host
42abe3cc5b
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).
134 lines
6.2 KiB
Python
134 lines
6.2 KiB
Python
"""用真实 codex 任务跑一轮 Harness² 改进（接线层 CLI）。
|
||
|
||
这会创建真实任务、花真实模型配额——所以默认只打印计划，不建任务：
|
||
先跑一遍不带 --yes 看计划，确认项目、任务数与预算无误后，再带
|
||
--yes 重跑。第一次请务必用测试项目，不要直接指生产项目。
|
||
|
||
密钥不走命令行参数（ps 可见）：用环境变量 CODEX_API_KEY。
|
||
|
||
示例：
|
||
export CODEX_API_KEY='...'
|
||
python3 examples/harness2_run.py --base-url http://127.0.0.1:8765 \\
|
||
--project-id <test-project> --prompt-file task.md \\
|
||
--base-harness-dir ./my-base-harness --source-dir ./task-source \\
|
||
--mode parallel --k 1 --budget 5 --output runs/h2-real-1
|
||
# 看计划无误后，加 --yes 重跑（同一 run_tag 可断点续跑；想从头
|
||
# 重跑就换 --run-tag 或不传，让它自动生成新的时间戳 tag）。
|
||
|
||
base-harness-dir：一个目录，里面就是 harness 树（AGENTS.md、
|
||
.agents/、scripts/、.codex/……），按原样读入；不存在或 AGENTS.md
|
||
为空时直接报错，不跑。
|
||
source-dir（可选）：原任务需要的源文件，会随 harness 一起 seed 进
|
||
每个探针任务的 worktree；与 harness 路径重名时 harness 优先。
|
||
"""
|
||
|
||
import argparse
|
||
import os
|
||
import sys
|
||
import tempfile
|
||
import time
|
||
from pathlib import Path
|
||
|
||
sys.path.insert(0, str(Path(__file__).resolve().parents[1]))
|
||
|
||
from cloud.harness2 import read_tree # noqa: E402
|
||
from cloud.harness2_tasks import ( # noqa: E402
|
||
HttpTaskClient, run_improvement_via_tasks,
|
||
)
|
||
|
||
|
||
def _flatten(root: Path) -> dict[str, bytes]:
|
||
out: dict[str, bytes] = {}
|
||
if not root or not root.is_dir():
|
||
return out
|
||
for p in sorted(root.rglob("*")):
|
||
if p.is_file() and not p.is_symlink() and "__pycache__" not in p.parts:
|
||
out[str(p.relative_to(root))] = p.read_bytes()
|
||
return out
|
||
|
||
|
||
def main():
|
||
ap = argparse.ArgumentParser()
|
||
ap.add_argument("--base-url", required=True, help="控制面地址，如 http://127.0.0.1:8765")
|
||
ap.add_argument("--project-id", required=True)
|
||
ap.add_argument("--prompt", default="")
|
||
ap.add_argument("--prompt-file", default="")
|
||
ap.add_argument("--task-id", default="h2-task-1")
|
||
ap.add_argument("--domain", default="general")
|
||
ap.add_argument("--base-harness-dir", required=True)
|
||
ap.add_argument("--source-dir", default="")
|
||
ap.add_argument("--mode", default="parallel", choices=["parallel", "sequential"])
|
||
ap.add_argument("--k", type=int, default=1)
|
||
ap.add_argument("--budget", type=float, default=None)
|
||
ap.add_argument("--run-tag", default="")
|
||
ap.add_argument("--output", default="")
|
||
ap.add_argument("--library-root", default="")
|
||
ap.add_argument("--model", default=None)
|
||
ap.add_argument("--effort", default=None)
|
||
ap.add_argument("--approval", default=None)
|
||
ap.add_argument("--environment-id", default=None)
|
||
ap.add_argument("--executor", default=None)
|
||
ap.add_argument("--no-time-limit", action="store_true")
|
||
ap.add_argument("--timeout", type=int, default=3600)
|
||
ap.add_argument("--yes", action="store_true", help="确认创建真实任务并消费配额")
|
||
args = ap.parse_args()
|
||
|
||
api_key = os.environ.get("CODEX_API_KEY")
|
||
if not api_key:
|
||
print("缺少 CODEX_API_KEY 环境变量（密钥不接受命令行参数）", file=sys.stderr)
|
||
return 2
|
||
prompt = args.prompt
|
||
if args.prompt_file:
|
||
prompt = Path(args.prompt_file).read_text(encoding="utf-8")
|
||
if not prompt.strip():
|
||
print("需要 --prompt 或 --prompt-file", file=sys.stderr)
|
||
return 2
|
||
base = read_tree(args.base_harness_dir)
|
||
if not base.get("AGENTS.md", "").strip():
|
||
print(f"base-harness-dir 里没有可用的 AGENTS.md：{args.base_harness_dir}",
|
||
file=sys.stderr)
|
||
return 2
|
||
run_tag = args.run_tag or time.strftime("h2-%Y%m%d-%H%M%S")
|
||
# 任务数（k=1 parallel）：1 base + 2 候选 + 1 终版探针 + 3 编辑
|
||
# + 1 选择 + 1 反思 = 9；sequential 同为 9；k 越大按 3k+? 线性涨，
|
||
# 上限 k=3 时 parallel 为 1 + 6 + 3 探针 + 9 编辑 + 1 选择 + 1 反思 = 21。
|
||
n_probes = 1 + 3 * args.k # 两种模式同款：1 base + 每轮 2 候选 + 1 终版
|
||
n_edits = 3 * args.k
|
||
plan = (f"计划：项目 {args.project_id}，任务 {args.task_id}，"
|
||
f"mode={args.mode} k={args.k}，run_tag={run_tag}，"
|
||
f"约 {n_probes} 个探针 + {n_edits} 个编辑 + 1 选择 + 1 反思 = "
|
||
f"{n_probes + n_edits + 2} 个真实任务，"
|
||
f"budget_usd={args.budget}，base 文件 {len(base)} 个。")
|
||
print(plan)
|
||
if not args.yes:
|
||
print("未带 --yes：只打印计划，不创建任何任务。确认无误后加 --yes 重跑。")
|
||
return 0
|
||
work_dir = Path(args.output) if args.output else Path(tempfile.mkdtemp(prefix="h2run-"))
|
||
client = HttpTaskClient(args.base_url, api_key)
|
||
result = run_improvement_via_tasks(
|
||
client, project_id=args.project_id, task_id=args.task_id, prompt=prompt,
|
||
base_harness=base, run_tag=run_tag, work_dir=work_dir, domain=args.domain,
|
||
source_files=_flatten(Path(args.source_dir)) if args.source_dir else None,
|
||
k=args.k, mode=args.mode,
|
||
library_root=Path(args.library_root) if args.library_root else None,
|
||
budget_usd=args.budget, model=args.model, effort=args.effort,
|
||
approval=args.approval, environment_id=args.environment_id,
|
||
executor=args.executor, no_time_limit=args.no_time_limit,
|
||
timeout_seconds=args.timeout)
|
||
d = result.to_dict()
|
||
chosen = result.mode_result.rollouts.get(result.selection.chosen_label)
|
||
print(f"ok={result.ok} chosen={result.selection.chosen_label} "
|
||
f"fallback={result.selection.fallback} "
|
||
f"fell_back={result.mode_result.fell_back or 'none'} "
|
||
f"budget_exceeded={result.budget_exceeded}")
|
||
print(f"deliverable={(chosen.deliverable_text if chosen else '')[:500]}")
|
||
print(f"total_cost_usd={result.mode_result.total_cost_usd} "
|
||
f"library_entry={result.library_entry}")
|
||
print(f"work_dir={work_dir}")
|
||
print(f"full_result={d}")
|
||
return 0 if result.ok else 1
|
||
|
||
|
||
if __name__ == "__main__":
|
||
raise SystemExit(main())
|