Skip to content

Solution for long term memory for agent coding CLIs and to facilitate handoff between different agent vendors

Rust

fix(llm): handle gpt-6 in the openai provider

Éverton Toffanetto60646d135bec
diff --git a/CHANGELOG.md b/CHANGELOG.md
index a50c4b8d..d0e49c11 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -8,11 +8,14 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
## [Unreleased]
### Fixed
-- The `codex`, `openai-oauth` and `copilot` providers no longer send
- `temperature` to `gpt-6-*` models. Codex answered these requests with
- `400 Unsupported parameter: temperature`, so any provider or fallback
- configured with `gpt-6-luna`, `gpt-6-sol` or `gpt-6-astra` failed on every
- call. GPT-6 models now get the same treatment as `gpt-5*` and `o*`. (#851)
+- Omitted `temperature` for `gpt-6-*` models in the `codex`,
+ `openai-oauth`, `copilot` and `openai` providers. Codex answered GPT-6
+ requests that carried `temperature` with `400 Unsupported parameter:
+ temperature`, which broke consolidation, lint and bootstrap on a
+ provider or fallback configured with `gpt-6-luna`, `gpt-6-sol` or
+ `gpt-6-astra`. The `openai` provider now also sends
+ `max_completion_tokens` for GPT-6 and no longer applies its local
+ 16,384-token cap to it, matching `gpt-5*`. (#851)
- `companions/ai-memory-macos/build.sh` no longer fails on machines whose
active developer directory is Command Line Tools only: SwiftUI `@State`
needs the `SwiftUIMacros` plugin shipped with full Xcode, so the script
diff --git a/crates/ai-memory-llm/src/copilot.rs b/crates/ai-memory-llm/src/copilot.rs
index 839ad5c2..3abf872d 100644
--- a/crates/ai-memory-llm/src/copilot.rs
+++ b/crates/ai-memory-llm/src/copilot.rs
@@ -21,6 +21,7 @@ use crate::embedding::{
};
use crate::error::{LlmError, LlmResult};
use crate::openai::{STRUCTURED_OUTPUT_SCHEMA_NAME, enforce_strict_object_schemas};
+use crate::openai_oauth::model_uses_default_temperature;
use crate::provider::LlmProvider;
use crate::response::{provider_error_body, response_json_limited, response_text_limited};
use crate::text::truncate_for_embedding;
@@ -695,11 +696,6 @@ fn build_chat_request<'a>(
}
}
-fn model_uses_default_temperature(model: &str) -> bool {
- let m = model.to_ascii_lowercase();
- m.starts_with("gpt-5") || m.starts_with("gpt-6") || m.starts_with('o')
-}
-
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
enum CopilotEndpoint {
ChatCompletions,
@@ -1328,6 +1324,19 @@ mod tests {
assert!(value.get("temperature").is_none());
}
+ #[test]
+ fn responses_request_omits_gpt6_temperature() {
+ let request = ChatRequest {
+ system: None,
+ messages: vec![crate::types::ChatMessage::user("hello")],
+ temperature: Some(0.2),
+ max_tokens: 123,
+ };
+ let value =
+ serde_json::to_value(build_responses_request("gpt-6-sol", &request, None)).unwrap();
+ assert!(value.get("temperature").is_none());
+ }
+
#[test]
fn structured_request_uses_json_schema_format() {
let request = ChatRequest::user_prompt("json please");
diff --git a/crates/ai-memory-llm/src/openai.rs b/crates/ai-memory-llm/src/openai.rs
index 95d25cbc..88186923 100644
--- a/crates/ai-memory-llm/src/openai.rs
+++ b/crates/ai-memory-llm/src/openai.rs
@@ -65,7 +65,7 @@ fn last_segment_is_version(url: &str) -> bool {
///
/// `Official` targets `api.openai.com` and honours the model-family
/// rules that the real OpenAI Chat Completions endpoint enforces:
-/// `max_completion_tokens` for gpt-5 / o-series, model-family output
+/// `max_completion_tokens` for gpt-5 / gpt-6 / o-series, model-family output
/// caps, omitted `temperature` for reasoning models, strict-mode JSON
/// schema normalisation.
///
@@ -388,7 +388,7 @@ impl OpenAiProvider {
} else {
(Some(capped), None)
};
- // gpt-5 and o-series reject any non-default temperature
+ // gpt-5, gpt-6 and o-series reject any non-default temperature
// with `Unsupported value: temperature does not support
// 0.2 with this model. Only the default (1) is
// supported.` The lint / consolidate / bootstrap call
@@ -576,17 +576,21 @@ pub(crate) fn enforce_strict_object_schemas(value: &mut serde_json::Value) {
/// Models that require `max_completion_tokens` instead of `max_tokens`.
/// OpenAI introduced this rename starting with the reasoning-capable o1
-/// family and made it mandatory across the gpt-5 line. Sending the legacy
+/// family and made it mandatory across the gpt-5 and gpt-6 lines. Sending the legacy
/// `max_tokens` to these models returns a 400 with
/// `Unsupported parameter: 'max_tokens'`.
fn model_requires_max_completion_tokens(model: &str) -> bool {
let m = model.to_ascii_lowercase();
- m.starts_with("gpt-5") || m.starts_with("o1") || m.starts_with("o3") || m.starts_with("o4")
+ m.starts_with("gpt-5")
+ || m.starts_with("gpt-6")
+ || m.starts_with("o1")
+ || m.starts_with("o3")
+ || m.starts_with("o4")
}
/// Models that reject any non-default `temperature` value.
///
-/// gpt-5 and the o-series reasoning models accept only the model
+/// gpt-5, gpt-6 and the o-series reasoning models accept only the model
/// default (1.0). Any caller-supplied value — including the 0.1-0.2
/// passed by lint / bootstrap / consolidation — returns a 400:
/// `Unsupported value: 'temperature' does not support 0.2 with this
@@ -653,12 +657,12 @@ impl ReasoningHost {
/// model-specific message — at which point the operator can lower
/// `max_tokens` or switch model. The cap exists to unblock the
/// common case (gpt-4o family at 16384), not to paper over every
-/// model. Reasoning models in the gpt-5 / o-series have much larger
+/// model. Reasoning models in the gpt-5 / gpt-6 / o-series have much larger
/// caps (128K+), so we leave their requests untouched.
fn max_output_tokens_for(model: &str) -> u32 {
if model_requires_max_completion_tokens(model) {
- // gpt-5 / o-series: documented at 128K output. Leave the
- // caller's value alone — they know what they're asking for.
+ // gpt-5 / o-series: documented at 128K output; gpt-6 gets the
+ // same handling. Leave the caller's value alone — they know what they're asking for.
u32::MAX
} else {
// gpt-4o family published cap. gpt-4-turbo / gpt-3.5 have a
@@ -1043,6 +1047,7 @@ mod tests {
assert!(model_requires_max_completion_tokens("o1-mini"));
assert!(model_requires_max_completion_tokens("o3"));
assert!(model_requires_max_completion_tokens("o4-mini"));
+ assert!(model_requires_max_completion_tokens("gpt-6-luna"));
}
#[test]
@@ -1117,6 +1122,25 @@ mod tests {
);
}
+ #[test]
+ fn build_request_omits_temperature_for_gpt6() {
+ let p = provider_for("gpt-6-luna");
+ let req_input = ChatRequest {
+ system: None,
+ messages: vec![ChatMessage {
+ role: Role::User,
+ content: "x".into(),
+ }],
+ max_tokens: 64_000,
+ temperature: Some(0.2),
+ };
+ let json = serde_json::to_value(p.build_request(&req_input, None)).unwrap();
+ assert!(json.get("temperature").is_none());
+ assert!(json.get("max_tokens").is_none());
+ // Above the 16,384 cap applied to non-reasoning models.
+ assert_eq!(json["max_completion_tokens"], 64_000);
+ }
+
#[test]
fn build_request_keeps_temperature_for_gpt4() {
// gpt-4 family accepts any temperature; forwarding the
diff --git a/crates/ai-memory-llm/src/openai_oauth.rs b/crates/ai-memory-llm/src/openai_oauth.rs
index a7632709..a26e7670 100644
--- a/crates/ai-memory-llm/src/openai_oauth.rs
+++ b/crates/ai-memory-llm/src/openai_oauth.rs
@@ -428,7 +428,10 @@ pub(crate) fn parse_sse_response(body: &str) -> LlmResult<CodexResponsesResponse
Ok(response)
}
-fn model_uses_default_temperature(model: &str) -> bool {
+/// Models on the ChatGPT/Codex and Copilot backends that reject a
+/// caller-supplied `temperature`. Shared by the `openai-oauth`, `codex`
+/// and `copilot` providers.
+pub(crate) fn model_uses_default_temperature(model: &str) -> bool {
let m = model.to_ascii_lowercase();
m.starts_with("gpt-5") || m.starts_with("gpt-6") || m.starts_with('o')
}