Build a task pool that keeps working when your usage resets

I want to build: [describe your idea in one sentence]. Don't start yet. Interview me to pin down the goal, scope, constraints and what "done" means, asking only a few key questions at a time. When you're done, write the conclusions into the "Goal" section of PLAN.md.

Read-only: don't change any files. Based on the "Goal" section of PLAN.md, research the relevant code, docs and current state. Write the facts, constraints and open questions into the "Research" section of PLAN.md, with the source for each.

Break the goal in PLAN.md into tasks and write them into the "Tasks" section. For each task, state what it should achieve, which files it touches, what it must not do, and the done condition (one check you can run directly). Each task must fit in a single session; mark which tasks depend on which.

Check that each approach in "Tasks" is feasible, then that they work together. Then review the overall design on these points: reliable and stable, failures get noticed, testable, quality backed by evidence, simple with no duplication. Mark each "pass" or "fail" with a reason and write the result into the "Review" section. If you find problems in the earlier sections, fix them there and tell me what you changed.

Turn the contents of "Tasks" in PLAN.md into a to-do list in POOL.md, in order. Put an unchecked checkbox in front of each item and keep: what it should achieve, files involved, what it must not do, and the done condition.

Work through the unchecked items in POOL.md in order, one at a time. After each item, run its done condition: if it passes, check it off, commit, and add a line at the end of POOL.md; if it fails, fix it, and if it still fails after two tries, mark the item "blocked", note why, and move on to the next. Don't ask me questions and don't change anything outside the list. When everything is done, write a summary at the end of POOL.md: what got done, what's blocked, and why.