feat(zarvis): codex-oauth provider — ChatGPT subscription via OAuth

crates/adapter-zarvis/src/agent.rs

@@ -692,6 +692,7 @@ impl ResolvedModel {
             provider::routing::Provider::OpenAI => "openai",
             provider::routing::Provider::Anthropic => "anthropic",
             provider::routing::Provider::Ollama => "ollama",
+            provider::routing::Provider::CodexOauth => "codex-oauth",
         }
     }
 }
@@ -733,6 +734,9 @@ pub fn resolve_model_from_spec(spec_str: &str) -> Result<ResolvedModel> {
             Box::new(provider::anthropic::Anthropic::from_env()?)
         }
         provider::routing::Provider::Ollama => Box::new(provider::ollama::Ollama::from_env()?),
+        provider::routing::Provider::CodexOauth => {
+            Box::new(provider::codex_oauth::CodexOauth::from_env()?)
+        }
     };
     Ok(ResolvedModel {
         model: spec.model,

crates/adapter-zarvis/src/context.rs

@@ -33,6 +33,13 @@ pub fn context_window_tokens(provider: &str, model: &str) -> usize {
         ("openai", _) => 32_000,
         ("anthropic", _) => 200_000,
         ("ollama", _) => 8_000,
+        // ChatGPT-subscription Codex backend. Same gpt-5* family,
+        // same advertised context window as the platform API — the
+        // billing pipe is what differs, not the model. Starting
+        // value; the runtime overflow-learn path in `model_limits.rs`
+        // will tighten if the subscription tier enforces something
+        // lower in practice.
+        ("codex-oauth", _) => 400_000,
         _ => 8_000,
     }
 }