Reasoning & Web Search

Configure provider-specific model reasoning and built-in web search for action: chat.completion.

Examples use OpenRouter; make the key reachable as shown in the overview.

Reasoning

The thinking object accepts:

Field	Type	Default	Description
enabled	boolean	false	Enable provider-specific reasoning mode.
effort	string	medium	Reasoning effort, from least to most: none, minimal, low, medium, high, xhigh, or max. Applies only when enabled is true.
budget_tokens	integer	provider-specific	Explicit reasoning token budget. Overrides effort where the provider accepts a budget.
include_in_output	boolean	false	Reserved. The current chat providers do not consistently apply this field.

Support and accepted limits depend on the provider and model:

Provider	Mapping
Anthropic	Uses adaptive thinking at the requested effort. With budget_tokens, sends a fixed thinking budget instead, which only models without adaptive thinking accept, such as Claude Haiku 4.5. With adaptive thinking, an unset max_tokens defaults to 32768 instead of 4096.
OpenAI, OpenCode, openai-codex	Uses reasoning effort. Sampling fields unsupported by a reasoning model are omitted.
Gemini	Uses a thinking level, or a token budget when budget_tokens is set.
OpenRouter	Maps the common settings to its unified reasoning configuration.
Z.AI	Turns thinking on or off. budget_tokens is ignored.
Local	Sends standard chat-completion fields; local-provider reasoning controls are not currently serialized.

A provider that lacks an effort level uses its nearest supported one, so a model fallback list that spans providers keeps working with one effort:

effort	Anthropic	OpenAI, OpenCode, openai-codex, OpenRouter	Gemini	Z.AI
none	Thinking omitted	none	minimal	Off
minimal	low	minimal	minimal	On
low	low	low	low	On
medium	medium	medium	medium	On
high	high	high	high	On
xhigh	xhigh	xhigh	high	On
max	max	xhigh	high	On

none turns thinking off where the provider and model allow it. Claude models that always think, such as Claude Opus 5.5, run at their default effort instead, and Gemini has no off level.

Claude models do not accept a forced tool call while reasoning. When a step needs one, such as a step with output_schema, Dagu offers the tool without forcing it for Anthropic and for Claude models through OpenRouter.

Web Search

Anthropic and Gemini can use provider-native search. OpenRouter uses its web-search plugin.

steps:
  - id: search_news
    action: chat.completion
    with:
      provider: openrouter
      model: deepseek/deepseek-v4-flash
      web_search:
        enabled: true
        max_uses: 3
      prompt: |
        What is the current stable version of the Linux kernel?
        Reply with just the version number.

For Anthropic, domain filtering is also available:

      web_search:
        enabled: true
        max_uses: 3
        allowed_domains:
          - example.com

Field	Type	Description
enabled	boolean	Enable the provider's built-in web-search integration.
max_uses	integer	Anthropic: maximum search invocations. OpenRouter: maximum results. Ignored by Gemini.
allowed_domains	array	Restrict results to these domains. Anthropic only.
blocked_domains	array	Exclude these domains. Anthropic only.
user_location	object	Approximate city, region, country, and timezone. Anthropic only.

For Anthropic, set either allowed_domains or blocked_domains, not both.

Web search cannot be combined with output_schema, whose answer comes from a tool call; search in an earlier step and pass the result on.

If a DAG whose tool name is web_search is also listed in tools, Dagu disables the built-in search integration for that request. The DAG tool remains available for the model to call; it is not called automatically.

Related

• LLM Completion for basic usage and configuration
• Providers & Endpoints for provider credentials and endpoints
• Tool Calling for exposing DAGs as model tools