/boost

Add advisor judgment during work and a bounded completion review

What it does

Boost gives the main agent access to a private advisor while it keeps ownership of the task. The agent can consult the advisor about a consequential decision or a failure that is not converging. A quiet observer can also suggest a correction at a normal model boundary.

/boost                 enable for this session
/boost <task>          enable and run this task
/boost off             disable for this session
/boost status          phase, consultations, usage, and intervention outcomes
/boost trace           recent advisor decisions

Fusion and Boost are mutually exclusive within a session. Enabling Boost disables Fusion there; other sessions keep their own settings.

During work

The main agent starts immediately. It can call consult_advisor with a focused question without handing implementation to another agent. Direct consultations are bounded to four per run. Automatic observation uses the trajectory to choose whether to review again after the next tool batch, after several batches, or at completion. New failures and user steering can reopen observation.

The advisor receives a bounded evidence bundle containing the task, recent user updates, tool outcomes, and earlier advice. It cannot execute commands or edit files. During automatic observation it may request up to two read-only research questions. Advice tied to superseded evidence or an earlier user request is withheld.

The configured background advisor model is used when available, followed by explicit advisor configuration and then the main model as a fallback. The main agent's selected model stays unchanged.

Completion

Substantive tasks receive a final intent and verification review. Simple social exchanges without tools bypass it. A material gap can trigger one repair cycle, followed by a review of that repair; there is no unbounded repair loop.

Advisor failures allow the main agent to continue and are recorded as degraded operation. Direct consultations and final review can add latency. /boost status reports advisor and research usage.

What the review establishes

Advisor judgments depend on the evidence supplied. A second opinion does not prove a fix; actual tests and inspection remain necessary. Use /bug-fix when a reproducible failure and regression test are needed.

Related