Codex CLI v0.137.0 Stable: Cloud Config Bundles, Multi-Agent Runtime Persistence, and Plugin JSON Output

Three days after v0.136.0 shipped the session archiving and --stdio app-server transport, v0.137.0 landed as a stable release on 4 June 2026.¹ The alpha previews (up to 0.137.0-alpha.4 on 3 June) introduced the skills extension crate and remote-control client management.² The stable cut adds enterprise cloud-managed configuration bundles, multi-agent v2 runtime persistence, machine-readable plugin output, and a clutch of TUI and platform fixes. This article covers everything that shipped.

What graduated from alpha

The features covered in the alpha preview article moved to stable with minimal API changes:

remoteControl/client/list and remoteControl/client/revoke — JSON-RPC methods for account-scoped device grant management remain unchanged.²
codex-skills-extension and codex-context-fragments crates — shared prompts, context fragments, and skills plumbing now live in dedicated crates, reducing codex-core coupling.³
Per-turn skill catalogue resolution — turn-input contributors resolve environment-aware skills per turn.²
Permission profile modernisation — sandbox defaults expressed as observable named profiles.²
Image generation in code-mode — exposure changed from DirectModelOnly to Direct, enabling nested code-mode access to gpt-image-2.²

If you tested the alpha, npm install -g @openai/codex@latest now points at 0.137.0 stable.

1 — Cloud-managed configuration bundles

The headline enterprise addition: admins on ChatGPT Business, Enterprise, or EDU plans can now push cloud-managed config bundles that the CLI fetches at sign-in.¹⁴

How it works

Codex assembles effective configuration through a precedence stack. Cloud-managed settings sit at the top:

flowchart TD
    A["Cloud-managed requirements<br/>(Business / Enterprise / EDU)"] --> B["macOS MDM preferences<br/>(com.openai.codex)"]
    B --> C["System requirements.toml<br/>(/etc/codex/requirements.toml)"]
    C --> D["managed_config.toml<br/>(defaults, user can override in-session)"]
    D --> E["User config.toml<br/>(~/.codex/config.toml)"]
    E --> F["CLI flags<br/>(--model, --approval-policy)"]
    style A fill:#e8f5e9,stroke:#2e7d32
    style B fill:#e3f2fd,stroke:#1565c0
    style C fill:#e3f2fd,stroke:#1565c0

Two mechanisms exist at the cloud layer:

Mechanism	Effect	User can override?
Requirements	Hard constraints (allowed sandbox modes, approval policies, network rules)	No
Managed defaults	Starting values applied at launch	Yes, during session

Example: locking down an EDU workspace

An administrator might push a cloud bundle equivalent to:

# requirements (enforced)
allowed_approval_policies = ["untrusted", "on-request"]
allowed_sandbox_modes = ["read-only", "workspace-write"]

[sandbox_workspace_write]
network_access = false

# managed defaults (soft)
approval_policy = "on-request"
sandbox_mode = "workspace-write"

Students can still switch between on-request and untrusted approval policies, but cannot escalate to full-auto. Network access stays off regardless of local config.toml settings.⁴

Monthly credit limits

Enterprise and EDU admin flows now surface monthly credit limits in the CLI, displaying remaining budget and workspace-specific usage-limit messages for credit and spend-cap failures.¹⁵

2 — Multi-Agent v2: runtime persistence across threads

Prior to v0.137.0, spawning a subagent created a fresh runtime context. If you changed approval policy to full-auto mid-session or toggled /permissions, child agents reverted to defaults.⁶

v0.137.0 fixes this: Multi-Agent v2 keeps the runtime choice with each thread and propagates it to spawned agents.¹

Concretely:

The parent turn’s live runtime overrides — sandbox mode, approval policy, /permissions changes, --yolo flag — are reapplied when Codex spawns a child.⁶
Follow-up task handling is cleaner: metadata defaults for spawned agents inherit the parent context rather than falling back to profile defaults.
Thread labels surface in approval overlays, so when many agents run concurrently you can tell which thread is requesting permission.

sequenceDiagram
    participant User
    participant Parent as Parent Thread
    participant Child as Spawned Agent

    User->>Parent: /permissions full-auto
    Parent->>Parent: runtime override stored
    Parent->>Child: spawn_agent("refactor auth module")
    Note over Child: Inherits full-auto + sandbox overrides
    Child->>Parent: result (no approval prompts)

This matters most in CSV batch workflows (spawn_agents_on_csv) where dozens of agents run in parallel — previously each one prompted for approval independently.⁶

3 — Machine-readable plugin output

codex plugin list now accepts a --json flag, emitting structured JSON instead of the human-readable table:¹

codex plugin list --json

[
  {
    "name": "codex-security",
    "version": "1.2.0",
    "source": "remote",
    "marketplace": "openai/codex-plugins",
    "skills": ["security-scan", "deep-security-scan", "security-diff-scan", "fix-finding"],
    "mcp_servers": 0
  },
  {
    "name": "my-team-standards",
    "version": "0.3.1",
    "source": "local",
    "marketplace": null,
    "skills": ["lint-config", "review-checklist"],
    "mcp_servers": 1
  }
]

The release also adds cached remote catalogue suggestions: when you run codex plugin search, results from the remote catalogue are cached locally, reducing network round-trips for repeated queries.¹⁷

Scripting example

Audit installed plugins in CI:

codex plugin list --json | jq -r '.[] | select(.source == "remote") | "\(.name)@\(.version)"'

Or verify a required plugin is present before a pipeline step:

if ! codex plugin list --json | jq -e '.[] | select(.name == "codex-security")' > /dev/null 2>&1; then
  echo "Required plugin codex-security not installed" >&2
  exit 1
fi

4 — Expanded code-mode tooling

Hosted web search and image generation tools are now available in more code-mode flows, with standalone web searches able to execute in parallel.¹

Previously, web_search inside a code-mode turn was serialised — each search blocked the next. v0.137.0 allows the model to dispatch multiple web searches concurrently when they are independent, reducing latency for research-heavy coding tasks.

Image generation via gpt-image-2 (graduated from alpha) is now accessible inside code-mode sessions without requiring a separate turn, enabling workflows like:

Generate a placeholder image for a UI component
Write the React component referencing the generated asset
Run the dev server to verify

All within a single code-mode turn.

5 — TUI ergonomic improvements

F13–F24 keybindings

Power users with extended keyboards or remapped layouts can now bind TUI actions to F13–F24 keys.¹ Configure them in config.toml:

[keybindings]
F13 = "toggle_reasoning"
F14 = "toggle_context_stats"

Paste in searchable menus

The session picker, plugin search, and slash-command menus now accept pasted text for filtering — previously, paste was silently dropped in these contexts.¹

Compact reasoning-only status item

A new model+reasoning status-line item displays model name and current reasoning effort in a single compact widget, replacing the need for two separate items:¹⁸

[status_line]
items = ["model+reasoning", "context_stats", "git_branch"]

6 — Bug fixes worth noting

Several fixes in this release address real workflow pain points:

Fix	Impact
Cancelled prompts restore drafts, attachments, and collaboration mode¹	No more lost work when you hit Escape after typing a long prompt
Slash-command filtering resets properly¹	`/` menu no longer shows stale results from a previous filter
Plugin manifest order preserved¹	Skills appear in the order the plugin author intended
Duplicate plugin installations deduplicated¹	`codex plugin install` no longer creates duplicates when run twice
Malformed skills fields treated as warnings¹	A single broken skill no longer prevents the entire plugin from loading
Permission requests carry environment identity¹	Approval prompts show which environment (local/remote) is requesting access
MITM proxy exports readable CA bundles¹	Corporate proxy users get PEM-formatted certificates instead of binary blobs

Platform-specific fixes

macOS: App launch reliability improved.¹
Windows: SQLite startup failures resolved; Bazel CI made Windows-compatible.¹
Thread resume: Pathless side-chat reloads and compressed rollouts handle edge cases more safely.¹

Upgrade path

npm install -g @openai/codex@latest
codex --version   # 0.137.0

If your organisation uses managed configuration, the cloud config bundle will apply automatically at next sign-in — no local file changes required.⁴

For teams pinning versions in CI:

npm install -g @openai/codex@0.137.0

What to watch next

The codex-skills-extension and codex-context-fragments crate extraction signals that plugin and skill loading are being decomposed for independent versioning.³ Combined with the plugin list --json output, this points towards a more scriptable, composable plugin ecosystem — potentially with lock-file-style dependency resolution in a future release.

The cloud-managed config bundle mechanism is currently Business/Enterprise/EDU only, but the underlying precedence architecture supports MDM profiles on macOS and system-level requirements.toml on Linux.⁴ Teams not on enterprise plans can achieve similar controls using the system-level files today.

Citations

OpenAI, “Release 0.137.0”, GitHub, 4 June 2026. https://github.com/openai/codex/releases/tag/rust-v0.137.0 ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷ ↩⁸ ↩⁹ ↩¹⁰ ↩¹¹ ↩¹² ↩¹³ ↩¹⁴ ↩¹⁵ ↩¹⁶ ↩¹⁷ ↩¹⁸ ↩¹⁹ ↩²⁰
OpenAI, “Release 0.137.0-alpha.4”, GitHub, 3 June 2026. https://github.com/openai/codex/releases/tag/rust-v0.137.0-alpha.4 ↩ ↩² ↩³ ↩⁴ ↩⁵
OpenAI, “Codex CLI Rust Workspace Structure”, Mintlify docs, 2026. https://mintlify.com/openai/codex/architecture/rust-crates ↩ ↩²
OpenAI, “Managed configuration – Codex”, OpenAI Developers, 2026. https://developers.openai.com/codex/enterprise/managed-configuration ↩ ↩² ↩³ ↩⁴
OpenAI, “Codex rate card”, OpenAI Help Center, 2026. https://help.openai.com/en/articles/20001106-codex-rate-card ↩
OpenAI, “Subagents – Codex”, OpenAI Developers, 2026. https://developers.openai.com/codex/subagents ↩ ↩² ↩³
Hashgraph Online, “awesome-codex-plugins”, GitHub, 2026. https://github.com/hashgraph-online/awesome-codex-plugins ↩
OpenAI, “Changelog – Codex”, OpenAI Developers, June 2026. https://developers.openai.com/codex/changelog ↩