Rinne v0.1.8: code graph, learn explainers and human controls
We just shipped v0.1.8 for Rinne. This release is about making long agent runs faster smarter and more inspectable: a persistent code graph so workers stop re-reading the whole repo every time, an interactive rinne learn explainer for any topic in your codebase, model routing with an escalation ladder, human-in-the-loop checkpoints and a live usage ledger so rate limits and token spend stop surprising you.
What shipped
v0.1.7 was about extensibility: MCP servers and Agent Skills. v0.1.8 is about the loop itself getting sharper. Rinne now indexes your code into a durable local graph, can generate a full interactive explainer for a topic, routes work by model capability with automatic escalation, lets you pin humans into the plan and tracks every token against live limits.
The workspace grows to eight crates with a new rinne-graph crate for tree-sitter indexing. Full notes and binaries are on the v0.1.8 release.
Persistent code graph
Rinne used to pay a full-repo tax on every run. It now keeps an incremental code graph in .rinne/state.db using tree-sitter and SQLite. Symbols (classes, functions, methods, imports) and call edges are extracted for Rust, TypeScript and Python. Call neighborhoods get mapped into worker prompts so an edit lands with architectural context instead of a blind file dump.
Inspect and rebuild it from the CLI or the TUI:
# Build or refresh the index
rinne graph index
# Counts of files, symbols and edges
rinne graph stats
# Symbols in a file
rinne graph symbols src/lib/loop.rs
# Callers and callees of a symbol
rinne graph neighborhood plan_task
# Opt out for a run
rinne --no-graphTUI mirrors: /index, /graph stats, /graph symbols <file>, /graph neighborhood <symbol>.
rinne learn
Ask Rinne to explain a concept in your codebase and get a self-contained interactive HTML walk-through. It reindexes the matching files and adjacent symbol clusters, resolves the topic to precise symbols, then assembles doc comments, code snippets and any matching CONTEXT.md sections into a narrative with a Mermaid flowchart of the code path.
# Generate an explainer for a topic
rinne learn "conductor routing"
# Open it in the browser when done
rinne learn "blackboard resume" --open
# Literal docs only, no LLM narration
rinne learn "rate limits" --no-ai