Concepts¶

Context engineering is the discipline of curating what enters an LLM's window: system prompts, tools, retrieval, history, and observations — under fixed attention limits.

Core techniques (this project)¶

Technique	Skill	What it does
Progressive disclosure	`context-fundamentals`	Load name + description first; full SKILL.md on activation
Observation masking	`context-optimization`	Replace verbose tool output with compact summaries
Hierarchical compaction	`context-compression`	Tier history by age/relevance
Handoff summaries	`context-compression`	Structured state for agent-to-agent transfer
Tool schema optimization	`tool-design`	Plain-language tool lists vs. heavy JSON
Context budgeting	`context-optimization`	Fit components under a token cap by priority
Filesystem offloading	`filesystem-context`	Large artifacts on disk; references in context

Progressive disclosure tiers¶

flowchart TB
  T1[Tier 1: Index — name + description]
  T2[Tier 2: SKILL.md body]
  T3[Tier 3: references/ deep dives]
  T1 -->|activation| T2
  T2 -->|needed| T3

Runtime primitives¶

Phase 2 exposes these as callable services (REST, MCP, SDK):

mask_observation
compact_session (hierarchical, handoff_summary, selective_retention)
budget_context
optimize_format
run_context_pipeline (combined optimization)

All run offline by default (deterministic fallback LLM for summarization paths).

Skill Router¶

Lexical, deterministic routing over skill names and descriptions — no network, no API keys. Used by the demo, MCP route_task, and REST POST /route.

Platform skills¶

Each supported agent platform (Copilot, Cursor, Kiro, Antigravity, Amazon Q) has three skills: context architecture, session management, and customization — documenting how that platform loads and manages context.