Local-First Agentic Pipeline

Ship code.
Let agents handle the rest.

Write a ticket. Superseded delegates implementation, testing, review, and release to AI agents — running locally in isolated worktrees with full feedback loops. Monitor everything from the web dashboard.

uv run superseded
Pipeline Stages
Spec Parse ticket requirements into structured spec
Plan Generate execution plan with task breakdown
Build Implement in isolated git worktree
Verify Run tests, lint, typecheck automatically
Review AI code review with quality gates
Ship Create PR via gh, ready to merge

Built for real workflows,
not toy demos.

Every feature exists because production pipelines need it — not because it looks good on a slide.

Feedback Loops

Stages retry on failure with error context injected into re-prompts. Click Retry from the dashboard to re-run. No silent failures.

Resilience

Execution Plans

The Plan stage writes structured plan.md. Build, Verify, and Review stages consume it — full traceability from intent to code.

Transparency

Worktree Isolation

Build, Verify, Review run in isolated git worktrees. Changes merge on success, discard on failure. Your main branch stays clean.

Safety

Docker Sandboxing

Run agents in isolated containers. Same environment every time — no host contamination, no dependency conflicts. Full workspace mount with your UID.

Isolation

Human Checkpoints

Pause before critical stages. Review a diff summary, then approve or reject from the web UI. Agents don't guess — you decide what ships.

Control

Progressive Context

Agents receive context in layers: AGENTS.md → docs → ticket → artifacts → rules → skill prompt → error context. Just enough, never too much.

Intelligence

Quality Enforcement

Critical review findings loop back to Build. Non-negotiable rules injected into every prompt. Code quality is not optional.

Standards

Multi-Repo Support

Tickets can target multiple repositories. SPEC/PLAN run once. BUILD/VERIFY/REVIEW fan out per repo. Ship creates a PR per repo.

Scale

Notifications

Push alerts on pipeline completion, failure, or stage transitions via ntfy.sh. Configure per-channel priority. Never miss a ship or a failure.

Awareness

Await Input

The Spec stage can pause and ask you questions. Answer in the web UI, and the pipeline resumes automatically. Agents clarify — they don't guess.

Collaboration

Metrics Dashboard

Track success rates, retry counts, and stage durations across all pipelines. Spot bottlenecks. Available as a web dashboard or JSON API.

Visibility

Web Dashboard

Full UI to create tickets, advance stages, review agent output, configure agents, and manage repos — all from the browser. Real-time SSE updates.

Control

From ticket to PR.
One command.

Write markdown specs, run the pipeline, get a pull request. No context-switching, no copy-paste.

superseded — pipeline run
$ superseded run 42

# Starting pipeline for issue #42: Add dark mode toggle

◉ SPEC done — parsed 4 requirements
⊞ PLAN done — 6 tasks, 2 files to touch
⚙ BUILD done — worktree feat/dark-mode-42
✓ VERIFY retry — 1 test failed, re-running…
✓ VERIFY done — 24/24 tests pass
⟳ REVIEW done — no critical findings
▷ SHIP donePR #43 created

# Or pause for human approval:
⊛ HITL await — 3 files changed, approve?…
✓ approved — continuing

✓ Pipeline complete — 1m 47s
$

Architecture

Designed for local-first operation with clear separation of concerns.

TICKETS Markdown specs + YAML frontmatter PIPELINE 6-stage engine spec → plan → build verify → review → ship feedback loops · retries DOCKER SANDBOX python:3.12-slim node:20-slim uvx opencode | npx claude WORKTREES Isolated git branches feat/* verify/* merge on success AGENTS CLI subprocess adapters Claude Code OpenCode extensible adapters PULL REQUEST via gh CLI ready to merge CONTEXT LAYERS AGENTS.md → docs → ticket → artifacts → rules → skill → error context STATE SQLite cache iteration tracking WEB UI HTMX + Alpine.js HITL CHECKPOINT approve | reject

Local-first. Agent-native. Opinionated.

Markdown tickets are the source of truth. SQLite is a cache/index. Agents run as CLI subprocesses in isolated worktrees. Every harness attempt is tracked and replayable.

  • Python 3.12+
  • FastAPI
  • HTMX
  • Alpine.js
  • Tailwind CSS
  • Jinja2
  • SQLite
  • aiosqlite
  • uv
  • Git Worktrees

Stop babysitting
your pipeline.

Write the spec. Let agents do the work.