agents-tools
Guides LazyLLM function-call tools, ToolManager, agents, sandbox metadata, SkillManager, MCP, search, SQL, HTTP, and built-in tools.
- License 许可
- Apache-2.0
- 最近更新
- 2026-09-03
- Stars
- ★ 351
- Forks
- 27
01 前言 Front Matter
- name
- agents-tools
- license
- Apache 2.0
- metadata
- { "disco-role": "operating" }
- description
- Guides LazyLLM function-call tools, ToolManager, agents, sandbox metadata, SkillManager, MCP, search, SQL, HTTP, and built-in tools.
- disable-model-invocation
- true
02 资源文件 4
- 参考文档references/agent-tool-workflows.md2.9 KB
- 参考文档references/mcp-and-skills.md2.0 KB
- 参考文档references/troubleshooting.md2.1 KB
- 脚本scripts/tool_agent_smoke.py1.8 KB
03 SKILL.md 正文
LazyLLM Agents and Tools
Use this sub-skill for LazyLLM function-call tools, fc_register, ToolManager, tool metadata, sandbox flags, file parameters, built-in tools, SQL/HTTP/search tools, ReactAgent, ReWOOAgent, PlanAndSolveAgent, SkillManager, skills CLI, and MCP integration.
Start here when
- The task mentions
fc_register,lazyllm.tool,ToolManager,execute_in_sandbox, tool schemas,input_files_parm, oroutput_files_parm. - The user wants to build a React/ReWOO/PlanAndSolve agent or debug tool-calling traces.
- The task uses
lazyllm skillscommands,SkillManager, repo/user skill directories, or skills agent examples. - The task mentions MCP clients/servers,
deploy mcp_server, npm/npx, or external tool processes. - The task asks about built-in search, SQL manager, HTTP nodes, sandboxed execution, or PowerMemory-style examples.
Files to read
- agent-tool-workflows.md for deterministic tool registration and agent planning recipes.
- mcp-and-skills.md for SkillManager, skills CLI, and MCP boundaries.
- troubleshooting.md for schema, sandbox, MCP, provider, and tool-call errors.
- scripts/tool_agent_smoke.py for a no-network check of tool registration metadata.
Safe workflow
- Register/inspect tools without an LLM first. Use
fc_registerand verify metadata. - Choose agent class only after tool contracts are stable.
ReactAgent: iterative tool use with retries.ReWOOAgent: plan then solve with tool observations.PlanAndSolveAgent: planning and solving roles with optional separate LLMs.
- Separate schema from execution. Tool registration and schema checks are safe; executing shell, HTTP, SQL, search, or MCP tools can be side-effecting.
- Treat LLM calls as optional. Agent reasoning needs an LLM/provider or local model configured by model-deployment.
- Treat MCP as external-process work. Do not run npm/npx/MCP servers unless the user approves the command and environment.
Verified signatures to remember
fc_register(f, *, rewrite_func=None, **kwargs)registers callable tools and metadata.ToolManager(tools, return_trace=False, sandbox=None)manages callable or named tools.SkillManager(dir=None, skills=None, max_skill_md_bytes=None, fs=None, sandbox=None)loads LazyLLM skills.ReactAgent(llm, tools=None, max_retries=5, return_trace=False, prompt=None, stream=False, return_last_tool_calls=False, skills=None, desc='', workspace=None, sandbox=None, force_summarize=False, force_summarize_context='', keep_full_turns=0, fs=None, skills_dir=None, enable_builtin_tools=True, extra_stop_condition=None, on_max_retries=None).ReWOOAgentandPlanAndSolveAgentaccept optionalplan_llm,solve_llm,tools, tracing, streaming, skills, sandbox, fs, and built-in tool controls.
Tool metadata facts
LazyLLM tests verify:
execute_in_sandboxdefaults to true and can be disabled.input_files_parm,output_files_parm, and staticoutput_filesmetadata are preserved.- unsupported
fc_registerparameters raise assertions. - writer/search/SQL/HTTP tools have local contract tests and should be validated before real external calls.
Handoff checklist
When completing an agent/tool task, provide:
- tool names, docstrings, parameter schema expectations, and sandbox/file metadata,
- whether the tool was registered and inspected without LLM calls,
- selected agent class and LLM/backend status,
- explicit side-effect classification for shell, HTTP, SQL, search, MCP, or remote services,
- local smoke command and optional/unrun external verification.
04 同仓库更多 Skills
passnet-triton-opt
Make a single PassNet Triton kernel fast AND numerically faithful: performance model (when a replacement can win at all), block/grid/warp tuning, autotune policy, launch overhead, and per-op numeric recipes to pass the dtype baseline tolerances. Use when a pass matches and is correct but speedup below expected, or when correctness fails by small numeric margins.
passnet-solve
END-TO-END playbook for solving one PassNet sample: analyze the computation graphs, decide the optimization strategy (what to fuse, what to replace, what to leave alone), drive the iteration loop, and maximize the sample score. This is the ENTRY skill — invoke it first for any PassNet optimization task; it tells you when to use passnet-pattern-fusion, passnet-triton-opt and passnet-feedback.
passnet-skill
PassNet GPU kernel optimization via compiler passes. Design and implement Triton-based optimization passes, create pass files under ./pass_dir/, self-evaluate with pass_evaluator, and iterate to maximize GPU speedup.
passnet-pattern-fusion
Author PassNet pass files: write patterns that actually MATCH the FX graph, pick fusion regions (which ops to absorb into one Triton kernel), and structure multi-pass submissions with the mandatory shared-dispatch architecture. Use when creating or fixing pattern()/replacement_args()/replacement_func() files, when a pass "failed to match", or when deciding how to fuse multiple kernels/ops.
passnet-orchestrate
Round/batch planning and decision support for PassNet graph optimization. Use for multi-sample triage, worker allocation, eval-budget policy, and post-eval keep/revert decisions. For solving one concrete sample end to end, use passnet-solve as the entry skill; consult this skill only when you need broader planning or a specific decision gate.
passnet-feedback
Fast, deterministic feedback for PassNet work: pre-flight pattern/match verification WITHOUT burning a GPU evaluation, per-node bottleneck analysis of a sample's graphs, and eval-log parsing into per-variant status + estimated score + failure classification. Use BEFORE every GPU evaluation (check_pattern), at the START of a sample (analyze_graph), and AFTER every evaluation (parse_eval_log).