Repository navigation
AgentPort 3.1.30
Haiku-only usage protection now blocks serialised requests larger than 256 KiB before forwarding, including repeated retries and both Codex compaction paths. The error identifies instruction, conversation and tool sizes. This safeguard does not truncate context or change native GPT settings. It is a protective byte limit, not a token or billing estimate, and may block legitimate long Haiku conversations. The underlying context inflation and ordinary-session cost efficiency remain under investigation. No paid provider request was made during verification.
Routed names in Codex lead with the provider, for example (OpenCode Go) Claude Haiku 5.5 and (OpenRouter) Qwen. Providers and models sort alphabetically in API Models, with confirmed free models grouped first. Certified Codex subagent routes retain priority slots so agent spawning remains available. Existing visibility selections are preserved.
NVIDIA NIM setup is available at API Models > Connect provider > NVIDIA NIM. Enter the NVIDIA key there, refresh the catalogue and select models. The provider listing does not imply free quota or access to every model.
Verified: 165 routing/catalogue/discovery regression tests; 107 desktop tests, with six platform/environment skips; syntax and TypeScript checks. Local mock integration tests confirmed zero upstream calls for oversized requests, retries and compaction, and one upstream call for a small request.
Available for Windows x64 and Apple Silicon macOS. The Mac preview supports Codex and API providers; local inference is not included. It is ad hoc signed, not Apple-notarised. Fully quit and reopen Codex after updating to refresh its model picker.