OpenIDE Learns to Call In Codex for Review Straight From a Claude Chat and Wait Out Rate Limits

A familiar situation: you spend a week working with one AI agent and tailor your workflows to it, and then your account gets blocked or you hit a limit at the worst possible moment. Moving to another agent means a new interface, new habits and rebuilding your workspace. The OpenIDE team is trying to take away that pain and make the agent as replaceable a tool as a linter or a debugger.

An important caveat: the material was published on Habr on behalf of the product's authors, so it is more a presentation of the update than an independent review. But the ideas in it are practical and worth examining.

One Chat, Two Agents

OpenIDE is a multilanguage environment supporting 10+ languages. It has built-in Claude Code, Codex, OpenCode, Cursor, Gemini CLI, Russian tools and more than a dozen other agents. You can also connect your own ACP-compatible agent. Some people work on a cloud subscription, others connect a local or corporate model through OpenCode. Code, Git, tests, the debugger and databases stay in place; only the one writing the code changes.

The main novelty of the update is threads. An example from the article: Claude added exclusions to a "Secret Santa" draw app and reports that everything is ready. Before sending the link to colleagues, you need to check edge cases: can a person draw themselves, what if there are no valid combinations, could the algorithm hang. Previously you would have had to open another chat, retell the task and transfer the history by hand.

Now Codex receives the history and replies in a separate branch on the right. The main dialogue with Claude isn't cluttered with side conversation. Once the comments are sorted out, the "Merge" button passes the result to Claude, and "Auto-merge" does it automatically. A thread can also be opened with the same agent via /thread to discuss an alternative algorithm. Branches survive an IDE restart, and they can be summarized, marked as resolved or turned into separate chats.

Named Presets and Questions About Code

The second idea is to save an "agent + model + effort" combination under a name. You click the star in the input field, save the preset as @reviewer, and then write "@reviewer check the changes". Direct mentions of @codex, @claude, @opus and @sonnet also work. Several names in one message hand out tasks to several agents at once. Choosing the executor becomes part of the request rather than a separate switch in settings.

The third feature is "Ask AI…" on a selected fragment. The answer appears as a card right under the code, where you can continue the dialogue or send it back to the main chat. If the code changed after the question, the card flags it, and the agent receives both versions. For review and analysis of other people's changes, this is more convenient than copying chunks into a chat.

Overnight Tasks and Limits

You can set "Send later", for example for three in the morning, and have the agent run the draw ten thousand times and write additional tests. The message waits in a queue and goes out at the scheduled time. For teams with local models, this is a way to shift the load to hours when they are used less.

When a limit is exhausted, instead of an error a card is shown: which limit ran out and when it resets. One minute after the reset, the IDE retries the request on its own, and the agent continues in the same session. The retry can be canceled, and a new message written before the reset will wait its turn in the queue.

Limitations and Takeaway

There are caveats. Everything described is available in the paid OpenIDE Pro version; only a two-month trial period is free. For delayed sending, the computer must be on and the IDE open. The claim that switching agents reduces changes to zero should be tested on your own projects: different agents have different behavioral quirks and tools.

The practical takeaway for developers and team leads: the "one agent writes, another reviews in a separate branch" pattern is useful regardless of the specific IDE. Different models catch different errors, and isolating side discussion protects the main chat's context. If you depend on a single vendor, plan a fallback in advance. OpenIDE's near-term plans include tracking file changes at each agent step, an improved AI Explorer and Telegram integration.