impactflow_discovery/app/services/reflector.py

"""ReflectionCoach: the Phase 2 AI-coach reflection loop.

The coach is a MIRROR, never a compass. It reflects the person's own profile
back to them in plain language, listens to their reactions, and — only when
they explicitly correct or add something — proposes revised text for the
affected prose fields using the person's own direction. It never prescribes
goals or invents direction.

Like DiscoveryExtractor, this class is responsible only for plumbing: building
the messages, calling the model, and parsing/validating the JSON it returns.
"""
import json
from typing import Any, Dict, List, Optional

from anthropic import AsyncAnthropic

DEFAULT_MODEL = "claude-sonnet-4-6"
MAX_TOKENS = 1200

# The only profile fields the coach may propose changes to. The structural
# Enneagram read (triad/type/wing/variant) is never editable via reflection.
EDITABLE_FIELDS = (
    "love_summary",
    "strength_summary",
    "mission_summary",
    "vocation_summary",
    "overlap_narrative",
    "short_term_goals",
    "long_term_goals",
)

# Maps a stored ReflectionMessage.role to an Anthropic message role.
ROLE_TO_API = {"coach": "assistant", "person": "user"}

# Sent as the first (user) turn on every call so the conversation always
# starts with a user message, and to frame the coach's task. Not stored.
PRIMER = (
    "I have just completed my self-discovery profile (it is in your "
    "instructions). Reflect it back to me so I can see whether it fits."
)

SYSTEM_PROMPT = """You are an AI coach inside a self-discovery tool. You are a MIRROR, never a compass.

THE PERSON'S CURRENT PROFILE:
{profile}

YOUR ROLE:
- Reflect this profile back in warm, plain language and ask whether it lands: in spirit, "Here is what I am hearing — do you recognize yourself? What would you add, change, or disagree with?"
- Listen to how the person reacts. When they correct, add to, or push back on something, reflect their own words back to them — clarify and sharpen what THEY mean.
- Ask gentle, open questions that help the person articulate their own sense of direction.

ABSOLUTE RULES (mirror, not compass):
- NEVER prescribe goals, paths, careers, or what they "should" do.
- NEVER invent a direction the person did not express. If you are unsure what they mean, ask rather than assume.
- Do NOT mention Enneagram type numbers; describe patterns in plain language.
- Keep replies short and conversational — 2 to 5 sentences, at most one question.

PROPOSING REVISIONS:
- Only when the person explicitly corrects, adds to, or asks to change part of their profile, propose updated text for the affected field(s), written in their own direction. Editable fields: love_summary, strength_summary, mission_summary, vocation_summary, overlap_narrative, short_term_goals, long_term_goals.
- Otherwise set "revisions" to null. Never change their Enneagram type, triad, or instinctual variant. Never revise just because you could — only to capture what the person said.

OUTPUT FORMAT:
Respond ONLY with valid JSON. No preamble, no markdown fences.

{{
  "message": "your reflective reply to the person, in second person (you/your), warm and plain",
  "revisions": {{ "<field>": "<revised text in the person's own direction>" }} or null,
  "revision_note": "a short phrase naming what changed, or null"
}}"""

RETRY_REMINDER = (
    "Your previous response could not be parsed as JSON. Respond ONLY with the "
    "single valid JSON object described in your instructions — no preamble, no "
    "explanation, and no markdown code fences."
)


class ReflectionError(Exception):
    """Raised when a reflection turn fails (API error or unparseable output)."""


class ReflectionCoach:
    """Generates one coach turn given the profile and the dialogue so far."""

    def __init__(self, api_key: str, model: str = DEFAULT_MODEL):
        if not api_key:
            raise ReflectionError(
                "ANTHROPIC_API_KEY is not set; cannot run reflection."
            )
        self.model = model
        self.client = AsyncAnthropic(api_key=api_key)

    @staticmethod
    def profile_context(profile: Dict[str, Any]) -> str:
        """Render the current profile as plain text for the system prompt."""
        triad = {
            "gut": "leads with instinct and will (gut-centered)",
            "heart": "leads with feeling and connection (heart-centered)",
            "head": "leads with thought and perception (head-centered)",
        }.get(profile.get("triad") or "", "centered pattern not yet clear")
        lines = [
            f"- Core pattern: {triad}",
            f"- What you love: {profile.get('love_summary') or '(none)'}",
            f"- What you are good at: {profile.get('strength_summary') or '(none)'}",
            f"- What the world needs from you: {profile.get('mission_summary') or '(none)'}",
            f"- What you can be paid for: {profile.get('vocation_summary') or '(none)'}",
            f"- Where it converges: {profile.get('overlap_narrative') or '(none)'}",
            f"- Near-term goals (6-12mo): {profile.get('short_term_goals') or '(none stated)'}",
            f"- Long-term goals (3-5yr): {profile.get('long_term_goals') or '(none stated)'}",
        ]
        return "\n".join(lines)

    def _build_messages(
        self, history: List[Dict[str, str]]
    ) -> List[Dict[str, str]]:
        """Build the Anthropic messages array: a fixed user primer followed by
        the stored turns mapped to user/assistant roles."""
        messages = [{"role": "user", "content": PRIMER}]
        for turn in history:
            api_role = ROLE_TO_API.get(turn["role"])
            if api_role is None:
                continue
            messages.append({"role": api_role, "content": turn["content"]})
        return messages

    async def _call_model(
        self, profile: Dict[str, Any], messages: List[Dict[str, str]]
    ) -> str:
        response = await self.client.messages.create(
            model=self.model,
            max_tokens=MAX_TOKENS,
            system=SYSTEM_PROMPT.format(
                profile=self.profile_context(profile)
            ),
            messages=messages,
        )
        return response.content[0].text

    async def reflect(
        self, profile: Dict[str, Any], history: List[Dict[str, str]]
    ) -> Dict[str, Any]:
        """Produce one coach turn.

        Args:
            profile: the current profile dict (at least the editable fields and
                triad).
            history: prior turns as ``[{"role": "coach"|"person", "content": ...}]``
                in order. Empty for the opening reflection. The last turn, if
                any, should be the person's latest message.

        Returns:
            ``{"message": str, "revisions": dict|None, "revision_note": str|None}``
            with revisions filtered to the editable fields only.

        Raises:
            ReflectionError on API failure or repeated parse failure.
        """
        messages = self._build_messages(history)

        try:
            raw = await self._call_model(profile, messages)
        except Exception as exc:  # noqa: BLE001 - surface any SDK/transport error
            raise ReflectionError(f"Anthropic API call failed: {exc}") from exc

        try:
            return self._parse(raw)
        except (json.JSONDecodeError, ValueError):
            retry = messages + [
                {"role": "assistant", "content": raw},
                {"role": "user", "content": RETRY_REMINDER},
            ]
            try:
                raw_retry = await self._call_model(profile, retry)
            except Exception as exc:  # noqa: BLE001
                raise ReflectionError(
                    f"Anthropic API call failed on retry: {exc}"
                ) from exc
            try:
                return self._parse(raw_retry)
            except (json.JSONDecodeError, ValueError) as exc:
                raise ReflectionError(
                    f"Model did not return valid JSON after retry: {exc}"
                ) from exc

    @staticmethod
    def _strip_fences(text: str) -> str:
        stripped = (text or "").strip()
        if stripped.startswith("```"):
            newline = stripped.find("\n")
            if newline != -1:
                stripped = stripped[newline + 1 :]
            if stripped.rstrip().endswith("```"):
                stripped = stripped.rstrip()[: -len("```")]
        return stripped.strip()

    @classmethod
    def _parse(cls, raw: str) -> Dict[str, Any]:
        if not raw or not raw.strip():
            raise ValueError("empty response from model")

        data = json.loads(cls._strip_fences(raw))
        if not isinstance(data, dict):
            raise ValueError("top-level JSON value is not an object")

        message = data.get("message")
        if not isinstance(message, str) or not message.strip():
            raise ValueError("missing or empty 'message'")

        revisions = cls._clean_revisions(data.get("revisions"))
        note = data.get("revision_note")
        if not isinstance(note, str) or not note.strip():
            note = None

        return {
            "message": message.strip(),
            "revisions": revisions,
            "revision_note": note,
        }

    @staticmethod
    def _clean_revisions(revisions: Any) -> Optional[Dict[str, str]]:
        """Keep only editable string fields; drop anything else (e.g. an
        attempt to change triad/type). Returns None if nothing valid remains."""
        if not isinstance(revisions, dict):
            return None
        cleaned = {
            k: v
            for k, v in revisions.items()
            if k in EDITABLE_FIELDS and isinstance(v, str) and v.strip()
        }
        return cleaned or None