Compare commits

..

4 Commits

Author SHA1 Message Date
d0be16d8e3 fix(openai): parse embedded 'tool_uses' JSON for gpt-5.4 parallel calls
Some checks failed
CI / Check (push) Has been cancelled
CI / Clippy (push) Has been cancelled
CI / Formatting (push) Has been cancelled
CI / Test (push) Has been cancelled
CI / Release Build (push) Has been cancelled
- Add static parse_tool_uses_json helper to extract embedded tool calls
- Update synchronous and streaming Responses API parsers to detect tool_uses blocks
- Strip tool_uses JSON from content to prevent raw JSON leakage to client
- Resolve lifetime issues by avoiding &self capture in streaming closure
2026-03-18 14:28:38 +00:00
83e0ad0240 fix(openai): flatten tools and tool_choice for Responses API
Some checks failed
CI / Check (push) Has been cancelled
CI / Clippy (push) Has been cancelled
CI / Formatting (push) Has been cancelled
CI / Test (push) Has been cancelled
CI / Release Build (push) Has been cancelled
- Map nested 'function' object to top-level fields
- Support string and object-based 'tool_choice' formats
- Fix 400 Bad Request 'Missing required parameter: tools[0].name'
2026-03-18 14:00:49 +00:00
275ce34d05 fix(openai): fix missing tools and instructions in Responses API
Some checks failed
CI / Check (push) Has been cancelled
CI / Clippy (push) Has been cancelled
CI / Formatting (push) Has been cancelled
CI / Test (push) Has been cancelled
CI / Release Build (push) Has been cancelled
- Add 'tools' and 'tool_choice' parameters to streaming Responses API
- Include 'name' field in message items for Responses API input
- Use string content for text-only messages to improve instruction following
- Fix subagents not triggering and files not being created
2026-03-18 13:51:36 +00:00
cb5b921550 feat(openai): implement tool support for gpt-5.4 via Responses API
Some checks failed
CI / Check (push) Has been cancelled
CI / Clippy (push) Has been cancelled
CI / Formatting (push) Has been cancelled
CI / Test (push) Has been cancelled
CI / Release Build (push) Has been cancelled
- Implement polymorphic 'input' structure for /responses endpoint
- Map 'tool' role to 'function_call_output' items
- Handle assistant 'tool_calls' as separate 'function_call' items
- Add synchronous and streaming parsers for function_call items
- Fix 400 Bad Request 'Invalid value: tool' error
2026-03-18 13:14:51 +00:00

View File

@@ -36,6 +36,57 @@ impl OpenAIProvider {
pricing: app_config.pricing.openai.clone(),
})
}
/// GPT-5.4 models sometimes emit parallel tool calls as a JSON block starting with
/// '{"tool_uses":' inside a text message instead of discrete function_call items.
/// This method attempts to extract and parse such tool calls.
pub fn parse_tool_uses_json(text: &str) -> Vec<crate::models::ToolCall> {
let mut calls = Vec::new();
if let Some(start) = text.find("{\"tool_uses\":") {
// Find the end of the JSON block by matching braces
let sub = &text[start..];
let mut brace_count = 0;
let mut end_idx = 0;
let mut found = false;
for (i, c) in sub.char_indices() {
if c == '{' { brace_count += 1; }
else if c == '}' {
brace_count -= 1;
if brace_count == 0 {
end_idx = i + 1;
found = true;
break;
}
}
}
if found {
let json_str = &sub[..end_idx];
if let Ok(val) = serde_json::from_str::<serde_json::Value>(json_str) {
if let Some(uses) = val.get("tool_uses").and_then(|u| u.as_array()) {
for (idx, u) in uses.iter().enumerate() {
let name = u.get("recipient_name")
.and_then(|v| v.as_str())
.unwrap_or("unknown")
// Strip "functions." prefix if present
.replace("functions.", "");
let arguments = u.get("parameters")
.map(|v| v.to_string())
.unwrap_or_else(|| "{}".to_string());
calls.push(crate::models::ToolCall {
id: format!("call_tu_{}_{}", uuid::Uuid::new_v4().to_string()[..8].to_string(), idx),
call_type: "function".to_string(),
function: crate::models::FunctionCall { name, arguments },
});
}
}
}
}
}
calls
}
}
#[async_trait]
@@ -112,10 +163,57 @@ impl super::Provider for OpenAIProvider {
let messages_json = helpers::messages_to_openai_json(&request.messages).await?;
let mut input_parts = Vec::new();
for m in &messages_json {
let mut role = m["role"].as_str().unwrap_or("user").to_string();
// Newer models (gpt-5, o1) prefer "developer" over "system"
if role == "system" {
role = "developer".to_string();
let role = m["role"].as_str().unwrap_or("user");
if role == "tool" {
input_parts.push(serde_json::json!({
"type": "function_call_output",
"call_id": m.get("tool_call_id").and_then(|v| v.as_str()).unwrap_or(""),
"output": m.get("content").and_then(|v| v.as_str()).unwrap_or("")
}));
continue;
}
if role == "assistant" && m.get("tool_calls").is_some() {
// Push message part if it exists
let content_val = m.get("content").cloned().unwrap_or(serde_json::json!(""));
if !content_val.is_null() && (content_val.is_array() && !content_val.as_array().unwrap().is_empty() || content_val.is_string() && !content_val.as_str().unwrap().is_empty()) {
let mut content = content_val.clone();
if let Some(text) = content.as_str() {
content = serde_json::json!([{ "type": "output_text", "text": text }]);
} else if let Some(arr) = content.as_array_mut() {
for part in arr {
if let Some(obj) = part.as_object_mut() {
if obj.get("type").and_then(|v| v.as_str()) == Some("text") {
obj.insert("type".to_string(), serde_json::json!("output_text"));
}
}
}
}
input_parts.push(serde_json::json!({
"type": "message",
"role": "assistant",
"content": content
}));
}
// Push tool calls as separate items
if let Some(tcs) = m.get("tool_calls").and_then(|v| v.as_array()) {
for tc in tcs {
input_parts.push(serde_json::json!({
"type": "function_call",
"call_id": tc["id"],
"name": tc["function"]["name"],
"arguments": tc["function"]["arguments"]
}));
}
}
continue;
}
let mut mapped_role = role.to_string();
if mapped_role == "system" {
mapped_role = "developer".to_string();
}
let mut content = m.get("content").cloned().unwrap_or(serde_json::json!([]));
@@ -127,12 +225,11 @@ impl super::Provider for OpenAIProvider {
if let Some(t) = part_obj.get("type").and_then(|v| v.as_str()) {
match t {
"text" => {
let new_type = if role == "assistant" { "output_text" } else { "input_text" };
let new_type = if mapped_role == "assistant" { "output_text" } else { "input_text" };
part_obj.insert("type".to_string(), serde_json::json!(new_type));
}
"image_url" => {
// Assistant typically doesn't have image_url in history this way, but for safety:
let new_type = if role == "assistant" { "output_image" } else { "input_image" };
let new_type = if mapped_role == "assistant" { "output_image" } else { "input_image" };
part_obj.insert("type".to_string(), serde_json::json!(new_type));
if let Some(img_url) = part_obj.remove("image_url") {
part_obj.insert("image".to_string(), img_url);
@@ -144,14 +241,20 @@ impl super::Provider for OpenAIProvider {
}
}
} else if let Some(text) = content.as_str() {
let new_type = if role == "assistant" { "output_text" } else { "input_text" };
content = serde_json::json!([{ "type": new_type, "text": text }]);
// If it's just a string, send it as a string instead of an array of objects
// as it's safer for standard conversational messages.
content = serde_json::json!(text);
}
input_parts.push(serde_json::json!({
"role": role,
let mut msg_item = serde_json::json!({
"type": "message",
"role": mapped_role,
"content": content
}));
});
if let Some(name) = m.get("name") {
msg_item["name"] = name.clone();
}
input_parts.push(msg_item);
}
let mut body = serde_json::json!({
@@ -174,7 +277,33 @@ impl super::Provider for OpenAIProvider {
}
if let Some(tools) = &request.tools {
body["tools"] = serde_json::json!(tools);
let flattened: Vec<serde_json::Value> = tools.iter().map(|t| {
let mut obj = serde_json::json!({
"type": t.tool_type,
"name": t.function.name,
});
if let Some(desc) = &t.function.description {
obj["description"] = serde_json::json!(desc);
}
if let Some(params) = &t.function.parameters {
obj["parameters"] = params.clone();
}
obj
}).collect();
body["tools"] = serde_json::json!(flattened);
}
if let Some(tool_choice) = &request.tool_choice {
match tool_choice {
crate::models::ToolChoice::Mode(mode) => {
body["tool_choice"] = serde_json::json!(mode);
}
crate::models::ToolChoice::Specific(specific) => {
body["tool_choice"] = serde_json::json!({
"type": specific.choice_type,
"name": specific.function.name,
});
}
}
}
let resp = self
@@ -200,18 +329,43 @@ impl super::Provider for OpenAIProvider {
// Normalize Responses API output into ProviderResponse
let mut content_text = String::new();
let mut tool_calls = Vec::new();
if let Some(output) = resp_json.get("output").and_then(|o| o.as_array()) {
for out in output {
if let Some(contents) = out.get("content").and_then(|c| c.as_array()) {
for item in contents {
if let Some(text) = item.get("text").and_then(|t| t.as_str()) {
if !content_text.is_empty() { content_text.push_str("\n"); }
content_text.push_str(text);
} else if let Some(parts) = item.get("parts").and_then(|p| p.as_array()) {
for p in parts {
if let Some(t) = p.as_str() {
let item_type = out.get("type").and_then(|v| v.as_str()).unwrap_or("");
match item_type {
"message" => {
if let Some(contents) = out.get("content").and_then(|c| c.as_array()) {
for item in contents {
if let Some(text) = item.get("text").and_then(|t| t.as_str()) {
if !content_text.is_empty() { content_text.push_str("\n"); }
content_text.push_str(t);
content_text.push_str(text);
}
}
}
}
"function_call" => {
let id = out.get("call_id")
.or_else(|| out.get("item_id"))
.or_else(|| out.get("id"))
.and_then(|v| v.as_str())
.unwrap_or("")
.to_string();
let name = out.get("name").and_then(|v| v.as_str()).unwrap_or("").to_string();
let arguments = out.get("arguments").and_then(|v| v.as_str()).unwrap_or("").to_string();
tool_calls.push(crate::models::ToolCall {
id,
call_type: "function".to_string(),
function: crate::models::FunctionCall { name, arguments },
});
}
_ => {
// Fallback for older/nested structure
if let Some(contents) = out.get("content").and_then(|c| c.as_array()) {
for item in contents {
if let Some(text) = item.get("text").and_then(|t| t.as_str()) {
if !content_text.is_empty() { content_text.push_str("\n"); }
content_text.push_str(text);
}
}
}
@@ -241,10 +395,20 @@ impl super::Provider for OpenAIProvider {
let completion_tokens = resp_json.get("usage").and_then(|u| u.get("completion_tokens")).and_then(|v| v.as_u64()).unwrap_or(0) as u32;
let total_tokens = resp_json.get("usage").and_then(|u| u.get("total_tokens")).and_then(|v| v.as_u64()).unwrap_or(0) as u32;
// GPT-5.4 parallel tool calls might be embedded in content_text as a JSON block
let embedded_calls = Self::parse_tool_uses_json(&content_text);
if !embedded_calls.is_empty() {
// Strip the JSON part from content_text to keep it clean
if let Some(start) = content_text.find("{\"tool_uses\":") {
content_text = content_text[..start].to_string();
}
tool_calls.extend(embedded_calls);
}
Ok(ProviderResponse {
content: content_text,
reasoning_content: None,
tool_calls: None,
tool_calls: if tool_calls.is_empty() { None } else { Some(tool_calls) },
prompt_tokens,
completion_tokens,
reasoning_tokens: 0,
@@ -379,10 +543,57 @@ impl super::Provider for OpenAIProvider {
let messages_json = helpers::messages_to_openai_json(&request.messages).await?;
let mut input_parts = Vec::new();
for m in &messages_json {
let mut role = m["role"].as_str().unwrap_or("user").to_string();
// Newer models (gpt-5, o1) prefer "developer" over "system"
if role == "system" {
role = "developer".to_string();
let role = m["role"].as_str().unwrap_or("user");
if role == "tool" {
input_parts.push(serde_json::json!({
"type": "function_call_output",
"call_id": m.get("tool_call_id").and_then(|v| v.as_str()).unwrap_or(""),
"output": m.get("content").and_then(|v| v.as_str()).unwrap_or("")
}));
continue;
}
if role == "assistant" && m.get("tool_calls").is_some() {
// Push message part if it exists
let content_val = m.get("content").cloned().unwrap_or(serde_json::json!(""));
if !content_val.is_null() && (content_val.is_array() && !content_val.as_array().unwrap().is_empty() || content_val.is_string() && !content_val.as_str().unwrap().is_empty()) {
let mut content = content_val.clone();
if let Some(text) = content.as_str() {
content = serde_json::json!([{ "type": "output_text", "text": text }]);
} else if let Some(arr) = content.as_array_mut() {
for part in arr {
if let Some(obj) = part.as_object_mut() {
if obj.get("type").and_then(|v| v.as_str()) == Some("text") {
obj.insert("type".to_string(), serde_json::json!("output_text"));
}
}
}
}
input_parts.push(serde_json::json!({
"type": "message",
"role": "assistant",
"content": content
}));
}
// Push tool calls as separate items
if let Some(tcs) = m.get("tool_calls").and_then(|v| v.as_array()) {
for tc in tcs {
input_parts.push(serde_json::json!({
"type": "function_call",
"call_id": tc["id"],
"name": tc["function"]["name"],
"arguments": tc["function"]["arguments"]
}));
}
}
continue;
}
let mut mapped_role = role.to_string();
if mapped_role == "system" {
mapped_role = "developer".to_string();
}
let mut content = m.get("content").cloned().unwrap_or(serde_json::json!([]));
@@ -394,12 +605,11 @@ impl super::Provider for OpenAIProvider {
if let Some(t) = part_obj.get("type").and_then(|v| v.as_str()) {
match t {
"text" => {
let new_type = if role == "assistant" { "output_text" } else { "input_text" };
let new_type = if mapped_role == "assistant" { "output_text" } else { "input_text" };
part_obj.insert("type".to_string(), serde_json::json!(new_type));
}
"image_url" => {
// Assistant typically doesn't have image_url in history this way, but for safety:
let new_type = if role == "assistant" { "output_image" } else { "input_image" };
let new_type = if mapped_role == "assistant" { "output_image" } else { "input_image" };
part_obj.insert("type".to_string(), serde_json::json!(new_type));
if let Some(img_url) = part_obj.remove("image_url") {
part_obj.insert("image".to_string(), img_url);
@@ -411,14 +621,20 @@ impl super::Provider for OpenAIProvider {
}
}
} else if let Some(text) = content.as_str() {
let new_type = if role == "assistant" { "output_text" } else { "input_text" };
content = serde_json::json!([{ "type": new_type, "text": text }]);
// If it's just a string, send it as a string instead of an array of objects
// as it's safer for standard conversational messages.
content = serde_json::json!(text);
}
input_parts.push(serde_json::json!({
"role": role,
let mut msg_item = serde_json::json!({
"type": "message",
"role": mapped_role,
"content": content
}));
});
if let Some(name) = m.get("name") {
msg_item["name"] = name.clone();
}
input_parts.push(msg_item);
}
let mut body = serde_json::json!({
@@ -441,6 +657,36 @@ impl super::Provider for OpenAIProvider {
}
}
if let Some(tools) = &request.tools {
let flattened: Vec<serde_json::Value> = tools.iter().map(|t| {
let mut obj = serde_json::json!({
"type": t.tool_type,
"name": t.function.name,
});
if let Some(desc) = &t.function.description {
obj["description"] = serde_json::json!(desc);
}
if let Some(params) = &t.function.parameters {
obj["parameters"] = params.clone();
}
obj
}).collect();
body["tools"] = serde_json::json!(flattened);
}
if let Some(tool_choice) = &request.tool_choice {
match tool_choice {
crate::models::ToolChoice::Mode(mode) => {
body["tool_choice"] = serde_json::json!(mode);
}
crate::models::ToolChoice::Specific(specific) => {
body["tool_choice"] = serde_json::json!({
"type": specific.choice_type,
"name": specific.function.name,
});
}
}
}
let url = format!("{}/responses", self.config.base_url);
let api_key = self.api_key.clone();
let model = request.model.clone();
@@ -475,6 +721,7 @@ impl super::Provider for OpenAIProvider {
// Responses API specific parsing for streaming
let mut content = String::new();
let mut finish_reason = None;
let mut tool_calls = None;
let event_type = chunk.get("type").and_then(|v| v.as_str()).unwrap_or("");
@@ -484,15 +731,35 @@ impl super::Provider for OpenAIProvider {
content.push_str(delta);
}
}
"response.output_text.done" => {
if let Some(text) = chunk.get("text").and_then(|v| v.as_str()) {
// Some implementations send the full text at the end
// We usually prefer deltas, but if we haven't seen them, this is the fallback.
// However, if we're already yielding deltas, we might not want this.
// For now, let's just use it as a signal that we're done.
finish_reason = Some("stop".to_string());
"response.item.delta" => {
if let Some(delta) = chunk.get("delta") {
let t = delta.get("type").and_then(|v| v.as_str()).unwrap_or("");
if t == "function_call" {
let call_id = delta.get("call_id")
.or_else(|| chunk.get("item_id"))
.and_then(|v| v.as_str());
let name = delta.get("name").and_then(|v| v.as_str());
let arguments = delta.get("arguments").and_then(|v| v.as_str());
tool_calls = Some(vec![crate::models::ToolCallDelta {
index: chunk.get("output_index").and_then(|v| v.as_u64()).unwrap_or(0) as u32,
id: call_id.map(|s| s.to_string()),
call_type: Some("function".to_string()),
function: Some(crate::models::FunctionCallDelta {
name: name.map(|s| s.to_string()),
arguments: arguments.map(|s| s.to_string()),
}),
}]);
} else if t == "message" {
if let Some(text) = delta.get("text").and_then(|v| v.as_str()) {
content.push_str(text);
}
}
}
}
"response.output_text.done" | "response.item.done" => {
finish_reason = Some("stop".to_string());
}
"response.done" => {
finish_reason = Some("stop".to_string());
}
@@ -514,12 +781,41 @@ impl super::Provider for OpenAIProvider {
}
}
if !content.is_empty() || finish_reason.is_some() {
// GPT-5.4 parallel tool calls might be embedded in content as a JSON block
let embedded_calls = Self::parse_tool_uses_json(&content);
if !embedded_calls.is_empty() {
// Strip the JSON part from content to keep it clean
if let Some(start) = content.find("{\"tool_uses\":") {
content = content[..start].to_string();
}
// Convert ToolCall to ToolCallDelta for streaming
let deltas: Vec<crate::models::ToolCallDelta> = embedded_calls.into_iter().enumerate().map(|(idx, tc)| {
crate::models::ToolCallDelta {
index: idx as u32,
id: Some(tc.id),
call_type: Some("function".to_string()),
function: Some(crate::models::FunctionCallDelta {
name: Some(tc.function.name),
arguments: Some(tc.function.arguments),
}),
}
}).collect();
if let Some(ref mut existing) = tool_calls {
existing.extend(deltas);
} else {
tool_calls = Some(deltas);
}
}
if !content.is_empty() || finish_reason.is_some() || tool_calls.is_some() {
yield ProviderStreamChunk {
content,
reasoning_content: None,
finish_reason,
tool_calls: None,
tool_calls,
model: model.clone(),
usage: None,
};