Reasoning Effort

Adds a reasoning-effort selector between Agent Zero's preset and agent-profile controls, plus an /effort slash command.

The plugin asks its authenticated local API for the active chat model's capabilities. The API reads LiteLLM's bundled model registry first, then falls back to live model metadata from Agent Zero API, Venice, and OpenRouter. Provider requests use Agent Zero's server-side configuration and are cached for 15 minutes. Provider-specific labels remain separate from the canonical values passed to LiteLLM.

Compatibility fallbacks cover GLM-5.3 (low, high, max) on Z.AI, Z.AI Coding, Agent Zero API, and Venice, plus gemini-4-argon (low, medium, high) on Agent Zero API and Venice. These fill gaps in LiteLLM and gateway metadata; native LiteLLM support takes precedence when available. The selected value is forwarded through extra_body so provider adapters cannot drop it; only direct Z.AI requests also enable thinking.

Argon's Google model documentation and Venice model page document these three levels. Venice's catalog currently reports supportsReasoningEffort: false; accepting a request does not establish that its upstream provider honors the effort.

You can enter a custom value when neither LiteLLM nor the provider publishes a list. Models explicitly marked unsupported by the provider show Unsupported instead. A selection applies only to the current chat and model; choosing Preset / provider default removes the override.

You can change reasoning effort while the agent is running. The selection applies to the next model turn; a request already in progress keeps its original effort.

A0 Connector

Version 0.4.0 registers an Effort choice control through Core's scoped connector_controls extension point. This requires an Agent Zero Core build advertising the composer_controls capability and a compatible a0-connector build. Older builds continue to use the WebUI normally.

The anchored popover beside provider/model offers the same advertised values, custom field, and preset/provider default as the WebUI. Labels, choices, validation, and per-chat/model persistence remain in this plugin; the connector only renders generic controls. Requests use Core's authenticated, CSRF-protected control API. Disabling or removing the plugin hides the control. Escape or clicking outside cancels without changing the setting.

Slash command

Version 0.4.1 adds /effort to the WebUI slash picker and A0 CLI command palette when Core's built-in Commands plugin is enabled. It runs entirely in this plugin, without an LLM request or plugin-specific connector code.

• /effort shows the current model, effective effort, its source, and available levels.
• /effort high selects an advertised level. Use canonical values such as xhigh for Extra and none for Off, when available.
• /effort default removes the current chat/model override and restores the preset/provider default.
• /effort custom <value> uses the same custom-value validation as the popover; it cannot bypass an explicitly unsupported model or compatibility-level restrictions.

Changes share the popover's per-chat/model state and apply to the next model turn. The CLI sends plugin commands through its normal chat path; use the popover to change effort immediately during an active run.

Install

Place this directory at usr/plugins/a0_reasoning_effort in Agent Zero, then reload the WebUI. No additional dependencies are required.

Check

From the Agent Zero root:

PYTHONPATH=. python usr/plugins/a0_reasoning_effort/helpers/reasoning_effort.py
PYTHONPATH=. python usr/plugins/a0_reasoning_effort/tests/test_active_run.py
PYTHONPATH=. python usr/plugins/a0_reasoning_effort/tests/test_command.py
node usr/plugins/a0_reasoning_effort/tests/test_active_run.mjs