Autonomous Software Development Pipeline
From requirements to production-ready code — orchestrated by AI agents, governed by safety gates, monitored in real-time.
Every stage is purpose-built, from environment validation through post-run retrospective. Stages can be toggled, looped, and customized.
Each agent has a focused role, curated tools, and strict guardrails. Opus handles reasoning-heavy work; Sonnet handles execution at speed.
Reads requirements, explores the codebase, and produces detailed implementation plans with file-level specificity.
Validates plans for feasibility, completeness, and correctness before execution begins. An optional quality gate.
Decomposes plans into tracked tasks with dependencies, parallel execution groups, and clear acceptance criteria.
Claims tasks and implements code using TDD. Writes failing tests first, then minimal code to pass. Parallel-ready.
Runs the full test suite, verifies coverage thresholds, and produces a proof status that gates the review stage.
Reviews code quality, manages the PR lifecycle, and enforces human approval gates. The only agent allowed to commit.
Analyzes completed runs and produces ranked observations with actionable improvement suggestions for future pipelines.
Every feature exists because an autonomous pipeline needs it — not because it demos well.
9 stages from preflight validation to post-run learning. Each stage loops, retries, and reports independently.
Guard hooks block dangerous operations. Test gates enforce coverage. Human approval checkpoints at every critical boundary.
WebSocket-based monitoring with 52 event types. Global mode monitors all projects from one tab. Token costs, Kanban boards, and live logs.
Tasks with dependencies, parallel execution groups, and full traceability through the Beads task system.
Per-agent token tracking, circuit breakers for runaway failures, and configurable budget thresholds.
Pause mid-stage, stop gracefully, resume from exactly where you left off. Full state persistence.
Implementers write failing tests first, then minimal code to pass. Guardian validates the full suite before PR.
Agent prompt overlays, model selection per agent, stage toggling, loop limits, and per-project overrides.
Global mode monitors all registered projects from a single browser tab. Sidebar picker with live status dots and one-click project switching.
Run N pipelines concurrently, each isolated in its own git worktree. Per-pipeline pause, stop, and resume. Configurable cleanup policies.
Four rework cycles — plan review, test failures, code review, and restart planning — each with configurable limits and bounded context.
Start from a text prompt, a detailed spec, a GitHub issue, or run multiple pipelines in parallel.
worca installs as a .claude/ folder. No global dependencies.
Launch the global dashboard to manage all your projects.
cd worca-cc/.claude/worca-ui && node server/index.js --global
Click + in the sidebar to register and install worca in one step.
http://127.0.0.1:3400
Describe your work and let the agents handle it.
run_pipeline.py --prompt "..."