hermes-agent/agent/secret_scope.py
Ben Barclay f538470cf4 feat(gateway): multiplex phase 2 — fail-closed profile credential isolation (Workstream A)
The credential gate. When multiplexing is active, a profile's secrets resolve
from a context-local scope, never the process-global os.environ (which in a
multiplexer may hold another profile's keys, and is inherited by every
subprocess spawned with env=dict(os.environ)).

- agent/secret_scope.py: get_secret() backed by a secret-scope contextvar.
  FAIL-CLOSED: when multiplex is active and no scope is installed, an unscoped
  read RAISES UnscopedSecretError instead of falling back to os.environ — a
  missed/new call site crashes loudly at that line rather than leaking a
  cross-profile value. Genuinely-global vars (HERMES_*, PATH, kanban paths,
  …) keep reading os.environ via an allowlist. load_env_file/build_profile_
  secret_scope parse a profile .env into an isolated dict WITHOUT mutating
  os.environ. Off by default => transparent os.getenv behavior.
- hermes_cli/runtime_provider.py: all credential/provider/base-url reads go
  through _getenv -> get_secret.
- agent/credential_pool.py: env fallbacks route through get_secret (the
  ~/.hermes/.env-first preference is preserved and already profile-correct via
  the home override).
- tools/mcp_tool.py: MCP config  interpolation resolves through
  get_secret, so a server's  picks up the routed profile's value.
- gateway/run.py: set_multiplex_active() at GatewayRunner init; per-turn .env
  reload is a no-op for credentials in multiplex mode (secrets come from the
  scope, not global env); _profile_runtime_scope context manager combines the
  HERMES_HOME override + secret scope; _run_agent wraps _run_agent_inner in
  that scope (resolved via _resolve_profile_home_for_source) when multiplexing.

Propagates into the agent worker thread for free via the existing
copy_context() in _run_in_executor_with_context.

Tests: 13 unit (fail-closed, scope isolation, global allowlist, .env parsing
without environ mutation) + 7 E2E (runtime_provider + MCP interpolation prove
two profiles isolated, unscoped read raises, globals still read environ).
2026-06-19 07:34:15 -07:00

205 lines
8.4 KiB
Python

"""Profile-scoped credential resolution for multi-profile gateway multiplexing.
The multiplexing gateway serves many profiles from one process. Each profile
has its own ``.env`` with its own provider keys and platform tokens, so we
**cannot** union them into the process-global ``os.environ`` (that would leak
profile A's keys to profile B's turns, and to every subprocess spawned with
``env=dict(os.environ)``).
This module provides a fail-closed, context-local secret scope:
- ``set_secret_scope(mapping)`` installs the active profile's secrets for the
current task (a contextvar, so it propagates into the agent's worker thread
via ``copy_context()`` exactly like the HERMES_HOME override).
- ``get_secret(name)`` reads from that scope. When multiplexing is **active**
and no scope is set, it RAISES rather than silently falling back to
``os.environ`` — an un-migrated or newly-added call site fails loud at that
exact line instead of leaking another profile's value. When multiplexing is
**off** (the default), it transparently reads ``os.environ`` so the
single-profile gateway and every non-gateway caller behave exactly as before.
Design rationale lives in ``docs/design/multiplexing-gateway.md`` (Workstream A).
"""
from __future__ import annotations
import os
from contextvars import ContextVar, Token
from pathlib import Path
from typing import Dict, Mapping, Optional
# ── multiplex-active flag ────────────────────────────────────────────────
# Process-global: set once at gateway startup when gateway.multiplex_profiles
# is true. Governs whether get_secret() fails closed on an unscoped read.
# A plain module global (not a contextvar): it describes the deployment mode,
# not a per-task value.
_MULTIPLEX_ACTIVE: bool = False
def set_multiplex_active(active: bool) -> None:
"""Mark whether the process is running as a profile multiplexer.
Called once at gateway startup. When True, ``get_secret`` fails closed on
an unscoped read instead of falling back to ``os.environ``.
"""
global _MULTIPLEX_ACTIVE
_MULTIPLEX_ACTIVE = bool(active)
def is_multiplex_active() -> bool:
"""Return whether the process is running as a profile multiplexer."""
return _MULTIPLEX_ACTIVE
# ── the secret scope contextvar ──────────────────────────────────────────
_SECRET_SCOPE: ContextVar[Optional[Mapping[str, str]]] = ContextVar(
"_SECRET_SCOPE", default=None
)
class UnscopedSecretError(RuntimeError):
"""Raised when a secret is read in multiplex mode with no scope installed.
This is the fail-closed signal: it means a credential read reached
``get_secret`` without a profile scope active, which in a multiplexer would
otherwise leak whichever profile's value happened to be in ``os.environ``.
The fix is to wrap the call path in ``set_secret_scope(...)`` (the per-turn
/ per-adapter profile scope), not to widen the allowlist.
"""
def set_secret_scope(secrets: Optional[Mapping[str, str]]) -> Token:
"""Install the active profile's secret mapping for the current context.
Returns a token for ``reset_secret_scope``. Pass ``None`` to clear.
"""
return _SECRET_SCOPE.set(secrets)
def reset_secret_scope(token: Token) -> None:
"""Restore the previous secret scope."""
_SECRET_SCOPE.reset(token)
def current_secret_scope() -> Optional[Mapping[str, str]]:
"""Return the active secret mapping, or None when no scope is installed."""
return _SECRET_SCOPE.get()
# ── genuinely-global env vars (NOT per-profile secrets) ──────────────────
# These are process/deployment-level settings, not profile credentials. They
# legitimately live in os.environ and must keep reading from it even in
# multiplex mode — routing them through the fail-closed path would wrongly
# crash. Anything matching is read from os.environ regardless of scope.
#
# Membership test is by exact name OR prefix (see _is_global_env). Keep this
# list tight: when in doubt a value is a profile secret, not a global.
_GLOBAL_ENV_EXACT = frozenset({
# Hermes runtime / deployment
"HERMES_HOME", "HERMES_PROFILE", "HERMES_GATEWAY_LOCK_DIR",
"HERMES_MAX_ITERATIONS", "HERMES_MAX_TOKENS", "HERMES_API_TIMEOUT",
"HERMES_REDACT_SECRETS", "HERMES_NOUS_TIMEOUT_SECONDS",
"_HERMES_GATEWAY",
# OS / interpreter
"PATH", "HOME", "USER", "LANG", "LC_ALL", "TZ", "PWD", "SHELL", "TMPDIR",
"VIRTUAL_ENV", "PYTHONPATH", "SSL_CERT_FILE",
# Kanban paths (per-board, not per-profile-secret)
"HERMES_KANBAN_DB", "HERMES_KANBAN_WORKSPACES_ROOT", "HERMES_KANBAN_BOARD",
})
_GLOBAL_ENV_PREFIXES = (
"HERMES_KANBAN_",
"HERMES_TELEGRAM_", # tuning knobs (batch delays, fallback toggles) — NOT the token
"TERMINAL_", # terminal/sandbox backend settings
)
def _is_global_env(name: str) -> bool:
"""Return True for genuinely process-global (non-profile-secret) env vars."""
if name in _GLOBAL_ENV_EXACT:
return True
return any(name.startswith(p) for p in _GLOBAL_ENV_PREFIXES)
def get_secret(name: str, default: Optional[str] = None) -> Optional[str]:
"""Resolve a credential by env-var name, honoring the active profile scope.
Resolution order:
1. Genuinely-global vars (``_is_global_env``) always read ``os.environ`` —
they are deployment settings, not profile secrets.
2. When a secret scope is installed (multiplexed turn), read from it; an
absent key returns ``default``. The scope is authoritative — we do NOT
fall through to ``os.environ``, because in a multiplexer ``os.environ``
may hold another profile's value.
3. No scope installed:
- multiplex INACTIVE (default deployment): read ``os.environ`` —
identical to the legacy ``os.getenv`` behavior every caller had before.
- multiplex ACTIVE: FAIL CLOSED. Raise ``UnscopedSecretError`` so the
missing scope is caught loudly instead of leaking a cross-profile value.
"""
if _is_global_env(name):
val = os.environ.get(name)
return val if val is not None else default
scope = _SECRET_SCOPE.get()
if scope is not None:
val = scope.get(name)
return val if val is not None else default
if _MULTIPLEX_ACTIVE:
raise UnscopedSecretError(
f"get_secret({name!r}) called with no profile secret scope active "
f"while multiplexing is on. This credential read must run inside a "
f"set_secret_scope(...) block (the per-turn / per-adapter profile "
f"scope). Reading os.environ here would risk leaking another "
f"profile's value. See docs/design/multiplexing-gateway.md "
f"(Workstream A)."
)
val = os.environ.get(name)
return val if val is not None else default
def load_env_file(env_path: Path) -> Dict[str, str]:
"""Parse a ``.env`` file into a plain dict WITHOUT touching ``os.environ``.
Used to load a profile's secrets into an isolated mapping for
``set_secret_scope``. Mirrors python-dotenv's basic parsing (KEY=VALUE,
``export`` prefix, ``#`` comments, optional matching quotes) but never
mutates the process environment — that isolation is the whole point.
"""
secrets: Dict[str, str] = {}
try:
text = env_path.read_text(encoding="utf-8")
except (FileNotFoundError, OSError, UnicodeDecodeError):
return secrets
for raw in text.splitlines():
line = raw.strip()
if not line or line.startswith("#"):
continue
if line.startswith("export "):
line = line[len("export "):].lstrip()
if "=" not in line:
continue
key, _, value = line.partition("=")
key = key.strip()
if not key:
continue
value = value.strip()
if len(value) >= 2 and value[0] == value[-1] and value[0] in ("'", '"'):
value = value[1:-1]
secrets[key] = value
return secrets
def build_profile_secret_scope(hermes_home: Path) -> Dict[str, str]:
"""Build a profile's secret mapping from its ``<home>/.env``.
Returns a fresh dict (safe to install via ``set_secret_scope``). Genuinely
global vars are intentionally NOT copied in — ``get_secret`` reads those
from ``os.environ`` directly, so the scope holds only profile secrets.
"""
return load_env_file(Path(hermes_home) / ".env")