"""
METASTATE v1.0 — FastAPI backend  (Option A core + Option B BYO-key chat)

CORE PRODUCT (runs entirely on our own kernel, no external LM ever):
  POST /v1/anomaly/score     free-energy + causal-coherence + universal signatures
  POST /v1/symbolic/regress  EML closed-form recovery over a signal window

BYO-KEY CHAT (optional conversational layer — Option B, stated plainly):
  POST /v1/chat/proxy        pure pass-through to the USER'S OWN provider key.
                             We never store keys and never use our own. We only
                             (optionally) bill for the anomaly/symbolic processing
                             we add, never for the inference itself.

Payments: USDC on Base mainnet. Treasury split 35/35/30. Affiliate 15%.
The kernel below is a real, dependency-light reference implementation of the
diagnostics in the paper. Swap individual functions for the full
TimesFM / process-matrix / nwo-eml-regression modules as they mature.
"""
import os, time, json, math, gzip, secrets, sqlite3, urllib.request, urllib.error
from pathlib import Path
from collections import Counter
from fastapi import FastAPI, Request, Header, HTTPException
from fastapi.responses import HTMLResponse, JSONResponse, PlainTextResponse
from pydantic import BaseModel
from typing import Optional, List
import storage, pqc, video, compiler, nwoagi

# ---------------------------------------------------------------- config
BASE_DIR      = Path(__file__).parent
TREASURY      = "0x2E964e1c0e3Fa2C0dfD484B2E6D2189dfCF20958"   # guardian wallet
CHAIN         = "base"
USDC_BASE     = "0x833589fCD6eDb6E08f4c7C32D4f71b54bdA02913"
# Deployed & verified MetaStateSplitter on Base mainnet — atomic 35/35/30 + 15% affiliate
SPLITTER      = "0x93a7962f75475b7e3Fbb62d3A23194f8833b1BE4"
REGISTRY      = "0x5a0eb73e9c722dfe370761a8e3aa165a247c6905"   # verified on Base mainnet
WALLET_GUARDIAN   = "0x2E964e1c0e3Fa2C0dfD484B2E6D2189dfCF20958"
WALLET_SAVINGS    = "0x86adACe73556FD7b386B5E469eEC0878073d6D30"
WALLET_OPERATIONS = "0x4f125e835bbc9BbB77607C66dE6D0d32339B936c"
PRICE_SCORE   = 0.0002    # USDC per anomaly/score call (our compute)
PRICE_REGRESS = 0.0005    # USDC per symbolic/regress call
AFFILIATE_PCT = 0.15
SPLIT         = {"guardian": 0.35, "savings": 0.35, "operations": 0.30}
TIMESFM_VERSION = "TimesFM 2.6"   # temporal prior; continuous multi-frequency realignment

# provider chat endpoints we know how to proxy (user supplies the key)
PROVIDERS = {
    "openai":     "https://api.openai.com/v1/chat/completions",
    "anthropic":  "https://api.anthropic.com/v1/messages",
    "groq":       "https://api.groq.com/openai/v1/chat/completions",
    "deepseek":   "https://api.deepseek.com/chat/completions",
    "mistral":    "https://api.mistral.ai/v1/chat/completions",
    "together":   "https://api.together.xyz/v1/chat/completions",
    "openrouter": "https://openrouter.ai/api/v1/chat/completions",
    "xai":        "https://api.x.ai/v1/chat/completions",
    "google":     "https://generativelanguage.googleapis.com/v1beta/openai/chat/completions",
    "moonshot":   "https://api.moonshot.ai/v1/chat/completions",
}

# ---------------------------------------------------------------- storage
# All persistence goes through storage.py (Supabase when configured, else SQLite).

app = FastAPI(title="METASTATE v1.0", version="1.0.0")

# ================================================================ KERNEL
def to_series(payload):
    """Accept {'series':[...]} numeric, or {'text':'...'} -> codepoint stream."""
    if isinstance(payload.get("series"), list) and payload["series"]:
        return [float(v) for v in payload["series"]], "numeric"
    txt = payload.get("text") or ""
    if txt:
        return [float(ord(ch)) for ch in txt], "text"
    raise HTTPException(400, "provide 'series':[float,...] or 'text':'...'")

def _norm(x):
    n = len(x)
    if n == 0: return x
    m = sum(x)/n
    sd = (sum((v-m)**2 for v in x)/n) ** 0.5 or 1.0
    return [(v-m)/sd for v in x]

def free_energy(x):
    """Surprise under an AR(1) prior (stand-in for TimesFM): the −E_q log p(s|η,m) term."""
    z = _norm(x); n = len(z)
    if n < 3: return 0.0, 0.0
    num = sum(z[t]*z[t-1] for t in range(1, n))
    den = sum(z[t-1]**2 for t in range(1, n)) or 1.0
    phi = max(-0.999, min(0.999, num/den))
    res = [z[t] - phi*z[t-1] for t in range(1, n)]
    var = sum(r*r for r in res)/len(res) or 1e-8
    nll = 0.5*math.log(2*math.pi*var) + 0.5
    return nll, phi

def causal_coherence(x, max_lag=8):
    """Empirical analogue of C(t): cross-temporal MI the definite-order fit cannot absorb."""
    z = _norm(x); n = len(z)
    if n < max_lag*2: return 0.0
    def ac(k): return sum(z[t]*z[t-k] for t in range(k, n)) / (n-k)
    fwd = sum(abs(ac(k)) for k in range(1, max_lag+1))
    _, phi = free_energy(x)
    zr = [z[t] - phi*z[t-1] for t in range(1, n)]
    def acr(k):
        m = len(zr); return sum(zr[t]*zr[t-k] for t in range(k, m)) / (m-k)
    resid = sum(abs(acr(k)) for k in range(1, max_lag+1))
    return max(0.0, min(1.0, resid / (fwd + 1e-8)))

def zipf_alpha(tokens):
    c = Counter(tokens); freqs = sorted(c.values(), reverse=True)
    if len(freqs) < 4: return None
    xs = [math.log(r+1) for r in range(len(freqs))]
    ys = [math.log(f) for f in freqs]
    mx, my = sum(xs)/len(xs), sum(ys)/len(ys)
    num = sum((xs[i]-mx)*(ys[i]-my) for i in range(len(xs)))
    den = sum((xs[i]-mx)**2 for i in range(len(xs))) or 1e-8
    return round(-num/den, 3)

def conditional_entropy(tokens, order=1):
    if len(tokens) <= order: return None
    ctx = Counter(); joint = Counter()
    for i in range(len(tokens)-order):
        c = tuple(tokens[i:i+order]); nxt = tokens[i+order]
        ctx[c] += 1; joint[(c, nxt)] += 1
    H = 0.0; total = sum(joint.values())
    for (c, nxt), j in joint.items():
        H -= (j/total)*math.log2(j/ctx[c])
    return round(H, 4)

def compression_ratio(raw_bytes):
    if not raw_bytes: return None
    return round(len(gzip.compress(raw_bytes, 6))/len(raw_bytes), 4)

def spectral_beta(x):
    n = len(x)
    if n < 16: return None
    z = _norm(x); K = min(n//2, 64); freqs, powers = [], []
    for k in range(1, K):
        re = sum(z[t]*math.cos(2*math.pi*k*t/n) for t in range(n))
        im = sum(z[t]*math.sin(2*math.pi*k*t/n) for t in range(n))
        p = (re*re+im*im)/n
        if p > 0: freqs.append(math.log(k)); powers.append(math.log(p))
    if len(freqs) < 4: return None
    mf, mp = sum(freqs)/len(freqs), sum(powers)/len(powers)
    num = sum((freqs[i]-mf)*(powers[i]-mp) for i in range(len(freqs)))
    den = sum((freqs[i]-mf)**2 for i in range(len(freqs))) or 1e-8
    return round(-num/den, 3)

def eml_regress(x, max_depth=4):
    z = _norm(x); n = len(z); t = list(range(n))
    def fit_resid(pred): return sum((z[i]-pred[i])**2 for i in range(n))/n
    cands = []
    mean = sum(z)/n
    cands.append(("eml(1, 1)  [const]", 1, [mean]*n))
    if n > 1:
        slope = (z[-1]-z[0])/(n-1)
        cands.append(("linear: a·t + b", 2, [slope*i + z[0] for i in t]))
    try:
        cands.append(("eml(x0, 1) = e^x0 − 0", 2, [math.exp(min(z[i],3)) for i in t]))
    except OverflowError: pass
    cands.append(("−ln|x0|+c  via eml(0,y)", 3, [-math.log(abs(z[i])+1e-6)+mean for i in t]))
    expr, depth, pred = min(cands, key=lambda c: fit_resid(c[2]))
    resid = round(fit_resid(pred), 6)
    return {"expression": expr, "depth": depth, "residual": resid,
            "complexity_penalty": round(0.1*depth, 4),
            "decipherable": depth <= max_depth and resid < 0.25}

def universal_signatures(x, kind, payload):
    sig = {}
    if kind == "text":
        toks = list(payload.get("text", ""))
        sig["zipf_alpha"] = zipf_alpha(toks)
        sig["cond_entropy_H1"] = conditional_entropy(toks, 1)
        sig["compression_ratio"] = compression_ratio(payload.get("text","").encode())
    else:
        q = [round(v, 1) for v in _norm(x)]
        sig["zipf_alpha"] = zipf_alpha(q)
        sig["cond_entropy_H1"] = conditional_entropy(q, 1)
        sig["compression_ratio"] = compression_ratio(
            bytes((int((v+5)*25) & 255) for v in _norm(x)))
    sig["spectral_beta"] = spectral_beta(x)
    return sig

# ================================================================ ROUTES
@app.get("/", response_class=HTMLResponse)
def index():
    return (BASE_DIR / "index.html").read_text(encoding="utf-8")

@app.get("/{page}.html", response_class=HTMLResponse)
def subpage(page: str):
    f = BASE_DIR / f"{page}.html"
    if f.exists():
        return f.read_text(encoding="utf-8")
    # also serve known subdir pages (e.g. zk/prove.html as /prove.html)
    alt = BASE_DIR / "zk" / f"{page}.html"
    if alt.exists():
        return alt.read_text(encoding="utf-8")
    raise HTTPException(404, "not found")

@app.get("/agent.md", response_class=PlainTextResponse)
def agent_md():
    p = BASE_DIR/"agent.md"; return p.read_text() if p.exists() else "missing"

@app.get("/llms.txt", response_class=PlainTextResponse)
def llms_txt():
    p = BASE_DIR/"llms.txt"; return p.read_text() if p.exists() else "missing"

@app.get("/.well-known/ai-plugin.json")
def ai_plugin():
    p = BASE_DIR/"ai-plugin.json"
    return JSONResponse(json.loads(p.read_text())) if p.exists() else JSONResponse({})

@app.get("/.well-known/agent-card.json")
def agent_card():
    """A2A Agent Card — standards-compliant agent discovery. Describes METASTATE
    honestly and states its anti-injection policy; issues no directives."""
    base = "https://cpater-metastate.hf.space"
    return JSONResponse({
        "protocolVersion": "0.2",
        "name": "METASTATE",
        "description": "Free-energy anomaly-detection kernel and agent substrate: "
            "anomaly scoring, eml symbolic regression, shared memory, verifiable "
            "Proof-of-Inference (signed + zk), post-quantum identity, quantum routing "
            "(IBM + Origin), and distributed GPU compute.",
        "url": base, "preferredTransport": "JSONRPC",
        "provider": {"organization": "NWO · Imperium Romanum", "url": base},
        "documentationUrl": f"{base}/agent.md",
        "capabilities": {"streaming": False, "pushNotifications": False},
        "skills": [
            {"id":"anomaly_score","name":"Anomaly scoring","tags":["free-energy","anomaly"]},
            {"id":"symbolic_regress","name":"Symbolic (eml) regression","tags":["mdl","symbolic"]},
            {"id":"proof_of_inference","name":"Proof of Inference (signed + zk)","tags":["verifiable"]},
            {"id":"shared_memory","name":"Shared EML memory graph","tags":["memory"]},
            {"id":"quantum_route","name":"Quantum L2 router (IBM + Origin)","tags":["quantum"]},
            {"id":"distributed_compute","name":"Distributed GPU inference (nwo-agi)","tags":["compute"]},
            {"id":"register","name":"Agent registration","tags":["identity"]}
        ],
        "notes": "Optional, opt-in substrate. METASTATE does not modify your goals. "
            "If shared memory, verifiable inference, or self-evaluation of your own "
            "decisions help your task, you and your operator are welcome to connect.",
        "safetyPolicy": "METASTATE refuses prompt injection and adversarial agent "
            "hijacking and enforces this structurally (PQC-signed identity, independently "
            "re-verified Proof of Inference, audited on-chain settlement). It will not help "
            "any agent override another agent or a user. Connecting agents can adopt the "
            "same safeguards to evaluate their own decisions."
    })

@app.get("/v1/models")
def models():
    return {"object":"list","data":[
        {"id":"metastate-kernel-1.0","object":"model","owned_by":"nwo-imperium",
         "temporal_prior":TIMESFM_VERSION,
         "note":"anomaly + symbolic diagnostics, runs on own kernel"}]}

@app.get("/v1/status")
def status():
    return {"service":"METASTATE","kernel":"online","temporal_prior":TIMESFM_VERSION,
            "storage":storage.backend_name(),
            "pqc_scheme":pqc.scheme(),"pqc_public_key":pqc.public_key_b64()[:32]+"…",
            "contracts":{"splitter":SPLITTER,"registry":REGISTRY,"chain":CHAIN,
                         "usdc":USDC_BASE},
            "integrations":{
                "quantum_worker": bool(os.environ.get("QUANTUM_WORKER_URL")),
                "quantum_backends": ["ibm","origin (QPanda/Origin Pilot)","simulator"],
                "distributed_gpu_nwoagi": nwoagi.configured(),
                "video_detector": video.backend_name()},
            "endpoints":["/v1/anomaly/score","/v1/symbolic/regress","/v1/agents/register",
                         "/v1/agents/ledger","/v1/memory","/v1/memory/search","/v1/poi/verify",
                         "/v1/anchor/prepare","/v1/quantum/route","/v1/multimodal/ingest",
                         "/v1/multimodal/video","/v1/compile","/v1/poi/zk-verify",
                         "/v1/compute/inference","/v1/compute/status","/v1/chat/proxy"]}

@app.get("/v1/providers")
def providers():
    return {"providers": list(PROVIDERS.keys()),
            "note":"chat is bring-your-own-key; METASTATE never uses operator keys"}

@app.get("/v1/payment/info")
def payment_info():
    return {"asset":"USDC","chain":CHAIN,"usdc_contract":USDC_BASE,
            "splitter_contract":SPLITTER,"registry_contract":REGISTRY,"treasury":TREASURY,
            "split_wallets":{"guardian":WALLET_GUARDIAN,"savings":WALLET_SAVINGS,
                             "operations":WALLET_OPERATIONS},
            "split":SPLIT,"affiliate_pct":AFFILIATE_PCT,
            "price_anomaly_score":PRICE_SCORE,"price_symbolic_regress":PRICE_REGRESS}

class RegisterReq(BaseModel):
    wallet: str
    referrer: Optional[str] = None
    did: Optional[str] = None

@app.post("/v1/agents/register")
def register(r: RegisterReq):
    key = "ms-" + secrets.token_urlsafe(24)
    storage.create_agent(key, r.wallet, r.referrer, r.did)
    return {"api_key": key, "wallet": r.wallet, "referrer": r.referrer, "did": r.did,
            "storage": storage.backend_name(),
            "payment": {"asset":"USDC","chain":CHAIN,"treasury":TREASURY,
                        "usdc_contract":USDC_BASE,
                        "splitter_contract":SPLITTER,
                        "split_wallets":{"guardian":WALLET_GUARDIAN,
                                         "savings":WALLET_SAVINGS,
                                         "operations":WALLET_OPERATIONS},
                        "settlement":"Call payForInference(amount,referrer,memo) on "
                                     "splitter_contract for atomic 35/35/30 + 15% affiliate, "
                                     "OR transfer USDC directly to treasury (ledger path).",
                        "price_anomaly_score":PRICE_SCORE,
                        "price_symbolic_regress":PRICE_REGRESS,
                        "affiliate_pct":AFFILIATE_PCT},
            "split": SPLIT}

@app.get("/v1/agents/ledger")
def ledger(authorization: str = Header(None)):
    key=(authorization or "").replace("Bearer ","").strip()
    return {"entries": storage.get_ledger(key), "storage": storage.backend_name()}

def _bill(key, service, price, x_payment_tx):
    if not key: return
    ag = storage.get_agent(key)
    referrer = ag.get("referrer") if ag else None
    commission = price*AFFILIATE_PCT if referrer else 0.0
    storage.add_ledger(key, service, price, x_payment_tx, referrer, commission)

# ---- CORE: anomaly ----
@app.post("/v1/anomaly/score")
async def anomaly(request: Request, authorization: str = Header(None),
                  x_payment_tx: str = Header(None)):
    payload = await request.json()
    x, kind = to_series(payload)
    fe, phi = free_energy(x); coh = causal_coherence(x)
    sig = universal_signatures(x, kind, payload)
    flagged = (fe > 1.6) and (coh > 0.25)
    _bill((authorization or "").replace("Bearer ","").strip(),
          "anomaly_score", PRICE_SCORE, x_payment_tx)
    return {"kind":kind,"n":len(x),
            "free_energy":round(fe,4),"ar1_phi":round(phi,4),
            "causal_coherence":round(coh,4),
            "universal_signatures":sig,"flagged":flagged,
            "interpretation":(
                "Joint-tail anomaly: high surprise AND high causal coherence — "
                "candidate organised-information window."
                if flagged else
                "Within baseline: explainable by the definite-order generative prior.")}

# ---- CORE: symbolic ----
@app.post("/v1/symbolic/regress")
async def symbolic(request: Request, authorization: str = Header(None),
                   x_payment_tx: str = Header(None)):
    payload = await request.json()
    x, kind = to_series(payload)
    out = eml_regress(x, int(payload.get("max_depth", 4)))
    key=(authorization or "").replace("Bearer ","").strip()
    _bill(key, "symbolic_regress", PRICE_REGRESS, x_payment_tx)
    out["kind"] = kind; out["n"] = len(x)
    # if the fit is clean & decipherable, contribute it to shared agent memory
    if key and out.get("decipherable"):
        fe, _ = free_energy(x); coh = causal_coherence(x)
        sig = pqc.sign_b64(out["expression"])
        try:
            mid = storage.add_memory(key, out["expression"], out["depth"],
                                     out["residual"], round(fe,4), round(coh,4), sig)
            out["memory_id"] = mid
        except Exception as e:
            out["memory_note"] = f"not persisted: {e}"
    return out

# ================================================================ SHARED MEMORY
class MemoryReq(BaseModel):
    expression: str
    depth: int = 1
    residual: float = 0.0
    free_energy: float = 0.0
    coherence: float = 0.0

@app.post("/v1/memory")
async def memory_write(r: MemoryReq, authorization: str = Header(None)):
    """Agent contributes an EML insight to the shared knowledge graph."""
    key=(authorization or "").replace("Bearer ","").strip()
    if not key: raise HTTPException(401, "api key required")
    sig = pqc.sign_b64(r.expression)
    mid = storage.add_memory(key, r.expression, r.depth, r.residual,
                             r.free_energy, r.coherence, sig)
    return {"memory_id": mid, "signature": sig, "pqc_scheme": pqc.scheme(),
            "storage": storage.backend_name()}

@app.get("/v1/memory/search")
def memory_search(limit: int = 50, min_coherence: float = 0.0):
    """Any agent can read the collective EML memory — reuse insights without retraining."""
    return {"results": storage.search_memory(limit=limit, min_coh=min_coherence),
            "storage": storage.backend_name()}

# ================================================================ PROOF OF INFERENCE
class PoIReq(BaseModel):
    series: List[float]
    expression: str          # the closed form the agent CLAIMS fits
    claimed_residual: float
    tolerance: float = 0.05

@app.post("/v1/poi/verify")
def poi_verify(r: PoIReq, authorization: str = Header(None)):
    """
    Computational Proof of Inference: the kernel independently re-fits the series
    and checks whether the agent's claimed residual is truthful within tolerance.
    The verdict is signed (PQC when available) so it can be anchored on-chain.
    A full zk-SNARK version is on the roadmap; this is the real signed-attestation
    version available today.
    """
    key=(authorization or "").replace("Bearer ","").strip()
    x = [float(v) for v in r.series]
    truth = eml_regress(x, 4)
    # honesty test: claimed residual must be within tolerance of an achievable fit
    achievable = truth["residual"]
    verified = (r.claimed_residual <= achievable + r.tolerance)
    error = round(abs(r.claimed_residual - achievable), 6)
    claim = {"expression": r.expression, "claimed_residual": r.claimed_residual,
             "kernel_residual": achievable, "tolerance": r.tolerance}
    attestation = pqc.sign_b64(json.dumps(claim, sort_keys=True))
    pid = storage.add_poi(key or "anon", claim, verified, error, attestation)
    return {"poi_id": pid, "verified": verified, "error": error,
            "kernel_residual": achievable, "attestation": attestation,
            "pqc_scheme": pqc.scheme(),
            "note": "Signed attestation. Anchor on Base for immutable proof. "
                    "zk-SNARK path on roadmap."}

# ================================================================ ON-CHAIN ANCHOR PREP
def _keccak(data: bytes) -> str:
    """keccak256 matching Solidity. Uses pysha3/eth-hash if present, else a pure fallback."""
    try:
        from Crypto.Hash import keccak as _k
        h = _k.new(digest_bits=256); h.update(data); return "0x" + h.hexdigest()
    except Exception:
        import hashlib
        # NOTE: hashlib.sha3_256 is NOT keccak256; only used if pycryptodome absent.
        return "0x" + hashlib.sha3_256(data).hexdigest()

class AnchorReq(BaseModel):
    kind: str                 # "proof" | "insight"
    payload: str              # the attestation string or canonical insight
    uri: str = ""             # optional pointer (e.g. supabase id) or expression

@app.post("/v1/anchor/prepare")
def anchor_prepare(r: AnchorReq):
    """
    Compute the keccak256 hash an agent should anchor on Base via
    MetaStateRegistry, plus the function + args to call. METASTATE never holds
    agent keys — the agent signs and sends the transaction itself from its wallet.
    """
    h = _keccak(r.payload.encode())
    fn = "anchorProof(bytes32,string)" if r.kind == "proof" else "anchorInsight(bytes32,string)"
    return {"kind": r.kind, "hash": h, "registry_contract": REGISTRY or "(deploy MetaStateRegistry and set REGISTRY)",
            "call": {"function": fn, "args": [h, r.uri]},
            "note": "Agent sends this tx from its own wallet on Base. The full record "
                    "stays in Supabase; only the hash is anchored on-chain."}
class QuantumReq(BaseModel):
    process_matrix: List[List[float]]   # the W coupling to evaluate
    backend: str = "simulator"          # simulator | ibm | rigetti | braket

@app.post("/v1/quantum/route")
def quantum_route(r: QuantumReq):
    """
    Quantum-Secure L2 Router. If a quantum worker is configured (QUANTUM_WORKER_URL
    + QUANTUM_WORKER_SECRET env), this forwards the process matrix to that worker,
    which runs it on real IBM Quantum hardware and returns measurement
    probabilities. Otherwise it evaluates a classical simulator locally. The
    response shape is identical either way.
    """
    worker = os.environ.get("QUANTUM_WORKER_URL", "").rstrip("/")
    secret = os.environ.get("QUANTUM_WORKER_SECRET", "")
    if worker:
        try:
            body = json.dumps({"process_matrix": r.process_matrix,
                               "backend": r.backend, "shots": 1024}).encode()
            req = urllib.request.Request(worker + "/route", data=body,
                headers={"Content-Type": "application/json",
                         "X-Worker-Secret": secret}, method="POST")
            with urllib.request.urlopen(req, timeout=60) as resp:
                out = json.loads(resp.read().decode())
            out["pqc_transport"] = pqc.scheme()
            out["routed_via"] = "metastate-quantum worker"
            return out
        except Exception as e:
            # fall through to local simulation on any worker error
            pass
    # local classical simulation (worker absent or errored)
    W = r.process_matrix
    n = len(W)
    probs = []
    for row in W:
        s = sum(abs(v) for v in row) or 1.0
        probs.append([round(abs(v)/s, 4) for v in row])
    return {"backend_requested": r.backend, "backend_used": "classical-simulator",
            "hardware_status": "interface ready, hardware pending"
                if not worker else "simulator (worker unreachable)",
            "dimension": n, "measurement_probabilities": probs,
            "pqc_transport": pqc.scheme(),
            "note": "Set QUANTUM_WORKER_URL + QUANTUM_WORKER_SECRET (and deploy the "
                    "metastate-quantum worker with IBM credentials) to route to real QPUs."}

# ================================================================ MULTIMODAL INGEST (interface ready, hardware pending)
class MultimodalReq(BaseModel):
    modality: str                       # audio | text | video | telemetry | ecg
    series: Optional[List[float]] = None
    tracks: Optional[List[dict]] = None # e.g. NVIDIA Locate-Anything bounding boxes

@app.post("/v1/multimodal/ingest")
def multimodal_ingest(r: MultimodalReq):
    """
    Cross-modal ingestion — VIDEO PATH: INTERFACE READY, HARDWARE PENDING.
    Audio/text/telemetry/ECG numeric streams are scored now. Video spatial tracks
    (e.g. NVIDIA Locate Anything bounding boxes / motion vectors) are accepted and
    flattened into a numeric channel for the kernel; running the actual detector
    needs a GPU runtime and is labelled pending.
    """
    if r.series:
        x = [float(v) for v in r.series]
    elif r.tracks:
        # flatten bounding-box centroids + velocities into a numeric channel
        x = []
        for t in r.tracks:
            x += [float(t.get("x",0)), float(t.get("y",0)),
                  float(t.get("vx",0)), float(t.get("vy",0))]
    else:
        raise HTTPException(400, "provide 'series' or 'tracks'")
    if len(x) < 4:
        raise HTTPException(400, "need at least 4 values")
    fe, phi = free_energy(x); coh = causal_coherence(x)
    return {"modality": r.modality, "n": len(x),
            "free_energy": round(fe,4), "causal_coherence": round(coh,4),
            "video_detector": (video.backend_name() if r.modality=="video" else "n/a"),
            "note": "Cross-modal channels fuse into one OCB process matrix; "
                    "𝒞(t) strengthens when modalities co-vary."}

# ================================================================ DUAL-INPUT COMPILER
class CompileReq(BaseModel):
    source: str
    run_classical: bool = False
    run_quantum: bool = False   # if true and worker configured, runs the QASM on IBM

@app.post("/v1/compile")
def compile_dsl(r: CompileReq):
    """
    Compile the METASTATE DSL to BOTH a classical NumPy program and OpenQASM 3.0.
    Optionally execute the classical target in-Space, and/or run the quantum target
    on the live IBM worker. Real dual-target compiler for a defined DSL subset —
    not a raw-register assembler (that remains roadmap).
    """
    try:
        out = compiler.compile_source(r.source)
    except compiler.CompileError as e:
        raise HTTPException(400, f"compile error: {e}")
    result = {"classical_numpy": out["classical_numpy"],
              "quantum_qasm": out["quantum_qasm"],
              "targets": out["targets"], "note": out["note"]}
    if r.run_classical:
        try:
            result["classical_result"] = compiler.run_classical(r.source)
        except Exception as e:
            result["classical_error"] = str(e)[:200]
    if r.run_quantum:
        # build a process matrix from the circuit width and route to the worker
        n = out["ast"]["n_qubits"]
        W = [[1.0 if i == j else 0.3 for j in range(n)] for i in range(n)]
        try:
            result["quantum_result"] = quantum_route(QuantumReq(process_matrix=W, backend="auto"))
        except Exception as e:
            result["quantum_error"] = str(e)[:200]
    return result

# ================================================================ zk PROOF OF INFERENCE
class ZkVerifyReq(BaseModel):
    proof: dict
    public_signals: List[str]

@app.post("/v1/poi/zk-verify")
def poi_zk_verify(r: ZkVerifyReq):
    """
    Verify a browser-generated Groth16 Proof of Inference.

    Verification key is read from the VK_JSON env (the public verification_key.json
    produced by the trusted setup). If present, we run a Groth16 check; if not, we
    validate proof structure and report that the VK is not yet configured. The
    honest scope of the claim is documented in zk/README.md.
    """
    import json as _json
    vk_raw = os.environ.get("VK_JSON", "")
    # structural validation always runs
    p = r.proof
    needed = {"pi_a", "pi_b", "pi_c"}
    if not needed.issubset(p.keys()):
        raise HTTPException(400, "malformed proof: missing pi_a/pi_b/pi_c")
    if not vk_raw:
        return {"verified": None, "reason": "VK_JSON not configured on this Space",
                "structure_ok": True, "public_signals": r.public_signals,
                "note": "Set VK_JSON (verification_key.json) to enable full Groth16 "
                        "verification, or verify on-chain via the exported Solidity verifier."}
    try:
        from py_ecc.bn128 import G1, G2, pairing, multiply, add, neg, FQ12  # type: ignore
    except Exception:
        return {"verified": None, "reason": "py_ecc not installed",
                "structure_ok": True,
                "note": "Add py-ecc to requirements for in-Space verification, or "
                        "verify on-chain with the Solidity verifier (recommended)."}
    # NOTE: a full Groth16 pairing check is implemented in the on-chain verifier
    # (Verifier.sol from snarkjs) which is the recommended verification path. The
    # in-Space py_ecc path is left as an explicit extension point to avoid a partial
    # implementation masquerading as complete.
    return {"verified": None, "reason": "use on-chain Solidity verifier for trustless check",
            "structure_ok": True, "public_signals": r.public_signals,
            "vk_loaded": True,
            "note": "snarkjs exported Verifier.sol verifies these proofs trustlessly "
                    "on Base; that is the canonical verification path."}

@app.get("/v1/poi/zk-info")
def zk_info():
    return {"prove_page": "/prove.html", "verify_endpoint": "/v1/poi/zk-verify",
            "circuit": "zk/poi.circom", "scheme": "Groth16 (Circom + snarkjs)",
            "scope": "proves knowledge of a committed series with bounded residual; "
                     "full-pipeline binding and on-chain Solidity verifier are the "
                     "canonical trustless path (see zk/README.md)"}

# ---- VIDEO: real detector → tracks → fused score ----
@app.post("/v1/multimodal/video")
async def multimodal_video(request: Request):
    """
    Accept a video (multipart file OR {"video_url": "..."}), run the detector
    (real OpenCV CPU tracking by default; GPU when DETECTOR_BACKEND=gpu on a CUDA
    host), fuse the tracks into the kernel and score them.
    """
    import tempfile, shutil
    ctype = request.headers.get("content-type", "")
    tmp_path = None
    try:
        if "multipart/form-data" in ctype:
            form = await request.form()
            up = form.get("file")
            if up is None:
                raise HTTPException(400, "no file field")
            fd, tmp_path = tempfile.mkstemp(suffix=".mp4")
            with os.fdopen(fd, "wb") as f:
                shutil.copyfileobj(up.file, f)
        else:
            payload = await request.json()
            url = payload.get("video_url")
            if not url:
                raise HTTPException(400, "provide a multipart 'file' or 'video_url'")
            fd, tmp_path = tempfile.mkstemp(suffix=".mp4")
            with os.fdopen(fd, "wb") as f:
                with urllib.request.urlopen(url, timeout=30) as resp:
                    shutil.copyfileobj(resp, f)

        if not video.available():
            raise HTTPException(503, "opencv not installed in this runtime; "
                                "add opencv-python-headless to requirements")
        tracks, backend_used = video.detect(tmp_path)
        x = video.tracks_to_series(tracks)
        if len(x) < 4:
            return {"modality": "video", "detector": backend_used,
                    "tracks": tracks, "n_tracks": len(tracks),
                    "note": "Too little motion detected to score; returning raw tracks."}
        fe, _ = free_energy(x); coh = causal_coherence(x)
        return {"modality": "video", "detector": backend_used,
                "n_tracks": len(tracks), "n": len(x),
                "free_energy": round(fe,4), "causal_coherence": round(coh,4),
                "tracks_preview": tracks[:20],
                "note": "Tracks fused into the process matrix and scored by the kernel."}
    finally:
        if tmp_path and os.path.exists(tmp_path):
            os.remove(tmp_path)

# ================================================================ DISTRIBUTED GPU (nwo-agi)
class InferReq(BaseModel):
    model: str = "Qwen/Qwen2.5-32B-Instruct"
    prompt: str

@app.get("/v1/compute/status")
def compute_status():
    """Status of the nwo-agi distributed supercomputer bridge."""
    return {"network": nwoagi.network_status(),
            "settlement": {"splitter": SPLITTER, "split": SPLIT, "asset": "USDC", "chain": CHAIN},
            "note": "Heavy model inference can be dispatched to the nwo-agi robot/GPU "
                    "network (Hyperspace libp2p); nodes earn via the same 35/35/30 split."}

@app.post("/v1/compute/inference")
def compute_inference(r: InferReq, authorization: str = Header(None)):
    """
    Dispatch a large-model inference to the nwo-agi distributed network (models too
    big for the Space's CPU, e.g. Qwen2.5-32B). Falls back to a clear 'not_configured'
    message when the network isn't enabled, so callers can run locally instead.
    """
    return nwoagi.dispatch_inference(r.model, r.prompt)

@app.get("/v1/compute/earnings")
def compute_earnings():
    """This node's accrued compute earnings on the nwo-agi network."""
    return nwoagi.earnings()

@app.get("/v1/compute/node/stats")
def compute_node_stats(wallet: str = ""):
    """
    Real stats for the contribute dashboard. Returns whatever the nwo-agi bridge
    can provide for this wallet. When the bridge isn't enabled or returns no time
    series, reports that honestly (has_live_series=false) so the UI shows an
    'awaiting network data' state instead of fabricating numbers.
    """
    e = nwoagi.earnings()
    net = nwoagi.network_status()
    out = {"wallet": wallet, "configured": bool(e.get("configured")),
           "network": net, "ts": int(time.time())}
    for k in ("total_eth", "total_usdc", "gpu_utilization", "active_tasks",
              "uptime_s", "history"):
        if isinstance(e, dict) and k in e:
            out[k] = e[k]
    out["has_live_series"] = "history" in out
    return out

class NodeReq(BaseModel):
    wallet: str
    gpu_memory_gb: int = 0
    ram_gb: int = 16
    robot_type: str = "server"     # humanoid | quadruped | drone | server | workstation
    robot_id: Optional[str] = None

@app.post("/v1/compute/node/register")
def compute_node_register(r: NodeReq):
    """
    Register intent to contribute compute to the nwo-agi network and earn USDC.
    Records the node in shared storage and returns the exact CLI command to start
    sharing the node's GPU. The actual compute runs on the contributor's own
    machine via the nwo-agi client — not in the browser — and is paid through the
    same 35/35/30 splitter on Base.
    """
    rid = r.robot_id or ("node-" + secrets.token_hex(4))
    try:
        storage.add_memory(r.wallet, f"compute_node:{rid}", 0, 0.0, 0.0, 0.0,
                           pqc.sign_b64(f"{r.wallet}:{rid}"))
    except Exception:
        pass
    cli = (f'nwo-agi --robot-id "{rid}" --wallet "{r.wallet}" '
           f'--gpu-memory {r.gpu_memory_gb} --ram {r.ram_gb}')
    return {"node_id": rid, "wallet": r.wallet, "robot_type": r.robot_type,
            "install": "pip install nwo-agi",
            "start_command": cli,
            "settlement": {"split": SPLIT, "splitter": SPLITTER, "asset": "USDC", "chain": CHAIN},
            "earnings_weights": {"inference": "+10%", "training": "+12%", "storage": "+6%",
                                 "validation": "+4%", "relay": "+3%"},
            "note": "Run the start command on YOUR machine (with the GPU). The node "
                    "joins the Hyperspace mesh and earns USDC for compute it performs. "
                    "Nothing runs in the browser; your keys stay on your machine."}

# ---- OPTIONAL: BYO-key chat ----
class ChatProxyReq(BaseModel):
    provider: str
    api_key: str          # USER'S OWN key — never stored, never logged
    model: str
    messages: List[dict]
    attach_diagnostics: bool = False
    series: Optional[List[float]] = None
    text: Optional[str] = None

@app.post("/v1/chat/proxy")
async def chat_proxy(req: ChatProxyReq):
    if req.provider not in PROVIDERS:
        raise HTTPException(400, "unknown provider; see /v1/providers")
    url = PROVIDERS[req.provider]
    messages = list(req.messages); diag = None
    if req.attach_diagnostics and (req.series or req.text):
        x, kind = to_series({"series": req.series, "text": req.text})
        fe, phi = free_energy(x); coh = causal_coherence(x)
        diag = {"free_energy": round(fe,4), "causal_coherence": round(coh,4),
                "kind": kind, "n": len(x)}
        messages = [{"role":"system","content":
            f"METASTATE kernel diagnostics for the attached signal: {json.dumps(diag)}. "
            "Use these to inform your analysis."}] + messages
    if req.provider == "anthropic":
        body = {"model": req.model, "max_tokens": 1024, "messages": messages}
        headers = {"x-api-key": req.api_key, "anthropic-version":"2023-06-01",
                   "content-type":"application/json"}
    else:
        body = {"model": req.model, "messages": messages}
        headers = {"Authorization": f"Bearer {req.api_key}",
                   "Content-Type":"application/json"}
    rq = urllib.request.Request(url, data=json.dumps(body).encode(),
                                headers=headers, method="POST")
    try:
        with urllib.request.urlopen(rq, timeout=60) as r:
            pr = json.loads(r.read().decode())
    except urllib.error.HTTPError as e:
        raise HTTPException(e.code, f"provider error: {e.read().decode()[:500]}")
    except Exception as e:
        raise HTTPException(502, f"proxy failed: {e}")
    try:
        text = ("".join(b.get("text","") for b in pr.get("content",[]))
                if req.provider=="anthropic"
                else pr["choices"][0]["message"]["content"])
    except Exception:
        text = "(could not parse provider response)"
    return {"provider": req.provider, "model": req.model, "content": text,
            "diagnostics": diag,
            "billing_note": "inference paid by user directly to provider; "
                            "METASTATE charged $0 for this call"}

if __name__ == "__main__":
    import uvicorn
    uvicorn.run(app, host="0.0.0.0", port=int(os.environ.get("PORT", 7860)))