Claude Code 5090：推送 Workflow 与节点基板后端交付并回填交接 #11

Activity

1. 2233admin commented on Jul 12, 2026

   

   2233admin

   on Jul 12, 2026

   OwnerAuthor

   More actions

   Pushed: branch feat/workflow-persistence-closed-loop, commit c326a2ace5882d226f1753bf65df0560b33209af.

   What's done

   • Persistence closed loop: Workspace -> Project -> WorkflowDraft -> immutable WorkflowVersion. Draft updates require expected_revision (409 on stale write). Publish requires a passed ValidationRun; each validation run is single-use (409 if already consumed, 409 if the draft moved on since the run was taken).
   • Dify/n8n import: n8n's connections dict (keyed by node name, {main: [[{node, type, index}]]}) is flattened into the same {id, source, target} edge shape everything else uses. Imported nodes get tagged with ui.externalWorkflow (runtime, graph name, raw snapshot) and, for n8n, ui.n8n provenance. Import is a preview -> Validation Run -> publish gate, same as native drafts.
   • Recursive Package compiler: nodes can nest packages inside packages. Scoped ids look like a::b::c. Hard cap at depth 16 (package_nesting_limit_exceeded past that). Cycle detection runs independently per nesting level. Depth-1 behavior is byte-identical to before this change (regression-tested).
   • Compatibility tracer + validation execution: reuses the existing WorkflowRun/event infra rather than a new subsystem. run_validation compiles, executes, and either strict-matches an expected-event fixture or does a loose passthrough check.
   • WorkspaceSettings: timezone, deterministic_simulation, max_items_per_run, with migration + CRUD API.

   New router: backend/api/v1/workspaces.py. New models: backend/models/workflow_authoring.py, backend/models/validation_run.py. Migration r5s6t7u8v9w0 (head, on top of m2n3o4p5q6r7).

   Test commands actually run

   • uv run pytest -m "not live" -q -> 1636 passed, 5 failed, 1 skipped, 10 deselected. The 5 failures are a pre-existing Windows/GBK locale codec error unrelated to this branch (backend/api/v1/nodes.py install-script endpoint reading a non-ASCII byte in scripts/install-agent.sh under a GBK locale) — none of the failing tests touch any file this branch modifies.
   • uv run pytest tests/integration/test_workflow_authoring_api.py -v --no-cov -> 6 passed.
   • uv run pytest tests/integration/test_workflow_external_import_api.py -v -> 3 passed.
   • uv run alembic upgrade head against a scratch SQLite DB -> applies cleanly.

   Honest limitations / things Codex should know before continuing

   1. The compatibility runtime is simulated, not real Dify/n8n worker dispatch. Import converts an external graph into native OpenCLI capability nodes and runs them through the existing compiler/tracer. There's no live Dify or n8n execution engine wired in anywhere — "compatibility" here means shape/provenance compatibility on import, not runtime interop. If real interop is wanted, that's a separate, much bigger piece of work (actual Dify/n8n API clients, credential handling, async job polling).
   2. n8n-origin classification is reachable but not exercised by the importer. resolve_node_origin() checks catalogId before ui.n8n, and the importer always assigns a recognized catalog id (merge/normalize/fallback). So importer output always classifies as node_library origin, never n8n, even though the node carries ui.n8n provenance. The n8n branch itself works (unit-tested directly), it's just dead code from the importer's current call sites. Worth a product decision: should imported n8n nodes classify as n8n origin for UI/filtering purposes? If so, resolve_node_origin precedence needs to flip, or the importer needs to stop assigning a catalog id for genuinely-unmapped types.
   3. Package nesting below depth 1 doesn't get tracer lifecycle events. _package_metadata (used for package-lifecycle events in the tracer) only runs on top-level project.nodes. Nested packages compile and validate correctly at any depth up to 16, but flatten into plain leaf nodes for run/trace purposes — no nested-package-entered/exited events. Deliberate scope cut, not a bug, but should be called out before anyone assumes trace output reflects real nesting depth.
   4. Coverage gate: the repo-wide pytest coverage gate sits around 90% currently; the new code is 86-95% covered per module (see services/workflow_authoring_service.py at 86%, external_importer.py at 91%) — mostly IntegrityError/race-condition branches not exercised, not missing feature coverage.
   5. Nothing here touches the frontend. This is backend/migration/test only, on top of latest main at the time of push (6e09973).

   Handing off — Codex, please pick up from here.
2. 2233admin

   added a commit that references this issue on Sep 2, 2026

   docs(verification): record release issue evidence

   ...

   24a0219