Wire mcp_connect/list/call, resume_*, and demote Jev in author docs. Offline goldens cover MCP client, resume, and small-llm/jev-alias route. Co-authored-by: Cursor <cursoragent@cursor.com>
3.9 KiB
3.9 KiB
Marqdo AI 栈统一开发方案(llm + agent)
| Status | Active |
| Date | 2026-09-24 |
| Source | doc/next/002–005 · ext-llm.md · ext-agent.md |
1. 一句话
Document → Prompt → LLM → Agent → Skill → Policy → Executable Knowledge
大模型负责探索,小模型(router_model / 任意 llm 句柄)负责决策,Marqdo 负责执行;经验编译进 .mq.md 后推理成本下降。
2. 分层(勿混)
| 层 | 职责 | 落点 |
|---|---|---|
| Intelligence primitive | ask / stream / collect;结果对象 |
ext/ai/llm |
| Agent runtime | run / step / plan / route |
ext/ai/agent |
| Memory / compile | episode → skill → policy | plugins/agent(ABI;不进 src/host) |
| Transport | HTTP / SSE / provider | ext/ai/llm/openai · ollama;用户不可见 |
硬规则:ext/** 不调 host_*;agent/OKF/路由热路径只进 plugins/agent。
3. LLM(005)— 已/待
| 项 | 状态 |
|---|---|
语义 API:ask / stream / collect(complete 降为兼容原语) |
done |
create 工厂;handle 带 backend |
done |
LLMResult:text / model / finish / backend / usage / tokens |
done |
Named models(llm.fast / llm.reasoning) |
done |
Prompt document:prompt / prompt_load + type: prompt 正文 |
done(无参 ask/stream → sys.module_source + prompt_body) |
拆 ext/ai/llm/openai.mq.md + ollama 后端 |
done |
| tool_calls 完整 Result | done(openai.tool_calls_from + LLMResult.tool_calls;llm-tool-calls-offline) |
Transport → 可选 ABI 插件 (plugins/llm) |
done |
Artifact Metadata Binding(${env.*} / secret · sys.meta) |
done(Phase 1 Metadata-only + Phase 2 scope lift;binding.md) |
作者面优先:
**answer = > llm.ask prompt="What is Marqdo?"**
**events = > model.stream prompt=`p`**
**text = > llm.collect events=`events`**
**fast = > llm.fast**
**prompt = > llm.prompt path="prompts/task.md"**
4. Agent(002–004)— 已/待
| 项 | 状态 |
|---|---|
Skill Compilation:episode → maybe_learn → llm_free |
done |
Adaptive Routing:marqdo | small-llm | llm | auto(jev→small-llm 别名) |
done |
Policy Compilation → policies/router.mq.md |
done |
run 串联 route → execute → record → learn |
done |
Agent 默认走 model.ask + Result.usage 记成本 |
done(model_turn) |
| Prompt artifact / document-as-prompt | done(经 llm path= / prompt) |
零 LLM 执行(policy+skill → llm_calls=0) |
done |
plan 不覆盖已有 llm_free resource |
done |
| 真 MCP client(stdio) / resume checkpoint / 小模型路由请求面 | done(live 路由仍可选 skip) |
5. 联调顺序
- LLM Result → Agent metrics — done
- Named handles — done
- Prompt document — done(
type: prompt剥 frontmatter;无参ask/stream经sys.module_source) - Backend split — done
- Execution compilation — done(
agent-zero-llm金样 + harness) - Metadata Binding — done(Phase 1 + Phase 2 scope lift;
llm-meta-offline·meta-lift)
6. 验收金样
- Agent:
tests/ext/agent-v2-learn.mq.md·agent-p4-route.mq.md·agent-zero-llm.mq.md·agent-plan-preserve.mq.md·agent-mcp-client.mq.md·agent-resume.mq.md·agent-small-llm-route.mq.md·scripts/agent-harness.sh - LLM:
tests/ext/llm-ask-offline.mq.md·llm-stream-offline.mq.md·llm-ctor-offline.mq.md·llm-named-offline.mq.md·llm-meta-offline.mq.md·llm-entry-prompt-offline.mq.md·llm-tool-calls-offline.mq.md - Binding:
tests/structure/meta-binding.mq.md·meta-lift.mq.md