Node authoring
Preview. Full cross-mode interoperability and real user paths are still being validated.
Entry: Longform → Workbench → Node authoring (长篇 → 长篇工作台 → 节点创作). Nodes place inputs, generation steps, and intermediate results on a canvas for inspection and composition.
When to use it
Once familiar with step-by-step longform, try a small workflow to break down a generation, reuse intermediate materials, or inspect upstream/downstream inputs.
For daily writing, you can continue using step-by-step longform.
Recommended trial
- Save a full backup and choose a small task.
- Check input requirements of official nodes/templates and connect required materials.
- Execute incrementally, checking actual inputs and outputs at each step.
- Use permitted candidate-confirmation entries to write to the official work.
- Return to the step-by-step interface and check the official result.
Official nodes reuse longform capabilities, works, and memory. Experimental-node drafts cannot directly become official settings.
Current limitations
A connected graph does not guarantee semantic correctness. The preview canvas is not a second independent work database. Correct invalid connections, stale inputs, or missing permissions as prompted before execution.
Continue: AI candidates and recovery.
Two node entry points
Node authoring in the longform workbench uses formal domain nodes. The workflow editor in the prompt library provides This node's template (本节点模板), material bindings, and AI hints (给 AI 的提示). These are separate entries: updates to prompt workflows do not establish that the domain canvas was updated.
Prompt workflows can use Forge Writing Workshop templates. Specify variables when connecting upstream material; a connection alone does not supply missing inputs.
Node maintenance and recovery
The following changes are delivered through PR #115. Use its full checks and merge state to assess delivery, and verify online deployment separately:
- Single-field nodes offer generation requirements (生成要求). The node library supports search, with experimental drafts hidden by default.
- Run to this node (运行到此节点) handles the selected node and its required ancestors. Other unfinished branches can remain drafts.
- Candidate edits and version selections save automatically. Mode changes wait for saving; failures retain input. Rejecting a candidate does not write to the work. Adopted and rejected results remain read-only.
- Nodes with a fixed target character offer adoption into that same character, visible in step mode. Changes to the original content or target prevent old candidates from overwriting it.
- After an interrupted refresh, recover saved results first. Recovery itself does not call a model; continuing production is a separate action.
- On phones, the library, canvas, node settings, and candidates/evidence have separate views. Switching views retains edits.
Automated tests use isolated model responses and do not establish literary quality. Full cross-mode acceptance of every official longform template remains pending. See development status.
Free graph editing and prompt selection
These operations are node preview improvements in PR #115; the online version depends on the corresponding deployment result.
- Click or drag a node from the library, or right-click/double-click empty canvas space to search and create one.
- Drag backward from an input port or forward from an output port. Release on empty canvas space, search compatible nodes and select one: the chosen port is connected automatically. You can also drag directly to a compatible port on an existing node.
- Select a node and disconnect an edge under Connected edges (已连接的线) before reconnecting it. Escape cancels the picker. Canvas controls zoom in/out or restore 100%.
- Drag from a generation node's Prompt input to add a Prompt template (Prompt 模板) node. Search built-in or personal templates, adjust template parameters, fill required variables and supplementary instructions, or expand the node-local prompt editor.
- The selected version and adjustments save with the graph and do not change global templates. Click Save (保存), confirm Saved (已保存), then refresh. Required inputs must be supplied before model calls.
- Temperature, output-token limits and candidate-count controls take effect through their corresponding ports. Rerun after upstream changes; old candidates cannot be adopted over new work.
The picker only lists compatible registered nodes. Experimental drafts require explicit opt-in and cannot write directly into the formal work.