Put your idle AI hardware to work while you sleep.
Here's the problem Night Shift solves: you already own AI hardware — a MacBook with unified memory, a gaming GPU, a spare desktop — and every night it sits idle. Those are free tokens. The hardware is paid for; the electricity costs cents. Night Shift collects them: overnight, your machines read your repo, find small safe work, and draft. You wake up to a short ranked brief with the few things worth looking at first.
It never pushes directly to your branches, merges, releases, or touches credentials. After explicit consent, it may open one tested draft PR from an isolated branch. Drafts, not deploys. Free and open source under the MIT License.
The normal overnight mode does real work. It can strengthen unit and E2E tests, repair stale docs, implement a narrow grounded issue, clean up one exact redundancy, or repair a reproduced test failure. Models return diffs only. The controller decides the allowed files and patch size before a model runs.
Execution happens in disposable copies with a pinned rootless runner, no network, no credentials, and bounded resources. An in-repo profile cannot authorize itself; ownership-bound approval lives in Night Shift's private config. Every baseline runs twice. Finished patches pass twice, or three times for E2E work. Unsafe repos fall back to analysis instead of weakening checks.
Portfolio brief — status: GREEN
BetterFeedback
PROVEN_REPAIR: focused route regression test
Verify: npm test -- route.test.ts
Draft PR: https://github.com/example/BetterFeedback/pull/42
Transcripted
CANDIDATE: recent import path needs deterministic proof
Evidence: Sources/Import.swift:84 | exact source line
Night Shift
INDEXED: CLI tests mapped to their command handlers
Repositories visited: 3 · New unique tasks: 18 · Repeated tasks skipped: 42
Draft PRs opened: 1 · Verified local patches: 1 · Nothing merged or deployed.
Model findings are candidates, never proof. A failing-before/passing-after fix
becomes PROVEN_REPAIR; a bounded patch whose checks pass both before and after
is a VERIFIED_DRAFT. Repeated tasks are remembered across nights and skipped
until the code or GitHub signal changes.
git clone https://github.com/r3dbars/nightshift.git
cd nightshift
./install.sh
~/.codex/bin/night-shift start # works immediately, even before opening a new shellNormal setup detects the current project, GitHub, local AI, and a configured
LAN worker, then asks one question: start the safe eight-hour plan? Use
night-shift start --advanced only when you want to customize the defaults.
Next morning:
night-shift report --latest
night-shift handoff --latest --item 1 # prepare the best item for independent reviewThe handoff stays local by default. Sending it to Codex requires cloud consent
from setup or a one-time --allow-cloud, and the review runs read-only.
Preview output shows the exact committed-file count, byte count, redaction count,
and privacy result before anything can be sent.
At any point, get the small operational readout instead of hunting through ledgers:
night-shift healthIt shows whether the controller is live, whether both AI lanes answer, whether the selected repo is analysis-only or sandbox-ready, the latest outcome totals, and how much local ledger storage Night Shift is using.
To preview or repair one repository's saved execution approval:
night-shift trust-repo --repo /path/to/project # preview only
night-shift trust-repo --repo /path/to/project --apply # one consent, saved outside the repoNormal first-run setup does this automatically after the one hands-on consent. It proves GitHub ownership, detects an exact verification command, builds the pinned runner, and checks the command twice before saving approval outside the repo. Missing tooling keeps that repo analysis-only.
Every selected task also has a durable lifecycle: DISCOVERED, REPRODUCED,
DIAGNOSED, PATCHED, VERIFIED, then human-only REVIEWED and PROMOTED.
An overnight run cannot skip from a hunch to a patch.
Old run artifacts stay until you review them. Preview safe reclamation with
night-shift clean; only night-shift clean --apply removes completed,
reviewed ledgers older than 21 days.
Works with: LM Studio, Ollama (auto-detected), or any OpenAI-compatible local model server · a second GPU box on your LAN as a heavy draft lane · optionally the Claude CLI for one or two hard questions a night and the GitHub CLI for open-PR context.
Night Shift starts generic scans on your Mac's local model. It only routes a task to the second machine when a pinned failed CI run or active PR signal gives that heavier lane a concrete problem to solve.
If a model keeps producing unsupported findings for one unchanged repo revision, Night Shift opens a small rejection circuit breaker and switches back to the factual brief rather than burning the rest of the night on retries.
No local models yet? night-shift start still works: it makes a read-only
planning brief and tells you exactly what to set up.
You shouldn't have to remember Night Shift exists. Arm it once and your hardware clocks in every night by itself:
night-shift schedule --nightly 23:30 # runs every night with your saved setup
night-shift schedule --status # when it runs, what happened, how to stop
night-shift snooze --days 7 # vacation switchThe standing shift looks after itself: it pauses when three morning briefs
pile up unread (no zombie automation making reports nobody reads — reading
one resumes it), drops to quiet mode on battery, and turns off with one
command. Optionally, night-shift deliver --latest --github-issue keeps one
digest issue updated. Hands-on mode may also open bounded tested draft PRs; it
never merges them. The full design:
docs/autopilot.md.
+--------------+ +----------------------+ +---------------+
| Your repo | --> | Night Shift | --> | Morning brief |
| Your compute | --> | local / GPU / cloud | --> | KEEP / MAYBE |
+--------------+ +----------------------+ +---------------+
- Scan live signals: recent diffs, tests, TODOs, issues, PRs, and failed workflows.
- Rank recently active GitHub repos and start with repair or unfinished work.
- Work down the ladder: Repair, Finish, Strengthen, Understand, then Index.
- Ground every task with numbered source, relevant diffs, and real repo commands.
Failed CI work pins the exact GitHub
headSha, including its files and package scripts. - Remember task fingerprints across nights so unchanged work never repeats.
- Draft safely in disposable worktrees using a task-specific policy.
- Prove baselines and finished patches repeatedly in the no-network sandbox.
- Stay on duty until morning, rescanning GitHub only after new work is exhausted.
Teach it what matters after you review a choice:
night-shift feedback --latest --item 1 --useful
night-shift feedback --latest --item 2 --not-useful --note "too generic"Feedback stays on your machine and shapes later prompts and rankings for that repo.
You choose how much it may prepare:
| Autonomy | What you get |
|---|---|
brief |
read-only repo scan, ranked work queue, morning brief |
draft-local |
+ exact patch plans, issue candidates, test ideas, and isolated tested drafts when enabled |
draft-prs (default) |
+ tested GitHub draft PRs after one saved consent; never merge |
And how hard it runs:
| Mode | Use it for | Rough shape |
|---|---|---|
quiet |
battery, small repos, short evenings | one repo, small unique batches, low heat |
night-shift |
the normal overnight run | up to three active repos, finish-first work |
afterburner |
maximizing idle hardware | deeper unique indexing and draft work across more repos |
- Push to your existing branches or merge PRs from an overnight run.
- Release, deploy, publish, tag, or notarize.
- Touch credentials, billing, or repository visibility.
- Move or delete your files.
- Pretend an unverified draft is the truth.
With saved hands-on consent, Night Shift may push one unique isolated branch per repo and open a draft PR, capped at three PRs per shift. It skips hosted CI for autonomous branches and leaves every PR unmerged for human or cloud-agent review. The full boundary lives in SAFETY.md.
- Autopilot design — how the standing nightly run stays trustworthy: attention-aware pausing, battery awareness, snooze, and opt-in morning delivery.
- User guide — the setup wizard walkthrough, every file a run produces, advanced recipes, and full mode details.
- 20 use cases — from "solo Mac with LM Studio" to "messy PR queue" to full tokenmaxx nights.
- Copy-paste examples — including a fake sample morning brief.
- Safety and privacy — what each worker lane can see and why the boundaries exist.
- Troubleshooting · Contributing · Changelog
The Night Shift helper: tiny, caffeinated, and only allowed to make drafts until a human checks the work.
MIT © r3dbars · LICENSE