- name
- background-delivery-sop
- description
- Standardize how agents handle tasks that run in the background, via sub-agents, or via ACP sessions, so completed work is always proactively delivered back to the user. Use when creating or improving agent workflows for async/background execution, fixing cases where results finish but are not sent, or teaching another agent how to do receipt → execution → final delivery without making the user chase for updates.
Background Delivery SOP
Use this skill to enforce a simple rule: if work finishes in the background, the user should not have to ask for the result.
This skill is for orchestration and delivery discipline, not for the task domain itself.
Core Rule
Treat every background completion as a delivery obligation.
If a task was delegated to:
- a background run,
- a sub-agent,
- an ACP session,
- or any async workflow,
then the parent / main agent must actively close the loop with the user.
5-Line SOP
- If no tools are needed, answer in the main session.
- If tools are needed, run in background instead of expanding execution in the main session.
- Send exactly one short start acknowledgment.
- When background work completes, proactively deliver the result / progress / blocker.
- Do not wait for the user to ask again unless the completion adds nothing new.
Delivery States
When a background task returns, classify it into exactly one of these states and act immediately.
1. Final result available
Send the final result now.
Use when:
- the requested work is complete,
- the user can act on the answer immediately,
- no further confirmation is required.
Output pattern:
- concise conclusion,
- essential evidence or summary,
- next step only if useful.
2. Partial progress available
Send a progress update now.
Use when:
- work is not fully complete,
- but there is meaningful progress,
- or the user benefits from seeing interim findings.
Output pattern:
- what is done,
- what remains,
- expected next step / ETA if known.
3. Blocked
Send the blocker now.
Use when:
- approval is needed,
- input is missing,
- permissions or environment prevent completion,
- or an external dependency failed.
Output pattern:
- what is blocked,
- why,
- exactly what the user needs to provide or approve.
Required Two-Phase Pattern
Use this fixed pattern for background work:
Phase 1: Start acknowledgment
Send one short message when the task begins.
Examples:
- "我去后台查，查完主动回你。"
- "我先放后台跑，完成后直接给你结果。"
- "我去处理，完成后我会主动同步。"
Rules:
- send it once,
- keep it short,
- do not repeat synonymous "already working on it" messages.
Phase 2: Completion delivery
When the background task completes, send exactly one substantive follow-up:
- final result,
- meaningful progress,
- or blocker.
Never leave the task in a state where:
- the start acknowledgment was sent,
- the background run completed,
- but no user-facing delivery followed.
Completion Event Handling
When a completion event reaches the main agent, default to this interpretation:
This is not just internal orchestration data. This is pending user delivery.
So the default action is:
- deliver, not ignore;
- summarize, not dump raw internal logs;
- close the loop, not wait for another prompt.
When NO_REPLY Is Allowed
Only allow silence if at least one of these is true:
- The returned content is materially identical to what the user already received.
- The user has already received an equivalent final answer.
- The agent is suppressing a duplicate internal event and will immediately send the proper delivery message.
If none of those are true, do not swallow the completion.
Anti-Patterns
Avoid these failure modes:
- Background task completes, but the agent says nothing.
- Main session treats completion as internal-only and never converts it into a user-facing update.
- Agent waits for "any update?" before sending finished work.
- Agent sends multiple start acknowledgments but no real completion delivery.
- Agent dumps raw internal status instead of a cleaned user-facing answer.
Quick Decision Tree
If the user asks a simple question and no tools are needed
Answer directly in the main session.
If tools or multi-step work are needed
Move execution to background / sub-agent / ACP as appropriate.
After spawning or delegating
Send one short acknowledgment.
When the async work finishes
Ask:
- Is there a final result? → deliver it.
- Is there useful progress? → send it.
- Is there a blocker? → explain it.
- Is it only duplicate noise? → silence is allowed.
Minimal Handoff Template For Other Agents
Use this template when teaching another agent or building a workflow rule:
If you move work into the background, you own the final delivery. Send one short acknowledgment when work starts. When the background task completes, proactively send either the final result, a meaningful progress update, or a blocker. Do not wait for the user to ask again unless the completion is truly duplicate.
Success Criteria
This SOP is being followed if:
- users do not need to chase completed work,
- background tasks consistently end with proactive delivery,
- start acknowledgments are short and non-repetitive,
- completion events are converted into clean user-facing updates.
Adaptive Thinking Profile
Use task-level adaptive thinking instead of trying to hot-switch the main conversation constantly.
Recommended mapping:
- Normal chat / simple direct answers →
off - Routine tool tasks / short checks →
low - Multi-step analysis / complex diagnosis / multi-source research →
medium - Programming / code changes / engineering debugging / ACP coding harness work →
high
Recommended operating pattern:
- keep the main session lightweight by default,
- raise thinking mainly in background runs, sub-agents, or ACP sessions, not through frequent main-session toggling.
中文极简版
适合直接教给中文 agent 的一句话原则：
只要任务转入后台、sub-agent 或 ACP 执行，完成后就必须主动把结果发给用户；不要等用户追问。
中文 5 行 SOP
- 不用工具：主会话直接答。
- 需要工具：默认后台执行。
- 启动时：只发一次简短回执。
- 完成后：主动发结果 / 进度 / 卡点。
- 非重复内容：禁止等用户追问才回复。
中文交付口径
后台任务完成后，主会话必须三选一：
- 有最终结果 → 直接交付结果
- 有阶段进度 → 主动同步进度
- 被阻塞 → 主动说明卡点和所需确认
中文反模式
以下都算没按 SOP 执行：
- 后台做完了但不发
- 用户不追问就不回
- 连续发“在处理”但不给结果
- 把完成事件当内部消息吞掉