'Audit all .eng/ git repos under ~/Code/ by default, or under provided roots — report which are dirty, which have unpushed commits, and which look uninitialized. Read-only. Use when the user asks about EngDirs/.eng repo status, unpushed commits, dirty working trees, or "what needs to be pushed/committed". Triggers: eng status, engdirs status, eng audit, what'\''s dirty, unpushed eng.'
Collect Copilot Chronicle session-store SQLite databases from local machine, WSL, and remote SSH hosts into a single staging directory, then query them with sqlite3. Use when the user wants to inspect, search, or export their Copilot chat history (sessions, turns, files, refs) across machines. Triggers: chronicle, session-store.db, copilot history, session history, search past sessions, what did I ask, find prompt.
Search and retrieve work information across M365 activity, internal engineering knowledge, and official Microsoft documentation. Use when gathering meeting summaries, email threads, chat messages, calendar events, internal docs, technical support guidance, or Microsoft/Azure product documentation. Covers tool selection, query strategies, and data-quality rules.
Workspace topology inspector and reconciler. Marker: my-workspace.yaml at root (also the index and the only loose file unless a file has its own index entry). Default mode is read-only. Reconcile only after proposing an exact plan and receiving explicit consent. Never auto-assigns vc_exception — propose, confirm per-dir. Keywords: workspace, conventions, layout, validate, is my workspace in a good state, topology, manifest, index.
Convention-aware worktree setup for a my-workspace topology. Naming: worktrees for repoX/ go in repoX.worktrees/<branch-name>/. Propose-approve-execute with standard preflight. Keywords: worktree, git worktree, isolated branch, worktree scaffold.
Diagnose and fix Copilot / VS Code agent files (`*.agent.md`) — especially their `tools:` list — and keep them correct as plugins install and MCP servers change. Default posture is a read-only "doctor": when an agent can't use a tool, an install scrambled its tools, an MCP server renamed tools, you're onboarding a new MCP server, or you're setting up a new agent, work out what's wrong and what the fix is — then apply it only when asked. Intent lives in a version-controlled store and is expanded mechanically into each agent's `tools:`. Keywords: agent files, agent tools, doctor mode, agent can't use tool, tool not available, toolsets, tools frontmatter, tool drift, MCP tool rename, onboard MCP server, new agent, expand toolset, assignments, agent-doctor, DrAgent, fix agent tools.
Run an LLM council: pose a question to multiple orthogonal models and/or specialist personas in parallel (via the harness's subagent tool), then summarize each stance and synthesize the responses with a separate chair model. Use when the user wants diverse model perspectives, a panel of specialist lenses (concise, skeptic, coder, prose), or cross-model review of a high-stakes or ambiguous question, or asks to ask the council, panel, or multiple models. Keywords: council, panel, specialists, personas, multiple models, diverse perspectives, ensemble, model comparison, cross-model review, fan-out, synthesize, second opinion, 3-flag, 2-horse.
Detect and fix AI slop in text — vague filler, fake authority, clichés, buzzwords, and structural patterns that signal low-value AI-generated content. Use when reviewing, editing, or writing any prose: docs, blog posts, emails, summaries, weekly updates, READMEs, or marketing copy. Keywords: slop, AI slop, cliché, filler, buzzwords, writing quality, lint, review text.