Context Pack
Context Pack is a distilled bundle of code, graph signals, and memory context tailored to a question. It keeps prompts lean while delivering high-signal context — including intelligent surfacing of lessons, intent detection, and compaction awareness.
What it includes
Code + files
The most relevant files, snippets, and symbols for the question.
Graph signals
Dependency and impact context that explains what connects to what.
Memory + decisions
Prior decisions, lessons, and preferences related to the query.
Enhanced Context
Context Pack goes beyond simple retrieval. It understands what you're doing and proactively surfaces what matters.
Intent detection
Semantic classification understands your intent — whether you're debugging, refactoring, or implementing. Context is shaped accordingly.
Lessons surfacing
Past mistakes and critical lessons appear proactively before you repeat them. The system learns from corrections.
Context pressure tracking
Know when your conversation is getting long. Get warnings before context limits affect quality.
Compaction awareness
Pre-compaction snapshots preserve critical state. Session context automatically restores after summarization.
How to use it (MCP)
Ask your AI tool to call context_smart withmode="pack" when you need code or file context.
context_smart(
user_message="What does the auth refresh worker do?",
format="minified",
max_tokens=400,
mode="pack",
distill=true
)If Context Pack is disabled, the API falls back to standard context_smart.
Token controls
max_tokens - hard budget for the response payload.
format="minified" - compact output for assistants.
distill=true - returns a compact summary when available.
When to use Context Pack
Code explanation
Summarize what a function or module does without pulling the whole file.
Refactor planning
Quickly identify what might break and what depends on it.
Long sessions
Context pressure tracking warns you before hitting limits. Snapshots preserve state through compaction.
Avoiding past mistakes
Lessons from previous sessions surface automatically before you repeat them.