Merge pull request #621 from voarsh2/pr/model-proxy-20260924

Normalize model metadata for OpenAI-compatible providers

Author: zemaj

code-rs/core/src/client.rs

@@ -68,7 +68,7 @@ use crate::error::RetryLimitReachedError;
 use crate::error::UnexpectedResponseError;
 use crate::error::UsageLimitReachedError;
 use crate::flags::CODEX_RS_SSE_FIXTURE;
-use crate::model_family::{find_family_for_model, ModelFamily};
+use crate::model_family::{derive_default_model_family, find_family_for_model, ModelFamily};
 use crate::model_provider_info::ModelProviderInfo;
 use crate::model_provider_info::WireApi;
 use crate::openai_tools::create_tools_json_for_responses_api;
@@ -372,6 +372,46 @@ impl ModelClient {
         }
     }
 
+    /// Resolve the model family used to build a request.
+    ///
+    /// The newer codex-rs path passes resolved `ModelInfo` into the request
+    /// builder. This older code-rs path still reconstructs families from the
+    /// slug, so keep that reconstruction provider-aware at the final boundary.
+    pub(crate) fn normalize_model_family_for_provider(
+        &self,
+        mut family: ModelFamily,
+    ) -> ModelFamily {
+        if !self.uses_first_party_openai_model_metadata() {
+            family.use_responses_lite = false;
+            family.prefer_websockets = false;
+        }
+        family
+    }
+
+    fn uses_first_party_openai_model_metadata(&self) -> bool {
+        // Provider-owned ModelInfo is the codex-rs source of truth. This
+        // compatibility check identifies the one built-in first-party path;
+        // proxy/custom endpoints must not inherit embedded OpenAI capabilities.
+        self.config.model_provider_id == "openai"
+            && self.provider.is_public_openai_responses_endpoint()
+    }
+
+    fn model_family_for_request(&self, prompt: &Prompt, model: &str) -> ModelFamily {
+        let family = prompt
+            .model_family_override
+            .clone()
+            .unwrap_or_else(|| {
+                if self.uses_first_party_openai_model_metadata() {
+                    find_family_for_model(model)
+                        .unwrap_or_else(|| self.config.model_family.clone())
+                } else {
+                    derive_default_model_family(model)
+                }
+            });
+
+        self.normalize_model_family_for_provider(family)
+    }
+
     /// Get the reasoning effort configuration
     pub fn get_reasoning_effort(&self) -> ReasoningEffortConfig {
         self.effort
@@ -778,11 +818,7 @@ impl ModelClient {
             .as_deref()
             .unwrap_or(self.config.model.as_str());
         let effective_effort = clamp_reasoning_effort_for_model(request_model, self.effort);
-        let request_family = prompt
-            .model_family_override
-            .clone()
-            .or_else(|| find_family_for_model(request_model))
-            .unwrap_or_else(|| self.config.model_family.clone());
+        let request_family = self.model_family_for_request(prompt, request_model);
         let store = should_store_responses(prompt, &self.provider, &request_family);
 
         let full_instructions = prompt.get_full_instructions(&request_family);
@@ -1265,11 +1301,7 @@ impl ModelClient {
             .as_deref()
             .unwrap_or(self.config.model.as_str());
         let effective_effort = clamp_reasoning_effort_for_model(request_model, self.effort);
-        let request_family = prompt
-            .model_family_override
-            .clone()
-            .or_else(|| find_family_for_model(request_model))
-            .unwrap_or_else(|| self.config.model_family.clone());
+        let request_family = self.model_family_for_request(prompt, request_model);
         let store = should_store_responses(prompt, &self.provider, &request_family);
 
         let full_instructions = prompt.get_full_instructions(&request_family);
@@ -2059,11 +2091,7 @@ impl ModelClient {
             .model_override
             .as_deref()
             .unwrap_or(self.config.model.as_str());
-        let family = prompt
-            .model_family_override
-            .clone()
-            .or_else(|| find_family_for_model(model_slug))
-            .unwrap_or_else(|| self.config.model_family.clone());
+        let family = self.model_family_for_request(prompt, model_slug);
         let session_id = prompt.session_id_override.unwrap_or(self.session_id);
         let session_id_str = session_id.to_string();
         let instructions = prompt.get_full_instructions(&family).into_owned();

code-rs/core/src/codex/session.rs

@@ -668,6 +668,9 @@ impl Session {
                 .model_override
                 .as_deref()
                 .unwrap_or(configured_model.as_str());
+            // The legacy path may start from embedded OpenAI metadata and then
+            // apply a remote override. Normalize after both steps so custom
+            // providers match codex-rs fallback ModelInfo behavior.
             let base_family = if let Some(family) = find_family_for_model(model_slug) {
                 family
             } else {
@@ -690,7 +693,9 @@ impl Session {
             } else {
                 base_family
             };
-            prompt.model_family_override = Some(family);
+            prompt.model_family_override = Some(
+                self.client.normalize_model_family_for_provider(family),
+            );
         }
         used_fallback_model_metadata
     }

code-rs/core/src/codex/streaming.rs

@@ -764,13 +764,26 @@ pub(super) async fn submission_loop(
                 tools_config.set_agent_models(agent_models);
 
                 let model_descriptions = model_guide_markdown_with_custom(&config.agents);
-                let remote_models_manager = auth_manager.as_ref().map(|mgr| {
-                    Arc::new(RemoteModelsManager::new(
-                        Arc::clone(mgr),
-                        provider.clone(),
-                        config.code_home.clone(),
-                    ))
-                });
+                // codex-rs gives each provider ownership of its model manager.
+                // code-rs still has one legacy RemoteModelsManager, so only
+                // wire it for the equivalent first-party/command-auth cases;
+                // ordinary OpenAI-compatible providers must not be treated as
+                // if their /models endpoint returned Codex ModelInfo records.
+                let should_create_remote_models_manager =
+                    (config.model_provider_id == "openai"
+                        && provider.is_public_openai_responses_endpoint())
+                        || provider.has_command_auth();
+                let remote_models_manager = should_create_remote_models_manager
+                    .then(|| {
+                        auth_manager.as_ref().map(|mgr| {
+                            Arc::new(RemoteModelsManager::new(
+                                Arc::clone(mgr),
+                                provider.clone(),
+                                config.code_home.clone(),
+                            ))
+                        })
+                    })
+                    .flatten();
                 if let Some(remote) = remote_models_manager.as_ref() {
                     let remote = Arc::clone(remote);
                     tokio::spawn(async move {
@@ -1481,8 +1494,10 @@ async fn spawn_review_thread(
 
     // Determine model + family for review mode.
     let review_model = config.review_model.clone();
-    let review_family = find_family_for_model(&review_model)
-        .unwrap_or_else(|| derive_default_model_family(&review_model));
+    let review_family = sess.client().normalize_model_family_for_provider(
+        find_family_for_model(&review_model)
+            .unwrap_or_else(|| derive_default_model_family(&review_model)),
+    );
 
     // Prepare a per-review configuration that favors deterministic feedback.
     let mut review_config = (*config).clone();
@@ -2183,7 +2198,10 @@ async fn maybe_run_auto_context_compaction(
     let mut raw_decision: Option<String> = None;
     for model in auto_context_judge_models() {
         prompt.model_override = Some(model.to_string());
-        prompt.model_family_override = Some(derive_default_model_family(model));
+        prompt.model_family_override = Some(
+            sess.client()
+                .normalize_model_family_for_provider(derive_default_model_family(model)),
+        );
 
         match tokio::time::timeout(
             std::time::Duration::from_secs(12),
@@ -3145,7 +3163,10 @@ async fn run_turn(
                     .unwrap_or_else(|| derive_default_model_family(&override_model))
             };
             prompt.model_override = Some(override_model);
-            prompt.model_family_override = Some(override_family);
+            prompt.model_family_override = Some(
+                tc.client
+                    .normalize_model_family_for_provider(override_family),
+            );
         }
 
         if used_fallback_model_metadata

code-rs/core/src/config.rs

@@ -1353,8 +1353,17 @@ impl Config {
             .or(cfg.context_mode)
             .or(Some(ContextMode::Auto));
 
-        let model_family =
+        let mut model_family =
             find_family_for_model(&model).unwrap_or_else(|| derive_default_model_family(&model));
+        if model_provider_id != "openai"
+            || !model_provider.is_public_openai_responses_endpoint()
+        {
+            // codex-rs resolves custom providers through their provider-owned
+            // ModelInfo manager; unknown custom-provider models use fallback
+            // metadata. Legacy code-rs starts from its embedded OpenAI catalog,
+            // so remove OpenAI-only Lite semantics at this compatibility seam.
+            model_family.use_responses_lite = false;
+        }
         let default_tool_output_max_bytes = model_family.tool_output_max_bytes();
 
         // Chat model reasoning effort (used when other flows follow the chat model).
@@ -2718,6 +2727,72 @@ model_verbosity = "high"
         Ok(())
     }
 
+    #[test]
+    fn model_metadata_uses_provider_appropriate_catalog_boundary() -> std::io::Result<()> {
+        let fixture = create_test_fixture()?;
+
+        let mut custom_cfg = fixture.cfg.clone();
+        custom_cfg
+            .profiles
+            .get_mut("gpt3")
+            .expect("gpt3 profile should exist")
+            .model = Some("gpt-5.6-luna".to_string());
+        let custom_config = Config::load_from_base_config_with_overrides(
+            custom_cfg,
+            ConfigOverrides {
+                config_profile: Some("gpt3".to_string()),
+                cwd: Some(fixture.cwd()),
+                ..Default::default()
+            },
+            fixture.code_home(),
+        )?;
+
+        assert!(!custom_config.model_family.use_responses_lite);
+        assert_eq!(
+            Some(crate::model_family::EXTENDED_CONTEXT_WINDOW_1M),
+            custom_config.model_context_window
+        );
+
+        let mut openai_cfg = fixture.cfg.clone();
+        openai_cfg
+            .profiles
+            .get_mut("o3")
+            .expect("o3 profile should exist")
+            .model = Some("gpt-5.6-luna".to_string());
+        let openai_config = Config::load_from_base_config_with_overrides(
+            openai_cfg,
+            ConfigOverrides {
+                config_profile: Some("o3".to_string()),
+                cwd: Some(fixture.cwd()),
+                ..Default::default()
+            },
+            fixture.code_home(),
+        )?;
+
+        assert!(openai_config.model_family.use_responses_lite);
+
+        let mut proxied_cfg = fixture.cfg.clone();
+        proxied_cfg.openai_base_url = Some("https://proxy.example.test/v1".to_string());
+        proxied_cfg
+            .profiles
+            .get_mut("o3")
+            .expect("o3 profile should exist")
+            .model = Some("gpt-5.6-luna".to_string());
+        let proxied_config = Config::load_from_base_config_with_overrides(
+            proxied_cfg,
+            ConfigOverrides {
+                config_profile: Some("o3".to_string()),
+                cwd: Some(fixture.cwd()),
+                ..Default::default()
+            },
+            fixture.code_home(),
+        )?;
+
+        assert!(!proxied_config.model_family.use_responses_lite);
+
+        Ok(())
+    }
+
     #[test]
     fn test_precedence_fixture_with_zdr_profile() -> std::io::Result<()> {
         let fixture = create_test_fixture()?;

code-rs/core/src/memories/mod.rs

@@ -557,11 +557,12 @@ async fn collect_json_response(session: &Session, prompt: &Prompt) -> anyhow::Re
 
 async fn resolve_model_family(session: &Session, model: &str) -> crate::model_family::ModelFamily {
     let base = find_family_for_model(model).unwrap_or_else(|| derive_default_model_family(model));
-    if let Some(remote) = session.remote_models_manager() {
+    let family = if let Some(remote) = session.remote_models_manager() {
         remote.apply_remote_overrides(model, base).await
     } else {
         base
-    }
+    };
+    session.client().normalize_model_family_for_provider(family)
 }
 
 fn build_stage1_user_prompt(

code-rs/core/src/model_family.rs

@@ -552,6 +552,7 @@ mod tests {
 
     use super::find_family_for_model;
     use super::parse_upstream_models;
+    use super::supports_extended_context;
 
     #[test]
     fn image_generation_support_tracks_image_input_modality() {
@@ -591,6 +592,16 @@ mod tests {
         assert!(!models.is_empty());
     }
 
+    #[test]
+    fn extended_context_support_includes_supported_gpt_5_6_variants() {
+        for model in ["gpt-5.4", "gpt-5.6-luna", "gpt-5.6-sol"] {
+            assert!(supports_extended_context(model), "{model} should support 1M context");
+        }
+
+        assert!(!supports_extended_context("gpt-5.6-terra"));
+        assert!(!supports_extended_context("gpt-5.5"));
+    }
+
     #[test]
     fn qwen3_coder_30b_quantized_slug_has_explicit_local_metadata() {
         let family = find_family_for_model("Qwen3-Coder-30B-A3B-Instruct-Q4_K_M")
@@ -636,7 +647,9 @@ pub const fn default_auto_compact_limit_for_context_window(context_window: u64)
 }
 
 pub fn supports_extended_context(model: &str) -> bool {
-    model.eq_ignore_ascii_case("gpt-5.4")
+    ["gpt-5.4", "gpt-5.6-luna", "gpt-5.6-sol"]
+        .iter()
+        .any(|supported| model.eq_ignore_ascii_case(supported))
 }
 
 pub fn resolve_context_mode_limits(

code-rs/core/src/remote_models/mod.rs

@@ -47,6 +47,10 @@ struct RemoteModelsState {
 ///
 /// Any error (disk, auth, network, parse) results in an empty remote model list
 /// so callers can safely fall back to built-in behaviour.
+///
+/// This is a compatibility layer for the older code-rs architecture. In
+/// codex-rs, the selected provider owns the model manager and decides whether
+/// `/models` is authoritative; the refresh gate below preserves that policy.
 #[derive(Debug)]
 pub struct RemoteModelsManager {
     state: RwLock<RemoteModelsState>,
@@ -174,9 +178,13 @@ impl RemoteModelsManager {
         identity: Option<String>,
         stale_etag: Option<String>,
     ) {
-        let auth_mode = auth.as_ref().map(|a| a.mode);
-        if !auth_mode.is_some_and(AuthMode::is_chatgpt) {
-            // Only the ChatGPT backend exposes the Codex `/models` schema.
+        let should_refresh = auth
+            .as_ref()
+            .is_some_and(CodexAuth::uses_codex_backend)
+            || self.provider.has_command_auth();
+        if !should_refresh {
+            // Match codex-rs: refresh provider-owned model catalogs only for
+            // Codex-backend auth or command-backed provider auth.
             return;
         }