feat(setup): Anthropic OAuth onboarding with setup-token support (#384)

* feat(setup): add Anthropic OAuth and Codex OAuth onboarding flows

Add OAuth token authentication as an alternative to API keys during
onboarding for both Anthropic (via `claude login`) and OpenAI/Codex
(via `~/.codex/auth.json`).

Key changes:
- New `AnthropicOAuthProvider` using `Authorization: Bearer` header
  (rig-core hardcodes `x-api-key` which rejects OAuth tokens)
- Wizard auth method selector: "Direct API Key" vs "OAuth Token"
  for both Anthropic and OpenAI providers
- Codex token extraction from `$CODEX_HOME/auth.json` / `~/.codex/auth.json`
- Claude Code sandbox sub-step in Docker setup (checks for credentials)
- Secret injection mappings for `ANTHROPIC_OAUTH_TOKEN` and `CODEX_OAUTH_TOKEN`
- `CODEX_OAUTH_TOKEN` falls back to `OPENAI_API_KEY` (same Bearer auth)

Supersedes #143 which had a broken auth flow (OAuth token sent as
x-api-key → 401). Credit to @bigguybobby for the original approach.

Co-Authored-By: Claude Opus 4.6 <[email protected]>

* fix: persist OAuth tokens in bootstrap .env and re-extract at startup

OAuth tokens stored only in the secrets DB were invisible to
Config::from_env() which runs before the DB connects (chicken-and-egg).

Two fixes:
1. write_bootstrap_env() now persists ANTHROPIC_OAUTH_TOKEN and
   CODEX_OAUTH_TOKEN to ~/.ironclaw/.env (same pattern as NEARAI_API_KEY)
2. main.rs re-extracts a fresh token from the OS credential store
   (macOS Keychain / ~/.claude/.credentials.json) before config resolution,
   handling token expiry (8-12h) gracefully

Co-Authored-By: Claude Opus 4.6 <[email protected]>

* fix: persist all LLM credentials in bootstrap .env, not just NEAR AI

All providers had the same chicken-and-egg issue: API keys stored in the
secrets DB were invisible to Config::from_env() which runs before DB
connects. Only NEARAI_API_KEY was written to bootstrap .env.

Now write_bootstrap_env() persists all credential env vars:
NEARAI_API_KEY, ANTHROPIC_API_KEY, ANTHROPIC_OAUTH_TOKEN, OPENAI_API_KEY,
CODEX_OAUTH_TOKEN, LLM_API_KEY, TINFOIL_API_KEY.

Also: setup_api_key_provider() now sets the env var during the wizard
session so write_bootstrap_env() can pick it up.

Co-Authored-By: Claude Opus 4.6 <[email protected]>

* fix: address security review findings for OAuth onboarding

- Extract "oauth-placeholder" to named OAUTH_PLACEHOLDER constant shared
  across config and wizard to prevent silent drift
- Document plaintext credential tradeoff in write_bootstrap_env (API keys
  stored with 0o600 permissions, recommend full-disk encryption)
- Add blocking "Press Enter" wait in Anthropic OAuth retry flow so user
  has time to run `claude login` in another terminal
- Add escape hatch from manual OAuth paste back to API key flow (empty
  input switches to setup_api_key_provider)
- Fix Retry-After header: parse u64 seconds into Duration before passing
  to LlmError::RateLimited
- Make config::llm module pub(crate) for constant visibility
- Use .bearer_auth() instead of manual format!("Bearer {}")
- Remove response body from debug log (may contain PII)
- Update Anthropic API version to 2024-10-22

Co-Authored-By: Claude Opus 4.6 <[email protected]>

* security: remove plaintext credentials from bootstrap .env

Credentials (API keys, OAuth tokens) were being written in plaintext to
~/.ironclaw/.env to work around a chicken-and-egg problem: Config::from_env()
runs before the encrypted secrets DB is connected.

Instead of storing secrets on disk, LlmConfig::resolve() now defers
gracefully when credentials are missing — it returns None for the provider
config instead of hard-erroring with MissingRequired. After the DB connects,
AppBuilder::build_all() loads secrets from encrypted storage via
inject_llm_keys_from_secrets() and re-resolves the config.

For Anthropic OAuth tokens (which expire in 8-12h), the secret injection
step also tries the OS credential store (macOS Keychain / Linux
credentials.json) for a fresh token, overriding the potentially stale
copy in the DB.

Changes:
- LlmConfig::resolve(): OpenAI, Anthropic, OpenAI-compatible, and Tinfoil
  all return None instead of MissingRequired when credentials are absent
- write_bootstrap_env(): no longer writes any credential env vars
- inject_llm_keys_from_secrets(): refreshes Anthropic OAuth from OS
  credential store before overlay is finalized
- main.rs: removed OAuth re-extraction hack (no longer needed)

Co-Authored-By: Claude Opus 4.6 <[email protected]>

* fix: load OS credential store tokens even without secrets DB

The OAuth token extraction from macOS Keychain / Linux credentials files
was only running inside inject_llm_keys_from_secrets(), which requires
the encrypted secrets DB. When no master key is configured, init_secrets()
returned early — skipping both DB secret loading AND OS credential store
extraction, leaving the Anthropic OAuth token unavailable.

Split into two paths:
- inject_llm_keys_from_secrets(): loads from encrypted DB + OS stores
- inject_os_credentials(): loads from OS stores only (no DB needed)

init_secrets() now calls inject_os_credentials() and re-resolves config
even in the no-master-key early-return path, so `claude login` tokens
are always available regardless of secrets DB state.

Co-Authored-By: Claude Opus 4.6 <[email protected]>

* fix: add anthropic-beta header required for OAuth authentication

Anthropic's api.anthropic.com requires the `anthropic-beta: oauth-2025-04-20`
header to accept OAuth Bearer tokens. Without it, the API returns 401
"OAuth authentication is currently not supported."

Also reverts API version to 2023-06-01 since the OAuth beta flag does
not support the 2024-10-22 version (returns 400 "not a valid version").

This was the same bug that caused PR #143's 401 errors — the beta header
was missing entirely.

Co-Authored-By: Claude Opus 4.6 <[email protected]>

* fix: Anthropic and OpenAI model resolution respects selected_model

The Anthropic and OpenAI config resolution ignored settings.selected_model
entirely, only checking the provider-specific env var (ANTHROPIC_MODEL,
OPENAI_MODEL) and falling back to a hardcoded default. This meant the
model chosen during onboarding wizard was silently overridden.

Now follows the same pattern as NearAI and OpenAI-compatible:
env var > settings.selected_model > hardcoded default.

Also deduplicated the Anthropic config construction (two identical
branches for API key vs OAuth now share model/base_url resolution).

Co-Authored-By: Claude Opus 4.6 <[email protected]>

* test: add provider resolution tests for all LLM backends

Covers deferred resolution (no credentials → None instead of error),
credential presence, model selection fallback chain, and OAuth token
routing for Anthropic, OpenAI, Tinfoil, Ollama, and NearAI.

Co-Authored-By: Claude Opus 4.6 <[email protected]>

* fix: handle nested tokens.access_token format in Codex auth.json

Codex CLI stores OAuth tokens in a nested format under
tokens.access_token (ChatGPT OAuth flow), not at the top level.
Also adds ENV_MUTEX to Codex token tests for thread safety.

Co-Authored-By: Claude Opus 4.6 <[email protected]>

* refactor: remove Codex OAuth onboarding (incompatible with OpenAI API)

Codex CLI OAuth tokens use a different endpoint
(chatgpt.com/backend-api/codex) and the Responses API wire format,
not api.openai.com with Chat Completions. The tokens lack the
model.request scope needed for the platform API, so they can't be
used as drop-in OPENAI_API_KEY replacements.

Removes: extract_codex_oauth_token(), wizard Codex OAuth flow,
CODEX_OAUTH_TOKEN env var support, and related tests.

OpenAI onboarding now uses direct API key only.

Co-Authored-By: Claude Opus 4.6 <[email protected]>

* style: fix formatting for CI (cargo fmt)

Co-Authored-By: Claude Opus 4.6 <[email protected]>

* fix: address Gemini review feedback

- Use ? operator for ANTHROPIC_MODEL/BASE_URL env resolution instead of
  .ok().flatten() to propagate ConfigErrors consistently
- Skip Tool messages without tool_call_id with a warning instead of
  using unwrap_or_default() which would send empty string to Anthropic
- Extract credential check into closure to reduce duplication in
  Claude Code sandbox setup

Co-Authored-By: Claude Opus 4.6 <[email protected]>

* refactor(review): address PR review feedback for OAuth onboarding

- Gate ANTHROPIC_OAUTH_TOKEN resolution to Anthropic provider only
  (was needlessly checked for all registry providers)
- Add 3 regression tests for OAuth config resolution:
  - oauth_token sets placeholder api_key
  - real api_key takes priority over oauth
  - non-Anthropic providers don't pick up oauth_token
- Validate OAuth token prefix (sk-ant-oat) in wizard to catch
  accidentally pasted API keys
- Improve error body read handling in AnthropicOAuthProvider
  (was silently swallowing read errors with unwrap_or_default)
- Remove extra blank line in write_bootstrap_env
- Remove stale blank line in RegistryProviderConfig doc comment

[skip-regression-check]

Co-Authored-By: Claude Opus 4.6 <[email protected]>

* fix: address PR #384 review comments

Blocker:
- Replace OnceLock<HashMap> with LazyLock<Mutex<HashMap>> for INJECTED_VARS
  so both inject_os_credentials() and inject_llm_keys_from_secrets() merge
  data instead of the second caller silently dropping its entries.

High:
- Add 401 retry with OS credential store re-extraction in
  AnthropicOAuthProvider, recovering from expired OAuth tokens (~8-12h)
  without manual intervention.
- Fix comment in app.rs: ~/.codex/auth.json → ~/.claude/.credentials.json.

Medium:
- Remove unsafe { std::env::set_var } from wizard; use thread-safe
  inject_single_var() overlay instead (safe on multi-threaded Tokio).
- Add post-init validation in AppBuilder: fail early with clear error when
  LLM_BACKEND is set but no credentials were resolved after secret injection.
- Add sk-ant-oat prefix validation in parse_oauth_access_token().
- Only route to AnthropicOAuthProvider when api_key is missing or equals
  OAUTH_PLACEHOLDER (API key takes priority over OAuth token).
- Teach fetch_anthropic_models() to use Bearer auth when only OAuth token
  is available (model listing no longer fails for OAuth-only users).

Low:
- Use optional_env() in wizard credential checks to read from injected
  overlay, not just raw env vars.

[skip-regression-check]

Co-Authored-By: Claude Opus 4.6 <[email protected]>

* style: cargo fmt

Co-Authored-By: Claude Opus 4.6 <[email protected]>

---------

Co-authored-by: Claude Opus 4.6 <[email protected]>
Co-authored-by: [email protected] <[email protected]>

src/config/helpers.rs

@@ -25,8 +25,13 @@ pub(crate) fn optional_env(key: &str) -> Result<Option<String>, ConfigError> {
     }
 
     // Fall back to thread-safe overlay (secrets injected from DB)
-    if let Some(val) = INJECTED_VARS.get().and_then(|map| map.get(key)) {
-        return Ok(Some(val.clone()));
+    if let Some(val) = INJECTED_VARS
+        .lock()
+        .unwrap_or_else(|p| p.into_inner())
+        .get(key)
+        .cloned()
+    {
+        return Ok(Some(val));
     }
 
     Ok(None)

src/config/llm.rs

@@ -9,6 +9,13 @@ use crate::llm::registry::{ProviderProtocol, ProviderRegistry};
 use crate::llm::session::SessionConfig;
 use crate::settings::Settings;
 
+/// Sentinel value used as `api_key` when only an OAuth token is present.
+///
+/// When we only have an OAuth token the provider factory in `llm/mod.rs`
+/// checks for this value and routes to `AnthropicOAuthProvider`, so this
+/// placeholder is never sent over the wire.
+pub const OAUTH_PLACEHOLDER: &str = "oauth-placeholder";
+
 /// Prompt cache retention policy for Anthropic.
 ///
 /// Controls Anthropic's automatic prompt caching via a top-level
@@ -66,6 +73,7 @@ pub struct RegistryProviderConfig {
     /// Provider identifier (e.g., "groq", "openai", "tinfoil").
     pub provider_id: String,
     /// API key (optional for some providers like Ollama).
+    /// For Anthropic OAuth, this is set to `OAUTH_PLACEHOLDER`.
     pub api_key: Option<SecretString>,
     /// Base URL for the API endpoint.
     pub base_url: String,
@@ -73,6 +81,9 @@ pub struct RegistryProviderConfig {
     pub model: String,
     /// Extra HTTP headers injected into every request.
     pub extra_headers: Vec<(String, String)>,
+    /// OAuth token for providers that support Bearer auth (e.g. Anthropic via `claude login`).
+    /// When set, the provider factory routes to the OAuth-specific provider implementation.
+    pub oauth_token: Option<SecretString>,
 }
 
 /// LLM provider configuration.
@@ -366,6 +377,22 @@ impl LlmConfig {
             Vec::new()
         };
 
+        // Resolve OAuth token (Anthropic-specific: `claude login` flow).
+        // Only check for OAuth token when the provider is actually Anthropic.
+        let oauth_token = if canonical_id == "anthropic" {
+            optional_env("ANTHROPIC_OAUTH_TOKEN")?.map(SecretString::from)
+        } else {
+            None
+        };
+        let api_key = if api_key.is_none() && oauth_token.is_some() {
+            // OAuth token present but no API key: use a placeholder so the
+            // config block is populated. The provider factory will route to
+            // the OAuth provider instead of rig-core's x-api-key client.
+            Some(SecretString::from(OAUTH_PLACEHOLDER.to_string()))
+        } else {
+            api_key
+        };
+
         Ok(RegistryProviderConfig {
             protocol,
             provider_id: canonical_id.to_string(),
@@ -373,6 +400,7 @@ impl LlmConfig {
             base_url,
             model,
             extra_headers,
+            oauth_token,
         })
     }
 }
@@ -677,8 +705,6 @@ mod tests {
 
     #[test]
     fn backend_alias_normalized_to_canonical_id() {
-        // When the user sets LLM_BACKEND to an alias (e.g., "open_ai"),
-        // LlmConfig.backend should resolve to the canonical ID ("openai").
         let _guard = ENV_MUTEX.lock().expect("env mutex poisoned");
         clear_openai_compatible_env();
         // SAFETY: Under ENV_MUTEX.
@@ -705,8 +731,6 @@ mod tests {
 
     #[test]
     fn unknown_backend_falls_back_to_openai_compatible() {
-        // An unrecognized LLM_BACKEND should fall back to the openai_compatible
-        // provider definition instead of erroring.
         let _guard = ENV_MUTEX.lock().expect("env mutex poisoned");
         clear_openai_compatible_env();
         // SAFETY: Under ENV_MUTEX.
@@ -717,7 +741,6 @@ mod tests {
 
         let settings = Settings::default();
         let cfg = LlmConfig::resolve(&settings).expect("resolve should succeed");
-        // Falls back to openai_compatible since "some_custom_provider" is unknown
         assert_eq!(cfg.backend, "openai_compatible");
         let provider = cfg.provider.expect("should have provider config");
         assert_eq!(provider.provider_id, "openai_compatible");
@@ -759,7 +782,6 @@ mod tests {
 
     #[test]
     fn base_url_resolution_priority() {
-        // Env var > settings > registry default
         let _guard = ENV_MUTEX.lock().expect("env mutex poisoned");
         clear_openai_compatible_env();
 
@@ -800,6 +822,119 @@ mod tests {
         }
     }
 
+    // ── OAuth resolution tests ──────────────────────────────────────
+
+    /// Clear all Anthropic-related env vars.
+    fn clear_anthropic_env() {
+        // SAFETY: Only called under ENV_MUTEX in tests.
+        unsafe {
+            std::env::remove_var("LLM_BACKEND");
+            std::env::remove_var("ANTHROPIC_API_KEY");
+            std::env::remove_var("ANTHROPIC_OAUTH_TOKEN");
+            std::env::remove_var("ANTHROPIC_MODEL");
+            std::env::remove_var("ANTHROPIC_BASE_URL");
+        }
+    }
+
+    #[test]
+    fn anthropic_oauth_token_sets_placeholder_api_key() {
+        use secrecy::ExposeSecret;
+
+        let _guard = ENV_MUTEX.lock().expect("env mutex poisoned");
+        clear_anthropic_env();
+        // SAFETY: Under ENV_MUTEX.
+        unsafe {
+            std::env::set_var("ANTHROPIC_OAUTH_TOKEN", "sk-ant-oat01-test-token");
+        }
+
+        let settings = Settings {
+            llm_backend: Some("anthropic".to_string()),
+            ..Default::default()
+        };
+        let cfg = LlmConfig::resolve(&settings).expect("resolve should succeed");
+        let provider = cfg.provider.expect("provider config should be present");
+
+        assert_eq!(
+            provider
+                .api_key
+                .as_ref()
+                .map(|k| k.expose_secret().to_string()),
+            Some(OAUTH_PLACEHOLDER.to_string()),
+            "api_key should be the OAuth placeholder when only OAuth token is set"
+        );
+        assert!(
+            provider.oauth_token.is_some(),
+            "oauth_token should be populated"
+        );
+        assert_eq!(
+            provider.oauth_token.as_ref().unwrap().expose_secret(),
+            "sk-ant-oat01-test-token"
+        );
+
+        clear_anthropic_env();
+    }
+
+    #[test]
+    fn anthropic_api_key_takes_priority_over_oauth() {
+        use secrecy::ExposeSecret;
+
+        let _guard = ENV_MUTEX.lock().expect("env mutex poisoned");
+        clear_anthropic_env();
+        // SAFETY: Under ENV_MUTEX.
+        unsafe {
+            std::env::set_var("ANTHROPIC_API_KEY", "sk-ant-real-key");
+            std::env::set_var("ANTHROPIC_OAUTH_TOKEN", "sk-ant-oat01-test-token");
+        }
+
+        let settings = Settings {
+            llm_backend: Some("anthropic".to_string()),
+            ..Default::default()
+        };
+        let cfg = LlmConfig::resolve(&settings).expect("resolve should succeed");
+        let provider = cfg.provider.expect("provider config should be present");
+
+        assert_eq!(
+            provider
+                .api_key
+                .as_ref()
+                .map(|k| k.expose_secret().to_string()),
+            Some("sk-ant-real-key".to_string()),
+            "real API key should take priority over OAuth placeholder"
+        );
+        assert!(
+            provider.oauth_token.is_some(),
+            "oauth_token should still be populated"
+        );
+
+        clear_anthropic_env();
+    }
+
+    #[test]
+    fn non_anthropic_provider_has_no_oauth_token() {
+        let _guard = ENV_MUTEX.lock().expect("env mutex poisoned");
+        clear_anthropic_env();
+        // SAFETY: Under ENV_MUTEX.
+        unsafe {
+            std::env::set_var("ANTHROPIC_OAUTH_TOKEN", "sk-ant-oat01-test-token");
+        }
+
+        let settings = Settings {
+            llm_backend: Some("openai".to_string()),
+            ..Default::default()
+        };
+        let cfg = LlmConfig::resolve(&settings).expect("resolve should succeed");
+        let provider = cfg.provider.expect("provider config should be present");
+
+        assert!(
+            provider.oauth_token.is_none(),
+            "non-Anthropic providers should not pick up ANTHROPIC_OAUTH_TOKEN"
+        );
+
+        clear_anthropic_env();
+    }
+
+    // ── Cache retention tests ───────────────────────────────────────
+
     #[test]
     fn cache_retention_from_str_primary_values() {
         assert_eq!(

src/config/mod.rs

@@ -13,7 +13,7 @@ mod embeddings;
 mod heartbeat;
 pub(crate) mod helpers;
 mod hygiene;
-mod llm;
+pub(crate) mod llm;
 mod routines;
 mod safety;
 mod sandbox;
@@ -24,7 +24,7 @@ mod tunnel;
 mod wasm;
 
 use std::collections::HashMap;
-use std::sync::OnceLock;
+use std::sync::{LazyLock, Mutex};
 
 use crate::error::ConfigError;
 use crate::settings::Settings;
@@ -53,7 +53,12 @@ pub use crate::llm::session::SessionConfig;
 /// Used by `inject_llm_keys_from_secrets()` to make API keys available to
 /// `optional_env()` without unsafe `set_var` calls. `optional_env()` checks
 /// real env vars first, then falls back to this overlay.
-static INJECTED_VARS: OnceLock<HashMap<String, String>> = OnceLock::new();
+///
+/// Uses `Mutex<HashMap>` instead of `OnceLock` so that both
+/// `inject_os_credentials()` and `inject_llm_keys_from_secrets()` can merge
+/// their data. Whichever runs first initialises the map; the second merges in.
+static INJECTED_VARS: LazyLock<Mutex<HashMap<String, String>>> =
+    LazyLock::new(|| Mutex::new(HashMap::new()));
 
 /// Main configuration for the agent.
 #[derive(Debug, Clone)]
@@ -285,6 +290,9 @@ impl Config {
 /// env-var-first resolution in `LlmConfig::resolve()`. Keys in the overlay
 /// 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
+/// credentials files) which don't require the secrets DB.
 pub async fn inject_llm_keys_from_secrets(
     secrets: &dyn crate::secrets::SecretsStore,
     user_id: &str,
@@ -292,7 +300,10 @@ pub async fn inject_llm_keys_from_secrets(
     // Static mappings for well-known providers.
     // The registry's setup hints define secret_name -> env_var mappings,
     // so new providers added to providers.json get injection automatically.
-    let mut mappings: Vec<(&str, &str)> = vec![("llm_nearai_api_key", "NEARAI_API_KEY")];
+    let mut mappings: Vec<(&str, &str)> = vec![
+        ("llm_nearai_api_key", "NEARAI_API_KEY"),
+        ("llm_anthropic_oauth_token", "ANTHROPIC_OAUTH_TOKEN"),
+    ];
 
     // Dynamically discover secret->env mappings from the provider registry.
     // Uses selectable() which deduplicates user overrides correctly.
@@ -331,5 +342,62 @@ pub async fn inject_llm_keys_from_secrets(
         }
     }
 
-    let _ = INJECTED_VARS.set(injected);
+    inject_os_credential_store_tokens(&mut injected);
+
+    merge_injected_vars(injected);
+}
+
+/// Load tokens from OS credential stores (no DB required).
+///
+/// Called unconditionally during startup — even when the encrypted secrets DB
+/// is unavailable (no master key, no DB connection). This ensures OAuth tokens
+/// from `claude login` (macOS Keychain / Linux credentials.json)
+/// are available for config resolution.
+pub fn inject_os_credentials() {
+    let mut injected = HashMap::new();
+    inject_os_credential_store_tokens(&mut injected);
+    merge_injected_vars(injected);
+}
+
+/// Merge new entries into the global injected-vars overlay.
+///
+/// New keys are inserted; existing keys are overwritten (later callers win,
+/// e.g. fresh OS credential store tokens override stale DB copies).
+fn merge_injected_vars(new_entries: HashMap<String, String>) {
+    if new_entries.is_empty() {
+        return;
+    }
+    match INJECTED_VARS.lock() {
+        Ok(mut map) => map.extend(new_entries),
+        Err(poisoned) => poisoned.into_inner().extend(new_entries),
+    }
+}
+
+/// Inject a single key-value pair into the overlay.
+///
+/// Used by the setup wizard to make credentials available to `optional_env()`
+/// without calling `unsafe { std::env::set_var }`.
+pub fn inject_single_var(key: &str, value: &str) {
+    match INJECTED_VARS.lock() {
+        Ok(mut map) => {
+            map.insert(key.to_string(), value.to_string());
+        }
+        Err(poisoned) => {
+            poisoned
+                .into_inner()
+                .insert(key.to_string(), value.to_string());
+        }
+    }
+}
+
+/// Shared helper: extract tokens from OS credential stores into the overlay map.
+fn inject_os_credential_store_tokens(injected: &mut HashMap<String, String>) {
+    // Try the OS credential store for a fresh Anthropic OAuth token.
+    // Tokens from `claude login` expire in 8-12h, so the DB copy may be stale.
+    // A fresh extraction from macOS Keychain / Linux credentials.json wins
+    // over the (possibly expired) copy stored in the encrypted secrets DB.
+    if let Some(fresh) = crate::config::ClaudeCodeConfig::extract_oauth_token() {
+        injected.insert("ANTHROPIC_OAUTH_TOKEN".to_string(), fresh);
+        tracing::debug!("Refreshed ANTHROPIC_OAUTH_TOKEN from OS credential store");
+    }
 }

src/config/sandbox.rs

@@ -233,9 +233,14 @@ impl ClaudeCodeConfig {
 /// Expected shape: `{"claudeAiOauth": {"accessToken": "sk-ant-oat01-..."}}`
 fn parse_oauth_access_token(json: &str) -> Option<String> {
     let creds: serde_json::Value = serde_json::from_str(json).ok()?;
-    creds["claudeAiOauth"]["accessToken"]
-        .as_str()
-        .map(String::from)
+    let token = creds["claudeAiOauth"]["accessToken"].as_str()?;
+    // Validate that the token looks like a real OAuth token before using it.
+    // Claude CLI tokens start with "sk-ant-oat".
+    if !token.starts_with("sk-ant-oat") {
+        tracing::debug!("Ignoring credential store token with unexpected prefix");
+        return None;
+    }
+    Some(token.to_string())
 }
 
 #[cfg(test)]
@@ -401,14 +406,14 @@ mod tests {
     fn parse_oauth_token_nested_extra_fields() {
         let json = r#"{
             "claudeAiOauth": {
-                "accessToken": "sk-ant-real-token",
+                "accessToken": "sk-ant-oat01-real-token",
                 "refreshToken": "rt-abc",
                 "expiresAt": 1700000000
             }
         }"#;
         assert_eq!(
             parse_oauth_access_token(json),
-            Some("sk-ant-real-token".to_string())
+            Some("sk-ant-oat01-real-token".to_string())
         );
     }
 
@@ -418,6 +423,12 @@ mod tests {
         assert_eq!(parse_oauth_access_token(json), None);
     }
 
+    #[test]
+    fn parse_oauth_token_rejects_invalid_prefix() {
+        let json = r#"{"claudeAiOauth": {"accessToken": "not-an-oauth-token"}}"#;
+        assert_eq!(parse_oauth_access_token(json), None);
+    }
+
     // ── default_claude_code_allowed_tools ───────────────────────────
 
     #[test]