Merge branch 'main' into alexcheema/trust-remote-code-cli-flag

The constant is the default for built-in models with known model cards,
which are trusted. Custom models added via API already default to
trust_remote_code=False in ModelCard.fetch_from_hf(). The CLI flag
overrides custom models only.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

src/exo/main.py

@@ -261,6 +261,13 @@ def main():
     if args.offline:
         logger.info("Running in OFFLINE mode — no internet checks, local models only")
 
+    # Set trust_remote_code override env var for runner subprocesses
+    if args.trust_remote_code:
+        os.environ["EXO_TRUST_REMOTE_CODE"] = "1"
+        logger.warning(
+            "--trust-remote-code enabled: models may execute arbitrary code during loading"
+        )
+
     # Set FAST_SYNCH override env var for runner subprocesses
     if args.fast_synch is True:
         os.environ["EXO_FAST_SYNCH"] = "on"
@@ -285,6 +292,7 @@ class Args(CamelCaseModel):
     no_downloads: bool = False
     offline: bool = False
     fast_synch: bool | None = None  # None = auto, True = force on, False = force off
+    trust_remote_code: bool = False
 
     @classmethod
     def parse(cls) -> Self:
@@ -336,6 +344,11 @@ class Args(CamelCaseModel):
             action="store_true",
             help="Run in offline/air-gapped mode: skip internet checks, use only pre-staged local models",
         )
+        parser.add_argument(
+            "--trust-remote-code",
+            action="store_true",
+            help="Allow models to execute custom code during tokenizer loading (security-sensitive, CLI-only)",
+        )
         fast_synch_group = parser.add_mutually_exclusive_group()
         fast_synch_group.add_argument(
             "--fast-synch",

src/exo/routing/router.py

@@ -220,6 +220,8 @@ def get_node_id_keypair(
     Obtains the :class:`Keypair` associated with this node-ID.
     Obtain the :class:`PeerId` by from it.
     """
+    # TODO(evan): bring back node id persistence once we figure out how to deal with duplicates
+    return Keypair.generate()
 
     def lock_path(path: str | bytes | PathLike[str] | PathLike[bytes]) -> Path:
         return Path(str(path) + ".lock")
@@ -239,6 +241,6 @@ def get_node_id_keypair(
 
         # if no valid credentials, create new ones and persist
         with open(path, "w+b") as f:
-            keypair = Keypair.generate()
+            keypair = Keypair.generate_ed25519()
             f.write(keypair.to_bytes())
             return keypair

src/exo/shared/constants.py

@@ -8,12 +8,12 @@ _EXO_HOME_ENV = os.environ.get("EXO_HOME", None)
 
 
 def _get_xdg_dir(env_var: str, fallback: str) -> Path:
-    """Get XDG directory, prioritising EXO_HOME environment variable if its set. On non-Linux platforms, default to ~/.exo. Cache home always prefers .cache/exo"""
+    """Get XDG directory, prioritising EXO_HOME environment variable if its set. On non-Linux platforms, default to ~/.exo."""
 
     if _EXO_HOME_ENV is not None:
         return Path.home() / _EXO_HOME_ENV
 
-    if sys.platform != "linux" and env_var != "XDG_CACHE_HOME":
+    if sys.platform != "linux":
         return Path.home() / ".exo"
 
     xdg_value = os.environ.get(env_var, None)
@@ -54,9 +54,10 @@ DASHBOARD_DIR = (
 # Log files (data/logs or cache)
 EXO_LOG_DIR = EXO_CACHE_HOME / "exo_log"
 EXO_LOG = EXO_LOG_DIR / "exo.log"
+EXO_TEST_LOG = EXO_CACHE_HOME / "exo_test.log"
 
 # Identity (config)
-EXO_NODE_ID_KEYPAIR = EXO_CACHE_HOME / "node_id.keypair"
+EXO_NODE_ID_KEYPAIR = EXO_CONFIG_HOME / "node_id.keypair"
 EXO_CONFIG_FILE = EXO_CONFIG_HOME / "config.toml"
 
 # libp2p topics for event forwarding

src/exo/worker/engines/mlx/constants.py

@@ -13,5 +13,6 @@ KV_CACHE_BITS: int | None = None
 
 DEFAULT_TOP_LOGPROBS: int = 5
 
-# TODO: We should really make this opt-in, but Kimi requires trust_remote_code=True
+# True for built-in models with known model cards; custom models added via API default to False
+# and can be overridden with the --trust-remote-code CLI flag.
 TRUST_REMOTE_CODE: bool = True

src/exo/worker/engines/mlx/utils_mlx.py

@@ -291,10 +291,14 @@ def shard_and_load(
 
 def get_tokenizer(model_path: Path, shard_metadata: ShardMetadata) -> TokenizerWrapper:
     """Load tokenizer for a model shard. Delegates to load_tokenizer_for_model_id."""
+    trust_remote_code = (
+        shard_metadata.model_card.trust_remote_code
+        or os.environ.get("EXO_TRUST_REMOTE_CODE") == "1"
+    )
     return load_tokenizer_for_model_id(
         shard_metadata.model_card.model_id,
         model_path,
-        trust_remote_code=shard_metadata.model_card.trust_remote_code,
+        trust_remote_code=trust_remote_code,
     )