[Silthus]

Problem

An agent building a workflow through the skill or MCP tools has no way to learn that the Generate text step exists, that its output fields need matching output_variable mappings, or that running it spends AI credits. A skill one PR behind the feature is a published document telling agents a paid operation is free, which is why these files ship in the same stack.

Top layer of the Generate text stack.

Changes

• graph-schema.md: a worked example for template-workflow-llm carrying the three non-discoverable facts. Output field keys must name declared variables. Each field needs an output_variable entry with result_path: fields.<key>, or the node runs, bills, and discards. result_path: text reaches the raw generation.
• graph-schema.md: the output_variable bullet now documents the array form and the label field, and the 5120-byte variable cap is stated where variables are described.
• SKILL.md: the description gains an AI/LLM trigger so the skill loads for this node, and the test-run step warns that mocking off spends AI credits.
• lifecycle-and-debugging.md: corrects the claim that the catalog never lists fixed-template steps. The authenticated catalog does list them; only the public mount strips hidden templates.
• tools.yaml: the test-run tool description names the Generate text step and its credit cost, with the two generated schema artifacts regenerated in the same commit.
• CONTRIBUTING.md: fixes a stale cyclotron.ts path, walks through the llmGenerate queue-parameter routing including the v2 wake path and backup ladder, and adds a short section on the per-template panel branch.

Every claim was written against the shipped code in the layers below and cite-checked by a second reviewer against the same files.

How did you test this code?

• Skill lint (build_skills.py --lint): 141 skills pass, no advisories in building-workflows.
• MCP tool generation reran clean; only the two schema JSON artifacts changed, and the reviewer confirmed workflows.ts carries no descriptions and is not stale.
• Not run: hogli build:skills through Django (needs the dev database; fails identically on a clean tree).

Automatic notifications

• Publish to changelog?

Docs update

This PR is the docs update for the stack. The posthog.com page is a named follow-up in the other repo.

🤖 Agent context

Autonomy: Human-driven (agent-assisted)

Opus 5 wrote the five files reading every claim from the shipped code, two Fable 5 reviewers checked factual accuracy against the source and writing quality (9 findings, including one factual overstatement about variable defaults), an Opus fix pass resolved them, and a Fable confirm pass verified each against the code. Skills invoked: /writing-user-facing-copy, /writing-pr-descriptions, /stacking-prs.

v2 revision (Fable 5, single session, human-directed): the queue-routing section now documents the wake event and the 1/3/5-minute backup marks that replaced the 2-second poll.