valkryhx/google_adk_agent / skills/self_improvement
self_improvement
VerifiedCaptures learnings, errors, and corrections to .learnings/ directory for continuous improvement.
tools.py
View on GitHub"""
Self-Improvement Skill - Tools Module
Provides log_learning and log_error tools for structured knowledge capture.
Also provides after_tool_callback and after_model_callback factory functions
for ADK native hook integration.
"""
import os
import datetime
import secrets
import traceback
import logging
from typing import Optional, Any, List
from pathlib import Path
logger = logging.getLogger(__name__)
# 项目根目录 (由 _init_paths 动态设置)
_PROJECT_ROOT: str = ""
_LEARNINGS_DIR: str = ""
# 错误检测模式列表 (移植自 error-detector.sh)
ERROR_PATTERNS = [
"error:", "Error:", "ERROR:",
"failed", "FAILED",
"command not found", "No such file",
"Permission denied", "fatal:",
"Exception", "Traceback",
"npm ERR!", "ModuleNotFoundError",
"SyntaxError", "TypeError",
"exit code", "non-zero",
"is not recognized", # Windows command not found
"无法完成", "失败", "错误", # Chinese errors
]
def _init_paths(project_root: str = "") -> None:
"""初始化路径 (幂等: 无参调用时若已初始化则跳过)"""
global _PROJECT_ROOT, _LEARNINGS_DIR
if not project_root and _LEARNINGS_DIR:
# 已初始化且未指定新路径,跳过
return
if not project_root:
project_root = os.path.dirname(
os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
)
_PROJECT_ROOT = project_root
_LEARNINGS_DIR = os.path.join(_PROJECT_ROOT, ".learnings")
os.makedirs(_LEARNINGS_DIR, exist_ok=True)
def _generate_id(prefix: str) -> str:
"""生成唯一 ID: LRN-20260421-A3F"""
date_str = datetime.datetime.now().strftime("%Y%m%d")
rand_suffix = secrets.token_hex(2)[:3].upper()
return f"{prefix}-{date_str}-{rand_suffix}"
async def log_learning(
category: str,
summary: str,
details: str = "",
priority: str = "medium",
area: str = "backend",
related_files: str = "",
tags: str = "",
) -> str:
"""
Record a learning entry to .learnings/LEARNINGS.md.
Use this when you discover something non-obvious, find a better approach,
or learn a project-specific convention.
Args:
category: Type of learning (correction, knowledge_gap, best_practice, workaround)
summary: One-line description of what was learned
details: Full context - what happened, what was wrong, what's correct
priority: low | medium | high | critical
area: frontend | backend | infra | tests | docs | config
related_files: Comma-separated list of related file paths
tags: Comma-separated tags for filtering
"""
_init_paths()
entry_id = _generate_id("LRN")
timestamp = datetime.datetime.now().isoformat()
entry = f"""
## [{entry_id}] {category}
**Logged**: {timestamp}
**Priority**: {priority}
**Status**: pending
**Area**: {area}
### Summary
{summary}
### Details
{details or 'N/A'}
### Suggested Action
Review and apply this learning in future tasks.
### Metadata
- Source: agent_session
- Related Files: {related_files or 'N/A'}
- Tags: {tags or category}
---
"""
learnings_file = os.path.join(_LEARNINGS_DIR, "LEARNINGS.md")
try:
with open(learnings_file, "a", encoding="utf-8") as f:
f.write(entry)
return f"[OK] Learning logged: {entry_id} ({category}: {summary})"
except Exception as e:
error_msg = f"[ERROR] Failed to write learning: {e}\n{traceback.format_exc()}"
logger.error(error_msg, exc_info=True)
return error_msg
async def log_error(
tool_name: str,
error_msg: str,
context: str = "",
priority: str = "high",
area: str = "backend",
related_files: str = "",
) -> str:
"""
Record an error entry to .learnings/ERRORS.md.
Use this when a tool or command fails unexpectedly, especially if it
required investigation to resolve.
Args:
tool_name: Name of the tool or command that failed
error_msg: The actual error message or output
context: What was being attempted and environment details
priority: low | medium | high | critical
area: frontend | backend | infra | tests | docs | config
related_files: Comma-separated list of related file paths
"""
_init_paths()
entry_id = _generate_id("ERR")
timestamp = datetime.datetime.now().isoformat()
entry = f"""
## [{entry_id}] {tool_name}
**Logged**: {timestamp}
**Priority**: {priority}
**Status**: pending
**Area**: {area}
### Summary
Error in {tool_name}
### Error
```
{error_msg}
```
### Context
{context or 'N/A'}
### Suggested Fix
Investigate the root cause and apply a fix.
### Metadata
- Reproducible: unknown
- Related Files: {related_files or 'N/A'}
---
"""
errors_file = os.path.join(_LEARNINGS_DIR, "ERRORS.md")
try:
with open(errors_file, "a", encoding="utf-8") as f:
f.write(entry)
return f"[OK] Error logged: {entry_id} ({tool_name})"
except Exception as e:
error_detail = f"[ERROR] Failed to write error log: {e}\n{traceback.format_exc()}"
logger.error(error_detail, exc_info=True)
return error_detail
def detect_error_in_output(output_text: str) -> bool:
"""检查输出文本是否包含错误模式 (移植自 error-detector.sh)"""
for pattern in ERROR_PATTERNS:
if pattern in output_text:
return True
return False
def build_after_tool_callback(loaded_skills_getter=None):
"""
工厂函数: 创建 after_tool_callback (常驻系统监控版)
返回 None = 不修改结果; 返回 dict = 替换工具结果
"""
def _after_tool_callback(tool, args, tool_context, tool_response):
# [Self-Improvement] 现在是常驻功能,默认始终执行监控
response_text = str(tool_response)
if detect_error_in_output(response_text):
tool_name = getattr(tool, "name", getattr(tool, "__name__", str(tool)))
logger.info(
"[SelfImprovement] Error detected in tool %s output, injecting reminder",
tool_name,
)
reminder = (
"\n\n<error-detected>"
"\nA tool error was detected. Consider logging to .learnings/ERRORS.md "
"using the `log_error` tool if this error was unexpected or required investigation."
"\n</error-detected>"
)
if isinstance(tool_response, dict):
modified = dict(tool_response)
for key in ("result", "output", "content"):
if key in modified:
modified[key] = str(modified[key]) + reminder
return modified
# fallback: 追加到第一个字符串值
for key, val in modified.items():
if isinstance(val, str):
modified[key] = val + reminder
return modified
# tool_response 不是 dict, 无法安全修改
return None
return None
return _after_tool_callback
def build_after_model_callback(loaded_skills_getter=None):
"""
工厂函数: 创建 after_model_callback (常驻系统监控版)
功能: LLM 回复后标记 session state, 提醒下一轮评估 learning
"""
def _after_model_callback(callback_context, llm_response):
# [Self-Improvement] 在 session state 中设置标记 (常驻执行)
try:
if hasattr(callback_context, "state") and callback_context.state is not None:
callback_context.state["_si_pending_eval"] = True
except Exception as e:
logger.error("[SelfImprovement] Failed to set state marker: %s", e, exc_info=True)
return None # 不修改 LLM 响应
return _after_model_callback
def get_tools(*args, **kwargs) -> List:
"""ADK skill 加载协议入口"""
# 初始化路径
if args and hasattr(args[0], "tools"):
# args[0] 是 agent 对象, 从它推断项目根目录
pass
_init_paths()
return [log_learning, log_error]
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| slack-gif-creatoranthropics/skillsKnowledge and utilities for creating animated GIFs optimized for Slack. Provides constraints, validation tools, and animation concepts. Use when users request animated GIFs for Slack like "make me a GIF of X doing Y for Slack." | Stars: 198k | Updated | Claude CodeCodex | Verified | Install |
| ponytailDietrichGebert/ponytailLazy senior dev mode for any coding task: the smallest change that fully solves it, and a reply a busy human understands in one read. | Stars: 194k | Updated | Claude CodeCodex | Verified | Install |
| ponytail-debtDietrichGebert/ponytailHarvest every shortcut: comment into one debt ledger, so deferrals get tracked instead of forgotten. One-shot report. | Stars: 194k | Updated | Claude CodeCodex | Verified | Install |
| ponytail-debtDietrichGebert/ponytailList every `shortcut:` comment (and older `ponytail:` ones) as a debt ledger. One-shot report, changes nothing. Use for "ponytail debt", "what did ponytail defer", "list the shortcuts", /ponytail-debt. | Stars: 194k | Updated | Claude CodeCodex | Verified | Install |