Skip to content

Edit your model ​

Edit supported constructor parameters, insert a Block or review a deletion directly from the canvas, or ask for a Python change in Canvas chat. Both routes produce a source draft for the same review and Accept decision.

Make a proposal ​

  1. Open your model and select the layers you want to change.
  2. Describe the change in Canvas chat. The agent uses the current model and selection as context.
  3. Review each proposed file in the native VS Code diff. Ask for revisions if needed; the draft keeps its other pending files.
  4. Choose Accept/Apply or Discard. Discard removes the whole draft.

Apply makes one undoable editor change, saves only the accepted files, and opens the model with its current saved setup. Preparing or revising a proposal does not save files or run the model. A failed save or intervening edit stops execution.

Canvas chat keeps the agent connection across replies and minimizing. Stop or a reconnect retains bounded graph evidence at its original revision; New chat clears it. Supported models default to medium effort, preserving your saved choice. Background preparation can use your agent’s provider before Send.

What to check ​

  • Shared definitions can affect more than one instance.
  • Grouping layers can change parameter paths and checkpoint keys.
  • A source review does not prove that the changed model will run with your inputs.

Apply rechecks every reviewed file against its original digest. Reopening a file or changing its editor version without changing its bytes does not block acceptance. If reviewed source bytes change, Apply keeps the draft open, clears the pending Accept indicator, and shows the refusal reason. Review the proposal again after its source or dependencies change.

Source positions and review limits ​

Navigation and edits account for non-ASCII text and emoji, including a final line without a newline. Form feeds and Unicode line separators above a class no longer shift edits by a line. Edits preserve CRLF line endings. A verified edit to a Sequential child targets that child.

Every replacement names its file and current digest. The complete review inventory, including unchanged dependencies, is limited to 256 files and 8,000,000 bytes; oversized reviews are refused without truncation. All proposed files share one review decision and one undoable edit.

When available, Run reviewed candidate uses the retained capture setup. Apply uses the current saved setup. Candidate checkpoint or input overrides do not update the saved profile. A checkpoint-load result is separate evidence from a successful forward.

Captured edits check source limits against the selected model file and all files referenced by captured source bindings. Unrelated Python files collected from the workspace do not disable editing. The relevant files must still fit the source budgets and match their captured digests.

In the local browser ​

Use Canvas chat to request a proposal. Nodes link to their source instead of opening inline editing forms. The Canvas agent can propose changes across up to 16 existing, graph-recorded workspace Python files; external MCP clients retain their supported propose_edit recipes.

Compare changes · Open and run the model

Supported transforms and source safety rules

Source editing reference covers captured values, grouping, proposal ownership, and host APIs.