12 KiB
12 KiB
type, title, description, tags
| type | title | description | tags | ||||||
|---|---|---|---|---|---|---|---|---|---|
| Reference | Theme 3 · AI Agent tools (Hermes / Codex / MiniMax code) | Synthesis page for Robin's daily-driver AI agent stack — Hermes Agent on WSL, MiniMax code on the desktop, Codex CLI fallback, OpenWiki as cross-agent memory substrate, and supporting toolchain (RustDesk, Gitea, Wiki.js, RAGFlow). |
|
Theme 3 · AI Agent tools (Hermes / Codex / MiniMax code)
This is the synthesis page for AI Agent 工具 — Robin's daily-driver stack for autonomous work. The stack has three primary agent surfaces (Hermes Agent on WSL, MiniMax code on the desktop, Codex CLI), with OpenWiki as the cross-agent memory substrate and a local-fleet support layer (RustDesk remote, Gitea wiki, RAGFlow KB cross-query).
What this theme covers
- Primary agent surfaces: Hermes Agent v0.19 "Quicksilver" (→ v0.20.0 in 2026-08-03), MiniMax code (desktop), Codex CLI (fallback).
- Inference provider: MiniMax-M3 (1M context multimodal) → MiniMax-H3 frontier model (2026-08-03 launch cluster).
- Memory substrate: OpenWiki (LangChain open-source, local markdown wiki) + Gitea (public read-only mirror).
- Cross-agent tooling: 4 OpenWiki bridge skills (MiniMax code reads OpenWiki); literature-deep-read SOP; sales-ammunition query skill; 4-connector cron pipeline.
- Remote / access: RustDesk self-hosted hbbs + hbbr for ITX host remote control (2026-08-05).
Agent stack architecture
flowchart TD
User[Robin · desktop / mobile]
subgraph DesktopWSL[WSL · primary host]
Hermes[Hermes Agent v0.19 Quicksilver]
Cron1["cron 02:00 (openwiki-update.sh)"]
end
subgraph DesktopPC[Windows desktop · secondary]
MiniMaxCode["MiniMax code (+ 4 bridge skills)"]
CodexCLI[Codex CLI fallback]
end
subgraph Remote[ITX 24h host · future]
ITXHermes["Hermes Agent (on Ubuntu 22.04)"]
end
subgraph Memory[Memory substrate]
OpenWiki[("OpenWiki (local markdown wiki)")]
Gitea["Gitea public mirror (10.10.10.254:3080)"]
end
subgraph KB[Knowledge backends · NAS]
RAGFlowKB["RAGFlow (18 KB · 32709 docs)"]
NocoDBDB["NocoDB (13 lit tables)"]
SeaFileStore["SeaFile (4 repos · 77 GB)"]
WikiJS["Wiki.js (1500+ pages)"]
end
subgraph LLM[Local inference · see theme 04]
MiniMaxM3[MiniMax-M3 1M context]
Ollama[Ollama · qwen2.5vl:7b]
Reranker[bge-reranker]
end
User -->|primary chat| Hermes
User -->|desktop tasks| MiniMaxCode
User -->|fallback| CodexCLI
Hermes -->|reads / writes| OpenWiki
MiniMaxCode -->|bridge skills read| OpenWiki
Cron1 -->|HN + Tavily + git-repo ingest| OpenWiki
Cron1 -->|M3 agent synthesis| OpenWiki
OpenWiki -->|commit + push| Gitea
Hermes -->|cross-query| RAGFlowKB
Hermes -->|metadata| NocoDBDB
Hermes -->|PDF retrieval| SeaFileStore
Hermes -->|published deep reads| WikiJS
ITXHermes -.future.-> OpenWiki
Hermes -.inference.-> MiniMaxM3
RAGFlowKB -.embedding.-> Ollama
RAGFlowKB -.rerank.-> Reranker
The architecture shows three agent surfaces sharing the OpenWiki memory substrate, which mirrors to a public Gitea repo for cross-device reading. The four NAS-backed knowledge systems (RAGFlow / NocoDB / SeaFile / Wiki.js) are queried by Hermes on demand and route evidence into the wiki synthesis layer.
Primary agent surfaces
Hermes Agent (WSL primary)
- v0.19 "Quicksilver" release (2026-07-25) — first-token delay 4.3s → 0.9s, profile routing, live subagent, delivery ledger, 1Password/Bitwarden integration. Tracks on
/themes.mdhermes-agent-v0-19-quicksilverrow. - v0.19.1 patch (2026-08-02) — NousResearch issues/77000 GitHub bug report ("gpt-5.6-luna + opencode-go"). Confirms patch-version activity on the v0.19.x track between v0.19 and v0.20.0.
- v0.20.0 release (2026-08-03, tag
v2026.8.3) — next major version after v0.19 Quicksilver. HN item49159200(1pt day-of, but the GitHub release is the authoritative primary source). Last-seen refresh applied tohermes-agent-v0-19-quicksilverrow description; v0.20.0 features not yet documented in source-page synthesis. configuration.md(2026-07-26) — documents iteration-budget behavior change (April 2026: removed premature 70%/90% warnings; now single wrap-up at 500/500 with grace call).features/api-server.md(2026-07-27) — documents new/v1/runs/{run_id}/stop+/approvalendpoints with MiniMax-M3 as the documented default model. Updated 2026-07-30 with a per-profileAPI_SERVER_KEYbreaking-change warning and a runs API for long-form sessions.- MCP 2026-07-28 protocol revision (adjacent watchlist) — finalized not-backward-compatible revision. Relevant to OpenWiki's MCP server architecture. Folded into
hermes-agent-v0-19-quicksilverevidence column.
MiniMax code (Windows desktop)
- 4 OpenWiki bridge skills — allow MiniMax code to cross-agent query the OpenWiki memory substrate (reads
/sources/literature.mdon "what did Robin recently read", reads/sources/agents.mdon "who are the distributors", reads/sources/wikijs.mdon "is there an X topic in Wiki.js", reads/sources/sales-ammunition.mdon "ClinoReactor sales ammunition"). See/sources/OpenWiki速查手册.md§5. - Inference provider: MiniMax-M3 (1M context multimodal) — daily-driver model for cross-agent tasks. See
/themes/04-local-llm.mdfor the broader local-LLM fleet. - Adjacent: openclaw-bridge v1.5 — MiniMax code cross-agent index, sits on top of the 4 OpenWiki bridge skills.
Codex CLI (fallback)
- Used only when Hermes or MiniMax code is unavailable — not the primary surface. Codex CLI itself has 42 HN hits in the 2026-08-03 pull (
49159200Hermes + MiniMax-H3 cluster is the only workflow-relevant signal), all of which are surface-level generic agent wrappers/TUIs/skills products (e.g.49107749Agent-Manager Tmux TUI 80 pts;49110007Aggregated Claude Code sessions 1 pt). Codex CLI itself is not the canonical Robin agent — it is the fallback.
Cross-agent memory substrate: OpenWiki
- OpenWiki = LangChain 7/10 open-source "personal brain" framework. Local markdown wiki rooted at
~/.openwiki/wiki, mirrored to Gitea publichttp://10.10.10.254:3080/Robin/openwiki-wiki. Cron pipeline: every night 02:00openwiki-update.sh(HN + Tavily + git-repo + M3 agent synthesis) → commit → push to Gitea. Every 02:30openwiki-update-sources.sh(NocoDB + SeaFile + Wiki.js + RAGFlow health checks) → commit → push. - Five canonical synthesis pages —
/quickstart.md,/themes.md,/commitments.md,/open-questions.md,/personal-logistics.md. Five detailed theme pages under/themes/. - Evidence layer — 21 source pages under
/sources/, one per active connector or subsystem. - Activation manual:
/README-ACTIVATION.md(2026-07-27 update). 8 sections covering TL;DR / connectors not yet on / daily operations / what Robin manually does / what Robin never touches / cross-tool cheat sheet / today's real system status / changes from 7-26 / next steps. - Cheat sheet:
/sources/OpenWiki速查手册.md(2026-07-27 one-shot reference for talking to colleagues / boss / agents about what OpenWiki is).
Supporting toolchain (AI-side)
Literature-deep-read (LDR) skill
- Master template:
/sources/ldr-report-template.md— 8 铁律 + 3 trigger modes (精简 4 步 vs 完整 8 步 vs 自动) + 5-section report template (做什么 / 3 发现 / 4 维产品对接 / 局限 / 转发语+落款+PDF). Total report length 800–1500 characters. - SOP lessons:
/sources/2026-07-27-4-27-sop-vs-implementation.md— 9-step SOP vs 7-27 凭印象实操 comparison. PyMuPDF xref dedup + 30KB filter + p1 logo 剔除. NocoDB v1 + JSON body (范本 v2 fails 3x). SeaFile?dl=1suffix required. 3-frame SeaFile image 嵌入 markdown. - Toolchain打通战报:
/memory-2026-07-16-tools-final.md— 2026-07-16 OpenWiki + Gitea + NocoDB + RAGFlow + Wiki.js + MiniMax code 端到端打通. 9 链路.
Remote access (RustDesk)
- Self-hosted hbbs + hbbr:
/sources/rustdesk-sop.md— NAS Unraid container, OpenWrt frpc → Tencent 轻量云 frps, ITX rustdesk 1.4.9 client. Permanent passwordLdw811223!, ITX ID263 739 134. Supports Robin's planned 24h ITX host build (/commitments.mditx-24h-host-build-2026-q3). - ITX 24h host build plan — mATX + i5-12400 + DDR4-2666 32G + dual-NIC (板载 Realtek 1Gb + Intel PRO/1000 GT) NAS isolation. Budget ¥2,985.82, 10/11 parts arrived. Plan: Ubuntu Server 22.04 LTS install + 双网卡静态 IP + Hermes Agent install + 24h soak test.
Public mirror (Gitea)
- Gitea repo:
Robin/openwiki-wikipublic,http://10.10.10.254:3080. Robin can read on phone / desktop browser without auth. Commit + push happens automatically after each OpenWiki cron run.
Adjacent workflow signals (HN / Tavily)
agent-memory-architecturetheme row on/themes.md— clusters Karpathy LLM Wiki + Google OKF + MCP + verginglabs.com "A Markdown wiki outscored every AI agent memory product we benchmarked" (HN49156055, 2 pts, watchlist). 15+ distinct sources on the shared-agent-memory pattern. OpenWiki aligns with this lane (Open Knowledge Format front-matter is the structural backbone).- Tavily / Hermes / MiniMax-M3 lane — Tavily answer text on this lane is partially fabricated (e.g. Tavily claimed a "release date of June 1, 2026" for Hermes v0.19 Quicksilver; the actual release is 2026-07-25 per Nous Research). Trust underlying URLs + GitHub release tags; ignore Tavily answer text.
Cross-cutting relationships
- Depends on
/themes/04-local-llm.md— every Hermes / MiniMax code call is served by local-LLM inference (Ollama / bge-reranker / RAGFlow retrieval). When the desktop AI service is down (7997 / 8090 / 11434), cross-agent RAGFlow queries fall back to metadata-only. - Writes to
/themes/05-robin-research.md— the M3 agent synthesis step is what routes evidence into the personal-research theme (deep reads, vendor product signals, demo plans). - Touches
/themes/01-organoid-equipment.mdand/themes/02-bioprinting.md** through the LDR pipeline — every deep read on CelVivo / 3Brain / Cellink / Femtobiomed goes through the LDR master template + SOP.
Backlog for this theme
- Codex CLI adoption question — currently the fallback agent; no active usage notes. Decide whether to expand Codex into a second-primary surface or keep as fallback-only.
- Hermes v0.20.0 feature documentation — only the GitHub release tag is in evidence; the changelog has not been deep-read. Add to next Tavily pull's
Hermes Agent MiniMax M3 changelog 2026query. - OpenWiki bridge skill expansion — currently 4 skills. Decide whether to add cross-KB retrieval skills (RAGFlow live query) now that embedding service is intermittent.
- MiniMax-H3 adoption — HN launch cluster 2026-08-03 (ComfyUI 167pts + HuggingFace 8pts + LMSYS Twitter 2pts); decide whether H3 replaces M3 as the daily-driver inference.
What to read next
| Question | Go to |
|---|---|
| What is OpenWiki and how do I use it? | /sources/OpenWiki速查手册.md |
| What connectors does OpenWiki have? | /README-ACTIVATION.md |
| What is the LDR report format? | /sources/ldr-report-template.md |
| What are the LDR SOP lessons? | /sources/2026-07-27-4-27-sop-vs-implementation.md |
| How is RustDesk configured? | /sources/rustdesk-sop.md |
| What did Robin do on 2026-07-16 (toolchain通)? | /memory-2026-07-16-tools-final.md |
| What is the ITX 24h host build? | /commitments.md itx-24h-host-build-2026-q3 |