custom

customPreset() in tui/internal/preset/preset.go is a user-supplied OpenAI-compatible template: model is empty until configured, the key slot is LLM_API_KEY, and the endpoint is user-supplied. Its vision capability inherits the configured LLM endpoint.

Template-specific settings

Whether images work is therefore unknown until the actual provider, model, protocol, and endpoint are identified. The vision tool still tries the current OpenAI-compatible endpoint, model, and credential by default instead of treating that uncertainty as setup-time manual-only.

Read the configured endpoint’s official documentation on demand. Useful protocol references are the OpenAI Chat API, Anthropic Messages API, or Gemini API, as applicable; none is a claim about an unspecified endpoint. No plan-level vision MCP can be asserted.

If direct vision fails, report the endpoint-specific limitation and let the agent choose an explicit skill. Do not guess credentials, switch providers, or auto-load/invoke an MCP. Verify all inherited fields in TUI source and inspect the saved manifest for the user’s actual configuration.

Responses transport

For provider: "custom" with api_compat: "openai", selecting the explicit wire_api: "responses" path reveals a separate Responses transport choice:

• http is the default and is omitted from the saved manifest.
• websocket writes responses_transport: "websocket" and opts into

Responses WebSocket v2.

The selector appears only in that exact scope. Changing provider, api_compat, or wire_api removes a stale responses_transport value before save. Selecting WebSocket does not claim that an arbitrary compatible endpoint supports it: a matching Kernel version performs the actual handshake and reports that the configured upstream does not support or rejected Responses WebSocket v2. It does not silently fall back to HTTP; choose http explicitly when the upstream lacks WebSocket support.

Responses reasoning effort

For provider: "custom" with api_compat: "openai", selecting the explicit wire_api: "responses" path reveals a reasoning-effort choice:

• default removes thinking from the saved manifest and preserves the

Kernel's existing custom-provider default of high.

• none, minimal, low, medium, high, and xhigh write the selected

value to thinking.

The selector appears only in that exact scope. Changing provider, api_compat, or wire_api removes a stale custom Responses thinking value before save. The Codex provider keeps its existing effort choices and its existing xhigh default.

Operations

For base URL/API-compat/model/capability declaration shape versus credentials, see reference/operations/endpoint-capabilities/SKILL.md.