File size: 9,600 Bytes
e4ab0d4 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 | # Node versioning β how to change a node without breaking saved workflows
> **β οΈ Scope β OPT-IN for Koolook nodes, MANDATORY for forks.**
> Backward compatibility is **not** a default constraint for Koolook-created
> nodes (the root `k_*.py` custom nodes); for those, **ignore everything below**
> and make the cleanest change β rename/drop/reorder freely β *unless* the
> maintainer says `check backward compatibility`. **For anything under
> `forks/`** (upstream ports/wrappers, including the Koolook-original nodes
> exposed via `SKIP_VERSION_SUFFIX`) **these rules still apply by default** β
> forks keep full back-compat discipline with no request needed. See
> [`CLAUDE.md`](../../CLAUDE.md) β *Change management*.
This is a hard-won list. ComfyUI's workflow format silently breaks in
several non-obvious ways when you change a node's inputs, outputs, or
class name. The rules below are what to follow whenever you touch a
**fork** node's `INPUT_TYPES` / `RETURN_TYPES` / `NODE_CLASS_MAPPINGS`,
or a **Koolook** node's once `check backward compatibility` has been
requested.
## How ComfyUI loads a saved workflow
When the frontend opens a workflow JSON, it reconstructs each node by:
1. **Looking up the node by its registered ID** (the key in
`NODE_CLASS_MAPPINGS`). If the lookup fails, the node renders as a
red "missing" stub and the workflow won't run.
2. **Mapping connections by input/output name.** Most cases.
3. **Mapping widget values by *position* in the widget array** β this is
the source of most breakage.
4. **Filling missing inputs with their default**, if they're declared in
`INPUT_TYPES["optional"]`. Missing *required* inputs β validation
error.
## What breaks vs what doesn't
| Change | Breaks saved workflows? | Why |
|---|---|---|
| Add a new input to `INPUT_TYPES["required"]` | **YES** | Old workflow has no value for it β validation fails |
| Add a new input to `INPUT_TYPES["optional"]` *at the end* | **safe** | Missing inputs use defaults |
| Add a new widget *between* existing widgets | **YES** | Widget array indices shift β values land in wrong widgets |
| Add a new widget *at the end* | mostly safe | Old workflows just don't have that index β uses default |
| Rename an existing input | **YES** | Saved value can't map by name |
| Rename the class / node ID | **YES** | Lookup fails entirely |
| Reorder existing combo options | **YES** for indexed serialization | Some workflows store combo as index, not string |
| Add new combo option *at the end* | **safe** | Existing string values still match |
| Remove an input | soft break | Saved value silently dropped, may lose semantic meaning |
| Change `RETURN_TYPES` order or types | **YES** | Downstream connections targeting `output[1]` now hit wrong type |
| Change a default value | "stealth break" | Old workflows have the old default baked in; new workflows get the new default β silent behavior divergence |
> "*Field gets messed up*" almost always = **widget position shift**.
> ComfyUI serializes widgets as `widgets_values: [val0, val1, val2]`.
> Insert a new widget at index 1 β every subsequent value shifts one
> slot β wrong values land in wrong widgets.
## The five rules
1. **Never rename a registered node ID.** When you really must, register
both old and new IDs pointing to the same class for a few releases
(deprecation alias), then remove the old one with a CHANGELOG note.
2. **Append new inputs/widgets to the END.** Never insert in the middle.
Order in `INPUT_TYPES` dict matters because Python preserves insertion
order and the frontend uses it.
3. **New inputs go in `INPUT_TYPES["optional"]` with a default**, not in
`required`. Old workflows that don't have the field will use the
default; new workflows can wire it up.
4. **Never reorder combo options.** Append new options. If you need to
remove one, leave it as a hidden alias or deprecation stub for a
release first.
5. **Treat `RETURN_TYPES` as immutable.** If you need to change outputs,
make a new node ID β `MyNodeV2`. The Comfy ecosystem (KJ Nodes, Was
Node Suite, Crystools) does this constantly.
## When you can't avoid a breaking change
| Pattern | Example |
|---|---|
| **Suffix-version a new node** | `EasyAIPipeline` β `EasyAIPipelineV2`. Both register, both work. |
| **Mark old as deprecated in display name** | Old: `Easy AI Pipeline (deprecated, use V2)` β display name change is fine, ID is what matters |
| **Document migration in CHANGELOG** | "Old `EasyAIPipeline` retained; new pipeline work should use `EasyAIPipelineV2` which adds X/Y/Z" |
| **Keep deprecated nodes for β₯2 minor versions** before removing |
This is exactly the pattern used for `EasyResize` β `EasyResize_Koolook`
in v0.1.6 (see [`../../CHANGELOG.md`](../../CHANGELOG.md)). Both IDs
register, both load, the bare-name one's display says "deprecated", and
the bare-name one will be removed in a future major release once the
deprecation has had time to propagate.
## Concrete pattern β the rename + alias
```python
# In whichever k_*.py file holds the class:
class MyNode:
""" ... existing implementation, unchanged ... """
# ...
# Register BOTH IDs against the same class:
NODE_CLASS_MAPPINGS = {
"MyNode_Koolook": MyNode, # new canonical ID
"MyNode": MyNode, # legacy alias (deprecated)
}
NODE_DISPLAY_NAME_MAPPINGS = {
"MyNode_Koolook": "My Node (Koolook)",
"MyNode": "My Node (deprecated, use 'My Node (Koolook)')",
}
```
That's it. Saved workflows that reference `MyNode` keep loading. New
workflows pick `MyNode_Koolook` from the search. The bare-name display
nudges users to switch.
## Concrete pattern β adding a new optional input
Suppose `EasyAIPipeline` currently has these inputs:
```python
"required": {
"shot_duration": ("INT", {...}),
"seed": ("INT", {...}),
"base_path": ("STRING", {...}),
"shot_name": ("STRING", {...}),
"ai_method": ("STRING", {...}),
"version": ("INT", {...}),
},
```
You want to add a new `pass_name` field. **Do not** add it to `required`
between existing fields β that breaks every saved workflow. Instead:
```python
"required": { ... unchanged ... },
"optional": {
"pass_name": ("STRING", {"default": "", "tooltip": "..."}),
},
```
- Old workflows that don't have `pass_name` use the default `""`.
- New workflows that do have it use the value the user sets.
- Output behavior must remain backward-compatible when `pass_name=""`
(i.e. equivalent to the pre-change behavior). If `pass_name=""` would
change the output filename, you need a `MyNodeV2` instead.
## Concrete pattern β output change
Suppose `EasyAIPipeline` currently returns `(output_path,)` and you want
to add a `(output_path, output_dir)` tuple. That's a `RETURN_TYPES`
change β downstream connections targeting `output[1]` would hit the new
output unexpectedly (or just exist where they didn't before).
Don't modify the existing class. Make a new one:
```python
class EasyAIPipeline_V2(EasyAIPipeline):
"""Same as EasyAIPipeline, plus emits the resolved output directory."""
RETURN_TYPES = ("STRING", "STRING")
RETURN_NAMES = ("output_path", "output_dir")
# ... reuse or override the execute method as needed ...
NODE_CLASS_MAPPINGS = {
"EasyAIPipeline": EasyAIPipeline, # unchanged
"EasyAIPipeline_V2": EasyAIPipeline_V2, # new
}
```
## Anti-patterns to refuse β *for forks, or after `check backward compatibility`*
These are anti-patterns for **any fork node**, and for a **Koolook node only
once the maintainer has asked** to preserve backward compatibility. For a
Koolook node with no such request, every one of them is the *correct* clean
change β append-vs-insert is the lone exception below.
- β "I'll just add the field in the middle for clarity, the dict order
doesn't matter." It does. Append. *(This one is worth following even
without a back-compat request β a mid-list widget insertion silently
scrambles the `widgets_values` of any workflow already on disk, including
your own dev graphs, for zero benefit. Appending is free.)*
- β "I'll rename the input to be clearer." For a fork node, or any node
under a back-compat request: add a new optional input with the clearer
name and deprecate the old one in display only. For a Koolook node with
no request, **just rename it.**
- β "It's a small change, no need to bump the node version." For a fork
node, or under a back-compat request, a surface change that could break a
saved workflow β new node ID. For a Koolook node with no request, change
in place and note it in the CHANGELOG.
- β "Nobody's using the old node yet." For a fork node, or under a
back-compat request, you can't assume that. For a Koolook node with no
request you *do* assume it β clean change, move on.
## Tooling β what we have, what's still missing
- β
**Workflow-fixture smoke test** β `tools/preflight_release.py`
check `workflows` walks `tests/workflows/*.json`, extracts node IDs
by both heuristic and AST-extracted set, and reports any reference
that no longer resolves. The `preflight-release` skill in
`.claude/skills/` runs all four checks before a release.
- β³ **`node-api-change` agent skill** β a check that would run before
any edit to `INPUT_TYPES` / `RETURN_TYPES` / class names, diff
against the rules above, and refuse unsafe changes. Same shape as
`license-pre-check`. Not built yet; scoped for the next time someone
changes a node's input/output surface.
|