Agent tooling · GitHub automation · instruction synthesis · 2026 · Solo — popularity-led research, conflict-detection algorithm, GitHub object API integration, TDD, interaction design, exact-SHA deployment
Skill Merge
A configure workspace that turns a pile of agent skills into one reviewable AGENTS.md pull request.
The problem
Agent teams are installing more skills than they can manually reconcile. Each Hermes or Claude skill ships as a Markdown instruction file with headings such as 'When to Use', 'Verification Checklist', and 'Pitfalls'. Once a repo contains dozens of skill files, repeated headings can contradict each other, and manually composing a single AGENTS.md becomes a slow, lossy editorial task. Skill Merge reads the real skill files from a GitHub repository, detects section-level conflicts, lets the operator choose pick-left, pick-right, or merge-both per conflict, and opens a draft PR containing the synthesized AGENTS.md. The nightly concept came from secure radar run radar-20260824-99dbfe9b: mattpocock/skills was daily rank #2 with 2,447 stars gained that day and 234,162 total stars. Rank #1 openai/codex was vetoed for NO_ORIGINAL_PRODUCT because adjacent session-to-commit ideas overlapped multiple recent portfolio builds. Skill Merge translates the new agent-skills wave into original merge tooling rather than copying the source repository's package manager, brand, interface, or skill content. Production verification used a private sandbox repo with two real .hermes/skills/*/SKILL.md files; the deployed app scanned the repo, detected conflicts in 'When to Use' and 'Verification Checklist', generated a preview, and opened draft PR https://github.com/bradleybeatz1313/skill-merge-sandbox-20260824-0711/pull/1 changing only AGENTS.md.
Architecture
Key decisions
Configure surface instead of another GitHub commit dashboard
Several recent builds also end in GitHub commits, so the design had to be structurally distinct. Skill Merge makes the primary surface a per-heading conflict configuration workspace: the visible work is choosing instruction precedence, not watching a generic API transaction complete.
TDD on the algorithm before UI code
Twenty-five tests were written and watched fail before implementation modules existed. They cover YAML frontmatter stripping, h1 fallback titles, h2 section extraction with h3 absorption, conflict detection for shared headings, identical-content suppression, three-way conflicts, classification, pick-left, pick-right, merge-both, line deduplication, and final Markdown synthesis.
Git object API instead of Contents-only shortcut
The app creates a blob, tree, commit, branch ref, and draft PR through GitHub REST. That object-model path makes the write atomic and reviewable, avoids direct mutation of the base branch, and exposes the exact commit SHA behind the result.
Blueprint visual language
The surface uses a cyan-on-near-black technical blueprint grid, mono typography, square edges, and side-by-side resolution bays. The one design risk is making conflict resolution feel like engineering drafting rather than a document editor, differentiating it from recent parchment, journal, and terminal-style Foundry builds.