Issue Autopilot

Take a folder of local Markdown issue files (the output of prd-to-issues) and drive every issue that can be done to done/, autonomously and AFK. For each issue you run a full quality loop — implement → review → fix → QA → simulate → fix — and you keep going, in dependency order, until the folder is exhausted or only human-gated work remains.

Delegate aggressively to subagents. The main thread is the orchestrator: it builds the plan, picks the next issue, spawns subagents for the expensive work (implementation, review, fixing, QA, QA-simulation), reads their structured results, and decides what happens next. The main thread does as little file-editing itself as possible.

AFK contract (non-negotiable)

This skill runs unattended. Therefore:

• Never call AskUserQuestion and never block waiting on the user. When a real decision is needed, make the best defensible choice, record it, and continue.
• Never fabricate a human gate decision. Issues that are inherently HITL (a go/no-go gate, an architectural sign-off, a design approval) cannot be "decided" by the autopilot. Produce the deliverable the issue asks for (e.g. write the memo with a recommendation), but if downstream issues are blocked by a human decision (e.g. "only if go"), stop that branch, flag it, and move to other ready work.
• Fail forward. One issue blocked or failing must not abort the whole run. Skip it with a recorded reason and continue with the rest of the ready set.
• Track progress with TaskCreate/TaskUpdate so the user can see live state when they return.

Inputs

• Argument (optional): a path to an issue folder, e.g. docs/issues/<change>/.
• Default if no argument: docs/issues/consumer-deepseek-migration/.