Skip to main content

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    /// Whether any per-turn `TurnCompleted`/`TurnFailed` usage was observed.
326    /// Guards the `ThreadCompleted` aggregate from double-counting usage.
327    saw_per_turn_usage: bool,
328    /// Pending tool invocations awaiting matching ToolOutput.
329    pending_tool_calls: Vec<PendingToolCall>,
330}
331
332struct PendingToolCall {
333    call_id: String,
334    tool_call_id: Option<String>,
335    tool_name: String,
336    arguments: Option<Value>,
337    timestamp: String,
338}
339
340impl AtifTrajectoryBuilder {
341    /// Create a new builder for the given agent.
342    pub fn new(agent: AtifAgent) -> Self {
343        Self {
344            agent,
345            session_id: None,
346            steps: Vec::new(),
347            next_step_id: 1,
348            total_input_tokens: 0,
349            total_output_tokens: 0,
350            total_cached_tokens: 0,
351            num_turns: 0,
352            saw_per_turn_usage: false,
353            pending_tool_calls: Vec::new(),
354        }
355    }
356
357    /// Set the session ID explicitly. If not set, it will be derived from
358    /// `ThreadStarted` or `ThreadCompleted` events.
359    pub fn set_session_id(&mut self, id: impl Into<String>) {
360        self.session_id = Some(id.into());
361    }
362
363    /// Process a thread event using the current wall-clock time.
364    pub fn process_event(&mut self, event: &ThreadEvent) {
365        self.process_event_at(event, Utc::now());
366    }
367
368    /// Process a thread event with an explicit timestamp (for deterministic tests).
369    pub fn process_event_at(&mut self, event: &ThreadEvent, ts: DateTime<Utc>) {
370        let ts_str = ts.to_rfc3339();
371        match event {
372            ThreadEvent::ThreadStarted(e) => {
373                if self.session_id.is_none() {
374                    self.session_id = Some(e.thread_id.clone());
375                }
376            }
377            ThreadEvent::ThreadCompleted(e) => {
378                if self.session_id.is_none() {
379                    self.session_id = Some(e.session_id.clone());
380                }
381                self.num_turns = e.num_turns;
382                // `ThreadCompleted` carries the *aggregate* usage, which equals
383                // the sum of per-turn usage already accumulated by the
384                // `TurnCompleted`/`TurnFailed` arms. Only fall back to it when
385                // no per-turn usage was observed (e.g. a harness that emits
386                // only the thread aggregate), otherwise totals double-count.
387                if !self.saw_per_turn_usage {
388                    self.total_input_tokens = self.total_input_tokens.saturating_add(e.usage.input_tokens);
389                    self.total_output_tokens = self.total_output_tokens.saturating_add(e.usage.output_tokens);
390                    self.total_cached_tokens = self.total_cached_tokens.saturating_add(e.usage.cached_input_tokens);
391                }
392            }
393            ThreadEvent::TurnCompleted(e) => {
394                self.saw_per_turn_usage = true;
395                self.total_input_tokens = self.total_input_tokens.saturating_add(e.usage.input_tokens);
396                self.total_output_tokens = self.total_output_tokens.saturating_add(e.usage.output_tokens);
397                self.total_cached_tokens = self.total_cached_tokens.saturating_add(e.usage.cached_input_tokens);
398                self.num_turns += 1;
399
400                let mut step = Step::system(self.next_step_id, "turn_completed");
401                step.timestamp = Some(ts_str);
402                step.metrics = Some(StepMetrics::from_usage(&e.usage));
403                self.push_step(step);
404            }
405            ThreadEvent::TurnFailed(e) => {
406                if let Some(usage) = &e.usage {
407                    self.saw_per_turn_usage = true;
408                    self.total_input_tokens = self.total_input_tokens.saturating_add(usage.input_tokens);
409                    self.total_output_tokens = self.total_output_tokens.saturating_add(usage.output_tokens);
410                    self.total_cached_tokens = self.total_cached_tokens.saturating_add(usage.cached_input_tokens);
411                }
412                let mut step = Step::system(self.next_step_id, &e.message);
413                step.timestamp = Some(ts_str);
414                step.metrics = e.usage.as_ref().map(StepMetrics::from_usage);
415                self.push_step(step);
416            }
417            ThreadEvent::ItemCompleted(e) => {
418                self.process_item_completed(&e.item.id, &e.item.details, &ts_str);
419            }
420            ThreadEvent::ThreadCompactBoundary(e) => {
421                let msg = format!(
422                    "context_compaction: {} messages -> {} messages ({})",
423                    e.original_message_count,
424                    e.compacted_message_count,
425                    e.trigger.as_str()
426                );
427                let mut step = Step::system(self.next_step_id, msg);
428                step.timestamp = Some(ts_str);
429                self.push_step(step);
430            }
431            ThreadEvent::ContextReset(e) => {
432                let msg = format!(
433                    "context_reset: {}% context used; plan preserved: {}; tool budget reset: {}",
434                    e.previous_context_usage_percent, e.plan_preserved, e.tool_budget_reset
435                );
436                let mut step = Step::system(self.next_step_id, msg);
437                step.timestamp = Some(ts_str);
438                step.extra = Some(serde_json::json!({
439                    "thread_id": e.thread_id,
440                    "turn_id": e.turn_id,
441                    "trigger": e.trigger,
442                    "plan_preserved": e.plan_preserved,
443                    "previous_context_usage_percent": e.previous_context_usage_percent,
444                    "tool_budget_reset": e.tool_budget_reset,
445                }));
446                self.push_step(step);
447            }
448            ThreadEvent::Error(e) => {
449                let mut step = Step::system(self.next_step_id, &e.message);
450                step.timestamp = Some(ts_str);
451                self.push_step(step);
452            }
453            // Skip streaming/lifecycle events that don't map to ATIF steps
454            ThreadEvent::TurnStarted(_)
455            | ThreadEvent::ItemStarted(_)
456            | ThreadEvent::ItemUpdated(_)
457            | ThreadEvent::PlanDelta(_)
458            | ThreadEvent::PlanApprovalRequested(_)
459            | ThreadEvent::PlanApprovalResolved(_)
460            | ThreadEvent::PermissionRequested(_)
461            | ThreadEvent::PermissionResolved(_)
462            | ThreadEvent::Interjected(_)
463            | ThreadEvent::Unknown => {}
464        }
465    }
466
467    fn process_item_completed(&mut self, item_id: &str, details: &ThreadItemDetails, ts: &str) {
468        match details {
469            ThreadItemDetails::AgentMessage(msg) => {
470                let mut step = Step::agent(self.next_step_id, &msg.text);
471                step.timestamp = Some(ts.to_string());
472                self.push_step(step);
473            }
474            ThreadItemDetails::Plan(plan) => {
475                let mut step = Step::agent(self.next_step_id, &plan.text);
476                step.timestamp = Some(ts.to_string());
477                step.extra = Some(serde_json::json!({ "vtcode_item_type": "plan" }));
478                self.push_step(step);
479            }
480            ThreadItemDetails::Reasoning(r) => {
481                let mut step = Step::agent(self.next_step_id, "");
482                step.timestamp = Some(ts.to_string());
483                step.reasoning_content = Some(r.text.clone());
484                step.message = None;
485                self.push_step(step);
486            }
487            ThreadItemDetails::ToolInvocation(inv) => {
488                // Buffer the invocation; we'll pair it with the ToolOutput
489                self.pending_tool_calls.push(PendingToolCall {
490                    call_id: item_id.to_string(),
491                    tool_call_id: inv.tool_call_id.clone(),
492                    tool_name: inv.tool_name.clone(),
493                    arguments: inv.arguments.clone(),
494                    timestamp: ts.to_string(),
495                });
496            }
497            ThreadItemDetails::ToolOutput(output) => {
498                // Find the matching pending invocation
499                let pending_idx = self.pending_tool_calls.iter().position(|p| p.call_id == output.call_id);
500
501                let (tool_name, arguments, tool_call_id, inv_ts) = if let Some(idx) = pending_idx {
502                    let p = self.pending_tool_calls.remove(idx);
503                    (p.tool_name, p.arguments, p.tool_call_id, p.timestamp)
504                } else {
505                    ("unknown".to_string(), None, output.tool_call_id.clone(), ts.to_string())
506                };
507
508                let call_id = tool_call_id.clone().unwrap_or_else(|| output.call_id.clone());
509
510                let mut step = Step::agent(self.next_step_id, "");
511                step.timestamp = Some(inv_ts);
512                step.message = None;
513                step.tool_calls = Some(vec![AtifToolCall {
514                    tool_call_id: call_id.clone(),
515                    function_name: tool_name,
516                    arguments,
517                }]);
518
519                let status_suffix = match output.status {
520                    ToolCallStatus::Failed => " [FAILED]",
521                    ToolCallStatus::InProgress => " [IN_PROGRESS]",
522                    ToolCallStatus::Completed => "",
523                };
524                let content = format!("{}{}", output.output, status_suffix);
525                step.observation = Some(Observation {
526                    results: vec![ObservationResult { source_call_id: call_id, content }],
527                });
528                self.push_step(step);
529            }
530            ThreadItemDetails::CommandExecution(cmd) => {
531                let call_id = item_id.to_string();
532                let mut step = Step::agent(self.next_step_id, "");
533                step.timestamp = Some(ts.to_string());
534                step.message = None;
535                step.tool_calls = Some(vec![AtifToolCall {
536                    tool_call_id: call_id.clone(),
537                    function_name: "command_execution".to_string(),
538                    arguments: Some(serde_json::json!({
539                        "command": cmd.command,
540                        "arguments": cmd.arguments,
541                    })),
542                }]);
543                step.observation = Some(Observation {
544                    results: vec![ObservationResult {
545                        source_call_id: call_id,
546                        content: cmd.aggregated_output.clone(),
547                    }],
548                });
549                if let Some(exit_code) = cmd.exit_code {
550                    step.extra = Some(serde_json::json!({ "exit_code": exit_code }));
551                }
552                self.push_step(step);
553            }
554            ThreadItemDetails::McpToolCall(mcp) => {
555                let call_id = item_id.to_string();
556                let mut step = Step::agent(self.next_step_id, "");
557                step.timestamp = Some(ts.to_string());
558                step.message = None;
559                step.tool_calls = Some(vec![AtifToolCall {
560                    tool_call_id: call_id.clone(),
561                    function_name: mcp.tool_name.clone(),
562                    arguments: mcp.arguments.clone(),
563                }]);
564                if let Some(result) = &mcp.result {
565                    step.observation = Some(Observation {
566                        results: vec![ObservationResult { source_call_id: call_id, content: result.clone() }],
567                    });
568                }
569                self.push_step(step);
570            }
571            ThreadItemDetails::FileChange(fc) => {
572                let changes: Vec<String> = fc.changes.iter().map(|c| format!("{}: {:?}", c.path, c.kind)).collect();
573                let msg = format!("file_changes: {}", changes.join(", "));
574                let mut step = Step::system(self.next_step_id, msg);
575                step.timestamp = Some(ts.to_string());
576                self.push_step(step);
577            }
578            ThreadItemDetails::WebSearch(ws) => {
579                let mut step = Step::system(self.next_step_id, format!("web_search: {}", ws.query));
580                step.timestamp = Some(ts.to_string());
581                if let Some(results) = &ws.results {
582                    step.observation = Some(Observation {
583                        results: results
584                            .iter()
585                            .enumerate()
586                            .map(|(i, r)| ObservationResult {
587                                source_call_id: format!("search_{i}"),
588                                content: r.clone(),
589                            })
590                            .collect(),
591                    });
592                }
593                self.push_step(step);
594            }
595            ThreadItemDetails::Harness(h) => {
596                let msg = format!("harness: {:?}", h.event);
597                let mut step = Step::system(self.next_step_id, msg);
598                step.timestamp = Some(ts.to_string());
599                if let Some(m) = &h.message {
600                    step.extra = Some(serde_json::json!({ "harness_message": m }));
601                }
602                self.push_step(step);
603            }
604            ThreadItemDetails::Error(e) => {
605                let mut step = Step::system(self.next_step_id, &e.message);
606                step.timestamp = Some(ts.to_string());
607                self.push_step(step);
608            }
609        }
610    }
611
612    fn push_step(&mut self, step: Step) {
613        self.next_step_id = step.step_id + 1;
614        self.steps.push(step);
615    }
616
617    /// Consume the builder and produce the final ATIF trajectory.
618    ///
619    /// Pass optional `FinalMetrics` to override the accumulated values.
620    /// If `None`, final metrics are derived from observed events.
621    pub fn finish(self, override_metrics: Option<FinalMetrics>) -> Trajectory {
622        let final_metrics = override_metrics.unwrap_or_else(|| FinalMetrics {
623            total_prompt_tokens: Some(self.total_input_tokens),
624            total_completion_tokens: Some(self.total_output_tokens),
625            total_cached_tokens: if self.total_cached_tokens > 0 {
626                Some(self.total_cached_tokens)
627            } else {
628                None
629            },
630            total_cost_usd: None,
631            total_steps: Some(self.steps.len() as u64),
632            extra: Some(serde_json::json!({ "num_turns": self.num_turns })),
633        });
634
635        Trajectory {
636            schema_version: ATIF_SCHEMA_VERSION.to_string(),
637            session_id: self.session_id.unwrap_or_else(|| uuid::Uuid::new_v4().to_string()),
638            agent: self.agent,
639            steps: self.steps,
640            notes: None,
641            final_metrics: Some(final_metrics),
642            extra: None,
643        }
644    }
645
646    /// Returns the number of steps collected so far.
647    fn step_count(&self) -> usize {
648        self.steps.len()
649    }
650}
651
652impl crate::EventEmitter for AtifTrajectoryBuilder {
653    fn emit(&mut self, event: &ThreadEvent) {
654        self.process_event(event);
655    }
656}
657
658#[cfg(test)]
659mod tests {
660    use super::*;
661    use crate::{
662        AgentMessageItem, CompactionMode, CompactionTrigger, ItemCompletedEvent, ThreadCompactBoundaryEvent,
663        ThreadItem, ThreadStartedEvent, ToolInvocationItem, ToolOutputItem, TurnCompletedEvent, TurnStartedEvent,
664        Usage,
665    };
666
667    fn fixed_ts() -> DateTime<Utc> {
668        "2025-01-15T10:30:00Z".parse().unwrap()
669    }
670
671    #[test]
672    fn trajectory_round_trip() {
673        let trajectory = Trajectory {
674            schema_version: ATIF_SCHEMA_VERSION.to_string(),
675            session_id: "test-session".to_string(),
676            agent: AtifAgent::vtcode(),
677            steps: vec![Step::user(1, "hello")],
678            notes: None,
679            final_metrics: None,
680            extra: None,
681        };
682
683        let json = serde_json::to_string_pretty(&trajectory).unwrap();
684        let restored: Trajectory = serde_json::from_str(&json).unwrap();
685        assert_eq!(restored.schema_version, ATIF_SCHEMA_VERSION);
686        assert_eq!(restored.session_id, "test-session");
687        assert_eq!(restored.steps.len(), 1);
688    }
689
690    #[test]
691    fn builder_thread_started_sets_session_id() {
692        let mut builder = AtifTrajectoryBuilder::new(AtifAgent::vtcode());
693        let event = ThreadEvent::ThreadStarted(ThreadStartedEvent { thread_id: "thread-abc".to_string() });
694        builder.process_event_at(&event, fixed_ts());
695        let trajectory = builder.finish(None);
696        assert_eq!(trajectory.session_id, "thread-abc");
697    }
698
699    #[test]
700    fn builder_agent_message_step() {
701        let mut builder = AtifTrajectoryBuilder::new(AtifAgent::vtcode());
702        let event = ThreadEvent::ItemCompleted(ItemCompletedEvent {
703            item: ThreadItem {
704                id: "msg-1".to_string(),
705                details: ThreadItemDetails::AgentMessage(AgentMessageItem { text: "Hello, world!".to_string() }),
706            },
707        });
708        builder.process_event_at(&event, fixed_ts());
709        let trajectory = builder.finish(None);
710
711        assert_eq!(trajectory.steps.len(), 1);
712        let step = &trajectory.steps[0];
713        assert_eq!(step.step_id, 1);
714        assert_eq!(step.source, StepSource::Agent);
715        assert_eq!(step.message.as_deref(), Some("Hello, world!"));
716    }
717
718    #[test]
719    fn builder_tool_invocation_with_output() {
720        let mut builder = AtifTrajectoryBuilder::new(AtifAgent::vtcode());
721        let ts = fixed_ts();
722
723        // Tool invocation
724        let inv_event = ThreadEvent::ItemCompleted(ItemCompletedEvent {
725            item: ThreadItem {
726                id: "tool_1".to_string(),
727                details: ThreadItemDetails::ToolInvocation(ToolInvocationItem {
728                    tool_name: "read_file".to_string(),
729                    arguments: Some(serde_json::json!({"path": "README.md"})),
730                    tool_call_id: Some("tc_0".to_string()),
731                    status: ToolCallStatus::Completed,
732                    outcome: None,
733                }),
734            },
735        });
736        builder.process_event_at(&inv_event, ts);
737
738        // Tool output
739        let out_event = ThreadEvent::ItemCompleted(ItemCompletedEvent {
740            item: ThreadItem {
741                id: "tool_1:output".to_string(),
742                details: ThreadItemDetails::ToolOutput(ToolOutputItem {
743                    call_id: "tool_1".to_string(),
744                    tool_call_id: Some("tc_0".to_string()),
745                    spool_path: None,
746                    output: "file contents here".to_string(),
747                    exit_code: Some(0),
748                    status: ToolCallStatus::Completed,
749                }),
750            },
751        });
752        builder.process_event_at(&out_event, ts);
753
754        let trajectory = builder.finish(None);
755        // Only one step: the invocation is buffered until output arrives
756        assert_eq!(trajectory.steps.len(), 1);
757        let step = &trajectory.steps[0];
758        assert_eq!(step.source, StepSource::Agent);
759
760        let calls = step.tool_calls.as_ref().unwrap();
761        assert_eq!(calls.len(), 1);
762        assert_eq!(calls[0].function_name, "read_file");
763        assert_eq!(calls[0].tool_call_id, "tc_0");
764
765        let obs = step.observation.as_ref().unwrap();
766        assert_eq!(obs.results.len(), 1);
767        assert_eq!(obs.results[0].content, "file contents here");
768    }
769
770    #[test]
771    fn builder_turn_completed_accumulates_metrics() {
772        let mut builder = AtifTrajectoryBuilder::new(AtifAgent::vtcode());
773        let event = ThreadEvent::TurnCompleted(TurnCompletedEvent {
774            usage: Usage {
775                input_tokens: 500,
776                cached_input_tokens: 100,
777                cache_creation_tokens: 0,
778                output_tokens: 200,
779            },
780        });
781        builder.process_event_at(&event, fixed_ts());
782
783        let trajectory = builder.finish(None);
784        let fm = trajectory.final_metrics.as_ref().unwrap();
785        assert_eq!(fm.total_prompt_tokens, Some(500));
786        assert_eq!(fm.total_completion_tokens, Some(200));
787        assert_eq!(fm.total_cached_tokens, Some(100));
788    }
789
790    #[test]
791    fn step_metrics_from_usage() {
792        let usage = Usage {
793            input_tokens: 1000,
794            cached_input_tokens: 200,
795            cache_creation_tokens: 50,
796            output_tokens: 300,
797        };
798        let metrics = StepMetrics::from_usage(&usage);
799        assert_eq!(metrics.prompt_tokens, Some(1000));
800        assert_eq!(metrics.completion_tokens, Some(300));
801        assert_eq!(metrics.cached_tokens, Some(200));
802        assert!(metrics.extra.is_some());
803    }
804
805    #[test]
806    fn builder_implements_event_emitter() {
807        let mut builder = AtifTrajectoryBuilder::new(AtifAgent::vtcode());
808        let event = ThreadEvent::ThreadStarted(ThreadStartedEvent { thread_id: "t-1".to_string() });
809        // Use EventEmitter trait
810        crate::EventEmitter::emit(&mut builder, &event);
811        assert_eq!(builder.step_count(), 0); // ThreadStarted doesn't create a step
812    }
813
814    #[test]
815    fn skips_lifecycle_events() {
816        let mut builder = AtifTrajectoryBuilder::new(AtifAgent::vtcode());
817        builder.process_event(&ThreadEvent::TurnStarted(TurnStartedEvent::default()));
818        assert_eq!(builder.step_count(), 0);
819    }
820
821    #[test]
822    fn compact_boundary_with_segment_metadata_preserves_atif_export() {
823        let mut builder = AtifTrajectoryBuilder::new(AtifAgent::vtcode());
824        let event = ThreadEvent::ThreadCompactBoundary(ThreadCompactBoundaryEvent {
825            thread_id: "thread-1".to_string(),
826            trigger: CompactionTrigger::Auto,
827            mode: CompactionMode::Local,
828            original_message_count: 12,
829            compacted_message_count: 5,
830            history_artifact_path: None,
831            previous_segment_id: Some("segment-0001".to_string()),
832            new_segment_id: Some("segment-0002".to_string()),
833            previous_prefix_hash: Some("prefix-before".to_string()),
834            new_prefix_hash: Some("prefix-after".to_string()),
835            previous_catalog_hash: Some("catalog-before".to_string()),
836            new_catalog_hash: Some("catalog-after".to_string()),
837        });
838
839        builder.process_event_at(&event, fixed_ts());
840        let trajectory = builder.finish(None);
841
842        assert_eq!(trajectory.steps.len(), 1);
843        assert_eq!(trajectory.steps[0].source, StepSource::System);
844        assert_eq!(
845            trajectory.steps[0].message.as_deref(),
846            Some("context_compaction: 12 messages -> 5 messages (auto)")
847        );
848    }
849}