# One query per line. Blank lines and lines starting with # are ignored.
#
# Each user picks from this list with a PRNG seeded from (run label, user
# index): random-looking and spread across users, but identical between two
# arms of an A/B so both see the same workload.
#
# WHY A LIST AND NOT ONE QUERY: a single repeated query gets its prompt prefix
# cached by the LLM provider and returns the same warm records every turn, so
# it overstates throughput. Use `PIPESHUB_QUERY` to force a single query only
# when you want both arms of an A/B identically biased -- never to quote a
# capacity number.
#
# REPLACE THESE for your own corpus, then check a run's answers: turns coming
# back with 0 citations are measuring the not-found path, not retrieval.

How do I set up an agent, and what configuration does it need?
Which AI model providers are supported and how are they configured?
How are spending limits and AI budgets applied to users and teams?
What does the connector integration playbook say about adding a new connector?
How is access control and permissions handled across connectors?
Summarise the task runner and concurrency architecture for background syncs.
What issues were found in the recent bug bash sessions?
What does the Asana SOC 3 report cover?
What has been discussed recently about MCP and the context layer?
What is on the PipesHub capabilities roadmap?
