Upstream's biggest new subsystem: background subagents, managed shell pool, and the task_stop / send_message model-facing agent control surface.

Upstream PRs (in dependency order)

• feat: background subagents with headless and SDK support QwenLM/qwen-code#3076 background subagents with headless and SDK support
• feat(cli): headless support and SDK task events for background agents QwenLM/qwen-code#3379 CLI headless support and SDK task events
• feat(core): model-facing agent control (task_stop, send_message, per-agent transcript) QwenLM/qwen-code#3471 model-facing agent control (task_stop, send_message, per-agent transcript)
• feat(cli): background-agent UI — pill, combined dialog, detail view QwenLM/qwen-code#3488 background-agent UI — pill, combined dialog, detail view
• feat(core): managed background shell pool with /tasks command QwenLM/qwen-code#3642 managed background shell pool with /tasks command
• feat(core): wire background shells into the task_stop tool QwenLM/qwen-code#3687 wire background shells into the task_stop tool
• feat(cli): wire background shells into combined Background tasks dialog QwenLM/qwen-code#3720 wire background shells into combined Background tasks dialog
• Add background agent resume and continuation QwenLM/qwen-code#3739 background agent resume and continuation

Why

Significant new capability surface — model-driven background work, parallel tool execution, status/control via /tasks. Maps closely to our subagent direction but expands it substantially.

Scope / effort

XL. Aggregated stats from prior investigation:

• 159 files touched (~100-110 unique after dedup)
• +13,109 / -1,703 lines (net +11,406 LOC)
• 72 conflict zones with our fork
• 7 of 9 PRs have intra-bundle dependencies — strict merge order required
• Permeates config.ts, agent-core.ts, nonInteractiveCli.ts, AppContainer.tsx, dialog management, tool scheduling

Decision point: Porting this locks us into upstream's subagent / background-task model. If we diverge further on agent lifecycle or task orchestration, future merges become harder. If we plan to converge with upstream on agents, this is the moment.

Prerequisites

• Architecture review: do we want upstream's task_stop/send_message API surface as our model-facing agent control?
• Plan: dedicated feature branch, do not interleave with other upstream merges.
• Estimate: 2-3 days integration + 1-2 follow-up PRs for fixes.

Links

• feat: background subagents with headless and SDK support QwenLM/qwen-code#3076
• feat(core): model-facing agent control (task_stop, send_message, per-agent transcript) QwenLM/qwen-code#3471
• feat(core): managed background shell pool with /tasks command QwenLM/qwen-code#3642
• Add background agent resume and continuation QwenLM/qwen-code#3739