v1.5.0

Added

• Lessons from tuning live agents, written without client specifics:
  • prompt-authoring.md: branches and fallbacks (clear-but-unexpected answers,
    accept branches, a general no-information line, per-question mismatch
    branches, routing fixes, absolute-reading prohibitions, scoping
    contradictions), content that must be said despite interruptions,
    mishearing (letter/digit codes, no echo, saying the limit aloud, separate
    ladders), persuasion ladders, knowledge-base search (query wording, facts
    that suppress search, the two-search rule, a search check at the end of the
    task), one-line flow examples, the greeting's own rendering, concrete values
    in examples, asking for silence, no in-call summaries, computed numbers in
    say-as, price tables, scoped pronunciation rules, business-hours gates that
    fail closed, and where to place new text
  • tools.md: tools that ask the customer something, lookup keys, failed
    transfers, returning decisions instead of raw values, exception lists in
    descriptions, sending the day and time together to resolve_date, and a
    caveat on model-side days_from_today comparisons
  • agents-and-teams.md: splitting one agent into a team
  • outcome-design.md: the labels that go wrong most, and schema pitfalls
    (integer enums, stored field order, tool results visible to post-call
    steps but not webhooks)
  • postprocessors.md, faq-authoring.md, troubleshooting.md: instruction
    wording, fixing a repeatedly wrong entry, greeting hang-ups, and cut-off
    numbers from malformed say-as tags

Changed

• The knowledge-base query guidance no longer suggests adding a brand name to
  queries; the company's own name measurably lowers retrieval.

• references/testing.md and templates/promptfoo-suite/: how to set up an
  automated conversation test suite for an agent with promptfoo. The template
  exports the agent and its tools through the client API, composes the prompt
  in the platform's order, plays whole conversations on the production model
  (google/gemma-4-26b-a4b-it, temperature 0.5, 768 output tokens), and tests
  the disposition outcomes on google/gemini-3.1-flash-lite. It includes
  reusable assertions (tool fired, silence after a transfer, no tool name
  spoken, spoken-Thai digit matching), a fictional example agent to check the
  setup, token and cost reporting, and guidance on keys, cost, what to test,
  and how to write assertions that survive run-to-run variation. Verified end
  to end: 8 conversations × 3 repeats and 4 disposition cases passing, with
  assertions confirmed to fail when their behaviour breaks.