rote: compile a skill into a deterministic pipeline

You orchestrate the rote CLI. It runs an LLM compiler agent over a source skill once, and emits a pipeline that runs forever after without an agent loop. Your job is to resolve the inputs, run the CLI, and interpret the output. You never classify nodes or write pipeline.yaml yourself; the CLI's compiler agent does that.

When this applies

Use it on a skill the user has already run many times and wants to run many more, unattended. Exploratory or one-off work should stay an agent loop: flexibility is the point there, and there is nothing proven to compile yet. Say so and stop if that is what you are looking at.

1. Identify the source skill

The source is a directory containing a SKILL.md, optionally with a references/ folder. The user names it, or you infer it from context: a skill just discussed, a path in the conversation, .claude/skills/* or skills/* in the project.

Confirm the resolved absolute path with the user before running. Compilation costs real time and tokens, so never guess and go. If the directory has no SKILL.md, stop and ask.

2. Pick a runtime target

Runtime	--runtime	Language	Choose when
DBOS (default)	dbos	Python	No orchestrator to deploy. SQLite for dev, Postgres for prod
Temporal	temporal	Python	You already operate a Temporal cluster
Plain Python	python	Python	Max legibility, stdlib only. Refuses pipelines with HITL gates
Cloudflare Workflows	cloudflare	TypeScript	Serverless, managed, wrangler deploy-ready
DBOS (TypeScript)	dbos-ts	TypeScript	Zero orchestrator on the TS side. Postgres only
Inngest	inngest	TypeScript	Mounting into an existing Node or Next.js app

If the user has no opinion and no existing infrastructure, use dbos. It is the default and the only Python target with zero standing infrastructure, so you can omit --runtime entirely.

3. Resolve the CLI

The CLI ships on PyPI as rote-cli and its executable is named rote. With uvx that means every invocation is uvx --from 'rote-cli>=0.12.1' rote <args>. Do not run uvx rote-cli ...; uvx looks for an executable named after the package, and the published wheel does not ship one.

uv --version                              # install uv first if missing
uvx --from 'rote-cli>=0.12.1' rote --version        # confirm the CLI resolves

If uv is missing, do not pipe a remote script into a shell. Ask the user to install it through their package manager (brew install uv, pipx install uv, or pip install uv) or to follow the official guide at https://docs.astral.sh/uv/getting-started/installation/ and choose the method they trust.

pip install rote-cli works too if the user prefers a virtualenv.

rote compile runs an LLM agent, so it needs a driver: Claude Code (claude) or Codex (codex) installed and authed, or ANTHROPIC_API_KEY for the in-process api driver. The default claude driver deliberately scrubs ANTHROPIC_API_KEY and ANTHROPIC_AUTH_TOKEN from the child environment so the run bills against the user's Claude subscription rather than per-token API charges. Do not "fix" auth by exporting an API key. If the user explicitly wants API billing, pass --agent api.

4. Run the compilation

uvx --from 'rote-cli>=0.12.1' rote compile <skill-dir> --runtime <runtime> --out <out-dir>

Pick an out-dir the user will find, such as ./compiled/<skill-name> next to the source skill, and make sure it does not clobber existing work.

Set expectations before launching. This is not a quick command: a realistic skill takes roughly 13 minutes of wall clock and 30 to 40 agent turns on Sonnet. Run it in the background, tell the user you did, and poll rather than blocking the session.

If the run exits nonzero, check whether <out-dir>/compiled/pipeline.yaml exists anyway. The CLI recovers completed work from transient subprocess failures and says so in its output. Surface stderr to the user either way.

5. Report the result

Read <out-dir>/compiled/pipeline.yaml and <out-dir>/compiled/compile-report.md, then summarize:

1. Node-kind table. Count nodes per kind and say what each means here:

   Kind	Meaning
   pure_function	Deterministic code. The LLM is gone
   external_call	Direct API call with retry and timeout
   llm_judge	Typed LLM signature, kept but bounded
   agent_loop	Still agentic, because the input is genuinely unbounded
   hitl_gate	Durable human approval point

2. Codified fraction. How many nodes no longer need an LLM, which nodes are mandatory, and what each HITL gate blocks on.

3. Where things landed. <out-dir>/compiled/ holds the IR, extracted/, signatures/, and the report. <out-dir>/runtime/<runtime>/ holds the deployable code.

4. Next steps. The extracted/* modules are scaffolds that raise NotImplementedError. The user fills in real client code, then deploys the runtime output.

Be honest in this summary. A pipeline that came out mostly agent_loop means the skill was not as deterministic as it looked, and the user should know that rather than hear a success story.

6. Serve compiled pipelines back to Claude

rote serve is one MCP server exposing every registered pipeline as a callable tool. It triggers deployed workflows; it does not host them. The full flow:

rote compile -> deploy the runtime -> rote register -> rote serve -> call from Claude

Register the pipeline once the runtime side is actually running (a DBOS app in worker mode, a Temporal worker, or a deployed Cloudflare Worker):

uvx --from 'rote-cli>=0.12.1' rote register <out-dir>
uvx --from 'rote-cli>=0.12.1' rote register <out-dir> --runtime temporal
uvx --from 'rote-cli>=0.12.1' rote register <out-dir> --runtime cloudflare --url https://<worker>.workers.dev

This upserts ~/.rote/registry.json. Re-registering updates in place. After recompiling a changed skill, register again: DBOS and Temporal workflow names derive from the pipeline content hash and must stay in sync with the emitted code.

Then add the server:

claude mcp add --scope user rote -- uvx --from 'rote-cli[serve,dbos]>=0.12.1' rote serve

Each registry entry becomes two tools, or three on DBOS: <name> starts a run and returns {workflow_id, status: "started"} immediately, since compiled pipelines run for minutes to days; <name>_status polls a run by workflow_id; and on DBOS <name>_signal resumes a run parked at a HITL gate, so Claude can deliver approvals itself.

Two caveats worth stating proactively. A DBOS run stuck in enqueued means the emitted app process is not running against the registered system database. And while Claude Code picks up newly registered pipelines immediately via the server's list_changed notification, Claude Desktop and claude.ai snapshot tools at connect time, so a pipeline registered mid-session appears there only after a reconnect.

Reference

• Repository and docs: https://github.com/trevhud/rote (Apache-2.0)
• Package: https://pypi.org/project/rote-cli/