Skip to content
SHRINESHRINE

Repository Context

Draft

Let the agent look up the codebase instead of rereading it.

What This Covers

Section titled “What This Covers”
  • Indexed knowledge of the code itself: symbols, definitions, references, call and dependency graphs
  • Exposed to agents as tools (usually MCP) or as a precomputed map
  • Distinct from Memory & Context, which holds agent and session state
  • Rule of thumb: repository context is derived from the code; memory is what the code cannot tell you

Key Principles

Section titled “Key Principles”
  • Query the index before reading whole files
  • Navigate by symbol, not by grep, when a semantic tool exists
  • The index must track the code; a stale graph is worse than none
  • Serves Consistency as Leverage: consistent code indexes cleanly and predicts well

Approaches

Section titled “Approaches”
  • LSP-backed navigation: go-to-definition, find-references, rename via the Language Server Protocol
  • Symbolic MCP tools: symbol lookup and symbol-level edits exposed to the agent
  • Code knowledge graph: parsed codebase stored as a queryable graph
  • Repo map: ranked summary of files and signatures sent with each request

Current Stack

Section titled “Current Stack”

Key Considerations

Section titled “Key Considerations”
  • How does the index refresh after edits, and can the agent detect staleness?
  • Does it cover every language in the repo?
  • What does one query cost in context compared with reading the file?

Anti-patterns

Section titled “Anti-patterns”
  • Storing session decisions in the code index, or code structure in memory
  • Dumping the full repo map into every prompt regardless of task
  • Trusting an index that has not been refreshed since the last large change
Section titled “Related”