ina

ina — Infinite Angel

한국어

Daemon-supervised coding angels that never stop.

Launch an agent. Go to sleep. Wake up to working code.

Quick Start • Skills • Pipeline • Acknowledgements

---

Quick Start

# 1. Install (binaries + auto-configuration)
curl -sSL https://raw.githubusercontent.com/jinto/infinite-angel/main/install.sh | sh
source ~/.zshrc  # or open a new terminal

# 2. Register daemon (auto-start on login)
ina install

# 3. Install skills (in Claude Code)
/plugin marketplace add https://github.com/jinto/infinite-angel
/plugin install ina

# 4. Example
/ina:autopilot Implement a login feature.

# Uninstall (removes daemon, hooks, HUD, MCP server)
ina uninstall          # keep config (~/.ina)
ina uninstall --purge  # remove everything

---

What makes ina different

Other Claude Code plugins are prompt libraries. ina is infrastructure.

	Prompt libraries	ina
Agent crashes	Gone forever	Daemon restarts + resumes
Context resets	Start over	Pipeline state preserved
Long-running tasks	Hope for the best	Monitored + alerts
Multiple agents	Manual	Registry + coordination

---

Skills

  EXPLORE       THINK         PLAN          BUILD (impl→review→commit)       SHIP
 ┌────────┐   ┌──────┐     ┌──────┐     ┌──────────────────────────┐     ┌──────┐
 │ Build  │   │ Idea │     │ Plan │     │ Code → 3-lane Review     │     │  PR  │
 │ it?    │──▶│ Spec │ ──▶ │Tasks │ ──▶ │ → Fix-first → Commit     │ ──▶ │Merge │
 └────────┘   └──────┘     └──────┘     └──────────────────────────┘     └──────┘
 GO/NO-GO/
  PIVOT

Skill	Description
explore	"Should I build this?" — inline market research + GO / NO-GO / PIVOT verdict
autopilot	Full pipeline: think → plan → build
think	Idea → spec (technical / business / improve)
overhaul	Codebase health check — full scan + codex review + improvement plan
plan	Consensus planning + TDD task breakdown
build	Implement → review → commit in one shot (3-lane review built-in)
review	Standalone 3-lane review (adversarial + security + simplify)
research	Multi-angle decomposition + parallel search
design	UI implementation + visual verification
test	Test runner + failure analysis + fix
ship	Create PR with auto-generated summary
guard	Safety guardrails for unattended execution

Don't know which skill to use? Just describe what you want — ina auto-selects the right skill via the built-in reference guide.

---

Usage Scenarios

"Should I even build this?"

/ina:explore I'm thinking of a SaaS tool for async code review

Detects Startup or Builder mode → questions with inline WebSearch → competitive landscape → premise challenge → GO / NO-GO / PIVOT verdict saved to .ina/specs/.

"I have a vague idea"

/ina:think I want to add user authentication

Socratic interview → multi-perspective validation (Architect/Critic/CEO) → spec document saved to .ina/specs/.

"Is this code any good?"

/ina:overhaul

Full codebase scan + parallel codex review → "If I rebuilt this from scratch, what would I do differently?" → improvement plan saved to .ina/specs/. No code modified.

"I have a spec, need a plan"

/ina:plan .ina/specs/20260405-1000-think-auth.md

Consensus planning (Planner → Architect → Critic) → TDD task breakdown → TASKS.md.

"I have tasks, just build it"

/ina:build

Implement → 3-lane review → commit, all in one shot. Auto-delegates: direct for 1 task, subagent for 2-3, team for 4+.

"Do everything from scratch"

/ina:autopilot Add user authentication with OAuth2

Full pipeline: think → plan → build (with review + commit). Crash-recoverable via .state/pipeline.json.

"Review before commit"

/ina:review

Parallel 3-lane review (adversarial + security + simplify) + fix-first auto-correction.

"Run tests and fix failures"

/ina:test

Root cause analysis + fix + re-run (max 5 cycles).

"Create a PR"

/ina:ship

Auto-generates summary from git log/diff + runs tests before PR creation.

---

Pipeline

explore ──GO──▶ think → plan → build → review → commit
   │                              ↑         │
PIVOT/NO-GO                       └─────────┘ (loopback max 3)
   │
   ▼
 stop or
re-explore

Crash recovery via .state/pipeline.json — the daemon restarts the agent and resumes from the recorded stage.

Multi-perspective validation — think and plan invoke Architect, Critic, and CEO subagents in parallel to validate specs and plans before execution.