Subagent activity exposes internal /root paths and repeats lifecycle states

Subagent activity exposes internal /root/... paths and repeats lifecycle states

Summary

Codex's multi-agent activity UI shows internal routing paths such as
/root/task1_quality_review instead of human-readable task names. In the same
activity sequence, lifecycle rows can repeat and appear out of order, for
example showing a task as both in progress and completed.

This affects the CLI transcript and other surfaces that consume the same
subagent activity data, rather than one integration or project configuration.

Environment

• Codex CLI: 0.154.0
• Platform: Linux
• Multi-agent workflow: enabled
• No custom [agents] display-name configuration is present.

Steps to reproduce

1. Start a Codex session with multi-agent collaboration enabled.
2. Ask the parent agent to delegate named tasks, for example
   task1_quality_review and task2_implementation.
3. Let one task complete while the next task starts.
4. Inspect the parent activity feed or terminal transcript.

Actual behavior

The feed renders canonical internal paths:

Started `/root/task1_quality_review`
Completed `/root/task1_quality_review`
Started `/root/task2_implementation`
Completed `/root/task2_implementation`

The same run can also render duplicate waiting/completed rows or leave an
earlier prose status saying a task is still in progress after its completion
row has appeared.

Expected behavior

Display a human-readable task title, for example:

Started Task 1: Quality review
Completed Task 1: Quality review
Started Task 2: Implementation
Completed Task 2: Implementation

Each task should have one coherent lifecycle. Replayed or duplicated events
should not create duplicate rows, and terminal states must not be followed by
older non-terminal states.

Investigation

The current TUI renderer constructs activity text directly from agent_path:

• codex-rs/tui/src/multi_agents.rs
  • sub_agent_activity_summary() formats Started \{agent_path}`andCompleted `{agent_path}``.
  • sub_agent_activity_title() does the same for rendered history rows.

The current upstream tests explicitly snapshot /root/reviewer as the
displayed label in codex-rs/tui/src/app/agent_status_feed_tests.rs.
This confirms the result is current product behavior rather than a stale local
installation or an Orca-specific formatting problem.

Suggested direction

Keep agent_path for routing, but carry or derive a separate display title:

1. Prefer the original task_name supplied to spawn_agent.
2. If no title is available, humanize the final path segment as a fallback.
3. Centralize lifecycle reduction per agent path/thread so duplicate events are
   ignored and terminal states cannot regress to waiting or working.
4. Use that display projection for the CLI, desktop activity feed, and other
   app-server consumers.

Related reports

• #24225: duplicate finished-subagent notifications
• #30813: canonical /root/... paths and inconsistent active-agent state
• #43689: subagent lifecycle/state visibility is inconsistent

Start in codex-rs/tui/src/multi_agents.rs, tracing sub_agent_activity_summary() and sub_agent_activity_title(), then inspect codex-rs/tui/src/app/agent_status_feed_tests.rs and its /root/reviewer snapshots. Confirm the activity projection preserves task names or humanized fallback labels, deduplicates lifecycle events, and prevents terminal states from regressing; done when the affected tests cover the expected coherent display.