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.
315#[must_use = "finish the trajectory before discarding the builder"]
316pub struct AtifTrajectoryBuilder {
317    completed_at: Option<String>,
318    agent: AtifAgent,
319    session_id: Option<String>,
320    steps: Vec<Step>,
321    next_step_id: u64,
322    // Running token accumulators for final metrics
323    total_input_tokens: u64,
324    total_output_tokens: u64,
325    total_cached_tokens: u64,
326    num_turns: usize,
327    /// Whether any per-turn `TurnCompleted`/`TurnFailed` usage was observed.
328    /// Guards the `ThreadCompleted` aggregate from double-counting usage.
329    saw_per_turn_usage: bool,
330    /// Pending tool invocations awaiting matching ToolOutput.
331    pending_tool_calls: Vec<PendingToolCall>,
332}
333
334struct PendingToolCall {
335    call_id: String,
336    tool_call_id: Option<String>,
337    tool_name: String,
338    arguments: Option<Value>,
339    timestamp: String,
340}
341
342impl AtifTrajectoryBuilder {
343    /// Create a new builder for the given agent.
344    pub fn new(agent: AtifAgent) -> Self {
345        Self {
346            agent,
347            completed_at: None,
348            session_id: None,
349            steps: Vec::new(),
350            next_step_id: 1,
351            total_input_tokens: 0,
352            total_output_tokens: 0,
353            total_cached_tokens: 0,
354            num_turns: 0,
355            saw_per_turn_usage: false,
356            pending_tool_calls: Vec::new(),
357        }
358    }
359
360    /// Set the session ID explicitly. If not set, it will be derived from
361    /// `ThreadStarted` or `ThreadCompleted` events.
362    pub fn set_session_id(&mut self, id: impl Into<String>) {
363        self.session_id = Some(id.into());
364    }
365
366    /// Process a thread event using the current wall-clock time.
367    pub fn process_event(&mut self, event: &ThreadEvent) {
368        self.process_event_at(event, Utc::now());
369    }
370
371    /// Process a thread event with an explicit timestamp (for deterministic tests).
372    pub fn process_event_at(&mut self, event: &ThreadEvent, ts: DateTime<Utc>) {
373        let first_step = self.steps.len();
374        let ts_str = ts.to_rfc3339();
375        match event {
376            ThreadEvent::ThreadStarted(e) => {
377                if self.session_id.is_none() {
378                    self.session_id = Some(e.thread_id.clone());
379                }
380            }
381            ThreadEvent::ThreadCompleted(e) => {
382                self.completed_at = e.completed_at.clone();
383                if self.session_id.is_none() {
384                    self.session_id = Some(e.session_id.clone());
385                }
386                self.num_turns = e.num_turns;
387                // `ThreadCompleted` carries the *aggregate* usage, which equals
388                // the sum of per-turn usage already accumulated by the
389                // `TurnCompleted`/`TurnFailed` arms. Only fall back to it when
390                // no per-turn usage was observed (e.g. a harness that emits
391                // only the thread aggregate), otherwise totals double-count.
392                if !self.saw_per_turn_usage {
393                    self.total_input_tokens = self.total_input_tokens.saturating_add(e.usage.input_tokens);
394                    self.total_output_tokens = self.total_output_tokens.saturating_add(e.usage.output_tokens);
395                    self.total_cached_tokens = self.total_cached_tokens.saturating_add(e.usage.cached_input_tokens);
396                }
397            }
398            ThreadEvent::TurnCompleted(e) => {
399                self.saw_per_turn_usage = true;
400                self.total_input_tokens = self.total_input_tokens.saturating_add(e.usage.input_tokens);
401                self.total_output_tokens = self.total_output_tokens.saturating_add(e.usage.output_tokens);
402                self.total_cached_tokens = self.total_cached_tokens.saturating_add(e.usage.cached_input_tokens);
403                self.num_turns += 1;
404
405                let mut step = Step::system(self.next_step_id, "turn_completed");
406                step.timestamp = Some(ts_str);
407                step.metrics = Some(StepMetrics::from_usage(&e.usage));
408                if !e.in_progress_exec_sessions.is_empty() {
409                    step.extra = Some(serde_json::json!({
410                        "in_progress_exec_sessions": e.in_progress_exec_sessions,
411                    }));
412                }
413                self.push_step(step);
414            }
415            ThreadEvent::TurnFailed(e) => {
416                if let Some(usage) = &e.usage {
417                    self.saw_per_turn_usage = true;
418                    self.total_input_tokens = self.total_input_tokens.saturating_add(usage.input_tokens);
419                    self.total_output_tokens = self.total_output_tokens.saturating_add(usage.output_tokens);
420                    self.total_cached_tokens = self.total_cached_tokens.saturating_add(usage.cached_input_tokens);
421                }
422                let mut step = Step::system(self.next_step_id, &e.message);
423                step.timestamp = Some(ts_str);
424                step.metrics = e.usage.as_ref().map(StepMetrics::from_usage);
425                self.push_step(step);
426            }
427            ThreadEvent::TurnBlocked(e) => {
428                if let Some(usage) = &e.usage {
429                    self.saw_per_turn_usage = true;
430                    self.total_input_tokens = self.total_input_tokens.saturating_add(usage.input_tokens);
431                    self.total_output_tokens = self.total_output_tokens.saturating_add(usage.output_tokens);
432                    self.total_cached_tokens = self.total_cached_tokens.saturating_add(usage.cached_input_tokens);
433                }
434                let mut step = Step::system(self.next_step_id, &e.message);
435                step.timestamp = Some(ts_str);
436                step.metrics = e.usage.as_ref().map(StepMetrics::from_usage);
437                step.extra = Some(serde_json::json!({
438                    "last_tool": e.last_tool,
439                    "blocked_streak": e.blocked_streak,
440                    "blocked_total": e.blocked_total,
441                    "consecutive_cap": e.consecutive_cap,
442                    "total_cap": e.total_cap,
443                    "recovery_active": e.recovery_active,
444                }));
445                self.push_step(step);
446            }
447            ThreadEvent::ItemCompleted(e) => {
448                self.process_item_completed(&e.item.id, &e.item.details, &ts_str);
449            }
450            ThreadEvent::ThreadCompactBoundary(e) => {
451                let msg = format!(
452                    "context_compaction: {} messages -> {} messages ({})",
453                    e.original_message_count,
454                    e.compacted_message_count,
455                    e.trigger.as_str()
456                );
457                let mut step = Step::system(self.next_step_id, msg);
458                step.timestamp = Some(ts_str);
459                self.push_step(step);
460            }
461            ThreadEvent::ContextReset(e) => {
462                let msg = format!(
463                    "context_reset: {}% context used; plan preserved: {}; tool budget reset: {}",
464                    e.previous_context_usage_percent, e.plan_preserved, e.tool_budget_reset
465                );
466                let mut step = Step::system(self.next_step_id, msg);
467                step.timestamp = Some(ts_str);
468                step.extra = Some(serde_json::json!({
469                    "thread_id": e.thread_id,
470                    "turn_id": e.turn_id,
471                    "trigger": e.trigger,
472                    "plan_preserved": e.plan_preserved,
473                    "previous_context_usage_percent": e.previous_context_usage_percent,
474                    "tool_budget_reset": e.tool_budget_reset,
475                }));
476                self.push_step(step);
477            }
478            ThreadEvent::Error(e) => {
479                let mut step = Step::system(self.next_step_id, &e.message);
480                step.timestamp = Some(ts_str);
481                self.push_step(step);
482            }
483            // Skip streaming/lifecycle events that don't map to ATIF steps
484            ThreadEvent::TurnStarted(e) => {
485                if let Some(context) = &e.context {
486                    let mut step = Step::agent(self.next_step_id, context.goal.as_deref().unwrap_or("Turn started"));
487                    step.source = StepSource::User;
488                    step.timestamp = Some(context.timestamp.clone());
489                    step.extra = Some(serde_json::json!({"execution_context": context}));
490                    self.push_step(step);
491                }
492            }
493            ThreadEvent::MatrixUpdated(_)
494            | ThreadEvent::ItemStarted(_)
495            | ThreadEvent::ItemUpdated(_)
496            | ThreadEvent::PlanDelta(_)
497            | ThreadEvent::PlanApprovalRequested(_)
498            | ThreadEvent::PlanApprovalResolved(_)
499            | ThreadEvent::PermissionRequested(_)
500            | ThreadEvent::PermissionResolved(_)
501            | ThreadEvent::Interjected(_)
502            | ThreadEvent::Unknown => {}
503        }
504        for step in self.steps.iter_mut().skip(first_step) {
505            let context = match event {
506                ThreadEvent::ItemCompleted(e) => e.item.context.as_deref(),
507                _ => None,
508            };
509            if let Some(context) = context {
510                step.timestamp = Some(context.timestamp.clone());
511                let extra = step.extra.get_or_insert_with(|| serde_json::json!({}));
512                if let Some(extra) = extra.as_object_mut() {
513                    let _ = extra.insert("item_context".into(), serde_json::json!(context));
514                }
515            }
516            match event {
517                ThreadEvent::TurnCompleted(e) => {
518                    if let Some(ts) = &e.completed_at {
519                        step.timestamp = Some(ts.to_string());
520                    }
521                }
522                ThreadEvent::TurnFailed(e) => {
523                    if let Some(ts) = &e.completed_at {
524                        step.timestamp = Some(ts.to_string());
525                    }
526                }
527                ThreadEvent::TurnBlocked(e) => {
528                    if let Some(ts) = &e.completed_at {
529                        step.timestamp = Some(ts.clone());
530                    }
531                }
532                _ => {}
533            }
534        }
535    }
536
537    fn process_item_completed(&mut self, item_id: &str, details: &ThreadItemDetails, ts: &str) {
538        match details {
539            ThreadItemDetails::Decision(d) => {
540                let mut step = Step::agent(self.next_step_id, &d.summary);
541                step.timestamp = Some(ts.to_owned());
542                step.extra = Some(
543                    serde_json::json!({"vtcode_item_type": "decision", "public_rationale": d.rationale, "alternatives": d.alternatives, "evidence_ids": d.evidence_ids}),
544                );
545                self.push_step(step);
546            }
547            ThreadItemDetails::AgentMessage(msg) => {
548                let mut step = Step::agent(self.next_step_id, &msg.text);
549                step.timestamp = Some(ts.to_string());
550                self.push_step(step);
551            }
552            ThreadItemDetails::Plan(plan) => {
553                let mut step = Step::agent(self.next_step_id, &plan.text);
554                step.timestamp = Some(ts.to_string());
555                step.extra = Some(serde_json::json!({ "vtcode_item_type": "plan" }));
556                self.push_step(step);
557            }
558            ThreadItemDetails::Reasoning(r) => {
559                let mut step = Step::agent(self.next_step_id, "");
560                step.timestamp = Some(ts.to_string());
561                if let Some(diagnostic) = &r.diagnostic {
562                    step.message = Some(diagnostic.summary.clone());
563                    step.extra = Some(serde_json::json!({ "diagnostic": diagnostic }));
564                } else {
565                    step.reasoning_content = Some(r.text.clone());
566                    step.message = None;
567                }
568                self.push_step(step);
569            }
570            ThreadItemDetails::ToolInvocation(inv) => {
571                // Buffer the invocation; we'll pair it with the ToolOutput
572                self.pending_tool_calls.push(PendingToolCall {
573                    call_id: item_id.to_string(),
574                    tool_call_id: inv.tool_call_id.clone(),
575                    tool_name: inv.tool_name.clone(),
576                    arguments: inv.arguments.clone(),
577                    timestamp: ts.to_string(),
578                });
579            }
580            ThreadItemDetails::ToolOutput(output) => {
581                // Find the matching pending invocation
582                let pending_idx = self.pending_tool_calls.iter().position(|p| p.call_id == output.call_id);
583
584                let (tool_name, arguments, tool_call_id, inv_ts) = if let Some(idx) = pending_idx {
585                    let p = self.pending_tool_calls.remove(idx);
586                    (p.tool_name, p.arguments, p.tool_call_id, p.timestamp)
587                } else {
588                    ("unknown".to_string(), None, output.tool_call_id.clone(), ts.to_string())
589                };
590
591                let call_id = tool_call_id.clone().unwrap_or_else(|| output.call_id.clone());
592
593                let mut step = Step::agent(self.next_step_id, "");
594                step.timestamp = Some(inv_ts);
595                step.message = None;
596                step.tool_calls = Some(vec![AtifToolCall {
597                    tool_call_id: call_id.clone(),
598                    function_name: tool_name,
599                    arguments,
600                }]);
601
602                let status_suffix = match output.status {
603                    ToolCallStatus::Failed => " [FAILED]",
604                    ToolCallStatus::InProgress => " [IN_PROGRESS]",
605                    ToolCallStatus::Completed => "",
606                };
607                let content = format!("{}{}", output.output, status_suffix);
608                step.observation = Some(Observation {
609                    results: vec![ObservationResult { source_call_id: call_id, content }],
610                });
611                self.push_step(step);
612            }
613            ThreadItemDetails::CommandExecution(cmd) => {
614                let call_id = item_id.to_string();
615                let mut step = Step::agent(self.next_step_id, "");
616                step.timestamp = Some(ts.to_string());
617                step.message = None;
618                step.tool_calls = Some(vec![AtifToolCall {
619                    tool_call_id: call_id.clone(),
620                    function_name: "command_execution".to_string(),
621                    arguments: Some(serde_json::json!({
622                        "command": cmd.command,
623                        "arguments": cmd.arguments,
624                    })),
625                }]);
626                step.observation = Some(Observation {
627                    results: vec![ObservationResult {
628                        source_call_id: call_id,
629                        content: cmd.aggregated_output.clone(),
630                    }],
631                });
632                if let Some(exit_code) = cmd.exit_code {
633                    step.extra = Some(serde_json::json!({ "exit_code": exit_code }));
634                }
635                self.push_step(step);
636            }
637            ThreadItemDetails::McpToolCall(mcp) => {
638                let call_id = item_id.to_string();
639                let mut step = Step::agent(self.next_step_id, "");
640                step.timestamp = Some(ts.to_string());
641                step.message = None;
642                step.tool_calls = Some(vec![AtifToolCall {
643                    tool_call_id: call_id.clone(),
644                    function_name: mcp.tool_name.clone(),
645                    arguments: mcp.arguments.clone(),
646                }]);
647                if let Some(result) = &mcp.result {
648                    step.observation = Some(Observation {
649                        results: vec![ObservationResult { source_call_id: call_id, content: result.clone() }],
650                    });
651                }
652                self.push_step(step);
653            }
654            ThreadItemDetails::FileChange(fc) => {
655                let changes: Vec<String> = fc.changes.iter().map(|c| format!("{}: {:?}", c.path, c.kind)).collect();
656                let msg = format!("file_changes: {}", changes.join(", "));
657                let mut step = Step::system(self.next_step_id, msg);
658                step.timestamp = Some(ts.to_string());
659                self.push_step(step);
660            }
661            ThreadItemDetails::WebSearch(ws) => {
662                let mut step = Step::system(self.next_step_id, format!("web_search: {}", ws.query));
663                step.timestamp = Some(ts.to_string());
664                if let Some(results) = &ws.results {
665                    step.observation = Some(Observation {
666                        results: results
667                            .iter()
668                            .enumerate()
669                            .map(|(i, r)| ObservationResult {
670                                source_call_id: format!("search_{i}"),
671                                content: r.clone(),
672                            })
673                            .collect(),
674                    });
675                }
676                self.push_step(step);
677            }
678            ThreadItemDetails::Harness(h) => {
679                let msg = format!("harness: {:?}", h.event);
680                let mut step = Step::system(self.next_step_id, msg);
681                step.timestamp = Some(ts.to_string());
682                let mut extra = serde_json::Map::new();
683                if let Some(m) = &h.message {
684                    let _ = extra.insert("harness_message".to_string(), Value::String(m.clone()));
685                }
686                if h.event == crate::HarnessEventKind::BackgroundSubprocessCompleted {
687                    for (key, value) in [
688                        ("task_id", h.task_id.as_ref()),
689                        ("session_id", h.session_id.as_ref()),
690                        ("exec_session_id", h.exec_session_id.as_ref()),
691                        ("status", h.status.as_ref()),
692                        ("transcript_path", h.transcript_path.as_ref()),
693                        ("archive_path", h.archive_path.as_ref()),
694                        ("error_category", h.error_category.as_ref()),
695                    ] {
696                        if let Some(value) = value {
697                            let _ = extra.insert(key.to_string(), Value::String(value.clone()));
698                        }
699                    }
700                    if let Some(exit_code) = h.exit_code {
701                        let _ = extra.insert("exit_code".to_string(), Value::from(exit_code));
702                    }
703                }
704                if !extra.is_empty() {
705                    step.extra = Some(Value::Object(extra));
706                }
707                self.push_step(step);
708            }
709            ThreadItemDetails::Error(e) => {
710                let mut step = Step::system(self.next_step_id, &e.message);
711                step.timestamp = Some(ts.to_string());
712                self.push_step(step);
713            }
714        }
715    }
716
717    fn push_step(&mut self, step: Step) {
718        self.next_step_id = step.step_id + 1;
719        self.steps.push(step);
720    }
721
722    /// Consume the builder and produce the final ATIF trajectory.
723    ///
724    /// Pass optional `FinalMetrics` to override the accumulated values.
725    /// If `None`, final metrics are derived from observed events.
726    #[must_use]
727    pub fn finish(self, override_metrics: Option<FinalMetrics>) -> Trajectory {
728        let final_metrics = override_metrics.unwrap_or_else(|| FinalMetrics {
729            total_prompt_tokens: Some(self.total_input_tokens),
730            total_completion_tokens: Some(self.total_output_tokens),
731            total_cached_tokens: if self.total_cached_tokens > 0 {
732                Some(self.total_cached_tokens)
733            } else {
734                None
735            },
736            total_cost_usd: None,
737            total_steps: Some(self.steps.len() as u64),
738            extra: Some(serde_json::json!({ "num_turns": self.num_turns })),
739        });
740
741        Trajectory {
742            schema_version: ATIF_SCHEMA_VERSION.to_string(),
743            session_id: self.session_id.unwrap_or_else(|| uuid::Uuid::new_v4().to_string()),
744            agent: self.agent,
745            steps: self.steps,
746            notes: None,
747            final_metrics: Some(final_metrics),
748            extra: self.completed_at.map(|timestamp| serde_json::json!({"completed_at":timestamp})),
749        }
750    }
751
752    /// Returns the number of steps collected so far.
753    pub fn step_count(&self) -> usize {
754        self.steps.len()
755    }
756}
757
758impl crate::EventEmitter for AtifTrajectoryBuilder {
759    fn emit(&mut self, event: &ThreadEvent) {
760        self.process_event(event);
761    }
762}
763
764#[cfg(test)]
765mod tests {
766    use super::*;
767    use crate::{
768        AgentMessageItem, CompactionMode, CompactionTrigger, HarnessEventItem, HarnessEventKind, ItemCompletedEvent,
769        ThreadCompactBoundaryEvent, ThreadItem, ThreadStartedEvent, ToolInvocationItem, ToolOutputItem,
770        TurnCompletedEvent, TurnStartedEvent, Usage,
771    };
772
773    fn fixed_ts() -> DateTime<Utc> {
774        "2025-01-15T10:30:00Z".parse().unwrap()
775    }
776
777    #[test]
778    fn terminal_timestamps_survive_export_without_synthetic_steps() {
779        let mut builder = AtifTrajectoryBuilder::new(AtifAgent::vtcode());
780        let timestamp = "2026-10-03T01:02:03Z";
781        let blocked = serde_json::from_value(
782            serde_json::json!({"type":"turn.blocked", "message":"Blocked", "completed_at":timestamp}),
783        )
784        .unwrap();
785        builder.process_event_at(&blocked, fixed_ts());
786        let completed = serde_json::from_value(serde_json::json!({"type":"thread.completed", "completed_at":timestamp, "thread_id":"thread", "session_id":"session", "subtype":"success", "outcome_code":"completed", "usage":Usage::default(), "num_turns":1})).unwrap();
787        builder.process_event_at(&completed, fixed_ts());
788        let trajectory = builder.finish(None);
789        assert_eq!(trajectory.steps.len(), 1);
790        assert_eq!(trajectory.steps[0].timestamp.as_deref(), Some(timestamp));
791        assert_eq!(trajectory.extra.as_ref().unwrap()["completed_at"], timestamp);
792    }
793
794    #[test]
795    fn explanation_metadata_and_public_rationale_preserve_recorded_time() {
796        let mut builder = AtifTrajectoryBuilder::new(AtifAgent::vtcode());
797        let context = crate::ItemContext {
798            task_id: "task-1".into(),
799            turn_id: "turn-1".into(),
800            actor_id: "root".into(),
801            parent_actor_id: None,
802            timestamp: "2026-10-03T01:02:03Z".into(),
803            activity: None,
804        };
805        builder.process_event_at(
806            &ThreadEvent::ItemCompleted(ItemCompletedEvent {
807                item: ThreadItem {
808                    id: "decision-1".into(),
809                    context: Some(Box::new(context.clone())),
810                    details: ThreadItemDetails::Decision(Box::new(crate::DecisionItem {
811                        summary: "Reuse parser".into(),
812                        rationale: "Preserve checks".into(),
813                        alternatives: vec!["Replace parser".into()],
814                        evidence_ids: vec!["read-1".into()],
815                    })),
816                },
817            }),
818            fixed_ts(),
819        );
820        let trajectory = builder.finish(None);
821        assert_eq!(trajectory.steps.len(), 1);
822        let step = &trajectory.steps[0];
823        assert_eq!(step.timestamp.as_deref(), Some(context.timestamp.as_str()));
824        let extra = step.extra.as_ref().unwrap();
825        assert_eq!(extra["public_rationale"], "Preserve checks");
826        assert_eq!(extra["item_context"]["task_id"], "task-1");
827        assert_eq!(extra["evidence_ids"][0], "read-1");
828    }
829
830    #[test]
831    fn trajectory_round_trip() {
832        let trajectory = Trajectory {
833            schema_version: ATIF_SCHEMA_VERSION.to_string(),
834            session_id: "test-session".to_string(),
835            agent: AtifAgent::vtcode(),
836            steps: vec![Step::user(1, "hello")],
837            notes: None,
838            final_metrics: None,
839            extra: None,
840        };
841
842        let json = serde_json::to_string_pretty(&trajectory).unwrap();
843        let restored: Trajectory = serde_json::from_str(&json).unwrap();
844        assert_eq!(restored.schema_version, ATIF_SCHEMA_VERSION);
845        assert_eq!(restored.session_id, "test-session");
846        assert_eq!(restored.steps.len(), 1);
847    }
848
849    #[test]
850    fn builder_thread_started_sets_session_id() {
851        let mut builder = AtifTrajectoryBuilder::new(AtifAgent::vtcode());
852        let event = ThreadEvent::ThreadStarted(ThreadStartedEvent { thread_id: "thread-abc".to_string() });
853        builder.process_event_at(&event, fixed_ts());
854        let trajectory = builder.finish(None);
855        assert_eq!(trajectory.session_id, "thread-abc");
856    }
857
858    #[test]
859    fn builder_agent_message_step() {
860        let mut builder = AtifTrajectoryBuilder::new(AtifAgent::vtcode());
861        let event = ThreadEvent::ItemCompleted(ItemCompletedEvent {
862            item: ThreadItem {
863                context: None,
864                id: "msg-1".to_string(),
865                details: ThreadItemDetails::AgentMessage(AgentMessageItem { text: "Hello, world!".to_string() }),
866            },
867        });
868        builder.process_event_at(&event, fixed_ts());
869        let trajectory = builder.finish(None);
870
871        assert_eq!(trajectory.steps.len(), 1);
872        let step = &trajectory.steps[0];
873        assert_eq!(step.step_id, 1);
874        assert_eq!(step.source, StepSource::Agent);
875        assert_eq!(step.message.as_deref(), Some("Hello, world!"));
876    }
877
878    #[test]
879    fn background_completion_preserves_identity_in_atif_extra() {
880        let mut builder = AtifTrajectoryBuilder::new(AtifAgent::vtcode());
881        let event = ThreadEvent::ItemCompleted(ItemCompletedEvent {
882            item: ThreadItem {
883                context: None,
884                id: "background-completion:task-1:exec-1".to_string(),
885                details: ThreadItemDetails::Harness(Box::new(HarnessEventItem {
886                    event: HarnessEventKind::BackgroundSubprocessCompleted,
887                    message: Some("completed successfully".to_string()),
888                    command: None,
889                    path: None,
890                    exit_code: Some(0),
891                    attempt: None,
892                    error_category: Some("background_subprocess".to_string()),
893                    duration_ms: None,
894                    task_id: Some("task-1".to_string()),
895                    session_id: Some("session-1".to_string()),
896                    exec_session_id: Some("exec-1".to_string()),
897                    status: Some("stopped".to_string()),
898                    transcript_path: Some("/tmp/transcript.jsonl".to_string()),
899                    archive_path: Some("/tmp/archive.json".to_string()),
900                })),
901            },
902        });
903
904        builder.process_event_at(&event, fixed_ts());
905        let trajectory = builder.finish(None);
906        let step = trajectory.steps.first().expect("background completion step");
907        let extra = step.extra.as_ref().expect("background completion metadata");
908        assert_eq!(step.message.as_deref(), Some("harness: BackgroundSubprocessCompleted"));
909        assert_eq!(extra["harness_message"], "completed successfully");
910        assert_eq!(extra["task_id"], "task-1");
911        assert_eq!(extra["session_id"], "session-1");
912        assert_eq!(extra["exec_session_id"], "exec-1");
913        assert_eq!(extra["status"], "stopped");
914        assert_eq!(extra["exit_code"], 0);
915        assert_eq!(extra["transcript_path"], "/tmp/transcript.jsonl");
916        assert_eq!(extra["archive_path"], "/tmp/archive.json");
917        assert_eq!(extra["error_category"], "background_subprocess");
918    }
919
920    #[test]
921    fn builder_tool_invocation_with_output() {
922        let mut builder = AtifTrajectoryBuilder::new(AtifAgent::vtcode());
923        let ts = fixed_ts();
924
925        // Tool invocation
926        let inv_event = ThreadEvent::ItemCompleted(ItemCompletedEvent {
927            item: ThreadItem {
928                context: None,
929                id: "tool_1".to_string(),
930                details: ThreadItemDetails::ToolInvocation(Box::new(ToolInvocationItem {
931                    tool_name: "read_file".to_string(),
932                    arguments: Some(serde_json::json!({"path": "README.md"})),
933                    tool_call_id: Some("tc_0".to_string()),
934                    status: ToolCallStatus::Completed,
935                    outcome: None,
936                })),
937            },
938        });
939        builder.process_event_at(&inv_event, ts);
940
941        // Tool output
942        let out_event = ThreadEvent::ItemCompleted(ItemCompletedEvent {
943            item: ThreadItem {
944                context: None,
945                id: "tool_1:output".to_string(),
946                details: ThreadItemDetails::ToolOutput(Box::new(ToolOutputItem {
947                    call_id: "tool_1".to_string(),
948                    tool_call_id: Some("tc_0".to_string()),
949                    spool_path: None,
950                    output: "file contents here".to_string(),
951                    exit_code: Some(0),
952                    status: ToolCallStatus::Completed,
953                })),
954            },
955        });
956        builder.process_event_at(&out_event, ts);
957
958        let trajectory = builder.finish(None);
959        // Only one step: the invocation is buffered until output arrives
960        assert_eq!(trajectory.steps.len(), 1);
961        let step = &trajectory.steps[0];
962        assert_eq!(step.source, StepSource::Agent);
963
964        let calls = step.tool_calls.as_ref().unwrap();
965        assert_eq!(calls.len(), 1);
966        assert_eq!(calls[0].function_name, "read_file");
967        assert_eq!(calls[0].tool_call_id, "tc_0");
968
969        let obs = step.observation.as_ref().unwrap();
970        assert_eq!(obs.results.len(), 1);
971        assert_eq!(obs.results[0].content, "file contents here");
972    }
973
974    #[test]
975    fn builder_turn_completed_accumulates_metrics() {
976        let mut builder = AtifTrajectoryBuilder::new(AtifAgent::vtcode());
977        let event = ThreadEvent::TurnCompleted(TurnCompletedEvent {
978            completed_at: None,
979            usage: Usage {
980                input_tokens: 500,
981                cached_input_tokens: 100,
982                cache_creation_tokens: 0,
983                output_tokens: 200,
984            },
985            in_progress_exec_sessions: Vec::new(),
986        });
987        builder.process_event_at(&event, fixed_ts());
988
989        let trajectory = builder.finish(None);
990        let fm = trajectory.final_metrics.as_ref().unwrap();
991        assert_eq!(fm.total_prompt_tokens, Some(500));
992        assert_eq!(fm.total_completion_tokens, Some(200));
993        assert_eq!(fm.total_cached_tokens, Some(100));
994    }
995
996    #[test]
997    fn builder_turn_completed_preserves_in_progress_sessions_for_resume() {
998        let mut builder = AtifTrajectoryBuilder::new(AtifAgent::vtcode());
999        let event = ThreadEvent::TurnCompleted(TurnCompletedEvent {
1000            completed_at: None,
1001            usage: Usage::default(),
1002            in_progress_exec_sessions: vec!["run-7".to_string()],
1003        });
1004        builder.process_event_at(&event, fixed_ts());
1005
1006        let trajectory = builder.finish(None);
1007        let step = trajectory.steps.last().expect("turn_completed step");
1008        let extra = step.extra.clone().expect("extra carries resume ids");
1009        assert_eq!(extra["in_progress_exec_sessions"], serde_json::json!(["run-7"]));
1010
1011        // Empty ids stay omitted so steady-state export is unchanged.
1012        let mut empty_builder = AtifTrajectoryBuilder::new(AtifAgent::vtcode());
1013        let empty = ThreadEvent::TurnCompleted(TurnCompletedEvent {
1014            completed_at: None,
1015            usage: Usage::default(),
1016            in_progress_exec_sessions: Vec::new(),
1017        });
1018        empty_builder.process_event_at(&empty, fixed_ts());
1019        let empty_trajectory = empty_builder.finish(None);
1020        assert!(empty_trajectory.steps.last().expect("step").extra.is_none());
1021    }
1022
1023    #[test]
1024    fn step_metrics_from_usage() {
1025        let usage = Usage {
1026            input_tokens: 1000,
1027            cached_input_tokens: 200,
1028            cache_creation_tokens: 50,
1029            output_tokens: 300,
1030        };
1031        let metrics = StepMetrics::from_usage(&usage);
1032        assert_eq!(metrics.prompt_tokens, Some(1000));
1033        assert_eq!(metrics.completion_tokens, Some(300));
1034        assert_eq!(metrics.cached_tokens, Some(200));
1035        assert!(metrics.extra.is_some());
1036    }
1037
1038    #[test]
1039    fn builder_implements_event_emitter() {
1040        let mut builder = AtifTrajectoryBuilder::new(AtifAgent::vtcode());
1041        let event = ThreadEvent::ThreadStarted(ThreadStartedEvent { thread_id: "t-1".to_string() });
1042        // Use EventEmitter trait
1043        crate::EventEmitter::emit(&mut builder, &event);
1044        assert_eq!(builder.step_count(), 0); // ThreadStarted doesn't create a step
1045    }
1046
1047    #[test]
1048    fn skips_lifecycle_events() {
1049        let mut builder = AtifTrajectoryBuilder::new(AtifAgent::vtcode());
1050        builder.process_event(&ThreadEvent::TurnStarted(TurnStartedEvent::default()));
1051        assert_eq!(builder.step_count(), 0);
1052    }
1053
1054    #[test]
1055    fn compact_boundary_with_segment_metadata_preserves_atif_export() {
1056        let mut builder = AtifTrajectoryBuilder::new(AtifAgent::vtcode());
1057        let event = ThreadEvent::ThreadCompactBoundary(Box::new(ThreadCompactBoundaryEvent {
1058            thread_id: "thread-1".to_string(),
1059            trigger: CompactionTrigger::Auto,
1060            mode: CompactionMode::Local,
1061            original_message_count: 12,
1062            compacted_message_count: 5,
1063            history_artifact_path: None,
1064            previous_segment_id: Some("segment-0001".to_string()),
1065            new_segment_id: Some("segment-0002".to_string()),
1066            previous_prefix_hash: Some("prefix-before".to_string()),
1067            new_prefix_hash: Some("prefix-after".to_string()),
1068            previous_catalog_hash: Some("catalog-before".to_string()),
1069            new_catalog_hash: Some("catalog-after".to_string()),
1070        }));
1071
1072        builder.process_event_at(&event, fixed_ts());
1073        let trajectory = builder.finish(None);
1074
1075        assert_eq!(trajectory.steps.len(), 1);
1076        assert_eq!(trajectory.steps[0].source, StepSource::System);
1077        assert_eq!(
1078            trajectory.steps[0].message.as_deref(),
1079            Some("context_compaction: 12 messages -> 5 messages (auto)")
1080        );
1081    }
1082}
vtcode_exec_events/atif.rs

vtcode_exec_events/
atif.rs
