dd3xp
362b7d601c
Content-overlay update of all GA framework files from upstream baseb4393af(2026-05-18) to1eb8c4c(2026-05-25). Applied as an overlay because GADesktop and upstream share no common git history. - 26 framework files brought to upstream: tui_v3 (new), fsapp rework, llmcore hardening, plugins/hooks lifecycle hooks, goal_mode prompt rewrite, supergrok_proxy, qq/tg/wechat/st apps, cost_tracker py3.9 compat, docs; removed obsolete install scripts. - Desktop frontend (frontends/desktop/*) and desktop_bridge.py left untouched. Upstream's only overlapping change (bridge mykey.txt-> mykey.py) is already present in our version, so it is a no-op.
104 lines
3.6 KiB
Python
104 lines
3.6 KiB
Python
# reflect/goal_mode.py — Goal Mode: 持续自驱直到预算耗尽
|
|
# 启动: set GOAL_STATE=temp/xxx.json && python agentmain.py --reflect reflect/goal_mode.py
|
|
# 配置: agent按SOP写好state json，通过环境变量GOAL_STATE指定路径
|
|
import os, json, time
|
|
|
|
INTERVAL = 3 # check间隔短，agent跑完立刻再检查
|
|
ONCE = False
|
|
|
|
_dir = os.path.dirname(os.path.abspath(__file__))
|
|
STATE_FILE = ''
|
|
def init(a):
|
|
global STATE_FILE
|
|
STATE_FILE = a.get('goal_state') or os.environ.get('GOAL_STATE') or os.path.join(_dir, '../temp/goal_state.json')
|
|
if not os.path.isabs(STATE_FILE): STATE_FILE = os.path.join(_dir, '..', STATE_FILE)
|
|
# --- state 管理 ---
|
|
def _load():
|
|
if not os.path.isfile(STATE_FILE): return None
|
|
with open(STATE_FILE, 'r', encoding='utf-8') as f:
|
|
return json.load(f)
|
|
|
|
def _save(state):
|
|
with open(STATE_FILE, 'w', encoding='utf-8') as f:
|
|
json.dump(state, f, ensure_ascii=False, indent=2)
|
|
|
|
# --- prompt 模板 ---
|
|
CONTINUATION_PROMPT = """[Goal Mode — 持续推进]
|
|
|
|
<objective>
|
|
{objective}
|
|
</objective>
|
|
|
|
⏱ 已用 {elapsed_min:.0f} 分钟，剩余约 {remaining_min:.0f} 分钟。第 {turn} 次唤醒。
|
|
|
|
你正在 Goal Mode 下工作：
|
|
1. 禁止说"已完成，是否继续"——预算没到就不准停，不准提前交付，围绕核心成果质量不要额外产出。
|
|
2. 在 cwd 下建立工作文件夹存放成果和进度，读取记忆中可能对成果质量有帮助的原则。
|
|
3. 核心完成后，选你认为最能提升成果质量的方向，深入打磨。
|
|
如果多轮都是同类型的小修——换一个完全不同的角度重新审视，
|
|
找到新的薄弱点再深入。包括但不限于测试/边界case/性能/安全/文档/代码质量。
|
|
4. 换角度的方法：假装你是第一次看到这个成果的使用者/审阅者/攻击者，
|
|
找到它最容易出问题的地方，然后修它。
|
|
5. 要为了目标持续推进，在工作文件夹中记录进度，不要更新全局记忆。
|
|
"""
|
|
|
|
BUDGET_LIMIT_PROMPT = """[Goal Mode — 预算耗尽，收口]
|
|
|
|
<objective>
|
|
{objective}
|
|
</objective>
|
|
|
|
⏱ 预算已耗尽（{budget_min:.0f} 分钟）。这是最后一轮。
|
|
|
|
请执行收口：
|
|
1. 总结本次 goal 的所有进展（列表）
|
|
2. 列出未完成的事项和建议的 next step
|
|
3. 确保工作文件夹中记录了关键成果
|
|
4. 清理一些确定无用的中间临时文件和不再用的进程
|
|
{done_prompt}
|
|
"""
|
|
|
|
# --- 主逻辑 ---
|
|
def check():
|
|
state = _load()
|
|
if state is None: return '/exit'
|
|
|
|
status = state.get('status', 'running')
|
|
if status != 'running': return '/exit'
|
|
|
|
start_time = state.get('start_time', time.time())
|
|
budget_sec = state.get('budget_seconds', 1800) # 默认30分钟
|
|
elapsed = time.time() - start_time
|
|
remaining = budget_sec - elapsed
|
|
turn = state.get('turns_used', 0) + 1
|
|
max_turns = state.get('max_turns', 50) # 防空转上限
|
|
|
|
# 预算耗尽或轮次上限
|
|
if remaining <= 0 or turn > max_turns:
|
|
state['status'] = 'wrapping_up'
|
|
_save(state)
|
|
return BUDGET_LIMIT_PROMPT.format(
|
|
objective=state['objective'],
|
|
budget_min=budget_sec / 60,
|
|
done_prompt=state.get('done_prompt', '')
|
|
)
|
|
|
|
# 正常continuation
|
|
state['turns_used'] = turn
|
|
_save(state)
|
|
return CONTINUATION_PROMPT.format(
|
|
objective=state['objective'],
|
|
elapsed_min=elapsed / 60,
|
|
remaining_min=remaining / 60,
|
|
turn=turn
|
|
)
|
|
|
|
def on_done(result):
|
|
state = _load()
|
|
if state is None: return
|
|
|
|
if state.get('status') == 'wrapping_up':
|
|
state['status'] = 'done_budget'
|
|
state['end_time'] = time.time()
|
|
_save(state)
|