pi-advisor-omp

OMP-compatible fork of @juicesharp/rpiv-advisor.

What it does

pi-advisor-omp adds an advisor() tool and /advisor command. The tool sends the current conversation branch to a configured stronger reviewer model through a direct side-call, with tools disabled for the reviewer.

It is not a before/after-every-action hook. The advisor should be used at phase boundaries where second judgment matters:

• after enough context has been gathered, before committing to a plan or architecture;
• before the first write/edit/destructive command on multi-file, model-routing, plugin-loading, data, security, or other high-stakes work;
• when stuck after repeated failures or contradictory evidence;
• before declaring substantive work done, after durable output and focused verification exist;
• before irreversible external-system, deployment, destructive git/filesystem, auth, provider, or public API decisions.

Claude Code ideas carried into this fork

The trigger design borrows Claude Code's fork/subagent shape without copying its runtime model:

• escalation is model-invoked instead of blindly automatic;
• context is branch-copied for cache locality;
• tool definitions stay disabled for the reviewer to avoid recursive tool chains;
• trigger policy values are prompt guidance, not lifecycle hooks: required means the executor is instructed to call advisor() at that gate, but OMP does not force-run the tool before an action;
• prompt guidance makes the reviewer a challenge reviewer, not a stricter linter;
• runtime guards reject nested advisor calls and repeated successful calls for the same session state.

Configuration

Run /advisor in OMP and select a reviewer model and effort. For JJ's OMP setup, the recommended reviewer is:

{
  "modelKey": "openai-codex:gpt-5.5",
  "effort": "xhigh"
}

Safe trigger policy example:

{
  "disabledForModels": [
    { "model": "openai-codex:gpt-5.5", "minEffort": "high" }
  ],
  "guidance": {
    "triggerPolicy": {
      "mode": "phase-gated",
      "planning": "remind",
      "beforeFirstEdit": "remind",
      "stuck": "required",
      "preDone": "remind",
      "highRisk": "required",
      "maxPerTurn": 1,
      "maxPerPhase": 1
    }
  },
  "modelKey": "openai-codex:gpt-5.5",
  "effort": "xhigh"
}

Trigger guidance is registered when the plugin loads. After editing guidance.triggerPolicy by hand, restart or reload OMP before expecting the model prompt to reflect it. Legacy configs using "auto" are accepted and treated as "required"; new configs should use "required" so nobody mistakes it for a runtime hook. Configuration is persisted at ~/.config/rpiv-advisor/advisor.json, matching upstream.

Important behavior

• /advisor configures the reviewer model; it does not itself review work.
• advisor() takes no parameters.
• The reviewer receives the current branch context plus tool inventory, but tools: [] prevents it from calling tools.
• The plugin blocks advisor availability for configured executor model/effort combinations.
• A nested advisor call returns an error instead of launching another reviewer call.
• After a successful advisor response, a second advisor call for the same session state returns a cooldown error; take a concrete research, implementation, or verification step first.