | /// an absolute filesystem path, never secret-bearing. |
| 165 | #[derive(Debug, Clone, Serialize, Deserialize, PartialEq, Eq)] |
| 166 | pub struct ReceiptRef { |
| 167 | pub kind: ReceiptKind, |
| 168 | pub reference: String, |
| 169 | /// Optional short human-readable label. |
| 170 | #[serde(default, skip_serializing_if = "Option::is_none")] |
| 171 | pub label: Option<String>, |
| 172 | } |
| 173 | |
| 174 | /// Dependency-neutral snapshot of a single agent run. |
| 175 | #[derive(Debug, Clone, Serialize, Deserialize, PartialEq, Eq)] |
| 176 | pub struct AgentRunSnapshot { |
| 177 | /// Serialized run identifier in the owner's existing scheme. |
| 178 | pub run_id: String, |
| 179 | /// Serialized identifier of the parent run/thread, when the owner |
| 180 | /// records one. |
| 181 | #[serde(default, skip_serializing_if = "Option::is_none")] |
| 182 | pub parent: Option<String>, |
| 183 | /// Which subsystem owns this run. |
| 184 | pub source: RunSource, |
| 185 | /// Neutral lifecycle state. |
| 186 | pub state: RunState, |
| 187 | /// Scalar budget facts; defaults to all-unknown. |
| 188 | #[serde(default)] |
| 189 | pub budget: BudgetSummary, |
| 190 | /// Terminal outcome; present exactly when `state == Terminal`. |
| 191 | #[serde(default, skip_serializing_if = "Option::is_none")] |
| 192 | pub terminal: Option<TerminalSummary>, |
| 193 | /// Durable references produced by the run. |
| 194 | #[serde(default, skip_serializing_if = "Vec::is_empty")] |
| 195 | pub refs: Vec<ReceiptRef>, |
| 196 | } |
| 197 | |
| 198 | impl AgentRunSnapshot { |
| 199 | /// `true` when the terminal summary and lifecycle state agree: |
| 200 | /// `state == Terminal` iff a terminal summary is present. |
| 201 | /// |
| 202 | /// Adapters uphold this by construction; tests assert it. |
| 203 | #[must_use] |
| 204 | pub fn is_coherent(&self) -> bool { |
| 205 | matches!(self.state, RunState::Terminal) == self.terminal.is_some() |
| 206 | } |
| 207 | } |
| 208 | |
| 209 | #[cfg(test)] |
| 210 | mod tests { |
| 211 | use super::*; |
| 212 | |
| 213 | fn full_snapshot() -> AgentRunSnapshot { |
| 214 | AgentRunSnapshot { |
| 215 | run_id: "agent_1234abcd".to_string(), |
| 216 | parent: Some("agent_00ff00ff".to_string()), |
| 217 | source: RunSource::Direct, |
| 218 | state: RunState::Terminal, |
| 219 | budget: BudgetSummary { |
| 220 | token_budget: Some(50_000), |
| 221 | tokens_used: Some(12_345), |
| 222 | steps_taken: Some(7), |
| 223 | max_steps: Some(40), |
| 224 | duration_ms: Some(93_000), |
| 225 | }, |
| 226 | terminal: Some(TerminalSummary { |
| 227 | outcome: TerminalOutcome::Completed, |
| 228 | ended_at_ms: Some(1_800_000_000_000), |
| 229 | detail: Some("verified build".to_string()), |
| 230 | }), |
| 231 | refs: vec![ReceiptRef { |
| 232 | kind: ReceiptKind::Artifact, |
| 233 | reference: "runs/agent_1234abcd/result.md".to_string(), |
| 234 | label: Some("result".to_string()), |
| 235 | }], |
| 236 | } |
| 237 | } |
| 238 | |
| 239 | fn minimal_snapshot() -> AgentRunSnapshot { |
| 240 | AgentRunSnapshot { |
| 241 | run_id: "job-42".to_string(), |
| 242 | parent: None, |
| 243 | source: RunSource::CoreJob, |
| 244 | state: RunState::Queued, |
| 245 | budget: BudgetSummary::default(), |
| 246 | terminal: None, |
| 247 | refs: Vec::new(), |
| 248 | } |
| 249 | } |
| 250 | |
| 251 | #[test] |
| 252 | fn full_snapshot_round_trips() { |
| 253 | let snapshot = full_snapshot(); |
| 254 | let json = serde_json::to_string(&snapshot).expect("serialize"); |
| 255 | let back: AgentRunSnapshot = serde_json::from_str(&json).expect("deserialize"); |
| 256 | assert_eq!(back, snapshot); |
| 257 | assert!(back.is_coherent()); |
| 258 | } |
| 259 | |
| 260 | #[test] |
| 261 | fn minimal_snapshot_round_trips_and_skips_empty_fields() { |
| 262 | let snapshot = minimal_snapshot(); |
| 263 | let json = serde_json::to_string(&snapshot).expect("serialize"); |
| 264 | // Optional/empty fields stay off the wire entirely. |
| 265 | assert!(!json.contains("parent")); |
| 266 | assert!(!json.contains("terminal")); |
| 267 | assert!(!json.contains("refs")); |
| 268 | assert!(!json.contains("token_budget")); |
| 269 | let back: AgentRunSnapshot = serde_json::from_str(&json).expect("deserialize"); |
| 270 | assert_eq!(back, snapshot); |
| 271 | assert!(back.is_coherent()); |
| 272 | } |
| 273 | |
| 274 | #[test] |
| 275 | fn enum_wire_names_are_snake_case_and_stable() { |
| 276 | assert_eq!( |
| 277 | serde_json::to_string(&RunSource::CoreJob).unwrap(), |
| 278 | "\"core_job\"" |
| 279 | ); |
| 280 | assert_eq!( |
| 281 | serde_json::to_string(&RunState::WaitingModel).unwrap(), |
| 282 | "\"waiting_model\"" |
| 283 | ); |
| 284 | assert_eq!( |
| 285 | serde_json::to_string(&TerminalOutcome::BudgetExhausted).unwrap(), |
| 286 | "\"budget_exhausted\"" |
| 287 | ); |
| 288 | assert_eq!( |
| 289 | serde_json::to_string(&ReceiptKind::Gate).unwrap(), |
| 290 | "\"gate\"" |
| 291 | ); |
| 292 | // Every state deserializes back to itself. |
| 293 | for state in [ |
| 294 | RunState::Queued, |
| 295 | RunState::Initializing, |
| 296 | RunState::Running, |
| 297 | RunState::WaitingModel, |
| 298 | RunState::WaitingTool, |
| 299 | RunState::WaitingInput, |
| 300 | RunState::Paused, |
| 301 | RunState::Stopping, |
| 302 | RunState::Terminal, |
| 303 | ] { |
| 304 | let json = serde_json::to_string(&state).unwrap(); |
| 305 | let back: RunState = serde_json::from_str(&json).unwrap(); |
| 306 | assert_eq!(back, state); |
| 307 | } |
| 308 | } |
| 309 | |
| 310 | #[test] |
| 311 | fn run_state_status_trait_partitions_all_states() { |
| 312 | for state in [ |
| 313 | RunState::Queued, |
| 314 | RunState::Initializing, |
| 315 | RunState::Running, |
| 316 | RunState::WaitingModel, |
| 317 | RunState::WaitingTool, |
| 318 | RunState::WaitingInput, |
| 319 | RunState::Paused, |
| 320 | RunState::Stopping, |
| 321 | RunState::Terminal, |
| 322 | ] { |
| 323 | let classifications = [state.is_terminal(), state.is_active(), state.is_paused()]; |
| 324 | assert_eq!( |
| 325 | classifications.iter().filter(|flag| **flag).count(), |
| 326 | 1, |
| 327 | "state {state:?} must be exactly one of terminal/active/paused" |
| 328 | ); |
| 329 | } |
| 330 | } |
| 331 | |
| 332 | #[test] |
| 333 | fn incoherent_snapshot_is_detectable() { |
| 334 | let mut snapshot = minimal_snapshot(); |
| 335 | snapshot.terminal = Some(TerminalSummary { |
| 336 | outcome: TerminalOutcome::Completed, |
| 337 | ended_at_ms: None, |
| 338 | detail: None, |
| 339 | }); |
| 340 | assert!(!snapshot.is_coherent()); |
| 341 | } |
| 342 | } |
| 343 |