codex/codex-rs/models-manager/src/manager.rs

use super::cache::FileModelsCache;
use crate::cache::ModelsCache;
use crate::cache::ModelsCacheEntry;
use crate::collaboration_mode_presets::builtin_collaboration_mode_presets;
use crate::config::ModelsManagerConfig;
use crate::model_info;
use chrono::Utc;
use codex_http_client::HttpClientFactory;
use codex_login::AuthManager;
use codex_protocol::auth::AuthMode;
use codex_protocol::config_types::CollaborationModeMask;
use codex_protocol::error::Result as CoreResult;
use codex_protocol::openai_models::ModelInfo;
use codex_protocol::openai_models::ModelPreset;
use codex_protocol::openai_models::ModelVisibility;
use codex_protocol::openai_models::ModelsResponse;
use std::fmt;
use std::future::Future;
use std::path::PathBuf;
use std::pin::Pin;
use std::sync::Arc;
use std::time::Duration;
use tokio::sync::RwLock;
use tokio::sync::TryLockError;
use tracing::Instrument as _;
use tracing::error;
use tracing::info;

const MODEL_CACHE_FILE: &str = "models_cache.json";
const DEFAULT_MODEL_CACHE_TTL: Duration = Duration::from_secs(300);

/// Remote endpoint used by the OpenAI-compatible model manager.
///
/// Implementations own provider-specific auth and transport details. The model
/// manager owns refresh policy, cache behavior, and catalog merging; it calls
/// this endpoint only when it decides a remote refresh should happen.
pub trait ModelsEndpointClient: fmt::Debug + Send + Sync {
    /// Opaque identity of the current provider and credentials, without resolving auth.
    ///
    /// Must change when account, email, plan, provider, or API credentials change.
    /// Return `None` when identity is unavailable; cached catalog reuse is then disabled.
    fn identity(&self) -> Option<String>;

    /// Returns whether this provider can authenticate command-scoped requests.
    fn has_command_auth(&self) -> bool;

    /// Returns whether the currently resolved auth can use Codex backend-only models.
    fn uses_codex_backend(&self) -> ModelsEndpointFuture<'_, bool>;

    /// Fetches the latest remote model catalog and optional ETag.
    fn list_models<'a>(
        &'a self,
        client_version: &'a str,
        http_client_factory: HttpClientFactory,
    ) -> ModelsEndpointFuture<'a, CoreResult<ModelsEndpointResponse>>;
}

/// Catalog and validation metadata captured using the credentials for one request.
#[derive(Debug)]
pub struct ModelsEndpointResponse {
    pub models: Vec<ModelInfo>,
    pub etag: Option<String>,
    /// Identity of the credentials actually used to fetch this catalog.
    /// Successful responses must always identify their provider and authentication scope.
    pub identity: String,
}

pub type ModelsEndpointFuture<'a, T> = Pin<Box<dyn Future<Output = T> + Send + 'a>>;

/// Strategy for refreshing available models.
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
pub enum RefreshStrategy {
    /// Always fetch from the network, ignoring cache.
    Online,
    /// Only use cached data, never fetch from the network.
    Offline,
    /// Use cache if available and fresh, otherwise fetch from the network.
    OnlineIfUncached,
}

impl RefreshStrategy {
    const fn as_str(self) -> &'static str {
        match self {
            Self::Online => "online",
            Self::Offline => "offline",
            Self::OnlineIfUncached => "online_if_uncached",
        }
    }
}

impl fmt::Display for RefreshStrategy {
    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
        f.write_str(self.as_str())
    }
}

type SharedModelsEndpointClient = Arc<dyn ModelsEndpointClient>;

/// Coordinates model discovery plus cached metadata on disk.
pub trait ModelsManager: fmt::Debug + Send + Sync {
    /// List all available models, refreshing according to the specified strategy.
    ///
    /// Returns model presets sorted by priority and filtered by auth mode and visibility.
    fn list_models(
        &self,
        refresh_strategy: RefreshStrategy,
        http_client_factory: HttpClientFactory,
    ) -> ModelsManagerFuture<'_, Vec<ModelPreset>> {
        Box::pin(
            async move {
                let catalog = self
                    .raw_model_catalog(refresh_strategy, http_client_factory)
                    .await;
                self.build_available_models(catalog.models)
            }
            .instrument(tracing::info_span!(
                "list_models",
                refresh_strategy = %refresh_strategy
            )),
        )
    }

    /// Return the active raw model catalog, refreshing according to the specified strategy.
    fn raw_model_catalog(
        &self,
        refresh_strategy: RefreshStrategy,
        http_client_factory: HttpClientFactory,
    ) -> ModelsManagerFuture<'_, ModelsResponse>;

    /// Return the current in-memory remote model catalog without refreshing or loading cache state.
    fn get_remote_models(&self) -> ModelsManagerFuture<'_, Vec<ModelInfo>>;

    /// Attempt to return the current in-memory remote model catalog without blocking.
    ///
    /// Returns an error if the internal lock cannot be acquired.
    fn try_get_remote_models(&self) -> Result<Vec<ModelInfo>, TryLockError>;

    /// Return the auth manager used for picker filtering.
    fn auth_manager(&self) -> Option<&AuthManager>;

    /// Build picker-ready presets from the active catalog snapshot.
    fn build_available_models(&self, mut remote_models: Vec<ModelInfo>) -> Vec<ModelPreset> {
        remote_models.sort_by_key(|model| model.priority);

        let mut presets: Vec<ModelPreset> = remote_models.into_iter().map(Into::into).collect();
        let uses_codex_backend = self
            .auth_manager()
            .is_some_and(AuthManager::current_auth_uses_codex_backend);
        presets = ModelPreset::filter_by_auth(presets, uses_codex_backend);

        ModelPreset::mark_default_by_picker_visibility(&mut presets);

        presets
    }

    /// List collaboration mode presets.
    ///
    /// Returns a static set of presets seeded with the configured model.
    fn list_collaboration_modes(&self) -> Vec<CollaborationModeMask>;

    /// Attempt to list models without blocking, using the current cached state.
    ///
    /// Returns an error if the internal lock cannot be acquired.
    fn try_list_models(&self) -> Result<Vec<ModelPreset>, TryLockError> {
        let remote_models = self.try_get_remote_models()?;
        Ok(self.build_available_models(remote_models))
    }

    // todo(aibrahim): should be visible to core only and sent on session_configured event
    /// Get the model identifier to use, refreshing according to the specified strategy.
    ///
    /// If `model` is provided, preserves it unless the implementation supports and the policy
    /// allows provider fallback. Otherwise selects the default based on auth mode and available
    /// models.
    fn get_default_model<'a>(
        &'a self,
        model: &'a Option<String>,
        allow_provider_model_fallback: bool,
        refresh_strategy: RefreshStrategy,
        http_client_factory: HttpClientFactory,
    ) -> ModelsManagerFuture<'a, String> {
        Box::pin(
            async move {
                if let Some(model) = model.as_ref() {
                    return model.to_string();
                }
                default_model_from_available(
                    self.list_models(refresh_strategy, http_client_factory)
                        .await,
                )
            }
            .instrument(tracing::info_span!(
                "get_default_model",
                model.provided = model.is_some(),
                allow_provider_model_fallback,
                refresh_strategy = %refresh_strategy
            )),
        )
    }

    // todo(aibrahim): look if we can tighten it to pub(crate)
    /// Look up model metadata, applying remote overrides and config adjustments.
    fn get_model_info<'a>(
        &'a self,
        model: &'a str,
        config: &'a ModelsManagerConfig,
    ) -> ModelsManagerFuture<'a, ModelInfo> {
        Box::pin(
            async move {
                let remote_models = self.get_remote_models().await;
                construct_model_info_from_candidates(model, &remote_models, config)
            }
            .instrument(tracing::info_span!("get_model_info", model = model)),
        )
    }

    /// Refresh models if the provided ETag differs from the cached ETag.
    ///
    /// Uses `Online` strategy to fetch latest models when ETags differ.
    fn refresh_if_new_etag(
        &self,
        etag: String,
        http_client_factory: HttpClientFactory,
    ) -> ModelsManagerFuture<'_, ()>;
}

pub type ModelsManagerFuture<'a, T> = Pin<Box<dyn Future<Output = T> + Send + 'a>>;

/// Shared model manager handle used across runtime services.
pub type SharedModelsManager = Arc<dyn ModelsManager>;

/// OpenAI-compatible model manager backed by bundled models, cache, and `/models`.
#[derive(Debug)]
pub struct OpenAiModelsManager {
    remote_models: RwLock<ModelsCacheEntry>,
    cache: Option<Arc<dyn ModelsCache>>,
    endpoint_client: SharedModelsEndpointClient,
    auth_manager: Option<Arc<AuthManager>>,
}

/// Static model manager backed by an authoritative in-process catalog.
#[derive(Debug)]
pub struct StaticModelsManager {
    remote_models: Vec<ModelInfo>,
    auth_manager: Option<Arc<AuthManager>>,
}

impl OpenAiModelsManager {
    /// Construct an OpenAI-compatible remote model manager.
    pub fn new(
        codex_home: PathBuf,
        endpoint_client: Arc<dyn ModelsEndpointClient>,
        auth_manager: Option<Arc<AuthManager>>,
    ) -> Self {
        let cache_path = codex_home.join(MODEL_CACHE_FILE);
        Self::new_with_optional_cache(
            Some(Arc::new(FileModelsCache::new(
                cache_path,
                DEFAULT_MODEL_CACHE_TTL,
            ))),
            endpoint_client,
            auth_manager,
        )
    }

    /// Construct an OpenAI-compatible model manager with caching disabled.
    pub fn new_without_cache(
        endpoint_client: Arc<dyn ModelsEndpointClient>,
        auth_manager: Option<Arc<AuthManager>>,
    ) -> Self {
        Self::new_with_optional_cache(/*cache*/ None, endpoint_client, auth_manager)
    }

    /// Constructs an OpenAI-compatible model manager with a caller-provided cache.
    ///
    /// The cache is consulted by cache-aware refresh strategies. Cache misses and backend errors
    /// fall back to the models endpoint, and cache write failures do not fail model discovery.
    pub fn new_with_cache(
        cache: Arc<dyn ModelsCache>,
        endpoint_client: Arc<dyn ModelsEndpointClient>,
        auth_manager: Option<Arc<AuthManager>>,
    ) -> Self {
        Self::new_with_optional_cache(Some(cache), endpoint_client, auth_manager)
    }

    fn new_with_optional_cache(
        cache: Option<Arc<dyn ModelsCache>>,
        endpoint_client: Arc<dyn ModelsEndpointClient>,
        auth_manager: Option<Arc<AuthManager>>,
    ) -> Self {
        let remote_models = load_remote_models_from_file().unwrap_or_default();
        Self {
            remote_models: RwLock::new(ModelsCacheEntry {
                fetched_at: Utc::now(),
                etag: None,
                client_version: Some(crate::client_version_to_whole()),
                identity: endpoint_client.identity(),
                models: remote_models,
            }),
            cache,
            endpoint_client,
            auth_manager,
        }
    }
}

impl StaticModelsManager {
    /// Construct a static model manager from an authoritative catalog.
    pub fn new(auth_manager: Option<Arc<AuthManager>>, model_catalog: ModelsResponse) -> Self {
        Self {
            remote_models: model_catalog.models,
            auth_manager,
        }
    }
}

impl ModelsManager for OpenAiModelsManager {
    fn raw_model_catalog(
        &self,
        refresh_strategy: RefreshStrategy,
        http_client_factory: HttpClientFactory,
    ) -> ModelsManagerFuture<'_, ModelsResponse> {
        Box::pin(OpenAiModelsManager::raw_model_catalog(
            self,
            refresh_strategy,
            http_client_factory,
        ))
    }

    fn get_remote_models(&self) -> ModelsManagerFuture<'_, Vec<ModelInfo>> {
        Box::pin(async move {
            let entry = self.remote_models.read().await;
            if entry.identity.is_some() && entry.identity == self.endpoint_client.identity() {
                entry.models.clone()
            } else {
                load_remote_models_from_file().unwrap_or_default()
            }
        })
    }

    fn try_get_remote_models(&self) -> Result<Vec<ModelInfo>, TryLockError> {
        let entry = self.remote_models.try_read()?;
        Ok(
            if entry.identity.is_some() && entry.identity == self.endpoint_client.identity() {
                entry.models.clone()
            } else {
                load_remote_models_from_file().unwrap_or_default()
            },
        )
    }

    fn auth_manager(&self) -> Option<&AuthManager> {
        self.auth_manager.as_deref()
    }

    fn list_collaboration_modes(&self) -> Vec<CollaborationModeMask> {
        builtin_collaboration_mode_presets()
    }

    fn refresh_if_new_etag(
        &self,
        etag: String,
        http_client_factory: HttpClientFactory,
    ) -> ModelsManagerFuture<'_, ()> {
        Box::pin(OpenAiModelsManager::refresh_if_new_etag(
            self,
            etag,
            http_client_factory,
        ))
    }
}

impl OpenAiModelsManager {
    async fn raw_model_catalog(
        &self,
        refresh_strategy: RefreshStrategy,
        http_client_factory: HttpClientFactory,
    ) -> ModelsResponse {
        if let Err(err) = self
            .refresh_available_models(refresh_strategy, &http_client_factory)
            .await
        {
            error!("failed to refresh available models: {err}");
        }
        ModelsResponse {
            models: self.get_remote_models().await,
        }
    }

    async fn refresh_if_new_etag(&self, etag: String, http_client_factory: HttpClientFactory) {
        let (identity, current_etag) = {
            let entry = self.remote_models.read().await;
            (entry.identity.clone(), entry.etag.clone())
        };
        if let Some(identity) = identity
            && Some(&identity) == self.endpoint_client.identity().as_ref()
            && current_etag.as_deref() == Some(etag.as_str())
        {
            if let Some(cache) = self.cache.as_ref()
                && let Err(err) = cache
                    .refresh_ttl(&crate::client_version_to_whole(), &identity, &etag)
                    .await
            {
                error!("failed to renew cache TTL: {err}");
            }
            return;
        }
        if let Err(err) = self
            .refresh_available_models(RefreshStrategy::Online, &http_client_factory)
            .await
        {
            error!("failed to refresh available models: {err}");
        }
    }

    /// Refresh available models according to the specified strategy.
    async fn refresh_available_models(
        &self,
        refresh_strategy: RefreshStrategy,
        http_client_factory: &HttpClientFactory,
    ) -> CoreResult<()> {
        if !self.should_refresh_models().await {
            if matches!(
                refresh_strategy,
                RefreshStrategy::Offline | RefreshStrategy::OnlineIfUncached
            ) {
                self.try_load_cache().await;
            }
            return Ok(());
        }
        match refresh_strategy {
            RefreshStrategy::Offline => {
                self.try_load_cache().await;
                Ok(())
            }
            RefreshStrategy::OnlineIfUncached => {
                if self.try_load_cache().await {
                    return Ok(());
                }
                self.fetch_and_update_models(http_client_factory).await
            }
            RefreshStrategy::Online => self.fetch_and_update_models(http_client_factory).await,
        }
    }

    async fn fetch_and_update_models(
        &self,
        http_client_factory: &HttpClientFactory,
    ) -> CoreResult<()> {
        let client_version = crate::client_version_to_whole();
        let ModelsEndpointResponse {
            models,
            etag,
            identity,
        } = self
            .endpoint_client
            .list_models(&client_version, http_client_factory.clone())
            .await?;
        if Some(&identity) != self.endpoint_client.identity().as_ref() {
            return Ok(());
        }
        let entry = ModelsCacheEntry {
            fetched_at: Utc::now(),
            etag,
            client_version: Some(client_version),
            identity: Some(identity),
            models,
        };
        if let Some(cache) = self.cache.as_ref()
            && let Err(err) = cache.store(&entry).await
        {
            error!("failed to write models cache: {err}");
        }
        self.apply_remote_models(entry).await;
        Ok(())
    }

    async fn should_refresh_models(&self) -> bool {
        self.endpoint_client.uses_codex_backend().await || self.endpoint_client.has_command_auth()
    }

    /// Publish only while the request identity still matches, including after async storage.
    async fn apply_remote_models(&self, mut entry: ModelsCacheEntry) -> bool {
        let mut current = self.remote_models.write().await;
        if entry.identity != self.endpoint_client.identity() {
            return false;
        }
        // Visible ChatGPT catalogs are authoritative; other catalogs augment bundled metadata.
        let remote_only = entry
            .models
            .iter()
            .any(|model| model.visibility == ModelVisibility::List)
            && self.auth_manager.as_ref().is_some_and(|auth_manager| {
                auth_manager
                    .auth_mode()
                    .is_some_and(AuthMode::has_chatgpt_account)
            });
        if !remote_only {
            let mut models = load_remote_models_from_file().unwrap_or_default();
            for model in entry.models {
                if let Some(index) = models
                    .iter()
                    .position(|existing| existing.slug == model.slug)
                {
                    models[index] = model;
                } else {
                    models.push(model);
                }
            }
            entry.models = models;
        }
        *current = entry;
        true
    }

    /// Attempt to satisfy the refresh from the cache when it matches the provider and TTL.
    async fn try_load_cache(&self) -> bool {
        let Some(cache) = self.cache.as_ref() else {
            return false;
        };
        let _timer =
            codex_otel::start_global_timer("codex.remote_models.load_cache.duration_ms", &[]);
        let client_version = crate::client_version_to_whole();
        info!(client_version, "models cache: evaluating cache eligibility");
        let Some(identity) = self.endpoint_client.identity() else {
            return false;
        };
        let cache_entry = match cache.load(&client_version).await {
            Ok(Some(cache_entry)) => cache_entry,
            Ok(None) => {
                info!("models cache: no usable cache entry");
                return false;
            }
            Err(err) => {
                error!("failed to load models cache: {err}");
                return false;
            }
        };
        if cache_entry.client_version.as_deref() != Some(client_version.as_str()) {
            info!(
                expected_version = client_version,
                cached_version = ?cache_entry.client_version,
                "models cache: cache version mismatch"
            );
            return false;
        }
        if cache_entry.identity.as_ref() != Some(&identity)
            || self.endpoint_client.identity().as_ref() != Some(&identity)
        {
            info!("models cache: provider or auth identity mismatch");
            return false;
        }
        self.apply_remote_models(cache_entry).await
    }
}

impl ModelsManager for StaticModelsManager {
    fn get_default_model<'a>(
        &'a self,
        model: &'a Option<String>,
        allow_provider_model_fallback: bool,
        refresh_strategy: RefreshStrategy,
        http_client_factory: HttpClientFactory,
    ) -> ModelsManagerFuture<'a, String> {
        Box::pin(
            async move {
                let available_models = self
                    .list_models(refresh_strategy, http_client_factory)
                    .await;
                let requested_model = model.as_deref();

                if allow_provider_model_fallback {
                    if requested_model_is_available(requested_model, &available_models)
                        && let Some(requested_model) = requested_model
                    {
                        return requested_model.to_string();
                    }
                    return default_model_from_available(available_models);
                }

                model
                    .clone()
                    .unwrap_or_else(|| default_model_from_available(available_models))
            }
            .instrument(tracing::info_span!(
                "get_default_model",
                model.provided = model.is_some(),
                allow_provider_model_fallback,
                refresh_strategy = %refresh_strategy
            )),
        )
    }

    fn raw_model_catalog(
        &self,
        _refresh_strategy: RefreshStrategy,
        _http_client_factory: HttpClientFactory,
    ) -> ModelsManagerFuture<'_, ModelsResponse> {
        Box::pin(async move {
            ModelsResponse {
                models: self.get_remote_models().await,
            }
        })
    }

    fn get_remote_models(&self) -> ModelsManagerFuture<'_, Vec<ModelInfo>> {
        Box::pin(async { self.remote_models.clone() })
    }

    fn try_get_remote_models(&self) -> Result<Vec<ModelInfo>, TryLockError> {
        Ok(self.remote_models.clone())
    }

    fn auth_manager(&self) -> Option<&AuthManager> {
        self.auth_manager.as_deref()
    }

    fn list_collaboration_modes(&self) -> Vec<CollaborationModeMask> {
        builtin_collaboration_mode_presets()
    }

    fn refresh_if_new_etag(
        &self,
        _etag: String,
        _http_client_factory: HttpClientFactory,
    ) -> ModelsManagerFuture<'_, ()> {
        Box::pin(async {})
    }
}

fn load_remote_models_from_file() -> Result<Vec<ModelInfo>, std::io::Error> {
    Ok(crate::bundled_models_response()?.models)
}

fn default_model_from_available(available: Vec<ModelPreset>) -> String {
    available
        .iter()
        .find(|model| model.is_default)
        .or_else(|| available.first())
        .map(|model| model.model.clone())
        .unwrap_or_default()
}

fn requested_model_is_available(
    requested_model: Option<&str>,
    available_models: &[ModelPreset],
) -> bool {
    requested_model.is_some_and(|requested_model| {
        available_models
            .iter()
            .any(|available_model| available_model.model == requested_model)
    })
}

fn find_model_by_longest_prefix(model: &str, candidates: &[ModelInfo]) -> Option<ModelInfo> {
    let mut best: Option<ModelInfo> = None;
    for candidate in candidates {
        if !model.starts_with(&candidate.slug) {
            continue;
        }
        let is_better_match = if let Some(current) = best.as_ref() {
            candidate.slug.len() > current.slug.len()
        } else {
            true
        };
        if is_better_match {
            best = Some(candidate.clone());
        }
    }
    best
}

fn find_model_by_namespaced_suffix(model: &str, candidates: &[ModelInfo]) -> Option<ModelInfo> {
    // Retry metadata lookup for a single namespaced slug like `namespace/model-name`.
    //
    // This only strips one leading namespace segment and only when the namespace looks
    // like a simple provider id to avoid broadly matching arbitrary aliases.
    let (namespace, suffix) = model.split_once('/')?;
    if suffix.contains('/') {
        return None;
    }
    if namespace.is_empty()
        || !namespace
            .chars()
            .all(|c| c.is_ascii_alphanumeric() || c == '_' || c == '-')
    {
        return None;
    }
    find_model_by_longest_prefix(suffix, candidates)
}

pub(crate) fn construct_model_info_from_candidates(
    model: &str,
    candidates: &[ModelInfo],
    config: &ModelsManagerConfig,
) -> ModelInfo {
    // First use the normal longest-prefix match. If that misses, allow a narrowly scoped
    // retry for namespaced slugs like `custom/gpt-5.3-codex`.
    let remote = find_model_by_longest_prefix(model, candidates)
        .or_else(|| find_model_by_namespaced_suffix(model, candidates));
    let model_info = if let Some(remote) = remote {
        ModelInfo {
            slug: model.to_string(),
            used_fallback_model_metadata: false,
            ..remote
        }
    } else {
        model_info::model_info_from_slug(model)
    };
    model_info::with_config_overrides(model_info, config)
}

#[cfg(test)]
#[path = "manager_tests.rs"]
mod tests;