Open Terminal from an agent command: agent-owned tabs grouped by thread / lineage

[humbertogontijo]

Problem

When an agent runs a shell tool (command_execution), expanding the row shows streamed output and Stop only interrupts the whole turn. There is no way to:

1. Open that run in something that feels like a terminal (watch live, scroll, optionally type / Ctrl+C).
2. Tell at a glance that the agent owns that session (vs my own Ctrl+J / right-panel terminals).
3. Find the right session when several threads or lineage children each have long-running commands (dev server, tests, watchers).

Today the surfaces are split: user terminals live in the thread terminal UI; most providers’ commands live only as tool cards (Codex / Claude / Cursor / …). Devin already runs through T3 terminals (docs); other agents do not get open/stop of a real PTY from the command row.

Related Ideas (not duplicates — different cuts of the same pain):

• #29136 — agent uses a real shared PTY you can watch and type into (no comments yet).
• #17456 — Processes page to see/stop agent shells globally.
• #21588 — Services in thread details (closed as dup of Processes page: see and manage dev servers, agent shells and project actions in one place #17456).
• #13503 — Background a foreground Bash (Ctrl+B), not Open Terminal.

Desired UX

From a running (or completed) command row in the activity log:

• An Open Terminal action (same spirit as Open Diff on file changes).
• Opens the existing terminal surface (drawer / right panel — whatever the product uses for thread terminals), not a one-off modal.
• The tab/session is visually agent-owned: badge or chrome that says the agent controls it (e.g. “Agent” / provider name), distinct from user-created terminals.
• Tabs (or sections) are grouped by thread / lineage: parent thread, then each child/subagent, so a nested Task’s shell is not mixed into an unrelated user tab.
• While the agent owns the session: user can watch; Cancel / Ctrl+C (or a Stop on that session) ends that command without needing to Stop the whole turn when the provider allows it.
• Optional later: user typing into the same PTY (full shared session as in Let agents use a real terminal in the thread that I can watch and type into #29136). v1 can be view + interrupt only if that is easier to approve.

Why this shape (approval-friendly)

• Reuses the thread terminal UI and grouping model instead of inventing a Processes mega-page in the first PR.
• Mirrors an accepted pattern elsewhere: expand/detail fetch + Open Diff for file changes (#27298 / discussion #17976).
• One problem for a first PR after approval: wire Open Terminal + agent-owned chrome + lineage grouping for providers that already have (or can attach) a T3 terminal id (starting with ACP/Devin client terminals). Broader “every Codex/Claude/Cursor shell becomes a PTY” is the Let agents use a real terminal in the thread that I can watch and type into #29136 architecture and should stay a follow-up unless maintainers want that in scope.

Proposed phased scope (need a pick)

Phase A — Open + badge + group (recommended first PR)

• Command row → Open Terminal focuses/creates an agent-owned terminal tab.
• Group tabs by thread / lineage section.
• Ship first where a terminalId (or equivalent) already exists (Devin / ACP client terminals).
• Other providers: keep today’s output panel until Phase B.

Phase B — Provider shells in T3 PTYs

• Codex / Claude / Cursor (etc.) run (or mirror) commands into thread terminals so Open Terminal works everywhere (Let agents use a real terminal in the thread that I can watch and type into #29136).

Phase C — Per-command cancel without killing the turn

• Stop on the command/session interrupts that tool only, when the provider supports it; turn-level Stop stays as today.

Out of scope for the first PR unless you say otherwise: full Processes page (#17456), Background-as-Ctrl+B (#13503), user typing into agent PTYs.

Questions for maintainers

1. Is Phase A the right first slice, or do you want Open Terminal only after Phase B (all major providers on PTYs)?
2. Should agent-owned tabs live in the same terminal drawer/panel as user tabs (with sections), or a separate “Agent” strip?
3. For cancel: is per-command interrupt in Phase A acceptable for ACP/Devin only, or must it wait for a cross-provider design?
4. Any hard no on user input into agent sessions in a later phase?

Ask

Please approve or redirect direction and scope (especially Phase A vs B-first). Per CONTRIBUTING I will not open an implementation PR until that approval is explicit here.