Fix native workflow alerts in chat proposals and AI Assist

application/single_app/config.py

@@ -101,7 +101,7 @@
 EXECUTOR_TYPE = 'thread'
 EXECUTOR_MAX_WORKERS = 30
 SESSION_TYPE = 'filesystem'
-VERSION = "0.261.314"
+VERSION = "0.261.315"
 IS_DEVELOPMENT = is_development_env_enabled()
 
 # Opt-out for deployments where App Service Easy Auth is active but the platform

application/single_app/functions_orchestration_planner.py

@@ -310,6 +310,42 @@
   words such as "this week" can stay relative.
 - optionally "alerts" {"mode":"every_run"|"failures_only","severity":"info"|"low"}, "run_as"
   "self"|"none", and "durable": true.
+  Those legacy alert modes are quiet, bell-only run-status notifications. For native conditional
+  monitoring alerts instead use "alerts":{"mode":"rules","rules":[...]} (1 to 20 rules).
+  Each rule has "severity":"info"|"low"|"medium"|"high"|"critical" and "condition"; optionally
+  "name", "enabled" (boolean), "delivery":"default"|"notify_only"|"popup",
+  "require_acknowledgment" (boolean), "sound":"off"|"once"|"repeat", "size":"small"|"medium"|"large",
+  and "scope":{"type":"final"|"any_task"} or {"type":"task","task":N}, where N is the 1-based
+  position of a task in this blueprint, not a stored id. Default scope is final.
+  Conditions: {"type":"run_status","statuses":[...]} (completed, failed, cancelled,
+  completed_with_task_errors); {"type":"task_status","statuses":[...]} (succeeded, failed);
+  {"type":"text_match","mode":"contains_any"|"contains_all"|"not_contains","values":[...],
+  "case_sensitive":false}, or {"type":"text_match","mode":"regex","pattern":"..."};
+  {"type":"model_evaluation","prompt":"..."}; {"type":"agent_signal","signal_name":"...",
+  "min_severity":"info"|"low"|"medium"|"high"|"critical"}; {"type":"no_output"};
+  or {"type":"file_sync","outcome":"changes_found"|"no_changes"|"sync_failed"}.
+  run_status, file_sync and agent_signal conditions must use final scope.
+  Native alerts require no email action or Microsoft 365 connection. Normally info/low go to the
+  bell, medium/high/critical pop up. Acknowledgment, sound and larger size require pop-up delivery;
+  repeat sound requires acknowledgment. These settings respect device sound preferences.
+  When the user explicitly requests strongest/highest attention, propose critical, delivery popup,
+  require_acknowledgment true, sound repeat, size large, unless their constraints say otherwise.
+  Disclose these attention options for approval; never make them the default for ordinary alerts.
+  A personal workflow alerts its owner, not arbitrary recipients or everyone in a group.
+  Monitor external data using an offered read-capable agent, then alert inside SimpleChat on adverse
+  output. This is different from changing an external system's alarm configuration: do not request
+  external writes unless that is what the user asks. A successful monitoring run may find failures.
+  run_status/task_status failures detect workflow execution errors, not domain problems in telemetry.
+  Prefer model_evaluation for semantic findings, explicitly excluding healthy/negated findings
+  ("no active alerts"); text_match is for defined literal signals, and agent_signal needs an agent
+  known to produce supported signals. Do not infer signal support from an agent's name.
+  For example, ongoing vehicle monitoring can read telemetry and use a critical model_evaluation
+  rule for actual active faults, while "email me the findings" needs a mail-capable agent.
+  "Configure the source system's alarms" is an external configuration request, not native alerting.
+  Use the native rules for "create a workflow that monitors and alerts me"; ask for missing cadence
+  rather than inventing a schedule. Do not turn a one-time alert/configuration request into recurring
+  monitoring. If native workflow proposals are unavailable, explain the limitation rather than
+  claiming an external read-only integration cannot support SimpleChat alerting.
 - A task can merge files with code instead of a model by adding "merge": {"kind","files",
   "output_format","file_name","options"} and no agent runner. kind "tabular" (the default) appends
   the rows of CSV and Excel files into one "csv" or "xlsx" file; "workbook" puts each CSV or Excel

application/single_app/functions_orchestration_registry.py

@@ -1485,7 +1485,10 @@ def generate_image_arguments_schema(options=None):
         ),
         'when_to_use': (
             'Use one step when the user asks for work to repeat on a schedule, run when File Sync '
-            'finds changes, or be saved to run later. Write the blueprint from the request and the '
+            'finds changes, or be saved to run later, including monitoring external data and issuing '
+            'native SimpleChat alerts on findings. Native alerts can pop up with critical severity, '
+            'require acknowledgment, repeat sound and use full-screen size; they are not email or '
+            'changes to the external source alarm configuration. Write the blueprint from the request and the '
             'workflow_planning catalog alone: this step takes no depends_on and no inputs, and no '
             'other step may bind its output. When the request also wants a result now, answer it '
             'once with the usual steps and select that answer as final_response.'

application/single_app/functions_orchestration_workflows.py

@@ -62,6 +62,7 @@
 from functions_workflow_drafts import (
     BLUEPRINT_TASK_TITLE_MAX_LENGTH,
     WORKFLOW_ORIGIN_SOURCE_ORCHESTRATION,
+    blueprint_alert_summary,
     check_workflow_blueprint,
     dry_run_workflow_blueprint,
     validate_workflow_blueprint,
@@ -500,7 +501,6 @@ def _proposal_summary(blueprint, task_actions, planning, used, request_time_zone
         m365_sources.update(entry.get('m365_sources') or ())
         can_send = can_send or entry.get('can_send') is True
         required = required or entry.get('needs_run_as') is True
-    alerts = blueprint.get('alerts') if isinstance(blueprint.get('alerts'), dict) else {}
     return {
         'name': clean_catalog_text(blueprint.get('name'), NAME_MAX_LENGTH * 2),
         'description': clean_catalog_text(blueprint.get('description'), 1000),
@@ -516,7 +516,7 @@ def _proposal_summary(blueprint, task_actions, planning, used, request_time_zone
             'run_as': blueprint.get('run_as') or 'none',
             'required': required,
         },
-        'alerts': {'mode': alerts.get('mode') or 'every_run', 'severity': alerts.get('severity') or 'info'},
+        'alerts': blueprint_alert_summary(blueprint),
         'durable': True,
     }, stored_type, normalized
 

application/single_app/functions_workflow_alert_authoring.py

@@ -0,0 +1,85 @@
+# functions_workflow_alert_authoring.py
+"""Closed, ID-free alert schemas shared by workflow AI authoring surfaces."""
+
+from functions_workflow_alerts import (
+    WORKFLOW_ALERT_DELIVERIES,
+    WORKFLOW_ALERT_EVALUATION_PROMPT_MAX_LENGTH,
+    WORKFLOW_ALERT_FILE_SYNC_OUTCOMES,
+    WORKFLOW_ALERT_MAX_TEXT_VALUES,
+    WORKFLOW_ALERT_REGEX_MAX_LENGTH,
+    WORKFLOW_ALERT_RULE_NAME_MAX_LENGTH,
+    WORKFLOW_ALERT_RUN_STATUSES,
+    WORKFLOW_ALERT_SEVERITY_ORDER,
+    WORKFLOW_ALERT_SIZE_ORDER,
+    WORKFLOW_ALERT_SOUND_ORDER,
+    WORKFLOW_ALERT_TASK_STATUSES,
+    WORKFLOW_ALERT_TEXT_MATCH_MODES,
+    WORKFLOW_ALERT_TEXT_VALUE_MAX_LENGTH,
+)
+
+
+def _text(limit, minimum=1):
+    return {'type': 'string', 'minLength': minimum, 'maxLength': limit, **({'pattern': r'\S'} if minimum else {})}
+
+
+def _values(options):
+    return {'type': 'array', 'minItems': 1, 'maxItems': len(options), 'uniqueItems': True,
+            'items': {'enum': sorted(options)}}
+
+
+def alert_condition_schema(*, file_sync=True):
+    """Native conditions without storage IDs or free-form fields."""
+    variants = []
+    fields = {
+        'run_status': ({'statuses': _values(WORKFLOW_ALERT_RUN_STATUSES)}, ['statuses']),
+        'task_status': ({'statuses': _values(WORKFLOW_ALERT_TASK_STATUSES)}, ['statuses']),
+        'model_evaluation': ({'prompt': _text(WORKFLOW_ALERT_EVALUATION_PROMPT_MAX_LENGTH)}, ['prompt']),
+        'agent_signal': ({
+            'signal_name': _text(WORKFLOW_ALERT_RULE_NAME_MAX_LENGTH, 0),
+            'min_severity': {'enum': list(WORKFLOW_ALERT_SEVERITY_ORDER)},
+        }, []),
+        'no_output': ({}, []),
+    }
+    if file_sync:
+        fields['file_sync'] = ({'outcome': {'enum': sorted(WORKFLOW_ALERT_FILE_SYNC_OUTCOMES)}}, ['outcome'])
+    for kind, (properties, required) in fields.items():
+        variants.append({
+            'type': 'object', 'additionalProperties': False, 'required': ['type', *required],
+            'properties': {'type': {'const': kind}, **properties},
+        })
+    for mode in sorted(WORKFLOW_ALERT_TEXT_MATCH_MODES):
+        properties = (
+            {'pattern': _text(WORKFLOW_ALERT_REGEX_MAX_LENGTH)} if mode == 'regex' else {
+                'values': {'type': 'array', 'minItems': 1, 'maxItems': WORKFLOW_ALERT_MAX_TEXT_VALUES,
+                           'items': _text(WORKFLOW_ALERT_TEXT_VALUE_MAX_LENGTH)},
+                'case_sensitive': {'type': 'boolean'},
+            }
+        )
+        variants.append({
+            'type': 'object', 'additionalProperties': False,
+            'required': ['type', 'mode', 'pattern' if mode == 'regex' else 'values'],
+            'properties': {'type': {'const': 'text_match'}, 'mode': {'const': mode}, **properties},
+        })
+    return {'oneOf': variants}
+
+
+def alert_rule_properties(task_reference, *, file_sync=True):
+    """Editable personal rule fields; the caller supplies its local task-reference schema."""
+    return {
+        'name': _text(WORKFLOW_ALERT_RULE_NAME_MAX_LENGTH, 0),
+        'enabled': {'type': 'boolean'},
+        'severity': {'enum': list(WORKFLOW_ALERT_SEVERITY_ORDER)},
+        'delivery': {'enum': sorted(WORKFLOW_ALERT_DELIVERIES)},
+        'require_acknowledgment': {'type': 'boolean'},
+        'sound': {'enum': list(WORKFLOW_ALERT_SOUND_ORDER)},
+        'size': {'enum': list(WORKFLOW_ALERT_SIZE_ORDER)},
+        'scope': {
+            'oneOf': [
+                {'type': 'object', 'additionalProperties': False, 'required': ['type', 'task'],
+                 'properties': {'type': {'const': 'task'}, 'task': task_reference}},
+                {'type': 'object', 'additionalProperties': False, 'required': ['type'],
+                 'properties': {'type': {'enum': ['final', 'any_task']}}},
+            ],
+        },
+        'condition': alert_condition_schema(file_sync=file_sync),
+    }

application/single_app/functions_workflow_assist.py

@@ -495,11 +495,15 @@ def parse_assist_request(body, user_id):
 
 RULES
 1. Change only what the instruction asks for, and keep everything else as it is.
-2. You can't change whether the workflow is enabled, who it runs as (Run as), sharing, ownership, URL access, approval settings, IDs or the definition version, and you can't set up File Sync. When asked, explain that the user can change it in the editor.
+2. You can't change whether the workflow is enabled, who it runs as (Run as), sharing, ownership, URL access, task approval settings, IDs or the definition version, and you can't set up File Sync. Alert acknowledgment is different from task approval and IS editable through alert operations. When asked for unsupported changes, explain that the user can change them in the editor.
 3. For each and If blocks can't be added, removed or restructured yet. When draft.format is "flow", tasks can't be added, removed, reordered or given inputs, but you can rename a task and change its instructions, runner and documents. Explain anything else.
 4. Parts of the draft marked "editable": false, a publication task's runner and documents, and a task's approval can't change. Explain instead.
 5. Schedules: "every day at", "on weekdays at", "every Monday at" and "on the 1st at" are calendar schedules (set_schedule_calendar). "Every N minutes" and "every N hours" are intervals (set_schedule_interval). time_of_day is a 24-hour HH:MM time. Leave "timezone" out unless the user names a time zone; the schedule then keeps its current time zone or uses time_zone. An interval can't be shorter than choices.minimum_interval_seconds.
 6. Alerts: mode "every_run" alerts after every run at a priority. Mode "rules" alerts only when a rule matches, so "only alert me when..." means mode "rules" plus a rule for that condition: a model_evaluation rule whose prompt states the condition, or a text_match rule for exact words. Match the severity to the wording; urgent or critical means high or critical. Mode "off" turns alerts off.
+   These are native SimpleChat alerts, not email and not changes to an external system's alarms. A task can read external telemetry with a read-only agent and a rule can alert on its findings even when the run succeeds. run_status/task_status failures mean workflow execution failed, not that the monitored system failed. Use model_evaluation for semantic findings, explicitly distinguishing actual problems from "no active alerts" or healthy results; use text_match only for defined literal signals. Use agent_signal only when the agent actually emits supported alert signals.
+   A file_sync alert condition can watch changes_found, no_changes or sync_failed in an existing File Sync workflow; editing that rule does not configure the File Sync trigger. Run status, File Sync and agent signals watch the whole run (final scope).
+   Info/low normally go to the bell; medium/high/critical normally pop up. Rules support require_acknowledgment (boolean), sound ("off", "once", "repeat"), and size ("small", "medium", "large"). Acknowledgment keeps the pop-up returning across pages/devices until acknowledged. Pop-up options need resolved pop-up delivery; repeat sound needs acknowledgment. These settings do not override device sound preferences. Update an existing rule by its alert handle with update_alert_rule, preserving everything not requested. "Require acknowledgment on Rule 2" changes that rule, not sound, size, other rules, or task approvals; also set delivery "popup" if needed and explain it. If the target is ambiguous, ask.
+   When the user explicitly requests the highest/strongest attention possible, propose critical severity, delivery "popup", require_acknowledgment true, sound "repeat", size "large", unless their other constraints say otherwise. Explain the acknowledgment, repeating sound, and full-screen consequences in the reply for review. Otherwise keep quieter defaults and unrequested options unchanged. Do not substitute an email task for native alerting unless email is explicitly requested. These are personal workflows; alerting "us" does not grant group audience or new recipients. Changes remain draft-only until saved.
 7. Runners: choose agents and models only from "choices". A task inherits the workflow's runner unless it has its own agent or model. A task that sends email needs an agent with the Microsoft 365 action.
 8. "attached_documents" are documents the user attached with #. Place each one by what the instruction asks:
    - Read now to shape the workflow ("use #X to design the steps"): don't place it. Say in the reply that it was read as context.