costaff

Activate for ANY request involving specialist delegation — single or multi-step. Contains the complete orchestration SOP: plan-and-confirm format, sequential execution rules, how to write a complete `request` argument, recovery from errors, retry limits, file delivery rules, and output presentation.

Activate when a multi-agent orchestration goes wrong. Contains: recovery from `ValueError: Tool '<name>' not found`, the full forbidden-tool reference for each specialist, retry limits and abandon criteria, recovery if you accidentally invoked two specialists in parallel, and the catalogue of historically-reproduced mistakes. Pair with `multi-agent-chain` (the always- loaded core skill) — that one has the happy-path rules; this one has the escape hatches.

Use when delegating report generation, PDF creation, chart visualization, or data interpretation to the business analysis expert. Load this skill before calling `business_analysis(request='...')` so you know what to send, what it returns, and which tools are internal to that specialist (and must NEVER be called directly).

Use when delegating any task to the coding expert — Python scripting, data analysis, SVM / ML algorithms, file I/O, package installation, git operations, or running tests. Load this skill before calling `coding(request='...')` so you know what to send and how to read the response.

Use when a task involves database operations — SQL queries, schema inspection, data retrieval, record insertion/update/deletion, or database health checks. Load this skill before calling `database(request='...')`. IMPORTANT: This specialist may not be deployed — check the registered tool spec first.

Run before marking any task as done. Parses the task spec for output file paths and verifies each file actually exists on disk. If any required file is missing, requeues the task instead of closing it. Skip for tasks with no file output.

Activate before executing any substantive work that produces artifacts or requires multi-step execution (coding, data analysis, report generation, multi-agent chains). Guides you to check past epics for context, register a new Epic/Story/Task when none exists, mark tasks as doing before execution, and close them after completion.

Activate when the user asks to create a new project, view project status, manage epics/stories/tasks, set execution priority, or query an agent's workload. Also activate when scheduling work as a project task with a cron (future execution).

Activate when the user wants the agent team to perform a recurring task automatically on a schedule — e.g. daily news summary, weekly report, nightly diary. Not for one-time reminders — use create-reminder for that.

Activate when the user wants a one-time message sent to them at a specific future time (e.g. "提醒我明天早上九點喝水"). Not for recurring work — use create-regular-work for that.

Activate when the user asks about team diary entries, standup summaries, recent agent activity, or when it is time to write the end-of-day diary for costaff_agent.