Repository guidance for this project lives in AGENTS.md — read it first. It is the single source shared by Claude Code, Codex, Copilot and any other agent, so that guidance never drifts between tools.
Task procedures live in .agents/skills/.
Add new repository guidance to AGENTS.md or a playbook, never to this file — anything written
here is invisible to the other tools.
Follow Human in the loop before publishing or connecting to a tenant. Agents may open only draft PRs, after human review, an explicit request, and the required validation evidence.
This is about acting unprompted. When the user asks you directly to connect to a tenant and run something, that is the authorization — do it, confirming each destructive command once. See Running against a tenant.
- Subagents:
.claude/agents/—permissions-auditor,docs-sync,api-surface-diff,issue-triage,cmdlet-scaffolder,release-prep,whatif-audit. Each one is a thin wrapper delegating to its playbook. The same set exists for Copilot (.github/agents/*.agent.md) and Codex (.codex/agents/*.toml), pointing at the same playbooks — keep the names in step when adding one. - Skills:
.claude/skills/—code-review,new-cmdlet. Same arrangement. - MCP:
.mcp.jsonconfigures the GitHub and Microsoft Learn servers. See.agents/README.md.
docs-sync, permissions-auditor, whatif-audit and api-surface-diff are read-only and fan out well across the
800+ cmdlets. For a whole-repository or multi-folder sweep, run them in parallel over separate
folders under src/Commands/. For one folder or a changed-file set, use a single agent or do the
check inline — each subagent adds to the session's cost.
docs-sync and api-surface-diff pin model: sonnet in their frontmatter because their work is
mechanical extraction. The others inherit the session model.