[endrix]

Stage 3, last step for the agent stack (P5e). wfpy-ide now holds the agent workflow UI and asks its own host for skills and presets (endrix/wfpy-ide#24, #26, #27, all merged). This deletes the platform's copies, including the original reason for the whole move: every product's diagram load scanned the skill and agent directories and could rebuild the model a second time.

Removed

• Server:
  • The per-load agent enrichment in WorkflowSourceModelStorage, about 400 lines:
    • the warm-or-defer scan of .wf/skills, .claude/skills, ~/.claude/skills and the .claude/agents directories, run on every diagram load of every product;
    • the wf:agentSkillsSnapshot / wf:claudeAgentsSnapshot snapshots on the root, and the skill status on @agent nodes;
    • its cache, and the second model refresh it dispatched when the scan found something.
  • The two discovery operation handlers (request-agent-skills-handler.ts, request-claude-agents-handler.ts), their registrations and their tests.
  • The refresh-agent-enrichment- polling prefix.
• Client:
  • the discovery requests, their answers and the answer handlers' caches in editing-action-handlers.ts;
  • src/agent-ui/ (agentUiModule), its 32 tests, which moved to wfpy-ide, and its path-import fixture.
• Toolkit tests: the enrichment-deferral test, and the applyAgentEnrichment stub seven other tests carried.

54 files changed: 178 lines added, 6,856 deleted.

Kept

• The contribution points (AnnotationEditor, NodeSectionContribution, AnnotationTemplate). Their tests now use a neutral panel harness (test/panel-harness/) and their own probes instead of the agent UI. The template tests' quoting cases moved to wfpy-ide with the templates.
• The cal:agentSkillStatus metadata key, because wfpy-ide's Skill row still writes it.

Review (/code-review medium), addressed

The review found no correctness bug in the deletions, and nothing left in the tree references the removed handlers, action kinds, caches or agentUiModule. Fixed:

• The load breadcrumb loses its agent-enrichment phases: the skillStatus/skills/agents buckets, setEnrichment and the enrichment= field, which would only have read 0 or n/a. Its test pins that they are gone.
• Docs that described the deleted scan: both READMEs' debugging note (including the one shipped in the VSIX), the caller-redelivery guard's comment, and a toolkit test's comment.
• CHANGELOG:
  • The release note names feat: the stock bundle no longer loads the agent workflow UI #137 and refactor: the platform no longer discovers agent skills and presets #139 among the changes that need the 0.9 bump.
  • The contribution-points entry no longer credits the platform with fixes to code that has left it.
• Tests:
  • The neutral harness drops the viewer-editor bridge defaults, which only the @viewer template used.
  • A node-section test that now passed trivially is gone.
  • A new test pins that a redraw a contribution asks for tells the annotation editors not to request discovery again. Only wfpy-ide's tests checked this after the agent UI left.
• Noted, wfpy-ide follow-up: its snapshot seeding is now dead, and its Skill row reads "not discovered" until the host answers.

Release

This is breaking for the API: the discovery actions and agentUiModule go, though agentUiModule was never released. It belongs to the 0.9 release, whose note [Unreleased] already carries.

Tests

• Local results:
  • typecheck: 5/5
  • neutrality: 5/5
  • build: OK
  • diagram-client: 410/410
  • diagram-server: 443/443
  • extension-core: 335/335
  • sidecar-toolkit: 357/358, the one failure being the known local timeout in variant-node-create-wizard, which passes in CI
• wfpy-ide main against this tree: 138/138, bundle build included.

https://claude.ai/code/session_01EToj6RJRMSSHQiqGNMU9wv