mirror of
https://github.com/agentscope-ai/ReMe.git
synced 2026-09-11 22:51:10 +00:00
### 1. Agent Wrapper(统一 Agent 后端抽象) - **`base_agent_wrapper.py`**:`reply()` 返回值从 `tuple[str, Any]` 改为 `dict`(含 `session_id` / `last_message` / `result` / 可选 `structured_output`);`reply_stream()` 改为产出统一的 `StreamChunk`。废弃 `add_tools()`,改为 `add_job_tools(names: list[str])`(按名解析 BaseJob)与 `add_skills()`;新增 `_resolve_job_tools()`、`_merged_kwargs()`、`_chunk()` 辅助方法及 `project_path` / `project_skills_root` 属性。 - **`as_agent_wrapper.py`(AgentScope 后端)**: - 会话持久化重写:`session_path` 落地到 `<vault>/<session_dir>/agentscope/`,`_load_state` 支持 `resume` / `session_id` / `fork_session`,并做 UUID 校验(`_validate_session_id`);`_cleanup_expired_sessions` 按天数清理过期会话。 - 新增内置工具集(`BypassAnalysisBash` + Edit/Glob/Grep/Read/Write),`BypassAnalysisBash` 绕过 AgentScope 自带 Bash 静态分析以让 permission_mode 生效;`_resolve_skills()` 把配置的 skill 暴露给后端,`_load_tool_env()` 注入项目 `.env`。 - `_event_to_chunk()` 把 20+ 种 AgentScope 事件(Reply/Text/Thinking/Data/ToolCall/ToolResult/ModelCall/ExceedMaxIters)归一化为 `StreamChunk`。 - **`cc_agent_wrapper.py`(Claude Code SDK 后端,+551 行)**: - 新增 `_CcFileSessionStore`:基于 vault 的文件型会话存储,实现 append(按 uuid 去重)/ load / list / delete / list_subkeys,并对路径做 `_safe_parts` + `resolve()` 防越界校验。 - `_build_options()`:统一构建 `ClaudeAgentOptions`,处理 skills、disallowed_tools(默认禁 `WebSearch`)、`.env` 注入、Claude Code 的 API 凭据解析(`_claude_code_api_env`,多级 base_url/api_key 回退)、`CLAUDE_CONFIG_DIR` 设置、skill 目录软链接(`_ensure_claude_skill_dir`)。 - `_raw_event_to_chunk()` / `_message_content_to_chunks()`:把 Anthropic 流式事件(message_start/delta/stop、content_block_*)与 SDK 消息块(AssistantMessage/UserMessage/ResultMessage/RateLimitEvent)转换为统一 `StreamChunk`;跟踪 block_id/block_type/tool_call_name 做关联;处理尾部 `"success"` 误报异常的吞掉逻辑。 ### 2. 统一流式协议(StreamChunk / ChunkEnum) - **`stream_chunk.py`**:`StreamChunk` 扩展为承载 AS + CC 双后端完整信息的统一结构,新增 `session_id` / `block_id` / `tool_call_id` / `tool_call_name` / `media_type` / `input_tokens` / `output_tokens` 等字段,纯文本流仍保持轻量。 - **`chunk_enum.py`**:补全生命周期标记 `REPLY_START` / `REPLY_END`,并文档化两套后端事件 → ChunkEnum 的映射。 ### 3. Index 模块重构(变化批次化 + dispatch) - 新增 `_change_batch.py`:`coalesce_changes()` 把同路径多次事件折叠为最终状态(结合 path 存在性判定),`bucket_changes()` 按 watchfiles.Change 分桶。 - 新增 `init_changes.py`(`InitChangesStep`):一次性扫描,对比 file_store / file_catalog 已索引节点计算 added/modified/deleted,写入 `context["changes"]` 后 dispatch。 - 新增 `update_changes.py`:抽象基类 `ChangeApplyStep` 统一 added/modified/deleted 处理与错误收集;`UpdateCatalogStep`(写 file_catalog)、`UpdateIndexStep`(写 file_store,含按后缀解析 chunker)。 - **`watch_changes.py`**:改用 `dispatch_step_specs`(基类提供的 `dispatch_steps()`),每批先 `coalesce_changes` 再 dispatch;默认参数调整(debounce 5000ms / step 1000ms / poll 5000ms)并暴露常量。 - 删除旧步骤:`clear_and_scan` / `foreach_dispatch` / `scan_changes` / `update_catalog`(旧) / `update_index`(旧);`clear_store.py` 取代 clear_and_scan。 ### 4. Evolve / Dream 模块(拆分为多步 pipeline) - 删除旧的单体 `auto_dream.py` / `dream.py` / `dream.yaml`,新增 `dream/` 子包,按 5 个步骤组织: - **`extract.py`**:扫描当日 day-index + daily 笔记,对比 file_catalog 找出 changed/deleted,调用 LLM 全局抽取 `units`(procedure/personal/wiki 三桶)与 `topics`,路径与桶做清洗/路由。 - **`integrate.py`**:逐个 unit 调用 LLM 写入 digest,结构化输出 `IntegrateOutcome`(CREATE/CORROBORATE/REFINE/CORRECT),失败 unit/路径收集回写。 - **`topics.py`**:写 `daily/<date>/interests.yaml`,结合当天已有 + 近 N 天做去重(`normalize_topic`),可走 LLM 或纯规则去重两条路径。 - **`proactive.py`**:读取当日 `interests.yaml`,作为主动推荐话题的入口。 - **`finish.py`**:把变更路径落盘到 dream file_catalog(checkpoint),渲染最终汇总摘要。 - 新增 `schema.py`(`DreamState` 等跨步骤共享状态与结构化输出模型)与 `utils.py`(状态存取、扫描打包、YAML 读写、结构化回复解析等公共函数)。 - `evolve/__init__.py` 导出全部新 step。 ### 5. auto_memory / auto_resource(适配新 Agent API) - **`auto_memory.py`**:会话路径迁移到 `<session_dir>/dialog/<session_id>.jsonl`;改用 `job_tools`;新增 `source_conversation` frontmatter 反向链接(`_session_link`);执行后刷新 day 索引(`refresh_day_index`),并对 session_id 做合法性校验。 - **`auto_resource.py`**:资源改用「同名 daily note」方案(`_compute_note_stem` 取文件 stem);批量处理 `changes: list[dict]`(`_handle_change` 逐项处理,返回逐项结果摘要);agent 会话 id 用稳定的 `uuid5`;同样刷新 day 索引。 ### 6. BaseStep 基类增强 - 新增 `dispatch_steps` / `dispatch_step_specs` 机制:`_resolve_dispatch_step()` 支持字符串或 dict 形式的 step spec,`dispatch_steps()` 复用当前 context 调用下游 step。 - 新增 `config_value()`:按 key 取 app config,缺失时回退 `ApplicationConfig` 默认值。 - 小幅清理:`language` 初始化、`copy()`、`Ref.__init__` 签名精简。 ### 7. Components 改动 - **`file_store/local_file_store.py`**:持久化改用 zstd 压缩(`.jsonl.zst`,通过新 `utils/jsonl_zst.py`);upsert 时先删除旧 chunk 的 keyword 文档;embedding 复用改为 `(text, embedding)` 键控,要求文本一致才复用;新增 `_matches_search_filter()` 对 vector/keyword 搜索做 path/path_prefix/metadata 的统一后过滤。 - **`keyword_index/bm25_index.py`**:索引文件名加入组件名 + tokenizer 指纹(sha256 前 12 位),快照/恢复时校验指纹防配置漂移;空索引 dump 时删除文件,加载失败抛错而非静默。 - **`file_chunker/markdown_file_chunker.py`**:弃用 `python-frontmatter`,改用内置 YAML 解析(非法 YAML 不阻断正文索引),并修正因 frontmatter 占用行号导致的 AST 行号偏移(`line_offset`)。 - **`cron_job.py`**:大幅简化(-187 行),由原来「dispatch 外部 job/step + 多种调度模式」改为「在自身 steps 上跑 cron 表达式」;`Application` 启动顺序随之调整为 base > stream > background > cron。 - 其余小调整:service(base/http/mcp)、file_graph、file_catalog、as_llm、as_embedding、tokenizer、prompt_handler、base_component 的签名/接口微调。 ### 8. Application 生命周期 - `_start()` 启动顺序明确为 components → base → stream → background → cron,启动失败会触发 `_close()` 回滚并 re-raise(不再吞异常)。 - 启动时创建 `session_dir` 目录;新增 `update_component()`(按类型/名就地更新已存在组件,不存在则报错)。 ### 9. File IO / 路径安全 - **`_path.py`**:`resolve_path` 增加 vault 越界防护(`is_relative_to` 校验),禁止 `.` / `..` 路径分量,支持 `allow_empty`。 - **`read.py`**:大文件(超过 `MAX_FILE_READ_BYTES`)走按行读取 `read_file_lines_safe`,避免一次性载入内存。 - **`_file_io.py` / `_daily_index.py` / `_path.py`** 等支持函数补齐(如 `refresh_day_index`、`read_file_lines_safe`)。 - **`env_utils.py`**:新增 `parse_env_file()`,`load_env()` 返回加载到的键值、支持 `override`、对无路径调用做幂等缓存。 ### 10. Config - `ApplicationConfig` 新增 `session_dir`(默认 `reme_session`)。 - `config_parser.py`:环境变量展开后做类型转换(`_convert_value`)、dot-notation 与 key=value 参数校验更严格、配置文件路径支持相对 `_CONFIG_DIR` 查找、根非 dict 报错。 - `default.yaml`:作业编排改用 `init_changes_step` + `dispatch_steps`(index/resource/digest 三个 watch loop 与 reindex);新增 `auto_dream`(4 步)、`proactive` 作业,移除旧 `dream`;file_catalog 增配 `resource` / `digest` / `dream` 实例;LLM 默认值与 Claude Code 凭据配置调整(tool_result_limit 50000、thinking_enable=false 等)。 ### 11. 其它 - 新增 `steps/common/add.py`(`AddStep` 算术 demo)、`channel/__init__.py` 与 common `__init__` 导出整理。 - 新增 4 篇文档:`docs4/auto_dream_logic_and_step_refactor.md`、`docs4/watch_loop_step_refactor_plan.md`、`docs4/todo.md`,以及 `reme_design.md` 更新。 **
146 lines
6.1 KiB
Python
146 lines
6.1 KiB
Python
"""``file_delete`` — hard-delete a file or folder under the vault (reports inbound refs).
|
|
|
|
Removes the file (or folder tree) from disk; the watcher then prunes
|
|
the affected chunks from all projections (vector / keyword / file_graph).
|
|
For **soft** deletion that keeps a file addressable, use
|
|
``property_update(status=archived)`` instead — that's the path
|
|
structure.md's Decay algorithm is designed around.
|
|
|
|
Reference reporting (no auto-fix). Before deleting, the inbound
|
|
wikilink count for each doomed ``.md`` file is captured from the
|
|
file_graph's reverse index (literal full-path matching, same rule
|
|
as the wikilink retarget helper). For folder deletes, sources that live inside
|
|
the same folder are filtered out — those links die alongside the
|
|
delete and can't surface as dangling. The remaining inbound list
|
|
is the agent's punch list for follow-up rewrites.
|
|
|
|
The delete itself is unconditional: this step does not rewrite
|
|
inbound references on the agent's behalf because there is no
|
|
canonical "new target" — the agent decides per-reference whether
|
|
to rewrite (via ``file_move`` to a merge target), edit the
|
|
citing prose, or accept dangling.
|
|
"""
|
|
|
|
import shutil
|
|
from pathlib import Path
|
|
|
|
from ._path import resolve_path
|
|
from ..base_step import BaseStep
|
|
from ...components import R
|
|
from ...utils.wikilink_handler import WikilinkHandler
|
|
|
|
|
|
def _is_inside(rel: str, folder_rel: str) -> bool:
|
|
"""``rel`` is the same as or nested under ``folder_rel`` (relative to the vault)."""
|
|
prefix = folder_rel.rstrip("/") + "/"
|
|
return rel == folder_rel or rel.startswith(prefix)
|
|
|
|
|
|
@R.register("delete_step")
|
|
class DeleteStep(BaseStep):
|
|
"""Hard-delete the path at ``path`` (file or folder, relative to the vault)."""
|
|
|
|
async def execute(self):
|
|
assert self.context is not None
|
|
path: str = self.context.get("path", "") or ""
|
|
assert path, "path is required"
|
|
payload = await self._delete(path)
|
|
if "error" in payload:
|
|
self.context.response.success = False
|
|
self.context.response.answer = f"Error: {payload['error']}"
|
|
self.logger.info(f"[{self.name}] delete failed path={path} error={payload['error']!r}")
|
|
elif payload.get("is_dir"):
|
|
self.context.response.success = True
|
|
self.context.response.answer = f"Deleted directory {path} ({len(payload['deleted_files'])} file(s))"
|
|
self.logger.info(
|
|
f"[{self.name}] deleted dir path={path} files={len(payload['deleted_files'])} "
|
|
f"inbound_files={payload['inbound']['files_touched']} "
|
|
f"inbound_links={payload['inbound']['links_total']}",
|
|
)
|
|
else:
|
|
self.context.response.success = True
|
|
self.context.response.answer = f"Deleted {path}"
|
|
self.logger.info(
|
|
f"[{self.name}] deleted file path={path} "
|
|
f"inbound_files={payload['inbound']['files_touched']} "
|
|
f"inbound_links={payload['inbound']['links_total']}",
|
|
)
|
|
self.context.response.metadata.update(payload)
|
|
|
|
async def _delete(self, path: str) -> dict:
|
|
if not path:
|
|
return {"path": path, "error": "not found"}
|
|
vault_dir = Path(self.file_store.vault_path or ".").resolve()
|
|
target, err = resolve_path(vault_dir, path)
|
|
if err or target is None:
|
|
return {"path": path, "error": err or "invalid path"}
|
|
if target.is_file():
|
|
inbound = await WikilinkHandler.find_inbound(self.file_store, target=path)
|
|
target.unlink()
|
|
return {
|
|
"path": path,
|
|
"deleted": True,
|
|
"is_dir": False,
|
|
"deleted_files": [path],
|
|
"inbound": {
|
|
"files_touched": inbound.get("files_touched", 0),
|
|
"links_total": inbound.get("links_total", 0),
|
|
"by_file": inbound.get("by_file", []),
|
|
},
|
|
}
|
|
|
|
if target.is_dir():
|
|
folder_rel = path.rstrip("/")
|
|
deleted_files: list[str] = []
|
|
per_target: list[dict] = []
|
|
external_sources: set[str] = set()
|
|
links_total = 0
|
|
|
|
for md in sorted(target.rglob("*.md")):
|
|
try:
|
|
rel = str(md.relative_to(vault_dir))
|
|
except ValueError:
|
|
continue
|
|
deleted_files.append(rel)
|
|
inbound = await WikilinkHandler.find_inbound(self.file_store, target=rel)
|
|
# Drop sources that also live inside the doomed folder —
|
|
# their links vanish with them and aren't actionable.
|
|
external = [row for row in inbound.get("by_file", []) if not _is_inside(row["path"], folder_rel)]
|
|
if not external:
|
|
continue
|
|
target_total = sum(row["count"] for row in external)
|
|
links_total += target_total
|
|
external_sources.update(row["path"] for row in external)
|
|
per_target.append(
|
|
{
|
|
"target": rel,
|
|
"files_touched": len(external),
|
|
"links_total": target_total,
|
|
"by_file": external,
|
|
},
|
|
)
|
|
|
|
# Also enumerate non-md files for the deleted_files report.
|
|
for entry in sorted(target.rglob("*")):
|
|
if not entry.is_file() or entry.suffix == ".md":
|
|
continue
|
|
try:
|
|
rel = str(entry.relative_to(vault_dir))
|
|
except ValueError:
|
|
continue
|
|
deleted_files.append(rel)
|
|
|
|
shutil.rmtree(target)
|
|
return {
|
|
"path": path,
|
|
"deleted": True,
|
|
"is_dir": True,
|
|
"deleted_files": sorted(deleted_files),
|
|
"inbound": {
|
|
"files_touched": len(external_sources),
|
|
"links_total": links_total,
|
|
"by_target": per_target,
|
|
},
|
|
}
|
|
|
|
return {"path": path, "error": "not found"}
|