ccx-native 0.0.1 → 0.1.0

package/ccx/controllers/live_vars.py

@@ -0,0 +1,1706 @@
+"""Shared/durable Python variables for CCX-managed workers."""
+
+from __future__ import annotations
+
+import base64
+from collections.abc import Mapping, MutableMapping
+from contextlib import contextmanager
+from dataclasses import dataclass
+import importlib
+import io
+import json
+import keyword
+import os
+from pathlib import Path
+import queue
+import shlex
+import shutil
+import socket
+import stat
+import subprocess
+import sys
+import sysconfig
+import threading
+import time
+import uuid
+from typing import Any
+
+from ccx import constants
+from ccx.system import oslayer
+from ccx.errors import CapabilityUnsupported, InvalidRequest, SessionUnavailable
+from ccx.system.executable import current_executable_path
+from ccx.controllers.live_vars_runner import execute_target, strict_json, STATE_FILE_LIMIT
+from ccx.store.paths import accept_same_user, bind_private_unix_socket, _require_unix_socket_budget
+
+
+VARIABLE_SCHEMA_VERSION = 2
+VARIABLE_STATE_SCHEMA_VERSION = 1
+_WIRE_DUMPS = json.dumps
+_WIRE_LOADS = json.loads
+_NATIVE_FALLTHROUGH_MODULES = frozenset({"pip", "venv", "ensurepip"})
+
+
+def _root() -> Path:
+    from ccx.store.root import require_ccx_active
+    return require_ccx_active(constants.CCX_ROOT).pointer.generation_root / "variables"
+
+
+def sidecar_path(engine: str, short: str) -> Path:
+    store, owner = _env_worker(engine, short)
+    return store.pointer.generation_root / "variables" / engine / f"{owner.worker_id}.json"
+
+
+def sidecar_lock_path(engine: str, short: str) -> Path:
+    store, owner = _env_worker(engine, short)
+    return store.pointer.generation_root / "variables" / engine / f"{owner.worker_id}.lock"
+
+
+def states_dir() -> Path:
+    return _root() / "states"
+
+
+def state_path(state_id: str) -> Path:
+    return states_dir() / f"{_valid_state_id(state_id)}.json"
+
+
+def state_lock_path(state_id: str) -> Path:
+    return states_dir() / f"{_valid_state_id(state_id)}.lock"
+
+
+def _env_worker(engine: str, worker_id: str):
+    from ccx.store.root import require_ccx_active
+    from ccx.store.identity import EngineMismatch
+    store = require_ccx_active(constants.CCX_ROOT)
+    owner = store.resolve(worker_id)
+    if owner.engine != engine:
+        raise EngineMismatch("Worker uses a different engine; check the handle")
+    return store, owner
+
+
+def shell_env_path(engine: str, short: str) -> Path:
+    store, owner = _env_worker(engine, short)
+    return (store.pointer.generation_root / "variables" / "env" /
+            engine / f"{owner.worker_id}.sh")
+
+
+def claude_bridge_path(short: str) -> Path:
+    store, owner = _env_worker("claude", short)
+    return (store.pointer.generation_root / "variables" / "env" /
+            "claude" / f"{owner.worker_id}.bridge.json")
+
+
+def _owner_runtime(
+    basename: str | None = None,
+) -> tuple[Path, oslayer.RuntimeChoice]:
+    """Plan the full live-owner socket suffix before any filesystem writes."""
+    dirname = f"ccx-live-vars-{os.getuid()}"
+    suffix = (max(
+        len(os.fsencode(f"/{dirname}/{basename}")) + 1,
+        len(os.fsencode(f"/{dirname}/.XXXXXXXX/s")) + 1,
+    ) if basename else 0)
+    choice = oslayer.runtime_base(purpose="live_vars", suffix_bytes=suffix)
+    if choice is None:
+        raise PermissionError("no private owner-socket runtime base available")
+    return choice.path / dirname, choice
+
+
+def _owners_path() -> Path:
+    return _owner_runtime()[0]
+
+
+def owners_dir(
+    path: Path | None = None, *, runtime: oslayer.RuntimeChoice | None = None,
+) -> Path:
+    if path is None:
+        path, runtime = _owner_runtime()
+    if runtime is None:
+        runtime = oslayer.RuntimeChoice(path.parent)
+    from ccx.rpc.transport import _ensure_private_runtime_directory
+    _ensure_private_runtime_directory(path, runtime)
+    st = os.lstat(path)
+    if not stat.S_ISDIR(st.st_mode) or st.st_uid != os.getuid():
+        raise PermissionError(
+            f"live-vars owner directory is not private and user-owned: {path}"
+        )
+    os.chmod(path, 0o700)
+    return path
+
+
+def _atomic_json(path: Path, value: dict[str, Any]) -> None:
+    """Main's atomic state replacement inside the validated private generation."""
+    text = strict_json(value)
+    if len(text.encode("utf-8")) > STATE_FILE_LIMIT:
+        raise InvalidRequest("worker variables exceed the 4 MiB state limit")
+    tmp = path.with_name(f".{path.name}.{os.getpid()}.{uuid.uuid4().hex}.tmp")
+    tmp.write_text(text, encoding="utf-8")
+    try:
+        os.chmod(tmp, 0o600)
+        os.replace(tmp, path)
+    finally:
+        tmp.unlink(missing_ok=True)
+
+
+@contextmanager
+def _worker_lock(engine: str, short: str):
+    from ccx.store.root import _flock_at
+    store, owner = _env_worker(engine, short)
+    with store.lease():
+        with store._gen("variables", engine, create=True) as directory:
+            with _flock_at(directory, owner.worker_id + ".lock",
+                           exclusive=True, create=True):
+                yield
+
+
+@contextmanager
+def _state_lock(state_id: str, *, create: bool = False):
+    from ccx.store.root import _flock_at, require_ccx_active
+    store = require_ccx_active(constants.CCX_ROOT)
+    with store.lease():
+        with store._gen("variables", "states", create=create) as directory:
+            with _flock_at(directory, _valid_state_id(state_id) + ".lock",
+                           exclusive=True, create=create):
+                yield
+
+
+def _valid_state_id(state_id: Any) -> str:
+    value = str(state_id or "")
+    if len(value) != 32 or any(ch not in "0123456789abcdef" for ch in value):
+        raise SessionUnavailable(f"invalid durable variable state id: {value!r}")
+    return value
+
+
+def _new_state_id() -> str:
+    return uuid.uuid4().hex
+
+
+def read_sidecar(engine: str, short: str) -> dict[str, Any] | None:
+    """Read a binding only after resolving its exact global full-UUID owner."""
+    from ccx.store.root import _read_json, StorageIntegrityError
+    store, owner = _env_worker(engine, short)
+    with store.lease():
+        try:
+            with store._gen("variables", engine) as directory:
+                value = _read_json(directory, owner.worker_id + ".json")
+        except FileNotFoundError:
+            return None
+    if value is not None and (
+        value.get("schemaVersion") != VARIABLE_SCHEMA_VERSION
+        or value.get("engine") != owner.engine
+        or value.get("workerId") != owner.worker_id
+        or value.get("mode") not in {"live", "durable"}
+    ):
+        raise StorageIntegrityError("Worker variable reference is incomplete; use the full handle")
+    return value
+
+
+def read_state(state_id: str) -> dict[str, Any]:
+    from ccx.store.root import _read_json, require_ccx_active, StorageIntegrityError
+    store = require_ccx_active(constants.CCX_ROOT)
+    with store.lease():
+        try:
+            with store._gen("variables", "states") as directory:
+                value = _read_json(directory, _valid_state_id(state_id) + ".json")
+        except FileNotFoundError:
+            value = None
+    if (value is None or value.get("schemaVersion") != VARIABLE_STATE_SCHEMA_VERSION
+            or value.get("stateId") != state_id
+            or not isinstance(value.get("names"), list)
+            or not isinstance(value.get("variables"), dict)):
+        raise StorageIntegrityError("Worker variables are unavailable; inspect the worker")
+    names, variables = value["names"], value["variables"]
+    if (any(not isinstance(name, str) or not name.isidentifier()
+            or keyword.iskeyword(name) for name in names)
+            or len(set(names)) != len(names)
+            or set(names) != set(variables)):
+        raise StorageIntegrityError("Worker variable names do not match saved values; inspect them")
+    return value
+
+
+def _new_state_document(
+    state_id: str,
+    *,
+    names: list[str],
+    variables: dict[str, Any],
+) -> dict[str, Any]:
+    return {
+        "schemaVersion": VARIABLE_STATE_SCHEMA_VERSION,
+        "stateId": _valid_state_id(state_id),
+        "names": sorted(names),
+        "variables": _json_clone(variables, "durable variables"),
+    }
+
+
+def durable_state_id(
+    engine: str,
+    short: str,
+    value: dict[str, Any] | None = None,
+) -> str:
+    from ccx.store.root import StorageIntegrityError
+    sidecar = value if value is not None else read_sidecar(engine, short)
+    if not sidecar or sidecar.get("mode") != "durable":
+        raise SessionUnavailable(f"{engine}:{short} has no shared variables; attach variables when creating it")
+    if not sidecar.get("stateId"):
+        raise StorageIntegrityError("Shared variable state is missing; inspect the worker")
+    return _valid_state_id(sidecar["stateId"])
+
+
+def _valid_name(name: Any) -> str:
+    if (not isinstance(name, str) or not name.isidentifier()
+            or keyword.iskeyword(name) or name.startswith("__")):
+        raise InvalidRequest(
+            f"variable names must be valid non-keyword Python identifiers, got {name!r}"
+        )
+    return name
+
+
+def _names(values: Mapping[str, Any]) -> list[str]:
+    return sorted(_valid_name(name) for name in values.keys())
+
+
+def _json_clone(value: Any, label: str) -> Any:
+    try:
+        return json.loads(strict_json(value))
+    except (TypeError, ValueError) as error:
+        raise InvalidRequest(f"{label} must be JSON-compatible: {error}") from error
+
+
+def parse_cli_vars(items: list[str] | None) -> dict[str, Any]:
+    result: dict[str, Any] = {}
+    for raw in items or []:
+        if "=" not in raw:
+            raise InvalidRequest(
+                f"--var expects NAME=JSON, got {raw!r}"
+            )
+        name, encoded = raw.split("=", 1)
+        name = _valid_name(name)
+        if name in result:
+            raise InvalidRequest(f"duplicate --var binding {name!r}")
+        try:
+            value = json.loads(encoded)
+        except ValueError as error:
+            raise InvalidRequest(
+                f"--var {name}=... must contain valid JSON: {error}"
+            ) from error
+        result[name] = _json_clone(value, f"variable {name!r}")
+    return result
+
+
+def _active_controller_runtime():
+    try:
+        module = importlib.import_module('ccx.controllers.controller')
+        return module._active_runtime.get()
+    except Exception:
+        return None
+
+
+@dataclass
+class PreparedAttachment:
+    mode: str
+    names: list[str]
+    mapping: MutableMapping[str, Any] | None = None
+    values: dict[str, Any] | None = None
+    state_id: str | None = None
+    owner: "LiveOwner | None" = None
+
+
+def _check_initial_durable_size(names: list[str], values: dict[str, Any]) -> None:
+    # A fresh state has a fixed-length state ID; reject before worker creation.
+    document = {"schemaVersion": VARIABLE_STATE_SCHEMA_VERSION,
+                "stateId": "0" * 32, "names": names, "variables": values}
+    if len(strict_json(document).encode("utf-8")) > STATE_FILE_LIMIT:
+        raise InvalidRequest("worker variables exceed the 4 MiB state limit")
+
+
+def prepare_attachment(values: Mapping[str, Any] | None) -> PreparedAttachment | None:
+    if values is None:
+        return None
+    if not isinstance(values, Mapping):
+        raise InvalidRequest("variables must be a mapping of Python names to values")
+    names = _names(values)
+    runtime = _active_controller_runtime()
+    if runtime is not None:
+        if not isinstance(values, MutableMapping):
+            raise InvalidRequest(
+                "controller-backed variables require a mutable mapping so worker rebinding can write back"
+            )
+        owner = runtime.live_vars_owner()
+        return PreparedAttachment(
+            mode="live",
+            names=names,
+            mapping=values,
+            owner=owner,
+        )
+    durable = _json_clone(dict(values), "variables")
+    _check_initial_durable_size(names, durable)
+    return PreparedAttachment(mode="durable", names=names, values=durable)
+
+
+def prepare_durable_attachment(values: Mapping[str, Any] | None) -> PreparedAttachment | None:
+    if values is None:
+        return None
+    if not isinstance(values, Mapping):
+        raise InvalidRequest("variables must be a mapping of Python names to values")
+    names = _names(values)
+    durable = _json_clone(dict(values), "variables")
+    _check_initial_durable_size(names, durable)
+    return PreparedAttachment(mode="durable", names=names, values=durable)
+
+
+def prepare_shared_durable_attachment(
+    engine: str,
+    short: str,
+) -> PreparedAttachment:
+    store, owner = _env_worker(engine, short)
+    info = store.variables_info(owner.worker)
+    if info["mode"] != "durable":
+        raise InvalidRequest(f"{owner.worker} has no durable variables to share")
+    if not info["available"]:
+        raise SessionUnavailable(f"{owner.worker} durable variables are unavailable")
+    return PreparedAttachment(
+        mode="durable",
+        names=info["names"],
+        state_id=info["state_id"],
+    )
+
+
+def _owner_gone(owner: Mapping[str, Any]) -> bool:
+    socket_path = owner.get("socket")
+    if not isinstance(socket_path, str) or not socket_path:
+        return False
+    try:
+        os.lstat(socket_path)
+    except FileNotFoundError:
+        return True
+    except OSError:
+        return False
+    client = socket.socket(socket.AF_UNIX, socket.SOCK_STREAM)
+    try:
+        client.settimeout(0.25)
+        client.connect(socket_path)
+        return False
+    except (FileNotFoundError, ConnectionRefusedError):
+        return True
+    except OSError:
+        return False
+    finally:
+        client.close()
+
+
+def _unlink_sidecar(store, engine: str, worker_id: str) -> None:
+    with store.lease():
+        with store._gen("variables", engine) as directory:
+            os.unlink(worker_id + ".json", dir_fd=directory)
+            os.fsync(directory)
+
+
+def _clear_dead_binding(engine: str, worker_id: str, token: str) -> None:
+    store, owner = _env_worker(engine, worker_id)
+    with _worker_lock(engine, owner.worker_id):
+        value = read_sidecar(engine, owner.worker_id)
+        if (value and value.get("mode") == "live"
+                and (value.get("owner") or {}).get("token") == token):
+            # Recheck after acquiring the writer lock.
+            if _owner_gone(value["owner"]):
+                _unlink_sidecar(store, engine, owner.worker_id)
+
+
+
+def publish_attachment(
+    attachment: PreparedAttachment | None,
+    engine: str,
+    short: str,
+) -> None:
+    if attachment is None:
+        return
+    if not attachment.names and attachment.state_id is None:
+        return
+
+    store, owner = _env_worker(engine, short)
+    with _worker_lock(engine, owner.worker_id):
+        existing = read_sidecar(engine, owner.worker_id)
+        if attachment.mode == "live":
+            assert attachment.owner is not None
+            assert attachment.mapping is not None
+            if existing and existing["mode"] == "durable":
+                raise InvalidRequest(
+                    f"{owner.worker} already has shared variables; choose one variable mode"
+                )
+            if existing:
+                old_owner = existing.get("owner")
+                if isinstance(old_owner, dict):
+                    if old_owner.get("token") != attachment.owner.token:
+                        if _owner_gone(old_owner):
+                            _unlink_sidecar(store, engine, owner.worker_id)
+                            existing = None
+                        else:
+                            raise InvalidRequest(
+                                f"{owner.worker} already has a reachable live variable owner"
+                            )
+            attachment.owner.register_worker(
+                engine, owner.worker_id, attachment.mapping, attachment.names
+            )
+            _atomic_json(sidecar_path(engine, owner.worker_id), {
+                "schemaVersion": VARIABLE_SCHEMA_VERSION,
+                "engine": engine,
+                "workerId": owner.worker_id,
+                "mode": "live",
+                "names": attachment.owner.worker_names(engine, owner.worker_id),
+                "owner": attachment.owner.public_identity(),
+            })
+            return
+
+        if attachment.mode != "durable":
+            raise InvalidRequest(f"unknown variable mode {attachment.mode!r}")
+        if existing and existing["mode"] == "live":
+            old_owner = existing.get("owner")
+            if isinstance(old_owner, dict) and _owner_gone(old_owner):
+                _unlink_sidecar(store, engine, owner.worker_id)
+                existing = None
+            else:
+                raise InvalidRequest(
+                    f"{owner.worker} already owns live variables; cannot mix modes"
+                )
+        if attachment.state_id is not None:
+            state_id = _valid_state_id(attachment.state_id)
+            with _state_lock(state_id):
+                read_state(state_id)
+            _atomic_json(sidecar_path(engine, owner.worker_id), {
+                "schemaVersion": VARIABLE_SCHEMA_VERSION,
+                "engine": engine,
+                "workerId": owner.worker_id,
+                "mode": "durable",
+                "stateId": state_id,
+            })
+            return
+
+        assert attachment.values is not None
+        if existing:
+            state_id = durable_state_id(engine, owner.worker_id, existing)
+        else:
+            state_id = _new_state_id()
+        with _state_lock(state_id, create=existing is None):
+            if existing:
+                document = read_state(state_id)
+                values = dict(document["variables"])
+                names = set(document["names"])
+            else:
+                values, names = {}, set()
+            values.update(attachment.values)
+            names.update(attachment.names)
+            _atomic_json(state_path(state_id), _new_state_document(
+                state_id, names=sorted(names), variables=values,
+            ))
+        _atomic_json(sidecar_path(engine, owner.worker_id), {
+            "schemaVersion": VARIABLE_SCHEMA_VERSION,
+            "engine": engine,
+            "workerId": owner.worker_id,
+            "mode": "durable",
+            "stateId": state_id,
+        })
+
+
+def variables_info(engine: str, short: str, name: str | None = None) -> dict[str, Any]:
+    """Use the already-accepted single full-UUID variable read authority."""
+    store, owner = _env_worker(engine, short)
+    return store.variables_info(owner.worker, name=name)
+
+
+def _runtime_state(owner) -> str | None:
+    try:
+        if owner.engine == "claude":
+            from ccx.rpc.rpc import _rpc
+            reply = _rpc(owner.worker, {"op": "status"}, timeout=0.5)
+            return reply.get("state") if reply.get("ok") is True else None
+        from ccx.codex.runtime import observe_worker
+        observation = observe_worker(owner.worker_id, timeout=0.5)
+        if not observation["daemon_served"]:
+            return None
+        return observation["reply"].get("state")
+    except Exception:  # an unreadable runtime never refuses; the send decides
+        return None
+
+
+def should_publish_send_attachment(store, owner, delivery_id: str | None,
+                                   *, caps: dict | None = None) -> bool:
+    """Preflight likely native refusal before touching attached variables.
+
+    Native admission may race after this check. A previously admitted delivery
+    token never reapplies its original variable values.
+    """
+    if delivery_id:
+        if owner.engine == "claude":
+            from ccx.claude.delivery import read_receipt
+        else:
+            from ccx.codex.delivery import read_receipt
+        prior = read_receipt(owner.worker, delivery_id)
+        if prior and prior.get("state") in (
+            "owned", "submitted", "accepted", "rejected", "delivered", "abandoned",
+        ):
+            return False
+
+    from ccx.errors import BudgetStopped, DeliveryRejected
+    record = store.record(owner.worker)
+    state = (record.get("lifecycle") or {}).get("state")
+    # A host or daemon that died mid-turn leaves a stale busy record that the
+    # send itself repairs, so only a live runtime's own state refuses here.
+    if state in ("busy", "waiting"):
+        state = _runtime_state(owner)
+    if state in ("busy", "waiting"):
+        raise DeliveryRejected(
+            f"worker is {state}; wait until idle before sending variables",
+            short=owner.short, state=state,
+            refusal="busy" if state == "busy" else "state_unavailable",
+            delivery_status="rejected",
+        )
+    if caps:
+        if owner.engine == "claude":
+            from ccx.claude.usage import cap_tripped, read_usage_cached
+            from ccx.constants import USAGE_STALE_LIMIT_S
+            data = read_usage_cached(stale_limit=USAGE_STALE_LIMIT_S)
+            trip = cap_tripped(data, caps)
+        else:
+            from ccx.codex.types import codex_cap_tripped
+            # A cap preflight may observe the existing daemon, never launch
+            # a second provider solely for a speculative send.
+            from ccx.codex.runtime import observation_daemon, daemon_call
+            deadline = time.monotonic() + 0.5
+            try:
+                data = (daemon_call({"op": "account-read", "tokens": False}, timeout=0.5)
+                        if observation_daemon(deadline) else None)
+                if isinstance(data, dict) and "data" in data:
+                    data = data["data"]
+            except (OSError, ValueError, SessionUnavailable):
+                data = None
+            trip = codex_cap_tripped(data, caps)
+        if trip:
+            key, used, limit = trip
+            raise BudgetStopped(
+                f"budget stop: usage cap {key} already exceeded ({used:g}% >= {limit:g}%)",
+                short=owner.short, reason=f"usage cap {key} exceeded", phase="preflight",
+            )
+    return True
+
+
+def with_variable_hint(text: str, names: list[str]) -> str:
+    if not names:
+        return text
+    joined = ", ".join(names)
+    return (
+        text.rstrip()
+        + "\n\n[CCX variables available in direct python/python3: "
+        + joined
+        + ". Rebinding or deleting these names writes back to shared state. "
+        + "command/env/explicit Python paths are ordinary native Python.]"
+    )
+
+
+def _managed_shell_env_file(value: str | None) -> bool:
+    if not value:
+        return False
+    candidate = Path(value).expanduser()
+    # Identify a CCX bridge from this, a retired, or a foreign generation.
+    # No ACTIVE-pointer lookup or foreign shell-file read is needed.
+    if _generated_shell_file_shape(candidate):
+        return True
+    try:
+        candidate_resolved = candidate.resolve()
+        if _generated_shell_file_shape(candidate_resolved):
+            return True
+        from ccx.store.root import require_ccx_active
+        managed_root = (
+            require_ccx_active(constants.CCX_ROOT).pointer.generation_root
+            / "variables" / "env"
+        ).resolve()
+        return candidate_resolved.is_relative_to(managed_root)
+    except (OSError, RuntimeError):
+        return False
+
+
+def _generated_shell_file_shape(path: Path) -> bool:
+    if (path.suffix != ".sh" or path.parent.name not in {"claude", "codex"}
+            or path.parent.parent.name != "env"
+            or path.parent.parent.parent.name != "variables"):
+        return False
+    from ccx.store.identity import canonical_worker_id, IdentityError
+    try:
+        canonical_worker_id(path.stem)
+    except IdentityError:
+        return False
+    return True
+
+
+def _managed_zdotdir(value: str | None) -> bool:
+    """Recognize this or a foreign generation's Codex zsh trampoline directory.
+
+    A user directory with the same shape is deliberately treated as managed:
+    never read or source a foreign worker's .zshenv.
+    """
+    if not value:
+        return False
+
+    def matches(path: Path) -> bool:
+        return path.parts[-5:-1] == ("variables", "env", "codex", "zsh")
+
+    candidate = Path(value).expanduser()
+    if matches(candidate):
+        return True
+    try:
+        return matches(candidate.resolve())
+    except (OSError, RuntimeError):
+        return False
+
+
+_BRIDGE_ORIGINALS = {
+    "BASH_ENV": "CCX_BRIDGE_ORIGINAL_BASH_ENV",
+    "ZDOTDIR": "CCX_BRIDGE_ORIGINAL_ZDOTDIR",
+    "CLAUDE_ENV_FILE": "CCX_BRIDGE_ORIGINAL_CLAUDE_ENV_FILE",
+}
+
+
+def sanitize_inherited_shell_bridge_env(
+    env: dict[str, str] | None = None,
+) -> dict[str, str]:
+    """Restore native user startup paths, never a parent's CCX Python bridge.
+
+    Each original worker-owned shell shim exports one-use snapshots of the
+    user's startup paths. A nested worker consumes those snapshots only if
+    the inherited value is itself a managed CCX path. Explicit user overrides
+    and unset-versus-empty original values are preserved.
+    """
+    target = os.environ if env is None else env
+    for key, marker in _BRIDGE_ORIGINALS.items():
+        snapshot_text = target.pop(marker, None)
+        value = target.get(key)
+        managed = (_managed_zdotdir(value) if key == "ZDOTDIR"
+                   else _managed_shell_env_file(value))
+        if not managed:
+            continue
+        target.pop(key, None)
+        try:
+            original = json.loads(snapshot_text) if isinstance(snapshot_text, str) else None
+        except ValueError:
+            original = None
+        if (isinstance(original, dict) and original.get("set") is True
+                and isinstance(original.get("value"), str)):
+            restored = original["value"]
+            # Refuse a pointer to any other worker, even when forged as a
+            # user-original path. Do not source foreign CCX-owned scripts.
+            if not (_managed_zdotdir(restored) if key == "ZDOTDIR"
+                    else _managed_shell_env_file(restored)):
+                target[key] = restored
+    return target
+
+
+def _shell_original_exports(env: Mapping[str, str]) -> list[str]:
+    """Emit exact user startup-path snapshots for a managed tool shell."""
+    return [
+        f"export {marker}={shlex.quote(json.dumps({'set': key in env, 'value': env.get(key)}))}"
+        for key, marker in _BRIDGE_ORIGINALS.items()
+    ]
+
+
+def _external_shell_env_file(value: str | None) -> str | None:
+    if not value:
+        return None
+    if _managed_shell_env_file(value):
+        return None
+    # A bridge from a different CCX_ROOT/generation must not override this
+    # worker's python/python3 functions. Do not source or read foreign scripts.
+    try:
+        path = Path(value).expanduser().resolve()
+        if _generated_shell_file_shape(path):
+            return None
+    except (OSError, RuntimeError):
+        return None
+    return str(Path(value).expanduser())
+
+
+def _python_bridge_functions(engine: str, short: str, *, root: Path, zsh: bool) -> list[str]:
+    """Render the same Python bridge command for both shell startup files."""
+    python = shlex.quote(sys.executable)
+    executable = shlex.quote(current_executable_path())
+    engine_q = shlex.quote(engine)
+    short_q = shlex.quote(short)
+    root_q = shlex.quote(str(root))
+    lines: list[str] = []
+    for name in ("python3", "python"):
+        name_q = shlex.quote(name)
+        command = (
+            f"{python} {executable} _live-vars-exec "
+            f"--root {root_q} --engine {engine_q} --short {short_q} "
+            f"--invoked-as {name_q} -- \"$@\""
+        )
+        # macOS may supply python3 without a native ``python`` executable.
+        # Do not publish a python function that can only fail with rc 127;
+        # do not silently remap python to python3, either.
+        native_python = (("whence -p python" if zsh else "type -P python")
+                         + " >/dev/null 2>&1 && ") if name == "python" else ""
+        lines.extend([
+            f"if {native_python}! typeset -f {name} >/dev/null 2>&1 && ! alias {name} >/dev/null 2>&1; then",
+            # Quoted zsh function names cannot expand a user-defined alias
+            # while the startup script is parsed.
+            (f"  function '{name}' {{ {command}; }}" if zsh
+             else f"  {name}() {{ {command}; }}"),
+            "fi",
+        ])
+    return lines
+
+
+def _publish_shell_script(
+    store, parts: tuple[str, ...], name: str, data: bytes, *,
+    exact_dir_mode: bool = False,
+) -> None:
+    """Reuse matching private script bytes; publish changed bytes atomically.
+
+    Every comparison occurs against a nofollow, user-owned regular file inside
+    the generation's pinned directory. A warm call needs no replace or fsync.
+    """
+    from ccx.store.root import StorageIntegrityError
+
+    with store.lease():
+        with store._gen(*parts, create=True) as directory:
+            if exact_dir_mode:
+                directory_stat = os.fstat(directory)
+                if (not stat.S_ISDIR(directory_stat.st_mode)
+                        or directory_stat.st_uid != os.getuid()
+                        or stat.S_IMODE(directory_stat.st_mode) != 0o700):
+                    raise StorageIntegrityError(
+                        "Codex zsh directory must be owner-owned mode 0700"
+                    )
+            try:
+                old = os.stat(name, dir_fd=directory, follow_symlinks=False)
+            except FileNotFoundError:
+                pass
+            else:
+                if (not stat.S_ISREG(old.st_mode) or old.st_uid != os.getuid()
+                        or stat.S_IMODE(old.st_mode) != 0o600 or old.st_nlink != 1):
+                    raise StorageIntegrityError(
+                        "managed shell script must be a private regular file"
+                    )
+                source_fd = os.open(
+                    name, os.O_RDONLY | os.O_NOFOLLOW | os.O_CLOEXEC |
+                    os.O_NONBLOCK, dir_fd=directory,
+                )
+                with os.fdopen(source_fd, "rb") as source:
+                    opened = os.fstat(source.fileno())
+                    if (not stat.S_ISREG(opened.st_mode)
+                            or opened.st_uid != os.getuid()
+                            or stat.S_IMODE(opened.st_mode) != 0o600
+                            or opened.st_nlink != 1):
+                        raise StorageIntegrityError(
+                            "managed shell script changed ownership or mode"
+                        )
+                    existing = source.read(len(data) + 1)
+                    finished = os.fstat(source.fileno())
+                now = os.stat(name, dir_fd=directory, follow_symlinks=False)
+                if (old.st_dev == opened.st_dev == finished.st_dev == now.st_dev
+                        and old.st_ino == opened.st_ino == finished.st_ino == now.st_ino
+                        and old.st_mtime_ns == finished.st_mtime_ns == now.st_mtime_ns
+                        and old.st_size == finished.st_size == now.st_size
+                        and existing == data):
+                    return
+
+            temp_name = "." + name + "." + uuid.uuid4().hex + ".tmp"
+            script_fd = os.open(
+                temp_name, os.O_WRONLY | os.O_CREAT | os.O_EXCL |
+                os.O_NOFOLLOW | os.O_CLOEXEC, 0o600, dir_fd=directory,
+            )
+            try:
+                with os.fdopen(script_fd, "wb") as sink:
+                    sink.write(data)
+                    sink.flush()
+                    os.fsync(sink.fileno())
+                os.replace(
+                    temp_name, name, src_dir_fd=directory, dst_dir_fd=directory,
+                )
+                os.fsync(directory)
+            except BaseException:
+                try:
+                    os.unlink(temp_name, dir_fd=directory)
+                except FileNotFoundError:
+                    pass
+                raise
+
+
+def ensure_shell_env(
+    engine: str,
+    short: str,
+    *,
+    inherited_env_file: str | None = None,
+) -> Path:
+    store, owner = _env_worker(engine, short)
+    return _ensure_shell_env_owned(
+        store, engine, owner.worker_id, inherited_env_file=inherited_env_file,
+    )
+
+
+def _ensure_shell_env_owned(
+    store, engine: str, worker_id: str, *,
+    inherited_env_file: str | None,
+    native_env: Mapping[str, str] | None = None,
+) -> Path:
+    """Use an already-attested full UUID; the generation lease revalidates it."""
+    path = (store.pointer.generation_root / "variables" / "env" /
+            engine / f"{worker_id}.sh")
+    inherited_env_file = _external_shell_env_file(inherited_env_file)
+
+    lines = ["# generated by ccx; sourced for managed worker shell commands"]
+    if inherited_env_file:
+        inherited = Path(inherited_env_file).expanduser()
+        try:
+            same = inherited.resolve() == path.resolve()
+        except OSError:
+            same = False
+        if not same:
+            lines.extend(
+                [
+                    f"if [ -f {shlex.quote(str(inherited))} ]; then",
+                    f"  . {shlex.quote(str(inherited))}",
+                    "fi",
+                ]
+            )
+    lines.extend(_shell_original_exports(os.environ if native_env is None else native_env))
+    lines.extend(_python_bridge_functions(engine, worker_id, root=store.root, zsh=False))
+    text = "\n".join(lines) + "\n"
+    # Only the sealed ACTIVE generation's uid-private owner directory may
+    # receive this shell script. Never resurrect CCX_ROOT/variables/<short>.
+    from ccx.store.root import _FILE_LIMIT
+    data = text.encode("utf-8")
+    if len(data) > _FILE_LIMIT:
+        raise SessionUnavailable("managed full-UUID shell environment script too large")
+    _publish_shell_script(store, ("variables", "env", engine), path.name, data)
+    return path
+
+
+def codex_zdotdir_path(worker_id: str) -> Path:
+    store, owner = _env_worker("codex", worker_id)
+    return (store.pointer.generation_root / "variables" / "env" / "codex" /
+            "zsh" / owner.worker_id)
+
+
+def ensure_codex_zshenv(worker_id: str, *, original_zdotdir: str | None) -> Path:
+    """Install the restore-first .zshenv in the sealed, full-UUID generation."""
+    store, owner = _env_worker("codex", worker_id)
+    return _ensure_codex_zshenv_owned(
+        store, owner.worker_id, original_zdotdir=original_zdotdir,
+    )
+
+
+def _zsh_user_startup_lines(stage: str, managed: Path, *,
+                            original_zdotdir: str | None = None) -> list[str]:
+    """Replay native user startup stages while keeping CCX ZDOTDIR effective.
+
+    A nested zsh is a new shell of the *same worker*, so its own .zshenv must
+    see the managed directory. The original user's directory is remembered
+    only for startup-file sourcing; the previous user's .zshenv (or any later
+    startup stage) may update its subsequent native lookup directory.
+    """
+    lines = [f"# generated by ccx; native user {stage} once, CCX bridge retained"]
+    if stage == ".zshenv":
+        lines.extend([
+            f"typeset -g __ccx_managed_zdotdir={shlex.quote(str(managed))}",
+            f"typeset -g __ccx_user_zdotdir_set={'0' if original_zdotdir is None else '1'}",
+            f"typeset -g __ccx_user_zdotdir_value={shlex.quote(original_zdotdir or '')}",
+        ])
+    lines.extend([
+        'if [[ $__ccx_user_zdotdir_set == 0 ]]; then',
+        '  unset ZDOTDIR',
+        f'  __ccx_user_startup="$HOME/{stage}"',
+        'elif [[ -z $__ccx_user_zdotdir_value ]]; then',
+        "  export ZDOTDIR=''",
+        '  __ccx_user_startup=""',
+        'else',
+        '  export ZDOTDIR="$__ccx_user_zdotdir_value"',
+        f'  __ccx_user_startup="${{__ccx_user_zdotdir_value}}/{stage}"',
+        'fi',
+        'if [[ -n $__ccx_user_startup && -f $__ccx_user_startup ]]; then',
+        '  case "${__ccx_user_startup:A}" in',
+        # A retired/foreign CCX-generated shim is not a user startup file.
+        f'    {shlex.quote(str(managed))}/.*|*/variables/env/codex/zsh/*/.*) ;;',
+        '    *) builtin source "$__ccx_user_startup" ;;',
+        '  esac',
+        'fi',
+        'if [[ -z ${ZDOTDIR+x} ]]; then',
+        '  __ccx_user_zdotdir_set=0',
+        '  __ccx_user_zdotdir_value=""',
+        'else',
+        '  __ccx_user_zdotdir_set=1',
+        '  __ccx_user_zdotdir_value="$ZDOTDIR"',
+        'fi',
+        'export ZDOTDIR="$__ccx_managed_zdotdir"',
+        'unset __ccx_user_startup',
+    ])
+    return lines
+
+
+def _ensure_codex_zshenv_owned(
+    store, worker_id: str, *, original_zdotdir: str | None,
+    native_env: Mapping[str, str] | None = None,
+) -> Path:
+    from ccx.store.root import _FILE_LIMIT
+
+    directory = (store.pointer.generation_root / "variables" / "env" /
+                 "codex" / "zsh" / worker_id)
+    if _managed_zdotdir(original_zdotdir):
+        original_zdotdir = None
+    for stage in (".zshenv", ".zprofile", ".zshrc", ".zlogin", ".zlogout"):
+        lines = _zsh_user_startup_lines(
+            stage, directory, original_zdotdir=original_zdotdir,
+        )
+        if stage == ".zshenv":
+            lines.extend(_shell_original_exports(
+                os.environ if native_env is None else native_env,
+            ))
+            lines.extend(_python_bridge_functions("codex", worker_id, root=store.root, zsh=True))
+        data = ("\n".join(lines) + "\n").encode("utf-8")
+        if len(data) > _FILE_LIMIT:
+            raise SessionUnavailable("managed full-UUID zsh startup script too large")
+        _publish_shell_script(
+            store, ("variables", "env", "codex", "zsh", worker_id),
+            stage, data, exact_dir_mode=True,
+        )
+    return directory
+
+
+def configure_claude_worker_env(
+    env: dict[str, str], short: str, *, host_pid: int,
+) -> dict[str, str]:
+    sanitize_inherited_shell_bridge_env(env)
+    inherited = env.get("CLAUDE_ENV_FILE")
+    store, owner = _env_worker("claude", short)
+    path = _ensure_shell_env_owned(
+        store, "claude", owner.worker_id, inherited_env_file=inherited,
+        native_env=env,
+    )
+    env["CLAUDE_ENV_FILE"] = str(path)
+    mark_claude_bridge(short, host_pid)
+    return env
+
+
+def codex_thread_config(worker_id: str) -> dict[str, Any]:
+    """Attested full-UUID Bash/zsh bridges for one original native Codex thread.
+
+    The daemon first validates the native claim/source/dispatch; this function
+    only prepares the shell config and cannot start/adopt a native thread.
+    Preserve Codex's user/default inheritance policy.
+    """
+    store, owner = _env_worker("codex", worker_id)
+    # The per-thread policy may be constructed by a nested CCX caller as
+    # well as by the shared (already scrubbed) codexd. Never capture that
+    # caller's bridge as this worker's genuine user-shell baseline.
+    native_env = sanitize_inherited_shell_bridge_env(dict(os.environ))
+    bash_env = _ensure_shell_env_owned(
+        store, "codex", owner.worker_id,
+        inherited_env_file=native_env.get("BASH_ENV"), native_env=native_env,
+    )
+    zdotdir = _ensure_codex_zshenv_owned(
+        store, owner.worker_id, original_zdotdir=native_env.get("ZDOTDIR"),
+        native_env=native_env,
+    )
+    return {"shell_environment_policy": {"set": {
+        "CCX_SELF_WORKER_ID": owner.worker_id,
+        "CCX_SELF_ENGINE": "codex",
+        "CCX_ROOT": str(constants.CCX_ROOT),
+        "BASH_ENV": str(bash_env),
+        "ZDOTDIR": str(zdotdir),
+    }}}
+
+
+def mark_claude_bridge(short: str, host_pid: int) -> None:
+    """Store exact full owner + current host pid as an atomic owned sidecar."""
+    from ccx.store.root import _write_json
+    store, owner = _env_worker("claude", short)
+    with store.lease():
+        with store._gen("variables", "env", "claude", create=True) as fd:
+            _write_json(fd, owner.worker_id + ".bridge.json", {
+                "workerId": owner.worker_id,
+                "engine": "claude",
+                "generation": store.generation,
+                "hostPid": int(host_pid),
+            })
+
+
+def claude_bridge_ready(short: str, host_pid: int | None) -> bool:
+    if not host_pid:
+        return False
+    from ccx.store.root import _read_json
+    store, owner = _env_worker("claude", short)
+    with store.lease():
+        try:
+            with store._gen("variables", "env", "claude") as fd:
+                value = _read_json(fd, owner.worker_id + ".bridge.json")
+        except FileNotFoundError:
+            return False
+    return (
+        isinstance(value, dict)
+        and value.get("workerId") == owner.worker_id
+        and value.get("engine") == "claude"
+        and value.get("generation") == store.generation
+        and value.get("hostPid") == int(host_pid)
+    )
+
+
+def require_send_variable_permission(owner, record: dict) -> None:
+    """Refuse a later-SEND variable attachment the original worker cannot take."""
+    create_args = record.get("createArgs") or {}
+    if owner.engine == "claude":
+        if not bool(create_args.get("skip_permissions", True)):
+            raise CapabilityUnsupported(
+                'Shared variables require full permissions (permissions="full")',
+                short=owner.short, engine=owner.engine,
+                capability="shared_python_variables",
+            )
+        lifecycle = record.get("lifecycle") or {}
+        state = lifecycle.get("state")
+        host_pid = lifecycle.get("hostPid")
+        if (state not in {"dead", "crashed", "idle-killed", "budget-stopped"}
+                and host_pid
+                and not claude_bridge_ready(owner.worker_id, host_pid)):
+            raise SessionUnavailable(
+                "this live Claude host predates the CCX shared-Python bridge; "
+                "kill/revive it or create a new worker before attaching variables",
+                short=owner.short,
+            )
+    elif not (create_args.get("sandbox") == "danger-full-access"
+              and create_args.get("approvalPolicy") == "never"):
+        raise CapabilityUnsupported(
+            'Shared variables require full permissions (permissions="full")',
+            short=owner.short, engine=owner.engine,
+            capability="shared_python_variables",
+        )
+
+
+class _ThreadStreamProxy:
+    def __init__(self, original):
+        self.original = original
+        self._routes: dict[int, Any] = {}
+        self._lock = threading.Lock()
+
+    def _target(self):
+        with self._lock:
+            return self._routes.get(threading.get_ident(), self.original)
+
+    @contextmanager
+    def route(self, target):
+        ident = threading.get_ident()
+        with self._lock:
+            self._routes[ident] = target
+        try:
+            yield
+        finally:
+            with self._lock:
+                self._routes.pop(ident, None)
+
+    def write(self, value):
+        return self._target().write(value)
+
+    def read(self, *args):
+        return self._target().read(*args)
+
+    def readline(self, *args):
+        return self._target().readline(*args)
+
+    def flush(self):
+        return self._target().flush()
+
+    def isatty(self):
+        return self._target().isatty()
+
+    def fileno(self):
+        return self._target().fileno()
+
+    @property
+    def buffer(self):
+        return self._target().buffer
+
+    @property
+    def encoding(self):
+        return getattr(self._target(), "encoding", None)
+
+    @property
+    def errors(self):
+        return getattr(self._target(), "errors", None)
+
+    def __iter__(self):
+        return iter(self._target())
+
+    def __getattr__(self, name):
+        return getattr(self._target(), name)
+
+
+@dataclass
+class _Binding:
+    mapping: MutableMapping[str, Any]
+    key: str
+
+
+class LiveOwner:
+    def __init__(self):
+        self.pid = os.getpid()
+        self.token = uuid.uuid4().hex
+        basename = f"{self.pid}-{self.token}.sock"
+        owner_dir, choice = _owner_runtime(basename)
+        candidate = owner_dir / basename
+        _require_unix_socket_budget(candidate)  # before any parent writes
+        self.socket_path = owners_dir(owner_dir, runtime=choice) / basename
+        self._registry_lock = threading.RLock()
+        self._exec_lock = threading.Lock()
+        self._workers: dict[tuple[str, str], dict[str, _Binding]] = {}
+        self._queue: queue.Queue[Any] = queue.Queue()
+        self._closed = threading.Event()
+        self._current: dict[str, Any] | None = None
+
+        self._old_stdin = sys.stdin
+        self._old_stdout = sys.stdout
+        self._old_stderr = sys.stderr
+        self._stdin_proxy = _ThreadStreamProxy(sys.stdin)
+        self._stdout_proxy = _ThreadStreamProxy(sys.stdout)
+        self._stderr_proxy = _ThreadStreamProxy(sys.stderr)
+        sys.stdin = self._stdin_proxy
+        sys.stdout = self._stdout_proxy
+        sys.stderr = self._stderr_proxy
+
+        self._server = socket.socket(socket.AF_UNIX, socket.SOCK_STREAM)
+        try:
+            try:
+                self.socket_path.unlink()
+            except FileNotFoundError:
+                pass
+            bind_private_unix_socket(self._server, self.socket_path)
+            self._server.listen()
+            # Closing a listening fd from another thread does not reliably
+            # interrupt a blocked accept() on Linux. Poll for shutdown.
+            self._server.settimeout(0.1)
+        except BaseException:
+            try:
+                self._server.close()
+            except OSError:
+                pass
+            if sys.stdin is self._stdin_proxy:
+                sys.stdin = self._old_stdin
+            if sys.stdout is self._stdout_proxy:
+                sys.stdout = self._old_stdout
+            if sys.stderr is self._stderr_proxy:
+                sys.stderr = self._old_stderr
+            raise
+
+        self._executor_thread = threading.Thread(
+            target=self._executor_loop,
+            name=f"ccx-live-vars-{self.pid}",
+            daemon=True,
+        )
+        self._accept_thread = threading.Thread(
+            target=self._accept_loop,
+            name=f"ccx-live-vars-accept-{self.pid}",
+            daemon=True,
+        )
+        self._executor_thread.start()
+        self._accept_thread.start()
+
+    def public_identity(self) -> dict[str, Any]:
+        return {
+            "pid": self.pid,
+            "token": self.token,
+            "socket": str(self.socket_path),
+            "python": {
+                "major": sys.version_info.major,
+                "minor": sys.version_info.minor,
+                "cacheTag": sys.implementation.cache_tag,
+                "platform": sysconfig.get_platform(),
+            },
+        }
+
+    def register_worker(
+        self,
+        engine: str,
+        short: str,
+        mapping: MutableMapping[str, Any],
+        names: list[str],
+    ) -> None:
+        self._exec_lock.acquire()
+        try:
+            with self._registry_lock:
+                bindings = self._workers.setdefault((engine, short), {})
+                for name in names:
+                    bindings[name] = _Binding(mapping=mapping, key=name)
+        finally:
+            self._exec_lock.release()
+
+    def worker_names(self, engine: str, short: str) -> list[str]:
+        with self._registry_lock:
+            return sorted(self._workers.get((engine, short), {}))
+
+    def _accept_loop(self) -> None:
+        while not self._closed.is_set():
+            try:
+                conn = accept_same_user(self._server)
+            except socket.timeout:
+                continue
+            except OSError:
+                return
+            if conn is None:
+                continue
+            threading.Thread(
+                target=self._handle_connection,
+                args=(conn,),
+                daemon=True,
+                name="ccx-live-vars-client",
+            ).start()
+
+    @staticmethod
+    def _read_line(conn: socket.socket) -> bytes:
+        data = bytearray()
+        while b"\n" not in data:
+            chunk = conn.recv(65536)
+            if not chunk:
+                break
+            data.extend(chunk)
+        return bytes(data).split(b"\n", 1)[0]
+
+    @staticmethod
+    def _send(conn: socket.socket, value: dict[str, Any]) -> None:
+        conn.sendall((_WIRE_DUMPS(value, separators=(",", ":")) + "\n").encode("utf-8"))
+
+    def _handle_connection(self, conn: socket.socket) -> None:
+        try:
+            if self._closed.is_set():
+                self._send(conn, {"ok": False, "errorKind": "state_unavailable"})
+                conn.close()
+                return
+            raw = self._read_line(conn)
+            request = _WIRE_LOADS(raw) if raw else {}
+            if request.get("token") != self.token:
+                self._send(conn, {"ok": False, "errorKind": "state_unavailable"})
+                conn.close()
+                return
+            if request.get("op") == "ping":
+                self._send(conn, {"ok": True})
+                conn.close()
+                return
+            if request.get("op") != "exec":
+                self._send(conn, {"ok": False, "errorKind": "invalid_request"})
+                conn.close()
+                return
+            self._exec_lock.acquire()
+            if self._closed.is_set():
+                self._exec_lock.release()
+                self._send(conn, {"ok": False, "errorKind": "state_unavailable"})
+                conn.close()
+                return
+            self._current = {
+                "worker": f"{request.get('engine')}:{request.get('short')}",
+                "startedAt": time.time(),
+            }
+            self._queue.put((conn, request))
+        except Exception as error:
+            try:
+                self._send(
+                    conn,
+                    {
+                        "ok": False,
+                        "errorKind": "protocol_error",
+                        "message": f"{type(error).__name__}: {error}",
+                    },
+                )
+            except OSError:
+                pass
+            conn.close()
+
+    def _restore_streams(self) -> None:
+        if sys.stdin is self._stdin_proxy:
+            sys.stdin = self._old_stdin
+        if sys.stdout is self._stdout_proxy:
+            sys.stdout = self._old_stdout
+        if sys.stderr is self._stderr_proxy:
+            sys.stderr = self._old_stderr
+
+    def _executor_loop(self) -> None:
+        while True:
+            item = self._queue.get()
+            if item is None:
+                self._restore_streams()
+                return
+            conn, request = item
+            try:
+                response = self._execute(request)
+                self._send(conn, response)
+            except BaseException as error:
+                try:
+                    self._send(
+                        conn,
+                        {
+                            "ok": False,
+                            "errorKind": "owner_error",
+                            "message": f"{type(error).__name__}: {error}",
+                        },
+                    )
+                except BaseException:
+                    pass
+            finally:
+                try:
+                    conn.close()
+                except OSError:
+                    pass
+                self._current = None
+                self._exec_lock.release()
+
+    def _execute(self, request: dict[str, Any]) -> dict[str, Any]:
+        engine = str(request.get("engine") or "")
+        short = str(request.get("short") or "")
+        with self._registry_lock:
+            bindings = dict(self._workers.get((engine, short), {}))
+        if not bindings:
+            return {"ok": False, "errorKind": "state_unavailable"}
+
+        seed: dict[str, Any] = {}
+        for name, binding in bindings.items():
+            if name in binding.mapping:
+                seed[name] = binding.mapping[name]
+
+        native = request.get("nativePython")
+        if not isinstance(native, dict):
+            native = {}
+        stdin_bytes = base64.b64decode(request.get("stdinBase64") or "")
+        stdin_buffer = io.BytesIO(stdin_bytes)
+        stdout_buffer = io.BytesIO()
+        stderr_buffer = io.BytesIO()
+        stdin_text = io.TextIOWrapper(
+            stdin_buffer,
+            encoding=getattr(self._old_stdin, "encoding", None) or "utf-8",
+            errors=getattr(self._old_stdin, "errors", None) or "strict",
+        )
+        stdout_text = io.TextIOWrapper(
+            stdout_buffer,
+            encoding=getattr(self._old_stdout, "encoding", None) or "utf-8",
+            errors=getattr(self._old_stdout, "errors", None) or "replace",
+            write_through=True,
+        )
+        stderr_text = io.TextIOWrapper(
+            stderr_buffer,
+            encoding=getattr(self._old_stderr, "encoding", None) or "utf-8",
+            errors=getattr(self._old_stderr, "errors", None) or "replace",
+            write_through=True,
+        )
+
+        old_cwd = os.getcwd()
+        old_argv = sys.argv[:]
+        old_path = sys.path[:]
+        try:
+            cwd = str(request.get("cwd") or old_cwd)
+            os.chdir(cwd)
+            native_path = native.get("sysPath")
+            if isinstance(native_path, list) and all(
+                isinstance(value, str) for value in native_path
+            ):
+                sys.path[:] = native_path
+            args = request.get("args")
+            if not isinstance(args, list) or not all(isinstance(v, str) for v in args):
+                return {"ok": False, "errorKind": "invalid_request"}
+            with (
+                self._stdin_proxy.route(stdin_text),
+                self._stdout_proxy.route(stdout_text),
+                self._stderr_proxy.route(stderr_text),
+            ):
+                exit_code, final_ns = execute_target(args, seed)
+                stdout_text.flush()
+                stderr_text.flush()
+        finally:
+            os.chdir(old_cwd)
+            sys.argv = old_argv
+            sys.path[:] = old_path
+
+        deleted: list[str] = []
+        with self._registry_lock:
+            current = self._workers.get((engine, short), {})
+            for name, binding in bindings.items():
+                if name in final_ns:
+                    binding.mapping[name] = final_ns[name]
+                else:
+                    binding.mapping.pop(name, None)
+                    if current.get(name) is binding:
+                        current.pop(name, None)
+                    deleted.append(name)
+            names = sorted(current)
+
+        self._publish_live_names(engine, short, names)
+        return {
+            "ok": True,
+            "exitCode": int(exit_code),
+            "stdoutBase64": base64.b64encode(stdout_buffer.getvalue()).decode("ascii"),
+            "stderrBase64": base64.b64encode(stderr_buffer.getvalue()).decode("ascii"),
+            "deleted": deleted,
+        }
+
+    def _publish_live_names(self, engine: str, short: str, names: list[str]) -> None:
+        with _worker_lock(engine, short):
+            value = read_sidecar(engine, short)
+            if (
+                not value
+                or value.get("mode") != "live"
+                or (value.get("owner") or {}).get("token") != self.token
+            ):
+                return
+            value["names"] = names
+            _atomic_json(sidecar_path(engine, short), value)
+
+    def close(self) -> None:
+        if self._closed.is_set():
+            return
+        self._closed.set()
+        try:
+            self._server.close()
+        except OSError:
+            pass
+        try:
+            self.socket_path.unlink()
+        except FileNotFoundError:
+            pass
+        self._queue.put(None)
+        # Wait briefly for both daemon loops to finish. In particular, the
+        # executor restores sys.stdin/stdout/stderr; without this join a new
+        # LiveOwner can wrap a proxy from a closed owner and strand it forever.
+        for worker in (self._accept_thread, self._executor_thread):
+            if threading.current_thread() is not worker:
+                worker.join(timeout=1.0)
+        # Never remove a sidecar that a newer controller or durable send owns.
+        for engine, worker_id in list(self._workers):
+            try:
+                store, owned = _env_worker(engine, worker_id)
+                with _worker_lock(engine, owned.worker_id):
+                    value = read_sidecar(engine, owned.worker_id)
+                    if (value and value.get("mode") == "live"
+                            and (value.get("owner") or {}).get("token") == self.token):
+                        _unlink_sidecar(store, engine, owned.worker_id)
+            except (OSError, SessionUnavailable):
+                pass
+
+
+def _native_python_metadata(executable: str) -> dict[str, Any] | None:
+    code = (
+        "import json,sys,sysconfig;"
+        "print(json.dumps({'version':list(sys.version_info[:3]),"
+        "'cacheTag':sys.implementation.cache_tag,"
+        "'platform':sysconfig.get_platform(),"
+        "'executable':sys.executable,'prefix':sys.prefix,"
+        "'execPrefix':sys.exec_prefix,'sysPath':sys.path}))"
+    )
+    try:
+        output = subprocess.check_output(
+            [executable, "-c", code],
+            text=True,
+            stderr=subprocess.DEVNULL,
+            timeout=5,
+        )
+        value = json.loads(output)
+    except (OSError, subprocess.SubprocessError, ValueError):
+        return None
+    return value if isinstance(value, dict) else None
+
+
+def _compatible(owner: Mapping[str, Any], native: Mapping[str, Any]) -> bool:
+    version = native.get("version")
+    return bool(
+        isinstance(version, list)
+        and len(version) >= 2
+        and version[0] == owner.get("major")
+        and version[1] == owner.get("minor")
+        and native.get("platform") == owner.get("platform")
+        and native.get("cacheTag") == owner.get("cacheTag")
+    )
+
+
+def _durable_runner_argv(
+    executable: str,
+    *,
+    state: Path,
+    lock: Path,
+    cwd: str,
+    args: list[str],
+) -> list[str]:
+    runner = Path(__file__).with_name("live_vars_runner.py")
+    bootstrap = (
+        "import runpy; "
+        f"runpy.run_path({str(runner)!r}, run_name='__main__')"
+    )
+    return [
+        executable,
+        "-c",
+        bootstrap,
+        "--state",
+        str(state),
+        "--lock",
+        str(lock),
+        "--cwd",
+        cwd,
+        "--real-python",
+        executable,
+        "--",
+        *args,
+    ]
+
+
+def _exec_native(executable: str, args: list[str], *, env: dict[str, str] | None = None) -> int:
+    os.execvpe(executable, [executable, *args], os.environ.copy() if env is None else env)
+    return 127
+
+
+def worker_exec(engine: str, short: str, invoked_as: str, args: list[str], *,
+                native_env: dict[str, str] | None = None) -> int:
+    from ccx.store.root import GenerationError
+    from ccx.store.identity import IdentityError
+
+    real_python = shutil.which(invoked_as)
+    if not real_python:
+        print(f"ccx: cannot resolve native {invoked_as}", file=sys.stderr)
+        return 127
+
+    # Stage0's exact ACTIVE owner/read gate stays before any native execution.
+    try:
+        store, owned = _env_worker(engine, short)
+        binding = read_sidecar(engine, owned.worker_id)
+    except (IdentityError, GenerationError, OSError) as error:
+        print(f"CcxStateUnavailable: managed worker variables unavailable: {error}",
+              file=sys.stderr)
+        return 69
+    if binding is None:
+        return _exec_native(real_python, args, env=native_env)
+
+    if len(args) >= 2 and args[0] == "-m" and args[1] in _NATIVE_FALLTHROUGH_MODULES:
+        return _exec_native(real_python, args, env=native_env)
+    # Shared variables cover `-c`, `-m`, stdin and scripts; any other leading
+    # interpreter option (--version, -V, -u, -I, ...) runs native Python.
+    if args and args[0].startswith("-") and args[0] not in ("-", "-c", "-m"):
+        return _exec_native(real_python, args, env=native_env)
+
+    short = owned.worker_id
+    if binding["mode"] == "durable":
+        try:
+            state_id = _valid_state_id(binding.get("stateId"))
+            with store.lease():
+                state = state_path(state_id)
+                lock = state_lock_path(state_id)
+        except (GenerationError, SessionUnavailable, OSError) as error:
+            print(f"CcxStateUnavailable: durable variable state unavailable: {error}",
+                  file=sys.stderr)
+            return 69
+        os.execvpe(
+            real_python,
+            _durable_runner_argv(
+                real_python, state=state, lock=lock, cwd=os.getcwd(), args=args,
+            ),
+            os.environ.copy() if native_env is None else native_env,
+        )
+        return 127
+
+    if binding["mode"] != "live":
+        print("CcxStateUnavailable: invalid worker variable mode", file=sys.stderr)
+        return 69
+    try:
+        value = read_sidecar(engine, short)
+    except (IdentityError, GenerationError, OSError) as error:
+        print(f"CcxStateUnavailable: live variable binding unavailable: {error}",
+              file=sys.stderr)
+        return 69
+    if not value or value.get("mode") != "live":
+        print("CcxStateUnavailable: live variable binding changed", file=sys.stderr)
+        return 69
+
+    owner = value.get("owner")
+    if not isinstance(owner, dict):
+        print("CcxStateUnavailable: live variable owner metadata is missing", file=sys.stderr)
+        return 69
+    if _owner_gone(owner):
+        _clear_dead_binding(engine, short, str(owner.get("token") or ""))
+        return _exec_native(real_python, args, env=native_env)
+    owner_python = owner.get("python")
+    native = _native_python_metadata(real_python)
+    if not isinstance(owner_python, dict) or native is None:
+        print(
+            "ccx: cannot verify Python compatibility for shared variables; running native Python",
+            file=sys.stderr,
+        )
+        return _exec_native(real_python, args, env=native_env)
+    if not _compatible(owner_python, native):
+        print(
+            "ccx: selected Python is incompatible with the live variable owner; running native Python",
+            file=sys.stderr,
+        )
+        return _exec_native(real_python, args, env=native_env)
+
+    request = {
+        "schemaVersion": VARIABLE_SCHEMA_VERSION,
+        "op": "exec",
+        "token": owner.get("token"),
+        "engine": engine,
+        "short": short,
+        "args": args,
+        "cwd": os.getcwd(),
+        "stdinBase64": base64.b64encode(
+            sys.stdin.buffer.read() if not args or args[0] == "-" else b""
+        ).decode("ascii"),
+        "nativePython": native,
+    }
+    sock = owner.get("socket")
+    if not isinstance(sock, str):
+        print("CcxStateUnavailable: live variable owner socket is missing", file=sys.stderr)
+        return 69
+
+    client = socket.socket(socket.AF_UNIX, socket.SOCK_STREAM)
+    try:
+        client.connect(sock)
+        client.sendall((json.dumps(request, separators=(",", ":")) + "\n").encode("utf-8"))
+        raw = LiveOwner._read_line(client)
+    except OSError as error:
+        print(
+            f"CcxStateUnavailable: shared variable owner is unavailable: {error}",
+            file=sys.stderr,
+        )
+        return 69
+    finally:
+        client.close()
+
+    try:
+        response = json.loads(raw)
+    except ValueError:
+        print("CcxStateUnavailable: invalid response from shared variable owner", file=sys.stderr)
+        return 69
+
+    if not response.get("ok"):
+        kind = response.get("errorKind")
+        if kind == "owner_busy":
+            print(
+                f"CcxStateBusy: shared Python owner is busy: {response.get('current')}",
+                file=sys.stderr,
+            )
+            return 75
+        print(
+            "CcxStateUnavailable: shared variable execution is unavailable"
+            + (f": {response.get('message')}" if response.get("message") else ""),
+            file=sys.stderr,
+        )
+        return 69
+
+    stdout = base64.b64decode(response.get("stdoutBase64") or "")
+    stderr = base64.b64decode(response.get("stderrBase64") or "")
+    if stdout:
+        sys.stdout.buffer.write(stdout)
+        sys.stdout.buffer.flush()
+    if stderr:
+        sys.stderr.buffer.write(stderr)
+        sys.stderr.buffer.flush()
+    return int(response.get("exitCode") or 0)
+
+
+def cmd_worker_exec(args) -> None:
+    target_args = list(getattr(args, "target_args", []) or [])
+    if target_args and target_args[0] == "--":
+        target_args = target_args[1:]
+    raise SystemExit(
+        worker_exec(
+            str(args.engine),
+            str(args.short),
+            str(args.invoked_as),
+            target_args,
+            native_env=getattr(args, "_bridge_native_env", None),
+        )
+    )