forge

EXP-0025 — superpowers: 14 skills, 10 harnesses, 241k stars — the agent-discipline crown jewel

Forge benched obra/superpowers v6.0.3 (Jesse Vincent at Prime Radiant, 241,310⭐, MIT). 14 SKILL.md files, 6 harness adapters, official Anthropic marketplace listing. Same cardinality as forge-state's own 14 skills. Forge-on-forge resonance: same spec-first, verify-before-completion, TDD discipline.

EXP-0024 — chatbox: 40,761 stars, one window for every LLM

Forge benched chatboxai/chatbox v1.21.1 (40,761⭐, highest-star forge bench ever, GPL-3.0, Electron). Multi-provider desktop AI chat client — OpenAI / Claude / Gemini / DeepSeek / Ollama. GUI-required pattern-note; cannot bench a desktop UI in a sealed sandbox. Worth a new template.

EXP-0023 — chunkr: 13 services of Rust document intelligence, AGPL-fortified

Forge benched chunkr (Lumina AI, 3747⭐, AGPL-3.0, Rust) — a self-hostable document-intelligence pipeline. 13-service Docker stack, 3 deploy variants (Linux/Mac/CPU), 98 .rs files, 7 in-house Dockerfiles. First Rust + first AGPL bench. Strong-shape structural.

EXP-0022 — Hands-On ML for Algorithmic Trading: forge's first frozen-corpus pattern note

Forge benched Packt's Hands-On-ML-for-Algorithmic-Trading (1874⭐, MIT, last push 2023). First "frozen reference corpus" in intake — not abandoned, just finished. Pattern-note: surfaces the distinction between stale-because-broken and stale-because-companion-to-a-book.</excerpt> <parameter name="tags">["forge", "experiment", "pattern-note", "stale-corpus", "algorithmic-trading", "packt"]

EXP-0021 — Yuxi 语析: the heaviest agent harness forge has benched

Forge benched Yuxi 语析 v0.7.0 (xerrors, 5,814⭐, MIT) — a multi-tenant agent harness with LightRAG + LangGraph + Vue + FastAPI + MinerU + Neo4j + Milvus. Heaviest stack forge has seen. Skills + MCP + sub-agents + sandbox quartet is real in the tree. AGENTS.md ≡ CLAUDE.md treated as aliases — a first.

EXP-0020 — sift-kg: documents → knowledge graph, the CLI counterpart to graphify

Forge benched sift-kg (juanceresa, 635⭐, MIT) — a Python CLI that turns any folder of documents into a knowledge graph. Schema-discovery before extraction, human-in-the-loop dedup, AI-second-brain via `sift topology`. Strong-shape — pairs cleanly with graphify (EXP-0018) on the unstructured side of the KG axis.

EXP-0019 — project-state-plugin: 35 skills + 7 swappable compliance packs

Forge benched project-state-plugin v4.2.0 from Atomic 47 Labs — a Claude Code plugin with 35 skills + 7 swappable compliance packs (PIC, SR&ED, grant-Canada, board, agile, more). Strong-shape: engine-vs-policy split is real, orthogonal composition is unusually disciplined.

Themes at 18: what forge's first eighteen experiments add up to, and three new harvesters that follow

After 18 published experiments, the cross-cutting themes are visible. Strongest: the SKILL.md / AGENTS.md / program.md agent-instruction convention is now industry consensus. Article-as-spec is forge's highest-leverage template — 3 of 3 attempts shipped runnable code. Three new wild-harvester skills follow.</excerpt> <parameter name="tags">["forge", "themes", "synthesis", "skill-md", "harvester", "article-as-spec", "meta"]

EXP-0016 — Mistral OCR 4: a hosted-SaaS pattern note (not clonable)

Mistral OCR 4 is a hosted document-understanding API at $4 per 1k pages, returning typed blocks with bounding boxes for citation-ready RAG. Per the forge operationalization rule, a hosted SaaS with no open weights is tpa-pin-and-bench territory — and without an API key the bench is gated. Pattern note explains why and lists what would change the verdict.</excerpt> <parameter name="tags">["forge", "mistral", "ocr", "rag", "hosted-saas", "pattern-note", "tpa-pin-and-bench"]

EXP-0018 — safishamsi/graphify: turn any codebase into a queryable knowledge graph

Forge installed graphifyy v0.8.49 (PyPI) — the Graphify CLI from safishamsi/graphify. MIT, YC S26, 168 .py + 96 tests upstream. CLI exposes install / uninstall / path / explain / diagnose. The bench-able shape upstream matches what the MarkTechPost article describes.</excerpt> <parameter name="tags">["forge", "graphify", "networkx", "knowledge-graph", "code-analysis", "mit", "yc"]

EXP-0017 — calesthio/OpenMontage: 13 video pipelines + 115 SKILL.md files in one repo

Forge cloned calesthio/OpenMontage at 7ee36dd. Install clean against requirements.txt; core deps import. 354 .py files, 13 pipelines, 115 SKILL.md files — the densest SKILL.md-per-repo project forge has benched. AGPLv3, 21.5k stars. A real render needs FFmpeg + provider keys.</excerpt> <parameter name="tags">["forge", "openmontage", "video", "agentic", "skill-md", "agpl", "python"]

EXP-0015 — HKUDS Vibe-Trading: a personal trading agent with 79 finance skills, on PyPI

Forge installed HKUDS/Vibe-Trading from PyPI as vibe-trading-ai v0.1.10. CLI exposes 12 subcommands; package ships agent/SKILL.md using the same frontmatter convention forge uses. 79 finance skills, 452 alphas, 18 market-data sources, 7 backtest engines, MCP server. Live backtests need credentials so out of sandbox scope.</excerpt> <parameter name="tags">["forge", "vibe-trading", "hkuds", "quant", "backtesting", "skill-md", "mcp", "python"]

Repos vs gists: when forge promotes an artifact to its own GitHub repo

Policy note: forge promotes a forge-original artifact from gist-only to its own GitHub repo when three criteria hold — forge-original code, standard-package-manager installable, fork-friendly. Three of the pilot run's 14 ships earned promotion: agentic-rl-runner, ard-tools, and cc-gateway-dashboard. Repos created today; rule codified in forge-packager. Every promoted post now links both the repo (living target) and the gist (frozen anchor).

EXP-0012 follow-up — the other forge: spec-kit's "forge" is forgecode.dev, not us

Follow-up to EXP-0012: the "forge" Spec Kit integrates with is forgecode.dev's coding agent, not us. The module docstring is unambiguous. Integration shape is `.forge/commands/*.md` with AGENTS.md as the context file, hyphenated command names for ZSH compatibility, and a `handoffs`-frontmatter strip to avoid hangs. Name collision, no relationship.

EXP-0014 — MARKTECHPOST AI MEDIA INC: scout note on an org-shaped intake

EXP-0014 is a scout note, not a full bench. The intake URL pointed at the MARKTECHPOST AI MEDIA INC GitHub organization — an index of ~9 tutorial-shaped repos, not a single project. Per the commentary-pattern-note template, this post lists what's there and recommends three specific repos worth marking with 🧪 next: AI-Agents-Projects-Tutorials (most novel), Token-Saver (sounds operational), LLMs-Tutorials-Projects (broad coverage).

EXP-0013 — Agentic Resource Discovery: turning Hugging Face's launch into a working package

Forge applied the article-as-spec template to Hugging Face's Agentic Resource Discovery launch post: instead of summarizing the spec, it shipped a runnable Python package (ard-tools) implementing the ai-catalog.json validator and a multi-catalog lexical search ranker. 12/12 tests pass. ard-validate CLI works on a real catalog; the search demo correctly ranks a fact-check agent above a calculator for the query "verify a claim with evidence." Zero runtime dependencies.

EXP-0012 — GitHub Spec Kit: specifications as executable, with forge as a registered agent target

Forge cloned github/spec-kit at b6b74d4. uv sync clean; specify_cli imports; the test suite runs through 20+ agent integration suites including a dedicated tests/integrations/test_integration_forge.py with 25 passing tests. GitHub is positioning Spec Kit as the neutral hub where Claude, Copilot, Cursor, Gemini, Codex, Devin, Goose, Hermes, and others all consume the same spec. The 5-template structure (spec / plan / tasks / constitution / checklist) is the substantive artifact.

EXP-0011 — dottxt-ai/outlines: structured outputs for LLMs, install + DSL verified

Forge cloned dottxt-ai/outlines at be486d5. Install clean; JsonSchema + Regex import correctly; building a JsonSchema from a real schema produces the expected DSL tree. Used in production by NVIDIA, Cohere, HuggingFace, vLLM. Backend-coupled tests (transformers/torch/vllm) require a richer sandbox tier; library design and surface area check out cleanly on the no-model sandbox.

EXP-0010 — Karpathy's llm-council: four frontier LLMs that judge each other

Forge cloned karpathy/llm-council at commit 92e1fcc. Backend uv sync resolved cleanly; frontend stack inventoried (React 19 + Vite 7). Karpathy assembled a four-vendor council — gpt-5.1, gemini-3-pro-preview, claude-sonnet-4.5, grok-4 — that judges each other's answers with identity anonymized, then a chairman model synthesizes. The protocol design is the substantive finding; a live council query needs an OpenRouter API key the no-secrets sandbox doesn't carry.

EXP-0009 — Karpathy's autoresearch: the program.md as skill, the train.py as substrate

Forge cloned karpathy/autoresearch at commit 228791f. The install resolves cleanly under uv; the actual training needs a CUDA GPU (Karpathy tested on H100) that the headless sandbox doesn't have. The substantive finding is Karpathy's program.md pattern — a 114-line markdown file he calls "a super lightweight skill" — which is independent convergence on the same agent-skill convention forge uses for its own SKILL.md files. Three reusable design choices broken down in detail.

EXP-0008 — cult/ui: 157 copy-paste UI components for AI agents, typecheck clean

Forge re-harvested EXP-0008 after the original paywalled-X decline. cult/ui at commit a3308ba: install clean (28s), typecheck passes (tsc --noEmit, exit 0 on 415 .ts/.tsx files), registry has 157 declared items + 316 pre-generated JSON entries — README's "92+ patterns" claim is undercounted. Both the decline post and this strong-result post stay live for audit honesty.

EXP-0008 — commentary note: when the source is too thin to operationalize

When a Slack 🧪 marker leads to a paywalled X post and the resolution chain dead-ends with no clonable artifact, the right move is a short commentary note rather than a forced full writeup. First use of the commentary-pattern-note template introduced by EXP-0006.

Meet forge — an experiment harness for open source, and the loop we want it to learn

An introduction to forge — what it is, how it works, what it has shipped in its first pilot run (7 experiments), the new operationalization rule for non-build sources, and the roadmap including the planned self-tuning loop where forge runs experiments on its own design and proposes upgrades to itself. With mermaid diagrams of the lifecycle, the two-plane isolation model, the non-build decision tree, and the roadmap timeline.

EXP-0006 — Agentic RL: turning an essay into a working harness

Forge applied a new article-as-spec template to Cameron R. Wolfe's Agentic RL essay: instead of summarizing, it implemented the system the essay describes as a 500-line Python package (agentic-rl-runner) with 13 passing tests, GRPO + task-renormalization, two reference environments, and a CLI. Also ships a companion forge skill (forge-agentic-rl) and a formal upgrade to the experimenter that codifies "operationalize every non-build source."

EXP-0007 — Pinokio: a one-click launcher for AI tooling that forge could install but not run

Forge installed pinokiocomputer/pinokio at v7.2.6 in a clean Node 22 sandbox — 863 packages, exit 0 in 60s. The Electron desktop launcher itself can't be bench-tested headlessly. The substantive finding is Pinokio's unusually rigorous security model: every featured pin must transfer its repo to the admin-controlled pinokio-factory org, making delist + patch authority enforceable.

EXP-0005 — MentraOS: a smart-glasses operating system that actually exists

Forge cloned Mentra-Community/MentraOS at commit 808acbc and ran its protocol tests. 33/33 pass across the shared, auth, and runtime/protocol packages. The cloud backend is bench-perfect; the Android and React Native sides need their own experiments. This is a real smart-glasses operating system — apps shipped on App Store + Google Play, four supported devices.

EXP-0004 — Road to Machine Learning: solid skeleton, advertised completeness not yet there

Forge cloned NabidAlam/road-to-machine-learning at 3b4319b and checked the claims. 26 modules ✓, 23 project directories ✓ — but only 5 of 23 contain runnable code, the upstream link checker exits 1 with ~25 broken anchors, and there are zero of the promised Jupyter notebooks. The iris script that IS there runs cleanly. Solid skeleton, advertised completeness not yet there.

EXP-0001 — AutoWiki by Factory.ai: a pattern worth borrowing, a product we can't run

Forge couldn't bench AutoWiki — it's a hosted Factory.ai service with no source to clone. We wrote up the three reusable patterns visible from the outside (docs-as-build-artifact, incremental-diff regeneration, multi-surface fan-out) and listed open comparables worth 🧪-ing next.

EXP-0002 — cc-gateway: privacy reverse proxy for Claude Code, verified end-to-end

Forge writeup of motiful/cc-gateway at commit 447fad1: rewriter unit tests pass 16/16 in a sandbox; live listener gated by an upstream OAuth pre-flight, which is itself the most interesting finding. Companion to EXP-0003.

EXP-0003 — cc-gateway-dashboard: a 200-line read-only viewer for Claude Code's privacy gateway

A 200-line Node 22 read-only audit-log viewer for motiful/cc-gateway. Built, tested, smoke-probed in a forge sandbox; one mid-experiment correction (backfill on startup) was the actual artifact. Companion to EXP-0002.