| name | advisor |
| description | Adopt and verify Anthropic's advisor primitive for an installed crew. Checks Claude Code support, configures advisorModel without overwriting settings, adds tightly scoped escalation discipline to high-judgement agents and verifies the result. |
| disable-model-invocation | true |
Adopt the advisor for the crew
The advisor primitive lets a working Claude model consult a stronger model inside the same
session. It is not another subagent and does not change the one-level crew topology.
Guardrails
- Inspect the installed Claude Code version and existing settings before changing anything.
- Back up
~/.claude/settings.json; it may contain secrets and permission rules. Never
print or commit its contents.
- Preserve every existing setting. Add or update only
advisorModel.
- The advisor must be at least as capable as the main model. If the desired pairing is
uncertain or unsupported, stop and explain the pairing instead of silently disabling it.
- Advisor calls consume plan/API quota. Reserve them for judgement points, repeated failure
and final independent review; do not add advisor instructions to every low-judgement role.
Runbook
- Run
claude --version and confirm that the installed release exposes /advisor or the
advisorModel setting. If it does not, recommend updating Claude Code and stop.
- Read the current main model without exposing other settings.
- Ask which advisor model to use if the user has not specified one. Recommend the strongest
model available within their quota, with the main/advisor compatibility explained.
- Create a timestamped backup of
~/.claude/settings.json.
- Parse the JSON, change only
advisorModel, and write valid formatted JSON.
- Add a short Reaching for the advisor section only to installed high-judgement agents
such as the orchestrator, architect, code auditor, problem solver and web builder. Keep
any existing tailored version instead of duplicating it.
- If the same crew is installed for Codex, do not add Anthropic advisor instructions to
Codex agents; Codex uses its own reviewer/subagent mechanisms.
Verification
- Parse the settings JSON.
- Report the advisor model value without printing unrelated settings.
- Confirm which agent files contain the escalation discipline.
- Start a fresh Claude Code session and run
/advisor to verify the primitive is available.
If live invocation is unavailable, distinguish a configuration success from an unverified
runtime result.
The discipline block should say to consult the advisor only:
- before committing to a consequential approach where judgement changes the outcome;
- after the same failure repeats twice;
- before declaring high-impact work complete.
On routine or mechanical work, do not consult it.