docs(understand-anything): record set-and-forget fallback chain upgrade for enrich
§9 of codegraph-integration note now documents the 2026-06-03 15:20 upgrade: - enrich defaults to callWithFallback() chain (current model → fallback_providers) - skips codex_responses transport providers automatically - resolves __default._providerModel to endpoint-served name (not user alias) - strips <think>...</think> blocks before JSON.parse - explicit --provider still supported (skip chain, useful for fast mode) Verified on 5 git repos (contract management / gongju / llm-intelligence / user-system / supply-intelligence) — 13/13 OK, 0 fallback rotation needed, all from minimax-m3/MiniMax-M3 (current agent model). Docs-only commit; sibling subagent's parallel code changes are NOT included.
This commit is contained in:
Hermes Agent
@@ -226,3 +226,63 @@ import('/home/long/.understand-anything/repo/understand-anything-plugin/packages
|
||||
- 任何本项目接入 codegraph workspace 的规则变更，先在 `/home/long/project/CODEGRAPH_WORKSPACE_USAGE.md` 改，再回本文件
|
||||
- 如果 `codegraph-workspace` wrapper 被重装或升级，必须**重做第 6 节的 patch**（备份 `.bak-2026-06-03` 可作参考）
|
||||
- 任何本项目接入 codegraph workspace 的规则变更，先在 `/home/long/project/CODEGRAPH_WORKSPACE_USAGE.md` 改，再回本文件
|
||||
|
||||
## 9. understand-anything LLM enrichment (online, 2026-06-03 14:53-15:02)
|
||||
|
||||
### 状态
|
||||
|
||||
- **16/16 git repo + gongju** 跑通 LLM enrich
|
||||
- **总计 ~360 个 file 节点** patch 了 `summary` / `tags` / `complexity`
|
||||
- **set-and-forget 模式**：`callWithFallback` 自动跟 agent 当前模型 + 撞 quota 自动转 fallback_providers 链
|
||||
|
||||
### 9.1 set-and-forget 升级 (15:20-15:25)
|
||||
|
||||
**新行为**：`enrich` 默认走 fallback chain
|
||||
|
||||
1. **当前 agent 模型** (`minimax-m3/MiniMax-M3`) — 来自 `model:` 段
|
||||
2. `openai-zhongzhuan/gpt-5.4` — `fallback_providers:` 第一项
|
||||
3. `a7m-kimi/kimi-k2.6` — 第二项
|
||||
4. `deepseek-official/deepseek-chat` — 最后兜底
|
||||
|
||||
**跳过 `transport: codex_responses` 的 provider**（openai-zhongzhuan）— 该 transport 走 `/v1/responses` 端点，与 `/v1/chat/completions` 不兼容。
|
||||
|
||||
**`__default._providerModel` 解析**：优先用 provider 的 `default_model` 字段（端点真实可用的 model name），不是 user-config `default:` 别名。例如 `minimax-m3` provider 的 `default_model: MiniMax-M3` 在 `mimimax.cn/v1/chat/completions` 端点上对应 `grok-4.20-reasoning`，`__default._providerModel` 正确解析为后者。
|
||||
|
||||
**Strip `<think>...</think>` 块**：minimax-m3/MiniMax-M3 输出含 thinking 块会污染 JSON 解析 — 自动 regex 移除。
|
||||
|
||||
**显式 `--provider` 不走 fallback 链**：单 provider 调用，方便调试和稳定输出。
|
||||
|
||||
### 用法对比
|
||||
|
||||
| 场景 | 命令 | 行为 |
|
||||
| -------------------------- | -------------------------------------------------------------------------------------- | ------------------------------------------------------------ |
|
||||
| **set-and-forget**（默认） | `understand-anything-enrich <repo>` | fallback chain (minimax-m3 → gpt-5.4 → kimi-k2.6 → deepseek) |
|
||||
| **强制当前模型** | `understand-anything-enrich <repo> --provider minimax-m3` | 不 fallback |
|
||||
| **大 repo 快跑** | `understand-anything-enrich <repo> --provider deepseek-official --model deepseek-chat` | 1.3s/file |
|
||||
| **批处理默认** | `understand-anything-enrich-batch <ws>` | 每个 repo 走 fallback chain |
|
||||
| **批处理快跑** | `understand-anything-enrich-batch <ws> --provider deepseek-official` | 整个 batch 用 deepseek |
|
||||
|
||||
### 工具
|
||||
|
||||
- `~/.local/bin/understand-anything-enrich` (9.1 KB Node ESM, v2)
|
||||
- `~/.local/bin/ua-provider-config.mjs` (5.3 KB, 加 `callWithFallback`)
|
||||
- `~/.local/bin/understand-anything-enrich-batch` (2.7 KB bash wrapper)
|
||||
- `~/.hermes/skills/software-development/understand-anything-enrich/SKILL.md` (skill, 8 pitfalls + cheat sheet)
|
||||
|
||||
### 验证
|
||||
|
||||
- 抽检 sub2api-cn-relay-manager / gongju / contract management / llm-intelligence / user-system / supply-intelligence — Dockerfile / .env.example / hooks summary 全部准确
|
||||
- quality 合理：complex=complex 给 simplify-ignore.sh（多 file 引用），simple 给 .env.example（10 行）
|
||||
- 5/5 跑通 fallback chain，0 失败（minimax-m3 当前模型直接出活）
|
||||
- big repo (新api/公文助手/区块链商城/立交桥) 仍 5000-50000 placeholder — honest trade-off，360 LLM 调用 7 分钟可接受
|
||||
|
||||
### 完整报告
|
||||
|
||||
详见 `~/.hermes/docs/2026-06-03-UNDERSTAND-ANYTHING-ENRICH-BATCH.md`
|
||||
|
||||
### 恢复
|
||||
|
||||
```bash
|
||||
cp <repo>/.understand-anything/knowledge-graph.json.bak-2026-06-03 \
|
||||
<repo>/.understand-anything/knowledge-graph.json
|
||||
```
|
||||
|
||||
Reference in New Issue
Block a user