Rto Render

A provenance-tagged knowledge graph for your codebase.

The Portuguese roteiros were the guarded pilot books of the Age of Discovery — accumulated route knowledge that made navigation repeatable. Roteiro does the same for a codebase: structure, intent, and context in one queryable store, for humans and AI agents alike.

Every edge in the graph records how it was produced:

Provenance	Source	Nature
derived	tree-sitter extraction (full symbol graph for Rust today; more languages rolling out via tree-sitter tags)	Deterministic — symbols, calls, imports
authored	ADRs, blueprints, // @rto: annotations	Curated intent, drift-checked in CI
inferred	Docs, PDFs, images, embeddings	Fuzzy suggestions with confidence scores

One SQLite store. One query surface. Three renderers: a docs website, an Obsidian vault, and an optional MCP server (--features mcp) — all build outputs of the same graph, so what humans review is what agents query. Offline by default — one optional, default-off feature can call a hosted model, and it is described below; git-native and content-addressed, so the graph is shareable across a team.

Getting started

cargo install roteiro          # lean default build; offline unless you ask it to fetch
cd your-repo
roteiro init                   # store + git hooks + AGENTS.md
roteiro sync                   # build the graph
roteiro review                 # graph-grounded review of your change

The default build needs no toolchain class beyond the C compiler Rust itself requires — no C++, no cmake, no libclang — and makes no network call on its own. It includes everything needed to prepare for working offline — roteiro model pull and roteiro security prefetch|status|run — so docs/OFFLINE_SETUP.md needs no special build. Nothing is fetched until you say yes to it, and security run still requires --allow-unsandboxed on every invocation, because it executes a third-party analyzer on this host with no isolation boundary. The analyzers themselves are yours to install (semgrep, osv-scanner, cargo install cargo-audit) — the guide has the commands, and roteiro security ingest accepts a report produced anywhere if you would rather install none of them. Local inference and serving remain opt-in (--features inference-local-models, --features serve).

One capability sends your repository's content elsewhere. It is off.

Everything above runs on your machine. One optional feature does not, and it is named here rather than left to be discovered: --features remote compiles the remote model tier (ADR-0019), which can send graph-derived context to a hosted model.

• It is not in the default build. cargo install roteiro cannot send anything, and no release will change that. It is included in --all-features, so a build made that way can — with consent.
• Enabling the feature does not enable the tier. A run must be granted by your own ~/.roteiro/config.toml and by the invocation (--allow-remote, or answering a prompt that shows you the exact bytes). Neither alone suffices. A committed roteiro.toml may deny it for a whole repository but can never grant it — a merged line must not authorise egress on a teammate's machine.
• Three commands can send, and each has to be told to. roteiro remote call is the one that exists to; roteiro spec draft --allow-remote drafts with the hosted model instead of a local one; roteiro serve --allow-remote makes it the model the Ask panel uses. Without the flag all three are local, and only remote call ever prompts — on the other two the flag is the only way, because a prompt on a command whose default is local turns a habituated "yes" into consent you never quite gave.
• serve --allow-remote grants for the life of the server process, not one request at a time, and that is a materially larger exposure than a one-shot command: every Ask that server answers sends context to the hosted model for as long as it runs, including requests from anyone else who can reach the port. serve binds loopback by default, your user config still has to have granted, and every call is on the ledger. The grant dies with the process and is never persisted.
• A refused --allow-remote stops the run. If you asked for the hosted model and a layer said no, Roteiro names the layer and does not answer from a local model instead — that would be a different answer with nothing to signal the change.
• What would be sent is inspectable first, and what did is recorded. roteiro remote dry-run prints the exact body and sends nothing; roteiro remote log reads the append-only ledger of what left, and when.
• Source code is not sent — function bodies are not in the graph. That is not the same as nothing identifying being sent: symbol names, file paths and captured prose go, they identify a codebase, and there is no redaction chokepoint on a prompt. roteiro remote status says so in full.
• It never falls back quietly. With the tier on and the endpoint unreachable, Roteiro fails with an error naming the endpoint rather than answering from a local model — a different model is a different answer.

So "nothing leaves the machine" is true of Roteiro as shipped and as configured, and is no longer a property of the software as a whole. ADR-0006 states the same sentence scoped to roteiro serve, which still exposes only installed models and still never downloads.

Planning to work on a train or a plane? Models and analyzer databases must be fetched once, deliberately, before you disconnect — docs/OFFLINE_SETUP.md is that one-time preparation, including the air-gapped route.

See https://roteiro.dev for the full guide (modes, local models, languages, config), and ADR-0001 plus docs/BUILD_PLAN.md for the design and roadmap. Contribution + review standards live in AGENTS.md.

Sandboxed analyzers (--features exec-boxlite)

Optional, off by default, and the only feature with build requirements the default install does not have. It runs semgrep inside a digest-pinned OCI image in a microVM — read-only worktree, no egress, no ambient credentials (ADR-0014).

1. protoc >= 3.12 must be on the build host. boxlite-shared's build script requires it unconditionally, and without it the failure is a build-script error rather than a missing-feature message:

sudo apt-get install -y protobuf-compiler   # Debian/Ubuntu
brew install protobuf                       # macOS

2. The sandbox runtime is verified at build time, whichever way you build. boxlite embeds a prebuilt runtime archive into the compiled library, which its own build script fetches with an unverified curl. Roteiro pins the SHA-256 of every file that archive contributes and checks them in rto-exec's build script — against what boxlite actually extracted, before anything links. A mismatch, a missing file or an unpinned extra file fails the build.

# Just build it. The runtime is fetched by boxlite over TLS and every extracted
# file is verified against the pins; the build says so on its own output.
cargo install roteiro --features exec-boxlite

# Then pull the pinned analyzer image. Only this build can: the image half of
# `prefetch` is compiled out of any binary without `exec-boxlite`.
roteiro security prefetch --analyzer semgrep --allow-download

The second step is separate for a reason of ordering rather than taste: the binary that can pull an image is the one the first step exists to produce. Because a run never pulls, skipping it means roteiro security run refuses with ImageNotProvisioned instead of fetching an image mid-scan.

For a build that touches no network at all, provision the archive first and name it. boxlite's curl then reads a local file and opens no socket, and the bytes are verified before they are extracted as well as after:

cargo install roteiro                                           # default set has `prefetch`
roteiro security prefetch --analyzer sandbox --allow-download    # digest-pinned, ~26 MB
BOXLITE_RUNTIME_URL="file://$HOME/.roteiro/security/boxlite-runtime/boxlite-runtime.tar.gz" \
  cargo install roteiro --features exec-boxlite

Both paths are verified; only the second is offline. Do not describe the first as air-gapped, and do not describe the second as the only safe one — the difference is egress, not integrity (ADR-0014 v1.3).

3. It embeds third-party binaries, some of them GPL-2.0 and LGPL-2.0, and distributing a binary built this way carries source-offer duties. prefetch prints the full notice before installing anything; it is also at crates/rto-exec/NOTICE-boxlite-runtime.md. A default build embeds none of it.

Platforms: darwin-arm64, linux-x64-gnu, linux-arm64-gnu. Running a scan also needs a usable hypervisor (/dev/kvm on Linux, Hypervisor.framework on Apple Silicon); where there is none, the run says so by name and the sandbox tests skip with a visible message.

Logging

By default Roteiro logs human-readable text to stdout, unchanged. You can additionally stream logs to a rotating file in an OpenTelemetry-shaped JSON format (groundwork for a future OTLP collector — see ADR-0011):

roteiro --log <cmd>                                  # file at $ROTEIRO_HOME/logs/roteiro.log
roteiro --log-file /var/log/roteiro.log <cmd>        # explicit path (enables it)
roteiro --log-rotation hourly --log-format otel <cmd>

Or set it once in roteiro.toml (or ~/.roteiro/config.toml):

[telemetry]
file = "~/.roteiro/logs/roteiro.log"   # unset ⇒ file logging OFF (stdout only)
rotation = "daily"                     # daily | hourly | minutely | never
format = "otel"                        # otel | json (alias) | text

Each flag also has an env var (ROTEIRO_LOG_FILE, ROTEIRO_LOG_ROTATION, ROTEIRO_LOG_FORMAT); ROTEIRO_LOG sets level filter directives (e.g. debug). Precedence is flag > env > project > user > default. Writes are non-blocking, so logging to disk never stalls a command. The OTLP network exporter and metrics are deferred; see the ADR for the field mapping and the seam.

Workspace

• crates/rto-graph — SQLite store, provenance model, content-addressed cache, extraction, sync, query, inference
• crates/rto-spec — house-style ADR/blueprint parsing, check (drift gate), importers, spec authoring
• crates/rto-render — docs site, Obsidian vault, MCP server (feature-gated)
• crates/rto-serve — local OpenAI-compatible model server (llama.cpp; feature-gated)
• crates/rto-llama — llama.cpp inference core (generation, embeddings, vision), shared by serving and internal uses
• crates/roteiro — umbrella CLI

License

Dual-licensed under MIT or Apache-2.0, at your option. Contributions are accepted under the same terms.