Codex CLI 0.50.0 Release: Strengthening Risk Assessment and MCP Reliability¶

Before you upgrade

codex --feedback now bundles richer diagnostics and attachments. Re-run the command on 0.50.0 whenever you submit tickets.
Enabling experimental_sandbox_command_assessment shows a model-generated risk level and category for commands that violate sandbox policy. Plan a rollout that matches your internal approval rules.
If you installed Codex CLI via Homebrew, back up your existing codex.json and MCP config before running brew upgrade codex.

Key takeaways in 3 minutes

Released on 2025-10-25 UTC (2025-10-26 JST). The theme is safer, more diagnosable automation.
/feedback captures richer logs and files, dramatically cutting the back-and-forth with support.
Sandbox-violating commands can surface risk ratings (experimental flag).
MCP clients jump to rmcp 0.8.3, handling HTTP 204 responses during initialization.
/status now warns when rate limit data is missing or stale, guiding users to the right dashboard.

Safety and diagnostics improvements in 0.50.0¶

PR #5536 adds a model-driven risk level (Low/Medium/High) and risk category whenever a command violates sandbox policy and requires approval.
The assessment is generated in a fresh conversation thread, so it is not biased by the ongoing session. Use it to tighten approval rules and reduce mis-clicks.
The feature is gated behind experimental_sandbox_command_assessment. Pilot it with a few high-risk workflows and document how the risk labels map to your compliance thresholds.

PRs #5661 and #5663 overhaul codex --feedback, bundling reproduction logs, annotated messages, and optional file uploads.
Update your internal runbooks: the UI differs from 0.49 and earlier, so screenshot-based manuals need a refresh.

/status shows explicit warnings when rate limits are missing or older than 15 minutes (#5570), preventing confusion during cold starts.
GitHub PAT validation errors and MCP startup timeouts now ship clearer instructions (#5595), shrinking the number of “Codex looks broken” tickets.
/mcp and mcp get redact environment variable values by default (#5648), keeping secrets out of shared logs.

PR #5542 upgrades to rmcp 0.8.3 so MCP servers that reply with HTTP 204 initialize cleanly.
If you rely on experimental_use_rmcp_client, a follow-up change (#5653) enforces an additional login after mcp add, blocking partially configured connections.

Introducing the CodexHttpClient wrapper (#5564) standardizes logging for HTTP method, URL, and status across Codex components—handy for SOC reviews and CI tracing.
A new codex/event/raw_item telemetry stream (#5578) improves automation observability. Capture a few samples for your analytics pipeline.
macOS builds now render the Option (⌥) symbol in key hints (#5582), keeping training materials consistent with the UI.

When using Homebrew, run the following. A formula-to-cask switch requires brew update first.

brew update
brew upgrade codex

Back up ~/.config/codex/, send a trial codex --feedback, and ensure attachments reach your support inbox.
Trigger at least two sandbox-approval flows and verify that the risk banner matches your incident response policy.
Reconnect MCP servers (especially GitHub). If your server returns HTTP 204 during initialization, confirm the handshake now succeeds.

Reference PRs

Coming from 0.47 or earlier? Pair this release with the Codex CLI 0.47 update guide to map infrastructure changes.
For sandbox governance, reuse the workflow templates in the Codex CLI 0.39–0.44 update guide and swap in the new risk banner screenshots.
Schedule a feedback log review one week after rollout to ensure teams adopt the new /feedback flow and that support tickets show the richer context.