AI coding tools
Codex CLI 0.152 Adds Per-Tool MCP Output Quotas and Fixes Lost Instructions in Long-Running Sessions
OpenAI has added output token limits for individual MCP tools and now allows app-server shell commands to run for more than an hour. The release also fixes an issue where user instructions and authorizations could be lost after conversation history was compacted, while the built-in planning tool is now disabled by default.
OpenAI released Codex CLI 0.152.0 on September 1. The main changes concern the boundaries around context, tools, and credentials in the agent runtime rather than the model itself. Previously, `tool_output_token_limit` applied a single limit across the entire session. The new release allows an `output_token_limit` to be configured for each MCP tool and preserves the same truncation rules when a session is resumed. This lets tools that require large result sets, such as code search, retain higher quotas while limiting list queries or chat integrations, reducing the risk that a single response will fill the context window.
The app-server’s `thread/shellCommand` now also supports client-configured deadlines longer than one hour, making it more suitable for full CI runs, cross-platform builds, or large data migrations. However, extending the timeout only addresses execution duration. Integrations must still provide cancellation, progress streaming, and orphaned-process cleanup themselves; a long deadline should not be treated as a job scheduler.
A more important correctness fix affects automatic approval reviews: after long conversations are compacted, user instructions, responses, and still-valid authorizations are now preserved. Previous versions could forget restrictions such as “do not modify this directory” during the second segment of a session—a rare but high-risk form of state loss. Cloud tasks have also begun rejecting untrusted backend URLs and disabling redirects to prevent stored credentials from being sent to other endpoints.
The upgrade includes another behavioral change: `update_plan` is no longer exposed to the model by default. Workflows that rely on the built-in plan view or use tool events to track progress must set `tools.update_plan.enabled = true`. Engineering teams should retest the completeness of results after MCP truncation, policy consistency before and after compaction, and how custom app-servers handle timeouts and URL validation, rather than merely confirming that the CLI starts successfully.