overnight-kit

Run a sequence of Claude Code prompts against your repo while you sleep. One commit per prompt, push, wait for CI (Vercel and/or GitHub Checks), let Claude fix a red build, open a PR — and only auto-merge if you say so. Halts are resumable with the same command.

npx overnight-kit init --plan "Add a /pricing page with the three tiers from config, a checkout button per tier, Playwright tests, and a README section."
npx overnight-kit run overnight.prompts.json --dry-run   # see exactly what would run; touches nothing
npx overnight-kit run overnight.prompts.json             # go to bed

The morning after, this is what lands in your inbox if you connect the morning digest: merged commits, CI results, spend, failed runs and the questions your agents could not answer. The kit itself is free and works without it.

What it does

Given a prompts.json (an ordered list of prompts, a branch, a base branch, a PR title), overnight run:

1. Checks out or creates branch_name from base_branch (adopts an existing branch when the ancestry checks pass; halts on a diverged branch, never force-pushes) and takes a PID lock per branch.
2. Runs each prompt with your own claude binary: claude -p [--model <model>] --permission-mode bypassPermissions --dangerously-skip-permissions "<prompt + acceptance criteria>", capturing output to ~/.overnight/logs/run-<ts>/prompt-<id>.log.
3. Classifies the outcome — commit (HEAD advanced, tree clean), noop (nothing to do), or failure (non-zero exit, empty log, a stream-error pattern such as stream interrupted / rate limit / context length exceeded, or files left uncommitted). A commit never redeems a stream error: the response was incomplete.
4. After every commit (when wait_for_ci is on): pushes, then polls vercel ls for a new deployment of the branch and gh api …/check-runs + the combined commit status for the sha. Red build → up to ci.fix_max_attempts fix prompts (Claude gets the failing logs; a fix commit re-enters the gate; a non-code cause becomes a written diagnosis and a halt). Timeouts warn and continue by default (ci.strict_timeout halts).
5. Writes state after every prompt (~/.overnight/state/<branch>.json), plus recovery/<branch>/last-success.md / last-failure.md and a recovery diff when work was left uncommitted. Every classification goes to decisions.log; every webhook attempt to notifications.log; every event to events.jsonl.
6. After the last prompt: push, gh pr create, and — only with merge_strategy or --merge — gh pr merge --auto --squash|--merge|--rebase. Sends a 🚨 notification with the PR link.

Everything is observable from your phone: Discord and/or Slack webhooks get a message per prompt (✅ committed with the commit subject, shortstat and Claude's last line; ℹ️ no-op; ⏭ skipped on resume; 🛑 halt with the reason, the last success, the response tail and the paths to the diagnostics).

Requirements

• Node 22+, git, gh (authenticated) for PRs and GitHub Checks, vercel (logged in, project linked) for the Vercel gate.
• claude on your PATH — Claude Code with your own subscription or API key. The kit ships no keys and calls no API of its own. overnight-kit is an independent open-source project and is not affiliated with or endorsed by Anthropic.
• macOS, Linux or WSL. (Unlike the bash original it does not wrap itself in caffeinate; run caffeinate -i npx overnight-kit run … on a laptop.)

prompts.json

{
  "project_path": "/abs/path/to/repo",
  "branch_name": "overnight/pricing-page",
  "base_branch": "main",
  "pr_title": "Pricing page",
  "pr_body": "Implemented overnight in 4 prompts.",
  "merge_strategy": "none",
  "default_model": "claude-opus-4-8[1m]",
  "wait_for_ci": true,
  "ci": {
    "vercel": true,
    "github_checks": true,
    "timeout_seconds": 900,
    "poll_seconds": 10,
    "skip_grace_seconds": 120,
    "fix_max_attempts": 3,
    "strict_timeout": false
  },
  "digest": { "url": "https://agentwares-digest.vercel.app/api/ingest" },
  "prompts": [
    {
      "id": "01-config",
      "label": "Tier config",
      "prompt": "…Commit when done.",
      "acceptance": ["`pnpm test` passes"]
    },
    {
      "id": "02-page",
      "label": "Pricing page",
      "prompt": "…",
      "depends_on": ["01-config"],
      "model": "claude-sonnet-5"
    }
  ]
}

• id is filename-safe and keys the resume state — do not rename completed ids. Duplicate ids are rejected before anything runs.
• acceptance is appended to the prompt text as a checklist; depends_on is informational (array order is what runs) and is validated by the prompt-maker.
• default_model / per-prompt model are passed to claude --model; omit both to use whatever your Claude Code defaults to.
• Secrets do not belong in the file: use OVERNIGHT_DISCORD_WEBHOOK_URL, OVERNIGHT_SLACK_WEBHOOK_URL, OVERNIGHT_DIGEST_TOKEN (the file's discord_webhook_url / slack_webhook_url / digest.token still work for private setups).
• claude-runner field names (wait_for_vercel, vercel_fix_max_attempts, vercel_fix_model, vercel_skip_grace_seconds) are accepted and mapped.

Prompt-maker (init / plan)

overnight init --plan "<one paragraph>" scans the repo (framework, package manager, test/lint/typecheck/build scripts, CI, CLAUDE.md/AGENTS.md) and writes overnight.prompts.json:

• with claude on PATH it asks your Claude Code for 3–12 ordered, dependency-aware tasks with acceptance criteria (validated, topologically sorted, verification commands appended);
• without it (or with --no-llm) a deterministic planner makes one task per sentence with sequential dependencies and the same repo-derived acceptance criteria.

It also installs the /overnight Claude Code skill at .claude/skills/overnight/SKILL.md (--no-skill to skip; overnight skill install to add it to any repo). overnight plan --plan "…" prints the file to stdout instead of writing it.

Flags

| Flag | Effect | | ---------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------ | | --dry-run | Walk the whole sequence: print each claude -p invocation, CI/merge settings, tree status. Executes nothing, writes no state. | | (none) | Fresh run, or implicit resume when a matching state file exists. | | --resume | Explicit resume. Skips completed prompts whose sha is reachable from HEAD, retries failed, runs pending/appended. | | --fresh | Delete the state file and run from prompt 0 (keeps the branch). | | --continue | Run the prompts on the currently checked-out branch; no branch creation, no state, no PR. | | --force-continue | Resume even when a completed prompt was edited (prefix-hash mismatch). | | --skip-current | Mark the failed prompt completed (with its recorded sha) and resume. | | --auto-stash / --auto-recover | Dirty pre-flight tree: stash it (left for review) / commit it as leftover from the previous prompt. | | --strict-preflight | Any dirty file halts, even ignored ones (lockfiles, dist/, *.log, …). | | --merge=<s> | Override merge_strategy for this run. | | --no-wait-for-ci / --no-ci-fix / --ci-fix-attempts=N / --strict-ci-timeout | CI gate overrides (--no-wait-for-vercel etc. still accepted). |

Environment: OVERNIGHT_HOME (default ~/.overnight; CLAUDE_RUNNER_HOME honoured), OVERNIGHT_DEFAULT_MODEL, OVERNIGHT_IGNORED_DIRTY_PATHS (replace) / OVERNIGHT_IGNORED_DIRTY_PATHS_EXTRA (append), OVERNIGHT_DISABLE_NOTIFY=1.

Resume rules

The state file stores the sha256 of the whole prompts array and of the completed prefix. On re-run: same JSON → resume; prompts appended at the tail → resume, the new ones run as pending; a completed prompt edited or reordered → halt (pass --fresh or --force-continue); a JSON with no id overlap → old state archived to <branch>.json.replaced-<ts> and everything runs. After a halt, last-failure.md opens with a resume strategy: plain --resume when HEAD did not advance; otherwise Option A (mark done, --skip-current), B (git reset --hard <before> then --resume, recommended unless the log shows tests/lint/build passing) or C (force-skip).

The morning digest (optional, $15/mo)

events.jsonl is the kit's structured log. Point digest.url at https://agentwares-digest.vercel.app/api/ingest with your ingest token in OVERNIGHT_DIGEST_TOKEN and each batch pushes its events when it finishes or halts (overnight digest push <events.jsonl> does it by hand). The digest also accepts session exports from Claude Code / Routines, Codex, Devin and Cursor background agents and emails one 7am summary across all of them. Summarisation runs on your Anthropic key (encrypted at rest); the fee covers hosting, history and the email.

Development

pnpm --filter overnight-kit test      # 62 offline tests: stub claude/gh/vercel in temp git repos
pnpm --filter overnight-kit build     # dist/cli.js (bin) + dist/index.js (library)

MIT. Part of agentwares; the paid products are agentcheck (monitoring for the agents you ship) and agentguard (spend caps and kill switch for the agents you run).