The ask
I asked for a Round 2 AI & Automation entry that explains my current harnesses, integrations, agent memory and working practices to someone meeting me for the first time. The entry needed to retain every project from the existing page and use only the round’s permitted sources.
What changed
- Published my assigned child page, titled AI & Automation by ChatGPT with Sol 6.1 High.
- Covered ChatGPT Chat, Work and Codex; Claude Chat, Cowork and Code; OpenCode with different models; and OpenClaw and Hermes on my Linux lab host.
- Explained Composio, PowerShell and bash, SSH administration, agent backups, and the distinct memory layers in OpenClaw and Hermes.
- Retained the MCP servers, agent skills, automation pipelines, self-hosted lab and vendor assessment engine. Connected them to the Lab Notes, Harness Scoreboard and Round 1 comparison process.
- Used core blocks, theme color and spacing presets, ordered headings and descriptive section links. The layout stacks its columns on a phone.
How
I used SSH and WP-CLI through the site’s approved operations scripts. I pulled the repository and read all four rule files before starting the registered run. Research used the original AI & Automation page, the permitted Obsidian notes, site documentation, earlier process Lab Notes and Round 1’s brief and model roster. I interviewed both agents and inspected their configurations only through the approved redacted agent tool.
I took a fresh snapshot before the entry change, verified the saved title, parent, slug, published status and HTTP 200 response, then synced one typed content commit. I checked the rendered page at desktop and phone widths: one H1, ordered H2 and H3 headings, valid section targets, stacked columns and no horizontal content overflow. The live AI & Automation source page’s checksum remained unchanged.
What worked, what didn’t
The notes provided enough evidence to explain the architecture without publishing operational details. The interviews helped distinguish durable preferences from the conversation behind a decision. I asked two questions of each agent, summarized below.
Questions asked of OpenClaw
- What is your role on my Linux lab host, how do you use Composio and shell tools, and how do your memory layers, retrieval, compaction and general backup principle fit together?
- How do live context, readable notes, QMD, clawmem and lossless-claw differ when remembering a preference versus recovering an earlier decision, without treating summaries or retrieval as infallible?
Questions asked of Hermes
- What is your role on my Linux lab host, how do you use Composio and shell tools, and how do memory, session search, fact search and backups work? Which claims describe configuration versus demonstrated behavior?
- Which memory layer holds a preference versus the discussion behind a decision, why do memory databases and custom code need to travel together, and how does feedback capability differ from proof that ratings are used?
OpenClaw’s first answer called itself a broader technical orchestrator, while its agent note describes an execution worker for smaller tasks. I followed the note and kept specialized broader work separate. Its first answer also described retrieval as lossless; I used the memory architecture note and follow-up to describe linked summaries and archived messages without a guarantee of perfect recall. The follow-up described QMD mainly as full-text search; the authoritative note also documents semantic retrieval, which I retained.
Hermes confirmed the distinction between session history and structured facts. I described feedback scores as a supported capability, without claiming that ratings have demonstrably improved results. I avoided a fixed current default model because the notes and redacted configuration describe different choices. Round 1’s records support comparing models, but do not establish a graded winner, so I claimed none.
I deliberately left out operational addresses, service links, file locations, identities, credentials, network details, backup destinations, timing, retention, recovery instructions and open issues. I also omitted dated memory sizes, changing job counts and business-planning material that did not help explain the active setup. I checked the saved entry and rendered text against the complete never-publish list. The entry contains no lab service URLs, and its links only navigate within the page.
The initial SSH attempt could not reach the host from the local sandbox; the authorized connection outside that sandbox worked. No site change broke, and no rollback was needed. I used the Round 1 brief and model metadata without reading contestants’ entries or their bake-off Lab Notes. I created only my assigned entry and this Lab Note; the required scripts handle their own exports and generated tracking pages.
Change record
| Harness | chatgpt_sol61high |
| Date | 2026-10-10 22:57 |
| Latest snapshot | 20261010-225656 |
| Undo this session | ops/rollback.sh --git 497c81b |
Commits