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::TurnBlocked(e) => {
418                if let Some(usage) = &e.usage {
419                    self.saw_per_turn_usage = true;
420                    self.total_input_tokens = self.total_input_tokens.saturating_add(usage.input_tokens);
421                    self.total_output_tokens = self.total_output_tokens.saturating_add(usage.output_tokens);
422                    self.total_cached_tokens = self.total_cached_tokens.saturating_add(usage.cached_input_tokens);
423                }
424                let mut step = Step::system(self.next_step_id, &e.message);
425                step.timestamp = Some(ts_str);
426                step.metrics = e.usage.as_ref().map(StepMetrics::from_usage);
427                step.extra = Some(serde_json::json!({
428                    "last_tool": e.last_tool,
429                    "blocked_streak": e.blocked_streak,
430                    "blocked_total": e.blocked_total,
431                    "consecutive_cap": e.consecutive_cap,
432                    "total_cap": e.total_cap,
433                    "recovery_active": e.recovery_active,
434                }));
435                self.push_step(step);
436            }
437            ThreadEvent::ItemCompleted(e) => {
438                self.process_item_completed(&e.item.id, &e.item.details, &ts_str);
439            }
440            ThreadEvent::ThreadCompactBoundary(e) => {
441                let msg = format!(
442                    "context_compaction: {} messages -> {} messages ({})",
443                    e.original_message_count,
444                    e.compacted_message_count,
445                    e.trigger.as_str()
446                );
447                let mut step = Step::system(self.next_step_id, msg);
448                step.timestamp = Some(ts_str);
449                self.push_step(step);
450            }
451            ThreadEvent::ContextReset(e) => {
452                let msg = format!(
453                    "context_reset: {}% context used; plan preserved: {}; tool budget reset: {}",
454                    e.previous_context_usage_percent, e.plan_preserved, e.tool_budget_reset
455                );
456                let mut step = Step::system(self.next_step_id, msg);
457                step.timestamp = Some(ts_str);
458                step.extra = Some(serde_json::json!({
459                    "thread_id": e.thread_id,
460                    "turn_id": e.turn_id,
461                    "trigger": e.trigger,
462                    "plan_preserved": e.plan_preserved,
463                    "previous_context_usage_percent": e.previous_context_usage_percent,
464                    "tool_budget_reset": e.tool_budget_reset,
465                }));
466                self.push_step(step);
467            }
468            ThreadEvent::Error(e) => {
469                let mut step = Step::system(self.next_step_id, &e.message);
470                step.timestamp = Some(ts_str);
471                self.push_step(step);
472            }
473            // Skip streaming/lifecycle events that don't map to ATIF steps
474            ThreadEvent::TurnStarted(_)
475            | ThreadEvent::ItemStarted(_)
476            | ThreadEvent::ItemUpdated(_)
477            | ThreadEvent::PlanDelta(_)
478            | ThreadEvent::PlanApprovalRequested(_)
479            | ThreadEvent::PlanApprovalResolved(_)
480            | ThreadEvent::PermissionRequested(_)
481            | ThreadEvent::PermissionResolved(_)
482            | ThreadEvent::Interjected(_)
483            | ThreadEvent::Unknown => {}
484        }
485    }
486
487    fn process_item_completed(&mut self, item_id: &str, details: &ThreadItemDetails, ts: &str) {
488        match details {
489            ThreadItemDetails::AgentMessage(msg) => {
490                let mut step = Step::agent(self.next_step_id, &msg.text);
491                step.timestamp = Some(ts.to_string());
492                self.push_step(step);
493            }
494            ThreadItemDetails::Plan(plan) => {
495                let mut step = Step::agent(self.next_step_id, &plan.text);
496                step.timestamp = Some(ts.to_string());
497                step.extra = Some(serde_json::json!({ "vtcode_item_type": "plan" }));
498                self.push_step(step);
499            }
500            ThreadItemDetails::Reasoning(r) => {
501                let mut step = Step::agent(self.next_step_id, "");
502                step.timestamp = Some(ts.to_string());
503                step.reasoning_content = Some(r.text.clone());
504                step.message = None;
505                self.push_step(step);
506            }
507            ThreadItemDetails::ToolInvocation(inv) => {
508                // Buffer the invocation; we'll pair it with the ToolOutput
509                self.pending_tool_calls.push(PendingToolCall {
510                    call_id: item_id.to_string(),
511                    tool_call_id: inv.tool_call_id.clone(),
512                    tool_name: inv.tool_name.clone(),
513                    arguments: inv.arguments.clone(),
514                    timestamp: ts.to_string(),
515                });
516            }
517            ThreadItemDetails::ToolOutput(output) => {
518                // Find the matching pending invocation
519                let pending_idx = self.pending_tool_calls.iter().position(|p| p.call_id == output.call_id);
520
521                let (tool_name, arguments, tool_call_id, inv_ts) = if let Some(idx) = pending_idx {
522                    let p = self.pending_tool_calls.remove(idx);
523                    (p.tool_name, p.arguments, p.tool_call_id, p.timestamp)
524                } else {
525                    ("unknown".to_string(), None, output.tool_call_id.clone(), ts.to_string())
526                };
527
528                let call_id = tool_call_id.clone().unwrap_or_else(|| output.call_id.clone());
529
530                let mut step = Step::agent(self.next_step_id, "");
531                step.timestamp = Some(inv_ts);
532                step.message = None;
533                step.tool_calls = Some(vec![AtifToolCall {
534                    tool_call_id: call_id.clone(),
535                    function_name: tool_name,
536                    arguments,
537                }]);
538
539                let status_suffix = match output.status {
540                    ToolCallStatus::Failed => " [FAILED]",
541                    ToolCallStatus::InProgress => " [IN_PROGRESS]",
542                    ToolCallStatus::Completed => "",
543                };
544                let content = format!("{}{}", output.output, status_suffix);
545                step.observation = Some(Observation {
546                    results: vec![ObservationResult { source_call_id: call_id, content }],
547                });
548                self.push_step(step);
549            }
550            ThreadItemDetails::CommandExecution(cmd) => {
551                let call_id = item_id.to_string();
552                let mut step = Step::agent(self.next_step_id, "");
553                step.timestamp = Some(ts.to_string());
554                step.message = None;
555                step.tool_calls = Some(vec![AtifToolCall {
556                    tool_call_id: call_id.clone(),
557                    function_name: "command_execution".to_string(),
558                    arguments: Some(serde_json::json!({
559                        "command": cmd.command,
560                        "arguments": cmd.arguments,
561                    })),
562                }]);
563                step.observation = Some(Observation {
564                    results: vec![ObservationResult {
565                        source_call_id: call_id,
566                        content: cmd.aggregated_output.clone(),
567                    }],
568                });
569                if let Some(exit_code) = cmd.exit_code {
570                    step.extra = Some(serde_json::json!({ "exit_code": exit_code }));
571                }
572                self.push_step(step);
573            }
574            ThreadItemDetails::McpToolCall(mcp) => {
575                let call_id = item_id.to_string();
576                let mut step = Step::agent(self.next_step_id, "");
577                step.timestamp = Some(ts.to_string());
578                step.message = None;
579                step.tool_calls = Some(vec![AtifToolCall {
580                    tool_call_id: call_id.clone(),
581                    function_name: mcp.tool_name.clone(),
582                    arguments: mcp.arguments.clone(),
583                }]);
584                if let Some(result) = &mcp.result {
585                    step.observation = Some(Observation {
586                        results: vec![ObservationResult { source_call_id: call_id, content: result.clone() }],
587                    });
588                }
589                self.push_step(step);
590            }
591            ThreadItemDetails::FileChange(fc) => {
592                let changes: Vec<String> = fc.changes.iter().map(|c| format!("{}: {:?}", c.path, c.kind)).collect();
593                let msg = format!("file_changes: {}", changes.join(", "));
594                let mut step = Step::system(self.next_step_id, msg);
595                step.timestamp = Some(ts.to_string());
596                self.push_step(step);
597            }
598            ThreadItemDetails::WebSearch(ws) => {
599                let mut step = Step::system(self.next_step_id, format!("web_search: {}", ws.query));
600                step.timestamp = Some(ts.to_string());
601                if let Some(results) = &ws.results {
602                    step.observation = Some(Observation {
603                        results: results
604                            .iter()
605                            .enumerate()
606                            .map(|(i, r)| ObservationResult {
607                                source_call_id: format!("search_{i}"),
608                                content: r.clone(),
609                            })
610                            .collect(),
611                    });
612                }
613                self.push_step(step);
614            }
615            ThreadItemDetails::Harness(h) => {
616                let msg = format!("harness: {:?}", h.event);
617                let mut step = Step::system(self.next_step_id, msg);
618                step.timestamp = Some(ts.to_string());
619                if let Some(m) = &h.message {
620                    step.extra = Some(serde_json::json!({ "harness_message": m }));
621                }
622                self.push_step(step);
623            }
624            ThreadItemDetails::Error(e) => {
625                let mut step = Step::system(self.next_step_id, &e.message);
626                step.timestamp = Some(ts.to_string());
627                self.push_step(step);
628            }
629        }
630    }
631
632    fn push_step(&mut self, step: Step) {
633        self.next_step_id = step.step_id + 1;
634        self.steps.push(step);
635    }
636
637    /// Consume the builder and produce the final ATIF trajectory.
638    ///
639    /// Pass optional `FinalMetrics` to override the accumulated values.
640    /// If `None`, final metrics are derived from observed events.
641    pub fn finish(self, override_metrics: Option<FinalMetrics>) -> Trajectory {
642        let final_metrics = override_metrics.unwrap_or_else(|| FinalMetrics {
643            total_prompt_tokens: Some(self.total_input_tokens),
644            total_completion_tokens: Some(self.total_output_tokens),
645            total_cached_tokens: if self.total_cached_tokens > 0 {
646                Some(self.total_cached_tokens)
647            } else {
648                None
649            },
650            total_cost_usd: None,
651            total_steps: Some(self.steps.len() as u64),
652            extra: Some(serde_json::json!({ "num_turns": self.num_turns })),
653        });
654
655        Trajectory {
656            schema_version: ATIF_SCHEMA_VERSION.to_string(),
657            session_id: self.session_id.unwrap_or_else(|| uuid::Uuid::new_v4().to_string()),
658            agent: self.agent,
659            steps: self.steps,
660            notes: None,
661            final_metrics: Some(final_metrics),
662            extra: None,
663        }
664    }
665
666    /// Returns the number of steps collected so far.
667    fn step_count(&self) -> usize {
668        self.steps.len()
669    }
670}
671
672impl crate::EventEmitter for AtifTrajectoryBuilder {
673    fn emit(&mut self, event: &ThreadEvent) {
674        self.process_event(event);
675    }
676}
677
678#[cfg(test)]
679mod tests {
680    use super::*;
681    use crate::{
682        AgentMessageItem, CompactionMode, CompactionTrigger, ItemCompletedEvent, ThreadCompactBoundaryEvent,
683        ThreadItem, ThreadStartedEvent, ToolInvocationItem, ToolOutputItem, TurnCompletedEvent, TurnStartedEvent,
684        Usage,
685    };
686
687    fn fixed_ts() -> DateTime<Utc> {
688        "2025-01-15T10:30:00Z".parse().unwrap()
689    }
690
691    #[test]
692    fn trajectory_round_trip() {
693        let trajectory = Trajectory {
694            schema_version: ATIF_SCHEMA_VERSION.to_string(),
695            session_id: "test-session".to_string(),
696            agent: AtifAgent::vtcode(),
697            steps: vec![Step::user(1, "hello")],
698            notes: None,
699            final_metrics: None,
700            extra: None,
701        };
702
703        let json = serde_json::to_string_pretty(&trajectory).unwrap();
704        let restored: Trajectory = serde_json::from_str(&json).unwrap();
705        assert_eq!(restored.schema_version, ATIF_SCHEMA_VERSION);
706        assert_eq!(restored.session_id, "test-session");
707        assert_eq!(restored.steps.len(), 1);
708    }
709
710    #[test]
711    fn builder_thread_started_sets_session_id() {
712        let mut builder = AtifTrajectoryBuilder::new(AtifAgent::vtcode());
713        let event = ThreadEvent::ThreadStarted(ThreadStartedEvent { thread_id: "thread-abc".to_string() });
714        builder.process_event_at(&event, fixed_ts());
715        let trajectory = builder.finish(None);
716        assert_eq!(trajectory.session_id, "thread-abc");
717    }
718
719    #[test]
720    fn builder_agent_message_step() {
721        let mut builder = AtifTrajectoryBuilder::new(AtifAgent::vtcode());
722        let event = ThreadEvent::ItemCompleted(ItemCompletedEvent {
723            item: ThreadItem {
724                id: "msg-1".to_string(),
725                details: ThreadItemDetails::AgentMessage(AgentMessageItem { text: "Hello, world!".to_string() }),
726            },
727        });
728        builder.process_event_at(&event, fixed_ts());
729        let trajectory = builder.finish(None);
730
731        assert_eq!(trajectory.steps.len(), 1);
732        let step = &trajectory.steps[0];
733        assert_eq!(step.step_id, 1);
734        assert_eq!(step.source, StepSource::Agent);
735        assert_eq!(step.message.as_deref(), Some("Hello, world!"));
736    }
737
738    #[test]
739    fn builder_tool_invocation_with_output() {
740        let mut builder = AtifTrajectoryBuilder::new(AtifAgent::vtcode());
741        let ts = fixed_ts();
742
743        // Tool invocation
744        let inv_event = ThreadEvent::ItemCompleted(ItemCompletedEvent {
745            item: ThreadItem {
746                id: "tool_1".to_string(),
747                details: ThreadItemDetails::ToolInvocation(ToolInvocationItem {
748                    tool_name: "read_file".to_string(),
749                    arguments: Some(serde_json::json!({"path": "README.md"})),
750                    tool_call_id: Some("tc_0".to_string()),
751                    status: ToolCallStatus::Completed,
752                    outcome: None,
753                }),
754            },
755        });
756        builder.process_event_at(&inv_event, ts);
757
758        // Tool output
759        let out_event = ThreadEvent::ItemCompleted(ItemCompletedEvent {
760            item: ThreadItem {
761                id: "tool_1:output".to_string(),
762                details: ThreadItemDetails::ToolOutput(ToolOutputItem {
763                    call_id: "tool_1".to_string(),
764                    tool_call_id: Some("tc_0".to_string()),
765                    spool_path: None,
766                    output: "file contents here".to_string(),
767                    exit_code: Some(0),
768                    status: ToolCallStatus::Completed,
769                }),
770            },
771        });
772        builder.process_event_at(&out_event, ts);
773
774        let trajectory = builder.finish(None);
775        // Only one step: the invocation is buffered until output arrives
776        assert_eq!(trajectory.steps.len(), 1);
777        let step = &trajectory.steps[0];
778        assert_eq!(step.source, StepSource::Agent);
779
780        let calls = step.tool_calls.as_ref().unwrap();
781        assert_eq!(calls.len(), 1);
782        assert_eq!(calls[0].function_name, "read_file");
783        assert_eq!(calls[0].tool_call_id, "tc_0");
784
785        let obs = step.observation.as_ref().unwrap();
786        assert_eq!(obs.results.len(), 1);
787        assert_eq!(obs.results[0].content, "file contents here");
788    }
789
790    #[test]
791    fn builder_turn_completed_accumulates_metrics() {
792        let mut builder = AtifTrajectoryBuilder::new(AtifAgent::vtcode());
793        let event = ThreadEvent::TurnCompleted(TurnCompletedEvent {
794            usage: Usage {
795                input_tokens: 500,
796                cached_input_tokens: 100,
797                cache_creation_tokens: 0,
798                output_tokens: 200,
799            },
800        });
801        builder.process_event_at(&event, fixed_ts());
802
803        let trajectory = builder.finish(None);
804        let fm = trajectory.final_metrics.as_ref().unwrap();
805        assert_eq!(fm.total_prompt_tokens, Some(500));
806        assert_eq!(fm.total_completion_tokens, Some(200));
807        assert_eq!(fm.total_cached_tokens, Some(100));
808    }
809
810    #[test]
811    fn step_metrics_from_usage() {
812        let usage = Usage {
813            input_tokens: 1000,
814            cached_input_tokens: 200,
815            cache_creation_tokens: 50,
816            output_tokens: 300,
817        };
818        let metrics = StepMetrics::from_usage(&usage);
819        assert_eq!(metrics.prompt_tokens, Some(1000));
820        assert_eq!(metrics.completion_tokens, Some(300));
821        assert_eq!(metrics.cached_tokens, Some(200));
822        assert!(metrics.extra.is_some());
823    }
824
825    #[test]
826    fn builder_implements_event_emitter() {
827        let mut builder = AtifTrajectoryBuilder::new(AtifAgent::vtcode());
828        let event = ThreadEvent::ThreadStarted(ThreadStartedEvent { thread_id: "t-1".to_string() });
829        // Use EventEmitter trait
830        crate::EventEmitter::emit(&mut builder, &event);
831        assert_eq!(builder.step_count(), 0); // ThreadStarted doesn't create a step
832    }
833
834    #[test]
835    fn skips_lifecycle_events() {
836        let mut builder = AtifTrajectoryBuilder::new(AtifAgent::vtcode());
837        builder.process_event(&ThreadEvent::TurnStarted(TurnStartedEvent::default()));
838        assert_eq!(builder.step_count(), 0);
839    }
840
841    #[test]
842    fn compact_boundary_with_segment_metadata_preserves_atif_export() {
843        let mut builder = AtifTrajectoryBuilder::new(AtifAgent::vtcode());
844        let event = ThreadEvent::ThreadCompactBoundary(ThreadCompactBoundaryEvent {
845            thread_id: "thread-1".to_string(),
846            trigger: CompactionTrigger::Auto,
847            mode: CompactionMode::Local,
848            original_message_count: 12,
849            compacted_message_count: 5,
850            history_artifact_path: None,
851            previous_segment_id: Some("segment-0001".to_string()),
852            new_segment_id: Some("segment-0002".to_string()),
853            previous_prefix_hash: Some("prefix-before".to_string()),
854            new_prefix_hash: Some("prefix-after".to_string()),
855            previous_catalog_hash: Some("catalog-before".to_string()),
856            new_catalog_hash: Some("catalog-after".to_string()),
857        });
858
859        builder.process_event_at(&event, fixed_ts());
860        let trajectory = builder.finish(None);
861
862        assert_eq!(trajectory.steps.len(), 1);
863        assert_eq!(trajectory.steps[0].source, StepSource::System);
864        assert_eq!(
865            trajectory.steps[0].message.as_deref(),
866            Some("context_compaction: 12 messages -> 5 messages (auto)")
867        );
868    }
869}
vtcode_exec_events/atif.rs

vtcode_exec_events/
atif.rs
