Codex CLI 0.125.0: Persistent Permissions & Token Reporting

Persistent Permission Profiles Across the Full Session Stack

One of the most frequently reported friction points in Codex CLI has been the need to re-approve tools and re-configure permission policies at the start of each new session. With version 0.125.0, OpenAI has resolved this: permission profiles now persist across the entire session lifecycle — including TUI sessions, individual user turns, MCP sandbox state, shell escalation events, and app-server APIs.

In practice, this means a developer who has configured a permission profile — for example, allowing a specific set of shell commands or MCP tools to execute without prompting — no longer needs to re-establish that policy after restarting the CLI, switching between environments, or connecting through the app-server. The profile follows the session context faithfully, which is especially valuable in long-running multi-agent workflows where approval interruptions create costly bottlenecks.

How It Works

Permission profiles now round-trip through all layers of the Codex execution stack. When a profile is set in a TUI session, it is serialised and carried forward through MCP sandbox creation, shell escalation events, and app-server RPC calls. This unification eliminates the previous inconsistency where a profile applied in one context would silently reset in another.

Reasoning-Token Visibility in codex exec --json

The codex exec --json command, which allows developers to run Codex tasks programmatically and consume structured output, now includes reasoning-token counts in its reported usage data. Previously, only standard input and output token counts were surfaced.

For teams running Codex in CI pipelines, automated workflows, or cost-monitoring dashboards, this addition provides the missing piece of the token accounting puzzle. Reasoning tokens — consumed during chain-of-thought processing — can represent a meaningful share of total spend when using high-reasoning models, and their absence from programmatic output made accurate cost attribution difficult.

Example Use Cases

• Cost attribution in CI pipelines: Teams running codex exec as part of automated code-review or test-generation workflows can now accurately track per-run token costs including reasoning overhead.
• Optimisation decisions: By observing reasoning-token counts across different task types, developers can calibrate reasoning level settings (introduced via the TUI Alt+comma / Alt+period controls in 0.124.0) for cost-efficiency.

Reliability Improvements

Beyond the two headline developer features, 0.125.0 addresses several reliability gaps in the app-server layer:

• The exec-server no longer drops buffered output after a process exits, ensuring that final output lines are always captured in programmatic invocations.
• WebSocket app-server clients are now less likely to disconnect during burst notification periods.
• Windows sandbox startup has been improved to handle multiple concurrent CLI versions and installed app directories more gracefully, with background PowerShell processes no longer spawning visible windows.
• The /review interrupt flow was fixed — exiting the TUI during a /review session no longer leaves the interface wedged at delegate startup.