LLM workflows

Built for coding agents

Point your agent at AGENTS.md and the JSON recipes. Discover once; query forever.

  1. Step 1

    User prompt

    Natural language — “what breaks if I change checkout?”

  2. Step 2

    Tool call

    Agent runs rg-build -f json … instead of dumping files into context.

  3. Step 3

    Graph facts

    schema_version JSON — blast, GQL rows, semantic hits, CPG mutations.

  4. Step 4

    Reason + edit

    LLM plans the change from facts, then verifies with check when needed.

Minimal agent loop

export REPO=/path/to/repo
$ rg-build -r "$REPO" discover .
$ rg-build -r "$REPO" -f json gql --macro-name all_functions unused | jq '.count'
$ rg-build -r "$REPO" -f json blast-radius "ShoppingCartService" --depth 3 \
| jq '{score: .metrics.score, callers: .metrics.direct_callers_count}'
Open AGENTS.mdRecipesInteractive demos