Files
hermes-agent/tools/code_kernel_remote.py
kshitijk4poor 30c44c7555 fix(code-kernel): make post-result cell cleanup best-effort
The rm of the cell result file runs after the runner has executed the cell.
A transport failure there propagated out of _run_remote_cell, so
_run_attached_cell evicted the kernel and re-raised, and the caller's
per-call fallback then ran the same code a second time.

Catch and debug-log that failure (a leftover cell_res_* is harmless because
seq is monotonic), so the atomic ship is the only remote call that can raise
inside the discard scope and the handler's "runs exactly once" comment holds.
2026-09-27 00:56:31 +05:30

440 lines
20 KiB
Python

"""Session-persistent kernels for REMOTE terminal backends (docker/ssh/modal).
Remote backends offer one primitive — ``env.execute(cmd)``, run-to-completion
— so the three things the local kernel gets from owning a child are rebuilt:
a detached runner (``nohup ... &``, PID recorded, ``kill -0`` probed per cell);
a file-based CELL protocol in the kernel dir (``cell_req_NNNNNN.json`` /
``cell_res_NNNNNN.json``), sibling to the unchanged file-based TOOL-RPC protocol
(req_/res_) whose host-side ``_rpc_poll_loop`` starts per cell with the calling
thread's context (= per-cell tool authority); and death detection — a failed
liveness probe reads as *kernel died: state lost* and the next call respawns,
never a hung poll (every wait is bounded by the cell timeout).
Same invariants as local: owner = approval session key with the ``::child::``
qualifier (one resolver in tools.code_kernel), same generated tool stubs, same
output post-processing in the caller. ``reset=true`` kills and respawns. Spawn
failure fails OPEN to the per-call path with a note.
"""
from __future__ import annotations
import atexit
import json
import logging
import secrets
import shlex
import threading
import time
import uuid
from dataclasses import dataclass, field
from typing import Any, Callable, Dict, List, Optional, Tuple
from tools.code_kernel import RUNNER_CELL_SOURCE, KernelRegistry
logger = logging.getLogger(__name__)
# Host poll interval for a cell result file; each poll is one env.execute round-trip
# (0.1-0.4s on ssh/docker), so this is a floor, not a rate.
_CELL_POLL_INTERVAL = 0.5
# The remote runner: a forever-loop that polls for cell request files, execs them in one
# persistent namespace, writes response files. Pure files + stdlib only (transport-agnostic);
# cells and tool-RPC share the kernel dir under distinct prefixes.
REMOTE_KERNEL_RUNNER_SOURCE = '''\
"""Auto-generated Hermes REMOTE session-kernel runner (file cell protocol)."""
import contextlib
import io
import json
import os
import sys
import time
import traceback
KDIR = os.environ["HERMES_KERNEL_DIR"]
CELLS = os.path.join(KDIR, "cells")
_CAPTURE_LIMIT = {capture_limit}
IDLE_EXIT_SECONDS = {idle_exit}
{cell_source}
def main():
execution_count = 0
last_activity = time.time()
while True:
pending = sorted(
f for f in os.listdir(CELLS)
if f.startswith("cell_req_") and f.endswith(".json")
)
if not pending:
if time.time() - last_activity > IDLE_EXIT_SECONDS:
return # self-reap: nobody is talking to us anymore
time.sleep(0.2)
continue
for name in pending:
req_path = os.path.join(CELLS, name)
try:
with open(req_path, "r", encoding="utf-8") as f:
request = json.load(f)
except Exception:
# Partially-written request (ship in progress): retry next tick.
continue
os.remove(req_path)
last_activity = time.time()
execution_count += 1
payload, _ = run_cell(request, execution_count)
res_name = name.replace("cell_req_", "cell_res_")
tmp = os.path.join(CELLS, res_name + ".tmp")
# Cell results carry the executed code's output: owner-only, even if
# the process umask is permissive.
fd = os.open(tmp, os.O_WRONLY | os.O_CREAT | os.O_TRUNC, 0o600)
with os.fdopen(fd, "w", encoding="utf-8") as f:
json.dump(payload, f, ensure_ascii=False)
os.replace(tmp, os.path.join(CELLS, res_name))
if payload["status"] == "exit":
return
if __name__ == "__main__":
main()
'''
def _sh(env, cmd: str, timeout: int = 15) -> str:
"""Run *cmd* on the remote from ``/`` and return its output text."""
result = env.execute(cmd, cwd="/", timeout=timeout)
return (result.get("output", "") if isinstance(result, dict) else "") or ""
@dataclass
class RemoteKernel:
"""Host-side record of one detached remote kernel process."""
env: Any
env_type: str
kernel_dir: str
pid: str
rpc_token: str
owner: str
last_used: float = field(default_factory=time.monotonic)
execution_count: int = 0
cell_seq: int = 0
# Cells currently running on this kernel. Reap/evict skip attached
# kernels: killing one mid-cell tears the runner out from under a live
# poll loop (same guard as tools.code_kernel, hermes-agent#101861).
attached: int = 0
# Owned by a live delegate_task child: exempt from LRU eviction (the child's teardown disposes it).
pinned: bool = False
def sh(self, cmd: str, timeout: int = 15) -> str:
return _sh(self.env, cmd, timeout)
def is_alive(self) -> bool:
"""Bounded liveness probe: kill -0 through the transport. Any transport
failure counts as dead — a dropped ssh connection and a dead runner are
indistinguishable from here, and both have the same correct answer (respawn)."""
try:
return "ALIVE" in self.sh(f"kill -0 {shlex.quote(self.pid)} 2>/dev/null && echo ALIVE")
except Exception:
return False
def kill(self) -> None:
"""Best-effort kill of the runner and its subprocesses, then rm -rf."""
q_pid = shlex.quote(self.pid)
for cmd, failure in (
# Kill the runner's children if the shell gave it a group, then the PID itself.
(f"pkill -TERM -P {q_pid} 2>/dev/null; kill {q_pid} 2>/dev/null; true",
"remote kernel kill failed (transport?)"),
(f"rm -rf {shlex.quote(self.kernel_dir)}", "remote kernel dir cleanup failed"),
):
try:
self.sh(cmd)
except Exception:
logger.debug(failure, exc_info=True)
def _kernel_key(owner: str, env_type: str, task_env_id: str, sandbox_tools: frozenset) -> Tuple:
"""The hermes_tools stub module is generated from ``sandbox_tools`` once, at spawn, so a kernel
is only reusable by calls with the SAME tool set; a different set gets its own kernel."""
return (owner, "remote", env_type, task_env_id, tuple(sorted(sandbox_tools)))
# Registry + lock shared-shape with code_kernel; teardown runs outside the lock.
_REGISTRY = KernelRegistry(lambda kernel: kernel.kill())
_REMOTE_KERNELS: Dict[Tuple, RemoteKernel] = _REGISTRY.kernels
def shutdown_all_remote_kernels() -> None:
_REGISTRY.shutdown()
def shutdown_remote_kernels_for_owner(owner: str) -> None:
"""Session-boundary disposal — wired to the same clear_session hook as
local kernels, so /new and session close reap both kinds."""
if owner:
_REGISTRY.shutdown(owner)
def shutdown_remote_kernels_where(owner_matches: Callable[[str], bool]) -> None:
"""Dispose every remote kernel whose owner satisfies the predicate (a finished child's kernels)."""
_REGISTRY.shutdown(owner_matches=owner_matches)
def _reap_unlocked(idle_timeout: int) -> List["RemoteKernel"]:
"""Pop idle-expired, unattached remote kernels; caller tears them down outside the lock. The
runner self-exits after the same idle window, so this clears the HOST-side entry — without it
the map grew one entry per never-revisited (owner, env_type, task_env_id) for the gateway's life."""
now = time.monotonic()
doomed = [key for key, kernel in _REMOTE_KERNELS.items()
if kernel.attached == 0 and now - kernel.last_used > idle_timeout]
return [_REMOTE_KERNELS.pop(key) for key in doomed]
def _evict_over_cap_unlocked(keep: Tuple) -> List["RemoteKernel"]:
"""Pop least-recently-used unattached remote kernels beyond the process-wide cap (the same
``max_session_kernels`` bound as local kernels, applied independently to this map)."""
from tools.code_kernel import _lifecycle_limits
cap, _ = _lifecycle_limits()
unpinned = [key for key in _REMOTE_KERNELS if not _REMOTE_KERNELS[key].pinned]
if len(unpinned) <= cap:
return []
by_age = sorted((key for key in unpinned if key != keep and _REMOTE_KERNELS[key].attached == 0),
key=lambda key: _REMOTE_KERNELS[key].last_used)
return [_REMOTE_KERNELS.pop(key) for key in by_age[: len(unpinned) - cap]]
atexit.register(shutdown_all_remote_kernels)
def _spawn_remote_kernel(env, env_type: str, owner: str, task_env_id: str,
sandbox_tools: frozenset, *, idle_exit: int) -> Optional[RemoteKernel]:
"""Start a detached kernel runner on the remote. None on failure (dir removed)."""
from tools.code_execution_rpc import _execute_checked, _private_dirs_cmd
from tools.code_execution_tool import (
MAX_STDOUT_BYTES, _ship_file_to_remote, _env_temp_dir,
_ship_env_file_and_launch, generate_hermes_tools_module,
)
kernel_dir = f"{_env_temp_dir(env)}/hermes_rkernel_{uuid.uuid4().hex[:12]}"
q_dir = shlex.quote(kernel_dir)
kernel = None
try:
# Private dirs: the kernel dir lives under a shared temp dir and carries
# the RPC token (in req files), tool results, and cell code/output.
# Fail closed on setup failure rather than ship secrets into a dir that
# stayed permissive.
_execute_checked(env, _private_dirs_cmd(kernel_dir, f"{kernel_dir}/cells",
f"{kernel_dir}/rpc"),
"remote kernel dir setup", timeout=15)
rpc_token = secrets.token_urlsafe(32)
_ship_file_to_remote(env, f"{kernel_dir}/kernel_runner.py", REMOTE_KERNEL_RUNNER_SOURCE.format(
cell_source=RUNNER_CELL_SOURCE, capture_limit=MAX_STDOUT_BYTES, idle_exit=idle_exit))
_ship_file_to_remote(env, f"{kernel_dir}/hermes_tools.py",
generate_hermes_tools_module(list(sandbox_tools), transport="file"))
# kernel.env is removed after sourcing: the runner's env keeps the
# values, so the token file need not sit at rest for the kernel's
# lifetime. runner.log is pre-created 600 so the launch redirect never
# lands at the remote's default umask. The inner `&` stays inside the
# subshell where `$!` resolves to the runner pid.
launch_cmd = _ship_env_file_and_launch(
env, kernel_dir, "kernel.env",
"rm -f ./kernel.env && touch runner.log && chmod 600 runner.log && "
'{ nohup python3 kernel_runner.py > runner.log 2>&1 & echo "PID:$!"; }',
rpc_dir=f"{kernel_dir}/rpc", rpc_token=rpc_token,
HERMES_KERNEL_DIR=kernel_dir, PYTHONPATH=kernel_dir)
started = _sh(env, launch_cmd, timeout=20)
pid = next((line.strip()[4:].strip() for line in started.splitlines()
if line.strip().startswith("PID:")), "")
if not pid.isdigit():
logger.warning("remote kernel spawn returned no PID: %r", started)
else:
candidate = RemoteKernel(env=env, env_type=env_type, kernel_dir=kernel_dir,
pid=pid, rpc_token=rpc_token, owner=owner)
if candidate.is_alive():
kernel = candidate
else:
# Died instantly (missing python3 was pre-checked by the caller,
# so this is unexpected) — surface the runner log.
try:
logger.warning("remote kernel died at spawn: %s",
_sh(env, f"cat {q_dir}/runner.log", timeout=10)[:500])
except Exception:
pass
except Exception:
logger.warning("remote kernel spawn failed", exc_info=True)
if kernel is None:
try:
_sh(env, f"rm -rf {q_dir}")
except Exception:
pass
return kernel
def _acquire_remote_kernel(env, env_type: str, owner: str, task_env_id: str,
sandbox_tools: frozenset, *, reset: bool,
idle_exit: int) -> Tuple[Optional[RemoteKernel], bool, bool, bool]:
"""Find/respawn the owner's kernel: (kernel|None, reused, state_reset, state_lost); reaps
idle-expired entries on the way in."""
key = _kernel_key(owner, env_type, task_env_id, sandbox_tools)
state_lost = state_reset = False
with _REGISTRY.lock:
expired = _reap_unlocked(idle_exit)
kernel = _REMOTE_KERNELS.get(key)
for doomed in expired:
doomed.kill()
if kernel is not None and reset:
_REGISTRY.discard(key, kernel)
kernel, state_reset = None, True
if kernel is not None and not kernel.is_alive():
# Transport drop, container restart, self-reaped on idle, OOM — all
# the same answer: report the loss, respawn fresh (kill is then only
# best-effort dir cleanup; the process is already gone).
_REGISTRY.discard(key, kernel)
kernel, state_lost = None, True
reused = kernel is not None
if kernel is None:
kernel = _spawn_remote_kernel(env, env_type, owner, task_env_id, sandbox_tools, idle_exit=idle_exit)
if kernel is not None:
from agent.delegation_context import is_delegated_child_context
kernel.pinned = is_delegated_child_context()
with _REGISTRY.lock:
_REMOTE_KERNELS[key] = kernel
return kernel, reused, state_reset, state_lost
def _run_remote_cell(kernel: RemoteKernel, code: str, timeout: int) -> Tuple[str, Dict[str, Any]]:
"""Ship one cell request and poll for its result: (cell status, payload)."""
from tools.code_execution_tool import _ship_file_to_remote
kernel.cell_seq += 1
seq = f"{kernel.cell_seq:06d}"
q_cells, q_res = shlex.quote(f"{kernel.kernel_dir}/cells"), shlex.quote(f"cell_res_{seq}.json")
# One round-trip: tmp write + rename publishes the request atomically.
_ship_file_to_remote(kernel.env, f"{kernel.kernel_dir}/cells/cell_req_{seq}.json",
json.dumps({"id": seq, "code": code}, ensure_ascii=False), atomic=True)
deadline = time.monotonic() + timeout
while time.monotonic() < deadline:
try:
body = kernel.sh(f"cat {q_cells}/{q_res} 2>/dev/null", timeout=20).strip()
except Exception:
# One flaky round-trip is not kernel death; liveness decides.
body = ""
if body:
try:
payload = json.loads(body)
status = payload.get("status", "error")
except ValueError:
payload, status = {}, "protocol-error"
try:
kernel.sh(f"rm -f {q_cells}/{q_res}", timeout=10)
except Exception:
# Best-effort: the cell already ran, so raising here would send
# the caller to its per-call fallback and run the code twice.
# A leftover result file is harmless (seq is monotonic).
logger.debug("remote kernel: cell result cleanup failed", exc_info=True)
return status, payload
time.sleep(_CELL_POLL_INTERVAL)
return "timeout", {}
def execute_in_remote_kernel(
code: str, *, env, env_type: str, task_env_id: str, sandbox_tools: frozenset,
timeout: int, max_tool_calls: int, reset: bool, idle_exit: int = 1800,
) -> Optional[Dict[str, Any]]:
"""Run one cell in the owner's remote kernel. Returns the raw cell result dict (caller
post-processes output), or ``None`` when no kernel could be spawned (caller falls open to
per-call). ``state_lost``/``state_reset``/``reused`` ride in the ``kernel`` sub-dict."""
from tools.code_kernel import _resolve_owner
owner = _resolve_owner(task_env_id)
kernel, reused, state_reset, state_lost = _acquire_remote_kernel(
env, env_type, owner, task_env_id, sandbox_tools, reset=reset, idle_exit=idle_exit)
if kernel is None:
return None # fail open to per-call
key = _kernel_key(owner, env_type, task_env_id, sandbox_tools)
kernel.last_used = time.monotonic()
with _REGISTRY.lock:
kernel.attached += 1
evicted = _evict_over_cap_unlocked(keep=key)
for doomed in evicted:
doomed.kill()
try:
return _run_attached_cell(kernel, key, code, env=env, task_env_id=task_env_id,
sandbox_tools=sandbox_tools, timeout=timeout, max_tool_calls=max_tool_calls,
reused=reused, state_reset=state_reset, state_lost=state_lost)
finally:
with _REGISTRY.lock:
kernel.attached -= 1
kernel.last_used = time.monotonic()
def _run_attached_cell(kernel: RemoteKernel, key: Tuple, code: str, *, env, task_env_id: str,
sandbox_tools: frozenset, timeout: int, max_tool_calls: int,
reused: bool, state_reset: bool, state_lost: bool) -> Dict[str, Any]:
from tools.code_execution_tool import _rpc_poll_loop
from tools.thread_context import propagate_context_to_thread
# Clean stale tool-RPC requests from a previous cell before arming this cell's poll loop, so
# a background thread the last cell leaked cannot smuggle a call into this authority window.
q_rpc = shlex.quote(kernel.kernel_dir + '/rpc')
try:
kernel.sh(f"rm -f {q_rpc}/req_* {q_rpc}/res_*", timeout=10)
except Exception:
pass
tool_call_counter, stop_event = [0], threading.Event()
# Per-cell RPC thread carrying THIS call's approval/session context — the remote analogue
# of CellAuthority: authority lives exactly as long as the cell's poll loop.
rpc_thread = threading.Thread(
target=propagate_context_to_thread(_rpc_poll_loop), daemon=True,
args=(env, f"{kernel.kernel_dir}/rpc", task_env_id, [], tool_call_counter,
max_tool_calls, sandbox_tools, stop_event, kernel.rpc_token))
rpc_thread.start()
cell_status, cell_payload = "no-result", {}
try:
cell_status, cell_payload = _run_remote_cell(kernel, code, timeout)
except Exception:
# The atomic ship is the only remote call here that can raise (the poll
# and result cleanup are best-effort), so the request never reached the runner and the
# caller's per-call fallback runs the code exactly once. Kill the
# kernel as the timeout path does: leaving it registered would let the
# next call reuse a kernel whose state silently missed this cell.
_REGISTRY.discard(key, kernel)
raise
finally:
stop_event.set()
rpc_thread.join(timeout=5)
kernel_info: Dict[str, Any] = {"reused": reused, "remote": True}
result: Dict[str, Any] = {
"status": "error", "stdout": cell_payload.get("stdout", ""), "stderr": cell_payload.get("stderr", ""),
"traceback": cell_payload.get("traceback", ""), "tool_calls_made": tool_call_counter[0], "kernel": kernel_info,
}
if cell_status in ("timeout", "protocol-error", "no-result"):
# No safe way to interrupt one cell in place (same contract as local): kill, report, respawn.
_REGISTRY.discard(key, kernel)
if cell_status == "timeout":
result["status"] = "timeout"
kernel_info.update(ended=True, state_lost=True, note=(
"Cell timed out; the remote session kernel was killed and its state was lost. The next call "
"starts a fresh kernel." if cell_status == "timeout"
else "Remote kernel protocol failure; kernel killed, state lost."))
return result
if cell_status == "exit":
_REGISTRY.discard(key, kernel)
kernel_info["ended"] = True
kernel.execution_count = kernel_info["execution_count"] = int(cell_payload.get("execution_count", 0) or 0)
if cell_status in ("ok", "exit"):
result["status"] = "success"
result["stdout_clipped"] = bool(cell_payload.get("stdout_clipped"))
result["stderr_clipped"] = bool(cell_payload.get("stderr_clipped"))
if state_reset:
kernel_info["state_reset"] = True
if state_lost:
kernel_info.update(state_lost=True, note=(
"The previous remote kernel was gone (transport drop, container "
"restart, or idle self-exit); state from earlier calls was lost and a fresh kernel was started."))
if cell_status == "error" and result["traceback"]:
result["error"] = result["traceback"].strip().splitlines()[-1]
return result
# ---- BEGIN PLUGIN-COMPAT (revert-scheduled; see COMPAT_MANIFEST.md) ----
# Names external plugins imported from this module before the Sep 2026 decomposition.
# Internal code MUST NOT use these (scripts/check_compat_pointers.py fails CI if it does).
# The whole block is removed by reverting the commit that added it.
import base64 # noqa: F401,E402
# ---- END PLUGIN-COMPAT ----