nightshift

The AI that works your on-call night shift — the safe kind.

nightshift watches your services, and when something breaks it diagnoses the root cause with an LLM, files one deduplicated ticket, and (optionally) drafts a fix. It never touches production on its own — every change needs a human's approval, and its entire cloud footprint is read-only.

Most "AI SRE" tools are Kubernetes-first and race toward full auto-remediation. nightshift is built for the messy reality most teams actually have — serverless & multi-account AWS, dozens of hand-deployed functions, no service mesh — and it's built to be trusted: read-only everywhere, human-in-the-loop by architecture.

   an error in your logs                     a person asks a question
            │                                          │
            ▼   push (automatic)                       ▼   pull (interactive)
   ┌──────────────────┐                       ┌──────────────────────┐
   │  Triage pipeline │                       │  Investigation agent │
   │  fingerprint →   │                       │  read logs + code →  │
   │  dedup → LLM →   │                       │  typed resolution →  │
   │  ticket + alert  │                       │  you approve         │
   └──────────────────┘                       └──────────────────────┘
            └──────────────┬─────────────────────────────┘
                           ▼
              one deduplicated, root-caused ticket
              (+ an optional PR you review and merge)

---

Why it's different

• Safe by design. Read-only in the cloud (logs:* and read metrics — nothing else). The only write in the whole system is opening a pull request, and that only happens on an approver's explicit click. A prompt-injected message can at most steer a diagnosis; worst case is a bad PR nobody merges. The allowlist and the identity-bound approval are enforced in code (core/safety.py), not prompt wording — and there's a test where a hijacked model asks to delete a table and can't.
• One bug → exactly one ticket. A two-layer dedup engine (a fingerprint that strips volatile noise + an atomic claim) means 500 copies of one crash collapse into a single ticket with a counter — not 500 pages at 3am.
• Fixes aren't always code. Resolutions are typed: code_pr, aws_infra (a copy-paste runbook), db_change (a reversible runbook), manual, or a plain answer. It proposes the right kind of fix, not always a diff.
• Provider-agnostic. Log source, LLM, ticket sink, and chat platform are all plugins. Ships with a zero-cloud demo path so you can try it before wiring anything.
• Cost-aware. The LLM is invoked only on new errors; repeats just bump a counter. Circuit breakers and per-service caps keep the bill proportional to real incidents, not traffic.

Try it in 30 seconds (no cloud, no API key)

git clone https://github.com/ranjan98/nightshift
cd nightshift
pip install -e .

nightshift demo          # the triage pipeline (what the GIF shows)
nightshift demo-agent    # the read-only investigation agent

demo feeds a sample crash log through the full pipeline — the level gate skips an INFO line, the ignore-list drops rate-limit noise, two real bugs are filed, and a replay of the same errors files nothing (dedup). demo-agent runs the tool-use agent over the same log. Both run fully offline; set ANTHROPIC_API_KEY to get real AI root-cause analysis and typed fixes.

Use it on your own logs

# triage a log file → console tickets (add creds to file real ones)
nightshift triage --logs app.log --component checkout-api

# file real GitHub issues + post Slack alerts
export GITHUB_TOKEN=ghp_... SLACK_BOT_TOKEN=xoxb-...
nightshift triage --logs app.log --github-repo you/ops --slack-channel C0XXXXXX

# ask the read-only agent to investigate (searches logs, reads source, reads a screenshot)
nightshift investigate "why are checkouts failing?" --logs app.log --src . --screenshot err.png

# render a self-contained HTML dashboard of a triage run
nightshift report --logs app.log --out report.html

See .env.example and nightshift.example.yaml for configuration, and docs/architecture.md for how the pieces fit.

The interactive Slack bot

nightshift/slack/ is a transport-agnostic Slack Events app: a person reports a problem in a channel → the agent investigates read-only → it replies in-thread with a diagnosis and, for a code_pr, an Approve button → only the configured approver's click opens a real pull request. Signature-verified, retry-deduped, and it acks in <3s then does the slow work async. Drop SlackApp.handle(body, headers, now) into Flask / FastAPI / a Lambda — no Slack SDK needed. See docs/slack.md.

Architecture

Layer	Plugin interface	Ships working	Also included
Where errors come from	LogSource	file (offline)	CloudWatch, GCP Logging, Datadog
The brain	LLMProvider	echo (offline)	Anthropic, OpenAI
Where tickets go	TicketSink	console	GitHub Issues, Jira, Linear
Where alerts land	ChatPlatform	console	Slack
Where fixes land	CodeRepo	—	GitHub PR (diff → blobs → tree → commit → PR)
Agent tools	read-only callables	LocalToolkit (logs + source)	cross-account CloudWatch
Dedup state	DedupStore	in-memory	DynamoDB atomic conditional put

The core (nightshift/core/) is provider-agnostic and dependency-free: fingerprint.py, dedup.py, safety.py, triage.py, agent.py, patch.py (a strict unified-diff engine that refuses rather than mis-applies). Every integration is stdlib-only where it can be (GitHub, Slack, Jira, Linear, Datadog all use urllib — no SDK); the heavier ones are optional extras (nightshift[anthropic,openai,aws,gcp]).

Roadmap — all shipped ✅

• Triage pipeline — fingerprint → two-layer dedup → LLM root-cause → ticket + alert, with level gate, ignore-list, and circuit breakers.
• Investigation agent — read-only tool loop, typed resolutions, identity-bound approve(), screenshot vision, local toolkit + CLI.
• Slack Events surface — report → diagnosis in-thread → Approve → real PR (signature-verified, retry-deduped, async).
• Self-healing PRs — a strict unified-diff engine + GitHub git-data-API PR opener; refuses to guess, falls back to posting the diff for manual apply.
• Providers — logs: CloudWatch · GCP · Datadog · file; brains: Anthropic · OpenAI · echo; tickets: GitHub · Jira · Linear · console; dedup: DynamoDB · memory.
• Reporting — a self-contained HTML dashboard (nightshift report).

Ideas and PRs for new providers welcome — each is one small class behind an interface.

Security

Read docs/security.md before deploying. The short version: the design's whole point is that the AI cannot take a destructive action — see the safety module, which enforces the read-only tool allowlist and the identity-bound approval in code, not just convention.

Development

pip install -e ".[dev]"
pytest                                   # 19 tests, <1s
python scripts/make_demo_gif.py          # regenerate assets/demo.gif (needs pillow)

License

MIT — see LICENSE.

---

Built by @ranjan98. Not affiliated with any employer; a clean-room implementation of ideas I've found useful for keeping serverless fleets healthy.