Carlos completes complex sections through conversation

[danielnaab]

User Story:

As a form filler (Carlos), in order to navigate complex conditional sections without confusion, I want to complete those sections through a conversational agent that explains questions, adapts the flow, and collects my answers

Preconditions:

• A published form exists with at least one section marked as conversational delivery mode (Slice 3)
• Static form filling works (Slice 5)

Acceptance Criteria:

• Sections marked as conversational in the FormSpec render as a chat interface instead of a static form
• The conversational agent asks questions in a natural order based on the DataCollectionSpec
• Agent skips irrelevant questions based on prior answers (evaluates conditions)
• Agent explains why data is needed when asked
• Agent flags sensitivity before collecting PII fields
• Collected answers are stored in the same Submission format as static form data
• Carlos can switch between conversational and static views for the same section
• Conversation state is preserved if Carlos navigates away and returns

Success Metrics:

• Completion rate for conversational vs. static sections
• User satisfaction (qualitative feedback)
• Correctness: conversational agent collects all required fields

Notes:

• Same DataCollectionSpec, different delivery experience — demonstrates the delivery spectrum
• Agent uses tool calls: collect_field, explain_field, skip_section
• CollectionState tracks: {field_name: value | null} for all fields in the section
• This is the filling agent pattern from the design spec

Definition of Done:

• Acceptance criteria met
• Threat model updated -- any new trust boundaries, data flows, or attack surfaces are reflected in catalog/architecture/threat-model.md
• Technical documentation updated -- architecture docs and decisions are current
• Filling agent has interface abstraction (swappable implementations)
• Tests pass
• Type checking passes
• CI pipeline green
• Deployed and demoable

[danielnaab]

This story has been started but is weak and in need of revisiting.

[danielnaab]

Acceptance criteria audit (2026-05-03)

5 of 8 criteria met — core conversational filling works and is deployed (PR #72).

Met

• ✅ Chat interface renders for conversational sections (flex-assistant component)
• ✅ Agent asks questions in natural order from DataCollectionSpec
• ✅ Agent skips irrelevant questions via evaluateCondition()
• ✅ Answers stored in standard Submission/FieldEntry format
• ✅ Conversation state persisted via SqliteConversationGateway

Not met

• ❌ Agent explains why data is needed — explain_field tool defined but execution is a no-op
• ❌ Agent flags PII sensitivity — SensitivityLevel exists in data model but filling agent ignores it
• ⚠️ Switch between chat/static views — navigation links exist but field state doesn't sync bidirectionally

Definition of Done

• ✅ Interface abstraction (FillingAgent with 2 implementations)
• ✅ Tests pass, type checking, CI green, deployed
• ❌ Threat model not updated for conversation state trust boundary
• ⚠️ Technical documentation partial