Update Chat Agent Scenarios

Consolidated parameter policy for a scenario.

This replaces the per-parameter ToolParameterConfig list with a cleaner, scenario-level interface for specifying parameter behavior:

• defaults: Fallback values if LLM doesn't provide them
• fixed: Always used, overrides any LLM-provided values
• required: Must be satisfied (by LLM, fixed, or default) before execution
• allowed: Enum constraints restricting valid values per parameter
• guidance: Per-parameter instruction for the model: what to ask the caller for or how to source the value. Appended to that parameter's description in the tool schema as "(Hint: ...)" by build_model_schema, taking precedence over an ask-caller prompt_hint. Ignored for fixed parameters, which the model never sees.

Example: ParameterPreset( defaults={"priority": "normal"}, fixed={"source": "voice_call"}, required=["contact_email"], allowed={"priority": ["low", "normal", "high"]}, guidance={"contact_email": "Ask the caller for their email address"} )

Typed annotation of a transfer scenario's primary destination.

Links the free-text destination (preset.fixed["to"]) to the entity it targets — a HumanAgent or a routing-environment node — so the routing-map projection resolves edges without string-matching phone numbers. Purely additive: the phone string remains the executable truth and the runtime ignores this field entirely.

Per-destination schedule override. Same day-buckets as the tenant default; None for a day-group means closed all day. Consumed by the shared business_hours normalizer after .model_dump().

friday is optional and key-presence-sensitive (see effective_group_days): a present key, including explicit None, closes or windows Friday on its own; a missing key leaves Friday inside weekdays. Unset friday is omitted from dumps so a legacy save cannot persist null and close Friday.

What the agent does when a gated destination is closed and no eligible transfer remains. Validation guarantees the type-specific field is present so the worker never hits an undefined terminal state.

Structured trigger specification for scenario matching.

This enables more sophisticated matching than raw string conditions:

• keyword: Match specific phrases ("I want to book", "transfer me")
• intent: Semantic intent matching ("booking_intent", "complaint")
• entity: Entity presence ("email_address", "phone_number")
• tool_request: LLM explicitly requests a tool by name
• always: Default fallback, always matches

The condition field in ScenarioBase remains for backward compatibility and is used as the LLM-facing description. TriggerSpec is for routing.

Example: TriggerSpec( type=TriggerType.INTENT, value="schedule_appointment", confidence_threshold=0.8 )

Configuration for what the agent says during tool execution.

The platform supports 4 message types during tool calls:

• start: When tool execution begins
• delayed: If execution takes longer than expected or user speaks
• complete: When tool succeeds
• failed: When tool fails

For each message:

• None = use platform default behavior
• "" (empty string) = silence (say nothing)
• "custom text" = say this specific message
• ["variant A", "variant B"] = random selection per call

Platform defaults:

• start: Random filler like "Hold on a sec", "One moment", "Just a sec"
• delayed: "Sorry, a few more seconds"
• complete: (none)
• failed: (none)