Advisor (Second-Model Review)

Purpose

A second model passively reviews each turn and injects brief notes - a cheap sanity check on the active model’s output without changing the main loop.

How it works

• src/advisor/runtime.ts orchestrates the reviews; transcript-recorder.ts captures the turn transcript; watchdog.ts bounds runaway review work.
• The advisor model is taken from the advisor model role; when enabled, each turn’s transcript goes to the advisor and its note is injected as a hidden follow-up/steering message (see src/session/agent-session.ts onTurnEnd).
• /advisor [on|off|status|dump [raw]] controls and inspects it.

Configuration

advisor:
  enabled: true        # master toggle
  subagents: false     # also review spawned task/eval subagents
  syncBacklog: ...     # synchronize any backlogged reviews

Real example

> /advisor on
  (every turn: a second model reviews, note injected)
> /advisor status
> /advisor dump
> /advisor off

Failure behavior

• Advisor failures do not break the main turn - review happens after the turn ends; errors are logged and the turn stands.
• /advisor status reflects the actual runtime state; a misconfigured advisor model role shows as such.

Limitations

• Adds latency and extra model calls per turn (cost multiplier).
• Reviews are injective notes; they can’t veto the main model’s output.