Desktop-bundled Codex: guardian_subagent approvals reviewer enabled without consent; background memory jobs drain plan quota while idle

Environment

• ChatGPT macOS desktop app with bundled Codex engine 0.146.0-alpha.9.2 (previously 0.145.0-alpha.30), macOS (Darwin 25.5.0)
• Standalone npm CLI on the same machine is 0.142.5 (stable) — not involved
• Paid ChatGPT plan; usage observed via the rate-limit meter

Issue 1 — Guardian approvals reviewer enabled without consent

After an app auto-update, approvals_reviewer = "guardian_subagent" appeared in my ~/.codex/config.toml. I never opted in (my own deliberate config edits are all annotated with dated comments; this line appeared unannotated).

Every command approval now spawns a guardian subagent session (source: {subagent: {other: "guardian"}}) that reads AGENTS.md plus a transcript excerpt and outputs {"outcome":"allow"}.

Observed cost per decision: ~14.5k input tokens (e.g. input_tokens 14,482 / cached 4,352 / output 81).

Over two days (Aug 2–3): 78 guardian sessions, ≈7.0M total tokens (≈1.4M non-cached), all counted against my plan limits, with no UI indication.

Issue 2 — Background memory pipeline retries on oversized transcripts while idle

The bundled app runs memory_stage1 jobs over past session transcripts (memories_1.sqlite: jobs / stage1_outputs). These run while the machine is idle, with no user action.

One job against a 30MB rollout fails with:

Codex ran out of room in the model's context window. Start a new thread or clear earlier history before retrying.

and is re-scheduled with hourly backoff (retry_at), burning tokens on each attempt. The queue table currently holds 186 jobs, 36 not done.

Issue 3 — No visibility

My usage meter drains while I type nothing. There is no surface in the app showing that guardian / memory / goals background features are consuming quota.

Requests

1. Make quota-consuming background features opt-in, or at least announce them with a visible toggle.
2. Skip or cap memory jobs on transcripts that exceed the context window instead of retrying.
3. Provide a per-feature usage breakdown so quota drains are explainable.