Advisor mode

Parent: Configuration.

Advisor mode gives the root agent an advisor tool backed by a second model. The executor keeps doing the work. The advisor only reviews the live session and returns guidance. Use a stronger or different model than the executor so the review adds a second judgment before the plan hardens, when the agent is stuck, and before it declares the work done.

How a call works

1. The executor calls advisor. The tool takes no parameters.
2. Rho builds a text transcript from the executor system prompt and the live session history, including the turn in flight. The agent cannot edit that payload.
3. Rho starts a one-shot advisor run on the configured advisor model. That run has no tools. It cannot read files, run commands, or change the workspace. This holds on both runtimes: a Claude Code advisor is spawned with --tools "".
4. While the request is in flight, the advisor tool card shows live status on the first line (waiting for provider, thinking, responding, or retrying provider) and streams guidance text into the card body as it arrives. Reasoning content stays hidden; only the phase changes to thinking.
5. The advisor finishes. Rho hands the final guidance back as an ordinary tool result. The executor keeps its turn and decides what to do next.

Rho runs the advisor itself. There is no server-side advisor and no provider beta flag, so any model in Rho's catalog can be the advisor on any provider. The advisor can also run on the installed claude binary instead, using your Claude Code subscription. See Claude Code as the advisor.

Turn it on

Advisor mode needs both switches:

1. advisor_mode = true
2. an explicit advisor model under [internal_agents.advisor]

With the mode on and no model, Rho shows advisor: no model in the status line and does not offer the tool. The advisor never falls back to the conversation model, because a reviewer that mirrors the executor adds nothing.

Ways to enable it:

• /advisor or /advisor on in the interactive TUI. Without a model, the command opens a model picker first. The mode turns on after you select one. esc leaves the mode off. /advisor off turns it off.
• /config → Agent behavior → Advisor mode, Advisor model, and Advisor reasoning
• /agents, then choose the advisor internal agent and pick a model
• a hand edit of ~/.rho/config.toml

[behavior]
advisor_mode = true

[internal_agents.advisor]
provider = "anthropic"
model = "claude-fable-5"
auth = "anthropic-api-key"
reasoning = "high"

reasoning is optional. When omitted, the advisor uses its built-in default (medium). Model aliases work in the advisor entry (model = "@deep").

Changes save at once and apply before the next turn. The session ID and history stay. You cannot toggle advisor mode while a run is active.

Claude Code as the advisor

The advisor model picker lists Claude Code alongside Rho's provider models:

claude-code/default
claude-code/fable
claude-code/opus
claude-code/sonnet
claude-code/haiku

Choosing one of those rows also chooses the runtime. Rho runs the review on the installed claude binary, and nothing else changes: the advisor still sees only the rendered transcript, still gets no tools, and still returns guidance text into the same card.

The rows appear only when the claude binary is on PATH. Sign in first with /login claude-code; Rho never stores a Claude token. A signed-out binary makes the advisor call fail with that same instruction, and the executor's turn survives it.

The equivalent config is:

[behavior]
advisor_mode = true

[internal_agents.advisor]
runtime = "claude-cli"
model = "opus"
reasoning = "high"

• runtime defaults to rho, so existing config keeps working unchanged.
• model is a Claude alias or full Claude model name, passed through as --model. Omit it to let Claude Code choose. Rho @alias references are not resolved here.
• provider and auth do not apply. Claude Code owns both, and Rho drops them with a warning if they appear.
• reasoning maps to Claude --effort, so only low, medium, high, xhigh, and max are offered.

Differences worth knowing before you choose it:

• Calls bill to your Claude subscription, not to a Rho provider, and they draw on the same 5-hour and 7-day windows as your own Claude Code use. Reported cost folds into the session total in the TUI. The usage ledger does not record them, matching delegated Claude runs.
• Each call spawns a process, so it starts slower than a provider request.
• The run is one turn with no tools and no session persistence, so it leaves nothing in your Claude session list.

What the advisor sees

The advisor receives a rendered transcript, not a free-form prompt from the executor:

• the executor system prompt, including the advisor steering text while the mode is on
• your requests
• assistant messages, tool calls, and tool results
• the live turn that issued the call

Large single items are clipped. When the whole body is still too large, Rho keeps the start of the session and the most recent work and elides the middle. Images become short placeholders such as [image: image/png].

When the executor should call it

While the mode is active and a model is set, Rho adds steering text to the executor system prompt. The tool description and that text both steer the executor to call advisor:

• before substantive work (writes, commits to an interpretation, or answers)
• when stuck (recurring errors, a plan that is not converging)
• when it considers a change of approach
• when it believes the task is complete, after it has made the deliverable durable

On short reactive tasks, a call is optional. On longer tasks, the prompt asks for at least one call before the approach hardens and one before the agent declares done.

In the TUI

While advisor mode is on, the status line names the reviewing model, for example advisor: anthropic/claude-fable-5, or advisor: claude-code/opus when the advisor runs on Claude Code. It stays out of the status line while the mode is off. An in-flight advisor call uses an agent-style card: advisor responding on the header with streamed guidance below. Finished advice stays on the same card as advisor completed, collapsed past the tool output limit and expandable with ctrl+o. /info also reports whether advisor mode is on.

Cost and scope

• Advisor calls bill to the advisor model's provider, or to your Claude subscription when the advisor runs on Claude Code.
• The usage ledger records provider calls under the advisor purpose. Claude Code calls are not ledger rows, the same as delegated Claude runs.
• Provider-reported advisor cost folds into the parent session total in the TUI.
• Automation runs honor advisor_mode, so rho run gets the same tool and steering text when a model is set.
• Subagents and workflow runs do not. The advisor reviews the root session. A child run has its own history and does not receive the tool.

Related settings

Setting	Role
[behavior].advisor_mode	Offers the advisor tool when a model is also set. Default: false.
[internal_agents.advisor]	The reviewer. On the default rho runtime: required provider, model, and auth. With runtime = "claude-cli": optional pass-through model only. Optional reasoning either way. No conversation-model fallback.
display.max_tool_output_lines	How many lines of advice show inline before the TUI collapses the card.

See also: /advisor, internal agent models, and the advisor tool.