Skip to content

Annex B — Product & feature reference

Annexes

Quick reference "cards" for the products and features tested by GH-300 and GH-600. Each notes what it is, when to use it, and the chapter it maps to.

GitHub Copilot surfaces

  • Inline suggestions — grey ghost-text completions as you type; accept with Tab. Best for in-flow coding. (Ch. 5)
  • Copilot Chat — conversational help with slash commands (/explain, /fix, /tests) and context references (#file, #selection). Best for explaining/refactoring specific code. (Ch. 3, 5)
  • GitHub Copilot CLI — standalone copilot terminal agent; interactive or programmatic (-p); plan mode; can edit files, run commands, open PRs; asks approval; run from trusted directories. (Ch. 5)
  • Agent mode — autonomous, goal-driven multi-file changes that run tools (tests/build) and iterate. (Ch. 6)
  • Copilot Edits — described multi-file edits you scope and accept/discard as a set. (Ch. 6)

Customization & extensibility

  • Instructions files — standing standards: .github/copilot-instructions.md (repo-wide); path-specific .github/instructions/*.instructions.md with applyTo; AGENTS.md (agent). Precedence: personal > repo

    org. (Ch. 6)

  • Prompt files — reusable, repeatable requests saved in the repo. (Ch. 6)
  • Model Context Protocol (MCP) — open standard to connect agents to external tools/data; GitHub MCP server built in. (Ch. 6, 10)
  • Custom agents / sub-agents — specialized agents (e.g., Explore, Task, Code review) the model can delegate to; defined in Markdown at user/repo/org level. (Ch. 6)
  • Spaces — curated, reusable/shareable context. (Ch. 6)
  • Spark — build an app from a description. (Ch. 6)

Administration & safeguards

  • Organization policies — enable/disable features and set availability across IDEs and github.com. (Ch. 8)
  • Copilot Code Review policies — govern availability of Copilot's PR review. (Ch. 6, 8)
  • Content exclusions — keep specified files/paths out of Copilot's context (input side). (Ch. 8)
  • Public-code duplication filter — suppress suggestions matching public code (output side). (Ch. 8)
  • Audit log events — records of Copilot admin/usage events; answer "who did what." (Ch. 8)
  • Copilot REST API — manage seats/subscriptions programmatically. (Ch. 8)

MCP governance (GH-600)

  • "MCP servers in Copilot" policy — whether MCP servers may run at all. (Ch. 10)
  • MCP allow list — managed settings — enterprise managed-settings.json; GA, strong, non-overridable (matches by name/URL/stdio). Preferred. (Ch. 10)
  • MCP allow list — custom registry — self-hosted; public preview; weaker/bypassable. (Ch. 10)

Agentic building blocks (GH-600)

  • Structured plan — explicit steps produced before acting; enables validation. (Ch. 9)
  • Agent memory — short-term / long-term / external; scope and apply expiry/pruning/reset. (Ch. 11)
  • Evaluation signals & scanning — quantitative/qualitative signals; CodeQL, dependency review, secret scanning as sources. (Ch. 12)
  • Multi-agent orchestration — patterns, isolation, conflict resolution, recovery, lifecycle. (Ch. 13)
  • Guardrails / autonomy levels / HITL — match control to risk; least privilege; block violations. (Ch. 14)

Cloud-agent guardrails at a glance (GH-600)

  • Only write-access users trigger it; pushes to a single branch; limited credentials (simple push). Opens draft PRs it can't approve/merge; workflows need approval; requester can't self-approve; extra approval for unattributed PRs; restricted internet; CodeQL + secret scanning (no GHAS license); signed, co-authored commits + session logs + audit events. (Ch. 10, 14)