1use std::time::Instant;
2
3use serde::Serialize;
4
5use super::api::{
6 vm_call_llm_full_streaming, LlmApiMode, LlmCallOptions, LlmRoutePolicy, OutputFormat,
7 ThinkingConfig,
8};
9use crate::value::{VmError, VmValue};
10
11const SMOKE_TEST_MAX_TOKENS: i64 = 32;
12
13#[derive(Clone, Debug, PartialEq, Eq)]
14pub struct ModelSmokeTestOptions {
15 pub model: String,
16 pub provider: Option<String>,
17 pub prompt: String,
18}
19
20#[derive(Clone, Debug, PartialEq, Serialize)]
21pub struct ModelSmokeTestResult {
22 pub model_id: String,
23 pub provider: String,
24 pub latency_ms: u64,
25 #[serde(skip_serializing_if = "Option::is_none")]
26 pub first_token_ms: Option<u64>,
27 pub input_tokens: i64,
28 pub output_tokens: i64,
29 pub estimated_cost_usd: f64,
30}
31
32pub async fn run_model_smoke_test(
33 options: ModelSmokeTestOptions,
34) -> Result<ModelSmokeTestResult, String> {
35 super::provider::register_default_providers();
36
37 let resolved = crate::llm_config::resolve_model_info(&options.model);
38 let model_id = resolved.id;
39 let provider = options
40 .provider
41 .as_deref()
42 .map(str::trim)
43 .filter(|provider| !provider.is_empty())
44 .map(str::to_string)
45 .unwrap_or(resolved.provider);
46 let api_key = super::helpers::resolve_api_key(&provider).map_err(vm_error_message)?;
47
48 if let Some(def) = crate::llm_config::provider_config(&provider) {
49 if super::supports_model_readiness_probe(&def) {
50 let readiness =
51 super::probe_openai_compatible_model(&provider, &model_id, &api_key).await;
52 if readiness.category == "model_missing" || readiness.category == "invalid_url" {
53 return Err(readiness.message);
54 }
55 }
56 }
57
58 let opts = LlmCallOptions {
59 provider: provider.clone(),
60 model: model_id.clone(),
61 api_key,
62 api_mode: LlmApiMode::ChatCompletions,
63 route_policy: LlmRoutePolicy::Manual,
64 fallback_chain: Vec::new(),
65 route_fallbacks: Vec::new(),
66 routing_decision: None,
67 routing_policy: None,
68 session_id: None,
69 reminders: None,
70 reminder_lifecycle: Vec::new(),
71 messages: vec![serde_json::json!({
72 "role": "user",
73 "content": options.prompt,
74 })],
75 system: None,
76 transcript_summary: None,
77 max_tokens: SMOKE_TEST_MAX_TOKENS,
78 temperature: None,
79 top_p: None,
80 top_k: None,
81 logprobs: false,
82 top_logprobs: None,
83 stop: None,
84 seed: None,
85 frequency_penalty: None,
86 presence_penalty: None,
87 fast: false,
88 output_format: OutputFormat::Text,
89 response_format: None,
90 json_schema: None,
91 output_schema: None,
92 output_validation: None,
93 schema_stream_abort: false,
94 thinking: ThinkingConfig::Disabled,
95 anthropic_beta_features: Vec::new(),
96 vision: false,
97 tools: None,
98 native_tools: None,
99 provider_tools: Vec::new(),
100 tool_choice: None,
101 tool_search: None,
102 cache: false,
103 timeout: None,
104 idle_timeout: None,
105 stream: true,
106 provider_overrides: None,
107 previous_response_id: None,
108 store: None,
109 background: None,
110 truncation: None,
111 compact: None,
112 include: None,
113 max_tool_calls: None,
114 budget: None,
115 prefill: None,
116 structural_experiment: None,
117 applied_structural_experiment: None,
118 };
119
120 let (delta_tx, mut delta_rx) = tokio::sync::mpsc::unbounded_channel::<String>();
121 let started = Instant::now();
122 let first_delta = tokio::spawn(async move { delta_rx.recv().await.map(|_| started.elapsed()) });
123 let result = vm_call_llm_full_streaming(&opts, delta_tx)
124 .await
125 .map_err(vm_error_message);
126 let latency_ms = duration_ms(started.elapsed());
127 let first_token_ms = first_delta.await.ok().flatten().map(duration_ms);
128 let result = result?;
129
130 Ok(ModelSmokeTestResult {
131 model_id: result.model.clone(),
132 provider: result.provider.clone(),
133 latency_ms,
134 first_token_ms,
135 input_tokens: result.input_tokens,
136 output_tokens: result.output_tokens,
137 estimated_cost_usd: super::calculate_cost_for_provider(
138 &result.provider,
139 &result.model,
140 result.input_tokens,
141 result.output_tokens,
142 ),
143 })
144}
145
146fn duration_ms(duration: std::time::Duration) -> u64 {
147 u64::try_from(duration.as_millis()).unwrap_or(u64::MAX)
148}
149
150fn vm_error_message(error: VmError) -> String {
151 match error {
152 VmError::CategorizedError { message, .. } => message,
153 VmError::Thrown(VmValue::String(message)) => message.to_string(),
154 VmError::Thrown(VmValue::Dict(dict)) => dict
155 .get("message")
156 .map(VmValue::display)
157 .unwrap_or_else(|| VmError::Thrown(VmValue::Dict(dict)).to_string()),
158 other => other.to_string(),
159 }
160}
161
162#[cfg(test)]
163mod tests {
164 use super::{run_model_smoke_test, ModelSmokeTestOptions};
165
166 #[tokio::test]
167 async fn mock_provider_smoke_test_reports_timing_tokens_and_cost() {
168 crate::llm::reset_llm_state();
169 let result = run_model_smoke_test(ModelSmokeTestOptions {
170 model: "mock".to_string(),
171 provider: Some("mock".to_string()),
172 prompt: "ping".to_string(),
173 })
174 .await
175 .expect("mock provider smoke test should not require network");
176
177 assert_eq!(result.model_id, "mock");
178 assert_eq!(result.provider, "mock");
179 assert_eq!(result.input_tokens, 4);
180 assert_eq!(result.output_tokens, 30);
181 assert_eq!(result.estimated_cost_usd, 0.0);
182 assert!(result.first_token_ms.is_some());
183 }
184}