Repo-local context for coding agents
AgentRail now includes an AGENTS.md file so future Codex, Copilot, Claude, or OpenCode sessions start with the same boundaries: static site, personal Beta, preserve release commands, verify before publishing.
Skills and instructions
AgentRail should produce task packets that any coding agent can execute and another agent can review.
The next useful layer is not more prompt text. It is repo-local instructions, reusable Skills, workitem state, and acceptance criteria that survive across Codex, Claude Code, Copilot, and future workers.
Skills make repeatable work portable
Claude Skills show a practical direction: reusable instructions plus supporting files for a domain workflow. AgentRail should turn website update, release verification, and review flows into portable skill-like packets.
Record, replay, and repo execution
Codex-style skills and replayable workflows matter because AgentRail's real value is repeatability: same task shape, same checks, same evidence, less operator memory.
Tool access needs an allowlist
MCP can expose files, services, browsers, and APIs. AgentRail should record which tool families are allowed for a workitem and which actions need human approval.
The page should generate dispatch briefs
The next website feature should be a local task-packet generator: goal, scope, repo path, target pages, agent choice, acceptance checks, PR checklist, and changelog note.
Task packet shape
Target: www.agentrail.tech Worker: Codex / Claude Code / Copilot Scope: update selected pages only Acceptance: URL 200 checks, install smoke test, mobile screenshot, evidence summary Human decision: review before merge and publish