mirror of
https://github.com/hansjone/oclaw.git
synced 2026-10-09 07:20:44 +08:00
oliver
97d5c1744a
Promote DSML markup from assistant text into native tool calls during streaming and completion, filter DSML from UI tokens, and document AIA_DSML_TEXT_TOOLS for local vLLM proxies. Co-authored-by: Cursor <cursoragent@cursor.com>
297 lines
13 KiB
Python
297 lines
13 KiB
Python
from __future__ import annotations
|
||
|
||
import hashlib
|
||
import threading
|
||
from pathlib import Path
|
||
from typing import Any
|
||
|
||
from svc.config.paths import PROJECT_ROOT
|
||
from runtime.memory_stage import render_memory_context_block
|
||
from runtime.dsml_tool_parse import dsml_text_tools_enabled
|
||
from runtime.project_context_prompt import build_project_context_block
|
||
from runtime.skill_role_binding import SKILL_ROLE_BINDING_KEY
|
||
from runtime.skills import workspace_skills_layout_signature
|
||
from runtime.skills_prompt import build_skills_catalog_block
|
||
from runtime.skills_workspace_lane import (
|
||
fs_safe_workspace_lane_segment,
|
||
workspace_lane_segment,
|
||
)
|
||
from runtime.types import OclawMemoryContext
|
||
from runtime.workspaces.experts import expert_workspace_signature_token
|
||
from runtime.prompt_templates import render_runtime_prompt
|
||
from runtime.tools.base import ToolRegistry
|
||
|
||
_EXECUTOR_STATIC_PROMPT_CACHE_LOCK = threading.Lock()
|
||
_EXECUTOR_STATIC_PROMPT_CACHE: dict[tuple[Any, ...], str] = {}
|
||
|
||
|
||
def _file_stat_signature(path: Path) -> str:
|
||
try:
|
||
st = path.stat()
|
||
return f"{path.as_posix()}:{int(st.st_mtime_ns)}:{int(st.st_size)}"
|
||
except Exception:
|
||
return f"{path.as_posix()}:missing"
|
||
|
||
|
||
def _session_bootstrap_content_signature(*, skill_binding_role: str | None, workspace_dir: str | None) -> str:
|
||
"""Best-effort signature for session-bootstrap content sources.
|
||
|
||
When this signature changes, cached executor system prompts are invalidated,
|
||
so the next turn rebuilds project context and re-runs bootstrap hooks.
|
||
"""
|
||
repo = Path(PROJECT_ROOT).resolve()
|
||
role = str(skill_binding_role or "").strip().lower()
|
||
ws = Path(workspace_dir).expanduser().resolve() if str(workspace_dir or "").strip() else None
|
||
|
||
targets: list[Path] = []
|
||
# session-bootstrap static source files
|
||
targets.append((repo / "skills" / "session-bootstrap" / "SOUL.md").resolve())
|
||
targets.append((repo / "skills" / "session-bootstrap" / "IDENTITY.md").resolve())
|
||
# wiki global memory/improvement sources
|
||
targets.append((repo / "data" / "wiki" / "improvement" / "learnings.md").resolve())
|
||
targets.append((repo / "data" / "wiki" / "improvement" / "errors.md").resolve())
|
||
targets.append((repo / "data" / "wiki" / "improvement" / "feature-requests.md").resolve())
|
||
targets.append((repo / "data" / "wiki" / "users" / "current.md").resolve())
|
||
|
||
core_dir = (repo / "data" / "wiki" / "core").resolve()
|
||
if core_dir.exists() and core_dir.is_dir():
|
||
for p in sorted(core_dir.glob("*.md"), key=lambda x: x.name.lower()):
|
||
targets.append(p.resolve())
|
||
|
||
if role:
|
||
role_dir = (repo / "data" / "wiki" / "experts" / role).resolve()
|
||
if role_dir.exists() and role_dir.is_dir():
|
||
for p in sorted(role_dir.glob("*.md"), key=lambda x: x.name.lower()):
|
||
targets.append(p.resolve())
|
||
|
||
if ws is not None:
|
||
mem_dir = (ws / "memory").resolve()
|
||
if mem_dir.exists() and mem_dir.is_dir():
|
||
mem_files = [p for p in mem_dir.glob("*.md") if p.is_file()]
|
||
mem_files.sort(key=lambda p: p.stat().st_mtime if p.exists() else 0.0, reverse=True)
|
||
if mem_files:
|
||
targets.append(mem_files[0].resolve())
|
||
|
||
h = hashlib.sha256()
|
||
for p in targets:
|
||
h.update(_file_stat_signature(p).encode("utf-8", errors="ignore"))
|
||
h.update(b"\n")
|
||
return h.hexdigest()[:24]
|
||
|
||
|
||
def _executor_prompt_settings_signature(store: Any) -> tuple[str, ...]:
|
||
keys = (
|
||
"AIA_SKILL_RUNTIME_ENABLED",
|
||
"AIA_SKILLS_PROMPT_IN_SYSTEM",
|
||
"AIA_SKILL_DISABLED_NAMES",
|
||
"AIA_SKILL_ROLE_BINDING_ENABLED",
|
||
"AIA_SKILL_ROLE_BINDING_MANAGER_INHERIT",
|
||
SKILL_ROLE_BINDING_KEY,
|
||
"AIA_PROJECT_CONTEXT_MAX_FILE_CHARS",
|
||
"AIA_PROJECT_CONTEXT_MAX_TOTAL_CHARS",
|
||
)
|
||
parts: list[str] = []
|
||
for key in keys:
|
||
try:
|
||
val = str(store.get_setting(key) or "")
|
||
except Exception:
|
||
val = ""
|
||
parts.append(f"{key}={val}")
|
||
return tuple(parts)
|
||
|
||
|
||
def _unified_skill_policy_guidance(*, dsml_recovery_enabled: bool = False) -> str:
|
||
# Global policy for all agents (including dynamic/ephemeral) — appended to base_system in build_executor_system_prompt.
|
||
if dsml_recovery_enabled:
|
||
tool_call_lines = (
|
||
"- 工具调用优先使用模型的原生 `tool_calls` 协议；若模型输出 DeepSeek DSML 标记，"
|
||
"运行时会自动转换，请勿在可见回复中重复或解释 DSML/XML 语法。\n"
|
||
)
|
||
else:
|
||
tool_call_lines = (
|
||
"- 工具调用必须通过模型的原生 `tool_calls` 协议返回；不要在正文输出 DSML/XML/工具协议 JSON 模板。\n"
|
||
"- 若当前模型链路不支持原生 `tool_calls`，请用自然语言明确说明“当前无法发起工具调用”，"
|
||
"并请求切换到支持该协议的链路；不要伪造或模拟工具调用。\n"
|
||
)
|
||
return (
|
||
"## 技能使用政策（Skill Usage Policy）：\n"
|
||
"- 每次会话启动后，在进行文件/路径相关操作前，必须先读取 `skills/_workspace/public/path-convention/SKILL.md` 了解当前路径规范。\n"
|
||
"- 如果用户问你有哪些技能（skill/技能），请直接根据已注入的技能目录及其 description 回答。\n"
|
||
"- 不要为了“列出技能”而去读取 SKILL.md。只有在你确实需要某个技能的详细使用说明时，才读取对应 SKILL.md。\n"
|
||
"- 当你需要技能细节时，请按目录中给出的 path 读取对应的 SKILL.md。\n"
|
||
"- 当对话涉及长期记忆、用户身份/偏好、项目背景延续、复发问题沉淀时，优先启用 wiki-first-autonomy 技能，并优先使用 memory_wiki_search/memory_wiki_get 检索上下文，再执行与回复。\n"
|
||
"- 使用 memory_wiki_search 时默认采用三段检索：先 `query` 精确检索；若结果不足再启用 `expand_query=true`；仍不足时加 `path_prefix` 定向到候选目录后重检。\n"
|
||
"- 三段检索建议参数：第一段 `limit=5~8`；第二段 `expand_query=true,max_rounds=2~3`；第三段在保留前述参数基础上增加 `path_prefix` 并分页（`offset`/`limit`）。\n"
|
||
"- 每段检索后先依据 `queries_attempted`、`total_hits_estimate`、`next_offset` 判断是否继续，避免一次无命中就直接下结论。\n"
|
||
"- memory wiki 请仅传相对 wiki 根的路径（相对 `data/wiki`），例如 `improvement/learnings.md`；不要传 `data/wiki/...` 或绝对路径。\n"
|
||
"- 当新增事实会影响后续决策时，完成当前任务后使用 memory_wiki_apply 写入结构化记忆，并用 memory_wiki_lint 做质量检查。\n"
|
||
"- 技能包由说明文档和可选文件组成。运行时不会自动执行技能 `scripts/` 目录下的文件；\n"
|
||
"- internal hooks 是独立系统，也不会自动执行这些脚本。\n"
|
||
"- 当用户明确要求运行技能脚本时，请使用项目允许的 terminal/bash/exec（或同类）工具执行；\n"
|
||
"- 如果脚本依赖相对路径（例如 `.learnings/`），请将工作目录设置为用户工作区。\n"
|
||
"- 在没有显式工具调用成功结果前，不要假设脚本已经执行。\n"
|
||
"- 在 Windows 上，`.sh` 可能需要 Git Bash、WSL 或等效环境。\n"
|
||
f"{tool_call_lines}"
|
||
"- 当用户目标是“安装 skill/技能”时，必须遵循 `oclaw-skill-manager` 的安装策略，并以其为唯一规范来源。\n"
|
||
"- 安装路径强约束：仅允许 `skill_auto_install`（`_workspace` lane）；不得改用任何非 auto 路径或脚本绕过。\n"
|
||
"- 严禁臆测前置条件：不要把未在规范中声明的环境变量、端口、服务启动状态当作必需前提。\n"
|
||
"- 若安装失败，仅输出可验证事实（至少包含 `error_code` 与 `detail`）和最小下一步，不得编造基础设施依赖。\n"
|
||
)
|
||
|
||
|
||
def _skill_catalog_lane_flags(
|
||
*,
|
||
skill_binding_role: str | None,
|
||
workspace_owner_session_id: str | None,
|
||
session_id: str | None,
|
||
) -> tuple[bool, str | None]:
|
||
role = str(skill_binding_role or "").strip().lower()
|
||
if role:
|
||
return True, fs_safe_workspace_lane_segment(role)
|
||
o = str(workspace_owner_session_id or "").strip()
|
||
s = str(session_id or "").strip()
|
||
if not o and not s:
|
||
return False, None
|
||
seg = workspace_lane_segment(workspace_owner_session_id=o or None, session_id=s or None)
|
||
return True, seg
|
||
|
||
|
||
def build_executor_system_prompt(
|
||
*,
|
||
store: Any,
|
||
tools: ToolRegistry | None,
|
||
base_url: str,
|
||
base_system: str,
|
||
memory_context: OclawMemoryContext | None,
|
||
lang: str,
|
||
workspace_dir: str | None = None,
|
||
skill_binding_role: str | None = None,
|
||
workspace_owner_session_id: str | None = None,
|
||
session_id: str | None = None,
|
||
model_id: str = "",
|
||
) -> str:
|
||
"""Build the final system string for the oclaw executor (memory block + skills catalog).
|
||
|
||
`lang` is reserved for future localized system fragments; current templates are bilingual/static.
|
||
"""
|
||
_ = lang # reserved for i18n extensions
|
||
final_system = get_executor_prompt_static(
|
||
store=store,
|
||
tools=tools,
|
||
base_url=base_url,
|
||
base_system=base_system,
|
||
workspace_dir=workspace_dir,
|
||
skill_binding_role=skill_binding_role,
|
||
workspace_owner_session_id=workspace_owner_session_id,
|
||
session_id=session_id,
|
||
model_id=model_id,
|
||
)
|
||
mem_block = render_memory_context_block(memory_context or OclawMemoryContext())
|
||
if not mem_block:
|
||
return final_system
|
||
return render_runtime_prompt(
|
||
"runtime/system_with_memory.md",
|
||
variables={"system_prompt": final_system, "memory_context": mem_block},
|
||
strict=True,
|
||
)
|
||
|
||
|
||
def get_executor_prompt_static(
|
||
*,
|
||
store: Any,
|
||
tools: ToolRegistry | None,
|
||
base_url: str,
|
||
base_system: str,
|
||
workspace_dir: str | None = None,
|
||
skill_binding_role: str | None = None,
|
||
workspace_owner_session_id: str | None = None,
|
||
session_id: str | None = None,
|
||
model_id: str = "",
|
||
) -> str:
|
||
excl, lane_seg = _skill_catalog_lane_flags(
|
||
skill_binding_role=skill_binding_role,
|
||
workspace_owner_session_id=workspace_owner_session_id,
|
||
session_id=session_id,
|
||
)
|
||
cache_key = (
|
||
str(base_url or "").strip(),
|
||
str(model_id or "").strip().lower(),
|
||
str(base_system or "").strip(),
|
||
str(workspace_dir or "").strip(),
|
||
str(skill_binding_role or "").strip().lower(),
|
||
expert_workspace_signature_token(),
|
||
_executor_prompt_settings_signature(store),
|
||
workspace_skills_layout_signature(),
|
||
bool(tools is not None),
|
||
int(bool(excl)),
|
||
str(lane_seg or ""),
|
||
_session_bootstrap_content_signature(skill_binding_role=skill_binding_role, workspace_dir=workspace_dir),
|
||
)
|
||
with _EXECUTOR_STATIC_PROMPT_CACHE_LOCK:
|
||
cached = _EXECUTOR_STATIC_PROMPT_CACHE.get(cache_key)
|
||
if isinstance(cached, str):
|
||
return cached
|
||
final_system = str(base_system or "").strip()
|
||
dsml_recovery = dsml_text_tools_enabled(base_url=str(base_url or ""), model_id=str(model_id or ""))
|
||
guide = _unified_skill_policy_guidance(dsml_recovery_enabled=dsml_recovery).strip()
|
||
if guide and guide not in final_system:
|
||
final_system = f"{final_system}\n\n{guide}".strip()
|
||
project_block = build_project_context_block(store=store, workspace_dir=workspace_dir)
|
||
if project_block:
|
||
final_system = f"{final_system}\n\n{project_block}".strip()
|
||
if tools is not None:
|
||
cat = build_skills_catalog_block(
|
||
store=store,
|
||
registry=tools,
|
||
base_url=str(base_url or ""),
|
||
skill_binding_role=skill_binding_role,
|
||
exclude_foreign_private_workspace_skills=excl,
|
||
private_workspace_lane_segment=lane_seg,
|
||
)
|
||
if cat.strip():
|
||
final_system = render_runtime_prompt(
|
||
"runtime/system_with_skills.md",
|
||
variables={"system_body": final_system, "skills_catalog": cat},
|
||
strict=True,
|
||
)
|
||
with _EXECUTOR_STATIC_PROMPT_CACHE_LOCK:
|
||
_EXECUTOR_STATIC_PROMPT_CACHE[cache_key] = final_system
|
||
if len(_EXECUTOR_STATIC_PROMPT_CACHE) > 256:
|
||
_EXECUTOR_STATIC_PROMPT_CACHE.clear()
|
||
return final_system
|
||
|
||
|
||
def warm_executor_prompt_cache(
|
||
*,
|
||
store: Any,
|
||
tools: ToolRegistry | None,
|
||
base_url: str,
|
||
role_base_systems: dict[str, str],
|
||
workspace_dir: str | None = None,
|
||
) -> dict[str, int]:
|
||
warmed = 0
|
||
for role, base_system in (role_base_systems or {}).items():
|
||
_ = get_executor_prompt_static(
|
||
store=store,
|
||
tools=tools,
|
||
base_url=base_url,
|
||
base_system=str(base_system or ""),
|
||
workspace_dir=workspace_dir,
|
||
skill_binding_role=str(role or "").strip().lower() or None,
|
||
workspace_owner_session_id=None,
|
||
session_id=None,
|
||
)
|
||
warmed += 1
|
||
return {"roles_warmed": int(warmed)}
|
||
|
||
|
||
def build_oclaw_executor_system_prompt(**kwargs: Any) -> str:
|
||
return build_executor_system_prompt(**kwargs)
|
||
|
||
|
||
__all__ = [
|
||
"build_executor_system_prompt",
|
||
"build_oclaw_executor_system_prompt",
|
||
"get_executor_prompt_static",
|
||
"warm_executor_prompt_cache",
|
||
]
|