Files
hack-house/hh/scripts/bench-native-harness.py
leetcrypt b998d28ebf fix(ai): repair native-harness bench — 3-tuple preflight + PTY-mirror stub
First functional bench run crashed 0/3 on two latent bench-only bugs:
(1) the Phase-4 token-count change made complete_with_tools return
(text, calls, usage), but the bench preflight still unpacked 2 values;
(2) make_bridge never stubbed _send_sbx_input, so the PTY-mirror
visibility feature hit ws=None and raised AttributeError on every tool
call. Stub it to capture mirrored lines for --verbose.

After the fix: 2/3 pass (write+read, script+run). mkdir+list still
fails — local-backend run_shell CWD divergence + 3B model churn, not a
working-memory regression.

Sprint: native-harness-working-memory (bench) — see also 46e5620, dc6317f, c83abbe

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
2026-06-27 10:16:30 -07:00

218 lines
7.8 KiB
Python

#!/usr/bin/env python3
"""Live bench/smoke for the native `!task` harness (docs/spec-native-harness.md, Phase 3).
Drives the real `AgentBridge._run_native` against a live Ollama model and the
`local` sandbox backend (host shell, scoped to a fresh temp workdir), with no chat
server or TUI in the loop. For each canonical task it reports: wall-clock latency,
model turns, tool calls, whether the expected artifact landed, and the final
answer. Also records a single chat-completion latency as the `simple`-harness
model-cost baseline (simple's execution needs the broker PTY, so only its model
call is comparable headlessly).
Usage: .venv/bin/python hh/scripts/bench-native-harness.py [--model qwen2.5:3b]
[--max-turns 5] [--threads 4]
Env: OLLAMA_HOST (default http://localhost:11434)
"""
from __future__ import annotations
import argparse
import asyncio
import os
import sys
import tempfile
import time
from pathlib import Path
# Make the repo importable when run from anywhere.
ROOT = Path(__file__).resolve().parents[2]
sys.path.insert(0, str(ROOT))
from cmd_chat.agent.bridge import AgentBridge # noqa: E402
from cmd_chat.agent.providers import ( # noqa: E402
Msg,
OllamaProvider,
ToolsUnsupported,
)
PER_TASK_TIMEOUT = 300.0 # hard ceiling so a runaway loop can't hang the bench
# (label, task text, check(workdir) -> bool)
TASKS = [
(
"write+read",
"create a file hello.txt containing exactly the text 'hello world', "
"then show its contents",
lambda d: (d / "hello.txt").is_file()
and "hello world" in (d / "hello.txt").read_text(),
),
(
"script+run",
"write a python script add.py that prints the sum of 2 and 3, then run it",
lambda d: (d / "add.py").is_file(),
),
(
"mkdir+list",
"make a directory named data and then list the files in the current directory",
lambda d: (d / "data").is_dir(),
),
]
class CountingOllama(OllamaProvider):
"""OllamaProvider that tallies tool-calling turns for the report."""
def __init__(self, *a, **k):
super().__init__(*a, **k)
self.tool_turns = 0
def complete_with_tools(self, system, messages, tools):
self.tool_turns += 1
return super().complete_with_tools(system, messages, tools)
def make_bridge(provider, workdir: str):
"""A headless AgentBridge wired to the local backend, with all network sends
stubbed to capture room output instead of encrypting to a websocket."""
b = AgentBridge(
"localhost", 0, name="bench", provider=provider,
no_tls=True, code_provider=provider, harness="native",
max_turns=provider_max_turns,
)
b.granted = True
b.sbx_engine = "local" # exec on the host shell (this process's CWD = workdir)
b.sbx_name = ""
b._chat: list[str] = []
async def cap_chat(ws, text):
b._chat.append(text)
async def noop(*a, **k):
pass
b._send_chat = cap_chat
b._send_typing = noop
b._send_stream = noop
async def cap_inject(ws, cmds): # simple-harness path (unused here, kept honest)
b._chat.append("[inject] " + " ; ".join(cmds))
b._inject = cap_inject
async def cap_sbx_input(ws, data): # PTY-mirror frames — broker-only in prod
b._chat.append("[pty] " + data.decode(errors="replace").rstrip("\n"))
b._send_sbx_input = cap_sbx_input
return b
provider_max_turns = 5 # set in main()
async def run_task(provider, label, task, check) -> dict:
workdir = Path(tempfile.mkdtemp(prefix=f"hh-bench-{label}-"))
cwd = os.getcwd()
os.chdir(workdir)
provider.tool_turns = 0
bridge = make_bridge(provider, str(workdir))
t0 = time.monotonic()
timed_out = False
err = None
try:
await asyncio.wait_for(
bridge._run_native(None, task, "andre"), timeout=PER_TASK_TIMEOUT)
except asyncio.TimeoutError:
timed_out = True
except Exception as e: # noqa: BLE001 — record, don't abort the suite
err = f"{type(e).__name__}: {e}"
finally:
os.chdir(cwd)
elapsed = time.monotonic() - t0
ok = False
try:
ok = bool(check(workdir)) and not timed_out and err is None
except Exception: # noqa: BLE001
ok = False
final = next((c for c in reversed(bridge._chat) if "(native) for" in c), "")
return {
"label": label, "ok": ok, "elapsed": elapsed, "turns": provider.tool_turns,
"timed_out": timed_out, "err": err, "workdir": str(workdir),
"chat": bridge._chat, "final": final,
}
async def main() -> int:
global provider_max_turns
ap = argparse.ArgumentParser()
ap.add_argument("--model", default="qwen2.5:3b")
ap.add_argument("--max-turns", type=int, default=5)
ap.add_argument("--threads", type=int, default=4)
ap.add_argument("--num-ctx", type=int, default=4096)
ap.add_argument("--verbose", action="store_true", help="print full chat per task")
args = ap.parse_args()
provider_max_turns = args.max_turns
host = os.environ.get("OLLAMA_HOST", "http://localhost:11434")
print(f"== native harness bench == model={args.model} host={host}")
print(f" max_turns={args.max_turns} threads={args.threads} num_ctx={args.num_ctx}\n")
provider = CountingOllama(
model=args.model, num_ctx=args.num_ctx, num_thread=args.threads, num_predict=512)
# 0. Wire preflight: does the model accept the `tools` field at all?
print("[preflight] probing tool support…", flush=True)
t0 = time.monotonic()
try:
text, calls, _usage = await asyncio.to_thread(
provider.complete_with_tools,
"You are a test.",
[{"role": "user", "content": "Call run_shell to echo hi."}],
__import__("cmd_chat.agent.bridge", fromlist=["NATIVE_TOOLS"]).NATIVE_TOOLS,
)
except ToolsUnsupported as e:
print(f" ✖ model rejects tools: {e}\n → native would degrade to simple. Stopping.")
return 1
except Exception as e: # noqa: BLE001
print(f" ✖ preflight error: {type(e).__name__}: {e}")
return 2
print(f" ✓ tools accepted in {time.monotonic()-t0:.1f}s "
f"(calls={len(calls)}, supports_tools={provider.supports_tools()})\n")
# Baseline: one plain chat completion (the only model cost simple pays).
t0 = time.monotonic()
try:
await asyncio.to_thread(provider.complete, "You are concise.",
[Msg("user", "say ok")])
base = time.monotonic() - t0
print(f"[baseline] one chat completion (≈ simple's model cost): {base:.1f}s\n")
except Exception as e: # noqa: BLE001
print(f"[baseline] chat completion failed: {e}\n")
results = []
for label, task, check in TASKS:
print(f"[task:{label}] {task}", flush=True)
r = await run_task(provider, label, task, check)
tag = "PASS" if r["ok"] else ("TIMEOUT" if r["timed_out"] else "FAIL")
print(f" → {tag} {r['elapsed']:.1f}s turns={r['turns']}"
+ (f" err={r['err']}" if r["err"] else ""))
if r["final"]:
print(f" final: {r['final'].splitlines()[-1][:160]}")
if args.verbose:
for c in r["chat"]:
print(" | " + c.replace("\n", " ")[:160])
print(flush=True)
results.append(r)
# Summary table.
print("== summary ==")
print(f"{'task':<14}{'result':<9}{'secs':>7}{'turns':>7}")
for r in results:
tag = "PASS" if r["ok"] else ("TIMEOUT" if r["timed_out"] else "FAIL")
print(f"{r['label']:<14}{tag:<9}{r['elapsed']:>7.1f}{r['turns']:>7}")
passes = sum(1 for r in results if r["ok"])
print(f"\n{passes}/{len(results)} passed")
return 0 if passes == len(results) else 3
if __name__ == "__main__":
raise SystemExit(asyncio.run(main()))