Generate ONEX Node

Status: SCAFFOLD-ONLY.

Every version of this document before 2.0.0 advertised "100% automated node generation", "ZERO manual work", and a BusinessLogicGenerator that "writes working code". None of that was true, and none of it was ever true in this repo. This skill is documented here as what it provably is: contract inference plus a template scaffold. Handler bodies are written by hand, or by a delegation stage that has not landed.

The overclaim mattered because it was load-bearing: other documents cited this skill as a working conversion pipeline. It is not one.

---

What actually exists today

Stage	Status	Where
Contract inference — derive a contract.yaml draft from existing node source	Live, AST-only, no LLM	omniintelligence/scripts/infer_contracts.py, whose own module docstring reads: "AST-only (no LLM calls). Scans Python files for classes inheriting from Node* base classes and generates contract.yaml drafts."
Template scaffold — stamp a node package from templates	Live	omnimarket/src/omnimarket/nodes/node_generate_node_effect/
LLM business-logic fill — write the handler body	DOES NOT EXIST here; live in omnimarket	nothing in omniclaude/src/ implements it. The omnimarket generation loop does fill bodies — see the live-run note below
AI quorum / multi-model consensus validation	DOES NOT EXIST	never ported; the --enable-quorum flag has no implementation behind it
Registration — the generated node is announced as an MCP tool	Live, but not gated on the node working	see below

What a live run actually produces (measured 2026-08-28, dev lane)

A real dispatch through omnimarket's node_generation_consumer was driven end-to-end on the dev lane and succeeded, so the honest account is more specific than "emits a shell":

• A handler body IS generated. The local model wrote correct, defensive Python for the requested task. This is not a stub.
• But it is not a conforming ONEX handler. The emitted body has the dict-in / dict-out shape handle(input_data), not the canonical typed signature handle(request: ModelX) -> ModelY.
• The emitted contract references models that were never generated. It names an input_model / output_model module that no run creates, and omits descriptor, handler_routing, node_version, inputs/outputs, and event_bus.
• Nothing is written to disk. The contract and handler exist only as strings inside event payloads. There is no importable package.
• Registration is not gated on the node working. The run registered the node as an MCP tool, and then its only sandbox invocation failed.

Net: a live run emits generated source text that is registered but not conforming, not staged, and not proven to run. An earlier documented attempt failed outright — its response did not satisfy the accepted code-output contract and no code was staged or accepted. Treat "the loop works" as meaning the dispatch path works, not the output is usable.

---

The documented entrypoint does not resolve

skills/generate_node/generate line 154 executes:

uv run omninode-generate "$prompt" "$@"

There is no omninode-generate console script. omniclaude/pyproject.toml's [project.scripts] declares only omniclaude-emit and omni-patterns; the command is absent from the project venv's bin/ and from PATH. The generate and regenerate wrappers in this directory are therefore not runnable as written — they fail at command resolution, after their prerequisite checks pass.

The components named in the old architecture diagram — ContractInferencer, HybridStrategy, BusinessLogicGenerator — have no implementation in omniclaude/src/. The only occurrence of those names in this repo is prose inside src/omniclaude/nodes/node_skill_generate_node_orchestrator/contract.yaml's description: field. They were never ported from the archived predecessor repo.

---

Where the working surface actually is

Node generation lives in omnimarket, not here, and is driven as a node dispatch rather than through this skill's wrapper scripts:

Node	Role
node_generation_consumer	The dispatch entrypoint. Declares runtime_dispatch.command_topic: onex.cmd.omnimarket.node-generation-requested.v1 with terminal success/failure events.
node_generate_node_effect	Template scaffold emission.
node_generated_code_validator	Validation of emitted code.
node_generated_node_publish_effect	Publication of the generated package.
node_rsd_fill_compute	A separate body-fill seam. Not the path a live dispatch takes — the fill measured above happens inside node_generation_consumer itself.

For contract inference specifically, call the AST tool directly:

# from the root of an omniintelligence checkout
uv run python scripts/infer_contracts.py --dry-run
uv run python scripts/infer_contracts.py --node NodePatternStorageEffect --execute

---

What you must still do by hand

After a scaffold run, expect to write:

• The handler body. This is the whole point of a node, and none of it is generated.
• Contract I/O models beyond what inference could derive from existing source. AST inference works from code that already exists; for a genuinely new node there is nothing to infer from.
• Tests. "Validation ensures quality" in the old text described a stage that does not write tests.

Budget this as writing a node with a pre-stamped skeleton — not as reviewing generated code.

---

Prerequisites

The generate wrapper checks for a workspace-root env var, uv, and OMNIBASE_INFRA_DB_URL before it fails on the missing CLI. The infrastructure the old document listed (PostgreSQL, Kafka/Redpanda, Consul) is what the omnimarket generation loop needs; it is not what makes this skill's wrapper work, because nothing makes this skill's wrapper work.

ZAI_API_KEY / ZAI_ENDPOINT / ZAI_MODEL are read by the regenerate wrapper's prompt-extraction path, which likewise terminates at the same missing CLI.

---

Honesty rule for this document

This skill sat in the stubs-behind-complete-docs failure class: a complete-looking document over a partial implementation. The rule going forward:

No capability claim in this file without a citation to code that implements it, or to a receipt that proves it ran. A percentage ("100% automated"), a duration ("10-25s per node"), or a quality score is a measurement; if no run produced it, it does not belong here. If the fill stage lands, this document changes in the same commit and cites the gate that proves it.

See Also

• omnimarket/src/omnimarket/nodes/node_generation_consumer/contract.yaml — the real dispatch surface.
• omniintelligence/scripts/infer_contracts.py — the AST contract inferencer.