codex exec --strict-config stopped on line 2. The error named not_a_real_key. No websocket was opened. The same file, with the same flag, handed to codex doctor --summary, did not name that field. The configuration row said config loaded, marked as a warning. Both processes exited 1. The exits were not the same failure.
That split is what Codex CLI 0.163.0 did on 19 November 2026. codex --version printed codex-cli 0.163.0. Doctor also noted that 0.170.0 was available. Nothing here was repeated on 0.170.0. The host was Linux x86_64. Each measurement used a new CODEX_HOME. The file under test was two lines:
model = "gpt-6.1-sol"
not_a_real_key = true
not_a_real_key is not a setting this binary documents. It was chosen because it cannot be mistaken for a renamed feature. No codex login was run. The disposable home had no auth.json. A second home, used only as a control, held the model line and nothing else.
The developer-commands page, fetched the same morning, says --strict-config errors when config.toml contains fields this Codex version does not recognize. It says the flag is supported by runtime commands such as codex, exec, review, resume, fork, app-server, and exec-server. The page does not say doctor is one of those commands. It does not say features, mcp, plugin, or sandbox reject the flag before they read the file. The help text printed by codex exec --help is the shorter sentence: error out when config.toml contains fields that are not recognized. codex doctor --help does not print that sentence. The flag is still accepted when it is placed before doctor.
The commands that stop
The flag belongs on the command that is about to run, not on a diagnostic that happens to share the word config.
From the disposable home:
codex exec --strict-config --skip-git-repo-check say hi
Exit 1. The text was:
Error loading config.toml:
<codex-home>/config.toml:2:1: unknown configuration field `not_a_real_key`
The caret sat under not_a_real_key. There was no session id, no reconnect counter, and no 401. The loader stopped before the transport.
codex review --strict-config --uncommitted, run in a disposable Git repository that held one committed file, printed the same loader error and exited 1. It did not review the tree. codex resume --last --strict-config and codex fork --last --strict-config printed the same loader error and exited 1. Neither command looked for a session. codex exec resume --last --strict-config did the same. An omitted session id is not reached when the file fails to load.
codex --strict-config app-server --listen stdio:// also exited 1, and it did not stay running. Its first line was Error: followed by the path, the line, and unknown configuration field. It did not use the Error loading config.toml: heading that exec, review, resume, and fork used. Both shapes name the field and the line. A wrapper that matches only one heading will miss the other.
The interactive codex command, with no subcommand, is on the documentation list. This run did not start a TUI. exec-server is on that list as well. This run did not start it. Treat those two as documented, not as measured here.
A rejected flag is not a field error
Several subcommands accept the flag at the top-level parser and then refuse it.
codex --strict-config features list
codex --strict-config mcp list
codex --strict-config plugin list
codex --strict-config sandbox --permission-profile :read-only --cd <workdir> -- /bin/true
Each exited 1. The features, mcp, and plugin lines were Error: --strict-config is not supported for codex features, and the same sentence with mcp or plugin in place of features. The sandbox line named codex sandbox. None of those four lines named not_a_real_key.
The mcp refusal is not caused by the unknown field. The control home, whose file contained only model = "gpt-6.1-sol", produced the same not supported error under codex --strict-config mcp list. Exit 1 from these four commands means the flag is not implemented there. It does not mean the file has a typo, and it does not mean the file is clean.
Without the flag, those commands do not mention the unknown key either. codex features list printed the inventory and exited 0. codex mcp list printed No MCP servers configured yet and exited 0. codex sandbox with :read-only, a --cd, and /bin/true exited 0 and wrote no stdout. A green features, mcp, or sandbox run is not a strict-config pass. The 11 November note on this site already recorded the features refusal. This run adds mcp, plugin, and sandbox, and it adds the control file that has no unknown key.
codex login --help and codex debug models --help exited 0 when --strict-config was placed before the subcommand, because help does not load config.toml. codex doctor --help, run without that flag, does not list it. Do not count a help exit as evidence that the subcommand honors the flag or rejects it.
Doctor records an ignore
codex --strict-config doctor loaded the bad file. It did not print not supported. It did not print unknown configuration field. The configuration check was a warning, not a failure. The failure in the report was credentials, the same failure the control home produced.
--summary is the view a job is most likely to store. On the bad home the notes included:
⚠ config config loaded
The configuration section repeated that warning row. The control home’s configuration section was:
✓ config loaded
The control notes did not mention config. The bad summary never printed not_a_real_key. A parser that looks for the field name in the summary will not find it. A parser that looks only for a failure mark will see the auth failure and miss the warning mark on the config row. The words config loaded are not a statement that every key was recognized.
The full report, without --summary, does name the field. Under the config details it printed startup warnings 1 and startup warning deprecated 1, then:
Codex is ignoring 1 unrecognized configuration setting. Check for typos or deprecated settings.
user (<codex-home>/config.toml): `not_a_real_key` is ignored.
The word in that sentence is ignored. It is not rejected. The counter is named deprecated. The key was never a setting in this binary’s help or in the configuration reference read the same morning. The counter still incremented. Do not treat that counter as a list of retired keys. It counted an invented one.
--json keeps the same facts in the config.load object. On the bad home, status was warning and summary was config loaded. details held the ignore sentence, startup warnings was 1, and startup warning deprecated was 1. On the control home, status was ok and summary was config loaded. The summary string is shared. The status is not. A job that stores the summary string and ignores status cannot tell these files apart. A job that treats any non-ok status as a strict-config failure will also be wrong: the status is warning, and the process still loaded the file.
Passing --strict-config did not change that doctor result. codex doctor --json without the flag, on the bad home, carried the same ignore sentence. The flag is a no-op on this subcommand. The full report is still useful. It is not the gate the help sentence describes.
Doctor’s own exit code is not the gate either. Both homes exited 1 because no credentials were present. Adding the unknown key did not change the exit. Removing it did not change the exit. If a later home has credentials, a warning-level config row can sit beside an exit 0. This run did not create that home. Do not infer a zero exit from a signed-in doctor. Read the config row.
A warning is not a stop
Without the flag, exec does not stop.
codex exec --skip-git-repo-check say hi
The process printed that it was ignoring one unrecognized setting, named not_a_real_key, and then opened the Responses transport. The disposable home had no credentials. The transport returned 401. That 401 is the missing login. It is not evidence about the unknown field. The field had already been ignored. The same file with --strict-config never reached the transport.
That is why the flag has to sit on the command you intend to run. A doctor warning, a features inventory, or an mcp list can all complete while exec is still willing to continue past the typo. The flag on exec, review, resume, fork, and app-server is the check that turns the ignore into a load error.
Verify the command you are about to run
Point CODEX_HOME at a copy of the home you mean to use, or at the home itself if you can afford the command to refuse to start. Do not point a test at a home whose sessions you need to keep if the command is resume or fork and the file is valid: those commands proceed once the file loads. With an unknown field they stop first. Confirm that stop on a copy if you are unsure the file is clean.
export CODEX_HOME=<codex-home>
codex exec --strict-config --skip-git-repo-check true
On 0.163.0 an unknown top-level key exits 1 and names the field before a session id is printed. A recognized file continues into authentication. This home had none, so a clean file would have failed later, at the transport, not at the loader. Read the first error. unknown configuration field is the strict-config result. 401 and no Codex credentials are a different result.
Use the same flag on review, resume, fork, and app-server when those are the commands the job will run. The app-server line starts with Error: rather than Error loading config.toml:. Match the field name, not only the heading.
Do not use features, mcp, plugin, or sandbox as that check. On this binary they exit 1 because the flag is unsupported, including when the file has no unknown key. Do not use doctor --summary as that check. It can warn with the words config loaded and never print the key. If you want the key named without stopping a later exec, read the full doctor report or the JSON config.load details, and treat ignored as a warning you still have to enforce yourself.
The 0.170.0 line in doctor’s update note is a version advertisement. It is not a measurement. Re-run the exec refusal on the binary you will actually launch before you treat this split as fixed.
Sources
- Codex CLI 0.163.0, version line
codex-cli 0.163.0, Linux x86_64, 19 November 2026.exec,review,resume,fork,exec resume,app-server,doctor,features,mcp,plugin, andsandboxran withCODEX_HOMEset to disposable directories. No login was run. codex exec --helpon that binary. The--strict-confighelp text says the flag errors out whenconfig.tomlcontains fields that are not recognized.codex doctor --helpdoes not list the flag.- Codex CLI developer commands, fetched 19 November 2026. The global-flag table says
--strict-configis supported by runtime commands such ascodex,exec,review,resume,fork,app-server, andexec-server.