Codex Desktop: automated agent/subagent runs each get a sidebar entry with no way to group or hide them

Start by reproducing a multi-agent Codex CLI task on macOS and inspect how its companion-task runs appear in the desktop sidebar. No implementation files or tests are named; done means automated agent and subagent runs are grouped, hidden, or filterable without removing access to them.

Edited after duplicate triage. This issue originally also covered CLI sessions being misfiled into the ChatGPT workspace, and coarse-grained ChatGPT/Codex history separation. Those are already tracked in #33910 and #33087 respectively, so I have removed them here and reacted to both. What remains is the part neither issue covers: automated agent and subagent runs as a distinct session class.

What version of the Codex App are you using (From "About Codex" dialog)?

26.715.21425 (build 5488) — the unified ChatGPT desktop app for macOS (Chat + Work + Codex)

What subscription do you have?

ChatGPT Pro

What platform is your computer?

Darwin 25.5.0 arm64 arm

What issue are you seeing?

Automated agent and subagent runs spawned by a Codex CLI task each get their own entry in the desktop app's sidebar, indistinguishable from sessions a person actually opened. A single CLI task that fans out to subagents can produce a dozen or more entries — in my case a long run of "Codex Companion Task" rows — none of which I started interactively or would ever want to reopen.

These are byproducts of one task, not conversations. But the sidebar treats them as peers of real sessions, so the more agent work you run, the harder your own sessions become to find. There is currently no way to collapse, group, hide, or filter them.

This is distinct from the two related issues already open:

• #33910 covers CLI sessions being routed to the wrong workspace (a misfiling bug).
• #33087 covers separating ChatGPT history from Codex history (a two-way split).

Neither addresses automated agent/subagent runs, which are a third class: machine-generated, high-volume, and rarely worth surfacing individually even when you are working entirely inside Codex. Fixing #33910 and #33087 would still leave the Codex sidebar full of companion-task rows.

What steps can reproduce the bug?

1. Run a Codex CLI task that spawns subagents or companion tasks (any multi-agent run reproduces this quickly).
2. Open the unified ChatGPT desktop app on macOS.
3. Look at the sidebar.

Observed: one entry per agent/subagent run, listed alongside sessions that were opened by hand, with nothing distinguishing the two.

What is the expected behavior?

Treat automated agent/subagent runs as their own session class and give them somewhere to live that is not the main list. Any of these would resolve it:

• Collapse the runs belonging to one parent task under that task, expandable on demand — this seems like the most natural fit, since they already have a parent-child relationship.
• Hide them from the sidebar by default, with a toggle or filter to show them.
• Mark them with a per-row origin indicator so they can at least be visually skipped, with filtering as a follow-up.

Grouping under the parent task would be the highest-value version, because it preserves access — the runs remain reachable when debugging a task — while removing them from the flat list.

Additional information

The app already distinguishes session origins internally (that is what the workspace routing in #33910 depends on), so this reads as a presentation and filtering affordance rather than new infrastructure.

Related: #33123 requests an unread-only sidebar filter. An origin-based filter would compose well with it.