Evidence record / codex
A dated record of one change, kept so the writing that cites it can be checked. Compiled from the sources listed below by the research run, not written for reading. The judgment lives in the signals and issues that cite it, below.
2026-08-10-codex-approve-for-me-puts-a-model-not-a-human-in-the-approval
--approve-for-me puts a model, not a human, in the approval seat.
A new --approve-for-me flag on the interactive and exec commands routes approval requests through automatic review instead of prompting the operator. It sets approval_policy="on-request" with the workspace-write sandbox, and propagates across root, exec, resume, and fork argument handling while preserving later subcommand permission overrides.
Channel: tagged-release. Ancestry: grep of the rust-v0.146.0...rust-v0.147.0 compare commit list finds exactly one match: b7a61066081644e0d8b2c0b4dbfd7408ac1514df 'Add an --approve-for-me CLI flag (#36373)'. PR #36373 merged 2026-07-31T18:28:24Z; first non-prerelease tag containing it is rust-v0.147.0.
Operator consequence: Watch, and decide deliberately before enabling it in CI. This is the flag that converts Codex's approval prompt from a human checkpoint into a model checkpoint on a single command-line switch, and it also silently pins the sandbox to workspace-write. If your governance story rests on 'a person approved every escalation', add --approve-for-me to whatever you lint agent invocations with. Note the ordering trap: subcommand permission overrides still win, so the effective policy is not always what the flag implies.
Receipt
Finding metadata
Run: 2026-08-10-weekly-digest-2026-08-03_2026-08-10-frontier-v0
Finding ID: 2026-08-10-codex-approve-for-me-puts-a-model-not-a-human-in-the-approval
Accepted signals
Source links
Primary links, including exact changelog lines when available.
Versioned source: run artifact