openai/codex
Lightweight coding agent that runs in your terminal
How codex is put together
openai/codex is the Codex CLI coding-agent monorepo, dominated by a large Rust workspace (codex-rs) of ~150 crates implementing the agent core, TUI, exec engine, sandboxing, MCP/app-server protocol layers, and extension system. A thin Node wrapper (codex-cli/bin/codex.js, published as @openai/codex) launches the native binary, and a TypeScript SDK lives under sdk/typescript. Communication flows through JSON-RPC-style protocols between clients (TUI/exec), an embedded or remote app-server, the core agent loop, and tool/sandbox executors. Build orchestration uses both Cargo and Bazel, with pnpm workspaces for JS tooling and devcontainers (including a hardened 'secure' profile with bubblewrap sandboxing and firewall allowlisting) for runtime isolation.
Languages
Frameworks
Infrastructure
Major components
codex-rs/core
The central agent loop: model prompting (embedded gpt-5.x prompt markdown), session orchestration, and coordination of tools and extensions.
codex-rs/app-server family (app-server, -protocol, -transport, -client, -daemon)
JSON-RPC application server exposing Codex sessions to CLI surfaces and external clients over sockets/stdio/UDS, including an in-process client facade.
codex-rs/tui
Terminal user interface front-end consuming the app-server client facade.
codex-rs/exec / exec-server / shell-command / apply-patch
Command execution engine, persistent exec server, shell command construction, and patch-application logic.
codex-rs sandboxing stack (linux-sandbox, bwrap, execpolicy, process-hardening, sandboxing, network-proxy)
OS-level isolation of agent-executed commands using bubblewrap namespaces/seccomp, execution policies, and outbound network control.
codex-rs/mcp layer (codex-mcp, mcp-server, rmcp-client, ext/mcp)
Model Context Protocol server/client integration allowing external tool servers to be attached to sessions.
codex-rs/auth & secrets (login, aws-auth, keyring-store, secrets, workload-identity)
Authentication flows, cloud/API credential handling, OS keyring storage, and workload identity.
sdk/typescript
Public TypeScript SDK built with tsup (ESM + dts) and linted with eslint/typescript-eslint.
The team shipped a wave of Guardian safety-review improvements, memory and history features for the terminal UI, and a batch of Windows and analytics refinements across 73 commits.
2 weeksWeek of 2026-08-24
Week of 2026-08-24latest
The team shipped a wave of Guardian safety-review improvements, memory and history features for the terminal UI, and a batch of Windows and analytics refinements across 73 commits.
Week of 2026-08-17
This large 327-commit week centered on tagging every piece of conversation content with 'content kinds' so annotations survive compaction, truncation, forks, and model switches, alongside Guardian review hardening and new computer-use configuration.
Over the past two weeks, the Codex team focused heavily on improving how the assistant's internal context is tracked and preserved—labeling every piece of conversation content with metadata so nothing gets lost during compaction, forks, or model switches. A second major theme was the 'Guardian' safety review system, which gained new scopes, better isolation, and tighter integration with tool calls. Rounding things out were quality-of-life features like Vim dot-repeat support, smarter thread titles, Windows path handling fixes, and expanded analytics for multi-agent workflows.
Week by week
2026-08-24The team shipped a wave of Guardian safety-review improvements, memory and history features for the terminal UI, and a batch of Windows and analytics refinements across 73 commits.latest6 changes
Feature
Dot-repeat comes to Vim mode
Users can now press the period key to repeat their last edit in the terminal's Vim-style editing mode.
Feature
Smarter thread titles
Threads now get descriptive titles generated from the conversation, and the /rename command suggests options based on what you've discussed.
Feature
Managed AWS keys for Bedrock
Teams using Amazon Bedrock can now supply centrally managed AWS access keys instead of local credentials.
Fix
Unicode paths survive on Windows
Batch aliases on Windows no longer break when file paths contain non-English characters.
Feature
Retry failed logins automatically
If a provider's authentication command fails on first try, Codex now retries it before giving up.
Refactor
Counting JSON without extra copies
Serialized message sizes are now measured without building temporary strings in memory, reducing overhead.
2026-08-17This large 327-commit week centered on tagging every piece of conversation content with 'content kinds' so annotations survive compaction, truncation, forks, and model switches, alongside Guardian review hardening and new computer-use configuration.6 changes
Refactor
Content kinds everywhere
User input, contextual fragments, and extension prompts are now consistently labeled with content types so important metadata isn't lost when messages are merged, truncated, or rolled back.
Feature
Browser and computer use configuration
New settings let users configure how Codex controls browsers and computers during automated tasks.
Feature
Node REPL as an MCP server
A new cua_repl option lets Codex run JavaScript through a Node-based REPL server for interactive scripting.
Fix
Guardian reviews cancel cleanly
Safety reviews now shut down properly when their associated tool calls are cancelled, avoiding orphaned background work.
Fix
Resumed threads archive fully
Archiving a thread tree now also shuts down any resumed child processes, preventing stray agents from lingering.
Chore
Package smoke tests
New end-to-end smoke tests verify that assembled Codex packages work before release.
Dependencies and code review
Dependency advisories
Security Watch
No known advisories across 1 scanned dependencies.
No known advisories in the scanned dependencies.
Code review
No concrete code-level issues confirmed.
Get this report every week for your repos.
GitZoid learns each repo, reports what changed, and flags what needs attention. One flat price for the whole team.
$19 a month, flat · First 10 outputs free · No card required