research-queue（研究队列）

Research Queue

Use this skill when work produces unresolved questions that deserve deliberate investigation instead of ad-hoc guessing.

Core workflow

1. Read QUESTIONS.md.
2. If the file is missing, initialize it using the format in references/queue-format.md.
3. Select at most one question unless the user explicitly asks for batch processing.
4. Prefer the highest-priority open question with the strongest practical value.
5. Investigate using only the minimum tools needed.
6. Update the question entry in place with:

- status - startedAt / completedAt when applicable - evidence summary - conclusion or blocked reason - memory note path if something durable was learned

1. If the final status is done or wontfix, move the entire question block from ## Active Questions to ## Completed Questions in the same edit pass.
2. Write durable findings to memory/YYYY-MM-DD.md when they are worth remembering.

Allowed investigation methods

Choose the lightest method that can answer the question:

• web_search for fast grounded search
• web_fetch for reading specific pages
• browser only when a real browser is needed
• read for local docs/config/code
• exec for bounded local experiments, scripts, benchmarks, math checks, and repro steps

Allowed local coding includes short Python, Rust, JavaScript, or shell experiments when they directly help answer the selected question.

Hard rules

• Investigate one question at a time by default.
• Do not silently delete old questions.
• Do not mark a question done without a concrete conclusion or result.
• Distinguish clearly between verified findings, hypotheses, and blocked work.
• When a question becomes done or wontfix, move it out of ## Active Questions immediately; do not leave completed items in the active section.
• Use plausible real timestamps only; never write a completedAt in the future relative to the current run.
• Do not modify production code or config unless the user explicitly asks for implementation work; research mode may create bounded scratch experiments or propose diffs instead.
• Do not let the queue turn into vague brainstorming sludge; rewrite unclear entries into concrete answerable questions.

Status model

Use only these statuses in QUESTIONS.md:

• open — not started yet
• investigating — currently being worked
• blocked — cannot finish without missing input/access/time
• done — answered well enough for now
• wontfix — intentionally dropped

First-use initialization

If QUESTIONS.md does not exist:

1. Read references/queue-format.md.
2. Create QUESTIONS.md from the canonical template.
3. Keep the initial headings and status conventions intact.
4. Add only the starter examples or the user’s real questions; do not pad it with fake research tasks.

Automation

Prefer OpenClaw cron when the user wants this queue processed automatically.

• Default to cron, not HEARTBEAT.
• Use HEARTBEAT only if the user explicitly wants main-session, drift-tolerant background attention.
• Be explicit that this means the OpenClaw cron tool, not Unix crontab.
• Prefer isolated cron runs so research work does not pollute the main conversation.
• For scheduling details and the minimal run prompt, read references/automation.md.

What good completion looks like

A finished question should leave behind:

• a stable answer or a bounded partial answer
• the evidence source or experiment basis
• completion time that matches the actual run window
• relocation of finished work into ## Completed Questions
• durable memory only if the result is likely to matter later

References

• Queue structure and canonical template: references/queue-format.md
• Automation and OpenClaw cron guidance: references/automation.md