leetcrypt
2091bcefed
The operator's capabilities contract was Claude-only — delivered via the hh-operator skill, which non-Claude runners cannot load. Extract it into a portable CAPABILITIES.md and make compose_directive runner-aware: the claude runner still loads its skill, every other runner gets the capabilities prompt inlined. Same contract, no skill machinery required. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
763 lines
35 KiB
Python
763 lines
35 KiB
Python
"""OperatorBridge — a headless room client a Claude Code session drives.
|
|
|
|
Inverts the `cmd_chat/agent` model: instead of the room calling a model per
|
|
message, this daemon owns the websocket and exposes the room as a local API.
|
|
A Claude Code session pumps it through the `hh-bridge` CLI (`read`/`say`/…),
|
|
which talks to a unix control socket served in this same event loop.
|
|
|
|
Phase 1 scope: join + roster + read (with long-poll) + say. Sandbox drive,
|
|
delegation and nesting are later phases — `_perm:acl`/`_sbx:status` frames are
|
|
*recorded* into the inbox for awareness but not acted on yet.
|
|
|
|
Reuses the protocol wholesale from `Client` (SRP, room key, encrypt/decrypt)
|
|
and mirrors `AgentBridge`'s reconnect/serve shape (cmd_chat/agent/bridge.py).
|
|
"""
|
|
|
|
from __future__ import annotations
|
|
|
|
import asyncio
|
|
import base64
|
|
import json
|
|
import os
|
|
import re
|
|
import time
|
|
|
|
import websockets
|
|
|
|
from ..client.client import Client
|
|
from . import bootstrap as boot
|
|
from . import sandbox as sbx
|
|
from .session import Session
|
|
|
|
|
|
class OperatorBridge(Client):
|
|
_PING_INTERVAL = 20.0
|
|
_PING_TIMEOUT = 60.0
|
|
_RECONNECT_MAX_BACKOFF = 30.0
|
|
_RECONNECT_STABLE_SECS = 30.0
|
|
MAX_EVENTS = 2000 # in-RAM inbox ring; full history lives in inbox.jsonl
|
|
TERM_CAP = 65536 # rolling relayed-PTY buffer (bytes) kept for `screen`
|
|
|
|
def __init__(self, server: str, port: int, name: str,
|
|
password: str | None = None, insecure: bool = False,
|
|
no_tls: bool = False, session: str | None = None,
|
|
trigger: str | None = None, budget: boot.Budget | None = None):
|
|
super().__init__(server, port, username=name, password=password,
|
|
insecure=insecure, no_tls=no_tls)
|
|
self.name = name
|
|
self.trigger = (trigger or "").strip()
|
|
self.session = Session(session or name)
|
|
# Recursion budget this operator may spend spawning nested operators.
|
|
self.budget = budget or boot.Budget()
|
|
|
|
# Inbox: monotonically-seq'd events + an async condition to wake waiters.
|
|
self.events: list[dict] = []
|
|
self.seq = 0
|
|
self.cond = asyncio.Condition()
|
|
|
|
# Mirrored room awareness (recorded only in Phase 1).
|
|
self.granted = False
|
|
self.can_sudo = False
|
|
self.sbx_engine: str | None = None
|
|
self.sbx_name: str = ""
|
|
|
|
# Co-located sandbox the operator launched itself (Phase 2). When set we
|
|
# exec straight into it (out-of-band); independent of the room's broker
|
|
# sandbox, which we can also drive when co-located + granted.
|
|
self._own_engine: str | None = None
|
|
self._own_name: str = ""
|
|
|
|
# Relay terminal buffer (Phase 6): rolling raw PTY bytes the broker
|
|
# relays as `_sbx:data`. Lets the operator *read* a remote/broker-owned
|
|
# sandbox it drives via `keys` — the output half of relay mode.
|
|
self._term = bytearray()
|
|
self._term_seq = 0 # bumps on each chunk so `watch` can poll for growth
|
|
|
|
# Live websocket (None while reconnecting); set in _connect_and_serve.
|
|
self._ws = None
|
|
self._stop = asyncio.Event()
|
|
self._seeded = False # backfill room history only on first init
|
|
self._inbox_fp = None # append handle to inbox.jsonl
|
|
self._addr_re = self._build_addr_re()
|
|
|
|
# ── addressing ───────────────────────────────────────────────────────
|
|
def _build_addr_re(self) -> re.Pattern:
|
|
toks = [re.escape(self.name), re.escape("@" + self.name)]
|
|
if self.trigger:
|
|
toks.append(re.escape(self.trigger))
|
|
return re.compile(r"(?<!\w)(" + "|".join(toks) + r")(?!\w)", re.IGNORECASE)
|
|
|
|
def _is_addressed(self, text: str) -> bool:
|
|
return bool(self._addr_re.search(text or ""))
|
|
|
|
# ── inbox ────────────────────────────────────────────────────────────
|
|
async def _emit(self, kind: str, **fields) -> None:
|
|
async with self.cond:
|
|
self.seq += 1
|
|
ev = {"seq": self.seq, "ts": round(time.time(), 3), "kind": kind, **fields}
|
|
self.events.append(ev)
|
|
if len(self.events) > self.MAX_EVENTS:
|
|
self.events = self.events[-self.MAX_EVENTS:]
|
|
if self._inbox_fp is not None:
|
|
try:
|
|
self._inbox_fp.write(json.dumps(ev) + "\n")
|
|
self._inbox_fp.flush()
|
|
except OSError:
|
|
pass
|
|
self.cond.notify_all()
|
|
|
|
# ── frame handling ───────────────────────────────────────────────────
|
|
async def _handle_frame(self, ws, raw) -> None:
|
|
try:
|
|
data = json.loads(raw)
|
|
except json.JSONDecodeError:
|
|
return
|
|
mtype = data.get("type")
|
|
if mtype == "init":
|
|
self.users = data.get("users", [])
|
|
await self._emit("roster", users=self._roster())
|
|
if not self._seeded:
|
|
for m in data.get("messages", []):
|
|
dec = self.decrypt_message(dict(m))
|
|
text = dec.get("text", "")
|
|
sender = dec.get("username", "?")
|
|
if (not text or text == "[decrypt failed]"
|
|
or text.startswith('{"_') or sender == self.name):
|
|
continue
|
|
await self._emit("message", **{"from": sender}, text=text,
|
|
addressed=False, backfill=True)
|
|
self._seeded = True
|
|
return
|
|
if mtype == "roster":
|
|
self.users = data.get("users", [])
|
|
await self._emit("roster", users=self._roster())
|
|
return
|
|
if mtype == "user_left":
|
|
left = data.get("user_id")
|
|
self.users = [u for u in self.users if u.get("user_id") != left]
|
|
await self._emit("roster", users=self._roster())
|
|
return
|
|
if mtype != "message":
|
|
return
|
|
msg = self.decrypt_message(data.get("data", {}))
|
|
text = msg.get("text", "")
|
|
sender = msg.get("username", "?")
|
|
if sender == self.name:
|
|
return
|
|
if text.startswith('{"_'):
|
|
await self._record_control(text)
|
|
return
|
|
await self._emit("message", **{"from": sender}, text=text,
|
|
addressed=self._is_addressed(text))
|
|
|
|
def _absorb_term(self, b64: str) -> None:
|
|
"""Append a relayed PTY chunk to the rolling terminal buffer (capped).
|
|
High-volume, so we don't emit an inbox event per chunk — `watch`/`screen`
|
|
poll `_term_seq` instead."""
|
|
if not b64:
|
|
return
|
|
try:
|
|
chunk = base64.b64decode(b64)
|
|
except (ValueError, TypeError):
|
|
return
|
|
self._term += chunk
|
|
if len(self._term) > self.TERM_CAP:
|
|
del self._term[:-self.TERM_CAP]
|
|
self._term_seq += 1
|
|
|
|
def _screen_text(self, tail: int | None, ansi: bool) -> str:
|
|
raw = bytes(self._term[-tail:] if tail else self._term)
|
|
s = raw.decode(errors="replace")
|
|
return s if ansi else sbx.strip_ansi(s)
|
|
|
|
async def _record_control(self, text: str) -> None:
|
|
"""Surface ACL + sandbox-status changes into the inbox (awareness only)
|
|
and absorb relayed PTY output (`_sbx:data`) into the terminal buffer.
|
|
Ignore other high-volume control frames (_ai, _ft)."""
|
|
try:
|
|
frame = json.loads(text)
|
|
except json.JSONDecodeError:
|
|
return
|
|
if frame.get("_sbx") == "data":
|
|
self._absorb_term(frame.get("b64", ""))
|
|
return
|
|
if frame.get("_sbx") == "status":
|
|
ready = frame.get("state") == "ready"
|
|
self.sbx_engine = frame.get("engine") if ready else None
|
|
self.sbx_name = (frame.get("name") or "") if ready else ""
|
|
await self._emit("sandbox", state=frame.get("state"),
|
|
engine=self.sbx_engine, name=self.sbx_name,
|
|
backend=frame.get("backend"))
|
|
elif frame.get("_perm") == "acl":
|
|
self.granted = self.name in frame.get("drivers", [])
|
|
self.can_sudo = self.name in frame.get("sudoers", [])
|
|
await self._emit("acl", owner=frame.get("owner"),
|
|
granted=self.granted, can_sudo=self.can_sudo,
|
|
drivers=frame.get("drivers", []))
|
|
|
|
def _roster(self) -> list[str]:
|
|
return [u.get("username", "?") for u in self.users]
|
|
|
|
# ── connection lifecycle (mirrors AgentBridge) ───────────────────────
|
|
async def run_async(self) -> None:
|
|
self.session.ensure_dir()
|
|
self.session.cleanup() # drop a stale socket from a prior run
|
|
self._inbox_fp = open(self.session.inbox_path, "a") # noqa: SIM115
|
|
server = await asyncio.start_unix_server(
|
|
self._serve_control, path=str(self.session.sock_path))
|
|
try:
|
|
os.chmod(self.session.sock_path, 0o600)
|
|
except OSError:
|
|
pass
|
|
await self._emit("system", event="starting",
|
|
host=self.server, port=self.port, user=self.name)
|
|
try:
|
|
await self._reconnect_loop()
|
|
finally:
|
|
server.close()
|
|
try:
|
|
await server.wait_closed()
|
|
except Exception: # noqa: BLE001
|
|
pass
|
|
if self._inbox_fp is not None:
|
|
self._inbox_fp.close()
|
|
self.session.cleanup()
|
|
|
|
async def _reconnect_loop(self) -> None:
|
|
backoff = 1.0
|
|
first = True
|
|
while not self._stop.is_set():
|
|
loop = asyncio.get_running_loop()
|
|
started = loop.time()
|
|
try:
|
|
await self._connect_and_serve(reconnect=not first)
|
|
except (KeyboardInterrupt, asyncio.CancelledError):
|
|
raise
|
|
except (websockets.ConnectionClosed, OSError) as e:
|
|
await self._emit("system", event="reconnecting",
|
|
reason=type(e).__name__)
|
|
except Exception as e: # noqa: BLE001 — auth/transient: report + retry
|
|
await self._emit("system", event="error",
|
|
reason=f"{type(e).__name__}: {e}")
|
|
else:
|
|
if not self._stop.is_set():
|
|
await self._emit("system", event="reconnecting",
|
|
reason="closed")
|
|
if self._stop.is_set():
|
|
break
|
|
if loop.time() - started >= self._RECONNECT_STABLE_SECS:
|
|
backoff = 1.0
|
|
first = False
|
|
try:
|
|
await asyncio.wait_for(self._stop.wait(), timeout=backoff)
|
|
except asyncio.TimeoutError:
|
|
pass
|
|
backoff = min(backoff * 2, self._RECONNECT_MAX_BACKOFF)
|
|
|
|
async def _connect_and_serve(self, reconnect: bool) -> None:
|
|
self.srp_authenticate() # fresh token each attempt; old one died on drop
|
|
url = f"{self.ws_url}/ws/chat?user_id={self.user_id}&ws_token={self.ws_token}"
|
|
async with websockets.connect(
|
|
url, ssl=self._ws_ssl_context(),
|
|
ping_interval=self._PING_INTERVAL, ping_timeout=self._PING_TIMEOUT,
|
|
) as ws:
|
|
self._ws = ws
|
|
self.running = True
|
|
announce = (f"{self.name} (operator) "
|
|
f"{'back online' if reconnect else 'online'} — Claude Code bridge")
|
|
await ws.send(self.room_fernet.encrypt(announce.encode()).decode())
|
|
await self._emit("system", event="connected", reconnect=reconnect)
|
|
try:
|
|
await self._serve(ws)
|
|
finally:
|
|
self._ws = None
|
|
self.running = False
|
|
|
|
async def _serve(self, ws) -> None:
|
|
async for raw in ws:
|
|
if not self.running:
|
|
break
|
|
try:
|
|
await self._handle_frame(ws, raw)
|
|
except (websockets.ConnectionClosed, asyncio.CancelledError):
|
|
raise
|
|
except Exception as e: # noqa: BLE001 — one bad frame mustn't drop us
|
|
await self._emit("system", event="frame_error",
|
|
reason=f"{type(e).__name__}: {e}")
|
|
|
|
# ── control socket (the CLI talks here) ──────────────────────────────
|
|
async def _serve_control(self, reader: asyncio.StreamReader,
|
|
writer: asyncio.StreamWriter) -> None:
|
|
try:
|
|
line = await reader.readline()
|
|
if not line:
|
|
return
|
|
try:
|
|
req = json.loads(line.decode())
|
|
except json.JSONDecodeError:
|
|
resp = {"ok": False, "error": "bad json"}
|
|
else:
|
|
resp = await self._dispatch(req)
|
|
writer.write((json.dumps(resp) + "\n").encode())
|
|
await writer.drain()
|
|
except (ConnectionResetError, BrokenPipeError):
|
|
pass
|
|
finally:
|
|
try:
|
|
writer.close()
|
|
except Exception: # noqa: BLE001
|
|
pass
|
|
|
|
async def _dispatch(self, req: dict) -> dict:
|
|
op = req.get("op")
|
|
if op == "ping":
|
|
return {"ok": True, "pong": True, "name": self.name,
|
|
"connected": self._ws is not None}
|
|
if op == "status":
|
|
return {"ok": True, "name": self.name,
|
|
"connected": self._ws is not None,
|
|
"granted": self.granted, "can_sudo": self.can_sudo,
|
|
"sbx_engine": self.sbx_engine, "sbx_name": self.sbx_name,
|
|
"users": self._roster(), "seq": self.seq,
|
|
"host": self.server, "port": self.port}
|
|
if op == "roster":
|
|
return {"ok": True, "users": self._roster()}
|
|
if op == "read":
|
|
return await self._op_read(req)
|
|
if op == "say":
|
|
return await self._op_say(req)
|
|
if op == "sbx":
|
|
return await self._op_sbx(req)
|
|
if op == "exec":
|
|
return await self._op_exec(req)
|
|
if op == "write":
|
|
return await self._op_write(req)
|
|
if op == "get":
|
|
return await self._op_get(req)
|
|
if op == "keys":
|
|
return await self._op_keys(req)
|
|
if op == "spawn":
|
|
return await self._op_spawn(req)
|
|
if op == "manifest":
|
|
return await self._op_manifest(req)
|
|
if op == "screen":
|
|
return await self._op_screen(req)
|
|
if op == "watch":
|
|
return await self._op_watch(req)
|
|
if op == "down":
|
|
asyncio.get_running_loop().call_soon(self._begin_shutdown)
|
|
return {"ok": True, "stopping": True}
|
|
return {"ok": False, "error": f"unknown op {op!r}"}
|
|
|
|
async def _op_read(self, req: dict) -> dict:
|
|
since = int(req.get("since", 0))
|
|
wait = bool(req.get("wait", False))
|
|
timeout = float(req.get("timeout", 30.0))
|
|
async with self.cond:
|
|
new = [e for e in self.events if e["seq"] > since]
|
|
if not new and wait:
|
|
try:
|
|
await asyncio.wait_for(self.cond.wait(), timeout)
|
|
except asyncio.TimeoutError:
|
|
pass
|
|
new = [e for e in self.events if e["seq"] > since]
|
|
return {"ok": True, "events": new, "seq": self.seq}
|
|
|
|
async def _op_say(self, req: dict) -> dict:
|
|
text = str(req.get("text", ""))
|
|
if not text:
|
|
return {"ok": False, "error": "empty text"}
|
|
if self._ws is None:
|
|
return {"ok": False, "error": "not connected to room"}
|
|
try:
|
|
await self._ws.send(self.room_fernet.encrypt(text.encode()).decode())
|
|
except Exception as e: # noqa: BLE001
|
|
return {"ok": False, "error": f"send failed: {type(e).__name__}: {e}"}
|
|
await self._emit("sent", **{"from": self.name}, text=text, addressed=False)
|
|
return {"ok": True}
|
|
|
|
# ── sandbox drive (Phase 2) ──────────────────────────────────────────
|
|
def _target(self) -> tuple[str | None, str, str | None]:
|
|
"""Pick the sandbox to exec into and report why.
|
|
|
|
Returns ``(engine, name, error)``. Prefer the operator's *own* container
|
|
(always ours to drive). Else, if the room granted us drive AND its broker
|
|
sandbox is a kind we can exec directly (we're co-located on this host),
|
|
target that. ``local`` is refused here — execing on the host is opt-in
|
|
only via our own `local` launch, never inherited from the room."""
|
|
if self._own_engine:
|
|
return self._own_engine, self._own_name, None
|
|
if self.granted and self.sbx_engine in ("docker", "podman", "multipass") \
|
|
and self.sbx_name:
|
|
return self.sbx_engine, self.sbx_name, None
|
|
if self.sbx_engine and not self.granted:
|
|
return None, "", "not a granted driver (owner must `/grant`)"
|
|
return None, "", "no sandbox — `sbx launch` one or join a room with one"
|
|
|
|
async def _op_sbx(self, req: dict) -> dict:
|
|
"""Manage the operator's own co-located container: launch / status / down."""
|
|
action = req.get("action", "status")
|
|
if action == "launch":
|
|
engine = req.get("engine") or sbx.pick_engine()
|
|
if not engine:
|
|
return {"ok": False, "error": "no container engine (need podman or docker)"}
|
|
image = req.get("image") or sbx.default_image(engine)
|
|
name = req.get("name") or f"hh-op-{self.name}"
|
|
ok, msg = await sbx.launch_container(engine, name, image)
|
|
if ok:
|
|
self._own_engine, self._own_name = engine, name
|
|
await self._emit("sandbox", state="own-ready", engine=engine,
|
|
name=name, image=image)
|
|
return {"ok": ok, "engine": engine, "name": name, "image": image,
|
|
"message": msg}
|
|
if action == "down":
|
|
if not self._own_engine:
|
|
return {"ok": False, "error": "no own sandbox to tear down"}
|
|
ok, msg = await sbx.teardown_container(self._own_engine, self._own_name)
|
|
if ok:
|
|
await self._emit("sandbox", state="own-down",
|
|
engine=self._own_engine, name=self._own_name)
|
|
self._own_engine, self._own_name = None, ""
|
|
return {"ok": ok, "message": msg}
|
|
# status
|
|
engine, name, err = self._target()
|
|
return {"ok": True, "target_engine": engine, "target_name": name,
|
|
"reason": err, "own_engine": self._own_engine,
|
|
"own_name": self._own_name, "room_engine": self.sbx_engine,
|
|
"room_name": self.sbx_name, "granted": self.granted}
|
|
|
|
async def _op_exec(self, req: dict) -> dict:
|
|
"""Run a shell command in the target sandbox, capture combined output +
|
|
exit code. The command runs under `sh -c` (a real shell, intended) inside
|
|
the container — the container is the blast radius."""
|
|
cmd = str(req.get("cmd", ""))
|
|
if not cmd:
|
|
return {"ok": False, "error": "empty cmd"}
|
|
engine, name, err = self._target()
|
|
if err:
|
|
return {"ok": False, "error": err}
|
|
prefix = sbx.exec_prefix(engine, name)
|
|
if prefix is None:
|
|
return {"ok": False, "error": f"can't address {engine}/{name}"}
|
|
out, rc = await sbx.exec_capture(prefix + ["sh", "-c", cmd])
|
|
await self._emit("exec", engine=engine, name=name, cmd=cmd, rc=rc)
|
|
return {"ok": rc == 0, "rc": rc, "output": out, "engine": engine, "name": name}
|
|
|
|
async def _op_write(self, req: dict) -> dict:
|
|
"""Write a file in the target sandbox. Content arrives on stdin (never
|
|
shell-parsed); the path is a positional arg. `mkdir -p` the parent first
|
|
so an absolute path into a new dir works. Mirrors the native harness."""
|
|
path = str(req.get("path", "")).strip()
|
|
if not path:
|
|
return {"ok": False, "error": "missing path"}
|
|
content = req.get("content", "")
|
|
data = content.encode() if isinstance(content, str) else bytes(content)
|
|
engine, name, err = self._target()
|
|
if err:
|
|
return {"ok": False, "error": err}
|
|
prefix = sbx.exec_prefix(engine, name)
|
|
if prefix is None:
|
|
return {"ok": False, "error": f"can't address {engine}/{name}"}
|
|
out, rc = await sbx.exec_capture(
|
|
prefix + ["sh", "-c", 'mkdir -p "$(dirname "$1")" && cat > "$1"',
|
|
"hh-write", path],
|
|
stdin=data)
|
|
await self._emit("write", engine=engine, name=name, path=path, rc=rc)
|
|
return {"ok": rc == 0, "rc": rc, "path": path, "bytes": len(data),
|
|
"output": out if (out or "").strip() else ""}
|
|
|
|
async def _op_get(self, req: dict) -> dict:
|
|
"""Read a file out of the target sandbox. Returns base64 (binary-safe);
|
|
the content is plumbing, so it's uncapped (unlike exec output)."""
|
|
path = str(req.get("path", "")).strip()
|
|
if not path:
|
|
return {"ok": False, "error": "missing path"}
|
|
engine, name, err = self._target()
|
|
if err:
|
|
return {"ok": False, "error": err}
|
|
prefix = sbx.exec_prefix(engine, name)
|
|
if prefix is None:
|
|
return {"ok": False, "error": f"can't address {engine}/{name}"}
|
|
raw, rc = await sbx.exec_capture(prefix + ["cat", "--", path], text=False)
|
|
if rc != 0:
|
|
return {"ok": False, "rc": rc,
|
|
"error": f"read failed (exit={rc}): {raw.decode(errors='replace')[:200]}"}
|
|
return {"ok": True, "rc": 0, "path": path,
|
|
"b64": base64.b64encode(raw).decode(), "bytes": len(raw)}
|
|
|
|
async def _op_keys(self, req: dict) -> dict:
|
|
"""Relay raw keystrokes into the room's *shared* PTY — the same
|
|
`_sbx:input` frame a human driver's terminal emits. The broker writes our
|
|
bytes to the shared shell iff we're a granted driver, so this is how the
|
|
operator drives a broker-owned sandbox (and tests interactive programs:
|
|
ctrl-c to stop, q to quit a pager, esc to leave vim). See
|
|
`sandbox.KEYS_HELP` for the byte vocabulary."""
|
|
if self._ws is None:
|
|
return {"ok": False, "error": "not connected to room"}
|
|
if not self.granted:
|
|
return {"ok": False, "error": "not a granted driver — keystrokes inert until `/grant`"}
|
|
spec = req.get("keys")
|
|
if spec in (None, "", []):
|
|
return {"ok": False, "error": "no keys", "help": sbx.KEYS_HELP}
|
|
try:
|
|
data = sbx.encode_keys(spec)
|
|
except ValueError as e:
|
|
return {"ok": False, "error": f"bad key spec: {e}"}
|
|
frame = json.dumps({"_sbx": "input", "b64": base64.b64encode(data).decode()})
|
|
try:
|
|
await self._ws.send(self.room_fernet.encrypt(frame.encode()).decode())
|
|
except Exception as e: # noqa: BLE001
|
|
return {"ok": False, "error": f"send failed: {type(e).__name__}: {e}"}
|
|
await self._emit("keys", bytes=len(data))
|
|
return {"ok": True, "bytes": len(data)}
|
|
|
|
# ── relay read + stop-condition engine (Phase 6) ─────────────────────
|
|
async def _op_screen(self, req: dict) -> dict:
|
|
"""Return the current relayed terminal buffer — the output half of relay
|
|
mode (drive with `keys`, read with `screen`). ANSI-stripped by default;
|
|
`tail` bounds the bytes returned."""
|
|
tail = req.get("tail")
|
|
tail = int(tail) if tail else None
|
|
ansi = bool(req.get("ansi", False))
|
|
return {"ok": True, "text": self._screen_text(tail, ansi),
|
|
"bytes": len(self._term), "seq": self._term_seq}
|
|
|
|
async def _op_watch(self, req: dict) -> dict:
|
|
"""Block until a stop condition fires, then report which one. Formalizes
|
|
the `/loop` autonomy stops so a driver can `keys` then wait for a result.
|
|
|
|
Conditions (first to fire wins):
|
|
- ``for`` regex matched in the chosen source ``in`` = "events" (default,
|
|
matches event text) or "screen" (matches relayed terminal output);
|
|
- ``idle`` seconds with no new event/output (quiescence);
|
|
- ``timeout`` seconds elapsed (the hard ceiling).
|
|
"""
|
|
pattern = req.get("for")
|
|
source = req.get("in", "events")
|
|
timeout = float(req.get("timeout", 30.0))
|
|
idle = req.get("idle")
|
|
idle = float(idle) if idle is not None else None
|
|
match_since = int(req.get("since", self.seq)) # fixed scan cursor
|
|
try:
|
|
rx = re.compile(pattern) if pattern else None
|
|
except re.error as e:
|
|
return {"ok": False, "error": f"bad regex: {e}"}
|
|
|
|
loop = asyncio.get_running_loop()
|
|
deadline = loop.time() + timeout
|
|
last_change = loop.time()
|
|
prev_seq, prev_term = self.seq, self._term_seq
|
|
while True:
|
|
if rx and source == "screen":
|
|
m = rx.search(self._screen_text(None, False))
|
|
if m:
|
|
return {"ok": True, "reason": "match", "where": "screen",
|
|
"match": m.group(0), "seq": self.seq}
|
|
if rx and source == "events":
|
|
for e in self.events:
|
|
if e["seq"] <= match_since:
|
|
continue
|
|
if rx.search(str(e.get("text", "")) or json.dumps(e)):
|
|
return {"ok": True, "reason": "match", "where": "events",
|
|
"event": e, "seq": self.seq}
|
|
now = loop.time()
|
|
if self.seq != prev_seq or self._term_seq != prev_term:
|
|
last_change, prev_seq, prev_term = now, self.seq, self._term_seq
|
|
if idle is not None and now - last_change >= idle:
|
|
return {"ok": True, "reason": "idle", "idle": idle, "seq": self.seq}
|
|
if now >= deadline:
|
|
return {"ok": True, "reason": "timeout", "matched": False,
|
|
"seq": self.seq}
|
|
await asyncio.sleep(min(0.3, max(0.0, deadline - now)))
|
|
|
|
# ── recursive spawn (Phase 5) ────────────────────────────────────────
|
|
async def _op_spawn(self, req: dict) -> dict:
|
|
"""Stand up a nested Claude Code operator against another room.
|
|
|
|
Guardrails, in order:
|
|
- **budget** — refuse unless this node may still spawn (depth>0 &
|
|
fanout>0); the child inherits a decremented budget, we record one spent
|
|
slot. This is the hard stop against a runaway tree.
|
|
- **creds gating** — we carry our own credentials to the child *only* when
|
|
`allow_creds` is explicitly set; otherwise the child authenticates
|
|
itself. Never implicit, never to a system we don't own.
|
|
- **dry_run** — default returns the full plan (argv, budget, creds
|
|
decision) and launches nothing, so a tree can be inspected before it's
|
|
grown.
|
|
|
|
`target` is "host" (Popen detached here) — sandbox/VM placement is the
|
|
operator's to run via `exec` using the returned plan."""
|
|
objective = str(req.get("objective", "")).strip()
|
|
if not objective:
|
|
return {"ok": False, "error": "spawn needs an objective"}
|
|
room = req.get("room") or {}
|
|
if not room.get("host") or not room.get("port"):
|
|
return {"ok": False, "error": "spawn needs room {host, port, name}"}
|
|
|
|
if not self.budget.can_spawn():
|
|
return {"ok": False, "error": "recursion budget exhausted "
|
|
f"(depth={self.budget.depth}, fanout={self.budget.fanout})"}
|
|
try:
|
|
child_budget = self.budget.descend()
|
|
except boot.BudgetExhausted as e:
|
|
return {"ok": False, "error": str(e)}
|
|
|
|
allow_creds = bool(req.get("allow_creds", False))
|
|
dry_run = bool(req.get("dry_run", True))
|
|
skip_perms = bool(req.get("skip_permissions", False))
|
|
config_dir = req.get("config_dir")
|
|
try:
|
|
runner = boot.get_runner(req.get("runner"))
|
|
except ValueError as e:
|
|
return {"ok": False, "error": str(e)}
|
|
|
|
stop = req.get("stop") or None
|
|
directive = boot.compose_directive(objective, room, child_budget,
|
|
stop=stop, runner=runner)
|
|
argv = boot.build_run_argv(directive, skip_permissions=skip_perms,
|
|
config_dir=config_dir, runner=runner)
|
|
creds = boot.plan_creds(config_dir or "~", allow=allow_creds, runner=runner)
|
|
present = boot.runner_present(runner)
|
|
plan = {"argv": argv, "budget": child_budget.as_dict(), "creds": creds,
|
|
"target": req.get("target", "host"),
|
|
"runner": runner.name, "runner_present": present}
|
|
|
|
if dry_run:
|
|
return {"ok": True, "dry_run": True, "plan": plan}
|
|
|
|
if not present:
|
|
return {"ok": False, "error": f"{runner.name} CLI not installed",
|
|
"install_plan": boot.install_plan(runner=runner), "plan": plan}
|
|
|
|
# Carry creds only behind the explicit gate.
|
|
if creds.get("staged"):
|
|
try:
|
|
dest = boot.Path(creds["dest"])
|
|
dest.parent.mkdir(parents=True, exist_ok=True)
|
|
dest.write_bytes(boot.Path(creds["src"]).read_bytes())
|
|
dest.chmod(0o600)
|
|
except OSError as e:
|
|
return {"ok": False, "error": f"creds staging failed: {e}"}
|
|
|
|
# Launch detached so the child outlives this daemon; logs to the session.
|
|
import subprocess
|
|
try:
|
|
logf = open(self.session.dir / f"spawn-{self.seq + 1}.log", "a") # noqa: SIM115
|
|
proc = subprocess.Popen(
|
|
argv, env=boot.child_env(config_dir, runner=runner),
|
|
stdout=logf, stderr=logf, stdin=subprocess.DEVNULL,
|
|
start_new_session=True, close_fds=True)
|
|
except OSError as e:
|
|
return {"ok": False, "error": f"spawn launch failed: {e}"}
|
|
|
|
self.budget = self.budget.spent() # consume one child slot here
|
|
await self._emit("spawn", objective=objective, pid=proc.pid,
|
|
budget=child_budget.as_dict(),
|
|
creds_staged=creds.get("staged", False))
|
|
return {"ok": True, "pid": proc.pid, "plan": plan,
|
|
"remaining": self.budget.as_dict()}
|
|
|
|
# ── agent-manifest: make a sandbox carry its own handoff record ──────────
|
|
async def _op_manifest(self, req: dict) -> dict:
|
|
"""Read/write the `.hh-agent/` manifest bundle *inside* the target
|
|
sandbox, so a VM travels with a self-describing handoff record.
|
|
|
|
Actions:
|
|
- ``push`` — build a bundle from the given fields and write all five
|
|
files under ``<root>/.hh-agent`` in the sandbox. This is what makes a
|
|
VM tradeable: the next agent reads AGENT.md and knows the purpose.
|
|
- ``pull`` — cat the canonical ``manifest.yaml`` back out, parsed, so a
|
|
handoff/spawn can resume exactly where the last one left off.
|
|
- ``update`` — pull, apply a state change (status/progress/done/todo/
|
|
blockers + a provenance entry), and re-push the rendered bundle.
|
|
|
|
All manifest logic lives in ``manifest.py``; the bridge only moves bytes
|
|
in and out of the sandbox via the same exec path as `write`/`get`."""
|
|
from . import manifest as mf
|
|
|
|
action = req.get("action", "push")
|
|
root = str(req.get("root", "/root")).rstrip("/") or "/"
|
|
bdir = f"{root}/{mf.MANIFEST_DIR}"
|
|
engine, name, err = self._target()
|
|
if err:
|
|
return {"ok": False, "error": err}
|
|
prefix = sbx.exec_prefix(engine, name)
|
|
if prefix is None:
|
|
return {"ok": False, "error": f"can't address {engine}/{name}"}
|
|
|
|
async def _put(path: str, data: bytes):
|
|
return await sbx.exec_capture(
|
|
prefix + ["sh", "-c", 'mkdir -p "$(dirname "$1")" && cat > "$1"',
|
|
"hh-manifest", path], stdin=data)
|
|
|
|
async def _read_manifest() -> mf.Manifest | None:
|
|
raw, rc = await sbx.exec_capture(
|
|
prefix + ["cat", "--", f"{bdir}/{mf.MANIFEST_FILE}"], text=False)
|
|
if rc != 0:
|
|
return None
|
|
return mf.Manifest.from_dict(mf._load(raw.decode(errors="replace")))
|
|
|
|
async def _write_bundle(m: mf.Manifest):
|
|
for fname, text in mf.bundle_files(m).items():
|
|
_, rc = await _put(f"{bdir}/{fname}", text.encode())
|
|
if rc != 0:
|
|
return False, fname
|
|
return True, ""
|
|
|
|
if action == "pull":
|
|
m = await _read_manifest()
|
|
if m is None:
|
|
return {"ok": False, "error": f"no manifest at {bdir}"}
|
|
return {"ok": True, "dir": bdir, "manifest": m.to_dict()}
|
|
|
|
if action == "push":
|
|
name_ = str(req.get("name") or "").strip()
|
|
if not name_:
|
|
return {"ok": False, "error": "manifest push needs a name"}
|
|
m = mf.Manifest(
|
|
name=name_, purpose=str(req.get("purpose", "")),
|
|
created_by=self.name,
|
|
vm={"engine": req.get("engine") or engine,
|
|
"image": req.get("image", ""),
|
|
"state_ref": req.get("state_ref", ""),
|
|
"entrypoint": req.get("entrypoint", "")},
|
|
setup=list(req.get("setup") or []),
|
|
usage=list(req.get("usage") or []),
|
|
goals={"objective": req.get("objective", ""),
|
|
"user_intent": req.get("user_intent", ""),
|
|
"stop_conditions": list(req.get("stop") or [])})
|
|
m.add_provenance(by=self.name, action="init", note=f"created {m.name}")
|
|
ok, bad = await _write_bundle(m)
|
|
if not ok:
|
|
return {"ok": False, "error": f"failed writing {bad} into sandbox"}
|
|
await self._emit("manifest", action="push", id=m.id, dir=bdir)
|
|
return {"ok": True, "id": m.id, "dir": bdir,
|
|
"files": list(mf.bundle_files(m))}
|
|
|
|
if action == "update":
|
|
m = await _read_manifest()
|
|
if m is None:
|
|
return {"ok": False, "error": f"no manifest at {bdir} (push first)"}
|
|
m.update_state(
|
|
status=req.get("status"), progress=req.get("progress"),
|
|
done=req.get("done") or None, todo=req.get("todo") or None,
|
|
blockers=req.get("blockers") or None,
|
|
append=not bool(req.get("replace", False)))
|
|
note = req.get("note")
|
|
if note:
|
|
m.add_provenance(by=self.name, action="update", note=str(note))
|
|
ok, bad = await _write_bundle(m)
|
|
if not ok:
|
|
return {"ok": False, "error": f"failed writing {bad} into sandbox"}
|
|
await self._emit("manifest", action="update", id=m.id, dir=bdir)
|
|
return {"ok": True, "id": m.id, "dir": bdir,
|
|
"status": m.state.get("status")}
|
|
|
|
return {"ok": False, "error": f"unknown manifest action {action!r}"}
|
|
|
|
def _begin_shutdown(self) -> None:
|
|
self._stop.set()
|
|
self.running = False
|
|
if self._ws is not None:
|
|
asyncio.create_task(self._ws.close())
|