physix/training/prompt.py

"""Prompt rendering and completion parsing for PhysiX-Live.

Responsibility:

- :func:`render_observation_for_prompt`: serialise a :class:`PhysiXObservation`
  into a compact, token-efficient string the agent can read.
- :func:`build_prompt`: combine the system message, grammar hint, and the
  current observation into a single chat-formatted prompt.
- :func:`parse_completion`: parse a raw model completion (which may contain a
  JSON object inside arbitrary text) into a :class:`PhysiXAction`.

This module imports nothing from :mod:`torch`, :mod:`unsloth`, or :mod:`trl`
so it can be tested on any machine.
"""

from __future__ import annotations

import json
import re
from typing import Any

from physix.models import (
    DEFAULT_MAX_TURNS,
    PhysiXAction,
    PhysiXObservation,
)
from physix.verifier.parser import GRAMMAR_HINT


SYSTEM_MESSAGE: str = (
    "You are an expert physicist. Your task is to discover the equation of "
    "motion that produced an observed trajectory. Each turn you propose a "
    "candidate equation; the environment simulates it and tells you how well "
    "the prediction matches observation. Refine your guess across turns based "
    "on the residual feedback. Keep equations as simple as possible.\n\n"
    + GRAMMAR_HINT
    + "\n\n"
    "Output a single JSON object with exactly these keys: "
    '"equation" (string, required), "params" (object of name->number, '
    'optional), "rationale" (short string, optional). Do not rename the '
    'keys: always emit "equation", never "eqn"/"ode"/"formula"/"expr". '
    'Example: {"equation": "d2y/dt2 = -9.81", "params": {}, '
    '"rationale": "free fall"}'
)


# Maximum number of trajectory samples shipped to the agent. We downsample
# from 100 to 12 to keep prompt size bounded; statistics carry the rest.
_TRAJECTORY_DOWNSAMPLE_COUNT: int = 12

# Maximum number of prior history entries surfaced. With 8 turns max budget,
# 7 prior turns is the upper bound; we cap at 5 to stay token-efficient.
_HISTORY_CAP: int = 5


def render_observation_for_prompt(obs: PhysiXObservation) -> str:
    """Render an observation as a compact string the agent can read.

    Format::

        SYSTEM_ID: free_fall_drag
        STATE_VARIABLES: y, vy
        HINT: <one-sentence physical context>
        STATS: y_min=-2.13 y_max=78.93 ... duration=6.00

        TRAJECTORY (12 samples downsampled from 100):
        t=0.000  y=78.93  vy=0.00
        t=0.500  y=77.71  vy=-4.84
        ...

        HISTORY (turns so far):
        turn=1  reward=0.42  [match=0.42 progress=0.00 simplicity=0.95 format=1.00]  equation=`d2y/dt2 = -9.81`
                mismatch: predicted y diverges past t=2.0s ...
        turn=2  ...

        TURN: 3 / 8 (5 turns remaining)

    HISTORY uses the literal string ``equation=`` (not a shorthand like
    ``eqn=``). Mid-strength chat models will mimic the field name they
    see in HISTORY when emitting the next turn's JSON, so the in-prompt
    name *must* match the JSON key the parser reads. Drift here silently
    produces ``{"eqn": ...}`` outputs that the parser ignores, scoring
    every post-first turn ``r_format=0`` even when the equation is
    perfect.
    """
    sections = [
        _render_metadata_block(obs),
        _render_trajectory_block(obs),
    ]
    if obs.history:
        sections.append(_render_history_block(obs))
    sections.append(_render_turn_footer(obs))
    return "\n\n".join(sections)


def build_prompt(obs: PhysiXObservation) -> list[dict[str, str]]:
    """Build a chat-format prompt list (system + user) for the model.

    The return value is the standard ``[{"role": "system", "content": ...},
    {"role": "user", "content": ...}]`` shape expected by Hugging Face
    chat-template tokenisers.
    """
    return [
        {"role": "system", "content": SYSTEM_MESSAGE},
        {"role": "user", "content": render_observation_for_prompt(obs)},
    ]


#: Field names we accept for the equation payload, in priority order. The
#: canonical key is ``equation`` and the system prompt asks for it
#: explicitly, but mid-strength chat models routinely substitute one of
#: these synonyms — especially after the first turn, where the model has
#: latched onto a different naming convention from its own pretraining
#: corpus. Treating these as missing produced silent ``r_format=0`` runs
#: even when the underlying equation was perfect; matching them
#: explicitly closes that hole without weakening the verifier (the
#: equation grammar itself remains strict).
_EQUATION_KEYS: tuple[str, ...] = (
    "equation",
    "eqn",
    "ode",
    "formula",
    "expression",
    "expr",
)

#: Same idea for the optional rationale payload. We never gate on this so
#: the cost of being permissive is zero.
_RATIONALE_KEYS: tuple[str, ...] = (
    "rationale",
    "reasoning",
    "explanation",
    "thought",
    "thoughts",
)

#: And for the params dict. Some models emit ``parameters`` instead.
_PARAMS_KEYS: tuple[str, ...] = ("params", "parameters", "constants")


def parse_completion(completion: str) -> PhysiXAction:
    """Parse a raw model completion into a :class:`PhysiXAction`.

    Scope is intentionally narrow: extract the first JSON object from the
    completion (which may be wrapped in markdown fences or surrounded by
    scratchpad text) and copy its fields verbatim into the action.

    The ``equation`` string is **not** rewritten or normalised here. The
    verifier in :mod:`physix.verifier.parser` defines the grammar, and any
    deviation must surface as a parse error so the env can score
    ``r_format=0`` and feed the failure back to the agent on the next turn.
    Rewriting equations upstream would silently change the agent's output
    and obscure that signal.

    Field-name aliases (``eqn``/``ode``/``formula``/...) are accepted in
    addition to ``equation``: refusing them produced a particularly
    confusing failure mode where every turn after the first scored
    ``r_format=0`` because the model latched onto the shorthand form
    used in the HISTORY block. We've fixed the prompt too, but accepting
    the synonyms is cheap defense-in-depth against future drift and
    against models with their own naming preferences.

    If no JSON object can be extracted (e.g. the model emitted free-form
    prose, or invalid JSON that even the JSON-aware decoder rejected),
    the action's ``equation`` is left **empty** so the verifier reports
    a clean ``Empty equation payload`` parse error and the env scores
    ``r_format=0``. The raw model text is preserved in ``rationale`` so
    the UI / training logs still show what was emitted, but it is
    *never* fed to the equation parser as if it were an equation —
    that produced misleading errors like ``Equation has no '=' sign:
    '{'`` which made the verifier look broken when the real fault was
    upstream.
    """
    payload = _extract_json_payload(completion)
    if payload is None:
        return PhysiXAction(equation="", rationale=completion.strip()[:500])

    normalized = _lowercase_keys(payload)
    equation = _first_string_value(normalized, _EQUATION_KEYS)
    rationale = _first_string_value(normalized, _RATIONALE_KEYS)
    params_raw = _first_value(normalized, _PARAMS_KEYS) or {}
    params = _coerce_params(params_raw)

    return PhysiXAction(
        equation=equation,
        params=params,
        rationale=rationale,
    )


def _lowercase_keys(payload: dict[str, Any]) -> dict[str, Any]:
    """Return ``payload`` with top-level keys lowercased.

    Some models emit ``"Equation"`` / ``"EQN"``; lowercasing once means
    the lookup tables above stay declarative.
    """
    return {str(k).lower(): v for k, v in payload.items()}


def _first_value(payload: dict[str, Any], keys: tuple[str, ...]) -> Any:
    for key in keys:
        if key in payload:
            return payload[key]
    return None


def _first_string_value(payload: dict[str, Any], keys: tuple[str, ...]) -> str:
    value = _first_value(payload, keys)
    if value is None:
        return ""
    return str(value).strip()


def _render_metadata_block(obs: PhysiXObservation) -> str:
    state_vars = ", ".join(obs.state_variables) or "(none)"
    stats_text = " ".join(f"{k}={v:.3g}" for k, v in obs.stats.items())
    return (
        f"SYSTEM_ID: {obs.system_id or 'unknown'}\n"
        f"STATE_VARIABLES: {state_vars}\n"
        f"HINT: {obs.hint}\n"
        f"STATS: {stats_text}"
    )


def _render_trajectory_block(obs: PhysiXObservation) -> str:
    samples = _downsample(obs.trajectory, _TRAJECTORY_DOWNSAMPLE_COUNT)
    lines = [f"TRAJECTORY ({len(samples)} samples downsampled from {len(obs.trajectory)}):"]
    for sample in samples:
        parts: list[str] = [f"t={sample['t']:.3f}"]
        for var in obs.state_variables:
            if var in sample:
                parts.append(f"{var}={sample[var]:.3f}")
        lines.append("  " + "  ".join(parts))
    return "\n".join(lines)


#: Order in which reward components are surfaced to the model. Match
#: matters most (it's the headline accuracy signal); format is last
#: because once it stabilises the others dominate the gradient. Stable
#: order also matters for the model's in-context retrieval: a fixed
#: column position is a reliable cue across turns.
_REWARD_COMPONENT_ORDER: tuple[str, ...] = (
    "match",
    "progress",
    "simplicity",
    "format",
)


def _render_history_block(obs: PhysiXObservation) -> str:
    """Render the most recent ``_HISTORY_CAP`` turns.

    Field name is ``equation=`` rather than ``eqn=`` deliberately:
    chat-tuned models tend to mimic the most-recent token spelling when
    emitting their own JSON, so we must use the same key here that the
    parser expects in the model's reply.

    Each turn's full *dense* reward breakdown is surfaced in addition to
    the scalar ``reward=`` total. The dense components are the same
    values the GRPO trainer optimises (``match``/``progress``/
    ``simplicity``/``format``), so showing them in-context lets the
    model attribute its own gains and losses turn-over-turn instead of
    having to infer them from the residual prose alone — e.g. it can
    see ``format=0.0`` after a parse error and prioritise grammar fixes,
    or ``match=0.62, progress=0.0`` after a stuck plateau and try a
    structurally different equation rather than tweaking the same
    coefficients.
    """
    recent = obs.history[-_HISTORY_CAP:]
    lines = ["HISTORY:"]
    for entry in recent:
        eqn = entry.get("equation", "")
        reward = float(entry.get("reward_total", 0.0))
        components_text = _format_reward_components(entry.get("reward_components"))
        lines.append(
            f"  turn={entry.get('turn')}  reward={reward:.3f}  "
            f"[{components_text}]  equation=`{eqn}`"
        )
        mismatch = entry.get("mismatch_summary", "")
        if mismatch:
            lines.append(f"    mismatch: {mismatch}")
    return "\n".join(lines)


def _format_reward_components(components: Any) -> str:
    """Render ``{match, progress, simplicity, format}`` as a compact line.

    Always emits all four fields in :data:`_REWARD_COMPONENT_ORDER`,
    defaulting to ``0.00`` when absent so the model never has to guess
    why a column is missing. Three-decimal formatting matches the
    server's history serialisation precision.
    """
    if not isinstance(components, dict):
        return " ".join(f"{name}=0.00" for name in _REWARD_COMPONENT_ORDER)
    parts: list[str] = []
    for name in _REWARD_COMPONENT_ORDER:
        try:
            value = float(components.get(name, 0.0))
        except (TypeError, ValueError):
            value = 0.0
        parts.append(f"{name}={value:.2f}")
    return " ".join(parts)


def _render_turn_footer(obs: PhysiXObservation) -> str:
    total = obs.turn + obs.turn_remaining or DEFAULT_MAX_TURNS
    return (
        f"TURN: {obs.turn + 1} / {total}  ({obs.turn_remaining} remaining)\n"
        "Emit the next hypothesis as JSON."
    )


def _downsample(samples: list[dict[str, float]], target: int) -> list[dict[str, float]]:
    if len(samples) <= target:
        return samples
    step = max(1, len(samples) // target)
    indices = list(range(0, len(samples), step))[:target]
    if indices[-1] != len(samples) - 1:
        indices[-1] = len(samples) - 1
    return [samples[i] for i in indices]


_JSON_DECODER = json.JSONDecoder()


def _extract_json_payload(text: str) -> dict[str, Any] | None:
    """Find the first ``{...}`` block in ``text`` that parses as a JSON object.

    Uses :meth:`json.JSONDecoder.raw_decode` so that braces appearing inside
    JSON *string* values (e.g. LaTeX like ``"\\frac{d vy}{dt}"``) do not
    confuse the scanner — a regex-based brace matcher would mis-balance
    here and return the whole completion as a malformed equation.
    """
    candidate = _strip_code_fences(text)

    for i, ch in enumerate(candidate):
        if ch != "{":
            continue
        try:
            payload, _ = _JSON_DECODER.raw_decode(candidate[i:])
        except json.JSONDecodeError:
            continue
        if isinstance(payload, dict):
            return payload
    return None


def _strip_code_fences(text: str) -> str:
    """Remove Markdown code-fence wrappers (```json``` / ```python``` / ```).

    This is *not* equation rewriting — it strips the outer fence syntax
    only, so the JSON-aware extractor below can find the object payload.
    """
    text = re.sub(r"```(?:json|python)?\s*", "", text)
    text = text.replace("```", "")
    return text


def _coerce_params(params_raw: Any) -> dict[str, float]:
    """Best-effort coercion of a raw params payload into ``dict[str, float]``."""
    if not isinstance(params_raw, dict):
        return {}
    out: dict[str, float] = {}
    for key, value in params_raw.items():
        try:
            out[str(key)] = float(value)
        except (TypeError, ValueError):
            continue
    return out