[endrix]

P5e, the @viewer step: the next wfpy-only piece of the client, behind a fourth contribution point (after annotation editors, node sections and wizard templates). Nothing changes yet.

Why

The double-click listener (viewer-mouse-listener.ts) did two things:

• for every product: open a node's definition (preview), unless the network navigation opens something for it;
• for wfpy only: for a task with a @viewer annotation, open its last tokens instead (open, diff, open with an editor, run a command), or a path the annotation declares.

What

• NodeOpener (src/node-opener.ts, exported). An opener has an id, and its open(node, { sourceUri }) answers a double-click with the actions to run, or undefined to leave the node to the platform. The listener multi-injects the openers and keeps one per id (the last binding wins, via lastPerKey). It asks them in the order their ids were first bound, before its own definition navigation, and the first that answers wins.
• The @viewer half moves verbatim into src/viewer-node-opener.ts (ViewerNodeOpener, id viewer), with its helpers and wf:viewer* argument names. Git's moved-code detection marks 33 of its 372 lines as new: the opener's own guard, its context lines, and the imports. It emits the same NavigateToExternalTargetActions, which Dialogram's host still turns into an editor, a diff or a command.
• The stock module binds it, so nothing changes. Next:
  1. wfpy-ide binds its own copy, as it did for the agent UI;
  2. the stock module stops binding this one.

Review (/code-review medium), addressed

The review confirmed the moved code is equivalent: the same order of checks, the same nodes reaching @viewer (it comes down to the type alone), the same messages and routes, and no change for actor or network nodes. Fixed:

• An empty answer is no answer. [] used to stop the loop and block the definition.
• An opener that throws is skipped, with a warning, instead of making the node open nothing.
• The contract says:
  • an opener without an id is dropped;
  • an opener's answer, unlike the platform's definition, does not stand back for a node whose first view the network navigation opens.
• One definition branch covers a task, or an actor not marked external. A test pins the marked-external actor.
• Comments:
  • ABSOLUTE_URI's explanation moved with it.
  • The stock module's line for the listener now describes what it does for every product.

Tests

• The three @viewer test files (viewer-declared-target, declared-target-resources, double-click-opens-one-thing) inject what the stock module binds and pass unchanged.
• New test/node-opener-seam.test.ts (9 tests):
  • an opener answers first, with the node and the diagram's source;
  • when it answers nothing, the definition opens;
  • it is asked for actors too, and the first answer wins;
  • a later opener for an id replaces the earlier one in its place;
  • through a real container: the stock viewer opener, a product's after it, a product's viewer replacing it, and none when nothing is bound.
• Container parity: the baseline accepts exactly the one NodeOpener constant binding.
• Local results (after merging refactor: the platform no longer discovers agent skills and presets #139):
  • client suite: 419/419
  • typecheck: 5/5
  • neutrality: 5/5

https://claude.ai/code/session_01EToj6RJRMSSHQiqGNMU9wv