code/autoslm/server/auth.py

"""Bearer auth for the managed control plane.

User authentication is freesolo API keys only — there is no native key system. A bearer
token equal to the operator's shared ``FREESOLO_INTERNAL_KEY`` resolves to the service
identity; any other token is verified against the freesolo backend and (on success)
resolved to a per-token user identity. ``AUTOSLM_SKIP_NET`` is the offline bypass that
disables the network verify (used by tests / air-gapped runs).
"""

from __future__ import annotations

import os
import threading
import time
import urllib.error
import urllib.request

from . import db

# Operators set this to the shared freesolo internal key; a bearer token equal to it
# authenticates as the service identity (see db.ensure_internal_key).
INTERNAL_KEY_ENV = "FREESOLO_INTERNAL_KEY"

# Freesolo USER-key acceptance: a user who `slm login`s with a freesolo API key sends it as
# the bearer to this control plane. Any non-internal token is verified against the freesolo
# backend and (on success) resolved to a per-token identity. AUTOSLM_SKIP_NET disables the
# network call entirely (offline bypass).
FREESOLO_BASE_URL_ENV = "FREESOLO_BASE_URL"
DEFAULT_FREESOLO_BASE_URL = "https://api.freesolo.co"
_VERIFY_TIMEOUT_S = 5.0
_VERIFY_CACHE_TTL_S = 300.0  # short TTL so it isn't a backend round-trip per request
# Negative verdicts get a much SHORTER TTL than positives. The freesolo verify endpoint
# returns 401 not only for a genuinely-bad key but also when the backend converts an
# auth-LOOKUP infra exception (authenticate_api_key failure) into a 401 — a transient outage.
# Caching such a negative for the full 300s would lock out an otherwise-valid key for 5
# minutes after the backend recovers. A short negative TTL keeps persistent bad tokens
# rate-limited (~30s, so they don't hammer the backend) while letting a transient 401 clear
# quickly. Positives keep the long TTL.
_VERIFY_CACHE_NEG_TTL_S = 30.0
# Upper bound on a bearer token we'll cache/verify. Real freesolo API keys are short; an
# arbitrarily long bearer is rejected up front so it can't bloat _verify_cache (keyed by the
# raw token) or produce an oversized outbound Authorization header.
_MAX_TOKEN_LEN = 256

# In-process verify cache: token -> (verified_bool, expires_at). Caches positives AND
# negatives so a burst of requests for the same token hits the backend at most once per TTL.
# Bounded: pruned of expired entries on every write and capped at _VERIFY_CACHE_MAX so a
# stream of unique bearer tokens can't grow it without bound (each token is a distinct key).
_verify_cache: dict[str, tuple[bool, float]] = {}
_verify_cache_lock = threading.Lock()
_VERIFY_CACHE_MAX = 1024


def _prune_verify_cache_locked(now: float) -> None:
    """Drop expired entries, then cap the cache size (oldest-expiry first).

    Caller must hold ``_verify_cache_lock``. Keeps the cache from growing unbounded as
    many distinct bearer tokens are verified over time.
    """
    for tok in [t for t, (_v, exp) in _verify_cache.items() if exp <= now]:
        del _verify_cache[tok]
    if len(_verify_cache) >= _VERIFY_CACHE_MAX:
        # Still over the cap after dropping expired entries: evict the soonest-to-expire
        # (oldest) entries until we're back under the cap.
        for tok, _exp in sorted(_verify_cache.items(), key=lambda kv: kv[1][1])[
            : len(_verify_cache) - _VERIFY_CACHE_MAX + 1
        ]:
            del _verify_cache[tok]


def _freesolo_verify(token: str) -> bool:
    """Verify a token against the freesolo backend (cached, short TTL, network errors = False).

    Never raises and never makes a network call when AUTOSLM_SKIP_NET is set — a swallowed
    error or a skipped call is treated as "not authenticated" (returns False), never a 500."""
    # Reject obviously-invalid oversized tokens before they touch the cache or the network.
    if not token or len(token) > _MAX_TOKEN_LEN:
        return False
    now = time.time()
    with _verify_cache_lock:
        cached = _verify_cache.get(token)
        if cached is not None and cached[1] > now:
            return cached[0]
    # Offline guard: with AUTOSLM_SKIP_NET set we never touch the network; treat as unverified.
    if os.environ.get("AUTOSLM_SKIP_NET"):
        return False
    base = os.environ.get(FREESOLO_BASE_URL_ENV) or DEFAULT_FREESOLO_BASE_URL
    url = f"{base.rstrip('/')}/api/auth/verify"
    req = urllib.request.Request(url, headers={"Authorization": f"Bearer {token}"})
    try:
        with urllib.request.urlopen(req, timeout=_VERIFY_TIMEOUT_S) as resp:
            verified = resp.status == 200
    except urllib.error.HTTPError as exc:
        # Only a DEFINITIVE rejection (4xx other than 429) is a verdict worth caching as a bad
        # key. A 5xx or 429 is a transient backend hiccup — treat it like a network error
        # (return False WITHOUT caching) so a valid key isn't locked out for the whole TTL
        # while the backend is briefly unhealthy.
        if exc.code >= 500 or exc.code == 429:
            return False
        verified = False
    except (urllib.error.URLError, OSError, ValueError):
        # A TRANSIENT network/connection error is NOT a verdict: don't cache it, so a valid
        # key isn't locked out for the whole TTL after the backend recovers.
        return False
    with _verify_cache_lock:
        # Prune expired entries and cap the size before inserting so unbounded distinct
        # tokens can't grow the cache.
        _prune_verify_cache_locked(now)
        # Pick the TTL by verdict: positives last the full TTL; a negative (which may be a
        # transient backend 401 rather than a real rejection) expires quickly so a valid key
        # isn't locked out for 5 minutes after the backend recovers.
        ttl = _VERIFY_CACHE_TTL_S if verified else _VERIFY_CACHE_NEG_TTL_S
        _verify_cache[token] = (verified, now + ttl)
    return verified


def authenticate(authorization: str | None) -> dict | None:
    """Resolve an ``Authorization: Bearer ...`` header to a key row.

    Freesolo keys are the only user auth. When the operator has configured
    ``FREESOLO_INTERNAL_KEY``, that shared internal key resolves to a single service
    identity. Any other token is verified against the freesolo backend and (on success)
    resolved to a per-token user identity so a user who ``slm login``s with their freesolo
    key can drive the control plane. ``AUTOSLM_SKIP_NET`` short-circuits the verify (no
    network call -> the token is treated as unverified)."""
    if not authorization or not authorization.startswith("Bearer "):
        return None
    token = authorization.removeprefix("Bearer ").strip()
    internal = os.environ.get(INTERNAL_KEY_ENV)
    if internal and token == internal:
        return db.lookup_key(token) or db.ensure_internal_key(token)
    # Any non-internal token is a freesolo USER key: verify it against the freesolo backend.
    # (AUTOSLM_SKIP_NET => no network call, returns False => authenticate returns None.)
    if _freesolo_verify(token):
        # A verified freesolo key gets its own per-token run-ownership identity.
        return db.lookup_key(token) or db.ensure_external_key(token)
    return None