← Back to the briefs
Coding

Codex 0.170.0 fixes child thread sandbox restrictions and isolates MCP environments

OpenAI released Codex 0.170.0, resolving issues where child threads lost sandbox restrictions and adding experimental support for identically named MCP servers across environments alongside TUI realtime voice negotiation.

Sandbox restrictions and permissions enforcement

OpenAI Codex 0.170.0 addresses critical boundary issues within execution environments. The update fixes permission updates and prevents child threads from shedding existing sandbox restrictions during runs. Alongside this change, project documentation now details sandbox permission profiles and filesystem access rules.

Platform-specific fixes landed for Windows environments as well. The release repairs file replacement procedures, ensures sandbox permissions persist across operations, and resolves runtime registration recovery after encountering service privilege errors.

Multi-environment MCP support and protocol stability

The release introduces experimental support for running identically named Model Context Protocol servers across separate environments. In parallel, OpenAI resolved race conditions that compromised MCP OAuth credentials during cancellation, authentication, and logout operations.

Transport reliability received attention across cloud sessions and network connections. Large WebSocket requests that exceed normal limits now fall back directly to HTTP without excess retries. Reconnection logic for Cloud conversations now preserves active recovery turns when clients drop and reconnect.

Interface features and telemetry updates

Within the terminal user interface, Codex adds WebRTC negotiation to support realtime voice calls directly from the console. For telemetry pipelines, the release introduces automatically refreshed authentication headers for OpenTelemetry HTTP exporters. Interface search logic also eliminates duplicate file listings and repairs thread search snippets when search queries contain whitespace.

Our assessment

Teams delegating autonomous tasks across nested subagents should verify that permission boundaries hold when spawning subprocesses. The bug where child threads lost sandbox restrictions illustrates the hazard of relying on implicit permission inheritance during nested runs. Operators running complex development setups across staging and production will benefit from testing the experimental multi-environment MCP server support, but credential flows should be monitored closely given recent OAuth concurrency fixes.

Source: OpenAI original publication.

Original source

Read the original material

OpenAI (opens in a new tab)

The article date belongs to Agentic Horizon. The source publication date is listed separately. Vendor claims remain attributed; we have not independently tested the reported capability.

Search the publication

Find a story by title, topic, or keyword.

Press Escape to close

Ask HorizonYour site guide
Your guide to the publication

Ask Horizon.

Ask a question about the publication.

Privacy