Port Google Research's Harness^2 method layer (model weights frozen, only the surrounding harness is optimized): Propose-Probe-Compose with INV/DOM/AGG editors, parallel/sequential modes, per-domain experience library, anonymous-product selection, k<=3 hard cap and budget gate. Execution goes only through the existing control-plane API (upload -> task with idempotency key -> poll -> trace/cost -> cancel on timeout); no new server endpoints, no codex flags composed, tasks inherit the normal executor/sandbox/quota/audit path. Wiring lessons from four real rounds against production (project wwww, ledger task): HttpTaskClient retries transient disconnects and 502/503/429 (3 attempts); editors get a worked file-block example, an explicit <no-change> marker and one feedback retry (.fmt2) instead of silently accepting notes-only replies; run_task unwraps the server's result_text JSON envelope (text field holds the real reply) — without this, line-anchored file-block parsing could never see blocks the agents were actually emitting. Final round (real4) fully green: all stages landed, selection non-fallback, experience入库, probe delivers the correct signed total 85.15. Tests: 35 harness tests (17 method + 18 wiring incl. fake-control-plane full round, idempotent replay, retry, format retry, envelope round). Deliberate deviations from upstream are documented in docs/harness2.md (no network-unrestricted/sandbox-disabled defaults, auto_apply off, no benchmark/scorer, probe deliverable = final reply text).
139 lines
5.6 KiB
Python
139 lines
5.6 KiB
Python
"""Harness² 融合演示:不花模型钱,跑通一轮完整改进。
|
|
|
|
用法(在项目根目录):
|
|
python3 examples/harness2_demo.py --mode parallel --k 1
|
|
python3 examples/harness2_demo.py --mode sequential --k 1 --output /tmp/h2demo
|
|
|
|
任务是对账本求和。base harness 的脚本只加正数(账本含一笔负数 credit,
|
|
base 会漏掉它);确定性编辑器去掉正数过滤;选择器按产物挑带符号和;
|
|
反思器把“正负都要算”写进经验库。跑完打印:被采纳的产物、回落情况、
|
|
经验库条目位置。真实接线时,这里的三个确定性类换成 codex 任务
|
|
(见 docs/harness2.md),执行器换成 Worker 的任务执行,循环本身不变。
|
|
"""
|
|
|
|
import argparse
|
|
import csv
|
|
import subprocess
|
|
import sys
|
|
import tempfile
|
|
from pathlib import Path
|
|
|
|
sys.path.insert(0, str(Path(__file__).resolve().parents[1]))
|
|
|
|
from cloud.harness2 import ( # noqa: E402
|
|
EditResult, Library, Rollout, improve_task, read_tree, write_tree,
|
|
)
|
|
|
|
BASE_SCRIPT = """import csv
|
|
import sys
|
|
from decimal import Decimal
|
|
with open(sys.argv[1]) as f:
|
|
amounts = [Decimal(r['amount']) for r in csv.DictReader(f)]
|
|
print(f"Total: {sum((a for a in amounts if a > 0), Decimal(0)):.2f}")
|
|
"""
|
|
|
|
BASE_TREE = {
|
|
"AGENTS.md": "Use scripts/summarize.py to total the ledger.\n",
|
|
"scripts/summarize.py": BASE_SCRIPT,
|
|
}
|
|
|
|
|
|
class LedgerAdapter:
|
|
task_id = "ledger-1"
|
|
domain = "accounting"
|
|
prompt = "Sum every signed amount in ledger.csv. Return Total: with two decimals."
|
|
|
|
def __init__(self, root):
|
|
self.root = Path(root)
|
|
self.root.mkdir(parents=True, exist_ok=True)
|
|
self.ledger = self.root / "ledger.csv"
|
|
self.ledger.write_text("amount\n100.25\n-20.10\n5.00\n", encoding="utf-8")
|
|
self.executions = 0
|
|
|
|
def base_harness(self):
|
|
return dict(BASE_TREE)
|
|
|
|
def execute(self, label, harness):
|
|
self.executions += 1
|
|
work = self.root / "runs" / label
|
|
work.mkdir(parents=True, exist_ok=True)
|
|
write_tree(harness, work / "harness")
|
|
proc = subprocess.run(
|
|
[sys.executable, str(work / "harness" / "scripts" / "summarize.py"),
|
|
str(self.ledger)], capture_output=True, text=True, timeout=10)
|
|
out = proc.stdout.strip()
|
|
return Rollout(label=label, ok=proc.returncode == 0,
|
|
delivered=out.startswith("Total:"), deliverable_text=out,
|
|
trajectory="[ACTION] Run scripts/summarize.py\n[OBSERVATION]\n" + out,
|
|
cost_usd=0.01)
|
|
|
|
def validate_tree(self, harness):
|
|
if "scripts/summarize.py" not in harness:
|
|
raise ValueError("The local executor requires scripts/summarize.py")
|
|
|
|
|
|
class FixEditor:
|
|
def edit(self, mission, context_dir, label):
|
|
harness_dir = Path(context_dir) / "harness"
|
|
script = harness_dir / "scripts" / "summarize.py"
|
|
script.write_text(script.read_text().replace(" if a > 0", ""), encoding="utf-8")
|
|
return EditResult(tree=read_tree(harness_dir),
|
|
notes="Keep signed amounts: the base total omitted credits.")
|
|
|
|
|
|
class PickSignedSelector:
|
|
def select(self, mission, context_dir):
|
|
for path in sorted((Path(context_dir) / "candidates").glob("*.md")):
|
|
if "Total: 85.15" in path.read_text():
|
|
return path.stem, "The signed total matches the source ledger.", 0.0
|
|
raise ValueError("no candidate with the signed total")
|
|
|
|
|
|
class LedgerReflector:
|
|
def reflect(self, mission, context_dir):
|
|
root = Path(context_dir)
|
|
changed = (root / "harness" / "scripts" / "summarize.py").read_text() != (
|
|
root / "base_harness" / "scripts" / "summarize.py").read_text()
|
|
if not changed:
|
|
return {}
|
|
return {
|
|
"cue": "When a ledger includes credits and debits",
|
|
"lesson": "Include both signs when summing in scripts/summarize.py.",
|
|
"evidence": "Removing the positive-only filter changed the probe total to "
|
|
"include the credits omitted by the base trajectory.",
|
|
"confidence": "high for this observed arithmetic contrast",
|
|
}
|
|
|
|
|
|
def main():
|
|
ap = argparse.ArgumentParser()
|
|
ap.add_argument("--mode", default="parallel", choices=["parallel", "sequential"])
|
|
ap.add_argument("--k", type=int, default=1)
|
|
ap.add_argument("--output", default="")
|
|
args = ap.parse_args()
|
|
tmp = None
|
|
root = Path(args.output) if args.output else Path(tempfile.mkdtemp(prefix="h2demo-"))
|
|
if not args.output:
|
|
tmp = root
|
|
adapter = LedgerAdapter(root)
|
|
library = Library(root / "library", "accounting")
|
|
result = improve_task(adapter, FixEditor(), selector=PickSignedSelector(),
|
|
reflector=LedgerReflector(), work_dir=root / "work",
|
|
k=args.k, mode=args.mode, library=library)
|
|
chosen = result.mode_result.rollouts[result.selection.chosen_label]
|
|
print(f"mode={result.mode} k={args.k} ok={result.ok}")
|
|
print(f"executions={adapter.executions} (base + candidates + finals)")
|
|
print(f"chosen={result.selection.chosen_label} deliverable={chosen.deliverable_text}")
|
|
print(f"fell_back={result.mode_result.fell_back or 'none'}")
|
|
print(f"library_entry={result.library_entry}")
|
|
print(f"experience={result.experience.get('lesson', '(none)')}")
|
|
print(f"work_dir={root}")
|
|
assert chosen.deliverable_text == "Total: 85.15", "demo must end at the signed total"
|
|
assert result.library_entry is not None, "demo must enter the library"
|
|
print("demo OK: signed total adopted and stored in the library")
|
|
_ = tmp
|
|
|
|
|
|
if __name__ == "__main__":
|
|
main()
|