Enhance MCP development features and introduce skills management

- Added configuration options for requiring human approval before applying LLM-generated MCP patches.
- Updated Docker setup to include skills directory.
- Integrated skills management into the backend, allowing for procedural guides and skill matching.
- Refactored database initialization to apply Alembic migrations.
- Enhanced task approval process to handle MCP patch applications with optional approval.
- Introduced new schemas for skills and updated existing APIs to support skills functionality.

This commit lays the groundwork for improved agent capabilities and better management of MCP development processes.
This commit is contained in:
2026-06-14 22:27:24 +03:00
parent 6185b9b85a
commit 0375b20bb4
30 changed files with 1733 additions and 151 deletions

View File

@@ -29,8 +29,10 @@ from __future__ import annotations
import os
from dataclasses import dataclass
from typing import TYPE_CHECKING
from app.inventory_catalog import CatalogEntry
if TYPE_CHECKING:
from app.inventory_catalog import CatalogEntry
# catalog_key → how secrets are stored / resolved
SLOT_AUTH: dict[str, dict] = {

View File

@@ -0,0 +1,208 @@
"""Apply or roll back MCP patches after human approval (worker-side)."""
from __future__ import annotations
import logging
from collections.abc import Awaitable, Callable
from typing import Any
from app.config import settings
from app.services import mcp_workspace
logger = logging.getLogger(__name__)
EmitFn = Callable[[int, str, str, dict | None], Awaitable[None]]
async def backup_proposal_files(server: str, edits: list[dict]) -> dict[str, str]:
"""Snapshot current workspace content before a proposed patch is applied."""
backups: dict[str, str] = {}
for edit in edits:
path = edit.get("path") or edit.get("file")
if not path:
continue
try:
backups[path] = await mcp_workspace.read_source_file(server, path)
except OSError:
backups[path] = ""
return backups
async def restore_file_backups(server: str, backups: dict[str, str]) -> list[str]:
restored: list[str] = []
for path, content in backups.items():
await mcp_workspace.write_source_file(server, path, content)
restored.append(path)
return restored
async def apply_and_reload_mcp(server: str, edits: list[dict], manager) -> dict[str, Any]:
written = await mcp_workspace.apply_file_edits(server, edits)
if not written:
return {"ok": False, "error": "No files written"}
await manager.reload_server(server)
return {"ok": True, "files": written, "tools": mcp_workspace.list_registered_tools(server)}
async def queue_git_push_approval(
*,
task_id: int,
server: str,
written: list[str],
commit_msg: str,
summary: str | None,
emit: EmitFn,
) -> dict[str, Any] | None:
if not settings.gitea_token or not written or task_id <= 0:
return None
from app.database import SessionLocal
from app.models.enums import ApprovalStatus
from app.models.task import ApprovalRequest
async with SessionLocal() as db:
db.add(
ApprovalRequest(
task_id=task_id,
tool_name="mcp_git_push",
tool_args={
"server": server,
"files": written,
"commit_message": commit_msg,
},
risk=summary,
status=ApprovalStatus.pending,
)
)
await db.commit()
await emit(
task_id,
"approval",
f"MCP patch applied — approve to push {server} to Gitea",
{"server": server, "tool_name": "mcp_git_push", "files": written},
)
return {"queued": True, "tool_name": "mcp_git_push"}
async def maybe_retry_tool_after_patch(
*,
task_id: int,
retry_context: dict | None,
manager,
emit: EmitFn,
) -> dict[str, Any] | None:
if not retry_context:
return None
device = retry_context.get("device") or "device"
server = retry_context.get("server") or ""
tool = retry_context.get("tool") or ""
args = retry_context.get("args") or {}
if not tool:
return None
from app.agent.mcp_helpers import tool_result_ok
from app.agent.tool_events import tool_event_payload
from app.mcp_manager.recovery import call_tool_resilient
await emit(task_id, "log", f"Retrying {tool} after approved MCP patch…", {"server": server, "tool": tool})
res = await call_tool_resilient(manager, server, tool, args, task_id=task_id, emit=emit)
ok = tool_result_ok(res)
await emit(
task_id,
"tool_call",
f"{device} {tool} (after approved MCP patch)",
tool_event_payload(
device=device,
tool=tool,
args=args,
status="done",
ok=ok,
output=res.get("text", ""),
),
)
return {"ok": ok, "result": res}
async def execute_approved_patch(
*,
task_id: int,
tool_args: dict,
manager,
emit: EmitFn | None = None,
push: bool | None = None,
) -> dict[str, Any]:
"""Apply a previously approved MCP patch, reload the server, optionally push/retry."""
from app.services.events import publish_event
async def _emit(tid: int, kind: str, message: str, payload: dict | None = None) -> None:
if emit:
await emit(tid, kind, message, payload)
elif tid:
await publish_event(tid, kind, message, payload)
server = tool_args.get("server") or ""
edits = tool_args.get("files") or []
commit_msg = tool_args.get("commit_message") or "agentic-os: MCP patch"
summary = tool_args.get("summary")
retry_context = tool_args.get("retry_context")
if not server or not edits:
return {"ok": False, "error": "Invalid MCP patch approval payload"}
await _emit(
task_id,
"mcp_dev",
f"Applying approved MCP patch to {server}",
{"server": server, "phase": "apply", "files": [e.get("path") for e in edits]},
)
reload_result = await apply_and_reload_mcp(server, edits, manager)
if not reload_result.get("ok"):
await _emit(task_id, "error", reload_result.get("error", "MCP reload failed"), reload_result)
return reload_result
await _emit(
task_id,
"mcp_dev",
f"MCP '{server}' reloaded — {len(reload_result.get('tools') or [])} tools registered",
{"server": server, "phase": "reload_done", "tools": reload_result.get("tools")},
)
written = reload_result.get("files") or []
should_push = push if push is not None else settings.mcp_dev_auto_push
push_result: dict[str, Any] | None = None
if should_push and settings.gitea_token:
push_result = await mcp_workspace.commit_and_push(server, commit_msg, written)
await _emit(
task_id,
"mcp_dev",
f"Git push {'ok' if push_result.get('ok') else 'failed'}: {push_result.get('commit') or push_result.get('error')}",
{"server": server, "phase": "push", **push_result},
)
elif settings.gitea_token and written and task_id > 0:
push_result = await queue_git_push_approval(
task_id=task_id,
server=server,
written=written,
commit_msg=commit_msg,
summary=summary,
emit=_emit,
)
retry_result = await maybe_retry_tool_after_patch(
task_id=task_id,
retry_context=retry_context,
manager=manager,
emit=_emit,
)
return {
"ok": True,
"summary": summary,
"reload": reload_result,
"push": push_result,
"retry": retry_result,
"tools": reload_result.get("tools"),
}

View File

@@ -0,0 +1,264 @@
"""Agent skills — procedural guides loaded from skills/*/SKILL.md on disk.
Skills complement troubleshooting rules: rules route devices and supply short hints;
skills inject deeper procedural knowledge into triage and reasoning prompts.
"""
from __future__ import annotations
import logging
import os
import re
from dataclasses import dataclass
import yaml
from app.config import settings
from app.services.troubleshooting_rules import _text_has_term
logger = logging.getLogger(__name__)
_skill_cache: dict[str, dict] = {}
_skill_dir_mtime: float = -1.0
@dataclass
class MatchedSkill:
id: str
name: str
description: str
body: str
priority: int
score: int
def as_dict(self) -> dict:
return {
"id": self.id,
"name": self.name,
"description": self.description,
"priority": self.priority,
"score": self.score,
"body_chars": len(self.body),
}
def skills_dir() -> str:
return os.environ.get("SKILLS_PATH", settings.skills_path)
def invalidate_skills_cache() -> None:
global _skill_cache, _skill_dir_mtime
_skill_cache = {}
_skill_dir_mtime = -1.0
def _parse_skill_file(path: str) -> dict | None:
try:
with open(path, encoding="utf-8") as fh:
raw = fh.read()
except OSError:
return None
meta: dict = {}
body = raw.strip()
if raw.startswith("---"):
parts = raw.split("---", 2)
if len(parts) >= 3:
try:
parsed = yaml.safe_load(parts[1])
meta = parsed if isinstance(parsed, dict) else {}
except yaml.YAMLError as exc:
logger.warning("invalid skill frontmatter %s: %s", path, exc)
meta = {}
body = parts[2].strip()
folder = os.path.basename(os.path.dirname(path))
skill_id = str(meta.get("id") or folder)
return {
"id": skill_id,
"name": str(meta.get("name") or skill_id),
"description": str(meta.get("description") or ""),
"priority": int(meta.get("priority") or 0),
"match": meta.get("match") or {},
"rule_ids": list(meta.get("rule_ids") or []),
"body": body,
"path": path,
}
def _dir_latest_mtime(root: str) -> float:
latest = -1.0
if not os.path.isdir(root):
return latest
for dirpath, _dirnames, filenames in os.walk(root):
for name in filenames:
if name != "SKILL.md":
continue
try:
latest = max(latest, os.path.getmtime(os.path.join(dirpath, name)))
except OSError:
continue
return latest
def load_skills(*, force: bool = False) -> dict[str, dict]:
"""Return skill_id → skill record (metadata + body)."""
global _skill_cache, _skill_dir_mtime
root = skills_dir()
mtime = _dir_latest_mtime(root)
if not force and _skill_cache and mtime == _skill_dir_mtime:
return _skill_cache
skills: dict[str, dict] = {}
if os.path.isdir(root):
for entry in sorted(os.listdir(root)):
skill_path = os.path.join(root, entry, "SKILL.md")
if not os.path.isfile(skill_path):
continue
record = _parse_skill_file(skill_path)
if record:
skills[record["id"]] = record
_skill_cache = skills
_skill_dir_mtime = mtime
return skills
def list_skills_summary() -> list[dict]:
return [
{
"id": s["id"],
"name": s["name"],
"description": s["description"],
"priority": s["priority"],
"rule_ids": s.get("rule_ids") or [],
"match": s.get("match") or {},
"path_hint": f"skills/{s['id']}/SKILL.md",
}
for s in sorted(load_skills().values(), key=lambda x: (-x["priority"], x["id"]))
]
def get_skill(skill_id: str) -> dict | None:
return load_skills().get(skill_id)
def read_skill_raw(skill_id: str) -> str:
skill = get_skill(skill_id)
if not skill:
return ""
try:
with open(skill["path"], encoding="utf-8") as fh:
return fh.read()
except OSError:
return ""
def save_skill_raw(skill_id: str, content: str) -> dict:
skill = get_skill(skill_id)
if not skill:
raise ValueError(f"Unknown skill: {skill_id}")
# Validate frontmatter parses before writing.
if content.startswith("---"):
parts = content.split("---", 2)
if len(parts) >= 3:
parsed = yaml.safe_load(parts[1])
if not isinstance(parsed, dict):
raise ValueError("Skill frontmatter must be a YAML mapping")
path = skill["path"]
with open(path, "w", encoding="utf-8") as fh:
fh.write(content if content.endswith("\n") else content + "\n")
invalidate_skills_cache()
updated = get_skill(skill_id)
if not updated:
raise ValueError(f"Skill {skill_id} missing after save")
return updated
def _skill_match_score(text: str, skill: dict, *, matched_rule_id: str | None) -> int:
match = skill.get("match") or {}
keywords = match.get("keywords") or []
all_of = match.get("all_of") or []
if all_of and not all(_text_has_term(text, t) for t in all_of):
keyword_score = 0
else:
keyword_score = 0
for kw in keywords:
if _text_has_term(text, kw):
keyword_score += 2 if " " in str(kw).strip() else 1
if all_of:
keyword_score += len(all_of) * 2
rule_boost = 0
rule_ids = skill.get("rule_ids") or []
if matched_rule_id and matched_rule_id in rule_ids:
rule_boost = 10
return keyword_score + rule_boost
def _truncate_body(body: str, max_chars: int) -> str:
if max_chars <= 0 or len(body) <= max_chars:
return body
return body[: max_chars - 20].rstrip() + "\n\n… (truncated)"
def match_skills(
title: str,
issue: str,
*,
matched_rule_id: str | None = None,
max_skills: int | None = None,
) -> list[MatchedSkill]:
"""Return up to *max_skills* best-matching procedural guides."""
limit = max_skills if max_skills is not None else settings.skills_max_matched
max_body = settings.skills_max_body_chars
text = f"{title} {issue}".lower()
skills = load_skills()
scored: list[tuple[int, int, dict]] = []
for skill in skills.values():
score = _skill_match_score(text, skill, matched_rule_id=matched_rule_id)
if score <= 0:
continue
scored.append((score, skill.get("priority") or 0, skill))
scored.sort(key=lambda t: (-t[0], -t[1], t[2]["id"]))
out: list[MatchedSkill] = []
for score, _prio, skill in scored[:limit]:
out.append(
MatchedSkill(
id=skill["id"],
name=skill["name"],
description=skill["description"],
body=_truncate_body(skill["body"], max_body),
priority=skill.get("priority") or 0,
score=score,
)
)
return out
def skill_guidance_block(matched: list[MatchedSkill | dict] | None) -> str:
if not matched:
return ""
lines = ["\n\n--- Agent skills (procedural guides) ---"]
for item in matched:
if isinstance(item, dict):
name = item.get("name", "")
sid = item.get("id", "")
description = item.get("description", "")
body = item.get("body", "")
else:
name, sid, description, body = item.name, item.id, item.description, item.body
lines.append(f"\n### Skill: {name} ({sid})")
if description:
lines.append(description)
if body:
lines.append(body)
lines.append(
"\nFollow these guides when investigating. Rules define device scope; "
"skills provide domain-specific procedure and proven root causes."
)
return "\n".join(lines)