What version of Codex CLI is running?
codex-cli 0.147.0
What subscription do you have?
Pro Lite
Which model were you using?
gpt-5.6-sol, reasoning high
What platform is your computer?
macOS 15.7.7, Apple Silicon, aarch64
What terminal emulator and version are you using (if applicable)?
Visual Studio Code integrated terminal, VS Code 1.133.0
Codex doctor report
Codex Doctor v0.147.0 · macos-aarch64
17 ok · 1 idle · 0 warn · 0 fail
Runtime: standalone Unix installation
Configuration: loaded successfully
Authentication: configured
Sandbox: restricted filesystem and restricted network
Approval policy: OnRequest
State databases: healthy
Thread inventory: healthy
WebSocket: connected
What issue are you seeing?
Registered custom agents cannot independently expose their runtime metadata.
The project has registered custom agents including:
- luna_reviewer
- terra_critical_reviewer
Their project configuration specifies:
- luna_reviewer: model gpt-5.6-luna, sandbox read-only
- terra_critical_reviewer: model gpt-5.6-terra, sandbox read-only
The parent Codex session is explicitly started with:
--sandbox read-only
--ask-for-approval on-request
The parent /status output confirms:
Permissions: Read Only (Ask for approval)
However, when a registered custom agent is started, the child reports that its identity, model, and reasoning are unavailable or cannot be independently observed.
Observed child runtime results include:
identity: unavailable
model: unavailable
reasoning: unavailable
raw sandbox: seatbelt or unavailable
inherited effective policy: read-only
Because the custom-agent identity and model cannot be independently verified, the project cannot treat the child as a valid registered Luna or Terra reviewer for its security-review gate.
This has occurred across fresh CLI sessions, after restarting Codex, after updating Codex, after running codex doctor --all, and after restarting the computer.
The child stops before reading project files. No project files, database, tests, or commits are involved in the runtime probe.
What steps can reproduce the bug?
Uploaded thread: 01a00370-fc32-7992-8018-9cd14ff82b90
What is the expected behavior?
Before reading project files, a registered custom agent should expose verifiable runtime metadata:
identity: terra_critical_reviewer
model: gpt-5.6-terra
reasoning: max
inherited effective policy: read-only
For luna_reviewer, the expected values are:
identity: luna_reviewer
model: gpt-5.6-luna
reasoning: max
inherited effective policy: read-only
The parent session should remain read-only, and the child should be identifiable as the registered agent that was selected.
Additional information
Codex CLI version: 0.147.0
Thread ID: 01a00370-fc32-7992-8018-9cd14ff82b90
The project configuration is present and parses successfully. The parent sandbox is correctly read-only. The issue is specifically that the child custom-agent runtime metadata is unavailable, not that the parent sandbox is writable.
Please clarify:
- Whether custom-agent identity, model, and reasoning metadata are expected to be visible in Codex CLI;
- Whether a feature flag, account entitlement, or configuration is required;
- Whether this is a known issue in Codex CLI 0.147.0;
- Whether a newer runtime fix is available;
- What exact diagnostic is required for a registered custom agent to prove its identity independently.
What version of Codex CLI is running?
codex-cli 0.147.0
What subscription do you have?
Pro Lite
Which model were you using?
gpt-5.6-sol, reasoning high
What platform is your computer?
macOS 15.7.7, Apple Silicon, aarch64
What terminal emulator and version are you using (if applicable)?
Visual Studio Code integrated terminal, VS Code 1.133.0
Codex doctor report
What issue are you seeing?
Registered custom agents cannot independently expose their runtime metadata.
The project has registered custom agents including:
Their project configuration specifies:
The parent Codex session is explicitly started with:
--sandbox read-only
--ask-for-approval on-request
The parent /status output confirms:
Permissions: Read Only (Ask for approval)
However, when a registered custom agent is started, the child reports that its identity, model, and reasoning are unavailable or cannot be independently observed.
Observed child runtime results include:
identity: unavailable
model: unavailable
reasoning: unavailable
raw sandbox: seatbelt or unavailable
inherited effective policy: read-only
Because the custom-agent identity and model cannot be independently verified, the project cannot treat the child as a valid registered Luna or Terra reviewer for its security-review gate.
This has occurred across fresh CLI sessions, after restarting Codex, after updating Codex, after running codex doctor --all, and after restarting the computer.
The child stops before reading project files. No project files, database, tests, or commits are involved in the runtime probe.
What steps can reproduce the bug?
Uploaded thread: 01a00370-fc32-7992-8018-9cd14ff82b90
What is the expected behavior?
Before reading project files, a registered custom agent should expose verifiable runtime metadata:
identity: terra_critical_reviewer
model: gpt-5.6-terra
reasoning: max
inherited effective policy: read-only
For luna_reviewer, the expected values are:
identity: luna_reviewer
model: gpt-5.6-luna
reasoning: max
inherited effective policy: read-only
The parent session should remain read-only, and the child should be identifiable as the registered agent that was selected.
Additional information
Codex CLI version: 0.147.0
Thread ID: 01a00370-fc32-7992-8018-9cd14ff82b90
The project configuration is present and parses successfully. The parent sandbox is correctly read-only. The issue is specifically that the child custom-agent runtime metadata is unavailable, not that the parent sandbox is writable.
Please clarify: