[Bug]: Cursor ACP backgrounds a shell, ends the turn, and never comes back

[Jonathanm10]

Before submitting

• I searched existing issues and did not find a duplicate.
• I included enough detail to reproduce or investigate the problem.

Area

apps/server

Steps to reproduce

1. Use the Cursor provider (agent acp / cursor-agent ACP).
2. Ask it to run a long shell with block_until_ms: 0 (or otherwise background the shell) and to end the turn saying it will come back when finished.
3. Wait well past when the shell would complete. Do not send another message.

Minimal bare-ACP repro (no T3 involved) also reproduces the silence:

# Prompt Cursor ACP to Shell `sleep 15 && echo BG_DONE` with block_until_ms: 0,
# reply BACKGROUND_STARTED_WILL_COME_BACK, end turn, no AwaitShell.
# After session/prompt returns stopReason=end_turn, wait 20s: zero session/update
# and zero cursor/* notifications follow. A later user prompt can report BG_DONE.

Expected behavior

After backgrounding work and promising to continue, the agent resumes when the shell finishes (same as Cursor’s interactive CLI harness), or clearly does not promise a comeback when the harness cannot wake it.

Actual behavior

session/prompt returns stopReason: end_turn. Nothing arrives afterward. The thread looks idle until the user prompts again; then Cursor can inspect the finished shell and report status.

Impact

Major degradation or frequent failure

Environment

• Cursor CLI / agent acp (reproduced on 2026.09.18-9a7762b)
• T3 Code Cursor provider (ACP adapter) — symptom matches; root cause is in Cursor ACP, not T3 dropping events

Additional context

Duplicate check (not duplicates):

Item	Why related / why not a duplicate
#6644 (closed)	Cursor subagent/Task Agents-panel + parent wake. Same “won’t continue without a nudge” class, but Task/subagent lifecycle — not background shells.
#9107 / #10295	Claude run_in_background shells missing from the timeline UI. Different provider and symptom.
#9139	Grok background shell task.* lifecycle. Different provider.
#13133	ACP assistant-segment fragmentation when background tools update mid-stream. Different failure mode.
#9047	Cursor adapter turn-lifecycle races / missing terminal events. Unrelated to post-end_turn shell wake.

Root cause (investigated against Cursor ACP + bare client):

• Interactive Cursor CLI wakes via backgroundTaskCompletionAction when background work finishes.
• Cursor ACP’s completion drain (subagent-completion-drain) only processes subagent completions and requeues non-subagent (shell) completions.
• A bare agent acp client (no T3) sees the same silence after end_turn, so T3 is not dropping a wake notification — Cursor ACP never emits one for shells.
• Shell tool calls often present as completed with empty stdout immediately when backgrounded, so T3 also has no live in_progress signal to keep backgroundLiveness honest.

T3-shaped options (not mutually exclusive):

1. Upstream / protocol: Cursor ACP should drain shell completions into a follow-up turn the way interactive CLI does (and the way ACP already does for subagents).
2. Adapter-boundary mitigation: Cursor-specific runtime instructions that the ACP harness will not resume after a backgrounded shell — await in-turn, or tell the user to send another message. Matches how T3 already patches harness gaps; does not invent synthetic user turns.
3. Not preferred: Host-invented synthetic sendTurn without a provider wake signal — fights orchestration and Claude/Grok patterns (those continue only when the provider keeps streaming / emits task lifecycle).

Contribution

• I would be open to helping implement this.

[juliusmarminge]

Triage

Confirmed on current main (f5ef0ddb9). This is a Cursor ACP hole, not T3 dropping a wake.

session/prompt returns stopReason: end_turn, and nothing follows. The Cursor notification consumer stays alive for the session (forked into the session scope, #6538), and sendTurn only then drains queued events and emits turn.completed, which clears activeTurnId. A later session/update would still be projected. A bare agent acp client seeing the same silence means there is no notification for T3 to miss.

What T3 cannot see:

• CursorAdapter forwards tool and content updates only. It never emits task.started / task.progress / task.completed. Sidebar backgroundLiveness and the session reaper are fed only from those events.
• Claude, Codex, Grok, and Antigravity do emit task.* for work that outlives the turn. Antigravity promotes an execute tool that is still inProgress when the turn ends. A Cursor shell that shows up already completed, with empty stdout, is not that signal, so there is nothing to promote.
• With backgroundLiveness left null, the thread looks idle. After 30 minutes the reaper can stop the provider session and kill the shell still running inside agent acp.

Upstream: Cursor's ACP docs only define cursor/task for subagent completion. There is no documented notification for a background shell finishing after end_turn. ACP allows further session/updates while the agent is idle; Cursor does not send one for this case. The CLI names in the report (backgroundTaskCompletionAction, subagent-completion-drain) are outside this repo and were not re-run here. They match that public contract.

T3 should not invent a follow-up sendTurn. Grok and Claude continue only when the provider keeps streaming or emits task lifecycle. A runtime-instruction warning ("this harness will not resume after a backgrounded shell") would be a provider-layer change; #11214 (Cursor Task → task.*) was closed on 2026-09-19 pending the V2 rewrite. cursor/task is also unhandled today; that is the subagent gap in #6644, and wiring it would not wake shell completions.

Not duplicates: #6644 (Task/subagent wake), #9047 (turn-lifecycle races), #13133 (assistant splits on mid-stream tool updates), #9107 / #10295 (Claude timeline), #9139 (Grok task lifecycle).

Until Cursor ACP emits a shell completion after end_turn, the workaround is to wait in-turn (do not use block_until_ms: 0 if you need the result this turn) or send another message. A promise to come back when the shell finishes will not be kept.

[juliusmarminge]

Thanks for reproducing the background-shell silence outside T3 too. Orchestrator V2 has merged in #2829.

Cursor execution now uses the official SDK rather than ACP. We are closing this transport-specific report as obsolete. The old ACP completion drain is no longer the path T3 runs, and any missing shell wake on the SDK path needs its own reproduction.

If the problem persists on a build containing V2, please open a fresh issue with the T3 version, provider version, reproduction steps, and relevant logs, and link this report.