[eric] workflows: a deleted workflow stays deleted, so the immortal 45-minute ghost can't write itself back to life

backend/apps/workflows/executor.py

@@ -213,9 +213,13 @@ async def execute(
     # of a toggled-off workflow running itself. Turn it back on to run it.
     p_live = storage.get_workflow(wf.id)
     p_refusal = None
-    if p_live is not None and p_live.deleted_at is not None:
+    if p_live is None:
+        # A hard delete leaves nothing to look up, and reading that as "no objection" is how a
+        # deleted workflow still ran to the end and then wrote itself back to life.
         p_refusal = "Workflow deleted"
-    elif p_live is not None and not p_live.schedule.enabled:
+    elif p_live.deleted_at is not None:
+        p_refusal = "Workflow deleted"
+    elif not p_live.schedule.enabled:
         p_refusal = "Workflow is paused"
     if p_refusal is not None:
         p_skipped = WorkflowRun(

backend/apps/workflows/storage.py

@@ -36,6 +36,13 @@ _missed_cache: list[MissedRun] = []
 _cache_loaded = False
 _paused = False
 
+# Ids deleted during this process's life. The cache hands out SHARED Workflow instances, so a run
+# already in flight when the user deletes still holds one and writes it back when it finishes, and
+# save_workflow used to recreate the file AND the cache entry, fully scheduled: the workflow rose
+# from the dead every time, which is exactly the "it never dies" field report. Only needs to live in
+# memory, because after a restart nothing holds a stale instance to write back.
+p_deleted_ids: set[str] = set()
+
 
 def _resolve_host_tz_name() -> str:
     """Best-effort IANA name for the host. Mirrors apps/service/client.py."""
@@ -158,8 +165,18 @@ def get_workflow(wid: str) -> Optional[Workflow]:
     return _workflow_cache.get(wid)
 
 
-def save_workflow(wf: Workflow) -> Workflow:
+def save_workflow(wf: Workflow, untrash: bool = False) -> Workflow:
     with _io_lock:
+        if wf.id in p_deleted_ids:
+            logger.info("ignoring a write-back for deleted workflow %s", wf.id)
+            return wf
+        # Trash is one-way too: only /restore passes untrash. Any other save carrying an older copy
+        # (a run that started before the user hit delete) would otherwise clear deleted_at and put
+        # the workflow back on the page with its schedule re-armed.
+        prior = _workflow_cache.get(wf.id)
+        if not untrash and prior is not None and prior.deleted_at is not None and wf.deleted_at is None:
+            logger.info("ignoring a write-back that would untrash workflow %s", wf.id)
+            return wf
         _ensure_dirs()
         _workflow_cache[wf.id] = wf
         p_atomic_write_json(_wf_path(wf.id), wf.model_dump(mode="json"))
@@ -171,6 +188,9 @@ def reload_workflow(wid: str) -> Optional[Workflow]:
     instances, so a handler that mutated one and then failed must roll back through here or the
     unsaved change lingers until any later save persists it by accident."""
     with _io_lock:
+        if wid in p_deleted_ids:
+            _workflow_cache.pop(wid, None)
+            return None
         path = _wf_path(wid)
         if not os.path.exists(path):
             _workflow_cache.pop(wid, None)
@@ -184,6 +204,7 @@ def reload_workflow(wid: str) -> Optional[Workflow]:
 def delete_workflow(wid: str) -> bool:
     with _io_lock:
         existed = wid in _workflow_cache
+        p_deleted_ids.add(wid)
         _workflow_cache.pop(wid, None)
         _runs_cache.pop(wid, None)
         wf_file = _wf_path(wid)
@@ -217,6 +238,8 @@ def list_all_runs(limit: int = 200) -> list[WorkflowRun]:
 
 def record_run(run: WorkflowRun) -> WorkflowRun:
     with _io_lock:
+        if run.workflow_id in p_deleted_ids:
+            return run
         _ensure_dirs()
         arr = _runs_cache.setdefault(run.workflow_id, [])
         # Replace prior entry with same id if we're updating an in-flight run.

backend/apps/workflows/workflows.py

@@ -906,7 +906,7 @@ async def restore_workflow(workflow_id: str):
     if not wf or wf.deleted_at is None:
         raise HTTPException(status_code=404, detail="Workflow not in trash")
     wf.deleted_at = None
-    storage.save_workflow(wf)
+    storage.save_workflow(wf, untrash=True)
     enriched = _enriched(wf)
     try:
         from backend.apps.agents.core.ws_manager import ws_manager
@@ -1453,6 +1453,10 @@ async def run_workflow_now(workflow_id: str, body: Optional[dict] = None):
     wf = storage.get_workflow(workflow_id)
     if not wf:
         raise HTTPException(status_code=404, detail="Workflow not found")
+    # The executor refuses a trashed workflow but writes no history for it, so without this the caller
+    # got run_id "" with a null status and no idea why nothing happened.
+    if wf.deleted_at is not None:
+        raise HTTPException(status_code=409, detail="This workflow is in Trash. Restore it to run it.")
     # executor.execute() owns the run record. Don't pre-create a stub here or we end up with two rows per manual fire (one orphan "running" row from this handler plus the real one from the executor).
     pre_ids = {r.id for r in storage.list_runs(wf.id, limit=10)}
     tested_signature = body.get("signature") if isinstance(body, dict) else None