yuroka-studio/nightshift-agents
Scheduled Claude Code agents that take one task from a backlog, open a PR with real test evidence, and never merge or deploy on their own.
ARCHETYPE
Fort Knox
Deny lists, pre-tool hooks, sandboxing. Nothing touches prod without a signature.
GUARDRAILS
4/5
Blocks destructive commands · Protects secrets · Pre-tool screening hook · No YOLO mode · details
Quality 61/100 Guardrails v2
Copy this rig
# review before running: this installs third-party code
$ npx degit yuroka-studio/nightshift-agents/.claude ./rig-nightshift-agents # inspect, then merge into .claude/ MCP servers are added to Claude Code at local scope; env vars are shown as YOUR_… placeholders — we never store values. Files are fetched with degit into a separate folder so you can review before merging.
$ npx degit yuroka-studio/nightshift-agents/.claude/skills/deliver-task .claude/skills/deliver-task $ curl -fsSL --create-dirs -o .claude/agents/architect.md https://raw.githubusercontent.com/yuroka-studio/nightshift-agents/main/.claude/agents/architect.md $ curl -fsSL --create-dirs -o .claude/agents/implementer.md https://raw.githubusercontent.com/yuroka-studio/nightshift-agents/main/.claude/agents/implementer.md $ curl -fsSL --create-dirs -o .claude/agents/reviewer.md https://raw.githubusercontent.com/yuroka-studio/nightshift-agents/main/.claude/agents/reviewer.md
Merge into .claude/settings.json (project) or ~/.claude/settings.json (user). Hook commands reference scripts in the source repo — copy those too.
{
"permissions": {
"deny": [
"Read(./.env)",
"Read(./.env.*)",
"Bash(git push --force:*)",
"Bash(git push -f:*)",
"Bash(gh repo delete:*)"
]
},
"hooks": {
"PreToolUse": [
{
"matcher": "Bash",
"hooks": [
{
"type": "command",
"command": "node \"${CLAUDE_PROJECT_DIR}/scripts/hooks/guard-bash.mjs\""
}
]
}
]
}
} Skills (1)
Subagents (3)
| architect model: opus | Read-only design advisor for genuinely hard decisions - cross-system architecture, tricky debugging, risky migrations. Expensive; do not use for routine tasks. |
| implementer model: sonnet | Implements exactly one factory task on its claude/* branch inside the files it owns, runs the project's tests, and reports real results. Use for routine feature and bugfix work. |
| reviewer model: sonnet | Independent read-only code reviewer for a factory task's diff. Checks correctness, security and regression risk against the task's acceptance criteria. Use before opening a PR for any change over a fe |
Hooks (1)
| event | matcher | runs |
|---|---|---|
| PreToolUse | Bash | node "${CLAUDE_PROJECT_DIR}/scripts/hooks/guard-bash.mjs" |
Permissions
deny (5)
Read(./.env)
Read(./.env.*)
Bash(git push --force:*)
Bash(git push -f:*)
Bash(gh repo delete:*)
ask (0)
—
allow (11)
Bash(node bin/factory.mjs:*)
Bash(npm test)
Bash(git status:*)
Bash(git diff:*)
Bash(git log:*)
Bash(git fetch:*)
Bash(git add:*)
Bash(git commit:*)
Bash(git checkout -b claude/:*)
Bash(git push -u origin claude/:*)
Bash(git push origin claude/:*)
Similar rigs
Saba-Var/claude-ward
A read-only tripwire that watches Claude Code's local config and reports suspicious changes.
Pragmatist 1.8k tok · · Q21
yopazerbot/brieltje
Smetlede and Briel (Lede, Belgium) in 3D: a walkable village in the browser, alone or with up to 16 friends. Explore, Paintball, Action, and endless waves of Zombies. Three.js and Blender, made with Claude Code and OpenAI Codex..
Pragmatist 4.7k tok · · Q8
Trolzie/claude-loop-systems-demo
A small demo of Markdown-as-system AI dev loops with Claude Code: issues → agents → skills → verification.
Pragmatist 480 tok · · Q15
homeofe/improvements
Portable, LLM-agnostic framework for multi-model AI workflows: model routing, structured agent handoff (AAHP), and a phased agent pipeline. MIT.
Orchestrator 1.4k tok · · Q24