nightshift
v0.4.0 · MITa Claude Code plugin

Your AI software team that ships while you sleep

A Claude Code plugin that turns one terminal into a full delivery team — product manager, architect, tech lead, engineers, and QA. It reads a Jira ticket and ships the spec, plan, code, and review.

Install in 60 seconds
/plugin marketplace add whimzyLive/nightshift-ai
/plugin install sdlc@nightshift
★ Star nightshift on GitHub

Terminal preview: /auto PROJ-142 runs the full delivery pipeline — reads the ticket, drafts the spec and plan, implements across specialized agents, passes the QA gate, and opens a reviewed pull request.

zsh — acme-api · claude code
00 specialized agents·00 slash commands·install in 00 seconds·free · MIT
01 · the other 80%

You don't lose time writing code. You lose it around the code.

The AI already writes the code. What still eats the sprint is the connective tissue: turning a vague ticket into a real spec, a spec into a plan, keeping the plan honest while you implement, then reviewing the result without rubber-stamping your own work.

Coding assistants handle the middle 20%. The other 80% — the process — stays manual, or gets skipped on vibes and shipped unreviewed.

Vague ticket, no spec

"Add auth" becomes whatever you remember at 4pm.

No enforced lifecycle

Runbooks exist. Nobody follows them without enforcement.

Self-review is a rubber stamp

Reviewing your own AI output isn't review.

02 · how it works

One command runs the whole lifecycle

Point it at a ticket. It triages the work, then runs each stage in order and closes the loop back to your tracker — spec before plan, plan before code, review before merge, tests as the gate.

/specTechnical specsolutions-architect
/planOrdered plantech-lead
/implImplementationprincipal-engineer
/reviewQuality gateqa-engineer
PRShip itauto
/auto PROJ-142
StepWhat runsWhat you get
1PRD + specThe ticket becomes a written spec
2PlanAn implementation plan, reviewed before code
3ImplementationCode written to the plan, on its own branch
4QA reviewIndependent review by a different agent than the author
5PR + ticket commentA reviewed PR, with the paper trail linked back to the ticket

/auto triages by size. Stories at or under the lightweight threshold (default ≤3 points) skip the spec and plan and go straight to implementation.

Drive a stage yourself. Every stage has its own verb: /spec /plan /impl /review.

03 · your day, split in two

Review by day. Ship by night.

"Ships while you sleep" is a workflow, not a slogan. The split is simple: your day is for decisions, the night is for execution.

☀ 09:00
🌙 23:00
↑ 07:00
☀ Day

You refine

Every story gets pre-refined with clear acceptance criteria, and the complicated ones get a full spec. You read the refined stories and specs, decide what's ready, and approve them. You make the calls. You don't write the implementation.

🌙 Night

The plugin implements

/auto picks up the tickets and specs you already approved, triages each one, and routes it to the right approach — following the practices your domain agents enforce for that repo (from its project-context.md). Nothing runs on work you haven't signed off.

↑ Morning

You review

You wake up to PRs ready for review. Each one addresses acceptance criteria you already agreed to, or a spec you already approved. You're reading a result you set up, not a surprise.

You kept every decision. The plugin did the execution. Prefer to drive a stage yourself? Each one has a verb: /refine-issue /spec /impl /review.

04 · the team

A team, not a megaprompt

"Do everything" agents hallucinate across roles and leave no trail. nightshift splits the work across 11 agents, each with a tight charter, its own prompt and tools, and a clean handoff to the next. Narrow charters mean fewer hallucinations and an auditable artifact at every stage.

$ nightshift --team --tree
you (human)— every gate returns here
├──refine feature/refine-feature · /prd
│ └──product-manager— PRD.md
├──refine story/refine-issue · /stories
│ └──scrum-master— stories
├──spec/spec
│ └──solutions-architect— spec.md
├──plan/plan
│ └──tech-lead— plan.md
├──implement/impl
│ └──principal-engineer— orchestrates, one agent at a time
│ ├──database-administrator— phase 1
│ ├──platform-engineer— phase 2
│ ├──sync-engineer— phase 3STANDBY
│ ├──web-engineer— phase 4
│ └──mobile-engineer— phase 5STANDBY
└──qa / review/review · /review-fix
└──qa-engineer— review → PR
// the org, as a tree
1 human · 11 agents

⊘ marks a hard gate — each phase ends there and control returns to you. Every ✶ is an agent with a tight charter. Hover any row to read its charter, contract, and source.

whimzyLive/nightshift-ai

⊘ every phase ends at a hard gate — control returns to you before the next one runs · hover a row for its charter

Meet the whole team — charters, handoffs, org chart →

05 · why different

Why builders choose it

The lifecycle is the product

Spec → plan → implement → review, enforced by the commands and the handoff protocol. Tests are the merge gate. Review is done by a different agent than the one who wrote the code.

Generic agents, per-repo config

No hardcoding. Every project fact lives in one project-context.md. The same plugin runs a Node monorepo, a Python service, or a mobile app.

Issue-tracker native

It reads the ticket, derives the branch, plan, and PR, and comments the result back to Jira and GitHub.

Free, open, yours to fork

Built on open Claude Code primitives, MIT-licensed. Fork it, extend it, swap a role. No lock-in, no paid tier gate.

06 · control

You decide how it gets built

Most AI dev tools abstract the process away until you can't see what you're getting — you only find out once the output lands, then negotiate your way back to what you meant. nightshift runs the opposite way: it keeps the software development lifecycle in front of you, with a hard gate at every phase that returns control before the next step runs.

// triage — every ticket takes the right-sized path
PROJ-142
estimate8 pts
1drag the estimate13
# .claude/project/project-context.md
lightweight_threshold:3
approval_mode:
assisted — every gate waits for your approval
full ceremony
story > threshold
/spec → /plan → /impl → /review
PROJ-142 · 8 pts
lightweight
story ≤ threshold — skips spec + plan
/impl → /review
defect
bugs never spec or plan
/impl → /review
PROJ-142 story · 8 pts > 3 → full ceremony — spec before plan, plan before code
// feel a gate — approve each phase yourself
refine
spec
plan
implement
review
// the same ticket, both ways — routed by the config above
one-shot AI
> build PROJ-142
🌙 nightshift · full ceremony
# route: full ceremony — one verb per phase, per project-context.md
> /refine-issue PROJ-142
scrum-master → refining PROJ-142 — story shape, binary acceptance criteria …
▌ working
Why an SDLC beats one-shot AI →
07 · questions

Questions builders ask first

No. It's a process engine. The wedge is the SDLC, not the code. Assistants do the middle 20%; nightshift runs the other 80% — spec, plan, review, and the handoffs between them.

/auto triages by size. Stories at or under the lightweight threshold (default ≤3 points) skip the spec and plan and go straight to implementation.

The agents are fully generic. Every project-specific fact lives in one project-context.md file. The same plugin runs across a Node monorepo, a Python service, or a mobile app — you configure it once per repo.

Review is done by a different agent than the author, tests are the merge gate, and every stage leaves an artifact you can read — PRD, spec, plan, review. Nothing merges on vibes.

Nothing. It's free and MIT-licensed. Adoption is the only metric we track.

Browse the full FAQ — workflow, setup, trust, licensing →

08 · install in 60 seconds

Put a ticket in tonight. Read a reviewed PR in the morning.

Install takes about a minute. Point it at a ticket and watch the spec, plan, code, and review land — with the paper trail linked back where your team already works.

/plugin marketplace add whimzyLive/nightshift-ai
/plugin install sdlc@nightshift