Codex CLI v0.146.0: Nameable, Pinnable, Forkable Threads and Agent Plugin Marketplaces#
Part of the Codex CLI Version Tracker series. | Codex on GitHub | Official Changelog
Codex CLI v0.146.0, the first release tracked since the 0.144.3 baseline, turns conversations into managed threads you can name, pin, and fork, and opens Agent Plugin marketplaces for Amazon Bedrock and Claude Code.
Threads become first-class#
The headline for daily use is thread management. In 0.146.0 you name a new session with /new or /clear, pin the threads that matter, and switch between side conversations without closing any of them. You can also fork a thread with its full paginated history, including temporary forks that never show up in the thread listing. That builds on 0.145.0, which introduced the experimental paginated thread history underneath: efficient resume, search, persisted names, sub-agent support, and per-thread memories. Together these move Codex from a single scrolling transcript toward a workspace of parallel, resumable conversations.
Plugins, marketplaces, and cross-tool import#
0.146.0 adds Agent Plugin manifests, workspace plugin publishing, and two new plugin marketplaces, one for Amazon Bedrock and one for Claude Code. The source side confirms the plumbing: new codex-agent-extension and codex-skills-extension-host-provider identifiers appear alongside the changelog entries. Pair that with 0.145.0's /import, which migrates settings, MCP servers, plugins, sessions, slash commands, and project-scoped memories from both Cursor and Claude Code, and Codex now has an explicit on-ramp from the two coding agents most of its target users already run.
Remote execution, skills, and providers#
The app-server can now reach remote Code Mode hosts over a WebSocket connection, extending the remote-execution surface that was already in the baseline; the source adds codex-code-mode-tool-names and new remote and workspace root gates to match. Skills discovery got safer and more resilient: Codex can enumerate executor-provided skills and read their associated resources, including ones you explicitly select, and it now retains more skills under a tight context budget and warns when a skill catalog has to be truncated.
On the model side, 0.146.0 enables standalone web search for compatible custom model providers. 0.145.0 stabilized the opt-in multi-agent V2 experience (configurable sub-agent models, reasoning levels, and concurrency), added experimental Amazon Bedrock login, and named GPT-5.6 Sol the default Bedrock model, a string that was already present in the source at the 0.144.3 baseline. The same 0.145.0 window added audio inputs and tool outputs plus streaming realtime V3, and the source dropped its codex-realtime-webrtc and codex-realtime-webrtc-tokio gates as the realtime path consolidated on the V3 streaming code.
Reliability work is heavy in this range. Proxies are now honored across authentication, plugin downloads, MCP authorization, remote execution, WebSockets, redirects, and LM Studio connections; MCP and Apps tool connections stay current across auth and config changes without restarting healthy servers; and submitted messages, final responses, approval settings, and imported timestamps survive interruptions, replay, imports, and forks. Windows navigation keys and reliable termination of sandboxed process trees were both fixed.
Inside the release#
The diff from rust-v0.144.3 to rust-v0.146.0 is 563 commits landed between July 20 and July 29, roughly 65 a day, from 59 distinct authors. The visible top contributors nearly all carry -oai or -openai staff handles (jif-oai leads the sampled set with 44 commits, ahead of bolinfest at 21 and a long staff tail), so this is a fully staffed sprint rather than two or three people carrying a subsystem. Almost all of it lands in one place: 284 of 300 sampled file changes sit under codex-rs, with roughly 10,300 lines added against 1,700 removed in the sample, so the release grows the surface rather than trimming it.
The new environment surface reinforces where that work went. Most of the added CODEX_* names are test scaffolding for proxy and WebSocket paths (CODEX_*_PROXY_TEST_*, CODEX_WEBSOCKET_*_PROBE_*), which lines up with the proxy-everywhere and WebSocket bug-fix bullets in the changelog.
What this tells us#
OpenAI is shaping Codex into a persistent, multi-thread, multi-agent workstation rather than a one-shot terminal assistant. Naming, pinning, forking, and side conversations are the same session-durability features that Claude Code has been building toward with resume and checkpointing, and the paginated history underneath is what makes them cheap to keep around.
The plugin and import work reads as a deliberate interop play. Shipping a Claude Code plugin marketplace and an /import that pulls Cursor and Claude Code configuration, MCP servers, and memories lowers the switching cost from exactly the two harnesses Codex competes with most directly. Combined with standalone web search for custom providers and Bedrock login, the harness is being built to run against models and setups that are not OpenAI's own.
The remaining weight is enterprise and remote execution: proxy handling across every network path, WebSocket Code Mode hosts, and managed multi-agent concurrency. That is the surface a company needs before it lets an agent run untrusted commands behind a corporate proxy, and it is where a fully staffed nine-day sprint chose to spend its 563 commits.
This analysis is based on publicly available open-source code and release metadata, conducted for educational and research purposes. All trademarks and software referenced belong to their respective owners.