OpenAI Codex release notes: CLI 0.139.0 adds web search in Code mode

OpenAI Codex CLI 0.139.0 is not a giant launch. It is more useful than that: Code mode gets direct web search, MCP tools get better schema support, and a few annoying CLI edge cases become harder to trip over.

Codex CLI is OpenAI's local coding agent: a terminal tool that can read a repo, reason about changes, and work inside a sandbox. A sandbox limits what the agent can do. An approval gate is the checkpoint where a person or policy approves higher-risk steps before they run. MCP, Model Context Protocol, is a way to connect the agent to external tools and data sources with clear schemas.

OpenAI Codex release notes: what changed in CLI 0.139.0?

The most visible change is that Code mode can now use standalone web search directly, including from nested JavaScript tool calls, and receive plaintext search results. That makes Codex more useful when the agent needs to check current documentation before it suggests a code change.

The same release makes tools and connectors less brittle: input schemas now preserve oneOf and allOf, and large schemas keep more shallow structure when compacted. For teams building MCP connections, that means fewer weird gaps between what a tool can do and what the agent thinks the tool can do.

There are also a few small workflow fixes worth noticing. codex doctor now includes editor and pager environment details in the local report, while redacting raw values in JSON output. codex plugin marketplace list --json shows which marketplace source each item came from. codex resume --last "..." and codex fork --last "..." now treat the text after --last as the initial prompt, not as a session ID.

Source: OpenAI Codex GitHub release rust-v0.139.0

Source: npm registry: @openai/codex latest

Why this matters for real teams

For Nordic organizations moving from chat to agent workflows, the web search change in Code mode is the interesting part. It can reduce the need to paste docs into chat by hand. But it also makes source hygiene more important: ask Codex to show the URLs it used, separate facts from assumptions, and stop before it changes auth, payments, environment variables, or permissions.

The schema changes matter almost as much for Tool Forge work. When AI agents connect to CRM, ticketing, internal APIs, or document workflows, the tool contracts need to be clear. oneOf and allOf sound technical, but in practice they help the agent choose the right version of a tool instead of guessing.

The safe integration line is still simple: use environment variables or secret managers, give the agent scoped API keys, limit permissions, redact sensitive values in output, and keep a run log a human can actually read.

What to test today

Human step: use a Codex CLI environment where you already run version 0.139.0. The sources verify the feature, but they do not describe a new slash command for web search. Treat this as a source-grounded Code mode task, not as a made-up button or menu instruction.

Short example: use the new Codex feature

Check our payment connector with source-grounded work in Codex CLI 0.139.0 Code mode. First verify the provider's current public documentation. Then read only the connector files, list the exact source URLs, and propose the smallest safe patch. Do not change secrets, auth scopes, or environment variables without asking for approval first.

Good output should include:

• Exact source URLs, not just the provider homepage.
• A clear line between verified facts and assumptions.
• The files that need review, plus what should stay untouched.
• A short list of approval gates before anything sensitive changes.

What we are watching next

This release is not a reason to rebuild everything. It is a good signal for the next step: Codex is getting better at fetching current context, understanding richer tools, and leaving clearer diagnostic traces. For Hammer readers, that is where agent work starts to become useful day to day: small, verifiable improvements with human control still in the loop.