vtcode_exec_events/
atif.rs

1//! Agent Trajectory Interchange Format (ATIF) types and builder.
2//!
3//! Implements the [ATIF specification](https://github.com/laude-institute/harbor/blob/main/docs/rfcs/0001-trajectory-format.md)
4//! v1.4 for logging complete agent interaction histories in a standardized,
5//! JSON-based format usable across debugging, visualization, SFT, and RL
6//! pipelines.
7//!
8//! # Overview
9//!
10//! ATIF provides a complete session trajectory: user messages, agent responses,
11//! tool executions, observations, and per-step/aggregate LLM metrics. The
12//! [`AtifTrajectoryBuilder`] converts live [`ThreadEvent`]
13//! streams into a finished [`Trajectory`].
14//!
15//! # Example
16//!
17//! ```rust
18//! use vtcode_exec_events::atif::*;
19//!
20//! let agent = AtifAgent::new("vtcode", env!("CARGO_PKG_VERSION"));
21//! let mut builder = AtifTrajectoryBuilder::new(agent);
22//!
23//! // Feed ThreadEvents as they arrive …
24//! // builder.process_event(&event);
25//!
26//! let trajectory = builder.finish(None);
27//! let json = serde_json::to_string_pretty(&trajectory).unwrap();
28//! ```
29
30use chrono::{DateTime, Utc};
31use serde::{Deserialize, Serialize};
32use serde_json::Value;
33
34use crate::{ThreadEvent, ThreadItemDetails, ToolCallStatus};
35
36/// Current ATIF schema version supported by this implementation.
37const ATIF_SCHEMA_VERSION: &str = "ATIF-v1.4";
38
39// ============================================================================
40// Core ATIF Types
41// ============================================================================
42
43/// Root-level ATIF trajectory object.
44#[derive(Debug, Clone, Serialize, Deserialize)]
45pub struct Trajectory {
46    /// ATIF schema version (e.g., "ATIF-v1.4").
47    schema_version: String,
48    /// Unique identifier for the entire agent run.
49    session_id: String,
50    /// Agent configuration for this trajectory.
51    agent: AtifAgent,
52    /// Ordered interaction steps.
53    steps: Vec<Step>,
54    /// Optional developer notes.
55    #[serde(skip_serializing_if = "Option::is_none")]
56    notes: Option<String>,
57    /// Aggregate metrics for the full trajectory.
58    #[serde(skip_serializing_if = "Option::is_none")]
59    pub final_metrics: Option<FinalMetrics>,
60    /// Optional custom root-level metadata.
61    #[serde(skip_serializing_if = "Option::is_none")]
62    extra: Option<Value>,
63}
64
65/// Agent configuration metadata.
66#[derive(Debug, Clone, Serialize, Deserialize)]
67pub struct AtifAgent {
68    /// Agent system name (e.g., "vtcode").
69    name: String,
70    /// Agent system version.
71    version: String,
72    /// Default LLM model used. Step-level `model_name` overrides this.
73    #[serde(skip_serializing_if = "Option::is_none")]
74    model_name: Option<String>,
75    /// Optional custom agent metadata.
76    #[serde(skip_serializing_if = "Option::is_none")]
77    extra: Option<Value>,
78}
79
80impl AtifAgent {
81    /// Create a new agent descriptor.
82    pub fn new(name: impl Into<String>, version: impl Into<String>) -> Self {
83        Self {
84            name: name.into(),
85            version: version.into(),
86            model_name: None,
87            extra: None,
88        }
89    }
90
91    /// Create a vtcode agent descriptor using the crate version.
92    pub fn vtcode() -> Self {
93        Self::new("vtcode", env!("CARGO_PKG_VERSION"))
94    }
95
96    /// Set the default model name.
97    pub fn with_model(mut self, model: impl Into<String>) -> Self {
98        self.model_name = Some(model.into());
99        self
100    }
101}
102
103/// The originator of a step.
104#[derive(Debug, Clone, Copy, Serialize, Deserialize, PartialEq, Eq)]
105#[serde(rename_all = "lowercase")]
106pub enum StepSource {
107    /// System prompt or system-initiated operation.
108    System,
109    /// User message.
110    User,
111    /// Agent response.
112    Agent,
113}
114
115/// Individual interaction step.
116#[derive(Debug, Clone, Serialize, Deserialize)]
117pub struct Step {
118    /// Ordinal index (starting from 1).
119    step_id: u64,
120    /// ISO 8601 timestamp.
121    #[serde(skip_serializing_if = "Option::is_none")]
122    timestamp: Option<String>,
123    /// Originator of this step.
124    source: StepSource,
125    /// LLM model used for this step (agent steps only).
126    #[serde(skip_serializing_if = "Option::is_none")]
127    model_name: Option<String>,
128    /// Step content — text message or array.
129    #[serde(skip_serializing_if = "Option::is_none")]
130    message: Option<String>,
131    /// Agent internal reasoning content.
132    #[serde(skip_serializing_if = "Option::is_none")]
133    reasoning_content: Option<String>,
134    /// Tool/function invocations (agent steps only).
135    #[serde(skip_serializing_if = "Option::is_none")]
136    tool_calls: Option<Vec<AtifToolCall>>,
137    /// Environment feedback after actions.
138    #[serde(skip_serializing_if = "Option::is_none")]
139    observation: Option<Observation>,
140    /// LLM operational metrics (agent steps only).
141    #[serde(skip_serializing_if = "Option::is_none")]
142    metrics: Option<StepMetrics>,
143    /// Custom step-level metadata.
144    #[serde(skip_serializing_if = "Option::is_none")]
145    extra: Option<Value>,
146}
147
148impl Step {
149    /// Create a user step.
150    fn user(step_id: u64, message: impl Into<String>) -> Self {
151        Self {
152            step_id,
153            timestamp: Some(Utc::now().to_rfc3339()),
154            source: StepSource::User,
155            model_name: None,
156            message: Some(message.into()),
157            reasoning_content: None,
158            tool_calls: None,
159            observation: None,
160            metrics: None,
161            extra: None,
162        }
163    }
164
165    /// Create an agent step.
166    fn agent(step_id: u64, message: impl Into<String>) -> Self {
167        Self {
168            step_id,
169            timestamp: Some(Utc::now().to_rfc3339()),
170            source: StepSource::Agent,
171            model_name: None,
172            message: Some(message.into()),
173            reasoning_content: None,
174            tool_calls: None,
175            observation: None,
176            metrics: None,
177            extra: None,
178        }
179    }
180
181    /// Create a system step.
182    fn system(step_id: u64, message: impl Into<String>) -> Self {
183        Self {
184            step_id,
185            timestamp: Some(Utc::now().to_rfc3339()),
186            source: StepSource::System,
187            model_name: None,
188            message: Some(message.into()),
189            reasoning_content: None,
190            tool_calls: None,
191            observation: None,
192            metrics: None,
193            extra: None,
194        }
195    }
196}
197
198/// Structured tool/function invocation.
199#[derive(Debug, Clone, Serialize, Deserialize)]
200pub struct AtifToolCall {
201    /// Unique identifier for the tool call.
202    tool_call_id: String,
203    /// Function/tool name.
204    function_name: String,
205    /// Arguments passed to the tool.
206    #[serde(skip_serializing_if = "Option::is_none")]
207    arguments: Option<Value>,
208}
209
210/// Environment feedback container.
211#[derive(Debug, Clone, Serialize, Deserialize)]
212pub struct Observation {
213    /// Results from tool calls or system operations.
214    results: Vec<ObservationResult>,
215}
216
217/// Individual observation result tied to a tool call.
218#[derive(Debug, Clone, Serialize, Deserialize)]
219pub struct ObservationResult {
220    /// Identifier of the originating tool call.
221    source_call_id: String,
222    /// Content/output of the observation.
223    content: String,
224}
225
226/// Per-step LLM operational metrics.
227#[derive(Debug, Clone, Serialize, Deserialize)]
228pub struct StepMetrics {
229    /// Total input tokens for this step (cached + non-cached).
230    #[serde(skip_serializing_if = "Option::is_none")]
231    prompt_tokens: Option<u64>,
232    /// Completion tokens generated.
233    #[serde(skip_serializing_if = "Option::is_none")]
234    completion_tokens: Option<u64>,
235    /// Subset of prompt_tokens that were cache hits.
236    #[serde(skip_serializing_if = "Option::is_none")]
237    cached_tokens: Option<u64>,
238    /// Estimated cost in USD for this step.
239    #[serde(skip_serializing_if = "Option::is_none")]
240    cost_usd: Option<f64>,
241    /// Log probabilities for completion tokens.
242    #[serde(skip_serializing_if = "Option::is_none")]
243    logprobs: Option<Vec<f64>>,
244    /// Completion token IDs for RL training.
245    #[serde(skip_serializing_if = "Option::is_none")]
246    completion_token_ids: Option<Vec<u64>>,
247    /// Prompt token IDs.
248    #[serde(skip_serializing_if = "Option::is_none")]
249    prompt_token_ids: Option<Vec<u64>>,
250    /// Custom metrics.
251    #[serde(skip_serializing_if = "Option::is_none")]
252    extra: Option<Value>,
253}
254
255impl StepMetrics {
256    /// Create metrics from vtcode Usage.
257    fn from_usage(usage: &crate::Usage) -> Self {
258        Self {
259            prompt_tokens: Some(usage.input_tokens),
260            completion_tokens: Some(usage.output_tokens),
261            cached_tokens: if usage.cached_input_tokens > 0 {
262                Some(usage.cached_input_tokens)
263            } else {
264                None
265            },
266            cost_usd: None,
267            logprobs: None,
268            completion_token_ids: None,
269            prompt_token_ids: None,
270            extra: if usage.cache_creation_tokens > 0 {
271                Some(serde_json::json!({
272                    "cache_creation_tokens": usage.cache_creation_tokens
273                }))
274            } else {
275                None
276            },
277        }
278    }
279}
280
281/// Trajectory-level aggregate metrics.
282#[derive(Debug, Clone, Default, Serialize, Deserialize)]
283pub struct FinalMetrics {
284    /// Sum of all prompt tokens across steps.
285    #[serde(skip_serializing_if = "Option::is_none")]
286    pub total_prompt_tokens: Option<u64>,
287    /// Sum of all completion tokens across steps.
288    #[serde(skip_serializing_if = "Option::is_none")]
289    pub total_completion_tokens: Option<u64>,
290    /// Sum of all cached tokens across steps.
291    #[serde(skip_serializing_if = "Option::is_none")]
292    pub total_cached_tokens: Option<u64>,
293    /// Total estimated cost in USD.
294    #[serde(skip_serializing_if = "Option::is_none")]
295    total_cost_usd: Option<f64>,
296    /// Total number of steps.
297    #[serde(skip_serializing_if = "Option::is_none")]
298    total_steps: Option<u64>,
299    /// Custom aggregate metrics.
300    #[serde(skip_serializing_if = "Option::is_none")]
301    extra: Option<Value>,
302}
303
304// ============================================================================
305// Builder — converts live ThreadEvent streams into ATIF Trajectory
306// ============================================================================
307
308/// Stateful collector that converts a live [`ThreadEvent`] stream into an
309/// ATIF-compliant [`Trajectory`].
310///
311/// Feed events via [`process_event`](Self::process_event) (timestamps at
312/// observation time) or [`process_event_at`](Self::process_event_at)
313/// (deterministic timestamps for tests). Call [`finish`](Self::finish) to
314/// produce the final trajectory.
315pub struct AtifTrajectoryBuilder {
316    agent: AtifAgent,
317    session_id: Option<String>,
318    steps: Vec<Step>,
319    next_step_id: u64,
320    // Running token accumulators for final metrics
321    total_input_tokens: u64,
322    total_output_tokens: u64,
323    total_cached_tokens: u64,
324    num_turns: usize,
325    /// Pending tool invocations awaiting matching ToolOutput.
326    pending_tool_calls: Vec<PendingToolCall>,
327}
328
329struct PendingToolCall {
330    call_id: String,
331    tool_call_id: Option<String>,
332    tool_name: String,
333    arguments: Option<Value>,
334    timestamp: String,
335}
336
337impl AtifTrajectoryBuilder {
338    /// Create a new builder for the given agent.
339    pub fn new(agent: AtifAgent) -> Self {
340        Self {
341            agent,
342            session_id: None,
343            steps: Vec::new(),
344            next_step_id: 1,
345            total_input_tokens: 0,
346            total_output_tokens: 0,
347            total_cached_tokens: 0,
348            num_turns: 0,
349            pending_tool_calls: Vec::new(),
350        }
351    }
352
353    /// Set the session ID explicitly. If not set, it will be derived from
354    /// `ThreadStarted` or `ThreadCompleted` events.
355    pub fn set_session_id(&mut self, id: impl Into<String>) {
356        self.session_id = Some(id.into());
357    }
358
359    /// Process a thread event using the current wall-clock time.
360    pub fn process_event(&mut self, event: &ThreadEvent) {
361        self.process_event_at(event, Utc::now());
362    }
363
364    /// Process a thread event with an explicit timestamp (for deterministic tests).
365    pub fn process_event_at(&mut self, event: &ThreadEvent, ts: DateTime<Utc>) {
366        let ts_str = ts.to_rfc3339();
367        match event {
368            ThreadEvent::ThreadStarted(e) => {
369                if self.session_id.is_none() {
370                    self.session_id = Some(e.thread_id.clone());
371                }
372            }
373            ThreadEvent::ThreadCompleted(e) => {
374                if self.session_id.is_none() {
375                    self.session_id = Some(e.session_id.clone());
376                }
377                // Accumulate aggregate usage
378                self.total_input_tokens = self.total_input_tokens.saturating_add(e.usage.input_tokens);
379                self.total_output_tokens = self.total_output_tokens.saturating_add(e.usage.output_tokens);
380                self.total_cached_tokens = self.total_cached_tokens.saturating_add(e.usage.cached_input_tokens);
381                self.num_turns = e.num_turns;
382            }
383            ThreadEvent::TurnCompleted(e) => {
384                self.total_input_tokens = self.total_input_tokens.saturating_add(e.usage.input_tokens);
385                self.total_output_tokens = self.total_output_tokens.saturating_add(e.usage.output_tokens);
386                self.total_cached_tokens = self.total_cached_tokens.saturating_add(e.usage.cached_input_tokens);
387                self.num_turns += 1;
388
389                let mut step = Step::system(self.next_step_id, "turn_completed");
390                step.timestamp = Some(ts_str);
391                step.metrics = Some(StepMetrics::from_usage(&e.usage));
392                self.push_step(step);
393            }
394            ThreadEvent::TurnFailed(e) => {
395                if let Some(usage) = &e.usage {
396                    self.total_input_tokens = self.total_input_tokens.saturating_add(usage.input_tokens);
397                    self.total_output_tokens = self.total_output_tokens.saturating_add(usage.output_tokens);
398                }
399                let mut step = Step::system(self.next_step_id, &e.message);
400                step.timestamp = Some(ts_str);
401                step.metrics = e.usage.as_ref().map(StepMetrics::from_usage);
402                self.push_step(step);
403            }
404            ThreadEvent::ItemCompleted(e) => {
405                self.process_item_completed(&e.item.id, &e.item.details, &ts_str);
406            }
407            ThreadEvent::ThreadCompactBoundary(e) => {
408                let msg = format!(
409                    "context_compaction: {} messages -> {} messages ({})",
410                    e.original_message_count,
411                    e.compacted_message_count,
412                    e.trigger.as_str()
413                );
414                let mut step = Step::system(self.next_step_id, msg);
415                step.timestamp = Some(ts_str);
416                self.push_step(step);
417            }
418            ThreadEvent::Error(e) => {
419                let mut step = Step::system(self.next_step_id, &e.message);
420                step.timestamp = Some(ts_str);
421                self.push_step(step);
422            }
423            // Skip streaming/lifecycle events that don't map to ATIF steps
424            ThreadEvent::TurnStarted(_)
425            | ThreadEvent::ItemStarted(_)
426            | ThreadEvent::ItemUpdated(_)
427            | ThreadEvent::PlanDelta(_)
428            | ThreadEvent::PlanApprovalRequested(_)
429            | ThreadEvent::PlanApprovalResolved(_)
430            | ThreadEvent::PermissionRequested(_)
431            | ThreadEvent::PermissionResolved(_)
432            | ThreadEvent::Interjected(_)
433            | ThreadEvent::Unknown => {}
434        }
435    }
436
437    fn process_item_completed(&mut self, item_id: &str, details: &ThreadItemDetails, ts: &str) {
438        match details {
439            ThreadItemDetails::AgentMessage(msg) => {
440                let mut step = Step::agent(self.next_step_id, &msg.text);
441                step.timestamp = Some(ts.to_string());
442                self.push_step(step);
443            }
444            ThreadItemDetails::Plan(plan) => {
445                let mut step = Step::agent(self.next_step_id, &plan.text);
446                step.timestamp = Some(ts.to_string());
447                step.extra = Some(serde_json::json!({ "vtcode_item_type": "plan" }));
448                self.push_step(step);
449            }
450            ThreadItemDetails::Reasoning(r) => {
451                let mut step = Step::agent(self.next_step_id, "");
452                step.timestamp = Some(ts.to_string());
453                step.reasoning_content = Some(r.text.clone());
454                step.message = None;
455                self.push_step(step);
456            }
457            ThreadItemDetails::ToolInvocation(inv) => {
458                // Buffer the invocation; we'll pair it with the ToolOutput
459                self.pending_tool_calls.push(PendingToolCall {
460                    call_id: item_id.to_string(),
461                    tool_call_id: inv.tool_call_id.clone(),
462                    tool_name: inv.tool_name.clone(),
463                    arguments: inv.arguments.clone(),
464                    timestamp: ts.to_string(),
465                });
466            }
467            ThreadItemDetails::ToolOutput(output) => {
468                // Find the matching pending invocation
469                let pending_idx = self.pending_tool_calls.iter().position(|p| p.call_id == output.call_id);
470
471                let (tool_name, arguments, tool_call_id, inv_ts) = if let Some(idx) = pending_idx {
472                    let p = self.pending_tool_calls.remove(idx);
473                    (p.tool_name, p.arguments, p.tool_call_id, p.timestamp)
474                } else {
475                    ("unknown".to_string(), None, output.tool_call_id.clone(), ts.to_string())
476                };
477
478                let call_id = tool_call_id.clone().unwrap_or_else(|| output.call_id.clone());
479
480                let mut step = Step::agent(self.next_step_id, "");
481                step.timestamp = Some(inv_ts);
482                step.message = None;
483                step.tool_calls = Some(vec![AtifToolCall {
484                    tool_call_id: call_id.clone(),
485                    function_name: tool_name,
486                    arguments,
487                }]);
488
489                let status_suffix = match output.status {
490                    ToolCallStatus::Failed => " [FAILED]",
491                    ToolCallStatus::InProgress => " [IN_PROGRESS]",
492                    ToolCallStatus::Completed => "",
493                };
494                let content = format!("{}{}", output.output, status_suffix);
495                step.observation = Some(Observation {
496                    results: vec![ObservationResult { source_call_id: call_id, content }],
497                });
498                self.push_step(step);
499            }
500            ThreadItemDetails::CommandExecution(cmd) => {
501                let call_id = item_id.to_string();
502                let mut step = Step::agent(self.next_step_id, "");
503                step.timestamp = Some(ts.to_string());
504                step.message = None;
505                step.tool_calls = Some(vec![AtifToolCall {
506                    tool_call_id: call_id.clone(),
507                    function_name: "command_execution".to_string(),
508                    arguments: Some(serde_json::json!({
509                        "command": cmd.command,
510                        "arguments": cmd.arguments,
511                    })),
512                }]);
513                step.observation = Some(Observation {
514                    results: vec![ObservationResult {
515                        source_call_id: call_id,
516                        content: cmd.aggregated_output.clone(),
517                    }],
518                });
519                if let Some(exit_code) = cmd.exit_code {
520                    step.extra = Some(serde_json::json!({ "exit_code": exit_code }));
521                }
522                self.push_step(step);
523            }
524            ThreadItemDetails::McpToolCall(mcp) => {
525                let call_id = item_id.to_string();
526                let mut step = Step::agent(self.next_step_id, "");
527                step.timestamp = Some(ts.to_string());
528                step.message = None;
529                step.tool_calls = Some(vec![AtifToolCall {
530                    tool_call_id: call_id.clone(),
531                    function_name: mcp.tool_name.clone(),
532                    arguments: mcp.arguments.clone(),
533                }]);
534                if let Some(result) = &mcp.result {
535                    step.observation = Some(Observation {
536                        results: vec![ObservationResult { source_call_id: call_id, content: result.clone() }],
537                    });
538                }
539                self.push_step(step);
540            }
541            ThreadItemDetails::FileChange(fc) => {
542                let changes: Vec<String> = fc.changes.iter().map(|c| format!("{}: {:?}", c.path, c.kind)).collect();
543                let msg = format!("file_changes: {}", changes.join(", "));
544                let mut step = Step::system(self.next_step_id, msg);
545                step.timestamp = Some(ts.to_string());
546                self.push_step(step);
547            }
548            ThreadItemDetails::WebSearch(ws) => {
549                let mut step = Step::system(self.next_step_id, format!("web_search: {}", ws.query));
550                step.timestamp = Some(ts.to_string());
551                if let Some(results) = &ws.results {
552                    step.observation = Some(Observation {
553                        results: results
554                            .iter()
555                            .enumerate()
556                            .map(|(i, r)| ObservationResult {
557                                source_call_id: format!("search_{i}"),
558                                content: r.clone(),
559                            })
560                            .collect(),
561                    });
562                }
563                self.push_step(step);
564            }
565            ThreadItemDetails::Harness(h) => {
566                let msg = format!("harness: {:?}", h.event);
567                let mut step = Step::system(self.next_step_id, msg);
568                step.timestamp = Some(ts.to_string());
569                if let Some(m) = &h.message {
570                    step.extra = Some(serde_json::json!({ "harness_message": m }));
571                }
572                self.push_step(step);
573            }
574            ThreadItemDetails::Error(e) => {
575                let mut step = Step::system(self.next_step_id, &e.message);
576                step.timestamp = Some(ts.to_string());
577                self.push_step(step);
578            }
579        }
580    }
581
582    fn push_step(&mut self, step: Step) {
583        self.next_step_id = step.step_id + 1;
584        self.steps.push(step);
585    }
586
587    /// Consume the builder and produce the final ATIF trajectory.
588    ///
589    /// Pass optional `FinalMetrics` to override the accumulated values.
590    /// If `None`, final metrics are derived from observed events.
591    pub fn finish(self, override_metrics: Option<FinalMetrics>) -> Trajectory {
592        let final_metrics = override_metrics.unwrap_or_else(|| FinalMetrics {
593            total_prompt_tokens: Some(self.total_input_tokens),
594            total_completion_tokens: Some(self.total_output_tokens),
595            total_cached_tokens: if self.total_cached_tokens > 0 {
596                Some(self.total_cached_tokens)
597            } else {
598                None
599            },
600            total_cost_usd: None,
601            total_steps: Some(self.steps.len() as u64),
602            extra: Some(serde_json::json!({ "num_turns": self.num_turns })),
603        });
604
605        Trajectory {
606            schema_version: ATIF_SCHEMA_VERSION.to_string(),
607            session_id: self.session_id.unwrap_or_else(|| uuid::Uuid::new_v4().to_string()),
608            agent: self.agent,
609            steps: self.steps,
610            notes: None,
611            final_metrics: Some(final_metrics),
612            extra: None,
613        }
614    }
615
616    /// Returns the number of steps collected so far.
617    fn step_count(&self) -> usize {
618        self.steps.len()
619    }
620}
621
622impl crate::EventEmitter for AtifTrajectoryBuilder {
623    fn emit(&mut self, event: &ThreadEvent) {
624        self.process_event(event);
625    }
626}
627
628#[cfg(test)]
629mod tests {
630    use super::*;
631    use crate::{
632        AgentMessageItem, CompactionMode, CompactionTrigger, ItemCompletedEvent, ThreadCompactBoundaryEvent,
633        ThreadItem, ThreadStartedEvent, ToolInvocationItem, ToolOutputItem, TurnCompletedEvent, TurnStartedEvent,
634        Usage,
635    };
636
637    fn fixed_ts() -> DateTime<Utc> {
638        "2025-01-15T10:30:00Z".parse().unwrap()
639    }
640
641    #[test]
642    fn trajectory_round_trip() {
643        let trajectory = Trajectory {
644            schema_version: ATIF_SCHEMA_VERSION.to_string(),
645            session_id: "test-session".to_string(),
646            agent: AtifAgent::vtcode(),
647            steps: vec![Step::user(1, "hello")],
648            notes: None,
649            final_metrics: None,
650            extra: None,
651        };
652
653        let json = serde_json::to_string_pretty(&trajectory).unwrap();
654        let restored: Trajectory = serde_json::from_str(&json).unwrap();
655        assert_eq!(restored.schema_version, ATIF_SCHEMA_VERSION);
656        assert_eq!(restored.session_id, "test-session");
657        assert_eq!(restored.steps.len(), 1);
658    }
659
660    #[test]
661    fn builder_thread_started_sets_session_id() {
662        let mut builder = AtifTrajectoryBuilder::new(AtifAgent::vtcode());
663        let event = ThreadEvent::ThreadStarted(ThreadStartedEvent { thread_id: "thread-abc".to_string() });
664        builder.process_event_at(&event, fixed_ts());
665        let trajectory = builder.finish(None);
666        assert_eq!(trajectory.session_id, "thread-abc");
667    }
668
669    #[test]
670    fn builder_agent_message_step() {
671        let mut builder = AtifTrajectoryBuilder::new(AtifAgent::vtcode());
672        let event = ThreadEvent::ItemCompleted(ItemCompletedEvent {
673            item: ThreadItem {
674                id: "msg-1".to_string(),
675                details: ThreadItemDetails::AgentMessage(AgentMessageItem { text: "Hello, world!".to_string() }),
676            },
677        });
678        builder.process_event_at(&event, fixed_ts());
679        let trajectory = builder.finish(None);
680
681        assert_eq!(trajectory.steps.len(), 1);
682        let step = &trajectory.steps[0];
683        assert_eq!(step.step_id, 1);
684        assert_eq!(step.source, StepSource::Agent);
685        assert_eq!(step.message.as_deref(), Some("Hello, world!"));
686    }
687
688    #[test]
689    fn builder_tool_invocation_with_output() {
690        let mut builder = AtifTrajectoryBuilder::new(AtifAgent::vtcode());
691        let ts = fixed_ts();
692
693        // Tool invocation
694        let inv_event = ThreadEvent::ItemCompleted(ItemCompletedEvent {
695            item: ThreadItem {
696                id: "tool_1".to_string(),
697                details: ThreadItemDetails::ToolInvocation(ToolInvocationItem {
698                    tool_name: "read_file".to_string(),
699                    arguments: Some(serde_json::json!({"path": "README.md"})),
700                    tool_call_id: Some("tc_0".to_string()),
701                    status: ToolCallStatus::Completed,
702                    outcome: None,
703                }),
704            },
705        });
706        builder.process_event_at(&inv_event, ts);
707
708        // Tool output
709        let out_event = ThreadEvent::ItemCompleted(ItemCompletedEvent {
710            item: ThreadItem {
711                id: "tool_1:output".to_string(),
712                details: ThreadItemDetails::ToolOutput(ToolOutputItem {
713                    call_id: "tool_1".to_string(),
714                    tool_call_id: Some("tc_0".to_string()),
715                    spool_path: None,
716                    output: "file contents here".to_string(),
717                    exit_code: Some(0),
718                    status: ToolCallStatus::Completed,
719                }),
720            },
721        });
722        builder.process_event_at(&out_event, ts);
723
724        let trajectory = builder.finish(None);
725        // Only one step: the invocation is buffered until output arrives
726        assert_eq!(trajectory.steps.len(), 1);
727        let step = &trajectory.steps[0];
728        assert_eq!(step.source, StepSource::Agent);
729
730        let calls = step.tool_calls.as_ref().unwrap();
731        assert_eq!(calls.len(), 1);
732        assert_eq!(calls[0].function_name, "read_file");
733        assert_eq!(calls[0].tool_call_id, "tc_0");
734
735        let obs = step.observation.as_ref().unwrap();
736        assert_eq!(obs.results.len(), 1);
737        assert_eq!(obs.results[0].content, "file contents here");
738    }
739
740    #[test]
741    fn builder_turn_completed_accumulates_metrics() {
742        let mut builder = AtifTrajectoryBuilder::new(AtifAgent::vtcode());
743        let event = ThreadEvent::TurnCompleted(TurnCompletedEvent {
744            usage: Usage {
745                input_tokens: 500,
746                cached_input_tokens: 100,
747                cache_creation_tokens: 0,
748                output_tokens: 200,
749            },
750        });
751        builder.process_event_at(&event, fixed_ts());
752
753        let trajectory = builder.finish(None);
754        let fm = trajectory.final_metrics.as_ref().unwrap();
755        assert_eq!(fm.total_prompt_tokens, Some(500));
756        assert_eq!(fm.total_completion_tokens, Some(200));
757        assert_eq!(fm.total_cached_tokens, Some(100));
758    }
759
760    #[test]
761    fn step_metrics_from_usage() {
762        let usage = Usage {
763            input_tokens: 1000,
764            cached_input_tokens: 200,
765            cache_creation_tokens: 50,
766            output_tokens: 300,
767        };
768        let metrics = StepMetrics::from_usage(&usage);
769        assert_eq!(metrics.prompt_tokens, Some(1000));
770        assert_eq!(metrics.completion_tokens, Some(300));
771        assert_eq!(metrics.cached_tokens, Some(200));
772        assert!(metrics.extra.is_some());
773    }
774
775    #[test]
776    fn builder_implements_event_emitter() {
777        let mut builder = AtifTrajectoryBuilder::new(AtifAgent::vtcode());
778        let event = ThreadEvent::ThreadStarted(ThreadStartedEvent { thread_id: "t-1".to_string() });
779        // Use EventEmitter trait
780        crate::EventEmitter::emit(&mut builder, &event);
781        assert_eq!(builder.step_count(), 0); // ThreadStarted doesn't create a step
782    }
783
784    #[test]
785    fn skips_lifecycle_events() {
786        let mut builder = AtifTrajectoryBuilder::new(AtifAgent::vtcode());
787        builder.process_event(&ThreadEvent::TurnStarted(TurnStartedEvent::default()));
788        assert_eq!(builder.step_count(), 0);
789    }
790
791    #[test]
792    fn compact_boundary_with_segment_metadata_preserves_atif_export() {
793        let mut builder = AtifTrajectoryBuilder::new(AtifAgent::vtcode());
794        let event = ThreadEvent::ThreadCompactBoundary(ThreadCompactBoundaryEvent {
795            thread_id: "thread-1".to_string(),
796            trigger: CompactionTrigger::Auto,
797            mode: CompactionMode::Local,
798            original_message_count: 12,
799            compacted_message_count: 5,
800            history_artifact_path: None,
801            previous_segment_id: Some("segment-0001".to_string()),
802            new_segment_id: Some("segment-0002".to_string()),
803            previous_prefix_hash: Some("prefix-before".to_string()),
804            new_prefix_hash: Some("prefix-after".to_string()),
805            previous_catalog_hash: Some("catalog-before".to_string()),
806            new_catalog_hash: Some("catalog-after".to_string()),
807        });
808
809        builder.process_event_at(&event, fixed_ts());
810        let trajectory = builder.finish(None);
811
812        assert_eq!(trajectory.steps.len(), 1);
813        assert_eq!(trajectory.steps[0].source, StepSource::System);
814        assert_eq!(
815            trajectory.steps[0].message.as_deref(),
816            Some("context_compaction: 12 messages -> 5 messages (auto)")
817        );
818    }
819}
vtcode_exec_events/atif.rs

vtcode_exec_events/
atif.rs
