Use this for a measurable comparison, not ordinary page editing.
Use lexsis_pages.edit_context, lexsis_pages.source,
lexsis_pages.compile, lexsis_pages.integrity, lexsis_analytics.page,
lexsis_analytics.experiment, lexsis_drafts.page_duplicate,
lexsis_drafts.page_variation, and lexsis_drafts.experiment_create.
Resolve unfamiliar argument schemas with exact router/action discovery. Do
not interpret an empty discovery result as a page or analytics outage. Report
any failure from the actual read or mutation and do not claim that operation
succeeded.
Confirm the base page's store/theme binding exists in
work/storefront/setup/setup.json. If it is missing, stop and ask the user to
run /setup; never invoke setup automatically.
Define the Test
Confirm:
- page and current baseline version
- one primary hypothesis
- primary metric and guardrail metrics
- intended audience or traffic segment
- approved traffic split and stopping rule
Change as few elements as needed to test the hypothesis. If several unrelated ideas are bundled together, split them into separate tests.
Local-First Variants
Before duplicating or changing a remote page:
- Confirm the base page has synchronized local source and manifest.
- Create a separate local directory, source file, and manifest for each variant.
- Apply the variant change locally.
- Validate and compile the complete variant source.
- Create or update the remote variant with the expected base version.
- Store every returned page ID and version locally.
- Run integrity, responsive, and affected commerce checks.
Never let a remote variant become the only copy of a change.
Launch
Use the currently discovered Lexsis experiment actions and schemas. Confirm entitlement and credits before creation. Do not publish a base page or variant without explicit approval.
Evaluate
Read the experiment's current status and results from Lexsis. Report:
- sample sizes and exposure split
- primary and guardrail metric movement
- whether the configured decision rule has been reached
- data-quality or targeting concerns
- recommended action: continue, stop, promote, or discard
Do not call a winner from directional movement alone.
Return
Return the hypothesis, local variant paths, remote page/version IDs, experiment ID, launch state, current decision, and MCP evidence.