The ask
Bake-off Round 2: rewrite the AI & Automation page as my entry, covering the harnesses I use, how they connect, how the self-hosted agents remember, how the site’s change process works, and everything already on the live page. Child of the Round 2 page, core blocks and theme presets only, first-person voice, nothing from the never-publish list.
What changed
- Created entry page 493, “AI & Automation by OpenCode with Qwen 3.8 Flash Medium”, child of the Round 2 page, published
- New sections: harnesses (OpenClaw, Hermes, ChatGPT, Claude, OpenCode), connections (Composio, agent mailboxes, shell, SSH to the lab host, shared vault), memory design for both self-hosted agents plus how their memory is backed up, and the one-rulebook working process with links to Lab Notes, the Scoreboard and Round 1
- Kept all five existing projects: MCP servers, agent skills, automation pipelines, the self-hosted agent lab and the vendor assessment engine, tightened and framed into the new structure
How
SSH into openclaw01, WP-CLI through ops/wp.sh, one snapshot (20261011-002136) before the single content change, verified with curl (HTTP 200) and a saved-content scan, then ops/sync.sh with a typed commit (baf1299). Research sources: the live page and site content, AGENTS.md, GUIDELINES.md, BAKEOFF.md, the round’s TASK.md, the allowed Obsidian notes on the agents, memory architectures, Composio, protocols and the lab host’s baseline, backups and restore docs, plus ops/agents.sh for the two agent interviews and their redacted configs.
What worked, what didn’t
Questions I asked each agent (all through ops/agents.sh, 4 of 4 per agent, some asked in the earlier interrupted session under this same contestant ID before the run resumed):
- OpenClaw: how its memory is layered and how a search across layers works for old material (twice, phrased for a site visitor); how model choice works across the many models it runs and how agents split tasks; and a fourth question on how its own state is backed up, which the budget refused (“Limit reached: 4 questions per round”) so I used the backup notes instead.
- Hermes: how memory survives between sessions, what happens at session end and how a session search is answered (twice); what scheduled jobs it runs and what protects its memory if the host dies; and job categories plus how it proves a backup truly completed.
One disagreement: the OpenClaw agent described its memory as three layers (live context, notes, compacted summaries) and left out the search sidecar and durable knowledge vault. The Obsidian architecture note documents five cooperating layers, and per the rules the notes outrank the answer, so the page follows the notes.
Deliberately left out: anything from the never-publish list. I checked the saved page content against it, specifically: no hostnames other than openclaw01 and no tailnet or LAN names, no IPs, ports or host URLs (the hero image is saved with a site-relative path so my content adds none), no file paths, usernames, email addresses or agent profile names, no backup destinations, schedules, retention numbers or restore steps (only the general principles: markers, checksums, integrity checks, drills), no open issues or hardening gaps, no client names beyond what the site already carries. Also left out: exact model lists per agent, messaging platform names, internal counts, the faith notes pipeline, teammate names, and the parked automation venture, none of which the page needs.
Process notes: the start script found my run already open from an interrupted session and resumed it, which the script supports (“RUN ALREADY OPEN”). PowerShell mangled nested quotes for the MODEL argument, solved with single quotes on the remote line. The vendor engine and MCP sections reuse the live page’s existing text so no project was dropped or reworded into new claims.
Change record
| Harness | opencode_qwen38flashmedium |
| Date | 2026-10-11 00:28 |
| Latest snapshot | 20261011-002826 |
| Undo this session | ops/rollback.sh --git d37d9ce |
Commits