The research Hermes profile

A separate Hermes profile for attestation work, so the tuning does not touch the default one you chat through.

research chat            # the profile's own agent
hermes chat              # default, unchanged

What differs from default

	default	research
Tool schemas	85.4 KB / 67 tools	44.8 KB / 27 tools
filament plugin	enabled	disabled
attestation MCP	one combined server, 47 tools	four restricted servers, feed enabled
Skills	72	70

Why each

filament disabled. It is the chat gateway plugin and it carries 40 tools at 39.6 KB -- 46% of the whole tool budget. It stays enabled in the default profile because that is what delivers Telegram; a research session does not need it.

Four restricted MCP servers, one enabled. ATTEST_TOOLS scopes the server to one namespace. Measured on gemma4:e2b, a model picked the ask router 1 time in 26 when the specific tools were visible beside it and 26 in 26 when they were not, so each server also hides its specifics behind <namespace>.tools unless ATTEST_EXPAND=1. Flip enabled in ~/.hermes/profiles/research/config.yaml to switch namespace.

arxiv and blogwatcher disabled. Their descriptions -- "Search arXiv papers" and "Monitor blogs and RSS/Atom feeds" -- collide directly with the feed tools. With four such skills listed alongside, routing fell from 6/6 to 3/6: "check my rss feeds" went to blogwatcher and "recent arxiv papers on rag" went to arxiv. Their SKILL.md files are renamed rather than deleted, so re-enabling is a mv.

Only these two were disabled. Per-skill attribution at n=6 was noise -- dropping arxiv alone scored worse than keeping it -- so the aggregate is the only trustworthy result, and the other 70 skills cost ~70 bytes each with no evidence they hurt.

What was deliberately NOT done

Skills were not mass-disabled. They load as a ~70-byte index entry and the body only loads on invoke: 68 skills cost 7 KB against 85 KB of tool schemas. Hiding tools is 12x the lever that hiding skills is.