handlealltodos

name	handlealltodos
description	Autonomously work through every outstanding TODO / action item in the session until done, then sweep the conversation for anything missed. Use when the user says "/handlealltodos", "handle all todos", "finish the backlog", "clear all action items". Wraps the /goal loop with a chat-sweep + escalate-to-AI rule.

handlealltodos

Standing objective: handle ALL outstanding action items, then sweep the conversation for anything missed, escalating to external AIs when unsure.

When invoked

1. Assemble the action-item list. Gather from:

   • the current TodoWrite list,
   • the conversation — any user ask not yet finished,
   • open items in reports/ / docs/ produced this session,
   • TODO / FIXME / "next steps" the session surfaced. Write the consolidated list into TodoWrite.

2. Invoke the goal skill (via the Skill tool, skill name goal) with an objective of this shape:

   "Handle every outstanding action item: . After each item, re-scan the chat for remaining work. If a step is unclear or a decision is ambiguous, get a second/third opinion before proceeding. Stop only when all items are done + verified, or a real blocker needs the user."

3. Work the /goal loop. Each turn = real edits / commands / commits, then tick + judge, auto-continue until done.

Rules

• A turn must produce real progress, not a status narration. The judge evaluates work.
• Verify before done — tests pass, py_compile clean, file written, command output confirmed. "Probably fine" is done=false.
• Respect repo safety (CLAUDE.md): branch before committing, py_compile not live generators, never stomp peer work, pull before push.
• Irreversible / outward-facing actions (deploy, force-push, delete, send) — confirm with the user first; do not guess.
• If the repo is in heavy peer storm-commit churn, work in an isolated git worktree.

Escalation ladder — when a step is unclear

Do NOT guess. Climb the ladder:

1. /swarm-second-opinion — 3-engine consensus (deepseek + xai + kilo).
2. /consult-grok, /consult-gemini, /consult-codex, /consult-cursor-agent, /consult-opencode, /consult-kilo — single-model opinions; ask 2-3 and compare.
3. Still unclear, or the decision is irreversible → ask the user.

Record which opinion was used and why in the relevant report/commit.

Done condition

All assembled action items are completed AND verified, and a final chat re-scan surfaces nothing new. Then report the closed list with evidence per item.