Workflow rules
Required procedures that keep Mozark workflows reliable.
The Mozark server ships mandatory workflow rules that any AI assistant retrieves before starting a task (get_workflow_rules). They exist to keep runs reliable and to stop the assistant from guessing. If the assistant seems to be asking a lot of questions, this is why.
Context
Resolve organisation / project / app once per session, before any project-scoped tool.
If only an account-wide default is available and you belong to several organisations, confirm before proceeding.
Any change you request in conversation is applied by actually calling the switch tool — a spoken "OK" doesn't change anything on its own.
Test generation
Always ask whether to use the existing knowledge base or add material first.
Always ask test type, polarity and feature name — never auto-select.
Always search the knowledge base before generating.
Never save unless you explicitly ask.
Test plans
List existing plans first; consult the Planner Agent alone; show matches before asking new-vs-existing; never create or update without confirmation.
Optics (device execution)
Pick a device from the free list; ask about capture options before starting (they can't be added later).
Never guess an app package — ask, or read it from the device.
Ask for the locator fallback order (text → XPath → AI fallback recommended) after probing the screen.
Inspect the screen before every step; narrate action, result, observation, expected result and next.
Verify each step with
assert_presence; never claim a pass without screenshot evidence.On failure, work down the ladder before reporting; never silently skip a step.
Export text-locator YAML with an assertion after every action; release the device.
Appium
List free devices and start sessions automatically — never ask you for a session ID.
Upload local builds via the streaming curl command; install by
mzfile://ID; never base64 a binary.Stop the session when done unless you ask to keep it open.
Test runs (Optics or Appium)
Start the test case in the run before step 1; mark every step; attach evidence on failures; finish the run with a terminal status.
Was this helpful?

