feat: make reasoning effort/summaries configurable (#1199)
Previous to this PR, we always set `reasoning` when making a request using the Responses API:d7245cbbc9/codex-rs/core/src/client.rs (L108-L111)Though if you tried to use the Rust CLI with `--model gpt-4.1`, this would fail with: ```shell "Unsupported parameter: 'reasoning.effort' is not supported with this model." ``` We take a cue from the TypeScript CLI, which does a check on the model name:d7245cbbc9/codex-cli/src/utils/agent/agent-loop.ts (L786-L789)This PR does a similar check, though also adds support for the following config options: ``` model_reasoning_effort = "low" | "medium" | "high" | "none" model_reasoning_summary = "auto" | "concise" | "detailed" | "none" ``` This way, if you have a model whose name happens to start with `"o"` (or `"codex"`?), you can set these to `"none"` to explicitly disable reasoning, if necessary. (That said, it seems unlikely anyone would use the Responses API with non-OpenAI models, but we provide an escape hatch, anyway.) This PR also updates both the TUI and `codex exec` to show `reasoning effort` and `reasoning summaries` in the header.
This commit is contained in:
@@ -1,3 +1,5 @@
|
||||
use crate::config_types::ReasoningEffort as ReasoningEffortConfig;
|
||||
use crate::config_types::ReasoningSummary as ReasoningSummaryConfig;
|
||||
use crate::error::Result;
|
||||
use crate::models::ResponseItem;
|
||||
use futures::Stream;
|
||||
@@ -52,25 +54,59 @@ pub enum ResponseEvent {
|
||||
|
||||
#[derive(Debug, Serialize)]
|
||||
pub(crate) struct Reasoning {
|
||||
pub(crate) effort: &'static str,
|
||||
pub(crate) effort: OpenAiReasoningEffort,
|
||||
#[serde(skip_serializing_if = "Option::is_none")]
|
||||
pub(crate) summary: Option<Summary>,
|
||||
pub(crate) summary: Option<OpenAiReasoningSummary>,
|
||||
}
|
||||
|
||||
/// See https://platform.openai.com/docs/guides/reasoning?api-mode=responses#get-started-with-reasoning
|
||||
#[derive(Debug, Serialize, Default, Clone, Copy)]
|
||||
#[serde(rename_all = "lowercase")]
|
||||
pub(crate) enum OpenAiReasoningEffort {
|
||||
Low,
|
||||
#[default]
|
||||
Medium,
|
||||
High,
|
||||
}
|
||||
|
||||
impl From<ReasoningEffortConfig> for Option<OpenAiReasoningEffort> {
|
||||
fn from(effort: ReasoningEffortConfig) -> Self {
|
||||
match effort {
|
||||
ReasoningEffortConfig::Low => Some(OpenAiReasoningEffort::Low),
|
||||
ReasoningEffortConfig::Medium => Some(OpenAiReasoningEffort::Medium),
|
||||
ReasoningEffortConfig::High => Some(OpenAiReasoningEffort::High),
|
||||
ReasoningEffortConfig::None => None,
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
/// A summary of the reasoning performed by the model. This can be useful for
|
||||
/// debugging and understanding the model's reasoning process.
|
||||
#[derive(Debug, Serialize)]
|
||||
/// See https://platform.openai.com/docs/guides/reasoning?api-mode=responses#reasoning-summaries
|
||||
#[derive(Debug, Serialize, Default, Clone, Copy)]
|
||||
#[serde(rename_all = "lowercase")]
|
||||
pub(crate) enum Summary {
|
||||
pub(crate) enum OpenAiReasoningSummary {
|
||||
#[default]
|
||||
Auto,
|
||||
#[allow(dead_code)] // Will go away once this is configurable.
|
||||
Concise,
|
||||
#[allow(dead_code)] // Will go away once this is configurable.
|
||||
Detailed,
|
||||
}
|
||||
|
||||
impl From<ReasoningSummaryConfig> for Option<OpenAiReasoningSummary> {
|
||||
fn from(summary: ReasoningSummaryConfig) -> Self {
|
||||
match summary {
|
||||
ReasoningSummaryConfig::Auto => Some(OpenAiReasoningSummary::Auto),
|
||||
ReasoningSummaryConfig::Concise => Some(OpenAiReasoningSummary::Concise),
|
||||
ReasoningSummaryConfig::Detailed => Some(OpenAiReasoningSummary::Detailed),
|
||||
ReasoningSummaryConfig::None => None,
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
/// Request object that is serialized as JSON and POST'ed when using the
|
||||
/// Responses API.
|
||||
#[derive(Debug, Serialize)]
|
||||
pub(crate) struct Payload<'a> {
|
||||
pub(crate) struct ResponsesApiRequest<'a> {
|
||||
pub(crate) model: &'a str,
|
||||
pub(crate) instructions: &'a str,
|
||||
// TODO(mbolin): ResponseItem::Other should not be serialized. Currently,
|
||||
@@ -88,6 +124,40 @@ pub(crate) struct Payload<'a> {
|
||||
pub(crate) stream: bool,
|
||||
}
|
||||
|
||||
pub(crate) fn create_reasoning_param_for_request(
|
||||
model: &str,
|
||||
effort: ReasoningEffortConfig,
|
||||
summary: ReasoningSummaryConfig,
|
||||
) -> Option<Reasoning> {
|
||||
let effort: Option<OpenAiReasoningEffort> = effort.into();
|
||||
let effort = effort?;
|
||||
|
||||
if model_supports_reasoning_summaries(model) {
|
||||
Some(Reasoning {
|
||||
effort,
|
||||
summary: summary.into(),
|
||||
})
|
||||
} else {
|
||||
None
|
||||
}
|
||||
}
|
||||
|
||||
pub fn model_supports_reasoning_summaries(model: &str) -> bool {
|
||||
// Currently, we hardcode this rule to decide whether enable reasoning.
|
||||
// We expect reasoning to apply only to OpenAI models, but we do not want
|
||||
// users to have to mess with their config to disable reasoning for models
|
||||
// that do not support it, such as `gpt-4.1`.
|
||||
//
|
||||
// Though if a user is using Codex with non-OpenAI models that, say, happen
|
||||
// to start with "o", then they can set `model_reasoning_effort = "none` in
|
||||
// config.toml to disable reasoning.
|
||||
//
|
||||
// Ultimately, this should also be configurable in config.toml, but we
|
||||
// need to have defaults that "just work." Perhaps we could have a
|
||||
// "reasoning models pattern" as part of ModelProviderInfo?
|
||||
model.starts_with("o") || model.starts_with("codex")
|
||||
}
|
||||
|
||||
pub(crate) struct ResponseStream {
|
||||
pub(crate) rx_event: mpsc::Receiver<Result<ResponseEvent>>,
|
||||
}
|
||||
|
||||
Reference in New Issue
Block a user