agent-skills

Use when you want a second, adversarial opinion on a pull request before merging it — spawns a child (or sibling, if you're already nested) session running the OPPOSITE agent provider (Claude <-> Codex) to review it, and has it raise concerns back to you via mail.

Use when you want a second opinion from a stronger model mid-task — before a non-obvious design decision, when stuck, or before declaring the work done — by spawning a short-lived advisor session and consulting it with send_message.

Use when the user explicitly asks to archive the current session as a board TODO for later, including /agent-skills:archive-to-todo in Claude Code. This ends the current session.

Use when deciding what to put on this session's visible checklist — the short list of work-in-progress the human sees on the Agent Wrangler board. Covers when an item is worth adding, how to phrase it, when to mark it done, and why this is never synced with your own internal planning tool.

Use when work should be linked on the Agent Wrangler board — a Jira issue or a GitHub pull request — to attach, update, or list those links. Covers which link types are supported, when to attach them, and session-vs-task scope.

Use when you get a "📬 New mail" notification pasted into your terminal, or when deciding whether to check your mailbox, read a message, or reply to a peer session. Covers read_mail vs list_mail and the no-reply-by-default norm.

Use when asked "what did I actually work on on date X" (standup notes, a weekly summary, "what did I do yesterday") and Agent Wrangler sessions are in scope — how to call the get_session_activity MCP tool, the source of truth for wrangler activity, for precise transcript-backed per-day detail instead of hand-parsing mappings.json/transcripts.

Use when a session needs to discover its own place in the Agent Wrangler hierarchy — its session id, who spawned it, whether/what it's nested under, and its task — via env vars and MCP tools, no filesystem access needed. Also use when deciding whether "parent" means board-nesting or launch lineage.

Use when you want to spin current work off into its own brand-new Agent Wrangler board session (not a sub-agent, not a fork) — how to call the spawn_session tool, write the handoff, and choose the right agent and model.

Use at the start of a session and whenever you have cross-cutting context to record — how to read and append the shared per-task memory file (AW_TASK_MEMORY) that the human and agent edit together across every repo the task touches.