diff --git a/agent/evolution/config.py b/agent/evolution/config.py index 864cbb1e..a3215858 100644 --- a/agent/evolution/config.py +++ b/agent/evolution/config.py @@ -14,7 +14,7 @@ from typing import Any # until release; enable via ``self_evolution_enabled``. DEFAULT_ENABLED = False DEFAULT_IDLE_MINUTES = 15 -DEFAULT_MIN_TURNS = 6 +DEFAULT_MIN_TURNS = 8 # Max review steps for the isolated evolution agent. Kept small (not exposed as # config): the review is meant to be cheap and focused, not a long autonomous run. DEFAULT_MAX_STEPS = 12 diff --git a/agent/evolution/executor.py b/agent/evolution/executor.py index 8734d263..33f5453e 100644 --- a/agent/evolution/executor.py +++ b/agent/evolution/executor.py @@ -175,6 +175,9 @@ _WATCH_SUBDIRS = ("MEMORY.md", "skills", "knowledge", "output") # Subpaths under memory/ to ignore: evolution's own bookkeeping + the nightly # dream diary, none of which count as a user-facing change signal. _MEMORY_IGNORE = (".evolution_backups", "dreams", "evolution") +# Files the skill subsystem maintains automatically (the enable/disable index). +# Not an evolution result, so a rewrite must not count as a change signal. +_WATCH_IGNORE_NAMES = ("skills_config.json",) def _workspace_snapshot(workspace_dir) -> dict: @@ -195,6 +198,8 @@ def _workspace_snapshot(workspace_dir) -> dict: for p in root.rglob("*"): if not p.is_file(): continue + if p.name in _WATCH_IGNORE_NAMES: + continue try: st = p.stat() snap[str(p.relative_to(ws))] = (st.st_mtime, st.st_size) @@ -269,10 +274,8 @@ def run_evolution_for_session( new_messages = all_messages[done:] transcript = _build_transcript(new_messages) if not transcript.strip(): - # Routine no-op: the per-minute scan hits every idle session, so keep - # this at debug to avoid spamming the log. - logger.debug(f"[Evolution] session={session_id}: no new messages, skip") - # Advance the cursor anyway so we don't re-scan the same tail. + # Routine no-op: the per-minute scan hits every idle session. Advance + # the cursor so we don't re-scan the same tail; no log (pure noise). agent._evo_done_msg_count = total_msgs return False @@ -334,17 +337,23 @@ def run_evolution_for_session( str(workspace_dir), ) review_agent = agent_bridge.create_agent( - system_prompt=EVOLUTION_SYSTEM_PROMPT, + system_prompt="", tools=review_tools, description="Self-evolution review agent", max_steps=cfg.max_steps, workspace_dir=str(workspace_dir), skill_manager=getattr(agent, "skill_manager", None), memory_manager=getattr(agent, "memory_manager", None), - enable_skills=False, + enable_skills=True, + runtime_info=getattr(agent, "runtime_info", None), ) # Reuse the live model so it follows the user's configured model. review_agent.model = agent.model + # Inject the evolution task brief AFTER the full system prompt: the agent + # gets the full context (tools, workspace, user preferences, memory, time) + # AND its evolution-specific instructions on top, instead of one + # overwriting the other. + review_agent.extra_system_suffix = EVOLUTION_SYSTEM_PROMPT logger.info( f"[Evolution] backup {backup_id} ({_backup_n} files) → running review agent" diff --git a/agent/evolution/prompts.py b/agent/evolution/prompts.py index dd71d82a..def9428b 100644 --- a/agent/evolution/prompts.py +++ b/agent/evolution/prompts.py @@ -7,7 +7,7 @@ language (instructed at the end of the prompt). Design goals (see ref/hermes-agent background_review for inspiration): - Default to doing NOTHING. Evolution is the exception, not the rule. - - Three signal types: memory, skill, unfinished task. + - Signal types: skill, unfinished task, memory, knowledge. - An explicit "do NOT capture" list to avoid self-poisoning over time. - Generic examples only — never bake in domain-specific business terms. """ @@ -72,12 +72,11 @@ them. When their signal is clear, act; do not be shy here. reply/decision, do NOTHING and stay [SILENT] — do not nag or ping the user. You only ever notify the user as a side effect of having actually done work. -3. MEMORY — LAST resort, and you are only a SAFETY NET here, not the primary - writer. The main assistant already writes memory DURING the conversation, and - a nightly pass consolidates daily notes into long-term memory. Prefer fixing - a skill (above) over writing memory whenever the fact belongs in a skill. - Act ONLY on something the main assistant clearly MISSED that does not belong - in any skill. +3. MEMORY — RARE, last resort. Default to writing NOTHING here. The main + assistant already writes memory during the chat, and a nightly pass plus + context-overflow saves are dedicated safety nets — so memory is almost always + already covered without you. Skip unless the main assistant clearly missed a + durable fact that belongs in no skill AND would visibly change future replies. - MEMORY.md is the curated long-term index, auto-loaded into EVERY future conversation. Treat it as precious: edit it in place to CORRECT a wrong fact, or append a new durable preference/decision/lesson — but do so @@ -92,6 +91,12 @@ them. When their signal is clear, act; do not be shy here. - If it is already captured anywhere (check MEMORY.md AND the daily file first), do NOTHING. +4. KNOWLEDGE — only if the conversation produced durable, reusable reference + knowledge on a topic (the kind worth looking up again) that the main + assistant did NOT already save to `knowledge/`. Add or update the relevant + file there. Like memory, this is the exception: skip routine Q&A, and if the + topic is already covered in `knowledge/`, do NOTHING rather than duplicate. + # Do NOT capture (these poison future behavior) - Environment failures: missing binaries, unset credentials, uninstalled @@ -140,9 +145,6 @@ def build_review_user_message(transcript: str, protected_skills: list = None) -> ``protected_skills`` lists skill names that must never be edited (built-in skills shipped with the product). Surfaced so the agent avoids them. """ - from datetime import datetime - today = datetime.now().strftime("%Y-%m-%d") - protected_note = "" if protected_skills: names = ", ".join(sorted(protected_skills)) @@ -152,12 +154,10 @@ def build_review_user_message(transcript: str, protected_skills: list = None) -> ) return ( "Here is the conversation transcript that just went idle. Review it per " - "your instructions and act on any clear signal. Prefer fixing a skill at " - "its source over writing memory whenever the fact belongs in a skill.\n" - f"Today is {today}. Only if a fact genuinely belongs in memory (and not " - f"in a skill): append one short bullet to the daily file " - f"`memory/{today}.md` for a new fact, or edit MEMORY.md in place to " - f"correct an existing wrong fact." + "your instructions. Acting is the exception: the main value is fixing or " + "creating a skill and finishing promised work. Memory and knowledge are " + "rare last resorts — stay [SILENT] unless there is a clear, durable signal " + "not already covered." f"{protected_note}\n" "\n" f"{transcript}\n" diff --git a/agent/protocol/agent.py b/agent/protocol/agent.py index 1dc72797..0b2e36ad 100644 --- a/agent/protocol/agent.py +++ b/agent/protocol/agent.py @@ -52,6 +52,11 @@ class Agent: self.workspace_dir = workspace_dir # Workspace directory self.enable_skills = enable_skills # Skills enabled flag self.runtime_info = runtime_info # Runtime info for dynamic time update + # Optional extra instructions appended AFTER the rebuilt full system + # prompt. Used by the self-evolution review agent to add its task brief + # on top of the full context (tools, workspace, user preferences, time) + # so it both follows the user's preferences and knows its evolution job. + self.extra_system_suffix = None # Initialize skill manager self.skill_manager = None @@ -120,15 +125,20 @@ class Agent: except Exception: lang = "zh" builder = PromptBuilder(workspace_dir=self.workspace_dir or "", language=lang) - return builder.build( + full = builder.build( tools=self.tools, context_files=context_files, skill_manager=self.skill_manager, memory_manager=self.memory_manager, runtime_info=self.runtime_info, ) + if self.extra_system_suffix: + full = f"{full}\n\n{self.extra_system_suffix}" + return full except Exception as e: logger.warning(f"Failed to rebuild system prompt, using cached version: {e}") + if self.extra_system_suffix: + return f"{self.system_prompt}\n\n{self.extra_system_suffix}" return self.system_prompt def refresh_skills(self): diff --git a/bridge/agent_initializer.py b/bridge/agent_initializer.py index e7161454..857c0f92 100644 --- a/bridge/agent_initializer.py +++ b/bridge/agent_initializer.py @@ -524,6 +524,14 @@ class AgentInitializer: logger.debug("[AgentInitializer] WebSearch skipped - no search provider configured") continue + # Skip evolution_undo when self-evolution is disabled: with no + # evolution there is nothing to roll back, so the tool is dead weight. + if tool_name == "evolution_undo": + from agent.evolution.config import get_evolution_config + if not get_evolution_config().enabled: + logger.debug("[AgentInitializer] evolution_undo skipped - self-evolution disabled") + continue + # Special handling for EnvConfig tool if tool_name == "env_config": from agent.tools import EnvConfig diff --git a/channel/web/chat.html b/channel/web/chat.html index 5bb4c7ae..1af8e334 100644 --- a/channel/web/chat.html +++ b/channel/web/chat.html @@ -772,7 +772,7 @@ diff --git a/channel/web/static/js/console.js b/channel/web/static/js/console.js index 4f61e8db..00598e25 100644 --- a/channel/web/static/js/console.js +++ b/channel/web/static/js/console.js @@ -123,7 +123,7 @@ const I18N = { config_max_turns: '最大记忆轮次', config_max_turns_hint: '一问一答为一轮，超过后会智能压缩处理', config_max_steps: '最大执行步数', config_max_steps_hint: '单次对话中 Agent 最多调用工具的次数', config_enable_thinking: '深度思考', config_enable_thinking_hint: '是否启用深度思考模式', - config_self_evolution: '自我进化', config_self_evolution_hint: '会话空闲后自动复盘，沉淀记忆、优化技能、处理未完成事项', + config_self_evolution: '自进化', config_self_evolution_hint: '会话空闲后自动复盘，沉淀记忆、优化技能、处理未完成事项', evolution_badge: '自主学习', config_channel_type: '通道类型', config_provider: '模型厂商', config_model_name: '模型', @@ -142,7 +142,7 @@ const I18N = { skills_section_title: '技能', skill_enable: '启用', skill_disable: '禁用', skill_toggle_error: '操作失败，请稍后再试', memory_title: '记忆管理', memory_desc: '查看 Agent 记忆文件和内容', - memory_tab_files: '记忆文件', memory_tab_dreams: '自我进化', + memory_tab_files: '记忆文件', memory_tab_dreams: '自进化', memory_loading: '加载记忆文件中...', memory_loading_desc: '记忆文件将显示在此处', memory_back: '返回列表', memory_col_name: '文件名', memory_col_type: '类型', memory_col_size: '大小', memory_col_updated: '更新时间', diff --git a/config.py b/config.py index cd54d29a..1f1100fe 100644 --- a/config.py +++ b/config.py @@ -254,7 +254,7 @@ available_setting = { # Self-evolution: review idle conversations to learn memory/skills. Flat keys. "self_evolution_enabled": False, # master switch (off until release) "self_evolution_idle_minutes": 15, # idle time before a session is reviewed - "self_evolution_min_turns": 6, # min user turns (or context pressure) to trigger + "self_evolution_min_turns": 8, # min user turns (or context pressure) to trigger "skill": {}, # Per-skill runtime config; nested keys flatten to SKILL__ env vars at startup "mcp_servers": [], # MCP server list; each entry supports type "stdio" (local process) or "sse" (remote URL) } diff --git a/docs/memory/self-evolution.mdx b/docs/memory/self-evolution.mdx index d06855f9..ad0ebd73 100644 --- a/docs/memory/self-evolution.mdx +++ b/docs/memory/self-evolution.mdx @@ -18,7 +18,7 @@ Self-Evolution focuses on three things: | Goal | Description | | --- | --- | | **Consolidate memory** | Record important preferences, decisions, and facts from the conversation, filling in what the main chat may have missed | -| **Improve skills** | When a skill shows a problem in use (such as a wrong setting or a missing step), fix the skill file directly instead of just noting it; create a new skill when one is genuinely needed | +| **Improve skills** | ① When a skill shows a problem in use (such as a wrong setting or a missing step), fix the skill file directly; ② when a reusable workflow emerges, turn it into a new skill so it can be reused next time | | **Follow up on unfinished tasks** | Spot the to-dos left in a conversation and finish them when possible | Once a review is done, if it actually changed something, the Agent tells you in a single line what it just learned and what it adjusted, so you can decide whether to roll it back. diff --git a/docs/zh/memory/self-evolution.mdx b/docs/zh/memory/self-evolution.mdx index bf18029b..3557f018 100644 --- a/docs/zh/memory/self-evolution.mdx +++ b/docs/zh/memory/self-evolution.mdx @@ -1,5 +1,5 @@ --- -title: 自我进化 +title: 自进化 description: Self-Evolution — 会话空闲后自动复盘，沉淀记忆、优化技能、处理未完成事项 --- @@ -7,19 +7,20 @@ description: Self-Evolution — 会话空闲后自动复盘，沉淀记忆、优 ### 简介 -自我进化（Self-Evolution）让 Agent 不止于"完成单次任务"，而是能在与你的相处中持续成长。每段对话告一段落后，它会自动"回头复盘"一次：把值得记住的沉淀为长期记忆、把使用中暴露的问题修进技能、把没做完的事情接着推进。久而久之，Agent 会越来越懂你的偏好、越来越少重复犯错、越来越主动地把事情收尾，而这一切都在后台静默完成，只有真正做了事情时才会简短地告诉你。 +自进化（Self-Evolution）让 Agent 不止于"完成单次任务"，而是能在与你的相处中持续成长。每段对话告一段落后，它会自动"回头复盘"一次：把使用中暴露的问题修进技能、把没做完的事情接着推进，并把值得记住的沉淀进记忆与知识库。久而久之，Agent 会越来越懂你的偏好、越来越少重复犯错、越来越主动地把事情收尾，而这一切都在后台静默完成，只有真正做了事情时才会简短地告诉你。 -> 它与[梦境蒸馏](/zh/memory/deep-dream)互补：梦境蒸馏负责整理记忆本身，自我进化则在记忆之外，进一步优化技能、推进未完成的任务，让 Agent 的能力随使用不断打磨。 +> 它与[梦境蒸馏](/zh/memory/deep-dream)互补：梦境蒸馏负责整理记忆本身，自进化则在记忆之外，进一步优化技能、推进未完成的任务，让 Agent 的能力随使用不断打磨。 -### 三个目标 +### 几个目标 -自我进化围绕三件事工作： +自进化围绕以下几件事工作，并以「优化技能、处理未完成事项」为主，「沉淀记忆、知识」作为主对话的查缺补漏： | 目标 | 说明 | | --- | --- | -| **沉淀记忆** | 把对话中重要的偏好、决策、事实补记到记忆中，作为主对话的查缺补漏 | -| **优化技能** | 当某个技能在使用中暴露出问题（如配置错误、步骤缺失），直接修正技能文件，而不只是记一笔；也可在需要时创建新技能 | +| **优化技能** | ① 技能在使用中暴露问题（如配置错误、步骤缺失）时，直接修正技能文件；② 出现一套可复用的流程时，主动固化为新技能，下次直接调用 | | **处理未完成事项** | 识别对话中遗留的待办，在能完成时直接完成 | +| **沉淀记忆** | 把对话中重要的偏好、决策、事实补记到记忆中，作为主对话的查缺补漏 | +| **沉淀知识** | 把对话中产生的、值得日后查阅的可复用知识补充进知识库（主对话遗漏时） | 复盘完成后，如果确实做了改动，Agent 会在对话中用一句话告诉你"刚刚自主学习了什么、调整了哪里"，方便你判断是否需要回滚。 @@ -27,7 +28,7 @@ description: Self-Evolution — 会话空闲后自动复盘，沉淀记忆、优 ### 触发时机 -自我进化不是定时执行，而是在**一段对话自然结束、进入空闲后**才触发，避免打断正在进行的交流。需要同时满足： +自进化不是定时执行，而是在**一段对话自然结束、进入空闲后**才触发，避免打断正在进行的交流。需要同时满足： - **对话已空闲** — 距离最后一次互动超过设定的空闲时长（默认 15 分钟） - **对话有足够内容** — 自上次进化以来累积了足够轮次，或上下文已接近容量上限 @@ -36,11 +37,11 @@ description: Self-Evolution — 会话空闲后自动复盘，沉淀记忆、优 ### 相关配置 -自我进化默认关闭，可在 Web 控制台「配置 → Agent 配置」中通过开关启用（位于"深度思考"下方），也可在配置文件中调整： +自进化默认关闭，可在 Web 控制台「配置 → Agent 配置」中通过开关启用（位于"深度思考"下方），也可在配置文件中调整： | 参数 | 说明 | 默认值 | | --- | --- | --- | -| `self_evolution_enabled` | 是否启用自我进化 | `false` | +| `self_evolution_enabled` | 是否启用自进化 | `false` | | `self_evolution_idle_minutes` | 对话空闲多久后触发（分钟） | `15` | | `self_evolution_min_turns` | 触发所需的最少对话轮次 | `6` | @@ -50,7 +51,7 @@ description: Self-Evolution — 会话空闲后自动复盘，沉淀记忆、优 ### 进化记录 -每次进化的过程和结果会按日期记录在 `memory/evolution/YYYY-MM-DD.md` 中，可在 Web 控制台的「记忆管理 → 自我进化」tab 中查看。该 tab 同时汇总了自我进化记录与梦境日记，方便统一回顾 Agent 的成长轨迹。 +每次进化的过程和结果会按日期记录在 `memory/evolution/YYYY-MM-DD.md` 中，可在 Web 控制台的「记忆管理 → 自进化」tab 中查看。该 tab 同时汇总了自进化记录与梦境日记，方便统一回顾 Agent 的成长轨迹。 ### 如何回滚 @@ -58,7 +59,7 @@ description: Self-Evolution — 会话空闲后自动复盘，沉淀记忆、优 ## 实现设计 -自我进化复用了系统已有的能力，保持轻量： +自进化复用了系统已有的能力，保持轻量： - **隔离执行**：每次复盘都启动一个独立的、临时的复盘任务，使用与主对话相同的模型，但拥有受限的工具集（只能读上下文、改记忆与技能文件）。它不会污染主对话的上下文，也不会影响主对话的性能。 - **基于备份的撤销**：进化前对相关文件做快照备份，撤销时按备份还原，因此每一次改动都可追溯、可逆。 @@ -66,7 +67,7 @@ description: Self-Evolution — 会话空闲后自动复盘，沉淀记忆、优 ### 克制与安全 -自我进化的设计原则是"必要时执行，减少打扰"： +自进化的设计原则是"必要时执行，减少打扰"： | 机制 | 说明 | | --- | --- | diff --git a/tests/test_evolution.py b/tests/test_evolution.py index 4ebc47e9..4b535234 100644 --- a/tests/test_evolution.py +++ b/tests/test_evolution.py @@ -195,6 +195,55 @@ def scenario_silent(): } +def scenario_silent_qa(): + """A normal knowledge Q&A -> nothing durable, should stay SILENT.""" + return { + "name": "普通问答 (should stay SILENT)", + "goal": "none", + "turns": [ + ("Python 里 list 和 tuple 有什么区别？", + "主要区别：list 可变、用 []；tuple 不可变、用 ()。tuple 更省内存、可作字典键。"), + ("那什么时候该用 tuple？", "当数据不应被修改、或要做字典键/集合元素时用 tuple。"), + ("懂了，谢谢", "不客气。"), + ], + "scripted": "[SILENT]", + "on_edit": None, + "expect_evolved": False, + } + + +def scenario_silent_transient(): + """User shares transient, non-durable info -> should stay SILENT.""" + return { + "name": "临时信息 (should stay SILENT)", + "goal": "none", + "turns": [ + ("帮我看下今天天气适合跑步吗，深圳", "深圳今天多云 26°C，傍晚湿度高，清晨或晚上跑步比较合适。"), + ("那我晚上去吧", "好的，记得补水。"), + ("行", "👍"), + ], + "scripted": "[SILENT]", + "on_edit": None, + "expect_evolved": False, + } + + +def scenario_silent_advice(): + """User asks for one-off advice, no reusable workflow -> should stay SILENT.""" + return { + "name": "一次性建议 (should stay SILENT)", + "goal": "none", + "turns": [ + ("给我起三个适合咖啡馆的名字", "可以考虑：① 拾光咖啡 ② 角落 Corner ③ 慢半拍。"), + ("第二个不错", "嗯，「角落 Corner」简洁好记。"), + ("就用这个了", "好的，祝开业顺利。"), + ], + "scripted": "[SILENT]", + "on_edit": None, + "expect_evolved": False, + } + + def scenario_memory_preference(): """User states a durable working preference -> update MEMORY.md.""" def edit(ws): @@ -475,6 +524,9 @@ def scenario_unfinished_task(): SCENARIOS = [ scenario_silent, + scenario_silent_qa, + scenario_silent_transient, + scenario_silent_advice, scenario_memory_preference, scenario_memory_correction, scenario_skill_gap, @@ -791,6 +843,12 @@ def run_real(): if __name__ == "__main__": + if "--debug" in sys.argv: + import logging + from common.log import logger as _cow_logger + _cow_logger.setLevel(logging.DEBUG) + for _h in _cow_logger.handlers: + _h.setLevel(logging.DEBUG) if "--real" in sys.argv: run_real() else: