9. Glossary

Terms are cross-referenced to the subsection where they're discussed.

Agent Mode — A high-autonomy IDE mode where Copilot plans, edits across files, runs commands, and iterates toward a goal with your approval. (2.3.1)

Agent session — A persistent, stateful run that holds agentic work together and carries context forward (CLI or IDE). (2.3.2)

Audit log — Org/enterprise record of governance events such as Copilot policy and configuration changes; an accountability tool, not a view of suggested code. (2.4.3)

Autopilot Mode — A CLI mode that executes steps without per-step confirmation; trades oversight for speed. (2.2.3)

Cloud coding agent — An asynchronous agent that works outside the IDE: assign it a GitHub issue and it opens a draft PR. Distinct from Agent Mode. (2.3.1)

Content exclusions — Admin-defined paths/patterns that Copilot must not use as context or suggestions; do not delete code; coverage can vary by editor. (2.1.3, 7.1.1)

Context window — The finite amount of context an LLM can consider at once; a structural limitation. (4.2.2)

Copilot CLI — A terminal-native coding agent (plan/autopilot modes, sub-agents, session memory); GA February 2026. (2.2.1)

Copilot code review — Agentic, project-aware AI review of pull requests; can pass fixes to the coding agent. (2.3.3)

Duplication detection (public-code matching filter) — Configurable policy that suppresses/flags suggestions matching public code; prerequisite for IP indemnity on Business/Enterprise. (7.2.1)

Edit Mode — A low-autonomy mode applying directed, scoped edits across files you specify. (2.3.1)

Few-shot prompting — Providing multiple input→output examples to lock in a pattern or format. (5.1.3)

Fill in the middle (FIM) — Prompt construction that lets the model complete code between existing code, not just append. (4.1.2)

GitHub AI Credits — Token-based usage billing that replaced premium requests on June 1, 2026; post-dates the exam snapshot. (Header currency note)

Hallucination — Plausible-looking but incorrect or nonexistent output (e.g., a fabricated API). (3.1.1)

Instructions files — Repository/org guidelines (e.g., copilot-instructions.md) that make Copilot follow team conventions. (2.3.4)

IP indemnity — Protection offered on Business/Enterprise for suggestions, contingent on the public-code matching filter being enabled. (7.1.2)

Model Context Protocol (MCP) — Open standard connecting AI tools to external data/tools; replaced proprietary Copilot Extensions; GA across major editors. (2.3.2)

Next edit suggestions — Predicts and suggests your likely next edit; included, no credit cost. (2.1.1)

Plan Mode — Drafts a step-by-step plan before execution for your approval; doesn't itself complete the work. (2.1.2, 2.3.1)

Prompt files — Reusable Markdown prompts in the workspace for consistent team responses. (2.3.5)

Proxy — GitHub-owned service on Azure that screens prompts inbound and post-processes responses outbound; an active checkpoint, not a relay. (4.1.3)

Slash commands — Chat shortcuts (e.g., /explain, /tests, /fix) that scope a request to a common intent. (2.3.5)

Spaces (Copilot Spaces) — Bundled, curated context Copilot can draw on for grounded answers. (2.3.4)

Spark (GitHub Spark) — Natural-language app builder (Pro+/Enterprise) that generates a working app with live preview. (2.3.4)

Sub-agents — Parallel workers a primary agent delegates parts of a task to (the CLI's "fleet" behavior). (2.3.2)

Telemetry (accept/dismiss) — Anonymized signal of whether suggestions are accepted; acceptance rate is GitHub's primary quality signal. (4.1.1, 4.2.1)

Zero-shot prompting — Instruction only, no examples; efficient for common, well-understood tasks. (5.1.3)