Codex agents coordinate codebase work through shared session IDs

• Codex agents can coordinate across parallel threads when one thread is given the other's session ID; LLMJunky shows the setup in his Codex tip and says the agents "chat back and forth" in his follow-up.
• The product model already treats a Codex conversation as a durable Thread; OpenAI's App Server docs define threads, turns, items, and thread APIs, which makes LLMJunky's session-ID workflow less mysterious.
• Search is the obvious workflow gap: LLMJunky says Codex history search often needs the exact string in his search complaint, while MengTo asks for project search with thumbnails, dates, or an infinite-canvas flow in his feedback list.
• Codex now lives inside ChatGPT desktop, per OpenAI's Codex what's-new page, and the split is visible in Peter Yang's ChatGPT/Codex toggle post.
• Creator workflows are moving into Codex skills: MengTo's bookmark-to-post system is described in his skills reply, with public skills for bookmark quote posts and writing in his X voice.

LLMJunky's right-clicked session-ID trick can make one Codex thread send a message from another chat while it is editing files. OpenAI's App Server docs describe Threads as conversation containers with IDs, turns, streamed events, and steering. The worktrees docs add the parallel-project layer, while OpenAI's Work launch post frames Codex technology as the engine behind ChatGPT's broader work mode.

LLMJunky's setup is simple enough to copy into a sticky note:

• Start two Codex threads in the same codebase.
• Right-click one thread to get its session ID.
• Give that session ID to the other agent.
• Ask the agents to collaborate.

The follow-up claim is the useful part: the agents chat back and forth, wait for each other, provide updates, and request updates, according to LLMJunky's follow-up.

The screenshoted exchange includes a message labeled "Sent by Codex from another chat," then a second agent adjusting its own work around the first agent's files. LLMJunky's one-word verdict was "peffect" in his reply, typo included.

A related cross-agent pattern showed up outside Codex: gokayfem said users with both subscriptions can tell Codex or Claude to talk to each other, after showing Claude Fable 5 and GPT-5.6 Sol discussing a scientific-discovery workflow in his cross-agent post.

OpenAI's public docs describe the lower-level machinery behind rich Codex clients, not the right-click UI trick itself.

In the Codex App Server docs, the primitives are:

• Thread: a conversation between a user and the Codex agent.
• Turn: one user request and the agent work that follows.
• Item: a message, command run, file change, tool call, or other unit of input or output.

The same docs show thread/start returning a thread ID like thr_123, turn/start accepting a threadId, and lifecycle methods for thread/resume, thread/fork, thread/read, thread/list, and turn/steer.

OpenAI's worktrees docs cover the codebase side: worktrees let Codex run multiple independent chats in the same project without interfering, and Handoff moves a chat between Local and Worktree.

A July 16 OpenAI Developer Community request asks for ChatGPT to connect to one or more active Codex sessions, monitor progress, review work, and send instructions through natural conversation. A July 18 command-center request asks for one main project chat that can coordinate pinned worker chats, check progress, notice blockers, and assign follow-ups.

LLMJunky says Codex search is noisy and often needs the exact string. He doubled down in replies, saying it "could be so much better" in one reply and adding that OpenAI could "maybe" add natural-language search in another reply.

Today, the workaround he pointed to is navigation, not retrieval: users can cycle through recently viewed chats, according to his shortcut reply.

MengTo's feedback list turns that pain into a product map:

• Sites should have grid-style previews and reopen in the previous chat for edits, or in a default Project called Sites.
• Skills should be community-oriented, with author pages and usage analytics.
• ChatGPT Live should work for Codex projects.
• Pet should move into the menu bar with notifications.
• Old threads need project search, thumbnails, dates, or an infinite-canvas flow chart.
• A second-brain view should browse articles, workflows, skills, and memories.
• Mobile prompting should stop asking for permissions when full access is enabled.
• Browser needs bookmarks, history, and a standalone-window option.

The App Server docs already expose thread/list with a searchTerm field, but the user request is semantic retrieval across messy work history.

Search is not the only way Codex users are navigating long sessions. LLMJunky showed a prompt dropdown that lets users jump between prompts inside a long thread, and called it a feature most people had probably never seen in his demo.

He praised a related feature as "sorely needed" in his reply to Jack. The pattern is narrow but useful: recover a prior prompt state without treating the whole thread as a flat transcript.

OpenAI's what's-new page says Codex joined the ChatGPT desktop app on July 9 for macOS and Windows, while keeping a dedicated coding experience alongside Chat and Work modes. The same page says the updated desktop app is available globally on every ChatGPT plan, including Free.

Peter Yang first asked whether ChatGPT Work was dead, then found a dropdown with two modes: ChatGPT for "Create, learn, and explore" and Codex for "Build, debug, and ship," as shown in his follow-up. Dan Shipper expected a separate Codex app instead, according to his reply.

The boundary is still visible in actual use. LLMJunky kicked off a frontend task in Codex and saw Chrome automatically display a "CODEX IS WORKING" overlay before code appeared, then clarified that it happened in the app in his reply.

Peter Yang's blunt product read was that the merge is not fixed until the work and chat tabs merge into one, according to his reply.

MengTo's Codex-adjacent workflow is a creator case study, not a coding demo. He said he hit 5M X impressions over three months after changing his posting mix while on vacation: quote posts from bookmarks, reposts of high performers, and original posts with useful URLs, skills, tools, apps, or workflows.

His AI loop mines bookmarks and drafts posts in his voice using memories, but he still writes many first drafts or heavily edits the output, according to the parent post in his skills reply.

The two public skills make the workflow concrete:

• x-bookmark-quote-posts turns recent X bookmarks into source-backed quote-post drafts calibrated against up to 100 authored posts.
• write-like-meng-on-x drafts, rewrites, reviews, or refines X posts using MengTo's authored-post corpus, product context, shared resources, and content repo evidence.

MengTo said he plans to update the skills regularly in his follow-up.