dgf1988/codewhale
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

fix(codex): keep Responses reasoning shape provider-native

Browse Source 
Codex Responses requests now clamp legacy minimal effort to low and carry a regression test that the request body uses reasoning.effort without DeepSeek thinking or chat-completions reasoning_effort fields.
This commit is contained in:
Hunter B
2026-06-12 02:27:59 -07:00
parent 8c690cb7bf
commit c1a48492b6
3 changed files with 92 additions and 50 deletions
Download Patch File Download Diff File Expand all files Collapse all files
CHANGELOG.md
+3 -1
View File
@@ -90,7 +90,9 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
`low`/`medium`/`high`/`xhigh` tiers. Startup, `/config`, and the model
picker now display Codex labels instead of leaking DeepSeek
`off`/`max` names, while Codex still reports as a Responses payload
provider.
provider. The Responses request builder also clamps legacy `minimal` input
to `low` and has regression coverage that Codex requests use
`reasoning.effort`, not DeepSeek `thinking` fields.
- **OpenAI Codex context metadata (#3070).** The `gpt-5.5` default and
CodeWhale aliases now use OpenAI's documented 1,050,000-token context window
and 128,000 max-output metadata for context pressure, prompts, and doctor
crates/tui/CHANGELOG.md
+3 -1
View File
@@ -90,7 +90,9 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
`low`/`medium`/`high`/`xhigh` tiers. Startup, `/config`, and the model
picker now display Codex labels instead of leaking DeepSeek
`off`/`max` names, while Codex still reports as a Responses payload
provider.
provider. The Responses request builder also clamps legacy `minimal` input
to `low` and has regression coverage that Codex requests use
`reasoning.effort`, not DeepSeek `thinking` fields.
- **OpenAI Codex context metadata (#3070).** The `gpt-5.5` default and
CodeWhale aliases now use OpenAI's documented 1,050,000-token context window
and 128,000 max-output metadata for context pressure, prompts, and doctor
crates/tui/src/client/responses.rs
+86 -48
View File
@@ -23,64 +23,63 @@ use super::{DeepSeekClient, ERROR_BODY_MAX_BYTES, bounded_error_text, system_to_
/// Base URL path for the Codex Responses endpoint.
const CODEX_RESPONSES_PATH: &str = "/codex/responses";
impl DeepSeekClient {
/// Build the Responses API request body from a `MessageRequest`.
fn build_responses_body(&self, request: &MessageRequest) -> Value {
let model = &request.model;
let mut body = json!({
"model": model,
"stream": true,
"store": false,
});
/// Build the Responses API request body from a `MessageRequest`.
fn build_responses_body(request: &MessageRequest) -> Value {
let model = &request.model;
let mut body = json!({
"model": model,
"stream": true,
"store": false,
});
// Instructions (system prompt). The Codex Responses backend rejects
// requests without instructions, so fall back to a minimal system
// prompt when the caller did not supply one.
let instructions = system_to_instructions(request.system.clone())
.filter(|text| !text.trim().is_empty())
.unwrap_or_else(|| "You are a helpful assistant.".to_string());
body["instructions"] = json!(instructions);
// Instructions (system prompt). The Codex Responses backend rejects
// requests without instructions, so fall back to a minimal system
// prompt when the caller did not supply one.
let instructions = system_to_instructions(request.system.clone())
.filter(|text| !text.trim().is_empty())
.unwrap_or_else(|| "You are a helpful assistant.".to_string());
body["instructions"] = json!(instructions);
// Convert messages to Responses input items.
let input = convert_messages_to_responses_input(request);
body["input"] = json!(input);
// Convert messages to Responses input items.
let input = convert_messages_to_responses_input(request);
body["input"] = json!(input);
// Convert tools to Responses function tools.
if let Some(tools) = request.tools.as_ref() {
let responses_tools: Vec<Value> =
tools.iter().map(tool_to_responses_function).collect();
if !responses_tools.is_empty() {
body["tools"] = json!(responses_tools);
body["tool_choice"] = json!("auto");
body["parallel_tool_calls"] = json!(true);
}
// Convert tools to Responses function tools.
if let Some(tools) = request.tools.as_ref() {
let responses_tools: Vec<Value> = tools.iter().map(tool_to_responses_function).collect();
if !responses_tools.is_empty() {
body["tools"] = json!(responses_tools);
body["tool_choice"] = json!("auto");
body["parallel_tool_calls"] = json!(true);
}
// Reasoning configuration. The Codex Responses backend accepts
// low/medium/high/xhigh, so provider-aware callers normalize inherited
// DeepSeek-only values before request construction: "off" becomes
// "low", and CodeWhale's "auto" falls back to "medium".
if let Some(raw) = request.reasoning_effort.as_deref()
&& let Some(effort) = codex_responses_reasoning_effort(raw)
{
body["reasoning"] = json!({
"effort": effort,
"summary": "auto",
});
}
// Include reasoning summaries in the stream.
body["include"] = json!(["reasoning.encrypted_content"]);
body
}
// Reasoning configuration. The Codex Responses backend accepts
// low/medium/high/xhigh, so provider-aware callers normalize inherited
// DeepSeek-only values before request construction: "off" becomes
// "low", and CodeWhale's "auto" falls back to "medium".
if let Some(raw) = request.reasoning_effort.as_deref()
&& let Some(effort) = codex_responses_reasoning_effort(raw)
{
body["reasoning"] = json!({
"effort": effort,
"summary": "auto",
});
}
// Include reasoning summaries in the stream.
body["include"] = json!(["reasoning.encrypted_content"]);
body
}
impl DeepSeekClient {
/// Handle a streaming Responses API request for the OpenAI Codex provider.
pub(super) async fn handle_responses_stream(
&self,
request: MessageRequest,
) -> Result<StreamEventBox> {
let body = self.build_responses_body(&request);
let body = build_responses_body(&request);
let url = format!("{}{}", self.base_url, CODEX_RESPONSES_PATH);
// The bearer Authorization header is already installed as a default
@@ -614,7 +613,7 @@ fn tool_to_responses_function(tool: &Tool) -> Value {
fn codex_responses_reasoning_effort(raw: &str) -> Option<&'static str> {
match raw.trim().to_ascii_lowercase().as_str() {
"off" | "disabled" | "none" | "false" => Some("low"),
"minimal" => Some("minimal"),
"minimal" => Some("low"),
"low" => Some("low"),
"high" => Some("high"),
"xhigh" | "max" | "maximum" => Some("xhigh"),
@@ -661,6 +660,7 @@ fn parse_responses_usage(val: &Value) -> Usage {
#[cfg(test)]
mod tests {
use super::*;
use crate::models::Message;
#[test]
@@ -670,10 +670,48 @@ mod tests {
assert_eq!(codex_responses_reasoning_effort("xhigh"), Some("xhigh"));
assert_eq!(codex_responses_reasoning_effort("high"), Some("high"));
assert_eq!(codex_responses_reasoning_effort("medium"), Some("medium"));
assert_eq!(codex_responses_reasoning_effort("minimal"), Some("low"));
assert_eq!(codex_responses_reasoning_effort("auto"), Some("medium"));
assert_eq!(codex_responses_reasoning_effort("off"), Some("low"));
}
#[test]
fn codex_responses_body_uses_responses_reasoning_not_deepseek_thinking() {
let request = MessageRequest {
model: "gpt-5.5".to_string(),
messages: vec![Message {
role: "user".to_string(),
content: vec![ContentBlock::Text {
text: "hello".to_string(),
cache_control: None,
}],
}],
max_tokens: 128,
system: None,
tools: None,
tool_choice: None,
metadata: None,
thinking: None,
reasoning_effort: Some("max".to_string()),
stream: None,
temperature: None,
top_p: None,
};
let body = build_responses_body(&request);
assert_eq!(
body.pointer("/reasoning/effort").and_then(Value::as_str),
Some("xhigh")
);
assert_eq!(
body.pointer("/reasoning/summary").and_then(Value::as_str),
Some("auto")
);
assert!(body.get("thinking").is_none());
assert!(body.get("reasoning_effort").is_none());
}
#[test]
fn responses_input_includes_user_role_tool_results() {
let request = MessageRequest {