[Silthus]

Problem

The runtime from the layers below is unreachable: no builder surface offers the Generate text node, so an author cannot configure a prompt, declare output fields, or see what a run would produce.

Layer 3 of the Generate text stack. It ships the builder panel.

Changes

• Adds an AI category to the workflow builder, gated on the new workflows-llm-action flag, with one node: Generate text.
• The panel renders the prompt (existing templating input, untouched), output-field rows, and a model picker pinned to the backend's six ids.
• Picking a variable for an output field derives an output_variable mapping with result_path: fields.<key>. Field edits regenerate only the rows they own; hand-added rows, including result_path: text for the raw generation, survive. One exported predicate decides ownership for both the regeneration and the "Set by this step" tag, so the tag and the delete rule cannot disagree.
• Run preview replaces Pick from response on this template: it runs the real prompt against the real model (the tooltip says it spends AI credits), keeps output_variable in the dispatch, and renders generated text, extracted fields, and resulting variable values. Preview expiry after 25 seconds renders as a notice, not a node failure.
• The node defaults to on_error: abort, carried through node creation, so a billed but failed generation never continues into a send with empty variables.
• A second preview click while one is in flight is blocked; the button resets on success and error.

No screenshot: the panel needs a loaded workflow plus a seeded template-workflow-llm, which sync_hog_function_templates does not provide in this environment, and no Storybook story covers the workflow editor panel. The panel is not rendered anywhere until the flag is on; reviewing it live behind the flag is the verification this PR could not do itself.

How did you test this code?

• Registry test: the category registers behind the flag with the right template id and on_error: abort.
• Logic tests: derived-mapping regeneration preserving hand-added rows, the collision between the tag predicate and the delete predicate (the review's blocker), preview response splitting, mock_async_functions: false dispatch, stale-preview drop after switching nodes, the double-submit guard, and the deadline-versus-skip notice.
• Model pin test: ids, order, and default match the backend list.
• on_error carry-through is driven through the real drop path, not asserted on the registry entry.
• Not run: visual verification, as above.

Automatic notifications

• Publish to changelog?

Docs update

Docs, skills, and MCP updates ship in the next layer of this stack.

🤖 Agent context

Autonomy: Human-driven (agent-assisted)

Same orchestration: Opus 5 TDD implementation, three Fable 5 review lenses (12 findings, one blocker: a listener that wrote output_fields to the workflow and immediately reverted it through a second persist path), Opus fix pass, Fable confirmation. Skills invoked: /tdd, /implement, /writing-tests, /writing-user-facing-copy, /adopting-generated-api-types (review lens), /writing-pr-descriptions, /stacking-prs. Session decisions: all business logic lives in a kea logic per repo rule; preview-expiry detection matches the runtime's exact strings rather than a structural guess after a reviewer showed the structural rule mislabeled filter-skipped previews.