Feature request: Continuous advisor with direct feedback into the running agent (Oh My Pi-style)

What variant of Codex are you using?

Codex App; the same capability would also be useful in CLI and IDE sessions.

What feature would you like to see?

Add an optional continuous advisor, modeled on Oh My Pi's advisor: a separately configured reviewer model watches the primary agent's progress and feeds actionable advice directly back into the running session.

The use case is catching missed requirements, faulty assumptions, unproductive loops, and premature completion while work is happening, without repeatedly asking for a review or copying feedback between threads.

Requested behavior:

• Enable or disable the advisor per session, with its model and reasoning effort configurable independently of the primary agent.
• Automatically review incremental transcript updates, including the user request, agent actions, tool results, and current changes. Allow read-only workspace inspection by default.
• Deliver concise findings directly to the primary agent at safe execution boundaries. Distinguish minor suggestions from concerns and blockers so serious problems can steer ongoing work.
• Keep the primary agent responsible for implementing fixes. Advisor advice must not grant approvals or bypass existing sandbox and permission controls.
• Support advisor-specific project instructions, similar to Oh My Pi's WATCHDOG.md, to define review priorities.
• Make advisor activity, findings, and usage visible. Deduplicate repeated advice and bound waiting so a slow or failing advisor does not indefinitely stall the primary agent.

For example, while Codex fixes a bug, the advisor notices that the same function has another affected caller and sends that finding to the primary session. Codex can investigate and fix the shared cause before declaring the task complete.

Additional information

Related requests:

• #36837 proposes bounded second opinions at selected decision checkpoints.
• #29530 proposes a conversational reviewer sidecar with review state and generated follow-up prompts.

This request focuses on an automatically running transcript observer with direct feedback into the active agent loop, as in Oh My Pi, rather than requiring a checkpoint consultation or user-mediated reviewer-to-worker handoff. These proposals could share underlying infrastructure.

No implementation files, tests, or entry points are named. First read related issues #36837 and #29530, then trace the Codex App, CLI, and IDE session boundaries. Done means an agreed design covering advisor configuration, transcript observation, safe-boundary feedback, permissions, visibility, deduplication, and bounded waiting.