Replies: 1 comment
|
The extension points look about right. The one thing I would not make primitive is A board item can be retried, forked, handed to a subagent, or resumed after a runtime restart. The old thread is still useful audit history. In BitFun we ended up needing a separate binding record: external task id, runtime session/thread id, attempt, repo/worktree identity, controller lease, and last event cursor. That changes approvals too: bind an approval to the attempt plus an action digest, not just the board item. Otherwise a stale panel can approve a newer retry by accident. If I had to stage this, I would ship native thread navigation, resumable lifecycle events, and scoped approvals before sidebar UI. A community UI can build around those; it cannot safely reconstruct them. We wrote up the process boundary here: https://github.com/GCWing/BitFun/blob/main/docs/architecture/agent-runtime-deployment-design.md |
Uh oh!
There was an error while loading. Please reload this page.
Summary
I would like to propose an open, community-extensible task coordination layer for Codex.
The goal is to support workflows commonly found in issue trackers and managed-agent platforms:
I am not asking OpenAI to build and maintain a complete project-management product.
A good outcome could be either:
Why existing capabilities are not yet sufficient
Codex already provides valuable building blocks:
codex app-serverwith JSON-RPC lifecycle eventsCommunity projects such as Threadlines demonstrate that App Server can power a complete independent Codex workspace.
However, the documented extension surfaces do not currently appear to provide a complete path for implementing this workflow inside Codex Desktop:
threadIdProposed extension points
UI contributions
ui.sidebarui.panelThread and task integration
threads.readthreads.createthreads.openthreads.linkExternalRecordthreads.subscribeLifecycleEvents should cover:
Project and worktree context
projects.readrepositories.readbranches.readworktrees.readApprovals
Automations
automations.readautomations.createautomations.updateautomations.pauseautomations.deleteNavigation
Security model
Each capability should require explicit permissions.
Plugins should not receive unrestricted access to threads, files, commands, or approvals. Workspace administrators should be able to limit installation, data access, execution, and write actions.
Sensitive actions should continue to require user confirmation.
Community contribution proposal
Individual contributors could build and maintain the taskboard, issue model, database, synchronization layer, and UI.
OpenAI would only need to provide and stabilize the integration contracts that cannot be implemented outside Codex:
An official sample plugin combining these capabilities would make community development substantially easier.
I would be willing to contribute implementation work, testing, documentation, Chinese localization, and feedback from a working taskboard prototype.
Related discussions and issues
This proposal connects those capabilities through a concrete, community-buildable workflow rather than requesting a duplicate implementation.
All reactions