ti-code-providers

Effect-native provider adapters and opt-in provider plugins for ti-code.

The current package includes OpenAI streaming with an opt-in official @effect/ai-openai Responses transport, OS-backed API-key storage with explicit file fallback, a local Obvious Responses SSE adapter with model discovery, an experimental OpenAI Responses WebSocket transport, an OpenCode Zen-only adapter and model helper, an opt-in Exa search client and plugin factory, ChatGPT auth/model helpers with Codex Responses WebSocket as the default transport, and the OpenAI Codex OAuth helper used by the CLI. The official Effect transport is opt-in while the existing transports remain available for compatibility. The Anthropic adapter remains deferred while the first vertical slice stays focused on the OpenAI/ChatGPT path.

Exports

• ti-code-providers - main provider surface.
• ti-code-providers/openai - OpenAI adapter with Chat Completions SSE by default plus optional Responses SSE via TI_OPENAI_TRANSPORT=responses-sse, Responses WebSocket via TI_OPENAI_TRANSPORT=responses-websocket, or official Effect Responses via TI_OPENAI_TRANSPORT=official-responses.
• ti-code-providers/obvious - local Obvious Responses SSE adapter and /models catalog reader, defaulting to the Obvious default model. Runtime config prefers OBVIOUS_* variables but also accepts direct-mode aliases used by external launchers (CODEX_DIRECT_*, OBDEX_DIRECT_*, and OPENAI_*).
• ti-code-providers/chatgpt - ChatGPT model and request helpers. The default transport is Codex Responses WebSocket with pre-event SSE fallback; pass transport: "sse" to force HTTP SSE.
• ti-code-providers/opencode - OpenCode Zen-only adapter, configured for https://opencode.ai/zen/v1 and the TI_OPENCODE_ZEN_* environment family.
• ti-code-providers/exa - Exa search client and opt-in plugin factory, wired only when EXA_API_KEY is present and a consumer explicitly adds the plugin.
• ti-code-providers/auth-storage - credential storage boundary: OAuth metadata remains file-backed, while API-key values prefer macOS Keychain or Linux Secret Service and fall back to the permissions-restricted auth file when unavailable.
• ti-code-providers/auth - API-key backend and command-runner contracts for Keychain, Secret Service, and explicit file fallback.
• ti-code-providers/openai-codex-oauth - provider-neutral OAuth flow.
• ti-code-providers/openai-codex-oauth-node - Node-specific OAuth helper.

Runtime

The public npm package targets Node.js 20 or newer. Runtime-specific portable bundles are produced by the CLI release workflow rather than this package.

The default OpenAI transport remains HTTP SSE. Responses WebSocket mode requires a runtime whose globalThis.WebSocket accepts handshake headers. Automatic stream replay/reconnect is not implemented in the provider layer; the official Effect transport supplies typed responses/errors while reconnect semantics need a proven no-replay design upstream.

The ChatGPT/Codex Responses WebSocket path is now the default ChatGPT transport and is covered by deterministic provider tests for URL normalization, header/frame construction, event decoding, close errors, and pre-event SSE fallback. A live gpt-5.5 probe passed on 2026-05-24, and the adapter strips max_output_tokens from ChatGPT WebSocket frames because the live endpoint rejected that field. Use the secret-safe probe when validating another account, model, runtime, or header shape:

pnpm exec tsx packages/providers/scripts/probe-chatgpt-codex-wss.ts

See docs/research/chatgpt-codex-wss-validation.md for the evidence boundary and inferred details.

OpenCode Zen is the Zen endpoint only. The Go-based OpenCode client is not part of this package or its public surface.