Skip to content

Auto-Advisor Mode

@advisor provides an independent second opinion on blocking decisions only — and only when genuinely warranted (frugality rule in advisor protocol). Non-blocking decisions proceed by stating assumptions.


Modes

ModeBehavior
off (default)@advisor is never dispatched; orchestrator decides alone. Manual @advisor calls still work.
lite@advisor is dispatched; presents both perspectives to the user to decide.
full@advisor is dispatched; FACTUAL questions with confidence >= 8 → auto-executes (max 10/session, then degrades to lite); otherwise presents to user (lite flow).

Switching

/auto-advisor off
/auto-advisor lite
/auto-advisor full

The auto-advisor-mode plugin writes the config before the LLM sees the command, so transitions are code-level reliable.


State persistence

  • Storage: autoAdvisorMode field in opencode.jsonc — no hidden state files, no environment variables. Values: off / lite / full.
  • Resolution: Project config (opencode.jsonc or .opencode/opencode.jsonc) → off (default). Purely project-scoped — no global fallback.
  • Project-only writes: /auto-advisor <mode> updates the field in the project's opencode.jsonc (preserving comments and other fields); never modifies global config.
  • Persists across sessions and processes, scoped to the individual project.

Red-team stance (adversarial design review)

An optional dispatch where @advisor argues AGAINST a proposal instead of balancing options:

  • Triggers: user asks explicitly ("压测这个方案" / "red team this" / "唱反调"), or orchestrator auto-triggers before irreversible design decisions (schema migration, public API contract, auth redesign, destructive data ops).
  • Output: verdict (HOLDS / HOLDS WITH CAVEATS / FAILS) + severity-ranked attack list + steelmanned defense.
  • On FAILS: orchestrator re-dispatches the design owner with attacks for rebuttal, then presents both to the user.
  • Auto-execute isolation: red-team output never carries a confidence score; code-level guard suppresses all auto-execute directives — adversarial verdicts can never trigger full-mode auto-execute.

Released under the MIT License.