Skip to content

feat: add Kimi Code provider #7222

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Closed
stebbins wants to merge 1 commit into from
Closed

feat: add Kimi Code provider #7222

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
22 changes: 19 additions & 3 deletions crates/goose/src/agents/agent.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1346,8 +1346,9 @@ impl Agent {
}
}

// Preserve thinking content from the original response
// Preserve thinking/reasoning content from the original response
// Gemini (and other thinking models) require thinking to be echoed back
// Kimi/DeepSeek require reasoning_content on assistant tool call messages
let thinking_content: Vec<MessageContent> = response.content.iter()
.filter(|c| matches!(c, MessageContent::Thinking(_)))
.cloned()
Expand All @@ -1361,10 +1362,25 @@ impl Agent {
messages_to_add.push(thinking_msg);
}

// Collect reasoning content to attach to tool request messages
let reasoning_content: Vec<MessageContent> = response.content.iter()
.filter(|c| matches!(c, MessageContent::Reasoning(_)))
.cloned()
.collect();

for (idx, request) in frontend_requests.iter().chain(remaining_requests.iter()).enumerate() {
if request.tool_call.is_ok() {
let request_msg = Message::assistant()
.with_id(format!("msg_{}", Uuid::new_v4()))
let mut request_msg = Message::assistant()
.with_id(format!("msg_{}", Uuid::new_v4()));

// Attach reasoning content to EVERY split tool request message.
// Providers like Kimi require reasoning_content on all assistant
// messages with tool_calls when thinking mode is enabled.
for rc in &reasoning_content {
request_msg = request_msg.with_content(rc.clone());
}

request_msg = request_msg
.with_tool_request_with_metadata(
request.id.clone(),
request.tool_call.clone(),
Expand Down
7 changes: 5 additions & 2 deletions crates/goose/src/model.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -98,8 +98,11 @@ static MODEL_SPECIFIC_LIMITS: Lazy<Vec<(&'static str, usize)>> = Lazy::new(|| {
("grok-4", 256_000),
("grok-code-fast-1", 256_000),
("grok", 131_072),
// other
("kimi-k2", 131_072),
// moonshot/kimi
("kimi-k2-0711", 131_072),
("kimi-k2", 262_144),
("kimi-for-coding", 262_144),
("kimi-code", 262_144),
]
});

Expand Down
29 changes: 29 additions & 0 deletions crates/goose/src/providers/declarative/kimi.json
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,29 @@
{
"name": "kimi",
"engine": "openai",
"display_name": "Kimi Code",
"description": "Kimi Code subscription models (powered by Kimi K2.5)",
"api_key_env": "KIMI_API_KEY",
"base_url": "https://api.kimi.com/coding/v1/chat/completions",
"models": [
{
"name": "kimi-for-coding",
"context_limit": 262144,
"input_token_cost": null,
"output_token_cost": null,
"currency": null,
"supports_cache_control": null
},
{
"name": "kimi-code",
"context_limit": 262144,
"input_token_cost": null,
"output_token_cost": null,
"currency": null,
"supports_cache_control": null
}
],
"headers": null,
"timeout_seconds": null,
"supports_streaming": true
}
38 changes: 27 additions & 11 deletions crates/goose/src/providers/formats/openai.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -82,7 +82,7 @@ pub fn format_messages(messages: &[Message], image_format: &ImageFormat) -> Vec<
let mut output = Vec::new();
let mut content_array = Vec::new();
let mut text_array = Vec::new();
let mut reasoning_text: Option<String> = None;
let mut reasoning_text = String::new();

for content in &message.content {
match content {
Expand Down Expand Up @@ -116,7 +116,7 @@ pub fn format_messages(messages: &[Message], image_format: &ImageFormat) -> Vec<
continue;
}
MessageContent::Reasoning(r) => {
reasoning_text = Some(r.text.clone());
reasoning_text.push_str(&r.text);
}
MessageContent::ToolRequest(request) => match &request.tool_call {
Ok(tool_call) => {
Expand Down Expand Up @@ -278,15 +278,11 @@ pub fn format_messages(messages: &[Message], image_format: &ImageFormat) -> Vec<
converted["content"] = json!(null);
}

// DeepSeek requires reasoning_content field when tool_calls are present
// Set it to the captured reasoning text, or empty string if not present
if converted.get("tool_calls").is_some() {
let reasoning = reasoning_text.unwrap_or_default();
converted["reasoning_content"] = json!(reasoning);
} else if let Some(reasoning) = reasoning_text {
if !reasoning.is_empty() {
converted["reasoning_content"] = json!(reasoning);
}
// Include reasoning_content only when non-empty.
// Kimi rejects empty reasoning_content (""), so we must omit it entirely
// when there's no reasoning to send.
if !reasoning_text.is_empty() {
converted["reasoning_content"] = json!(reasoning_text);
}

if converted.get("content").is_some() || converted.get("tool_calls").is_some() {
Expand Down Expand Up @@ -542,6 +538,7 @@ where
use futures::StreamExt;

let mut accumulated_reasoning: Vec<Value> = Vec::new();
let mut accumulated_reasoning_content = String::new();

'outer: while let Some(response) = stream.next().await {
if response.as_ref().is_ok_and(|s| s == "data: [DONE]") {
Expand All @@ -562,6 +559,9 @@ where
if let Some(details) = &chunk.choices[0].delta.reasoning_details {
accumulated_reasoning.extend(details.iter().cloned());
}
if let Some(rc) = &chunk.choices[0].delta.reasoning_content {
accumulated_reasoning_content.push_str(rc);
}
}

let mut usage = extract_usage_with_output_tokens(&chunk);
Expand Down Expand Up @@ -602,6 +602,9 @@ where
if let Some(details) = &tool_chunk.choices[0].delta.reasoning_details {
accumulated_reasoning.extend(details.iter().cloned());
}
if let Some(rc) = &tool_chunk.choices[0].delta.reasoning_content {
accumulated_reasoning_content.push_str(rc);
}
if let Some(delta_tool_calls) = &tool_chunk.choices[0].delta.tool_calls {
for delta_call in delta_tool_calls {
if let Some(index) = delta_call.index {
Expand Down Expand Up @@ -642,6 +645,10 @@ where
};

let mut contents = Vec::new();
if !accumulated_reasoning_content.is_empty() {
contents.push(MessageContent::reasoning(&accumulated_reasoning_content));
accumulated_reasoning_content.clear();
}
let mut sorted_indices: Vec<_> = tool_call_data.keys().cloned().collect();
sorted_indices.sort();

Expand Down Expand Up @@ -835,6 +842,15 @@ pub fn create_request(
payload["stream_options"] = json!({"include_usage": true});
}

// Merge provider-specific request parameters (e.g., reasoning_effort, thinking config)
if let Some(ref params) = model_config.request_params {
if let Some(payload_obj) = payload.as_object_mut() {
for (key, value) in params {
payload_obj.insert(key.clone(), value.clone());
}
}
}

Ok(payload)
}

Expand Down