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                if !e.in_progress_exec_sessions.is_empty() {
404                    step.extra = Some(serde_json::json!({
405                        "in_progress_exec_sessions": e.in_progress_exec_sessions,
406                    }));
407                }
408                self.push_step(step);
409            }
410            ThreadEvent::TurnFailed(e) => {
411                if let Some(usage) = &e.usage {
412                    self.saw_per_turn_usage = true;
413                    self.total_input_tokens = self.total_input_tokens.saturating_add(usage.input_tokens);
414                    self.total_output_tokens = self.total_output_tokens.saturating_add(usage.output_tokens);
415                    self.total_cached_tokens = self.total_cached_tokens.saturating_add(usage.cached_input_tokens);
416                }
417                let mut step = Step::system(self.next_step_id, &e.message);
418                step.timestamp = Some(ts_str);
419                step.metrics = e.usage.as_ref().map(StepMetrics::from_usage);
420                self.push_step(step);
421            }
422            ThreadEvent::TurnBlocked(e) => {
423                if let Some(usage) = &e.usage {
424                    self.saw_per_turn_usage = true;
425                    self.total_input_tokens = self.total_input_tokens.saturating_add(usage.input_tokens);
426                    self.total_output_tokens = self.total_output_tokens.saturating_add(usage.output_tokens);
427                    self.total_cached_tokens = self.total_cached_tokens.saturating_add(usage.cached_input_tokens);
428                }
429                let mut step = Step::system(self.next_step_id, &e.message);
430                step.timestamp = Some(ts_str);
431                step.metrics = e.usage.as_ref().map(StepMetrics::from_usage);
432                step.extra = Some(serde_json::json!({
433                    "last_tool": e.last_tool,
434                    "blocked_streak": e.blocked_streak,
435                    "blocked_total": e.blocked_total,
436                    "consecutive_cap": e.consecutive_cap,
437                    "total_cap": e.total_cap,
438                    "recovery_active": e.recovery_active,
439                }));
440                self.push_step(step);
441            }
442            ThreadEvent::ItemCompleted(e) => {
443                self.process_item_completed(&e.item.id, &e.item.details, &ts_str);
444            }
445            ThreadEvent::ThreadCompactBoundary(e) => {
446                let msg = format!(
447                    "context_compaction: {} messages -> {} messages ({})",
448                    e.original_message_count,
449                    e.compacted_message_count,
450                    e.trigger.as_str()
451                );
452                let mut step = Step::system(self.next_step_id, msg);
453                step.timestamp = Some(ts_str);
454                self.push_step(step);
455            }
456            ThreadEvent::ContextReset(e) => {
457                let msg = format!(
458                    "context_reset: {}% context used; plan preserved: {}; tool budget reset: {}",
459                    e.previous_context_usage_percent, e.plan_preserved, e.tool_budget_reset
460                );
461                let mut step = Step::system(self.next_step_id, msg);
462                step.timestamp = Some(ts_str);
463                step.extra = Some(serde_json::json!({
464                    "thread_id": e.thread_id,
465                    "turn_id": e.turn_id,
466                    "trigger": e.trigger,
467                    "plan_preserved": e.plan_preserved,
468                    "previous_context_usage_percent": e.previous_context_usage_percent,
469                    "tool_budget_reset": e.tool_budget_reset,
470                }));
471                self.push_step(step);
472            }
473            ThreadEvent::Error(e) => {
474                let mut step = Step::system(self.next_step_id, &e.message);
475                step.timestamp = Some(ts_str);
476                self.push_step(step);
477            }
478            // Skip streaming/lifecycle events that don't map to ATIF steps
479            ThreadEvent::TurnStarted(_)
480            | ThreadEvent::ItemStarted(_)
481            | ThreadEvent::ItemUpdated(_)
482            | ThreadEvent::PlanDelta(_)
483            | ThreadEvent::PlanApprovalRequested(_)
484            | ThreadEvent::PlanApprovalResolved(_)
485            | ThreadEvent::PermissionRequested(_)
486            | ThreadEvent::PermissionResolved(_)
487            | ThreadEvent::Interjected(_)
488            | ThreadEvent::Unknown => {}
489        }
490    }
491
492    fn process_item_completed(&mut self, item_id: &str, details: &ThreadItemDetails, ts: &str) {
493        match details {
494            ThreadItemDetails::AgentMessage(msg) => {
495                let mut step = Step::agent(self.next_step_id, &msg.text);
496                step.timestamp = Some(ts.to_string());
497                self.push_step(step);
498            }
499            ThreadItemDetails::Plan(plan) => {
500                let mut step = Step::agent(self.next_step_id, &plan.text);
501                step.timestamp = Some(ts.to_string());
502                step.extra = Some(serde_json::json!({ "vtcode_item_type": "plan" }));
503                self.push_step(step);
504            }
505            ThreadItemDetails::Reasoning(r) => {
506                let mut step = Step::agent(self.next_step_id, "");
507                step.timestamp = Some(ts.to_string());
508                step.reasoning_content = Some(r.text.clone());
509                step.message = None;
510                self.push_step(step);
511            }
512            ThreadItemDetails::ToolInvocation(inv) => {
513                // Buffer the invocation; we'll pair it with the ToolOutput
514                self.pending_tool_calls.push(PendingToolCall {
515                    call_id: item_id.to_string(),
516                    tool_call_id: inv.tool_call_id.clone(),
517                    tool_name: inv.tool_name.clone(),
518                    arguments: inv.arguments.clone(),
519                    timestamp: ts.to_string(),
520                });
521            }
522            ThreadItemDetails::ToolOutput(output) => {
523                // Find the matching pending invocation
524                let pending_idx = self.pending_tool_calls.iter().position(|p| p.call_id == output.call_id);
525
526                let (tool_name, arguments, tool_call_id, inv_ts) = if let Some(idx) = pending_idx {
527                    let p = self.pending_tool_calls.remove(idx);
528                    (p.tool_name, p.arguments, p.tool_call_id, p.timestamp)
529                } else {
530                    ("unknown".to_string(), None, output.tool_call_id.clone(), ts.to_string())
531                };
532
533                let call_id = tool_call_id.clone().unwrap_or_else(|| output.call_id.clone());
534
535                let mut step = Step::agent(self.next_step_id, "");
536                step.timestamp = Some(inv_ts);
537                step.message = None;
538                step.tool_calls = Some(vec![AtifToolCall {
539                    tool_call_id: call_id.clone(),
540                    function_name: tool_name,
541                    arguments,
542                }]);
543
544                let status_suffix = match output.status {
545                    ToolCallStatus::Failed => " [FAILED]",
546                    ToolCallStatus::InProgress => " [IN_PROGRESS]",
547                    ToolCallStatus::Completed => "",
548                };
549                let content = format!("{}{}", output.output, status_suffix);
550                step.observation = Some(Observation {
551                    results: vec![ObservationResult { source_call_id: call_id, content }],
552                });
553                self.push_step(step);
554            }
555            ThreadItemDetails::CommandExecution(cmd) => {
556                let call_id = item_id.to_string();
557                let mut step = Step::agent(self.next_step_id, "");
558                step.timestamp = Some(ts.to_string());
559                step.message = None;
560                step.tool_calls = Some(vec![AtifToolCall {
561                    tool_call_id: call_id.clone(),
562                    function_name: "command_execution".to_string(),
563                    arguments: Some(serde_json::json!({
564                        "command": cmd.command,
565                        "arguments": cmd.arguments,
566                    })),
567                }]);
568                step.observation = Some(Observation {
569                    results: vec![ObservationResult {
570                        source_call_id: call_id,
571                        content: cmd.aggregated_output.clone(),
572                    }],
573                });
574                if let Some(exit_code) = cmd.exit_code {
575                    step.extra = Some(serde_json::json!({ "exit_code": exit_code }));
576                }
577                self.push_step(step);
578            }
579            ThreadItemDetails::McpToolCall(mcp) => {
580                let call_id = item_id.to_string();
581                let mut step = Step::agent(self.next_step_id, "");
582                step.timestamp = Some(ts.to_string());
583                step.message = None;
584                step.tool_calls = Some(vec![AtifToolCall {
585                    tool_call_id: call_id.clone(),
586                    function_name: mcp.tool_name.clone(),
587                    arguments: mcp.arguments.clone(),
588                }]);
589                if let Some(result) = &mcp.result {
590                    step.observation = Some(Observation {
591                        results: vec![ObservationResult { source_call_id: call_id, content: result.clone() }],
592                    });
593                }
594                self.push_step(step);
595            }
596            ThreadItemDetails::FileChange(fc) => {
597                let changes: Vec<String> = fc.changes.iter().map(|c| format!("{}: {:?}", c.path, c.kind)).collect();
598                let msg = format!("file_changes: {}", changes.join(", "));
599                let mut step = Step::system(self.next_step_id, msg);
600                step.timestamp = Some(ts.to_string());
601                self.push_step(step);
602            }
603            ThreadItemDetails::WebSearch(ws) => {
604                let mut step = Step::system(self.next_step_id, format!("web_search: {}", ws.query));
605                step.timestamp = Some(ts.to_string());
606                if let Some(results) = &ws.results {
607                    step.observation = Some(Observation {
608                        results: results
609                            .iter()
610                            .enumerate()
611                            .map(|(i, r)| ObservationResult {
612                                source_call_id: format!("search_{i}"),
613                                content: r.clone(),
614                            })
615                            .collect(),
616                    });
617                }
618                self.push_step(step);
619            }
620            ThreadItemDetails::Harness(h) => {
621                let msg = format!("harness: {:?}", h.event);
622                let mut step = Step::system(self.next_step_id, msg);
623                step.timestamp = Some(ts.to_string());
624                let mut extra = serde_json::Map::new();
625                if let Some(m) = &h.message {
626                    let _ = extra.insert("harness_message".to_string(), Value::String(m.clone()));
627                }
628                if h.event == crate::HarnessEventKind::BackgroundSubprocessCompleted {
629                    for (key, value) in [
630                        ("task_id", h.task_id.as_ref()),
631                        ("session_id", h.session_id.as_ref()),
632                        ("exec_session_id", h.exec_session_id.as_ref()),
633                        ("status", h.status.as_ref()),
634                        ("transcript_path", h.transcript_path.as_ref()),
635                        ("archive_path", h.archive_path.as_ref()),
636                        ("error_category", h.error_category.as_ref()),
637                    ] {
638                        if let Some(value) = value {
639                            let _ = extra.insert(key.to_string(), Value::String(value.clone()));
640                        }
641                    }
642                    if let Some(exit_code) = h.exit_code {
643                        let _ = extra.insert("exit_code".to_string(), Value::from(exit_code));
644                    }
645                }
646                if !extra.is_empty() {
647                    step.extra = Some(Value::Object(extra));
648                }
649                self.push_step(step);
650            }
651            ThreadItemDetails::Error(e) => {
652                let mut step = Step::system(self.next_step_id, &e.message);
653                step.timestamp = Some(ts.to_string());
654                self.push_step(step);
655            }
656        }
657    }
658
659    fn push_step(&mut self, step: Step) {
660        self.next_step_id = step.step_id + 1;
661        self.steps.push(step);
662    }
663
664    /// Consume the builder and produce the final ATIF trajectory.
665    ///
666    /// Pass optional `FinalMetrics` to override the accumulated values.
667    /// If `None`, final metrics are derived from observed events.
668    pub fn finish(self, override_metrics: Option<FinalMetrics>) -> Trajectory {
669        let final_metrics = override_metrics.unwrap_or_else(|| FinalMetrics {
670            total_prompt_tokens: Some(self.total_input_tokens),
671            total_completion_tokens: Some(self.total_output_tokens),
672            total_cached_tokens: if self.total_cached_tokens > 0 {
673                Some(self.total_cached_tokens)
674            } else {
675                None
676            },
677            total_cost_usd: None,
678            total_steps: Some(self.steps.len() as u64),
679            extra: Some(serde_json::json!({ "num_turns": self.num_turns })),
680        });
681
682        Trajectory {
683            schema_version: ATIF_SCHEMA_VERSION.to_string(),
684            session_id: self.session_id.unwrap_or_else(|| uuid::Uuid::new_v4().to_string()),
685            agent: self.agent,
686            steps: self.steps,
687            notes: None,
688            final_metrics: Some(final_metrics),
689            extra: None,
690        }
691    }
692
693    /// Returns the number of steps collected so far.
694    pub fn step_count(&self) -> usize {
695        self.steps.len()
696    }
697}
698
699impl crate::EventEmitter for AtifTrajectoryBuilder {
700    fn emit(&mut self, event: &ThreadEvent) {
701        self.process_event(event);
702    }
703}
704
705#[cfg(test)]
706mod tests {
707    use super::*;
708    use crate::{
709        AgentMessageItem, CompactionMode, CompactionTrigger, HarnessEventItem, HarnessEventKind, ItemCompletedEvent,
710        ThreadCompactBoundaryEvent, ThreadItem, ThreadStartedEvent, ToolInvocationItem, ToolOutputItem,
711        TurnCompletedEvent, TurnStartedEvent, Usage,
712    };
713
714    fn fixed_ts() -> DateTime<Utc> {
715        "2025-01-15T10:30:00Z".parse().unwrap()
716    }
717
718    #[test]
719    fn trajectory_round_trip() {
720        let trajectory = Trajectory {
721            schema_version: ATIF_SCHEMA_VERSION.to_string(),
722            session_id: "test-session".to_string(),
723            agent: AtifAgent::vtcode(),
724            steps: vec![Step::user(1, "hello")],
725            notes: None,
726            final_metrics: None,
727            extra: None,
728        };
729
730        let json = serde_json::to_string_pretty(&trajectory).unwrap();
731        let restored: Trajectory = serde_json::from_str(&json).unwrap();
732        assert_eq!(restored.schema_version, ATIF_SCHEMA_VERSION);
733        assert_eq!(restored.session_id, "test-session");
734        assert_eq!(restored.steps.len(), 1);
735    }
736
737    #[test]
738    fn builder_thread_started_sets_session_id() {
739        let mut builder = AtifTrajectoryBuilder::new(AtifAgent::vtcode());
740        let event = ThreadEvent::ThreadStarted(ThreadStartedEvent { thread_id: "thread-abc".to_string() });
741        builder.process_event_at(&event, fixed_ts());
742        let trajectory = builder.finish(None);
743        assert_eq!(trajectory.session_id, "thread-abc");
744    }
745
746    #[test]
747    fn builder_agent_message_step() {
748        let mut builder = AtifTrajectoryBuilder::new(AtifAgent::vtcode());
749        let event = ThreadEvent::ItemCompleted(ItemCompletedEvent {
750            item: ThreadItem {
751                id: "msg-1".to_string(),
752                details: ThreadItemDetails::AgentMessage(AgentMessageItem { text: "Hello, world!".to_string() }),
753            },
754        });
755        builder.process_event_at(&event, fixed_ts());
756        let trajectory = builder.finish(None);
757
758        assert_eq!(trajectory.steps.len(), 1);
759        let step = &trajectory.steps[0];
760        assert_eq!(step.step_id, 1);
761        assert_eq!(step.source, StepSource::Agent);
762        assert_eq!(step.message.as_deref(), Some("Hello, world!"));
763    }
764
765    #[test]
766    fn background_completion_preserves_identity_in_atif_extra() {
767        let mut builder = AtifTrajectoryBuilder::new(AtifAgent::vtcode());
768        let event = ThreadEvent::ItemCompleted(ItemCompletedEvent {
769            item: ThreadItem {
770                id: "background-completion:task-1:exec-1".to_string(),
771                details: ThreadItemDetails::Harness(Box::new(HarnessEventItem {
772                    event: HarnessEventKind::BackgroundSubprocessCompleted,
773                    message: Some("completed successfully".to_string()),
774                    command: None,
775                    path: None,
776                    exit_code: Some(0),
777                    attempt: None,
778                    error_category: Some("background_subprocess".to_string()),
779                    duration_ms: None,
780                    task_id: Some("task-1".to_string()),
781                    session_id: Some("session-1".to_string()),
782                    exec_session_id: Some("exec-1".to_string()),
783                    status: Some("stopped".to_string()),
784                    transcript_path: Some("/tmp/transcript.jsonl".to_string()),
785                    archive_path: Some("/tmp/archive.json".to_string()),
786                })),
787            },
788        });
789
790        builder.process_event_at(&event, fixed_ts());
791        let trajectory = builder.finish(None);
792        let step = trajectory.steps.first().expect("background completion step");
793        let extra = step.extra.as_ref().expect("background completion metadata");
794        assert_eq!(step.message.as_deref(), Some("harness: BackgroundSubprocessCompleted"));
795        assert_eq!(extra["harness_message"], "completed successfully");
796        assert_eq!(extra["task_id"], "task-1");
797        assert_eq!(extra["session_id"], "session-1");
798        assert_eq!(extra["exec_session_id"], "exec-1");
799        assert_eq!(extra["status"], "stopped");
800        assert_eq!(extra["exit_code"], 0);
801        assert_eq!(extra["transcript_path"], "/tmp/transcript.jsonl");
802        assert_eq!(extra["archive_path"], "/tmp/archive.json");
803        assert_eq!(extra["error_category"], "background_subprocess");
804    }
805
806    #[test]
807    fn builder_tool_invocation_with_output() {
808        let mut builder = AtifTrajectoryBuilder::new(AtifAgent::vtcode());
809        let ts = fixed_ts();
810
811        // Tool invocation
812        let inv_event = ThreadEvent::ItemCompleted(ItemCompletedEvent {
813            item: ThreadItem {
814                id: "tool_1".to_string(),
815                details: ThreadItemDetails::ToolInvocation(Box::new(ToolInvocationItem {
816                    tool_name: "read_file".to_string(),
817                    arguments: Some(serde_json::json!({"path": "README.md"})),
818                    tool_call_id: Some("tc_0".to_string()),
819                    status: ToolCallStatus::Completed,
820                    outcome: None,
821                })),
822            },
823        });
824        builder.process_event_at(&inv_event, ts);
825
826        // Tool output
827        let out_event = ThreadEvent::ItemCompleted(ItemCompletedEvent {
828            item: ThreadItem {
829                id: "tool_1:output".to_string(),
830                details: ThreadItemDetails::ToolOutput(Box::new(ToolOutputItem {
831                    call_id: "tool_1".to_string(),
832                    tool_call_id: Some("tc_0".to_string()),
833                    spool_path: None,
834                    output: "file contents here".to_string(),
835                    exit_code: Some(0),
836                    status: ToolCallStatus::Completed,
837                })),
838            },
839        });
840        builder.process_event_at(&out_event, ts);
841
842        let trajectory = builder.finish(None);
843        // Only one step: the invocation is buffered until output arrives
844        assert_eq!(trajectory.steps.len(), 1);
845        let step = &trajectory.steps[0];
846        assert_eq!(step.source, StepSource::Agent);
847
848        let calls = step.tool_calls.as_ref().unwrap();
849        assert_eq!(calls.len(), 1);
850        assert_eq!(calls[0].function_name, "read_file");
851        assert_eq!(calls[0].tool_call_id, "tc_0");
852
853        let obs = step.observation.as_ref().unwrap();
854        assert_eq!(obs.results.len(), 1);
855        assert_eq!(obs.results[0].content, "file contents here");
856    }
857
858    #[test]
859    fn builder_turn_completed_accumulates_metrics() {
860        let mut builder = AtifTrajectoryBuilder::new(AtifAgent::vtcode());
861        let event = ThreadEvent::TurnCompleted(TurnCompletedEvent {
862            usage: Usage {
863                input_tokens: 500,
864                cached_input_tokens: 100,
865                cache_creation_tokens: 0,
866                output_tokens: 200,
867            },
868            in_progress_exec_sessions: Vec::new(),
869        });
870        builder.process_event_at(&event, fixed_ts());
871
872        let trajectory = builder.finish(None);
873        let fm = trajectory.final_metrics.as_ref().unwrap();
874        assert_eq!(fm.total_prompt_tokens, Some(500));
875        assert_eq!(fm.total_completion_tokens, Some(200));
876        assert_eq!(fm.total_cached_tokens, Some(100));
877    }
878
879    #[test]
880    fn builder_turn_completed_preserves_in_progress_sessions_for_resume() {
881        let mut builder = AtifTrajectoryBuilder::new(AtifAgent::vtcode());
882        let event = ThreadEvent::TurnCompleted(TurnCompletedEvent {
883            usage: Usage::default(),
884            in_progress_exec_sessions: vec!["run-7".to_string()],
885        });
886        builder.process_event_at(&event, fixed_ts());
887
888        let trajectory = builder.finish(None);
889        let step = trajectory.steps.last().expect("turn_completed step");
890        let extra = step.extra.clone().expect("extra carries resume ids");
891        assert_eq!(extra["in_progress_exec_sessions"], serde_json::json!(["run-7"]));
892
893        // Empty ids stay omitted so steady-state export is unchanged.
894        let mut empty_builder = AtifTrajectoryBuilder::new(AtifAgent::vtcode());
895        let empty = ThreadEvent::TurnCompleted(TurnCompletedEvent {
896            usage: Usage::default(),
897            in_progress_exec_sessions: Vec::new(),
898        });
899        empty_builder.process_event_at(&empty, fixed_ts());
900        let empty_trajectory = empty_builder.finish(None);
901        assert!(empty_trajectory.steps.last().expect("step").extra.is_none());
902    }
903
904    #[test]
905    fn step_metrics_from_usage() {
906        let usage = Usage {
907            input_tokens: 1000,
908            cached_input_tokens: 200,
909            cache_creation_tokens: 50,
910            output_tokens: 300,
911        };
912        let metrics = StepMetrics::from_usage(&usage);
913        assert_eq!(metrics.prompt_tokens, Some(1000));
914        assert_eq!(metrics.completion_tokens, Some(300));
915        assert_eq!(metrics.cached_tokens, Some(200));
916        assert!(metrics.extra.is_some());
917    }
918
919    #[test]
920    fn builder_implements_event_emitter() {
921        let mut builder = AtifTrajectoryBuilder::new(AtifAgent::vtcode());
922        let event = ThreadEvent::ThreadStarted(ThreadStartedEvent { thread_id: "t-1".to_string() });
923        // Use EventEmitter trait
924        crate::EventEmitter::emit(&mut builder, &event);
925        assert_eq!(builder.step_count(), 0); // ThreadStarted doesn't create a step
926    }
927
928    #[test]
929    fn skips_lifecycle_events() {
930        let mut builder = AtifTrajectoryBuilder::new(AtifAgent::vtcode());
931        builder.process_event(&ThreadEvent::TurnStarted(TurnStartedEvent::default()));
932        assert_eq!(builder.step_count(), 0);
933    }
934
935    #[test]
936    fn compact_boundary_with_segment_metadata_preserves_atif_export() {
937        let mut builder = AtifTrajectoryBuilder::new(AtifAgent::vtcode());
938        let event = ThreadEvent::ThreadCompactBoundary(Box::new(ThreadCompactBoundaryEvent {
939            thread_id: "thread-1".to_string(),
940            trigger: CompactionTrigger::Auto,
941            mode: CompactionMode::Local,
942            original_message_count: 12,
943            compacted_message_count: 5,
944            history_artifact_path: None,
945            previous_segment_id: Some("segment-0001".to_string()),
946            new_segment_id: Some("segment-0002".to_string()),
947            previous_prefix_hash: Some("prefix-before".to_string()),
948            new_prefix_hash: Some("prefix-after".to_string()),
949            previous_catalog_hash: Some("catalog-before".to_string()),
950            new_catalog_hash: Some("catalog-after".to_string()),
951        }));
952
953        builder.process_event_at(&event, fixed_ts());
954        let trajectory = builder.finish(None);
955
956        assert_eq!(trajectory.steps.len(), 1);
957        assert_eq!(trajectory.steps[0].source, StepSource::System);
958        assert_eq!(
959            trajectory.steps[0].message.as_deref(),
960            Some("context_compaction: 12 messages -> 5 messages (auto)")
961        );
962    }
963}
vtcode_exec_events/atif.rs

vtcode_exec_events/
atif.rs
