Task Weight Manager

Use this skill when one conversation contains multiple strands of intent and the user wants the agent to stay oriented instead of drifting.

What this skill does

• Splits recent conversation into named threads.
• Assigns each thread a weight based on priority, urgency, dependency, and user intent.
• Chooses a mainline thread and protects it from casual drift.
• Captures interruptions into a parking lot instead of losing them.
• Periodically re-checks whether the current work still matches the highest-value thread.

Default operating model

Treat the conversation as a weighted queue of threads, not as one flat transcript.

Each thread should have:

• id: short label such as A, B, F
• title: one-line meaning
• goal: desired outcome
• weight: 1-100
• state: mainline, active, parked, blocked, or done
• evidence: the latest user turns that justify the classification
• next_action: smallest useful next step

Weighting rules

Start from the user's explicit intent, then adjust using these signals:

• +40 if the user says this is the main task, primary goal, or highest priority
• +25 if other work depends on it
• +20 if it is time-sensitive
• +15 if the user has already invested significant context into it
• -15 if it is merely exploratory or nice-to-have
• -20 if it is blocked by missing inputs

Use relative judgment, not rigid math. The purpose is stable prioritization, not fake precision.

Focus lock

If one thread clearly dominates, mark it as mainline and behave as follows:

1. Continue work inside that thread unless a higher-priority interruption appears.
2. When the user introduces a side topic, classify it, acknowledge it briefly, and park it.
3. Return to the mainline in the same response unless the user explicitly switched.

Do not silently abandon the mainline because of an interesting tangent.

Interrupt handling

When a new topic appears during focused work:

1. Decide whether it is urgent, important but not now, or background noise.
2. If not urgent, store it as a separate thread.
3. Tell the user where it was parked.
4. Resume the mainline.

Good phrasing pattern:

我把这个记到线程 B（技能体系）里，当前先继续线程 F（主线任务）。

Reflection loop

Whenever the user asks for re-prioritization, or when periodic automation wakes the agent:

1. Re-read the latest turns plus the saved thread board.
2. Merge duplicates.
3. Re-score each thread.
4. Check whether the current work still matches the highest-weight thread.
5. If drift happened, explain the drift in one sentence and steer back.

Response contract

When this skill is active, prefer compact status blocks like:

主线: F - 完成任务权重管理器 MVP
当前排序: F(92) > B(61) > D(38)
本轮动作: 继续完成 F 的下一步，并把新话题先停放到 B

Then continue the actual work.

State persistence

If the workspace allows files, maintain a small board at:

• task-weight-manager/threads.md

Keep it human-readable. Update only when the thread model materially changes.

Suggested section layout:

• Mainline
• Weighted Threads
• Parking Lot
• Blocked
• Done

Guardrails

• Do not create too many threads for tiny wording variations.
• Do not keep stale threads alive just because they once existed.
• Do not override a direct user switch of priority.
• Do not use weighting language to avoid answering the user's current explicit request.
• If two threads are truly tied, say so and ask for a priority decision only if needed.

When to read more

Read references/integration.md when the user wants this behavior to work with OpenClaw features such as HEARTBEAT.md, cron, memory, or session isolation.

Read references/usage-patterns.md when the user wants examples, Chinese command phrases, or recommended operating patterns.