[ask-a-sailor] ABA-informed dialogue layer: Coach + Reflect modes, Social Stories schema

[Copilot]

Adds ABA-informed prompt architecture for youth sailing coaching, informed by ASD-iLLM (Lai et al., EMNLP 2025) and the ABA-with-GPT HuggingFace blueprint.

Prompt modes (src/prompts/system.py)

• Coach — in-the-moment, actionable guidance (sensory-aware tips, positive reinforcement, short structured responses)
• Reflect — post-session metacognitive processing (open-ended questions, effort-over-outcome framing, longer reflective responses)
• Shared _BOUNDARY_BLOCK injected into both: declines diagnosis, redirects to BCBA/pediatrician/counselor
• Original SYSTEM_PROMPT unchanged

Social Stories schema (src/prompts/social_stories.py)

Pydantic model with the six specified fields:

from prompts.social_stories import SocialStory, TonePreset, DisclosureIntent

story = SocialStory(
    setting="Lakewood Yacht Club dock, Saturday morning",
    roles=["sailor (child)", "instructor"],
    objective="Preview the steps for rigging an Opti",
    sensory_profile="Loud halyards, wind, cold spray",
    tone_preset=TonePreset.calm,
    disclosure_intent=DisclosureIntent.normalising,
)

Tests (tests/test_prompts.py)

35 tests across four classes: prompt distinctness, boundary-hygiene phrase checks on both ABA prompts, persona keyword differentiation (Coach vs Reflect), and schema validation including enum serialisation roundtrip.

Original prompt

---

This section details on the original issue you should resolve

<issue_title>[ask-a-sailor] ABA-informed dialogue layer: Coach + Reflect modes, Social Stories schema</issue_title>
<issue_description>ASD-iLLM (Lai et al., EMNLP 2025) fine-tuned a 7B model on 8,035 ABA-based clinical dialogues and outperformed all existing models. The 'ABA with GPT' HuggingFace blueprint (Weinberger + Molnar, huggingface.co/blog/Clock070303/aba-for-gpt) provides full prompt architecture. Work: (1) Two system prompt modes in src/prompts/system.py -- Coach (in-the-moment) and Reflect (post-session metacognitive), (2) Social Stories corpus schema: fields = setting, roles, objective, sensory_profile, tone_preset, disclosure_intent, (3) boundary hygiene: agent explicitly declines diagnosis, redirects to professionals, (4) tests verifying persona differences. This is the foundation for the ASD/youth coaching product. Refs: github.com/Shuzhong-Lai/ASD-iLLM, Noora RCT doi:10.1007/s10803-025-06734-x</issue_description>

Comments on the Issue (you are @copilot in this section)

• Fixes [ask-a-sailor] ABA-informed dialogue layer: Coach + Reflect modes, Social Stories schema #16

---

🔒 GitHub Advanced Security automatically protects Copilot coding agent pull requests. You can protect all pull requests by enabling Advanced Security for your repositories. Learn more about Advanced Security.