user advisor

Advisor

The advisor lets the agent ask a stronger model for strategic guidance mid-task. It is one model call over the live conversation: no tools, no child loop, no file changes. It is not a sub-agent; for delegated work see Sub-agent delegation.

When it helps

• before committing to an implementation or review approach
• when requirements are ambiguous or tradeoffs matter
• after a failed check when the next fix is unclear
• before declaring a plan, implementation, or review complete

The advisor steers; it does not replace tests, reviews, or local evidence. The agent decides what to do with its answer.

Enabling it

The advisor is off by default in interactive and --exec runs:

advisor:
  enabled: true

models:
  profiles:
    default:
      default_model: openai/<default-model-id>
      advisor: anthropic/<stronger-model-id>

The profile's advisor model falls back to its default_model. Oneshot phases always have the advisor available, whatever advisor.enabled says. Caps, output-token limit, and timeout are listed under Configuration.

Who can call it

• The main agent.
• code, documentor, review, and evaluate sub-agents. explore, research, sanity_check, and vision cannot.

The agent can pass a question and a list of files to include verbatim, since the advisor cannot read files itself. Both are optional; a bare call asks for guidance on the conversation as it stands.

Usage caps

• advisor.max_uses_per_run caps calls by the main agent across the whole session (default 3).
• advisor.max_uses_per_sub_agent caps calls by each sub-agent over its lifetime, including after follow_up (default 1).

Once a cap is reached, further calls do not reach the model. The agent gets:

advisor budget exhausted for this session (N/N); proceed on your own judgment

The tool stays listed after the cap, so the available tools do not change mid-session.

In the TUI

Each advisor call shows as a collapsible advisor panel with the model and the use count.

Cost and safety

• Each call sends the caller's conversation to the advisor model, so calls late in a long session can be large and slow. The caps bound cost and latency; advisor.timeout applies to each attempt.
• The advisor has no tools and cannot change local state.
• It uses the credentials of the provider configured for the advisor model.