v1.0.0 - Parallel MCP orchestration engine

AppForge v1.0.0 — a parallel, MCP-coordinated multi-agent orchestration engine. A real MCP state server and a pool of independent OS worker processes drive a product idea through a six-phase dependency graph (Clarify → Design → Code → Test → Deploy → Iterate), with human approval gates and automatic budget-driven model downgrade.

What is in it

• MCP state server (backend/engine/state_server.py) — FastMCP over streamable-HTTP wrapping a single-writer SQLite (WAL) store. It is the only writer of run state and exposes coordination as MCP tools.
• Independent worker processes (backend/engine/worker.py) — separate OS processes claiming ready tasks via a guarded single-statement claim plus versioned CAS. Lease/heartbeat and a reaper give at-least-once execution with exactly-once effect.
• Six-phase task DAG (config/phases.yaml) with a pure scheduler resolving readiness, gates and budget.
• 16 agents (config/agents.yaml) — deterministic mock mode by default, real-Anthropic mode behind one env flag.
• Approval gates after Clarify and Design; rejection re-opens the phase.
• Budget enforcement with live claim-time downgrade — 85% downgrades, 95% pauses for acknowledgement, 100% hard-stops.
• Live web UI — Socket.IO bridge to a React + React Flow graph view.

Evidence, not assertion

Every headline claim has a runnable proof test (see the table in the README), including test_concurrency_no_collision.py, in which 8 real worker subprocesses drain 12 contended tasks with each executed exactly once. A fully documented run is committed at docs/runs/2026-07-24/.

Verified at this commit

All five CI jobs green on d24ccab: backend (3.11), backend (3.12), frontend, e2e, validate-config.

• Backend suite 156 passed, coverage 86.99% (gate 70%)
• Frontend 28 passed
• Browser e2e 4 passed (Playwright, chromium)
• ruff and black clean

Smoke-tested from a genuine fresh clone of this tag: git clone && uv sync && uv run appforge run "..." drives all six phases to complete and exits 0.

Known gaps, stated plainly

• Simulated token cost. The downgrade mechanism is real and proven; the dollar figures driving it are placeholders, not real Anthropic usage metadata.
• Mock by design. The suite and the documented run use deterministic mock agents. AppForge is an orchestration architecture, not a finished app generator.
• Single-user web bridge. The live UI targets local single-user use; multi-tenant lifecycle hardening is unbuilt.
• Shutdown noise. A CancelledError traceback prints on CLI teardown. Cosmetic — the run reports done and exits 0.
• Historical docs. docs/Roadmap.md, docs/CoreDesignDocument.md and the dated Status-*.md files describe the retired LangGraph-era design and are kept as history only. docs/STATUS.md is canonical.

MIT licensed.