stakpak_shared/models/integrations/
openai.rs

1//! OpenAI provider configuration and chat message types
2//!
3//! This module contains:
4//! - Configuration types for OpenAI provider
5//! - OpenAI model enums with pricing info
6//! - Chat message types used throughout the TUI
7//! - Tool call types for agent interactions
8//!
9//! Note: Low-level API request/response types are in `libs/ai/src/providers/openai/`.
10
11use crate::models::llm::{
12    GenerationDelta, LLMMessage, LLMMessageContent, LLMMessageImageSource, LLMMessageTypedContent,
13    LLMTokenUsage, LLMTool,
14};
15use crate::models::model_pricing::{ContextAware, ContextPricingTier, ModelContextInfo};
16use serde::{Deserialize, Serialize};
17use serde_json::{Value, json};
18use uuid::Uuid;
19
20// =============================================================================
21// Provider Configuration
22// =============================================================================
23
24/// Configuration for OpenAI provider
25#[derive(Serialize, Deserialize, Clone, Debug, Default, PartialEq)]
26pub struct OpenAIConfig {
27    pub api_endpoint: Option<String>,
28    pub api_key: Option<String>,
29}
30
31impl OpenAIConfig {
32    /// Create config with API key
33    pub fn with_api_key(api_key: impl Into<String>) -> Self {
34        Self {
35            api_key: Some(api_key.into()),
36            api_endpoint: None,
37        }
38    }
39
40    /// Create config from ProviderAuth (only supports API key for OpenAI)
41    pub fn from_provider_auth(auth: &crate::models::auth::ProviderAuth) -> Option<Self> {
42        match auth {
43            crate::models::auth::ProviderAuth::Api { key } => Some(Self::with_api_key(key)),
44            crate::models::auth::ProviderAuth::OAuth { .. } => None, // OpenAI doesn't support OAuth
45        }
46    }
47
48    /// Merge with credentials from ProviderAuth, preserving existing endpoint
49    pub fn with_provider_auth(mut self, auth: &crate::models::auth::ProviderAuth) -> Option<Self> {
50        match auth {
51            crate::models::auth::ProviderAuth::Api { key } => {
52                self.api_key = Some(key.clone());
53                Some(self)
54            }
55            crate::models::auth::ProviderAuth::OAuth { .. } => None, // OpenAI doesn't support OAuth
56        }
57    }
58}
59
60// =============================================================================
61// Model Definitions
62// =============================================================================
63
64/// OpenAI model identifiers
65#[derive(Serialize, Deserialize, Debug, Clone, PartialEq, Default)]
66pub enum OpenAIModel {
67    // Reasoning Models
68    #[serde(rename = "o3-2025-04-16")]
69    O3,
70    #[serde(rename = "o4-mini-2025-04-16")]
71    O4Mini,
72
73    #[default]
74    #[serde(rename = "gpt-5-2025-08-07")]
75    GPT5,
76    #[serde(rename = "gpt-5.1-2025-11-13")]
77    GPT51,
78    #[serde(rename = "gpt-5-mini-2025-08-07")]
79    GPT5Mini,
80    #[serde(rename = "gpt-5-nano-2025-08-07")]
81    GPT5Nano,
82
83    Custom(String),
84}
85
86impl OpenAIModel {
87    pub fn from_string(s: &str) -> Result<Self, String> {
88        serde_json::from_value(serde_json::Value::String(s.to_string()))
89            .map_err(|_| "Failed to deserialize OpenAI model".to_string())
90    }
91
92    /// Default smart model for OpenAI
93    pub const DEFAULT_SMART_MODEL: OpenAIModel = OpenAIModel::GPT5;
94
95    /// Default eco model for OpenAI
96    pub const DEFAULT_ECO_MODEL: OpenAIModel = OpenAIModel::GPT5Mini;
97
98    /// Default recovery model for OpenAI
99    pub const DEFAULT_RECOVERY_MODEL: OpenAIModel = OpenAIModel::GPT5Mini;
100
101    /// Get default smart model as string
102    pub fn default_smart_model() -> String {
103        Self::DEFAULT_SMART_MODEL.to_string()
104    }
105
106    /// Get default eco model as string
107    pub fn default_eco_model() -> String {
108        Self::DEFAULT_ECO_MODEL.to_string()
109    }
110
111    /// Get default recovery model as string
112    pub fn default_recovery_model() -> String {
113        Self::DEFAULT_RECOVERY_MODEL.to_string()
114    }
115}
116
117impl ContextAware for OpenAIModel {
118    fn context_info(&self) -> ModelContextInfo {
119        let model_name = self.to_string();
120
121        if model_name.starts_with("o3") {
122            return ModelContextInfo {
123                max_tokens: 200_000,
124                pricing_tiers: vec![ContextPricingTier {
125                    label: "Standard".to_string(),
126                    input_cost_per_million: 2.0,
127                    output_cost_per_million: 8.0,
128                    upper_bound: None,
129                }],
130                approach_warning_threshold: 0.8,
131            };
132        }
133
134        if model_name.starts_with("o4-mini") {
135            return ModelContextInfo {
136                max_tokens: 200_000,
137                pricing_tiers: vec![ContextPricingTier {
138                    label: "Standard".to_string(),
139                    input_cost_per_million: 1.10,
140                    output_cost_per_million: 4.40,
141                    upper_bound: None,
142                }],
143                approach_warning_threshold: 0.8,
144            };
145        }
146
147        if model_name.starts_with("gpt-5-mini") {
148            return ModelContextInfo {
149                max_tokens: 400_000,
150                pricing_tiers: vec![ContextPricingTier {
151                    label: "Standard".to_string(),
152                    input_cost_per_million: 0.25,
153                    output_cost_per_million: 2.0,
154                    upper_bound: None,
155                }],
156                approach_warning_threshold: 0.8,
157            };
158        }
159
160        if model_name.starts_with("gpt-5-nano") {
161            return ModelContextInfo {
162                max_tokens: 400_000,
163                pricing_tiers: vec![ContextPricingTier {
164                    label: "Standard".to_string(),
165                    input_cost_per_million: 0.05,
166                    output_cost_per_million: 0.40,
167                    upper_bound: None,
168                }],
169                approach_warning_threshold: 0.8,
170            };
171        }
172
173        if model_name.starts_with("gpt-5") {
174            return ModelContextInfo {
175                max_tokens: 400_000,
176                pricing_tiers: vec![ContextPricingTier {
177                    label: "Standard".to_string(),
178                    input_cost_per_million: 1.25,
179                    output_cost_per_million: 10.0,
180                    upper_bound: None,
181                }],
182                approach_warning_threshold: 0.8,
183            };
184        }
185
186        ModelContextInfo::default()
187    }
188
189    fn model_name(&self) -> String {
190        match self {
191            OpenAIModel::O3 => "O3".to_string(),
192            OpenAIModel::O4Mini => "O4-mini".to_string(),
193            OpenAIModel::GPT5 => "GPT-5".to_string(),
194            OpenAIModel::GPT51 => "GPT-5.1".to_string(),
195            OpenAIModel::GPT5Mini => "GPT-5 Mini".to_string(),
196            OpenAIModel::GPT5Nano => "GPT-5 Nano".to_string(),
197            OpenAIModel::Custom(name) => format!("Custom ({})", name),
198        }
199    }
200}
201
202impl std::fmt::Display for OpenAIModel {
203    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
204        match self {
205            OpenAIModel::O3 => write!(f, "o3-2025-04-16"),
206            OpenAIModel::O4Mini => write!(f, "o4-mini-2025-04-16"),
207            OpenAIModel::GPT5Nano => write!(f, "gpt-5-nano-2025-08-07"),
208            OpenAIModel::GPT5Mini => write!(f, "gpt-5-mini-2025-08-07"),
209            OpenAIModel::GPT5 => write!(f, "gpt-5-2025-08-07"),
210            OpenAIModel::GPT51 => write!(f, "gpt-5.1-2025-11-13"),
211            OpenAIModel::Custom(model_name) => write!(f, "{}", model_name),
212        }
213    }
214}
215
216// =============================================================================
217// Message Types (used by TUI)
218// =============================================================================
219
220/// Message role
221#[derive(Debug, Serialize, Deserialize, Clone, PartialEq, Default)]
222#[serde(rename_all = "lowercase")]
223pub enum Role {
224    System,
225    Developer,
226    User,
227    #[default]
228    Assistant,
229    Tool,
230}
231
232impl std::fmt::Display for Role {
233    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
234        match self {
235            Role::System => write!(f, "system"),
236            Role::Developer => write!(f, "developer"),
237            Role::User => write!(f, "user"),
238            Role::Assistant => write!(f, "assistant"),
239            Role::Tool => write!(f, "tool"),
240        }
241    }
242}
243
244/// Model info for tracking which model generated a message
245#[derive(Debug, Serialize, Deserialize, Clone, PartialEq)]
246pub struct ModelInfo {
247    /// Provider name (e.g., "anthropic", "openai")
248    pub provider: String,
249    /// Model identifier (e.g., "claude-sonnet-4-20250514", "gpt-4")
250    pub id: String,
251}
252
253/// Chat message
254#[derive(Debug, Serialize, Deserialize, Clone, PartialEq, Default)]
255pub struct ChatMessage {
256    pub role: Role,
257    pub content: Option<MessageContent>,
258    #[serde(skip_serializing_if = "Option::is_none")]
259    pub name: Option<String>,
260    #[serde(skip_serializing_if = "Option::is_none")]
261    pub tool_calls: Option<Vec<ToolCall>>,
262    #[serde(skip_serializing_if = "Option::is_none")]
263    pub tool_call_id: Option<String>,
264    #[serde(skip_serializing_if = "Option::is_none")]
265    pub usage: Option<LLMTokenUsage>,
266
267    // === Extended fields for session tracking ===
268    /// Unique message identifier
269    #[serde(skip_serializing_if = "Option::is_none")]
270    pub id: Option<String>,
271    /// Model that generated this message (for assistant messages)
272    #[serde(skip_serializing_if = "Option::is_none")]
273    pub model: Option<ModelInfo>,
274    /// Cost in dollars for this message
275    #[serde(skip_serializing_if = "Option::is_none")]
276    pub cost: Option<f64>,
277    /// Why the model stopped: "stop", "tool_calls", "length", "error"
278    #[serde(skip_serializing_if = "Option::is_none")]
279    pub finish_reason: Option<String>,
280    /// Unix timestamp (ms) when message was created/sent
281    #[serde(skip_serializing_if = "Option::is_none")]
282    pub created_at: Option<i64>,
283    /// Unix timestamp (ms) when assistant finished generating
284    #[serde(skip_serializing_if = "Option::is_none")]
285    pub completed_at: Option<i64>,
286    /// Plugin extensibility - unstructured metadata
287    #[serde(skip_serializing_if = "Option::is_none")]
288    pub metadata: Option<serde_json::Value>,
289}
290
291impl ChatMessage {
292    pub fn last_server_message(messages: &[ChatMessage]) -> Option<&ChatMessage> {
293        messages
294            .iter()
295            .rev()
296            .find(|message| message.role != Role::User && message.role != Role::Tool)
297    }
298
299    pub fn to_xml(&self) -> String {
300        match &self.content {
301            Some(MessageContent::String(s)) => {
302                format!("<message role=\"{}\">{}</message>", self.role, s)
303            }
304            Some(MessageContent::Array(parts)) => parts
305                .iter()
306                .map(|part| {
307                    format!(
308                        "<message role=\"{}\" type=\"{}\">{}</message>",
309                        self.role,
310                        part.r#type,
311                        part.text.clone().unwrap_or_default()
312                    )
313                })
314                .collect::<Vec<String>>()
315                .join("\n"),
316            None => String::new(),
317        }
318    }
319}
320
321/// Message content (string or array of parts)
322#[derive(Debug, Serialize, Deserialize, Clone, PartialEq)]
323#[serde(untagged)]
324pub enum MessageContent {
325    String(String),
326    Array(Vec<ContentPart>),
327}
328
329impl MessageContent {
330    pub fn inject_checkpoint_id(&self, checkpoint_id: Uuid) -> Self {
331        match self {
332            MessageContent::String(s) => MessageContent::String(format!(
333                "<checkpoint_id>{checkpoint_id}</checkpoint_id>\n{s}"
334            )),
335            MessageContent::Array(parts) => MessageContent::Array(
336                std::iter::once(ContentPart {
337                    r#type: "text".to_string(),
338                    text: Some(format!("<checkpoint_id>{checkpoint_id}</checkpoint_id>")),
339                    image_url: None,
340                })
341                .chain(parts.iter().cloned())
342                .collect(),
343            ),
344        }
345    }
346
347    pub fn extract_checkpoint_id(&self) -> Option<Uuid> {
348        match self {
349            MessageContent::String(s) => s
350                .rfind("<checkpoint_id>")
351                .and_then(|start| {
352                    s[start..]
353                        .find("</checkpoint_id>")
354                        .map(|end| (start + "<checkpoint_id>".len(), start + end))
355                })
356                .and_then(|(start, end)| Uuid::parse_str(&s[start..end]).ok()),
357            MessageContent::Array(parts) => parts.iter().rev().find_map(|part| {
358                part.text.as_deref().and_then(|text| {
359                    text.rfind("<checkpoint_id>")
360                        .and_then(|start| {
361                            text[start..]
362                                .find("</checkpoint_id>")
363                                .map(|end| (start + "<checkpoint_id>".len(), start + end))
364                        })
365                        .and_then(|(start, end)| Uuid::parse_str(&text[start..end]).ok())
366                })
367            }),
368        }
369    }
370}
371
372impl std::fmt::Display for MessageContent {
373    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
374        match self {
375            MessageContent::String(s) => write!(f, "{s}"),
376            MessageContent::Array(parts) => {
377                let text_parts: Vec<String> =
378                    parts.iter().filter_map(|part| part.text.clone()).collect();
379                write!(f, "{}", text_parts.join("\n"))
380            }
381        }
382    }
383}
384
385impl Default for MessageContent {
386    fn default() -> Self {
387        MessageContent::String(String::new())
388    }
389}
390
391/// Content part (text or image)
392#[derive(Debug, Serialize, Deserialize, Clone, PartialEq)]
393pub struct ContentPart {
394    pub r#type: String,
395    #[serde(skip_serializing_if = "Option::is_none")]
396    pub text: Option<String>,
397    #[serde(skip_serializing_if = "Option::is_none")]
398    pub image_url: Option<ImageUrl>,
399}
400
401/// Image URL with optional detail level
402#[derive(Debug, Serialize, Deserialize, Clone, PartialEq)]
403pub struct ImageUrl {
404    pub url: String,
405    #[serde(skip_serializing_if = "Option::is_none")]
406    pub detail: Option<String>,
407}
408
409// =============================================================================
410// Tool Types (used by TUI)
411// =============================================================================
412
413/// Tool definition
414#[derive(Debug, Serialize, Deserialize, Clone, PartialEq)]
415pub struct Tool {
416    pub r#type: String,
417    pub function: FunctionDefinition,
418}
419
420/// Function definition for tools
421#[derive(Debug, Serialize, Deserialize, Clone, PartialEq)]
422pub struct FunctionDefinition {
423    pub name: String,
424    pub description: Option<String>,
425    pub parameters: serde_json::Value,
426}
427
428impl From<Tool> for LLMTool {
429    fn from(tool: Tool) -> Self {
430        LLMTool {
431            name: tool.function.name,
432            description: tool.function.description.unwrap_or_default(),
433            input_schema: tool.function.parameters,
434        }
435    }
436}
437
438/// Tool choice configuration
439#[derive(Debug, Clone, PartialEq)]
440pub enum ToolChoice {
441    Auto,
442    Required,
443    Object(ToolChoiceObject),
444}
445
446impl Serialize for ToolChoice {
447    fn serialize<S>(&self, serializer: S) -> Result<S::Ok, S::Error>
448    where
449        S: serde::Serializer,
450    {
451        match self {
452            ToolChoice::Auto => serializer.serialize_str("auto"),
453            ToolChoice::Required => serializer.serialize_str("required"),
454            ToolChoice::Object(obj) => obj.serialize(serializer),
455        }
456    }
457}
458
459impl<'de> Deserialize<'de> for ToolChoice {
460    fn deserialize<D>(deserializer: D) -> Result<Self, D::Error>
461    where
462        D: serde::Deserializer<'de>,
463    {
464        struct ToolChoiceVisitor;
465
466        impl<'de> serde::de::Visitor<'de> for ToolChoiceVisitor {
467            type Value = ToolChoice;
468
469            fn expecting(&self, formatter: &mut std::fmt::Formatter) -> std::fmt::Result {
470                formatter.write_str("string or object")
471            }
472
473            fn visit_str<E>(self, value: &str) -> Result<ToolChoice, E>
474            where
475                E: serde::de::Error,
476            {
477                match value {
478                    "auto" => Ok(ToolChoice::Auto),
479                    "required" => Ok(ToolChoice::Required),
480                    _ => Err(serde::de::Error::unknown_variant(
481                        value,
482                        &["auto", "required"],
483                    )),
484                }
485            }
486
487            fn visit_map<M>(self, map: M) -> Result<ToolChoice, M::Error>
488            where
489                M: serde::de::MapAccess<'de>,
490            {
491                let obj = ToolChoiceObject::deserialize(
492                    serde::de::value::MapAccessDeserializer::new(map),
493                )?;
494                Ok(ToolChoice::Object(obj))
495            }
496        }
497
498        deserializer.deserialize_any(ToolChoiceVisitor)
499    }
500}
501
502#[derive(Debug, Serialize, Deserialize, Clone, PartialEq)]
503pub struct ToolChoiceObject {
504    pub r#type: String,
505    pub function: FunctionChoice,
506}
507
508#[derive(Debug, Serialize, Deserialize, Clone, PartialEq)]
509pub struct FunctionChoice {
510    pub name: String,
511}
512
513/// Tool call from assistant
514#[derive(Debug, Serialize, Deserialize, Clone, PartialEq)]
515pub struct ToolCall {
516    pub id: String,
517    pub r#type: String,
518    pub function: FunctionCall,
519    /// Opaque provider-specific metadata (e.g., Gemini thought_signature)
520    #[serde(skip_serializing_if = "Option::is_none")]
521    pub metadata: Option<serde_json::Value>,
522}
523
524/// Function call details
525#[derive(Debug, Serialize, Deserialize, Clone, PartialEq)]
526pub struct FunctionCall {
527    pub name: String,
528    pub arguments: String,
529}
530
531/// Tool call result status
532#[derive(Debug, Serialize, Deserialize, Clone, PartialEq)]
533pub enum ToolCallResultStatus {
534    Success,
535    Error,
536    Cancelled,
537}
538
539/// Tool call result
540#[derive(Debug, Serialize, Deserialize, Clone, PartialEq)]
541pub struct ToolCallResult {
542    pub call: ToolCall,
543    pub result: String,
544    pub status: ToolCallResultStatus,
545}
546
547/// Streaming progress info for a tool call being generated by the LLM
548#[derive(Debug, Clone, Serialize, Deserialize, PartialEq)]
549pub struct ToolCallStreamInfo {
550    /// Tool name (may be empty if not yet streamed)
551    pub name: String,
552    /// Estimated token count of arguments streamed so far (~chars/4)
553    pub args_tokens: usize,
554    /// Optional description extracted from arguments (best-effort, may be None if JSON incomplete)
555    #[serde(skip_serializing_if = "Option::is_none")]
556    pub description: Option<String>,
557}
558
559/// Tool call result progress update
560#[derive(Debug, Clone, Serialize, Deserialize, PartialEq)]
561pub struct ToolCallResultProgress {
562    pub id: Uuid,
563    pub message: String,
564    /// Type of progress update for specialized handling
565    #[serde(skip_serializing_if = "Option::is_none")]
566    pub progress_type: Option<ProgressType>,
567    /// Structured task updates for task wait progress
568    #[serde(skip_serializing_if = "Option::is_none")]
569    pub task_updates: Option<Vec<TaskUpdate>>,
570    /// Overall progress percentage (0.0 - 100.0)
571    #[serde(skip_serializing_if = "Option::is_none")]
572    pub progress: Option<f64>,
573}
574
575/// Type of progress update
576#[derive(Debug, Clone, Serialize, Deserialize, PartialEq)]
577pub enum ProgressType {
578    /// Command output streaming
579    CommandOutput,
580    /// Task wait progress with structured updates
581    TaskWait,
582    /// Generic progress
583    Generic,
584}
585
586/// Structured task status update
587#[derive(Debug, Clone, Serialize, Deserialize, PartialEq)]
588pub struct TaskUpdate {
589    pub task_id: String,
590    pub status: String,
591    #[serde(skip_serializing_if = "Option::is_none")]
592    pub description: Option<String>,
593    #[serde(skip_serializing_if = "Option::is_none")]
594    pub duration_secs: Option<f64>,
595    #[serde(skip_serializing_if = "Option::is_none")]
596    pub output_preview: Option<String>,
597    /// Whether this is a target task being waited on
598    #[serde(default)]
599    pub is_target: bool,
600    /// Pause information for paused subagent tasks
601    #[serde(skip_serializing_if = "Option::is_none")]
602    pub pause_info: Option<TaskPauseInfo>,
603}
604
605/// Pause information for subagent tasks awaiting approval
606#[derive(Debug, Clone, Serialize, Deserialize, PartialEq)]
607pub struct TaskPauseInfo {
608    /// The agent's message before pausing
609    #[serde(skip_serializing_if = "Option::is_none")]
610    pub agent_message: Option<String>,
611    /// Pending tool calls awaiting approval
612    #[serde(skip_serializing_if = "Option::is_none")]
613    pub pending_tool_calls: Option<Vec<crate::models::async_manifest::PendingToolCall>>,
614}
615
616// =============================================================================
617// Chat Completion Types (used by TUI)
618// =============================================================================
619
620/// Chat completion request
621#[derive(Debug, Serialize, Deserialize, Clone, PartialEq)]
622pub struct ChatCompletionRequest {
623    pub model: String,
624    pub messages: Vec<ChatMessage>,
625    #[serde(skip_serializing_if = "Option::is_none")]
626    pub frequency_penalty: Option<f32>,
627    #[serde(skip_serializing_if = "Option::is_none")]
628    pub logit_bias: Option<serde_json::Value>,
629    #[serde(skip_serializing_if = "Option::is_none")]
630    pub logprobs: Option<bool>,
631    #[serde(skip_serializing_if = "Option::is_none")]
632    pub max_tokens: Option<u32>,
633    #[serde(skip_serializing_if = "Option::is_none")]
634    pub n: Option<u32>,
635    #[serde(skip_serializing_if = "Option::is_none")]
636    pub presence_penalty: Option<f32>,
637    #[serde(skip_serializing_if = "Option::is_none")]
638    pub response_format: Option<ResponseFormat>,
639    #[serde(skip_serializing_if = "Option::is_none")]
640    pub seed: Option<i64>,
641    #[serde(skip_serializing_if = "Option::is_none")]
642    pub stop: Option<StopSequence>,
643    #[serde(skip_serializing_if = "Option::is_none")]
644    pub stream: Option<bool>,
645    #[serde(skip_serializing_if = "Option::is_none")]
646    pub temperature: Option<f32>,
647    #[serde(skip_serializing_if = "Option::is_none")]
648    pub top_p: Option<f32>,
649    #[serde(skip_serializing_if = "Option::is_none")]
650    pub tools: Option<Vec<Tool>>,
651    #[serde(skip_serializing_if = "Option::is_none")]
652    pub tool_choice: Option<ToolChoice>,
653    #[serde(skip_serializing_if = "Option::is_none")]
654    pub user: Option<String>,
655    #[serde(skip_serializing_if = "Option::is_none")]
656    pub context: Option<ChatCompletionContext>,
657}
658
659impl ChatCompletionRequest {
660    pub fn new(
661        model: String,
662        messages: Vec<ChatMessage>,
663        tools: Option<Vec<Tool>>,
664        stream: Option<bool>,
665    ) -> Self {
666        Self {
667            model,
668            messages,
669            frequency_penalty: None,
670            logit_bias: None,
671            logprobs: None,
672            max_tokens: None,
673            n: None,
674            presence_penalty: None,
675            response_format: None,
676            seed: None,
677            stop: None,
678            stream,
679            temperature: None,
680            top_p: None,
681            tools,
682            tool_choice: None,
683            user: None,
684            context: None,
685        }
686    }
687}
688
689#[derive(Debug, Serialize, Deserialize, Clone, PartialEq)]
690pub struct ChatCompletionContext {
691    pub scratchpad: Option<Value>,
692}
693
694#[derive(Debug, Serialize, Deserialize, Clone, PartialEq)]
695pub struct ResponseFormat {
696    pub r#type: String,
697}
698
699#[derive(Debug, Serialize, Deserialize, Clone, PartialEq)]
700#[serde(untagged)]
701pub enum StopSequence {
702    String(String),
703    Array(Vec<String>),
704}
705
706/// Chat completion response
707#[derive(Debug, Serialize, Deserialize, Clone, PartialEq)]
708pub struct ChatCompletionResponse {
709    pub id: String,
710    pub object: String,
711    pub created: u64,
712    pub model: String,
713    pub choices: Vec<ChatCompletionChoice>,
714    pub usage: LLMTokenUsage,
715    #[serde(skip_serializing_if = "Option::is_none")]
716    pub system_fingerprint: Option<String>,
717    pub metadata: Option<serde_json::Value>,
718}
719
720#[derive(Debug, Serialize, Deserialize, Clone, PartialEq)]
721pub struct ChatCompletionChoice {
722    pub index: usize,
723    pub message: ChatMessage,
724    pub logprobs: Option<LogProbs>,
725    pub finish_reason: FinishReason,
726}
727
728#[derive(Debug, Serialize, Deserialize, Clone, PartialEq)]
729#[serde(rename_all = "snake_case")]
730pub enum FinishReason {
731    Stop,
732    Length,
733    ContentFilter,
734    ToolCalls,
735}
736
737#[derive(Debug, Serialize, Deserialize, Clone, PartialEq)]
738pub struct LogProbs {
739    pub content: Option<Vec<LogProbContent>>,
740}
741
742#[derive(Debug, Serialize, Deserialize, Clone, PartialEq)]
743pub struct LogProbContent {
744    pub token: String,
745    pub logprob: f32,
746    pub bytes: Option<Vec<u8>>,
747    pub top_logprobs: Option<Vec<TokenLogprob>>,
748}
749
750#[derive(Debug, Serialize, Deserialize, Clone, PartialEq)]
751pub struct TokenLogprob {
752    pub token: String,
753    pub logprob: f32,
754    pub bytes: Option<Vec<u8>>,
755}
756
757// =============================================================================
758// Streaming Types
759// =============================================================================
760
761#[derive(Debug, Serialize, Deserialize, Clone, PartialEq)]
762pub struct ChatCompletionStreamResponse {
763    pub id: String,
764    pub object: String,
765    pub created: u64,
766    pub model: String,
767    pub choices: Vec<ChatCompletionStreamChoice>,
768    pub usage: Option<LLMTokenUsage>,
769    pub metadata: Option<serde_json::Value>,
770}
771
772#[derive(Debug, Serialize, Deserialize, Clone, PartialEq)]
773pub struct ChatCompletionStreamChoice {
774    pub index: usize,
775    pub delta: ChatMessageDelta,
776    pub finish_reason: Option<FinishReason>,
777}
778
779#[derive(Debug, Serialize, Deserialize, Clone, PartialEq)]
780pub struct ChatMessageDelta {
781    #[serde(skip_serializing_if = "Option::is_none")]
782    pub role: Option<Role>,
783    #[serde(skip_serializing_if = "Option::is_none")]
784    pub content: Option<String>,
785    #[serde(skip_serializing_if = "Option::is_none")]
786    pub tool_calls: Option<Vec<ToolCallDelta>>,
787}
788
789#[derive(Debug, Serialize, Deserialize, Clone, PartialEq)]
790pub struct ToolCallDelta {
791    pub index: usize,
792    pub id: Option<String>,
793    pub r#type: Option<String>,
794    pub function: Option<FunctionCallDelta>,
795    /// Opaque provider-specific metadata (e.g., Gemini thought_signature)
796    #[serde(skip_serializing_if = "Option::is_none")]
797    pub metadata: Option<serde_json::Value>,
798}
799
800#[derive(Debug, Serialize, Deserialize, Clone, PartialEq)]
801pub struct FunctionCallDelta {
802    pub name: Option<String>,
803    pub arguments: Option<String>,
804}
805
806// =============================================================================
807// Conversions
808// =============================================================================
809
810impl From<LLMMessage> for ChatMessage {
811    fn from(llm_message: LLMMessage) -> Self {
812        let role = match llm_message.role.as_str() {
813            "system" => Role::System,
814            "user" => Role::User,
815            "assistant" => Role::Assistant,
816            "tool" => Role::Tool,
817            "developer" => Role::Developer,
818            _ => Role::User,
819        };
820
821        let (content, tool_calls, tool_call_id) = match llm_message.content {
822            LLMMessageContent::String(text) => (Some(MessageContent::String(text)), None, None),
823            LLMMessageContent::List(items) => {
824                let mut text_parts = Vec::new();
825                let mut tool_call_parts = Vec::new();
826                let mut tool_result_id: Option<String> = None;
827
828                for item in items {
829                    match item {
830                        LLMMessageTypedContent::Text { text } => {
831                            text_parts.push(ContentPart {
832                                r#type: "text".to_string(),
833                                text: Some(text),
834                                image_url: None,
835                            });
836                        }
837                        LLMMessageTypedContent::ToolCall {
838                            id,
839                            name,
840                            args,
841                            metadata,
842                        } => {
843                            tool_call_parts.push(ToolCall {
844                                id,
845                                r#type: "function".to_string(),
846                                function: FunctionCall {
847                                    name,
848                                    arguments: args.to_string(),
849                                },
850                                metadata,
851                            });
852                        }
853                        LLMMessageTypedContent::ToolResult {
854                            tool_use_id,
855                            content,
856                        } => {
857                            if tool_result_id.is_none() {
858                                tool_result_id = Some(tool_use_id);
859                            }
860                            text_parts.push(ContentPart {
861                                r#type: "text".to_string(),
862                                text: Some(content),
863                                image_url: None,
864                            });
865                        }
866                        LLMMessageTypedContent::Image { source } => {
867                            text_parts.push(ContentPart {
868                                r#type: "image_url".to_string(),
869                                text: None,
870                                image_url: Some(ImageUrl {
871                                    url: format!(
872                                        "data:{};base64,{}",
873                                        source.media_type, source.data
874                                    ),
875                                    detail: None,
876                                }),
877                            });
878                        }
879                    }
880                }
881
882                let content = if !text_parts.is_empty() {
883                    Some(MessageContent::Array(text_parts))
884                } else {
885                    None
886                };
887
888                let tool_calls = if !tool_call_parts.is_empty() {
889                    Some(tool_call_parts)
890                } else {
891                    None
892                };
893
894                (content, tool_calls, tool_result_id)
895            }
896        };
897
898        ChatMessage {
899            role,
900            content,
901            name: None,
902            tool_calls,
903            tool_call_id,
904            usage: None,
905            ..Default::default()
906        }
907    }
908}
909
910impl From<ChatMessage> for LLMMessage {
911    fn from(chat_message: ChatMessage) -> Self {
912        let mut content_parts = Vec::new();
913
914        match chat_message.content {
915            Some(MessageContent::String(s)) => {
916                if !s.is_empty() {
917                    content_parts.push(LLMMessageTypedContent::Text { text: s });
918                }
919            }
920            Some(MessageContent::Array(parts)) => {
921                for part in parts {
922                    if let Some(text) = part.text {
923                        content_parts.push(LLMMessageTypedContent::Text { text });
924                    } else if let Some(image_url) = part.image_url {
925                        let (media_type, data) = if image_url.url.starts_with("data:") {
926                            let parts: Vec<&str> = image_url.url.splitn(2, ',').collect();
927                            if parts.len() == 2 {
928                                let meta = parts[0];
929                                let data = parts[1];
930                                let media_type = meta
931                                    .trim_start_matches("data:")
932                                    .trim_end_matches(";base64")
933                                    .to_string();
934                                (media_type, data.to_string())
935                            } else {
936                                ("image/jpeg".to_string(), image_url.url)
937                            }
938                        } else {
939                            ("image/jpeg".to_string(), image_url.url)
940                        };
941
942                        content_parts.push(LLMMessageTypedContent::Image {
943                            source: LLMMessageImageSource {
944                                r#type: "base64".to_string(),
945                                media_type,
946                                data,
947                            },
948                        });
949                    }
950                }
951            }
952            None => {}
953        }
954
955        if let Some(tool_calls) = chat_message.tool_calls {
956            for tool_call in tool_calls {
957                let args = serde_json::from_str(&tool_call.function.arguments).unwrap_or(json!({}));
958                content_parts.push(LLMMessageTypedContent::ToolCall {
959                    id: tool_call.id,
960                    name: tool_call.function.name,
961                    args,
962                    metadata: tool_call.metadata,
963                });
964            }
965        }
966
967        // Handle tool result messages: when role is Tool and tool_call_id is present,
968        // convert the content to a ToolResult content part. This is the generic
969        // intermediate representation - each provider's conversion layer handles
970        // the specifics (e.g., Anthropic converts to user role with tool_result blocks)
971        if chat_message.role == Role::Tool
972            && let Some(tool_call_id) = chat_message.tool_call_id
973        {
974            // Extract content as string for the tool result
975            let content_str = content_parts
976                .iter()
977                .filter_map(|p| match p {
978                    LLMMessageTypedContent::Text { text } => Some(text.clone()),
979                    _ => None,
980                })
981                .collect::<Vec<_>>()
982                .join("\n");
983
984            // Replace content with a single ToolResult
985            content_parts = vec![LLMMessageTypedContent::ToolResult {
986                tool_use_id: tool_call_id,
987                content: content_str,
988            }];
989        }
990
991        LLMMessage {
992            role: chat_message.role.to_string(),
993            content: if content_parts.is_empty() {
994                LLMMessageContent::String(String::new())
995            } else if content_parts.len() == 1 {
996                match &content_parts[0] {
997                    LLMMessageTypedContent::Text { text } => {
998                        LLMMessageContent::String(text.clone())
999                    }
1000                    _ => LLMMessageContent::List(content_parts),
1001                }
1002            } else {
1003                LLMMessageContent::List(content_parts)
1004            },
1005        }
1006    }
1007}
1008
1009impl From<GenerationDelta> for ChatMessageDelta {
1010    fn from(delta: GenerationDelta) -> Self {
1011        match delta {
1012            GenerationDelta::Content { content } => ChatMessageDelta {
1013                role: Some(Role::Assistant),
1014                content: Some(content),
1015                tool_calls: None,
1016            },
1017            GenerationDelta::Thinking { thinking: _ } => ChatMessageDelta {
1018                role: Some(Role::Assistant),
1019                content: None,
1020                tool_calls: None,
1021            },
1022            GenerationDelta::ToolUse { tool_use } => ChatMessageDelta {
1023                role: Some(Role::Assistant),
1024                content: None,
1025                tool_calls: Some(vec![ToolCallDelta {
1026                    index: tool_use.index,
1027                    id: tool_use.id,
1028                    r#type: Some("function".to_string()),
1029                    function: Some(FunctionCallDelta {
1030                        name: tool_use.name,
1031                        arguments: tool_use.input,
1032                    }),
1033                    metadata: tool_use.metadata,
1034                }]),
1035            },
1036            _ => ChatMessageDelta {
1037                role: Some(Role::Assistant),
1038                content: None,
1039                tool_calls: None,
1040            },
1041        }
1042    }
1043}
1044
1045#[cfg(test)]
1046mod tests {
1047    use super::*;
1048
1049    #[test]
1050    fn test_serialize_basic_request() {
1051        let request = ChatCompletionRequest {
1052            model: "gpt-4".to_string(),
1053            messages: vec![
1054                ChatMessage {
1055                    role: Role::System,
1056                    content: Some(MessageContent::String(
1057                        "You are a helpful assistant.".to_string(),
1058                    )),
1059                    name: None,
1060                    tool_calls: None,
1061                    tool_call_id: None,
1062                    usage: None,
1063                    ..Default::default()
1064                },
1065                ChatMessage {
1066                    role: Role::User,
1067                    content: Some(MessageContent::String("Hello!".to_string())),
1068                    name: None,
1069                    tool_calls: None,
1070                    tool_call_id: None,
1071                    usage: None,
1072                    ..Default::default()
1073                },
1074            ],
1075            frequency_penalty: None,
1076            logit_bias: None,
1077            logprobs: None,
1078            max_tokens: Some(100),
1079            n: None,
1080            presence_penalty: None,
1081            response_format: None,
1082            seed: None,
1083            stop: None,
1084            stream: None,
1085            temperature: Some(0.7),
1086            top_p: None,
1087            tools: None,
1088            tool_choice: None,
1089            user: None,
1090            context: None,
1091        };
1092
1093        let json = serde_json::to_string(&request).unwrap();
1094        assert!(json.contains("\"model\":\"gpt-4\""));
1095        assert!(json.contains("\"messages\":["));
1096        assert!(json.contains("\"role\":\"system\""));
1097    }
1098
1099    #[test]
1100    fn test_llm_message_to_chat_message() {
1101        let llm_message = LLMMessage {
1102            role: "user".to_string(),
1103            content: LLMMessageContent::String("Hello, world!".to_string()),
1104        };
1105
1106        let chat_message = ChatMessage::from(llm_message);
1107        assert_eq!(chat_message.role, Role::User);
1108        match &chat_message.content {
1109            Some(MessageContent::String(text)) => assert_eq!(text, "Hello, world!"),
1110            _ => panic!("Expected string content"),
1111        }
1112    }
1113
1114    #[test]
1115    fn test_llm_message_to_chat_message_tool_result_preserves_tool_call_id() {
1116        let llm_message = LLMMessage {
1117            role: "tool".to_string(),
1118            content: LLMMessageContent::List(vec![LLMMessageTypedContent::ToolResult {
1119                tool_use_id: "toolu_01Abc123".to_string(),
1120                content: "Tool execution result".to_string(),
1121            }]),
1122        };
1123
1124        let chat_message = ChatMessage::from(llm_message);
1125        assert_eq!(chat_message.role, Role::Tool);
1126        assert_eq!(chat_message.tool_call_id.as_deref(), Some("toolu_01Abc123"));
1127        assert_eq!(
1128            chat_message.content,
1129            Some(MessageContent::Array(vec![ContentPart {
1130                r#type: "text".to_string(),
1131                text: Some("Tool execution result".to_string()),
1132                image_url: None,
1133            }]))
1134        );
1135    }
1136
1137    #[test]
1138    fn test_chat_message_to_llm_message_tool_result() {
1139        // Test that Tool role messages with tool_call_id are converted to ToolResult content
1140        // This is critical for Anthropic compatibility - the provider layer converts
1141        // role="tool" to role="user" with tool_result content blocks
1142        let chat_message = ChatMessage {
1143            role: Role::Tool,
1144            content: Some(MessageContent::String("Tool execution result".to_string())),
1145            name: None,
1146            tool_calls: None,
1147            tool_call_id: Some("toolu_01Abc123".to_string()),
1148            usage: None,
1149            ..Default::default()
1150        };
1151
1152        let llm_message: LLMMessage = chat_message.into();
1153
1154        // Role should be preserved as "tool" - provider layer handles conversion
1155        assert_eq!(llm_message.role, "tool");
1156
1157        // Content should be a ToolResult with the tool_call_id
1158        match &llm_message.content {
1159            LLMMessageContent::List(parts) => {
1160                assert_eq!(parts.len(), 1, "Should have exactly one content part");
1161                match &parts[0] {
1162                    LLMMessageTypedContent::ToolResult {
1163                        tool_use_id,
1164                        content,
1165                    } => {
1166                        assert_eq!(tool_use_id, "toolu_01Abc123");
1167                        assert_eq!(content, "Tool execution result");
1168                    }
1169                    _ => panic!("Expected ToolResult content part, got {:?}", parts[0]),
1170                }
1171            }
1172            _ => panic!(
1173                "Expected List content with ToolResult, got {:?}",
1174                llm_message.content
1175            ),
1176        }
1177    }
1178
1179    #[test]
1180    fn test_chat_message_to_llm_message_tool_result_empty_content() {
1181        // Test tool result with empty content
1182        let chat_message = ChatMessage {
1183            role: Role::Tool,
1184            content: None,
1185            name: None,
1186            tool_calls: None,
1187            tool_call_id: Some("toolu_02Xyz789".to_string()),
1188            usage: None,
1189            ..Default::default()
1190        };
1191
1192        let llm_message: LLMMessage = chat_message.into();
1193
1194        assert_eq!(llm_message.role, "tool");
1195        match &llm_message.content {
1196            LLMMessageContent::List(parts) => {
1197                assert_eq!(parts.len(), 1);
1198                match &parts[0] {
1199                    LLMMessageTypedContent::ToolResult {
1200                        tool_use_id,
1201                        content,
1202                    } => {
1203                        assert_eq!(tool_use_id, "toolu_02Xyz789");
1204                        assert_eq!(content, ""); // Empty content
1205                    }
1206                    _ => panic!("Expected ToolResult content part"),
1207                }
1208            }
1209            _ => panic!("Expected List content with ToolResult"),
1210        }
1211    }
1212
1213    #[test]
1214    fn test_chat_message_to_llm_message_assistant_with_tool_calls() {
1215        // Test that assistant messages with tool_calls are converted correctly
1216        let chat_message = ChatMessage {
1217            role: Role::Assistant,
1218            content: Some(MessageContent::String(
1219                "I'll help you with that.".to_string(),
1220            )),
1221            name: None,
1222            tool_calls: Some(vec![ToolCall {
1223                id: "call_abc123".to_string(),
1224                r#type: "function".to_string(),
1225                function: FunctionCall {
1226                    name: "get_weather".to_string(),
1227                    arguments: r#"{"location": "Paris"}"#.to_string(),
1228                },
1229                metadata: None,
1230            }]),
1231            tool_call_id: None,
1232            usage: None,
1233            ..Default::default()
1234        };
1235
1236        let llm_message: LLMMessage = chat_message.into();
1237
1238        assert_eq!(llm_message.role, "assistant");
1239        match &llm_message.content {
1240            LLMMessageContent::List(parts) => {
1241                assert_eq!(parts.len(), 2, "Should have text and tool call");
1242
1243                // First part should be text
1244                match &parts[0] {
1245                    LLMMessageTypedContent::Text { text } => {
1246                        assert_eq!(text, "I'll help you with that.");
1247                    }
1248                    _ => panic!("Expected Text content part first"),
1249                }
1250
1251                // Second part should be tool call
1252                match &parts[1] {
1253                    LLMMessageTypedContent::ToolCall { id, name, args, .. } => {
1254                        assert_eq!(id, "call_abc123");
1255                        assert_eq!(name, "get_weather");
1256                        assert_eq!(args["location"], "Paris");
1257                    }
1258                    _ => panic!("Expected ToolCall content part second"),
1259                }
1260            }
1261            _ => panic!("Expected List content"),
1262        }
1263    }
1264}
stakpak_shared/models/integrations/openai.rs

stakpak_shared/models/integrations/
openai.rs
