Merge pull request #1466 from nearai/staging-promote/3da9810e-23351687636

chore: promote staging to staging-promote/cba1bc37-23334371795 (2026-03-20 16:12 UTC)
2026-08-25 14:53:34 +00:00 · 2026-03-23 12:00:31 -07:00
parent cba1bc3799 3da9810e87
commit ddf64e8485
20 changed files with 2477 additions and 14 deletions
@@ -4,7 +4,7 @@ DATABASE_POOL_SIZE=10

 # LLM Provider
 # LLM_BACKEND=nearai           # default
-# Possible values: nearai, ollama, openai_compatible, openai, anthropic, tinfoil
+# Possible values: nearai, ollama, openai_compatible, openai, anthropic, tinfoil, openai_codex
 # LLM_REQUEST_TIMEOUT_SECS=120  # Increase for local LLMs (Ollama, vLLM, LM Studio)

 # === Anthropic Direct ===
@@ -92,6 +92,13 @@ NEARAI_AUTH_URL=https://private.near.ai
 #   long  = 1-hour TTL, 2.0× (200%) write surcharge
 # ANTHROPIC_CACHE_RETENTION=short

+# === OpenAI Codex (ChatGPT subscription, OAuth) ===
+# LLM_BACKEND=openai_codex
+# OPENAI_CODEX_MODEL=gpt-5.3-codex              # default
+# OPENAI_CODEX_CLIENT_ID=app_EMoamEEZ73f0CkXaXp7hrann  # override (rare)
+# OPENAI_CODEX_AUTH_URL=https://auth.openai.com  # override (rare)
+# OPENAI_CODEX_API_URL=https://chatgpt.com/backend-api/codex  # override (rare)
+
 # For full provider setup guide see docs/LLM_PROVIDERS.md

 # Channel Configuration
@@ -696,6 +696,7 @@ impl AppBuilder {
        // fail early with a clear error instead of a confusing runtime failure.
        if self.config.llm.backend != "nearai"
            && self.config.llm.backend != "bedrock"
+            && self.config.llm.backend != "openai_codex"
            && self.config.llm.provider.is_none()
        {
            let backend = &self.config.llm.backend;
@@ -239,6 +239,17 @@ pub enum Command {
    )]
    Import(ImportCommand),

+    /// Authenticate with a provider (re-login)
+    #[command(
+        about = "Authenticate with a provider",
+        long_about = "Re-authenticate with an LLM provider.\nExample: ironclaw login --openai-codex"
+    )]
+    Login {
+        /// Authenticate with OpenAI Codex (ChatGPT subscription)
+        #[arg(long)]
+        openai_codex: bool,
+    },
+
    /// Run as a sandboxed worker inside a Docker container (internal use).
    /// This is invoked automatically by the orchestrator, not by users directly.
    #[command(hide = true)]
@@ -24,6 +24,7 @@ Commands:
  status      Show system status
  completion  Generate completions
  import      Import from other AI systems
+  login       Authenticate with a provider
  help        Print this message or the help of the given subcommand(s)

 Options:
@@ -23,6 +23,7 @@ Commands:
  logs        View and manage gateway logs
  status      Show system status
  completion  Generate completions
+  login       Authenticate with a provider
  help        Print this message or the help of the given subcommand(s)

 Options:
@@ -27,6 +27,7 @@ Commands:
  status      Show system status
  completion  Generate completions
  import      Import from other AI systems
+  login       Authenticate with a provider
  help        Print this message or the help of the given subcommand(s)

 Options:
@@ -26,6 +26,7 @@ Commands:
  logs        View and manage gateway logs
  status      Show system status
  completion  Generate completions
+  login       Authenticate with a provider
  help        Print this message or the help of the given subcommand(s)

 Options:
@@ -37,6 +37,7 @@ impl LlmConfig {
            },
            provider: None,
            bedrock: None,
+            openai_codex: None,
            request_timeout_secs: 120,
            cheap_model: None,
            smart_routing_cascade: false,
@@ -72,8 +73,12 @@ impl LlmConfig {
            backend_lower == "nearai" || backend_lower == "near_ai" || backend_lower == "near";
        let is_bedrock =
            backend_lower == "bedrock" || backend_lower == "aws_bedrock" || backend_lower == "aws";
+        let is_openai_codex = backend_lower == "openai_codex"
+            || backend_lower == "openai-codex"
+            || backend_lower == "codex";

-        if !is_nearai && !is_bedrock && registry.find(&backend_lower).is_none() {
+        if !is_nearai && !is_bedrock && !is_openai_codex && registry.find(&backend_lower).is_none()
+        {
            tracing::warn!(
                "Unknown LLM backend '{}'. Will attempt as openai_compatible fallback.",
                backend
@@ -126,8 +131,8 @@ impl LlmConfig {
            smart_routing_cascade: parse_optional_env("SMART_ROUTING_CASCADE", true)?,
        };

-        // Resolve registry provider config (for non-NearAI, non-Bedrock backends)
-        let provider = if is_nearai || is_bedrock {
+        // Resolve registry provider config (for non-NearAI, non-Bedrock, non-Codex backends)
+        let provider = if is_nearai || is_bedrock || is_openai_codex {
            None
        } else {
            Some(Self::resolve_registry_provider(
@@ -174,6 +179,38 @@ impl LlmConfig {
            None
        };

+        // Resolve OpenAI Codex config
+        let openai_codex = if is_openai_codex {
+            // Model: OPENAI_CODEX_MODEL > OPENAI_MODEL > settings.selected_model > default
+            let model = optional_env("OPENAI_CODEX_MODEL")?
+                .or(optional_env("OPENAI_MODEL")?)
+                .or_else(|| settings.selected_model.clone())
+                .unwrap_or_else(|| "gpt-5.3-codex".to_string());
+            let auth_endpoint = optional_env("OPENAI_CODEX_AUTH_URL")?
+                .unwrap_or_else(|| "https://auth.openai.com".to_string());
+            validate_base_url(&auth_endpoint, "OPENAI_CODEX_AUTH_URL")?;
+            let api_base_url = optional_env("OPENAI_CODEX_API_URL")?
+                .unwrap_or_else(|| "https://chatgpt.com/backend-api/codex".to_string());
+            validate_base_url(&api_base_url, "OPENAI_CODEX_API_URL")?;
+            let client_id = optional_env("OPENAI_CODEX_CLIENT_ID")?
+                .unwrap_or_else(|| "app_EMoamEEZ73f0CkXaXp7hrann".to_string());
+            let session_path = optional_env("OPENAI_CODEX_SESSION_PATH")?
+                .map(PathBuf::from)
+                .unwrap_or_else(|| ironclaw_base_dir().join("openai_codex_session.json"));
+            let token_refresh_margin_secs =
+                parse_optional_env("OPENAI_CODEX_REFRESH_MARGIN_SECS", 300)?;
+            Some(OpenAiCodexConfig {
+                model,
+                auth_endpoint,
+                api_base_url,
+                client_id,
+                session_path,
+                token_refresh_margin_secs,
+            })
+        } else {
+            None
+        };
+
        let request_timeout_secs = parse_optional_env("LLM_REQUEST_TIMEOUT_SECS", 120)?;

        // Generic cheap model (works with any backend).
@@ -189,6 +226,8 @@ impl LlmConfig {
                "nearai".to_string()
            } else if is_bedrock {
                "bedrock".to_string()
+            } else if is_openai_codex {
+                "openai_codex".to_string()
            } else if let Some(ref p) = provider {
                p.provider_id.clone()
            } else {
@@ -198,6 +237,7 @@ impl LlmConfig {
            nearai,
            provider,
            bedrock,
+            openai_codex,
            request_timeout_secs,
            cheap_model,
            smart_routing_cascade,
@@ -1069,4 +1109,159 @@ mod tests {
            std::env::remove_var("LLM_REQUEST_TIMEOUT_SECS");
        }
    }
+
+    // ── OpenAI Codex tests ──────────────────────────────────────────
+
+    /// Clear all openai-codex-related env vars.
+    fn clear_openai_codex_env() {
+        // SAFETY: Only called under ENV_MUTEX in tests.
+        unsafe {
+            std::env::remove_var("LLM_BACKEND");
+            std::env::remove_var("OPENAI_CODEX_MODEL");
+            std::env::remove_var("OPENAI_MODEL");
+        }
+    }
+
+    #[test]
+    fn openai_codex_resolves_config() {
+        let _guard = ENV_MUTEX.lock().expect("env mutex poisoned");
+        clear_openai_codex_env();
+
+        let settings = Settings {
+            llm_backend: Some("openai_codex".to_string()),
+            ..Default::default()
+        };
+
+        let cfg = LlmConfig::resolve(&settings).expect("resolve should succeed");
+        assert_eq!(cfg.backend, "openai_codex");
+        let codex = cfg.openai_codex.expect("codex config should be present");
+        assert_eq!(codex.model, "gpt-5.3-codex"); // default
+        assert!(
+            cfg.provider.is_none(),
+            "codex should not use registry provider"
+        );
+    }
+
+    #[test]
+    fn openai_codex_model_env_resolution() {
+        let _guard = ENV_MUTEX.lock().expect("env mutex poisoned");
+        clear_openai_codex_env();
+        // SAFETY: Under ENV_MUTEX.
+        unsafe {
+            std::env::set_var("OPENAI_CODEX_MODEL", "o3-pro");
+        }
+
+        let settings = Settings {
+            llm_backend: Some("openai_codex".to_string()),
+            ..Default::default()
+        };
+
+        let cfg = LlmConfig::resolve(&settings).expect("resolve should succeed");
+        let codex = cfg.openai_codex.expect("codex config should be present");
+        assert_eq!(codex.model, "o3-pro");
+
+        // SAFETY: Under ENV_MUTEX.
+        unsafe {
+            std::env::remove_var("OPENAI_CODEX_MODEL");
+        }
+    }
+
+    #[test]
+    fn openai_codex_falls_back_to_openai_model() {
+        let _guard = ENV_MUTEX.lock().expect("env mutex poisoned");
+        clear_openai_codex_env();
+        // SAFETY: Under ENV_MUTEX.
+        unsafe {
+            std::env::set_var("OPENAI_MODEL", "gpt-4o");
+        }
+
+        let settings = Settings {
+            llm_backend: Some("openai_codex".to_string()),
+            ..Default::default()
+        };
+
+        let cfg = LlmConfig::resolve(&settings).expect("resolve should succeed");
+        let codex = cfg.openai_codex.expect("codex config should be present");
+        assert_eq!(codex.model, "gpt-4o");
+
+        // SAFETY: Under ENV_MUTEX.
+        unsafe {
+            std::env::remove_var("OPENAI_MODEL");
+        }
+    }
+
+    #[test]
+    fn openai_codex_falls_back_to_selected_model() {
+        let _guard = ENV_MUTEX.lock().expect("env mutex poisoned");
+        clear_openai_codex_env();
+
+        let settings = Settings {
+            llm_backend: Some("openai_codex".to_string()),
+            selected_model: Some("gpt-4o-mini".to_string()),
+            ..Default::default()
+        };
+
+        let cfg = LlmConfig::resolve(&settings).expect("resolve should succeed");
+        let codex = cfg.openai_codex.expect("codex config should be present");
+        assert_eq!(codex.model, "gpt-4o-mini");
+    }
+
+    /// Regression: SSRF validation on OPENAI_CODEX_API_URL (#1103).
+    #[test]
+    fn openai_codex_rejects_ssrf_api_url() {
+        let _guard = ENV_MUTEX.lock().expect("env mutex poisoned");
+        clear_openai_codex_env();
+        // SAFETY: Under ENV_MUTEX.
+        unsafe {
+            std::env::set_var(
+                "OPENAI_CODEX_API_URL",
+                "http://169.254.169.254/latest/meta-data",
+            );
+        }
+
+        let settings = Settings {
+            llm_backend: Some("openai_codex".to_string()),
+            ..Default::default()
+        };
+
+        let err = LlmConfig::resolve(&settings).unwrap_err();
+        let msg = err.to_string();
+        assert!(
+            msg.contains("OPENAI_CODEX_API_URL"),
+            "error should reference the field name: {msg}"
+        );
+
+        // SAFETY: Under ENV_MUTEX.
+        unsafe {
+            std::env::remove_var("OPENAI_CODEX_API_URL");
+        }
+    }
+
+    /// Regression: SSRF validation on OPENAI_CODEX_AUTH_URL (#1103).
+    #[test]
+    fn openai_codex_rejects_ssrf_auth_url() {
+        let _guard = ENV_MUTEX.lock().expect("env mutex poisoned");
+        clear_openai_codex_env();
+        // SAFETY: Under ENV_MUTEX.
+        unsafe {
+            std::env::set_var("OPENAI_CODEX_AUTH_URL", "http://10.0.0.1");
+        }
+
+        let settings = Settings {
+            llm_backend: Some("openai_codex".to_string()),
+            ..Default::default()
+        };
+
+        let err = LlmConfig::resolve(&settings).unwrap_err();
+        let msg = err.to_string();
+        assert!(
+            msg.contains("OPENAI_CODEX_AUTH_URL"),
+            "error should reference the field name: {msg}"
+        );
+
+        // SAFETY: Under ENV_MUTEX.
+        unsafe {
+            std::env::remove_var("OPENAI_CODEX_AUTH_URL");
+        }
+    }
 }
@@ -54,7 +54,7 @@ pub use self::transcription::TranscriptionConfig;
 pub use self::tunnel::TunnelConfig;
 pub use self::wasm::WasmConfig;
 pub use crate::llm::config::{
-    BedrockConfig, CacheRetention, LlmConfig, NearAiConfig, OAUTH_PLACEHOLDER,
+    BedrockConfig, CacheRetention, LlmConfig, NearAiConfig, OAUTH_PLACEHOLDER, OpenAiCodexConfig,
    RegistryProviderConfig,
 };
 pub use crate::llm::session::SessionConfig;
@@ -377,7 +377,7 @@ pub(crate) fn resolve_owner_id(settings: &Settings) -> Result<String, ConfigErro
 /// are read by `optional_env()` before falling back to `std::env::var()`,
 /// so explicit env vars always win.
 ///
-/// Also loads tokens from OS credential stores (macOS Keychain, Linux
+/// Also loads tokens from OS credential stores (macOS Keychain / Linux
 /// credentials files) which don't require the secrets DB.
 pub async fn inject_llm_keys_from_secrets(
    secrets: &dyn crate::secrets::SecretsStore,
@@ -13,6 +13,9 @@ Multi-provider LLM integration with circuit breaker, retry, failover, and respon
 | `nearai_chat.rs` | NEAR AI Chat Completions provider (dual auth: session token or API key) |
 | `codex_auth.rs` | Reads Codex CLI `auth.json`, extracts tokens, refreshes ChatGPT OAuth access tokens |
 | `codex_chatgpt.rs` | Custom Responses API provider for Codex ChatGPT backend (`/backend-api/codex`) |
+| `openai_codex_provider.rs` | OpenAI Codex Responses API client (SSE streaming, JWT auth, subscription billing) |
+| `openai_codex_session.rs` | OAuth 2.0 session manager for OpenAI Codex (device code flow, token persistence) |
+| `token_refreshing.rs` | Token-refreshing `LlmProvider` decorator for OpenAI Codex (pre-emptive refresh, zero-cost billing) |
 | `reasoning.rs` | `Reasoning` struct, `ReasoningContext`, `RespondResult`, `ActionPlan`, `ToolSelection`; thinking-tag stripping; `SILENT_REPLY_TOKEN` |
 | `session.rs` | NEAR AI session token management with disk + DB persistence, OAuth login flow |
 | `circuit_breaker.rs` | Circuit breaker: Closed → Open → HalfOpen state machine |
@@ -38,6 +41,7 @@ Set via `LLM_BACKEND` env var:
 | `openai_compatible` | Any OpenAI-compatible endpoint | `LLM_BASE_URL`, `LLM_API_KEY`, `LLM_MODEL` |
 | `tinfoil` | Tinfoil TEE inference | `TINFOIL_API_KEY`, `TINFOIL_MODEL` |
 | `bedrock` | AWS Bedrock (requires `--features bedrock`) | `BEDROCK_REGION`, `BEDROCK_MODEL`, `AWS_PROFILE` |
+| `openai_codex` | OpenAI Codex (ChatGPT subscription) | `OPENAI_CODEX_MODEL`, `OPENAI_CODEX_CLIENT_ID` |

 Codex auth reuse:
 - Set `LLM_USE_CODEX_AUTH=true` to load credentials from `~/.codex/auth.json` (override with `CODEX_AUTH_PATH`).
@@ -148,9 +152,27 @@ To add a new provider:

 Set `LLM_EXTRA_HEADERS=Key:Value,Key2:Value2` to inject headers into every request. Useful for OpenRouter attribution (`HTTP-Referer`, `X-Title`). Invalid header names/values are skipped with a warning (not a fatal error).

+## OpenAI Codex Provider
+
+Uses the Responses API at `chatgpt.com/backend-api/codex/responses` with ChatGPT subscription OAuth tokens (zero API cost — billing through subscription).
+
+**Auth flow:** Device code OAuth via `auth.openai.com/api/accounts/deviceauth/*` endpoints. On first run, displays a code for the user to enter at a URL. Tokens are persisted to `~/.ironclaw/openai_codex_session.json` (mode 0600) and auto-refreshed before expiry.
+
+**Provider chain:** `OpenAiCodexProvider` → `TokenRefreshingProvider` (pre-emptive refresh + retry on 401) → standard decorator chain. The `TokenRefreshingProvider` intercepts `AuthFailed`/`SessionExpired` errors, refreshes the OAuth token, and retries once.
+
+**Key differences from other providers:**
+- Uses Responses API (not Chat Completions) — SSE streaming with different event types
+- System messages are sent as `instructions` field, not in `input` array
+- Tool schemas are normalized via `normalize_schema_strict()` for OpenAI strict mode
+- `cost_per_token()` returns `(0, 0)` — subscription-based billing
+- `set_model()` returns error — model is fixed at construction time
+- Image attachments are silently dropped with a warning log
+
+**Env vars:** `OPENAI_CODEX_MODEL` (default: `gpt-5.3-codex`), `OPENAI_CODEX_CLIENT_ID`, `OPENAI_CODEX_AUTH_URL`, `OPENAI_CODEX_API_URL`.
+
 ## Provider Chain Construction

-`build_provider_chain()` in `mod.rs` is the single source of truth for assembling decorators. The chain is:
+`build_provider_chain()` in `mod.rs` is the single source of truth for assembling decorators. It creates the base provider (dispatching to `create_openai_codex_provider()` for codex, `create_llm_provider()` for everything else), then applies all decorators inline:

 ```
 Raw provider
@@ -0,0 +1,34 @@
+//! Shared test helpers for OpenAI Codex provider tests.
+
+#![cfg(test)]
+
+use crate::config::OpenAiCodexConfig;
+
+/// Build a minimal JWT for testing (header.payload.signature).
+pub(crate) fn make_test_jwt(account_id: &str) -> String {
+    use base64::Engine;
+    let engine = base64::engine::general_purpose::URL_SAFE_NO_PAD;
+
+    let header = engine.encode(b"{\"alg\":\"RS256\",\"typ\":\"JWT\"}");
+    let payload_json = serde_json::json!({
+        "sub": "user123",
+        "https://api.openai.com/auth": {
+            "chatgpt_account_id": account_id,
+        },
+    });
+    let payload = engine.encode(payload_json.to_string().as_bytes());
+    let sig = engine.encode(b"fake-signature");
+    format!("{header}.{payload}.{sig}")
+}
+
+/// Build a test `OpenAiCodexConfig` with a given session path.
+pub(crate) fn test_codex_config(session_path: std::path::PathBuf) -> OpenAiCodexConfig {
+    OpenAiCodexConfig {
+        model: "gpt-5.3-codex".to_string(),
+        auth_endpoint: "https://auth.openai.com".to_string(),
+        api_base_url: "https://chatgpt.com/backend-api/codex".to_string(),
+        client_id: "test_client_id".to_string(),
+        session_path,
+        token_refresh_margin_secs: 300,
+    }
+}
@@ -9,6 +9,7 @@ use std::path::PathBuf;

 use secrecy::SecretString;

+use crate::bootstrap::ironclaw_base_dir;
 use crate::llm::registry::ProviderProtocol;
 use crate::llm::session::SessionConfig;

@@ -102,6 +103,36 @@ pub struct RegistryProviderConfig {
    pub unsupported_params: Vec<String>,
 }

+/// Configuration for OpenAI Codex (ChatGPT subscription OAuth).
+#[derive(Debug, Clone)]
+pub struct OpenAiCodexConfig {
+    /// Model to use (default: "gpt-5.3-codex").
+    pub model: String,
+    /// OAuth authorization server (default: "https://auth.openai.com").
+    pub auth_endpoint: String,
+    /// Responses API base URL (default: "https://chatgpt.com/backend-api/codex").
+    pub api_base_url: String,
+    /// OAuth client ID (default: OpenAI's public Codex client).
+    pub client_id: String,
+    /// Path to session file (default: ~/.ironclaw/openai_codex_session.json).
+    pub session_path: PathBuf,
+    /// Seconds before expiry to proactively refresh (default: 300).
+    pub token_refresh_margin_secs: u64,
+}
+
+impl Default for OpenAiCodexConfig {
+    fn default() -> Self {
+        Self {
+            model: "gpt-5.3-codex".to_string(),
+            auth_endpoint: "https://auth.openai.com".to_string(),
+            api_base_url: "https://chatgpt.com/backend-api/codex".to_string(),
+            client_id: "app_EMoamEEZ73f0CkXaXp7hrann".to_string(),
+            session_path: ironclaw_base_dir().join("openai_codex_session.json"),
+            token_refresh_margin_secs: 300,
+        }
+    }
+}
+
 /// Configuration for AWS Bedrock (native Converse API).
 #[derive(Debug, Clone)]
 pub struct BedrockConfig {
@@ -134,6 +165,8 @@ pub struct LlmConfig {
    pub provider: Option<RegistryProviderConfig>,
    /// AWS Bedrock config (populated when backend=bedrock, requires --features bedrock).
    pub bedrock: Option<BedrockConfig>,
+    /// OpenAI Codex config (populated when backend=openai_codex).
+    pub openai_codex: Option<OpenAiCodexConfig>,
    /// HTTP request timeout in seconds for LLM API calls.
    /// Default: 120. Increase for local LLMs (Ollama, vLLM, LM Studio) that
    /// need more time for prompt evaluation on consumer hardware.
@@ -20,6 +20,8 @@ pub mod error;
 pub mod failover;
 mod nearai_chat;
 pub mod oauth_helpers;
+pub mod openai_codex_provider;
+pub mod openai_codex_session;
 mod provider;
 mod reasoning;
 pub mod recording;
@@ -29,6 +31,10 @@ pub mod retry;
 mod rig_adapter;
 pub mod session;
 pub mod smart_routing;
+mod token_refreshing;
+
+#[cfg(test)]
+mod codex_test_helpers;

 pub mod image_models;
 pub mod models;
@@ -37,12 +43,14 @@ pub mod vision_models;

 pub use circuit_breaker::{CircuitBreakerConfig, CircuitBreakerProvider};
 pub use config::{
-    BedrockConfig, CacheRetention, LlmConfig, NearAiConfig, OAUTH_PLACEHOLDER,
+    BedrockConfig, CacheRetention, LlmConfig, NearAiConfig, OAUTH_PLACEHOLDER, OpenAiCodexConfig,
    RegistryProviderConfig,
 };
 pub use error::LlmError;
 pub use failover::{CooldownConfig, FailoverProvider};
 pub use nearai_chat::{DEFAULT_MODEL, ModelInfo, NearAiChatProvider, default_models};
+pub use openai_codex_provider::OpenAiCodexProvider;
+pub use openai_codex_session::{OpenAiCodexSession, OpenAiCodexSessionManager};
 pub use provider::{
    ChatMessage, CompletionRequest, CompletionResponse, ContentPart, FinishReason, ImageUrl,
    LlmProvider, ModelMetadata, Role, ToolCall, ToolCompletionRequest, ToolCompletionResponse,
@@ -59,6 +67,7 @@ pub use retry::{RetryConfig, RetryProvider};
 pub use rig_adapter::RigAdapter;
 pub use session::{SessionConfig, SessionManager, create_session_manager};
 pub use smart_routing::{SmartRoutingConfig, SmartRoutingProvider, TaskComplexity};
+pub use token_refreshing::TokenRefreshingProvider;

 use std::sync::Arc;

@@ -97,6 +106,15 @@ pub async fn create_llm_provider(
        }
    }

+    if config.backend == "openai_codex" {
+        return Err(LlmError::RequestFailed {
+            provider: "openai_codex".to_string(),
+            reason:
+                "OpenAI Codex uses a dedicated factory path. Use build_provider_chain() instead of create_llm_provider()."
+                    .to_string(),
+        });
+    }
+
    let reg_config = config
        .provider
        .as_ref()
@@ -374,6 +392,47 @@ fn create_ollama_from_registry(
    Ok(Arc::new(adapter))
 }

+/// Create an OpenAI Codex provider with OAuth authentication.
+///
+/// This is async because it needs to ensure authentication before
+/// creating the provider (which requires a valid Bearer token).
+///
+/// Uses the Responses API (`chatgpt.com/backend-api/codex/responses`)
+/// instead of the Chat Completions API, matching OpenClaw's approach.
+async fn create_openai_codex_provider(
+    config: &LlmConfig,
+) -> Result<Arc<dyn LlmProvider>, LlmError> {
+    let codex = config
+        .openai_codex
+        .as_ref()
+        .ok_or_else(|| LlmError::AuthFailed {
+            provider: "openai_codex".to_string(),
+        })?;
+
+    let session_mgr = Arc::new(OpenAiCodexSessionManager::new(codex.clone())?);
+    session_mgr.ensure_authenticated().await?;
+
+    let token = session_mgr.get_access_token().await?;
+
+    let provider = Arc::new(OpenAiCodexProvider::new(
+        &codex.model,
+        &codex.api_base_url,
+        token.expose_secret(),
+        config.request_timeout_secs,
+    )?);
+
+    tracing::info!(
+        "Using OpenAI Codex (Responses API, model: {}, base: {})",
+        codex.model,
+        codex.api_base_url,
+    );
+
+    Ok(Arc::new(TokenRefreshingProvider::new(
+        provider,
+        session_mgr,
+    )))
+}
+
 /// Create a cheap/fast LLM provider for lightweight tasks (heartbeat, routing, evaluation).
 ///
 /// Resolution order:
@@ -460,7 +519,11 @@ pub async fn build_provider_chain(
    ),
    LlmError,
 > {
-    let llm = create_llm_provider(config, session.clone()).await?;
+    let llm: Arc<dyn LlmProvider> = if config.backend == "openai_codex" {
+        create_openai_codex_provider(config).await?
+    } else {
+        create_llm_provider(config, session.clone()).await?
+    };
    tracing::debug!("LLM provider initialized: {}", llm.model_name());

    // 1. Retry
@@ -632,6 +695,7 @@ mod tests {
            request_timeout_secs: 120,
            cheap_model: None,
            smart_routing_cascade: true,
+            openai_codex: None,
        }
    }

@@ -347,5 +347,6 @@ pub(crate) fn build_nearai_model_fetch_config() -> crate::config::LlmConfig {
        request_timeout_secs: 120,
        cheap_model: None,
        smart_routing_cascade: false,
+        openai_codex: None,
    }
 }
@@ -0,0 +1,731 @@
+//! OAuth 2.0 session manager for OpenAI Codex (ChatGPT subscription).
+//!
+//! Supports two auth flows:
+//! - **Device Code** (primary): Works on headless servers, no browser needed.
+//! - **Browser PKCE** (fallback): Standard OAuth for local machines.
+//!
+//! Tokens are persisted to `~/.ironclaw/openai_codex_session.json` and
+//! auto-refreshed before expiry.
+
+use chrono::{DateTime, Utc};
+use reqwest::Client;
+use reqwest::header::{HeaderMap, HeaderValue, USER_AGENT};
+use secrecy::SecretString;
+use serde::{Deserialize, Serialize};
+use tokio::sync::{Mutex, RwLock};
+
+use crate::config::OpenAiCodexConfig;
+use crate::error::LlmError;
+
+/// Persisted OAuth session data.
+///
+/// Note: `Debug` is manually implemented to redact tokens.
+#[derive(Serialize, Deserialize)]
+pub struct OpenAiCodexSession {
+    pub(crate) access_token: String,
+    pub(crate) refresh_token: String,
+    pub(crate) expires_at: DateTime<Utc>,
+    pub(crate) created_at: DateTime<Utc>,
+}
+
+impl std::fmt::Debug for OpenAiCodexSession {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        f.debug_struct("OpenAiCodexSession")
+            .field("access_token", &"[REDACTED]")
+            .field("refresh_token", &"[REDACTED]")
+            .field("expires_at", &self.expires_at)
+            .field("created_at", &self.created_at)
+            .finish()
+    }
+}
+
+/// Request body for the device code usercode endpoint.
+#[derive(Debug, Serialize)]
+struct UserCodeRequest {
+    client_id: String,
+}
+
+/// Response from the device code usercode endpoint.
+#[derive(Debug, Deserialize)]
+struct UserCodeResponse {
+    /// Unique ID for this device auth session.
+    device_auth_id: String,
+    /// Code the user enters in their browser.
+    user_code: String,
+    /// URL where the user enters the code (may not be present).
+    #[serde(default = "default_verification_uri")]
+    verification_uri: String,
+    /// Polling interval in seconds (OpenAI sends this as a string).
+    #[serde(
+        default = "default_interval",
+        deserialize_with = "deserialize_string_or_u64"
+    )]
+    interval: u64,
+    /// Expiry timestamp (OpenAI sends `expires_at` as ISO-8601).
+    #[serde(default)]
+    expires_at: Option<String>,
+    /// Seconds until the device code expires (standard field, may not be present).
+    #[serde(default)]
+    expires_in: Option<u64>,
+}
+
+fn default_verification_uri() -> String {
+    "https://auth.openai.com/codex/device".to_string()
+}
+
+fn default_interval() -> u64 {
+    5
+}
+
+/// Deserialize a value that may be either a string or a number as u64.
+fn deserialize_string_or_u64<'de, D>(deserializer: D) -> Result<u64, D::Error>
+where
+    D: serde::Deserializer<'de>,
+{
+    use serde::de;
+
+    struct StringOrU64;
+    impl<'de> de::Visitor<'de> for StringOrU64 {
+        type Value = u64;
+        fn expecting(&self, formatter: &mut std::fmt::Formatter) -> std::fmt::Result {
+            formatter.write_str("a string or integer")
+        }
+        fn visit_u64<E: de::Error>(self, v: u64) -> Result<u64, E> {
+            Ok(v)
+        }
+        fn visit_str<E: de::Error>(self, v: &str) -> Result<u64, E> {
+            v.parse().map_err(de::Error::custom)
+        }
+    }
+    deserializer.deserialize_any(StringOrU64)
+}
+
+impl UserCodeResponse {
+    /// Get the expiry duration in seconds, from either `expires_in` or `expires_at`.
+    fn expires_in_secs(&self) -> u64 {
+        if let Some(secs) = self.expires_in {
+            return secs;
+        }
+        if let Some(ref ts) = self.expires_at
+            && let Ok(dt) = chrono::DateTime::parse_from_rfc3339(ts)
+        {
+            let remaining = dt.signed_duration_since(Utc::now()).num_seconds();
+            return remaining.max(0) as u64;
+        }
+        900 // default 15 minutes
+    }
+}
+
+/// Request body for polling the device auth token endpoint.
+#[derive(Debug, Serialize)]
+struct DeviceTokenPollRequest {
+    device_auth_id: String,
+    user_code: String,
+}
+
+/// Successful response from the device auth token endpoint.
+/// Returns an authorization code + PKCE pair for the final token exchange.
+#[derive(Debug, Deserialize)]
+struct DeviceAuthCodeResponse {
+    authorization_code: String,
+    #[allow(dead_code)]
+    code_challenge: String,
+    code_verifier: String,
+}
+
+/// Response from the final OAuth token exchange.
+#[derive(Debug, Deserialize)]
+struct TokenResponse {
+    access_token: String,
+    #[serde(default)]
+    refresh_token: String,
+    #[serde(default)]
+    expires_in: u64,
+    #[serde(default)]
+    #[allow(dead_code)]
+    token_type: String,
+}
+
+/// Manages OpenAI Codex OAuth sessions with persistence and auto-refresh.
+pub struct OpenAiCodexSessionManager {
+    config: OpenAiCodexConfig,
+    client: Client,
+    session: RwLock<Option<OpenAiCodexSession>>,
+    renewal_lock: Mutex<()>,
+}
+
+impl OpenAiCodexSessionManager {
+    /// Create a new session manager. Tries to load existing session from disk.
+    ///
+    /// # Errors
+    ///
+    /// Returns `LlmError` if the HTTP client cannot be constructed.
+    pub fn new(config: OpenAiCodexConfig) -> Result<Self, LlmError> {
+        let mut headers = HeaderMap::new();
+        headers.insert(
+            USER_AGENT,
+            HeaderValue::from_static(concat!("ironclaw/", env!("CARGO_PKG_VERSION"))),
+        );
+        let client = Client::builder()
+            .default_headers(headers)
+            .timeout(std::time::Duration::from_secs(30))
+            .build()
+            .map_err(|e| LlmError::RequestFailed {
+                provider: "openai_codex".into(),
+                reason: format!("HTTP client build failed: {e}"),
+            })?;
+
+        let mgr = Self {
+            config,
+            client,
+            session: RwLock::new(None),
+            renewal_lock: Mutex::new(()),
+        };
+
+        // Try synchronous load from disk during construction
+        if let Ok(data) = std::fs::read_to_string(&mgr.config.session_path)
+            && let Ok(session) = serde_json::from_str::<OpenAiCodexSession>(&data)
+            && let Ok(mut guard) = mgr.session.try_write()
+        {
+            *guard = Some(session);
+            tracing::info!(
+                "Loaded OpenAI Codex session from {}",
+                mgr.config.session_path.display()
+            );
+        }
+
+        Ok(mgr)
+    }
+
+    /// Check if we have a session (may be expired).
+    pub async fn has_session(&self) -> bool {
+        self.session.read().await.is_some()
+    }
+
+    /// Check if the current access token needs refreshing.
+    pub async fn needs_refresh(&self) -> bool {
+        let guard = self.session.read().await;
+        match guard.as_ref() {
+            None => true,
+            Some(s) => {
+                let margin =
+                    chrono::Duration::seconds(self.config.token_refresh_margin_secs as i64);
+                Utc::now() + margin >= s.expires_at
+            }
+        }
+    }
+
+    /// Get the current access token, refreshing if needed.
+    ///
+    /// If the token is within the refresh margin, silently refreshes first.
+    /// If no session exists, returns an AuthFailed error.
+    pub async fn get_access_token(&self) -> Result<SecretString, LlmError> {
+        if self.needs_refresh().await {
+            let has_refresh = self
+                .session
+                .read()
+                .await
+                .as_ref()
+                .map(|s| !s.refresh_token.is_empty())
+                .unwrap_or(false);
+            if has_refresh {
+                self.refresh_tokens().await?;
+            } else {
+                return Err(LlmError::AuthFailed {
+                    provider: "openai_codex".to_string(),
+                });
+            }
+        }
+
+        let guard = self.session.read().await;
+        guard
+            .as_ref()
+            .map(|s| SecretString::from(s.access_token.clone()))
+            .ok_or_else(|| LlmError::AuthFailed {
+                provider: "openai_codex".to_string(),
+            })
+    }
+
+    /// Ensure we have a valid session. Loads from disk, refreshes, or prompts login.
+    pub async fn ensure_authenticated(&self) -> Result<(), LlmError> {
+        // Try loading from disk if we don't have a session
+        if !self.has_session().await {
+            let _ = self.load_session().await;
+        }
+
+        if !self.has_session().await {
+            // No session at all -- need to authenticate
+            return self.device_code_login().await;
+        }
+
+        if self.needs_refresh().await {
+            // Try refresh; if it fails, re-authenticate
+            match self.refresh_tokens().await {
+                Ok(()) => Ok(()),
+                Err(e) => {
+                    tracing::info!("Token refresh failed ({}), re-authenticating...", e);
+                    self.device_code_login().await
+                }
+            }
+        } else {
+            Ok(())
+        }
+    }
+
+    /// Run OpenAI's device code auth flow.
+    ///
+    /// Uses OpenAI's custom `/api/accounts/deviceauth/*` endpoints (not the standard
+    /// Auth0 `/oauth/device/code` which is behind Cloudflare managed challenge).
+    ///
+    /// Flow:
+    /// 1. POST `/api/accounts/deviceauth/usercode` → get device_auth_id + user_code
+    /// 2. Poll POST `/api/accounts/deviceauth/token` → get authorization_code + PKCE
+    /// 3. Exchange via POST `/oauth/token` → get access_token + refresh_token
+    pub async fn device_code_login(&self) -> Result<(), LlmError> {
+        let _guard = self.renewal_lock.lock().await;
+
+        let auth_base = format!("{}/api/accounts", self.config.auth_endpoint);
+
+        // Step 1: Request device code
+        let usercode_url = format!("{}/deviceauth/usercode", auth_base);
+        let resp = self
+            .client
+            .post(&usercode_url)
+            .json(&UserCodeRequest {
+                client_id: self.config.client_id.clone(),
+            })
+            .send()
+            .await
+            .map_err(|e| LlmError::SessionRenewalFailed {
+                provider: "openai_codex".to_string(),
+                reason: format!("Device code request failed: {}", e),
+            })?;
+
+        if !resp.status().is_success() {
+            let status = resp.status();
+            let body = resp.text().await.unwrap_or_default();
+            return Err(LlmError::SessionRenewalFailed {
+                provider: "openai_codex".to_string(),
+                reason: format!("Device code request failed: HTTP {} -- {}", status, body),
+            });
+        }
+
+        let body_text = resp
+            .text()
+            .await
+            .map_err(|e| LlmError::SessionRenewalFailed {
+                provider: "openai_codex".to_string(),
+                reason: format!("Failed to read device code response: {}", e),
+            })?;
+        tracing::debug!("Device code response received ({} bytes)", body_text.len());
+        let device: UserCodeResponse =
+            serde_json::from_str(&body_text).map_err(|e| LlmError::SessionRenewalFailed {
+                provider: "openai_codex".to_string(),
+                reason: format!(
+                    "Failed to parse device code response: {} ({} bytes)",
+                    e,
+                    body_text.len()
+                ),
+            })?;
+
+        // Step 2: Display code to user
+        println!();
+        println!("===========================================================");
+        println!("               OpenAI Codex Authentication                  ");
+        println!("===========================================================");
+        println!();
+        println!("  1. Open this URL in any browser:");
+        println!("     {}", device.verification_uri);
+        println!();
+        println!("  2. Enter this code:");
+        println!();
+        println!("              [  {}  ]", device.user_code);
+        println!();
+        let expires_secs = device.expires_in_secs();
+        println!(
+            "  Waiting for authorization... (expires in {} min)",
+            expires_secs / 60
+        );
+        println!("===========================================================");
+        println!();
+
+        // Step 3: Poll for authorization code
+        let poll_url = format!("{}/deviceauth/token", auth_base);
+        let mut interval = std::time::Duration::from_secs(device.interval.max(5));
+        let deadline = tokio::time::Instant::now() + std::time::Duration::from_secs(expires_secs);
+
+        let auth_code = loop {
+            tokio::time::sleep(interval).await;
+
+            if tokio::time::Instant::now() >= deadline {
+                return Err(LlmError::SessionRenewalFailed {
+                    provider: "openai_codex".to_string(),
+                    reason: "Device code authorization timed out".to_string(),
+                });
+            }
+
+            let resp = self
+                .client
+                .post(&poll_url)
+                .json(&DeviceTokenPollRequest {
+                    device_auth_id: device.device_auth_id.clone(),
+                    user_code: device.user_code.clone(),
+                })
+                .send()
+                .await
+                .map_err(|e| LlmError::SessionRenewalFailed {
+                    provider: "openai_codex".to_string(),
+                    reason: format!("Token poll request failed: {}", e),
+                })?;
+
+            let status = resp.status();
+            if status.is_success() {
+                let code_resp: DeviceAuthCodeResponse =
+                    resp.json()
+                        .await
+                        .map_err(|e| LlmError::SessionRenewalFailed {
+                            provider: "openai_codex".to_string(),
+                            reason: format!("Failed to parse auth code response: {}", e),
+                        })?;
+                break code_resp;
+            }
+
+            // 403 = authorization_pending, keep polling
+            // 404 = device code not found / not enabled
+            if status == reqwest::StatusCode::FORBIDDEN {
+                continue;
+            }
+
+            if status == reqwest::StatusCode::NOT_FOUND {
+                return Err(LlmError::SessionRenewalFailed {
+                    provider: "openai_codex".to_string(),
+                    reason: "Device code login is not enabled. Please check your OpenAI account settings.".to_string(),
+                });
+            }
+
+            // Slow down on 429, cap at 60s to avoid unbounded growth
+            if status == reqwest::StatusCode::TOO_MANY_REQUESTS {
+                interval = (interval + std::time::Duration::from_secs(5))
+                    .min(std::time::Duration::from_secs(60));
+                continue;
+            }
+
+            let body = resp.text().await.unwrap_or_default();
+            return Err(LlmError::SessionRenewalFailed {
+                provider: "openai_codex".to_string(),
+                reason: format!("Device auth poll failed: HTTP {} -- {}", status, body),
+            });
+        };
+
+        // Step 4: Exchange authorization code for tokens (form-encoded, per Auth0 spec)
+        let token_url = format!("{}/oauth/token", self.config.auth_endpoint);
+        let resp = self
+            .client
+            .post(&token_url)
+            .form(&[
+                ("grant_type", "authorization_code"),
+                ("code", &auth_code.authorization_code),
+                ("code_verifier", &auth_code.code_verifier),
+                ("client_id", &self.config.client_id),
+                (
+                    "redirect_uri",
+                    &format!("{}/deviceauth/callback", self.config.auth_endpoint),
+                ),
+            ])
+            .send()
+            .await
+            .map_err(|e| LlmError::SessionRenewalFailed {
+                provider: "openai_codex".to_string(),
+                reason: format!("Token exchange failed: {}", e),
+            })?;
+
+        if !resp.status().is_success() {
+            let status = resp.status();
+            let body = resp.text().await.unwrap_or_default();
+            return Err(LlmError::SessionRenewalFailed {
+                provider: "openai_codex".to_string(),
+                reason: format!("Token exchange failed: HTTP {} -- {}", status, body),
+            });
+        }
+
+        let token_resp: TokenResponse =
+            resp.json()
+                .await
+                .map_err(|e| LlmError::SessionRenewalFailed {
+                    provider: "openai_codex".to_string(),
+                    reason: format!("Failed to parse token response: {}", e),
+                })?;
+
+        let session = OpenAiCodexSession {
+            access_token: token_resp.access_token,
+            refresh_token: token_resp.refresh_token,
+            expires_at: Utc::now()
+                + chrono::Duration::seconds(if token_resp.expires_in > 0 {
+                    token_resp.expires_in
+                } else {
+                    tracing::warn!("Token response has expires_in=0, defaulting to 3600s");
+                    3600
+                } as i64),
+            created_at: Utc::now(),
+        };
+
+        self.save_session(&session).await?;
+        self.set_session(session).await;
+
+        println!();
+        println!("Authentication successful!");
+        println!();
+        Ok(())
+    }
+
+    /// Refresh the access token using the refresh token.
+    pub async fn refresh_tokens(&self) -> Result<(), LlmError> {
+        let _guard = self.renewal_lock.lock().await;
+
+        // Double-check: another task may have refreshed while we waited on the lock
+        if !self.needs_refresh().await {
+            return Ok(());
+        }
+
+        let refresh_token = {
+            let guard = self.session.read().await;
+            guard
+                .as_ref()
+                .map(|s| s.refresh_token.clone())
+                .ok_or_else(|| LlmError::AuthFailed {
+                    provider: "openai_codex".to_string(),
+                })?
+        };
+
+        let token_url = format!("{}/oauth/token", self.config.auth_endpoint);
+        let resp = self
+            .client
+            .post(&token_url)
+            .form(&[
+                ("grant_type", "refresh_token"),
+                ("refresh_token", refresh_token.as_str()),
+                ("client_id", self.config.client_id.as_str()),
+            ])
+            .send()
+            .await
+            .map_err(|e| LlmError::SessionRenewalFailed {
+                provider: "openai_codex".to_string(),
+                reason: format!("Token refresh request failed: {}", e),
+            })?;
+
+        if !resp.status().is_success() {
+            let status = resp.status();
+            let body = resp.text().await.unwrap_or_default();
+            return Err(LlmError::SessionRenewalFailed {
+                provider: "openai_codex".to_string(),
+                reason: format!("Token refresh failed: HTTP {} -- {}", status, body),
+            });
+        }
+
+        let token_resp: TokenResponse =
+            resp.json()
+                .await
+                .map_err(|e| LlmError::SessionRenewalFailed {
+                    provider: "openai_codex".to_string(),
+                    reason: format!("Failed to parse refresh response: {}", e),
+                })?;
+
+        let session = OpenAiCodexSession {
+            access_token: token_resp.access_token,
+            refresh_token: token_resp.refresh_token,
+            expires_at: Utc::now()
+                + chrono::Duration::seconds(if token_resp.expires_in > 0 {
+                    token_resp.expires_in
+                } else {
+                    tracing::warn!("Token response has expires_in=0, defaulting to 3600s");
+                    3600
+                } as i64),
+            created_at: Utc::now(),
+        };
+
+        self.save_session(&session).await?;
+        self.set_session(session).await;
+
+        tracing::debug!("OpenAI Codex token refreshed successfully");
+        Ok(())
+    }
+
+    /// Save session data to disk with restrictive permissions.
+    pub async fn save_session(&self, session: &OpenAiCodexSession) -> Result<(), LlmError> {
+        if let Some(parent) = self.config.session_path.parent() {
+            tokio::fs::create_dir_all(parent).await.map_err(|e| {
+                LlmError::Io(std::io::Error::new(
+                    e.kind(),
+                    format!("Failed to create session directory: {}", e),
+                ))
+            })?;
+        }
+
+        let json =
+            serde_json::to_string_pretty(session).map_err(|e| LlmError::SessionRenewalFailed {
+                provider: "openai_codex".to_string(),
+                reason: format!("Failed to serialize session: {}", e),
+            })?;
+
+        tokio::fs::write(&self.config.session_path, &json)
+            .await
+            .map_err(|e| {
+                LlmError::Io(std::io::Error::new(
+                    e.kind(),
+                    format!("Failed to write session file: {}", e),
+                ))
+            })?;
+
+        #[cfg(unix)]
+        {
+            use std::os::unix::fs::PermissionsExt;
+            let perms = std::fs::Permissions::from_mode(0o600);
+            tokio::fs::set_permissions(&self.config.session_path, perms)
+                .await
+                .map_err(|e| {
+                    LlmError::Io(std::io::Error::new(
+                        e.kind(),
+                        format!("Failed to set permissions: {}", e),
+                    ))
+                })?;
+        }
+
+        Ok(())
+    }
+
+    /// Load session from disk.
+    pub async fn load_session(&self) -> Result<(), LlmError> {
+        let data = tokio::fs::read_to_string(&self.config.session_path)
+            .await
+            .map_err(|e| {
+                LlmError::Io(std::io::Error::new(
+                    e.kind(),
+                    format!("Failed to read session file: {}", e),
+                ))
+            })?;
+
+        let session: OpenAiCodexSession =
+            serde_json::from_str(&data).map_err(|e| LlmError::SessionRenewalFailed {
+                provider: "openai_codex".to_string(),
+                reason: format!("Failed to parse session file: {}", e),
+            })?;
+
+        let mut guard = self.session.write().await;
+        *guard = Some(session);
+        tracing::info!(
+            "Loaded OpenAI Codex session from {}",
+            self.config.session_path.display()
+        );
+        Ok(())
+    }
+
+    /// Set session directly (for testing or after auth).
+    pub async fn set_session(&self, session: OpenAiCodexSession) {
+        let mut guard = self.session.write().await;
+        *guard = Some(session);
+    }
+
+    /// Handle a 401 response by refreshing, or re-authenticating.
+    pub async fn handle_auth_failure(&self) -> Result<(), LlmError> {
+        match self.refresh_tokens().await {
+            Ok(()) => Ok(()),
+            Err(_) => self.device_code_login().await,
+        }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use crate::llm::codex_test_helpers::test_codex_config as test_config;
+    use tempfile::tempdir;
+
+    #[tokio::test]
+    async fn test_save_and_load_session() {
+        let dir = tempdir().unwrap();
+        let path = dir.path().join("session.json");
+        let config = test_config(path.clone());
+
+        let mgr = OpenAiCodexSessionManager::new(config).unwrap();
+
+        // No session initially
+        assert!(!mgr.has_session().await);
+
+        // Save a session
+        let session = OpenAiCodexSession {
+            access_token: "access_abc".to_string(),
+            refresh_token: "refresh_xyz".to_string(),
+            expires_at: chrono::Utc::now() + chrono::Duration::hours(1),
+            created_at: chrono::Utc::now(),
+        };
+        mgr.save_session(&session).await.unwrap();
+        mgr.set_session(session).await;
+
+        assert!(mgr.has_session().await);
+
+        // Load from disk in a new manager
+        let config2 = test_config(path);
+        let mgr2 = OpenAiCodexSessionManager::new(config2).unwrap();
+        mgr2.load_session().await.unwrap();
+        assert!(mgr2.has_session().await);
+    }
+
+    #[tokio::test]
+    async fn test_needs_refresh_when_near_expiry() {
+        let dir = tempdir().unwrap();
+        let config = test_config(dir.path().join("session.json"));
+        let mgr = OpenAiCodexSessionManager::new(config).unwrap();
+
+        // Token expiring in 2 minutes (margin is 300s = 5 min)
+        let session = OpenAiCodexSession {
+            access_token: "access_abc".to_string(),
+            refresh_token: "refresh_xyz".to_string(),
+            expires_at: chrono::Utc::now() + chrono::Duration::minutes(2),
+            created_at: chrono::Utc::now(),
+        };
+        mgr.set_session(session).await;
+
+        assert!(mgr.needs_refresh().await);
+    }
+
+    #[test]
+    fn device_code_parse_error_redacts_body() {
+        // Regression: the parse error used to include raw body_text which could
+        // contain sensitive auth data. Now it only shows byte count.
+        let body_text = r#"{"secret_token":"sk-12345","error":"unexpected"}"#;
+        let err: Result<UserCodeResponse, _> = serde_json::from_str(body_text);
+        assert!(err.is_err());
+        let e = err.unwrap_err();
+        let error_msg = format!(
+            "Failed to parse device code response: {} ({} bytes)",
+            e,
+            body_text.len()
+        );
+        assert!(
+            !error_msg.contains("sk-12345"),
+            "error message must not contain raw body: {error_msg}"
+        );
+        assert!(
+            error_msg.contains("bytes"),
+            "error message should show byte count"
+        );
+    }
+
+    #[tokio::test]
+    async fn test_no_refresh_when_fresh() {
+        let dir = tempdir().unwrap();
+        let config = test_config(dir.path().join("session.json"));
+        let mgr = OpenAiCodexSessionManager::new(config).unwrap();
+
+        // Token expiring in 30 minutes (margin is 300s = 5 min)
+        let session = OpenAiCodexSession {
+            access_token: "access_abc".to_string(),
+            refresh_token: "refresh_xyz".to_string(),
+            expires_at: chrono::Utc::now() + chrono::Duration::minutes(30),
+            created_at: chrono::Utc::now(),
+        };
+        mgr.set_session(session).await;
+
+        assert!(!mgr.needs_refresh().await);
+    }
+}
@@ -132,7 +132,7 @@ fn round_f32_to_f64(val: f32) -> f64 {
 ///
 /// This is applied as a clone-and-transform at the provider boundary so the
 /// original tool definitions remain unchanged for other providers.
-fn normalize_schema_strict(schema: &JsonValue) -> JsonValue {
+pub(crate) fn normalize_schema_strict(schema: &JsonValue) -> JsonValue {
    let mut schema = schema.clone();
    normalize_schema_recursive(&mut schema);
    schema
@@ -0,0 +1,191 @@
+//! Token-refreshing LlmProvider decorator for OpenAI Codex.
+//!
+//! Wraps an `OpenAiCodexProvider` and:
+//! - Pre-emptively refreshes the OAuth access token before each call if near expiry
+//! - Updates the inner provider's token after refresh (no client rebuild needed)
+//! - Retries once on `AuthFailed` / `SessionExpired` after refreshing
+//! - Overrides `cost_per_token()` to return (0, 0) since billing is through subscription
+
+use std::sync::Arc;
+
+use async_trait::async_trait;
+use rust_decimal::Decimal;
+use secrecy::ExposeSecret;
+
+use crate::error::LlmError;
+use crate::llm::openai_codex_provider::OpenAiCodexProvider;
+use crate::llm::openai_codex_session::OpenAiCodexSessionManager;
+use crate::llm::provider::{
+    CompletionRequest, CompletionResponse, LlmProvider, ModelMetadata, ToolCompletionRequest,
+    ToolCompletionResponse,
+};
+
+/// Decorator that refreshes OAuth tokens before API calls and reports zero cost.
+///
+/// The inner `OpenAiCodexProvider` manages its own token state, so after a
+/// refresh we just call `update_token()` -- no client rebuild is needed.
+pub struct TokenRefreshingProvider {
+    inner: Arc<OpenAiCodexProvider>,
+    session: Arc<OpenAiCodexSessionManager>,
+}
+
+impl TokenRefreshingProvider {
+    pub fn new(inner: Arc<OpenAiCodexProvider>, session: Arc<OpenAiCodexSessionManager>) -> Self {
+        Self { inner, session }
+    }
+
+    /// Push a fresh token from the session manager into the inner provider.
+    async fn update_inner_token(&self) -> Result<(), LlmError> {
+        let token = self.session.get_access_token().await?;
+        self.inner.update_token(token.expose_secret()).await?;
+        tracing::debug!("Updated inner provider token after refresh");
+        Ok(())
+    }
+
+    /// Best-effort pre-emptive token refresh before an API call.
+    ///
+    /// If refresh fails (e.g., no refresh token), we log and continue so the
+    /// actual request still fires and the retry-on-auth-failure path can kick in.
+    async fn ensure_fresh_token(&self) {
+        if self.session.needs_refresh().await {
+            match self.session.refresh_tokens().await {
+                Ok(()) => {
+                    if let Err(e) = self.update_inner_token().await {
+                        tracing::warn!(
+                            "Pre-emptive token update failed: {e}, will retry on auth failure"
+                        );
+                    }
+                }
+                Err(e) => {
+                    tracing::warn!(
+                        "Pre-emptive token refresh failed: {e}, will retry on auth failure"
+                    );
+                }
+            }
+        }
+    }
+}
+
+#[async_trait]
+impl LlmProvider for TokenRefreshingProvider {
+    fn model_name(&self) -> &str {
+        self.inner.model_name()
+    }
+
+    fn cost_per_token(&self) -> (Decimal, Decimal) {
+        (Decimal::ZERO, Decimal::ZERO)
+    }
+
+    async fn complete(&self, request: CompletionRequest) -> Result<CompletionResponse, LlmError> {
+        self.ensure_fresh_token().await;
+
+        match self.inner.complete(request.clone()).await {
+            Err(LlmError::AuthFailed { .. } | LlmError::SessionExpired { .. }) => {
+                tracing::info!("Auth failure during complete(), refreshing and retrying once");
+                self.session.handle_auth_failure().await?;
+                self.update_inner_token().await?;
+                self.inner.complete(request).await
+            }
+            other => other,
+        }
+    }
+
+    async fn complete_with_tools(
+        &self,
+        request: ToolCompletionRequest,
+    ) -> Result<ToolCompletionResponse, LlmError> {
+        self.ensure_fresh_token().await;
+
+        match self.inner.complete_with_tools(request.clone()).await {
+            Err(LlmError::AuthFailed { .. } | LlmError::SessionExpired { .. }) => {
+                tracing::info!(
+                    "Auth failure during complete_with_tools(), refreshing and retrying once"
+                );
+                self.session.handle_auth_failure().await?;
+                self.update_inner_token().await?;
+                self.inner.complete_with_tools(request).await
+            }
+            other => other,
+        }
+    }
+
+    async fn list_models(&self) -> Result<Vec<String>, LlmError> {
+        self.ensure_fresh_token().await;
+        self.inner.list_models().await
+    }
+
+    async fn model_metadata(&self) -> Result<ModelMetadata, LlmError> {
+        self.ensure_fresh_token().await;
+        self.inner.model_metadata().await
+    }
+
+    fn active_model_name(&self) -> String {
+        self.inner.model_name().to_string()
+    }
+
+    fn effective_model_name(&self, requested_model: Option<&str>) -> String {
+        self.inner.effective_model_name(requested_model)
+    }
+
+    fn set_model(&self, model: &str) -> Result<(), LlmError> {
+        self.inner.set_model(model)
+    }
+
+    fn calculate_cost(&self, _input_tokens: u32, _output_tokens: u32) -> Decimal {
+        Decimal::ZERO
+    }
+
+    fn cache_write_multiplier(&self) -> Decimal {
+        self.inner.cache_write_multiplier()
+    }
+
+    fn cache_read_discount(&self) -> Decimal {
+        self.inner.cache_read_discount()
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use crate::llm::codex_test_helpers::{make_test_jwt, test_codex_config};
+    use crate::llm::openai_codex_session::OpenAiCodexSessionManager;
+    use tempfile::tempdir;
+
+    fn make_provider_and_session() -> (TokenRefreshingProvider, tempfile::TempDir) {
+        let dir = tempdir().unwrap();
+        let config = test_codex_config(dir.path().join("session.json"));
+        let jwt = make_test_jwt("acct_test");
+        let inner = Arc::new(
+            OpenAiCodexProvider::new(&config.model, &config.api_base_url, &jwt, 300)
+                .expect("provider creation should succeed"),
+        );
+        let session = Arc::new(OpenAiCodexSessionManager::new(config).unwrap());
+        (TokenRefreshingProvider::new(inner, session), dir)
+    }
+
+    #[test]
+    fn test_model_name_delegates() {
+        let (provider, _dir) = make_provider_and_session();
+        assert_eq!(provider.model_name(), "gpt-5.3-codex");
+    }
+
+    #[test]
+    fn test_cost_per_token_zero() {
+        let (provider, _dir) = make_provider_and_session();
+        let (input, output) = provider.cost_per_token();
+        assert_eq!(input, Decimal::ZERO);
+        assert_eq!(output, Decimal::ZERO);
+    }
+
+    #[test]
+    fn test_calculate_cost_zero() {
+        let (provider, _dir) = make_provider_and_session();
+        assert_eq!(provider.calculate_cost(1000, 500), Decimal::ZERO);
+    }
+
+    #[test]
+    fn test_active_model_name_delegates() {
+        let (provider, _dir) = make_provider_and_session();
+        assert_eq!(provider.active_model_name(), "gpt-5.3-codex");
+    }
+}
@@ -139,6 +139,47 @@ async fn async_main() -> anyhow::Result<()> {
            )
            .await;
        }
+        Some(Command::Login { openai_codex }) => {
+            init_cli_tracing();
+            if *openai_codex {
+                // Resolve codex config so OPENAI_CODEX_* env overrides are
+                // honoured even when LLM_BACKEND isn't set to openai_codex.
+                let codex_config = {
+                    let config = Config::from_env()
+                        .await
+                        .map_err(|e| anyhow::anyhow!("{}", e))?;
+                    config.llm.openai_codex.unwrap_or_else(|| {
+                        use ironclaw::llm::OpenAiCodexConfig;
+                        let mut cfg = OpenAiCodexConfig::default();
+                        if let Ok(v) = std::env::var("OPENAI_CODEX_AUTH_URL") {
+                            cfg.auth_endpoint = v;
+                        }
+                        if let Ok(v) = std::env::var("OPENAI_CODEX_API_URL") {
+                            cfg.api_base_url = v;
+                        }
+                        if let Ok(v) = std::env::var("OPENAI_CODEX_CLIENT_ID") {
+                            cfg.client_id = v;
+                        }
+                        if let Ok(v) = std::env::var("OPENAI_CODEX_SESSION_PATH") {
+                            cfg.session_path = std::path::PathBuf::from(v);
+                        }
+                        cfg
+                    })
+                };
+                let mgr = ironclaw::llm::OpenAiCodexSessionManager::new(codex_config)
+                    .map_err(|e| anyhow::anyhow!("{}", e))?;
+                mgr.device_code_login()
+                    .await
+                    .map_err(|e| anyhow::anyhow!("{}", e))?;
+                println!(
+                    "OpenAI Codex authentication complete. Set LLM_BACKEND=openai_codex to use it."
+                );
+            } else {
+                println!("Specify a provider to authenticate with:");
+                println!("  ironclaw login --openai-codex   (ChatGPT subscription)");
+            }
+            return Ok(());
+        }
        Some(Command::Onboard {
            skip_auth,
            channels_only,
@@ -3,7 +3,7 @@
 //! The wizard guides users through:
 //! 1. Database connection
 //! 2. Security (secrets master key)
-//! 3. Inference provider (NEAR AI, Anthropic, OpenAI, Ollama, OpenAI-compatible)
+//! 3. Inference provider (NEAR AI, Anthropic, OpenAI, OpenAI Codex, Ollama, OpenAI-compatible)
 //! 4. Model selection
 //! 5. Embeddings
 //! 6. Channel configuration
@@ -1083,8 +1083,10 @@ impl SetupWizard {
            print_info(&format!("Current provider: {}", display));
            println!();

-            let is_known =
-                current == "nearai" || current == "bedrock" || registry.is_known(&current);
+            let is_known = current == "nearai"
+                || current == "bedrock"
+                || current == "openai_codex"
+                || registry.is_known(&current);

            if is_known && confirm("Keep current provider?", true).map_err(SetupError::Io)? {
                if current == "bedrock" {
@@ -1093,6 +1095,10 @@ impl SetupWizard {
                    print_info("Keeping existing AWS Bedrock configuration.");
                    return Ok(());
                }
+                if current == "openai_codex" {
+                    print_info("Keeping existing OpenAI Codex configuration.");
+                    return Ok(());
+                }
                return self.run_provider_setup(&current, &registry).await;
            }

@@ -1107,7 +1113,7 @@ impl SetupWizard {
        print_info("Select your inference provider:");
        println!();

-        // Build menu: NearAI first, then all registry providers with setup hints, then Bedrock
+        // Build menu: NearAI first, then OpenAI Codex, then registry providers, then Bedrock
        let selectable = registry.selectable();
        let mut options: Vec<String> = Vec::with_capacity(2 + selectable.len());
        let mut provider_ids: Vec<String> = Vec::with_capacity(2 + selectable.len());
@@ -1115,6 +1121,9 @@ impl SetupWizard {
        options.push("NEAR AI          - multi-model access via NEAR account".to_string());
        provider_ids.push("nearai".to_string());

+        options.push("OpenAI Codex     - ChatGPT subscription (Plus/Pro/Max)".to_string());
+        provider_ids.push("openai_codex".to_string());
+
        for def in &selectable {
            let label = format!(
                "{:<17}- {}",
@@ -1158,6 +1167,10 @@ impl SetupWizard {
            return self.setup_nearai().await;
        }

+        if provider_id == "openai_codex" {
+            return self.setup_openai_codex().await;
+        }
+
        let def = registry
            .find(provider_id)
            .ok_or_else(|| SetupError::Config(format!("Unknown provider: {}", provider_id)))?;
@@ -1490,6 +1503,29 @@ impl SetupWizard {
        Ok(())
    }

+    /// OpenAI Codex (ChatGPT subscription) setup: device code OAuth flow.
+    async fn setup_openai_codex(&mut self) -> Result<(), SetupError> {
+        self.settings.llm_backend = Some("openai_codex".to_string());
+        if self.settings.selected_model.is_some() {
+            self.settings.selected_model = None;
+        }
+
+        use crate::config::OpenAiCodexConfig;
+        use crate::llm::OpenAiCodexSessionManager;
+
+        let config = OpenAiCodexConfig::default();
+
+        let mgr = OpenAiCodexSessionManager::new(config).map_err(|e| {
+            SetupError::Config(format!("OpenAI Codex session manager init failed: {}", e))
+        })?;
+        mgr.device_code_login().await.map_err(|e| {
+            SetupError::Config(format!("OpenAI Codex authentication failed: {}", e))
+        })?;
+
+        print_success("OpenAI Codex configured (ChatGPT subscription)");
+        Ok(())
+    }
+
    /// Generic Ollama-style setup: just needs a base URL, no API key.
    fn setup_ollama_generic(
        &mut self,
@@ -2963,6 +2999,7 @@ impl SetupWizard {
                "ollama" => "Ollama",
                "openai_compatible" => "OpenAI-compatible",
                "bedrock" => "AWS Bedrock",
+                "openai_codex" => "OpenAI Codex",
                other => other,
            };
            println!("  Provider: {}", display);