TUI renders duplicate sub-agent activity cells (Started/Interacted/Interrupted) for every multi-agent v2 tool call since 0.146.0

Note: I understand code contributions are by invitation per docs/contributing.md. I'm filing this bug report with full root-cause analysis and a tested fix on my fork — happy to open a PR if invited, or the analysis above should make an internal fix trivial.

What version of Codex CLI is running?

Reproduced against current openai/codex main / 0.146.x+. Bug present in 0.146.0, 0.146.1, and the 0.148.0-alpha line per source analysis below. Run codex --version locally if needed.

What subscription do you have?

N/A — protocol-level bug, not plan-specific.

Which model were you using?

Any model using multi-agent v2 sub-agent tools (protocol-level, not model-specific). Observed with GPT-5.6 Sol / Sol Ultra.

What platform is your computer?

Linux 6.8.0-137-generic x86_64 x86_64

What terminal emulator and version are you using (if applicable)?

Ghostty; also reproduced in other terminals.

Codex doctor report

not available

What issue are you seeing?

Whenever a session uses multi-agent v2 sub-agent tools (spawn_agent, send_message, followup_task, interrupt_agent), the TUI transcript renders each sub-agent activity cell twice, back to back. Example:

• Started `/root/bundle_c_gate_patch_review`
• Started `/root/bundle_c_gate_patch_review`

The same duplication affects "Interacted with path" and "Interrupted path" cells.

Root cause (from source analysis of this repo):

1. PR #35363 ("Include item start times in completion events", merged 2026-07-25) changed emit_sub_agent_activity in codex-rs/core/src/tools/handlers/multi_agents_v2.rs from emitting only an item-completed notification to emitting both emit_turn_item_started and emit_turn_item_completed for the same SubAgentActivity item (so the completion can carry started_at_ms).

2. The TUI has rendered a SubAgentActivity history cell on both notification paths since PR #27007: the ItemStarted handler in codex-rs/tui/src/chatwidget/protocol.rs and the ItemCompleted path via handle_thread_item in codex-rs/tui/src/chatwidget/replay.rs. Unlike CommandExecution/McpToolCall/CollabAgentToolCall, SubAgentActivity has no status field to branch on, so both paths print identical cells. The ItemStarted arm was dead code while the item was completion-only (as designed in #27007) — #35363 made it live and created the collision.

First shipped in rust-v0.146.0 (2026-07-29); still present in 0.146.1, in 0.148.0-alpha.6, and on main as of 2026-08-10. The /agent status feed is not affected (it dedupes by item id), and codex exec human output never rendered these items on ItemStarted — the TUI transcript is the only affected surface.

What steps can reproduce the bug?

1. Run the codex TUI (0.146.0 or later) with a model/config where multi-agent v2 sub-agent tools are available.
2. Give the agent any task that causes it to spawn a sub-agent (spawn_agent), or send input to / interrupt one.
3. Observe the transcript: every "Started agent_path" (and Interacted with / Interrupted) cell appears twice consecutively.

What is the expected behavior?

Exactly one transcript cell per sub-agent activity, matching pre-0.146.0 behavior and matching session replay (which reconstructs from completed items only).

Additional information

Fix ready on fork — rebased on latest openai/codex main, can open PR immediately if invited.

• Branch: BenItBuhner/codex:fix/duplicate-subagent-activity-cells (rebased onto openai/codex main as of d10939327, 2026-08-10)
• Change: one-arm fix in codex-rs/tui/src/chatwidget/protocol.rs — stop rendering SubAgentActivity on ItemStarted; keep rendering on ItemCompleted via replay.rs. Core's dual emission left intact (#35363 added it deliberately for started_at_ms).
• Regression test: sub_agent_activity_started_and_completed_renders_one_cell in codex-rs/tui/src/chatwidget/tests/app_server.rs — feeds ItemStarted + ItemCompleted for one item, asserts a single cell. Fails without the fix, passes with it.
• Verified: just test -p codex-tui green (3354 passed); just fix -p codex-tui clean.

I understand code contributions are by invitation per docs/contributing.md — happy to open a PR with this if invited, or the analysis above should make an internal fix trivial.

Start in codex-rs/tui/src/chatwidget/protocol.rs and compare the SubAgentActivity handling with the completion path in codex-rs/tui/src/chatwidget/replay.rs. Run the sub-agent activity scenario in codex-rs/tui/src/chatwidget/tests/app_server.rs, especially sub_agent_activity_started_and_completed_renders_one_cell. Done means one transcript cell is rendered for each activity while replay and the codex-tui test suite remain passing.