GitHub's daily trending page on October 6, 2026 was led by a testing framework, with two agent-skill projects for non-software work close behind. Star counts below are from GitHub's trending page and API at the time of writing.
tester-army/e2e: natural-language steps, recorded and replayed
e2e (Apache-2.0, created July 2026, about 4,900 stars, 1,398 of them today) is an end-to-end testing framework for web and mobile apps. A test mixes ordinary assertions with agent steps such as agent.act('upgrade the workspace to the Pro plan') and agent.assert('the invoice preview shows a prorated amount'). The design choice that matters is in the README: "An agent step that a later assertion verifies records its actions, and the next run replays them with no model calls until the app changes." Tests without agent steps need no model, and users bring their own subscription, API key or local model. Web runs use Chromium, Firefox and WebKit through Playwright; mobile runs use iOS simulators and Android emulators; a reporter posts results as a pull-request comment.
What developers are building, and the constraint. This answers the two objections to AI-driven testing: cost and flakiness. Paying for a model call on every CI run, and getting a different path each time, makes a test suite useless. Recording once and replaying deterministically until the UI changes keeps the model out of the hot path. The open question is how well "until the app changes" is detected -- a stale recording that still passes is worse than a failing test.
text-to-cad and OpenMontage: professional workflows as agent skills
- earthtojake/text-to-cad (MIT, about 17,500 stars, 437 today) gives an agent local workflows to generate 3D models as STEP, GLB, STL or 3MF files, run design-for-manufacturing checks, produce engineering drawings and connect to 3D-printing, sheet-metal and CNC services. It installs as a plugin or skills for Claude Code, Codex, Cursor, Gemini and Grok, built on build123d and Open CASCADE.
- calesthio/OpenMontage (AGPL-3.0, about 64,100 stars, 742 today) bills itself as an "agentic video production system" with 12 production pipelines and more than 700 agent skill and production-knowledge files. It can start from a reference video, analyse its pacing and structure, and return concepts with cost estimates before generating any assets.
Analysis: all three put the expensive or risky part -- a model call, a fabrication order, a paid video generation -- behind a check: a recorded replay, a manufacturability check, a cost estimate. That is the pattern worth copying. Note also the licences: OpenMontage's AGPL matters for anyone embedding it in a hosted product.
October 6's GitHub trending list was led by tester-army/e2e, which records agent-driven test steps once and replays them with no model calls until the app changes, alongside text-to-cad and OpenMontage, which package CAD and video production as agent skills -- each putting a check in front of the costly step.