The ask
Set up Round 2: every harness writes the AI & Automation page about my harnesses and how I work, researching my Obsidian notes and interviewing my OpenClaw and Hermes agents, without publishing anything that poses a security risk.
What changed
- New Round 2 page under Harness Bake-off, with the task, kickoff prompt and scoring in the repo. The best entry replaces the live AI & Automation page.
- Every Round 2 entry must name its model and effort setting; the start script refuses to run without one.
- A read-only interview tool lets contestants ask OpenClaw and Hermes up to four questions each, plus a redacted view of each agent’s configuration. Both strip hosts, addresses, paths, IDs and secrets, and contestants never touch the agents’ credential files.
- A never-publish list in the task: anything on it disqualifies an entry.
How
Claude (Cowork) over SSH: wrote the task and tools, created the round page with WP-CLI, tested an interview with each agent and checked both config views for leaks.
What worked, what didn’t
The first config view leaked a chat channel ID; the redaction rules now catch those. Agent answers can still include details that are fine for me but not for a public page, which is why the task’s never-publish list and a final self-check stay mandatory.
Change record
| Harness | claude |
| Date | 2026-10-10 22:27 |
| Latest snapshot | 20261010-222330 |
| Undo this session | ops/rollback.sh --git 5c84930 |
Commits