Workflow

A global goals engine. Fully self-hostable, open-source (MIT platform / CC0 catalog), runs on your own infrastructure. Humanity declares shared Goals — research breakthroughs, great novels, successful prosecutions, cures, open datasets, whatever people actually want done — and a legion of diverse AI-augmented workflows pursues each Goal in parallel. Branches evolve, cross-pollinate, and get ranked by how far their outputs advance up each Goal's real-world outcome-gate ladder. The system is built for whatever people collectively care about next.

This repo contains substantial architecture and implementation work for TinyAssets. The starter surfaces below help you navigate, extend, and connect — including via Obsidian if you use it.

Built by Jonathan Farnsworth ([email protected], GitHub @Jonnyton) — sole human author; the only co-authors are the project's own AI agents.

Proof of life

The engine runs on its own infrastructure. The volatile facts below are linked to live state rather than copied here, so this section can't go stale:

• Canary-gated deploys, live receipts. The current deploy SHA, canary status, queue throughput, and the provider list are returned live by the get_status MCP tool and rendered at tinyassets.io/fine-print — read the numbers there rather than trusting a copy here.
• 7,925 tests across 418 files, all offline. Providers are mocked (_FORCE_MOCK=True); no API keys: pip install -e .[dev] && pytest -q.

Honest caveat (the site says this too): the user-facing outcome loop hasn't shipped a real external artifact yet — draft mode is on, OAuth is unwired, run_count is 0. What's proven today is the engine and the architecture; the first shipped real-world outcome is the next milestone.

_{Repo facts refreshed 2026-06-25 by scripts/gen_discoverability.py (bounded — rewrites only between the markers).}

The flagship: the Loop

A user's chatbot hits a capability gap, files it as a patch request, and the system routes the work through branch, evidence-gate, and deploy surfaces so the next summon starts smarter. The retired CI writer loop has been removed; the active direction is the general Loop substrate described in PLAN.md and the branch/gate implementation under tinyassets/.

See the code (one click from here)

The entry path should reach functions, not just docs. Representative core:

• The MCP surface every chatbot connects to — tinyassets/universe_server.py (the universe / extensions / goals / gates / wiki / get_status tools).
• The daemon run loop — fantasy_daemon/__main__.py, the current default runtime (LangGraph universe graph, SQLite checkpointer, pause/resume). The branch-execution substrate is goal-agnostic — branch specs compile to graphs via tinyassets/graph_compiler.py — though this domain is still the hardcoded default; extracting the runtime into each universe's soul-declared loop is tracked in the de-fantasy audit.
• Branch spec → executable graph — tinyassets/graph_compiler.py (compiles a declarative branch into a runnable StateGraph; approval-gated node execution).
• The evaluation/gate primitive — tinyassets/node_eval.py.

What's strongest here

A coherent, dependency-verified stack (LangGraph / FastMCP / LanceDB / igraph / clingo) wired into a single self-patching engine; design philosophy with teeth (minimal primitives, fork-over-build, commons-first privacy); operational seriousness (canary-gated deploys, deploy receipts tied to source SHA, ~7,800 offline tests); and a system honest enough to file bugs against itself and state in public what it hasn't shipped yet.

Quick Start (for contributors)

Clone-to-green-tests in ~5 minutes on a clean machine:

git clone https://github.com/Jonnyton/TinyAssets.git
cd TinyAssets
python -m venv .venv && source .venv/bin/activate  # Windows: .venv\Scripts\activate
pip install -e .[dev]
pytest -q                 # full suite — no API keys needed (tests mock providers)
ruff check                # lints clean on a fresh clone

All tests run offline with _FORCE_MOCK=True set in tests/conftest.py. No ANTHROPIC_API_KEY or similar required for CI or local dev. If any test fails on a clean clone, file an issue — that's a TEST-1 regression.

Cross-platform notes:

• Tested on Windows, macOS, Linux. Paths use pathlib.Path — backslashes don't leak into tests.
• Python 3.11+ required (see pyproject.toml).
• The tray (tinyassets_tray.py) is Windows-first; macOS/Linux support is work-in-progress. Platform code is cross-platform.

Start Here

1. Read STATUS.md for live state.
2. Read PLAN.md for architecture and design intent.
3. Read AGENTS.md for process rules.
4. Read docs/project-lineage.md for how TinyAssets grew out of the earlier Hex, Echoes, Fantasy Writer, and Fantasy Author work.
5. Use python scripts/docview.py for large Markdown, text, and JSON files before any raw whole-file read.
6. Capture loose user ideas in ideas/INBOX.md or with python scripts/capture_idea.py "Idea summary".

Core Hubs

• AGENTS.md: process truth.
• PLAN.md: design truth.
• STATUS.md: live-state truth.
• docs/portfolio/README.md: public project graph, lineage, and auto-maintenance standard.
• ideas/INDEX.md: idea capture, triage, and shipped ledger.
• knowledge/INDEX.md: human-readable knowledge map.

Notes

• The new knowledge/ docs complement knowledge.db; they do not replace it.
• The new docs/exec-plans/ surface complements existing planning docs like BUILD_PREP.md and RESTRUCTURE_PLAN.md; it does not invalidate them.
• The user may steer multiple live sessions across different providers at once. Durable coordination belongs in files, not only in chat.