Context Management

How Context Pilot manages what the AI sees — the core concept that makes everything else work.

Overview

Every message sent to the LLM includes a set of context elements — the system prompt, tool definitions, file contents, memories, logs, conversation history, and more. Context Pilot makes these elements visible, measurable, and manageable.

ℹ️ Key insight In most AI tools, context is invisible — you can't see what the AI knows. In Context Pilot, every context element is a visible panel with a measured token count.

Context Elements

Context elements are ordered by priority. The LLM sees them in this order:

Type	Description	Example
System prompt	Active agent's instructions	~100 tokens
Tool definitions	All enabled tool schemas	~10K tokens
Core panels	Memories, callbacks, library, scratchpad	~3K tokens
File panels	Opened source files	Variable
Dynamic panels	Git, search results, console output	Variable
History panels	Previous conversation segments	Variable
Tree	Directory structure with descriptions	~2K tokens
Todos	Active task list	Variable
Logs	Timestamped event log	Variable
Conversation	Current chat messages	Growing

Panels

Panels are the visual representation of context elements in the sidebar. Each panel has:

ID — e.g., P5, P11, P23
Type — file, console, git, search, etc.
Token count — how much context space it uses
Content — visible in the main panel area

Core panels (P1-P7)

These cannot be closed and are always present:

P1 — Todo list
P2 — Library (agents, skills, commands)
P3 — Statistics (context usage, git changes)
P4 — Tools (all 58 tools with status)
P5 — Directory tree
P6 — Memories
P7 — Spine (notifications, auto-continuation)

Dynamic panels (P8+)

Created by tool calls — file opens, git commands, search results, console output. The AI can close these with Close_panel to free context space.

Token Budget

Context Pilot tracks token usage across all elements. The Statistics panel (P3) shows current usage vs. the budget:

Budget — maximum tokens the model supports (~200K for Claude Sonnet)
Threshold — soft limit where context should be managed (~140K)
Current usage — sum of all context element tokens

⚠️ Context overflow When context approaches the budget, the AI should close stale panels, summarize histories, and remove unnecessary elements. If it doesn't, you can ask it to clean up.

Context Management Strategies

The AI manages its own context

Context Pilot's AI is instructed to manage context proactively:

Close files after editing (they served their purpose)
Close old conversation histories (preserving info as logs/memories first)
Close search result panels after extracting relevant information
Close console panels after checking output

You can help

Ask the AI to "clean your context" when things feel cluttered
Use the sidebar to see which panels are consuming the most tokens
The Statistics panel (P3) shows accumulated costs per element

Element Lifecycle

Created — a tool call opens a file, runs a command, creates a panel
Active — the element is visible to the AI and counted in the token budget
Auto-refreshed — some panels (git, github) refresh periodically
Closed — removed from context via Close_panel or auto-suicide

Auto-suicide

Some panels clean themselves up automatically:

File panels — suicide when the file is deleted from disk
Console panels — suicide when the session is lost or a newer callback panel replaces it
Watchers — suicide when their target console session vanishes