Context Engine

Coding agents waste tokens re-reading whole files to find the few lines that matter. AgentRail's context engine fixes that. It builds a local index of your repo and, for each task, compiles a context pack: a bounded set of line ranges, each with a cited reason it was picked.

A local index

agentrail init (or agentrail context index) walks your repo and builds two things on your machine:

• A BM25 keyword index over your source, for lexical retrieval.
• A code graph — a deterministic relationship model over files, symbols, imports, references, and tests.

Indexing stays local. Nothing about your source leaves your machine to retrieve context — the engine runs repo-adjacent, not in a cloud crawler.

Context packs, not file dumps

When an agent picks up a task, the engine compiles a context pack: instead of pasting whole files, it selects the exact line ranges the task needs and attaches a reason to each pick — a symbol definition, a graph expansion, a BM25 keyword match. That bounded, cited selection is what the agent reads, and the source of the measured token win.

agentrail context query "where is the rate limiter applied to ingest?"

Structural navigation

The code graph also answers structural questions directly, without an LLM:

agentrail context def RateLimiter        # where a symbol is defined
agentrail context callers RateLimiter    # who calls it
agentrail context impact RateLimiter     # what a change would touch
agentrail context ast path/to/file.ts    # parsed structure of a file

Use agentrail context sources to see what the index covers.

Next: Bounded Execution — how an agent turns a context pack into a reviewed change.