copilot-session-management

Copilot Session Management

GitHub Copilot CLI holds every user message, model response, tool call, and tool result in a fixed-size context window. This skill covers the operational lifecycle: monitoring usage, compaction, checkpoints, and session hygiene — the levers that keep long-running Copilot sessions coherent.

When to Activate

Activate this skill when:

• A Copilot CLI session is long-running and the agent seems to forget earlier parts of the conversation
• Planning a multi-phase task (scaffold → implement → test → PR) in a single session
• Deciding whether to continue a session, compact proactively, or start fresh
• Debugging contradictions between current agent behavior and earlier decisions
• Auditing what was preserved across a compaction

Do not activate this skill for adjacent work owned by other skills:

• Do not activate for how Copilot builds prompts and retrieves context: copilot-context-architecture.
• Do not activate for designing compression strategies in your own agent systems: context-compression.
• Do not activate for diagnosing attention-level failures: context-degradation.

Core Concepts

What fills the context window

Four accumulating components:

1. System instructions and tool definitions — fixed overhead, always present.
2. User messages — every prompt sent.
3. Model responses — everything Copilot says back.
4. Tool calls and results — both request and output; tool results are often the largest contributor (long file reads, verbose command output).

Monitoring: /context

The /context slash command displays a visual breakdown: System/Tools (fixed overhead), Messages (conversation history), Free Space, and Buffer (reserved portion that triggers automatic management). Use it when a session is long, when Copilot seems forgetful, or to check whether compaction has occurred or is imminent.

Compaction

• Automatic trigger: at ~80% of capacity, compaction starts in the background, leaving a ~20% buffer so tool calls keep running. If context fills to ~95% before compaction finishes, the CLI pauses briefly until it completes.
• Manual trigger: /compact at any time — useful before starting a new phase of work. Esc cancels.
• Mechanism: snapshot the conversation → send the full history to the model with a summarization prompt (capturing goals, work done, key technical details, important files, next steps) → replace history with the summary plus original user instructions and current plan/to-do state → retain messages added during background compaction.
• What is lost: exact wording of messages, full command outputs, minor early decisions. Compaction is summarization — irreversible and lossy. Without it, the only fallback would be silently dropping old messages.