Expert Context Pack (ECP) v1.0
Persistent, maintainable, and testable expert context for AI agents. ECP is a backward-compatible extension to Agent Skills that adds portable citations, eval-gated refresh, and auditable provenance — without vendor lock-in.
Key features
Persistent expert context
Store domain knowledge that persists across agent sessions — no constant rediscovery.
Portable citations
Ground every answer in verifiable evidence: file path, line ranges, and revision hash.
Eval-gated maintenance
Refresh context incrementally with built-in test suites and optional rollback on failure.
Offline-first
Works entirely offline. Network features are optional and explicitly controlled.
Multiple indexing backends
Keyword, SQLite FTS, vector embeddings, or hybrid retrieval with Reciprocal Rank Fusion (RRF).
Portable packaging
Create distributable ZIP packages with hash verification and portability constraints.
What is ECP?
Expert Context Pack (ECP) is a backward-compatible extension to the Agent Skills format.
It adds a standardized expert/ directory to any skill, containing:
EXPERT.yaml— manifest declaring sources, artifacts, policy, and evals- Indexed context — chunked, searchable knowledge from your codebase or documents
- Maintenance policy — controls for refresh cadence, rollback, budgets, and retention
- Evaluation suites — automated tests validating response shape and citation integrity
Agents that don’t understand ECP simply ignore expert/. ECP-aware runtimes can leverage it to
produce grounded, citable answers.
Use cases
Organizations
Package institutional knowledge into portable experts that can be validated before deployment.
LLM agents
Ground responses with citable evidence from your codebase — every claim traces to source.
Security-conscious environments
Offline operation, explicit secrets/PII controls, and auditable provenance by design.