Thinking effort#

Each vendor's API controls reasoning depth with a different parameter (an effort string, reasoning_effort, a token budget). OpenProgram unifies them into a single effort slider.

Levels#

The framework defines off plus six levels:

minimal · low · medium · high · xhigh · max

Each model supports only a subset, and the UI shows the levels the model actually supports. For example, Opus 4.8 has five levels from low to max, DeepSeek reasoner is not adjustable (it always reasons at full effort), and the GitHub Copilot path has no thinking support. Level data comes from each provider's declaration (the thinking block in provider.json, with per-model overrides); providers whose endpoint reports capabilities — Anthropic's /v1/models, the Codex subscription models endpoint — refresh the per-model levels on Fetch. Providers with no declaration at all fall back to the three levels low / medium / high.

Defaults#

The default level is declared by each provider (default_effort in provider.json):

How to change it#

Three layers from outermost to innermost, with inner layers overriding outer ones:

1. Agent configuration thinking_effort: the agent's default level (agent settings in the Web UI).
2. Project configuration: a project-level default that overrides the agent default.
3. Per turn: the level picker in the chat UI (next to the Web UI input box), or the /effort command in the TUI; applies turn by turn to the current session.

The selected level persists with the session, and the provider translates the framework level into the vendor API's actual parameter before sending each request.

For the full record of mapping rules and detection strategy, see the design notes.