1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
//! Evaluation runner that orchestrates the full eval pipeline.
//!
//! The runner creates agents via an [`AgentFactory`], captures their execution
//! trajectories, scores them with an [`EvaluatorRegistry`], and aggregates results.
use tokio_util::sync::CancellationToken;
use tracing::{info, warn};
use swink_agent::{
Agent, AssistantMessage, ContentBlock, Cost, ModelSpec, StopReason, Usage, UserMessage,
};
use crate::error::EvalError;
use crate::evaluator::EvaluatorRegistry;
use crate::score::{Score, Verdict};
use crate::trajectory::TrajectoryCollector;
use crate::types::{
EvalCase, EvalCaseResult, EvalMetricResult, EvalSet, EvalSetResult, EvalSummary, Invocation,
TurnRecord,
};
/// Factory that creates a configured [`Agent`] for each eval case.
///
/// Decouples the runner from agent construction so the caller controls
/// model selection, tools, and system prompt.
pub trait AgentFactory: Send + Sync {
/// Create an agent and cancellation token for the given eval case.
///
/// Implementers that honor `case.budget` must translate those constraints
/// via [`BudgetConstraints::to_policies`](crate::BudgetConstraints::to_policies)
/// and attach the resulting policies with
/// [`AgentOptions::with_pre_turn_policy`](swink_agent::AgentOptions::with_pre_turn_policy).
fn create_agent(&self, case: &EvalCase) -> Result<(Agent, CancellationToken), EvalError>;
}
/// Orchestrates evaluation: runs agents, captures trajectories, and scores results.
pub struct EvalRunner {
registry: EvaluatorRegistry,
}
impl EvalRunner {
/// Create a runner with a custom evaluator registry.
#[must_use]
pub const fn new(registry: EvaluatorRegistry) -> Self {
Self { registry }
}
/// Create a runner pre-loaded with built-in evaluators.
#[must_use]
pub fn with_defaults() -> Self {
Self::new(EvaluatorRegistry::with_defaults())
}
/// Run a single eval case and return the scored result.
pub async fn run_case(
&self,
case: &EvalCase,
factory: &dyn AgentFactory,
) -> Result<EvalCaseResult, EvalError> {
info!(case_id = %case.id, case_name = %case.name, "running eval case");
let (mut agent, _cancel) = factory.create_agent(case)?;
// Build user messages from the case.
let messages: Vec<_> = case
.user_messages
.iter()
.map(|text| {
swink_agent::AgentMessage::Llm(swink_agent::LlmMessage::User(UserMessage {
content: vec![ContentBlock::Text { text: text.clone() }],
timestamp: swink_agent::now_timestamp(),
cache_hint: None,
}))
})
.collect();
// Run the agent and collect the trajectory. Budget enforcement is
// attached by the factory via agent loop policies.
let stream = agent.prompt_stream(messages)?;
let invocation = TrajectoryCollector::collect_from_stream(stream).await;
// Score.
let metric_results = self.registry.evaluate(case, &invocation);
let verdict = if metric_results.iter().all(|r| r.score.verdict().is_pass()) {
Verdict::Pass
} else {
Verdict::Fail
};
info!(
case_id = %case.id,
verdict = ?verdict,
metrics = metric_results.len(),
"eval case complete"
);
Ok(EvalCaseResult {
case_id: case.id.clone(),
invocation,
metric_results,
verdict,
})
}
/// Run an entire eval set and return aggregated results.
pub async fn run_set(
&self,
eval_set: &EvalSet,
factory: &dyn AgentFactory,
) -> Result<EvalSetResult, EvalError> {
info!(
set_id = %eval_set.id,
cases = eval_set.cases.len(),
"running eval set"
);
let mut case_results = Vec::with_capacity(eval_set.cases.len());
let mut total_cost = Cost::default();
let mut total_usage = Usage::default();
let mut total_duration = std::time::Duration::ZERO;
let mut passed = 0usize;
let mut failed = 0usize;
for case in &eval_set.cases {
match self.run_case(case, factory).await {
Ok(result) => {
total_cost += result.invocation.total_cost.clone();
total_usage += result.invocation.total_usage.clone();
total_duration += result.invocation.total_duration;
if result.verdict.is_pass() {
passed += 1;
} else {
failed += 1;
}
case_results.push(result);
}
Err(e) => {
warn!(case_id = %case.id, error = %e, "eval case failed with error — recording failure and continuing");
failed += 1;
case_results.push(EvalCaseResult {
case_id: case.id.clone(),
invocation: Invocation {
turns: vec![TurnRecord {
turn_index: 0,
assistant_message: AssistantMessage {
content: vec![],
provider: String::new(),
model_id: String::new(),
usage: Usage::default(),
cost: Cost::default(),
stop_reason: StopReason::Error,
error_message: Some(e.to_string()),
error_kind: None,
timestamp: swink_agent::now_timestamp(),
cache_hint: None,
},
tool_calls: vec![],
tool_results: vec![],
duration: std::time::Duration::ZERO,
}],
total_usage: Usage::default(),
total_cost: Cost::default(),
total_duration: std::time::Duration::ZERO,
final_response: None,
stop_reason: StopReason::Error,
model: ModelSpec::new("unknown", "unknown"),
},
metric_results: vec![EvalMetricResult {
evaluator_name: "error".to_string(),
score: Score::fail(),
details: Some(e.to_string()),
}],
verdict: Verdict::Fail,
});
}
}
}
let summary = EvalSummary {
total_cases: eval_set.cases.len(),
passed,
failed,
total_cost,
total_usage,
total_duration,
};
info!(
set_id = %eval_set.id,
passed = summary.passed,
failed = summary.failed,
total = summary.total_cases,
"eval set complete"
);
Ok(EvalSetResult {
eval_set_id: eval_set.id.clone(),
case_results,
summary,
timestamp: swink_agent::now_timestamp(),
})
}
}