0798db7941
The operator brain was Claude-only: spawn launched the claude CLI and let its own loop drive the room. This adds the other first-class brain — a native tool-calling loop (cmd_chat/operator/harness.py) that runs ANY function-calling Provider (cmd_chat.ai) against an OPERATOR_TOOLS schema wired to the existing bridge control verbs (say/exec/write/get/keys/screen/watch/manifest). No CLI install, no creds carry. The harness is a driver, not a new side-effect surface: every tool handler sends the same control-socket request the hh-bridge CLI already sends, so the bridge keeps enforcing grant-before-drive, the sandbox blast radius, and the recursion budget. Layer-1 capabilities come from CAPABILITIES.md (the same portable contract), and a per-agent token ceiling is enforced as a hard stop. Providers that can't function-call are refused with a clear message rather than degrading to prose. New `operate` verb joins via the existing daemon socket and runs the loop with a --profile or --provider/--model brain. Chat's native loop is left untouched (self-contained harness over the shared Provider core, not a risky extraction). Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
450 lines
20 KiB
Python
450 lines
20 KiB
Python
"""Harness-mode operator (Phase 2) — run *any* function-calling model as a
|
|
hack-house operator.
|
|
|
|
The CLI-subprocess operator (``bootstrap`` runner registry) spawns a vendor
|
|
agent CLI and lets *its* own loop drive the room. This module is the other
|
|
first-class brain: it runs the native tool-calling loop ourselves, against the
|
|
shared Provider core (``cmd_chat.ai``), with the **operator toolset** wired to
|
|
the bridge daemon's control verbs. So a raw Ollama / OpenAI-compatible model —
|
|
no CLI install, no creds carry — becomes a first-class operator.
|
|
|
|
It is deliberately a *driver*, not a new side-effect surface: every tool handler
|
|
just sends the same control-socket request the `hh-bridge` CLI already sends
|
|
(``say``/``exec``/``write``/``get``/``keys``/``screen``/``watch``/``manifest``/
|
|
``spawn``). The bridge keeps enforcing grant-before-drive, the sandbox blast
|
|
radius, and the recursion budget — the harness only chooses *which* verb to call.
|
|
|
|
The read→think→act loop owns the **read**: it polls room events between turns and
|
|
feeds them to the model as observations, so the model spends its turns deciding
|
|
and acting rather than remembering to poll. Capabilities (Layer-1) come straight
|
|
from ``CAPABILITIES.md`` via :func:`bootstrap.load_capabilities` — the same
|
|
portable contract the Claude skill loads — so harness and CLI operators share one
|
|
source of truth.
|
|
"""
|
|
|
|
from __future__ import annotations
|
|
|
|
import json
|
|
import re
|
|
from dataclasses import dataclass, field
|
|
from typing import Callable
|
|
|
|
from cmd_chat.ai.providers import ToolsUnsupported
|
|
|
|
from . import bootstrap as boot
|
|
from .cli_client import request as _socket_request
|
|
|
|
# The operator toolset — each maps 1:1 to a bridge control verb the daemon
|
|
# already implements. Schema is the Ollama/OpenAI `tools` shape (same as the chat
|
|
# agent's NATIVE_TOOLS), so any provider that does function calling can drive it.
|
|
OPERATOR_TOOLS = [
|
|
{
|
|
"type": "function",
|
|
"function": {
|
|
"name": "say",
|
|
"description": "Speak one line into the room. Report what you are doing "
|
|
"and what you found — the room is the coordination bus.",
|
|
"parameters": {
|
|
"type": "object",
|
|
"properties": {
|
|
"text": {"type": "string", "description": "The chat line to send."},
|
|
},
|
|
"required": ["text"],
|
|
},
|
|
},
|
|
},
|
|
{
|
|
"type": "function",
|
|
"function": {
|
|
"name": "exec",
|
|
"description": "Run a shell command in the shared sandbox out-of-band "
|
|
"(does not disturb the live terminal). Returns combined "
|
|
"stdout+stderr and the exit code.",
|
|
"parameters": {
|
|
"type": "object",
|
|
"properties": {
|
|
"command": {"type": "string", "description": "The shell command to run."},
|
|
},
|
|
"required": ["command"],
|
|
},
|
|
},
|
|
},
|
|
{
|
|
"type": "function",
|
|
"function": {
|
|
"name": "write_file",
|
|
"description": "Create or overwrite a file in the sandbox with exact content.",
|
|
"parameters": {
|
|
"type": "object",
|
|
"properties": {
|
|
"path": {"type": "string", "description": "Destination file path."},
|
|
"content": {"type": "string", "description": "Full file content."},
|
|
},
|
|
"required": ["path", "content"],
|
|
},
|
|
},
|
|
},
|
|
{
|
|
"type": "function",
|
|
"function": {
|
|
"name": "get_file",
|
|
"description": "Read and return the contents of a file in the sandbox.",
|
|
"parameters": {
|
|
"type": "object",
|
|
"properties": {
|
|
"path": {"type": "string", "description": "File path to read."},
|
|
},
|
|
"required": ["path"],
|
|
},
|
|
},
|
|
},
|
|
{
|
|
"type": "function",
|
|
"function": {
|
|
"name": "keys",
|
|
"description": "Inject keystrokes into the room's shared live terminal "
|
|
"(drive the PTY). Use for interactive programs; prefer "
|
|
"exec for one-shot commands.",
|
|
"parameters": {
|
|
"type": "object",
|
|
"properties": {
|
|
"text": {"type": "string", "description": "Literal text to type."},
|
|
"enter": {"type": "boolean",
|
|
"description": "Press Enter after the text (default true)."},
|
|
},
|
|
"required": ["text"],
|
|
},
|
|
},
|
|
},
|
|
{
|
|
"type": "function",
|
|
"function": {
|
|
"name": "screen",
|
|
"description": "Print the relayed shared-sandbox terminal buffer "
|
|
"(what everyone watching the PTY sees).",
|
|
"parameters": {
|
|
"type": "object",
|
|
"properties": {
|
|
"tail": {"type": "integer",
|
|
"description": "Last N bytes only (optional)."},
|
|
},
|
|
},
|
|
},
|
|
},
|
|
{
|
|
"type": "function",
|
|
"function": {
|
|
"name": "watch",
|
|
"description": "Block until a regex fires in room events or the terminal "
|
|
"screen — how you wait on a teammate's report or a build "
|
|
"result instead of sleeping.",
|
|
"parameters": {
|
|
"type": "object",
|
|
"properties": {
|
|
"pattern": {"type": "string", "description": "Regex to match."},
|
|
"source": {"type": "string", "enum": ["events", "screen"],
|
|
"description": "Where to match (default: events)."},
|
|
"timeout": {"type": "number",
|
|
"description": "Give up after this many seconds."},
|
|
},
|
|
"required": ["pattern"],
|
|
},
|
|
},
|
|
},
|
|
{
|
|
"type": "function",
|
|
"function": {
|
|
"name": "manifest",
|
|
"description": "Stamp the sandbox's .hh-agent handoff record so work and "
|
|
"memory hand off to the next agent or human.",
|
|
"parameters": {
|
|
"type": "object",
|
|
"properties": {
|
|
"action": {"type": "string", "enum": ["push", "update"],
|
|
"description": "push a fresh record or update the state."},
|
|
"status": {"type": "string", "description": "Current status line."},
|
|
"progress": {"type": "string", "description": "Progress note."},
|
|
"done": {"type": "string", "description": "A completed item."},
|
|
"todo": {"type": "string", "description": "A next item."},
|
|
},
|
|
"required": ["action"],
|
|
},
|
|
},
|
|
},
|
|
]
|
|
|
|
# Tool names that drive the sandbox/room mutably — used only for messaging; the
|
|
# bridge is the real authority on grant-before-drive.
|
|
_TOOL_NAMES = {(t["function"]["name"]) for t in OPERATOR_TOOLS}
|
|
|
|
# DONE marker (same convention as the chat native loop): a text-only turn whose
|
|
# first line starts with this means the model considers the objective met.
|
|
_DONE_RE = re.compile(r"^\s*DONE:", re.I)
|
|
|
|
DEFAULT_MAX_TURNS = 12
|
|
DEFAULT_TOKEN_CEILING = 20000 # cumulative prompt+completion tokens (hard stop)
|
|
DEFAULT_MAX_NUDGES = 2
|
|
_OBS_CAP = 4096 # bytes of any single observation fed back to the model
|
|
|
|
HARNESS_SYSTEM_TAIL = (
|
|
"\n\nYou act ONLY by calling the provided tools (say/exec/write_file/get_file/"
|
|
"keys/screen/watch/manifest) — never by describing an action in prose. Work in "
|
|
"small steps and read each tool result before the next call. You may read and "
|
|
"`say` immediately, but exec/write_file/get_file/keys only work once the room "
|
|
"owner has granted you drive (the bridge will return an error until then — wait "
|
|
"and watch for the grant). When the objective is fully met, reply with a single "
|
|
"line starting 'DONE:' and a one-sentence summary, and do NOT call another tool."
|
|
)
|
|
|
|
|
|
def compose_system(objective: str, *, stop: list[str] | None = None,
|
|
budget: "boot.Budget | None" = None,
|
|
role_overlay: str | None = None) -> str:
|
|
"""The harness operator's system prompt: portable Layer-1 capabilities +
|
|
objective + stop conditions + (optional) role overlay + recursion budget +
|
|
the tool-only operating rule. One screen of capabilities, behaviour overlaid —
|
|
no Claude-isms (the 'use the skill' line lives only in the Claude path)."""
|
|
stop = stop or ["objective met", "owner says stop/leave", "idle past remit"]
|
|
parts = [boot.load_capabilities()]
|
|
if role_overlay:
|
|
parts.append(role_overlay.strip())
|
|
parts.append(f"OBJECTIVE: {objective}")
|
|
parts.append("Stop conditions (say a short sign-off, then stop when any holds): "
|
|
+ "; ".join(stop) + ".")
|
|
if budget is not None:
|
|
parts.append(
|
|
f"Recursion budget you inherit: depth={budget.depth}, "
|
|
f"fanout={budget.fanout}, cost≈${budget.cost_usd:.2f}. Spawn at most "
|
|
"`fanout` children and only while depth>0; pass each a decremented budget.")
|
|
return "\n\n".join(parts) + HARNESS_SYSTEM_TAIL
|
|
|
|
|
|
def _clip(s: str, cap: int = _OBS_CAP) -> str:
|
|
if len(s) <= cap:
|
|
return s
|
|
return s[:cap] + f"\n…[clipped {len(s) - cap} bytes]"
|
|
|
|
|
|
def _format_events(events: list[dict], me: str) -> str:
|
|
"""Turn raw bridge inbox events into a compact observation block the model
|
|
reads as room state. Drops our own echoes and high-volume noise."""
|
|
lines: list[str] = []
|
|
for ev in events:
|
|
kind = ev.get("kind")
|
|
if kind == "message":
|
|
who = ev.get("from", "?")
|
|
if who == me:
|
|
continue
|
|
mark = " (@you)" if ev.get("addressed") else ""
|
|
lines.append(f"[{who}{mark}] {ev.get('text', '')}")
|
|
elif kind == "acl":
|
|
state = "GRANTED drive" if ev.get("granted") else "drive NOT granted"
|
|
lines.append(f"[room] you are now {state}.")
|
|
elif kind == "sandbox":
|
|
lines.append(f"[room] sandbox {ev.get('state')} "
|
|
f"({ev.get('engine') or '—'}/{ev.get('name') or '—'}).")
|
|
elif kind == "roster":
|
|
lines.append(f"[room] members: {', '.join(ev.get('users', []))}.")
|
|
return "\n".join(lines)
|
|
|
|
|
|
@dataclass
|
|
class HarnessResult:
|
|
final: str = ""
|
|
turns: int = 0
|
|
tokens: int = 0
|
|
reason: str = "" # why the loop ended
|
|
tool_calls: int = 0
|
|
|
|
|
|
@dataclass
|
|
class OperatorHarness:
|
|
"""Drives a room with a function-calling Provider over the bridge control
|
|
socket. Side-effect-free to construct; :meth:`run` does the work. ``request_fn``
|
|
is injectable so unit tests can drive the whole loop with no daemon/network."""
|
|
|
|
provider: object
|
|
sock_path: str
|
|
objective: str
|
|
stop: list[str] | None = None
|
|
budget: "boot.Budget | None" = None
|
|
role_overlay: str | None = None
|
|
max_turns: int = DEFAULT_MAX_TURNS
|
|
token_ceiling: int = DEFAULT_TOKEN_CEILING
|
|
max_nudges: int = DEFAULT_MAX_NUDGES
|
|
me: str = "operator"
|
|
request_fn: Callable[[str, dict, float], dict] | None = None
|
|
poll_timeout: float = 20.0
|
|
|
|
_cursor: int = field(default=0, init=False)
|
|
|
|
# ── control-socket plumbing ──────────────────────────────────────────
|
|
def _call(self, obj: dict, read_timeout: float = 35.0) -> dict:
|
|
fn = self.request_fn or _socket_request
|
|
return fn(self.sock_path, obj, read_timeout)
|
|
|
|
def _drain_events(self, *, wait: bool, timeout: float) -> list[dict]:
|
|
resp = self._call({"op": "read", "since": self._cursor,
|
|
"wait": wait, "timeout": timeout},
|
|
read_timeout=timeout + 5.0)
|
|
events = resp.get("events", []) or []
|
|
if events:
|
|
self._cursor = events[-1].get("seq", self._cursor)
|
|
elif resp.get("seq") is not None:
|
|
self._cursor = max(self._cursor, int(resp["seq"]))
|
|
return events
|
|
|
|
# ── tool dispatch (each maps to a bridge verb) ───────────────────────
|
|
def _run_tool(self, name: str, args: dict) -> str:
|
|
if name == "say":
|
|
r = self._call({"op": "say", "text": str(args.get("text", ""))})
|
|
return "(said)" if r.get("ok") else f"error: {r.get('error')}"
|
|
if name == "exec":
|
|
r = self._call({"op": "exec", "cmd": str(args.get("command", ""))})
|
|
out = _clip(r.get("output", ""))
|
|
return f"exit={r.get('rc')}\n{out}" if not r.get("error") else f"error: {r['error']}"
|
|
if name == "write_file":
|
|
r = self._call({"op": "write", "path": str(args.get("path", "")),
|
|
"content": args.get("content", "")})
|
|
return (f"wrote {r.get('path')} ({r.get('bytes')} bytes)"
|
|
if r.get("ok") else f"error: {r.get('error')}")
|
|
if name == "get_file":
|
|
import base64
|
|
r = self._call({"op": "get", "path": str(args.get("path", ""))})
|
|
if not r.get("ok"):
|
|
return f"error: {r.get('error')}"
|
|
try:
|
|
raw = base64.b64decode(r.get("b64", "")).decode(errors="replace")
|
|
except Exception: # noqa: BLE001
|
|
raw = "[binary]"
|
|
return _clip(raw)
|
|
if name == "keys":
|
|
toks = [f"text:{args.get('text', '')}"]
|
|
if args.get("enter", True):
|
|
toks.append("enter")
|
|
r = self._call({"op": "keys", "keys": toks})
|
|
return "(keys sent)" if r.get("ok") else f"error: {r.get('error')}"
|
|
if name == "screen":
|
|
r = self._call({"op": "screen", "tail": args.get("tail"), "ansi": False})
|
|
return _clip(r.get("text", "")) if r.get("ok") else f"error: {r.get('error')}"
|
|
if name == "watch":
|
|
timeout = float(args.get("timeout", 30.0))
|
|
r = self._call({"op": "watch", "for": args.get("pattern"),
|
|
"in": args.get("source", "events"), "idle": None,
|
|
"timeout": timeout}, read_timeout=timeout + 5.0)
|
|
return _clip(json.dumps(r))
|
|
if name == "manifest":
|
|
req = {"op": "manifest", "action": args.get("action", "update")}
|
|
for k in ("status", "progress"):
|
|
if args.get(k) is not None:
|
|
req[k] = args[k]
|
|
for k in ("done", "todo"):
|
|
if args.get(k):
|
|
req[k] = [args[k]]
|
|
r = self._call(req)
|
|
return _clip(json.dumps(r))
|
|
return f"error: unknown tool '{name}'"
|
|
|
|
@staticmethod
|
|
def _calls_to_wire(calls: list[dict]) -> list[dict]:
|
|
"""Round-trip the model's tool calls back into the wire `tool_calls` shape
|
|
so the next turn sees its own request alongside the result."""
|
|
return [{"type": "function",
|
|
"function": {"name": c.get("name", ""),
|
|
"arguments": c.get("arguments", {})}}
|
|
for c in calls]
|
|
|
|
# ── the loop ──────────────────────────────────────────────────────────
|
|
def run(self) -> HarnessResult:
|
|
cwt = getattr(self.provider, "complete_with_tools", None)
|
|
if cwt is None:
|
|
return HarnessResult(
|
|
reason="provider-no-tools",
|
|
final="[refused — this provider cannot do function calling; the "
|
|
"harness operator needs a tool-capable model]")
|
|
system = compose_system(self.objective, stop=self.stop, budget=self.budget,
|
|
role_overlay=self.role_overlay)
|
|
|
|
# Seed: snapshot current room state (full ring) as the first observation.
|
|
try:
|
|
seed = self._drain_events(wait=False, timeout=0.0)
|
|
except Exception as e: # noqa: BLE001
|
|
return HarnessResult(reason="bridge-unreachable",
|
|
final=f"[bridge unreachable: {e}]")
|
|
obs = _format_events(seed, self.me) or "(room is quiet)"
|
|
messages: list[dict] = [
|
|
{"role": "user",
|
|
"content": f"You have joined the room. Current state:\n{obs}\n\n"
|
|
f"Begin working toward the objective."}
|
|
]
|
|
|
|
res = HarnessResult()
|
|
nudges = 0
|
|
while res.turns < self.max_turns + self.max_nudges:
|
|
res.turns += 1
|
|
try:
|
|
text, calls, usage = cwt(system, messages, OPERATOR_TOOLS)
|
|
except ToolsUnsupported as e:
|
|
return HarnessResult(
|
|
turns=res.turns, tokens=res.tokens, reason="tools-unsupported",
|
|
final=f"[refused — model rejected tool calling: {e}]")
|
|
except Exception as e: # noqa: BLE001
|
|
res.reason, res.final = "provider-error", f"[ai error: {e}]"
|
|
return res
|
|
|
|
res.tokens += int((usage or {}).get("prompt_eval_count", 0) or 0)
|
|
res.tokens += int((usage or {}).get("eval_count", 0) or 0)
|
|
|
|
if not calls:
|
|
if _DONE_RE.match(text or ""):
|
|
res.final = re.sub(_DONE_RE, "", text, count=1).strip()
|
|
res.reason = "done"
|
|
return res
|
|
if nudges >= self.max_nudges:
|
|
res.final = (text or "").strip() or "[stopped — model stalled]"
|
|
res.reason = "stalled"
|
|
return res
|
|
nudges += 1
|
|
messages.append({"role": "assistant", "content": text or ""})
|
|
# Poll for any room activity to give the nudge fresh context.
|
|
fresh = []
|
|
try:
|
|
fresh = self._drain_events(wait=True, timeout=self.poll_timeout)
|
|
except Exception: # noqa: BLE001
|
|
pass
|
|
nudge_obs = _format_events(fresh, self.me)
|
|
messages.append({
|
|
"role": "user",
|
|
"content": (("Room update:\n" + nudge_obs + "\n\n") if nudge_obs else "")
|
|
+ "If the objective is met, reply with a single 'DONE: <summary>' "
|
|
"line. Otherwise call the next tool to make progress."})
|
|
continue
|
|
|
|
if res.tokens >= self.token_ceiling:
|
|
res.final = ("[stopped — per-agent token ceiling "
|
|
f"{self.token_ceiling} reached]")
|
|
res.reason = "token-ceiling"
|
|
return res
|
|
|
|
# Echo the assistant's tool-call turn, then run each call and feed the
|
|
# results back as tool messages so the next turn sees the outcome.
|
|
messages.append({"role": "assistant", "content": text or "",
|
|
"tool_calls": self._calls_to_wire(calls)})
|
|
for c in calls:
|
|
name, cargs = c.get("name", ""), c.get("arguments") or {}
|
|
result = self._run_tool(name, cargs)
|
|
res.tool_calls += 1
|
|
messages.append({"role": "tool", "content": _clip(result)})
|
|
# After acting, fold any new room events into the next observation.
|
|
try:
|
|
fresh = self._drain_events(wait=False, timeout=0.0)
|
|
except Exception: # noqa: BLE001
|
|
fresh = []
|
|
obs = _format_events(fresh, self.me)
|
|
if obs:
|
|
messages.append({"role": "user", "content": "Room update:\n" + obs})
|
|
|
|
res.final = res.final or "[stopped — reached turn cap]"
|
|
res.reason = res.reason or "turn-cap"
|
|
return res
|