Unattended runner for MIP task lists. Works one ready task at a time, opens a Draft PR for each, and stops cleanly at a usage limit or a real blocker. Only when typed by the owner.
# mip-overnight
This skill works through `docs/MIPs/MIP-NNNN.tasks.md` files while the owner is away. It never
merges. The owner reviews in the morning.
## Before the run
Make sure that `.claude/settings.json` has `permissions.deny` for `Bash(gh pr merge*)` and
`Bash(gh pr close*)`. The deny rule holds for the whole run. If it is missing, stop. Tell the
owner to add it.
## Arguments
`$ARGUMENTS`: one or more MIP numbers, in order. With no number, select the easiest ready MIP:
1. Candidates are MIPs with a task issue that has the `agent-ready` label.
2. Remove each candidate that a missing earlier phase blocks.
3. Sort by effort (S, M, L, XL). Prefer the verdict "cheap win". Then the lowest number.
4. If the MIP has no tasks file, run the `mip-tasks` steps first.
Say the selected MIP and the reason in the first line and in each PR body.
## Goal
Each row of each tasks file has an open Draft PR, or a log line that says it is blocked. At the
end, print `GOAL_COMPLETE: <summary>` or `GOAL_FAILED: <reason>`.
## Rules
- Work a task only when its issue has `agent-ready`. Otherwise log it as "not ready".
- One task is one branch and one Draft PR, stacked as in `mip-tasks`.
- Before each commit, the full gate is green. Each commit has the `Tested:` trailer.
- Never merge or close a PR. Never push to `main`. Never use `--force` without `-with-lease`.
- Never start a paid resource.
- Two failures on one task: log it as blocked and go to the next task.
- If `gh` has no login: push the branch, and write the exact `gh pr create` command into
`GH_POST_MORTEM.md` (ignored by git). Then continue.
## Effort
- A mechanical task (one small file, configuration) starts at low effort.
- A real task starts at the default effort.
- After one failure, raise the effort. Change the model only after two more failures, and only
if the usage budget permits it. Go back to the default after that task.
- Write in each PR body which effort and model did the task.
## Usage guard
Before each task, read the usage of the five-hour window and of the weekly window. For each
window, project the usage at the end of the window from the current rate.
- Projection below 75% (five-hour) or 85% (weekly): continue.
- Above that: finish the current task, then stop.
- Above 90% (five-hour) or 95% (weekly): stop now, after the checkpoint below.
If no usage data is available, use a dollar budget that the owner gave at the start.
## Checkpoint
Before a stop, the current task is in one of two states only:
1. **Done**: green, committed, pushed, and its PR open (or its command logged).
2. **Abandoned**: the work tree is clean again, with one line in the report that says why.
Never stop with a half-built tree and no note.
h0ffmann/skills