Skip to content

hermes image: a connection's skill is shown to an agent that holds the connection (FIN-509) - #1296

Draft
alexlwn123 wants to merge 1 commit into
t3/fin-509-6-six-connectionsfrom
t3/fin-509-7-skill-filter
Draft

alexlwn123 wants to merge 1 commit into
t3/fin-509-6-six-connectionsfrom
t3/fin-509-7-skill-filter

Conversation

@alexlwn123

Copy link
Copy Markdown
Collaborator

What and why

Hermes lists every installed skill's name and description in an agent's prompt, and the image installs the whole managed set. With a skill per connection that grows with the catalog: at a hundred connections, about 5,000 tokens in every chat, connected or not.

Each managed skill that names a connection (metadata: fragment: connection) is now disabled, by name, in the profile of an agent whose credentials hold no such provider (Hermes' skills.disabled), and shown once they do. It is read again whenever a profile's config is written: at a connect or disconnect, and on the models cadence, so it follows a newly released skill too.

The image's, as the platform rule has it. No platform or wire change. 3 files, +90.

Gaps

  • Not run in the image. skills.disabled is read from a local copy of Hermes' source (agent/skill_utils.py, get_disabled_skill_names), not confirmed on the pinned v0.21.7 image, nor that it covers skills from external_dirs. The Docker and real-Hermes lanes are owed.
  • An agent with no skill for a service can no longer name that service from its skills list. The computer's page says its owner can connect more in Settings.
  • Google's skill is now hidden from an agent without Google connected. Before, it was listed and said to ask the owner.

What ran

On the stack's tip (8f123596e, on master e6478e699), on the Mac:

  • cargo xtask check: source guards, 155 JavaScript files parse, 38 JavaScript test files pass, then it stops at fragment-cli --test mcp, which fails the same way on master on this machine (it writes a config folder macOS does not read).
  • The rest by hand: cargo test --workspace --all-features --no-fail-fast: 1,445 passed, 2 failed, both environmental here (fragment-cli --test mcp; fragment-e2e hosted::tests::the_plan_follows_what_each_section_needs, which needs Chrome).
  • cargo clippy --workspace --all-targets --all-features -- -D warnings, the cell's Wasm clippy, and images/ clippy: clean. images/: cargo test -p hermes-boot, 89 passed.

Not run, on purpose for now: any e2e lane, the Docker image tests, the Wasm size delta, and a Codex review. This is a draft until those are in. Nothing here was tried against a real provider or a real agent.

The stack (FIN-509)

  1. structural tests · 2. labels in the row · 3. fragment api · 4. the shared shape, with Google, Linear and Salesforce · 5. sign-in scopes · 6. six connections · 7. the skill filter · 8. GitHub. Each is based on the one before, so its diff is its own.

…e connection (FIN-509)

Each managed skill that names a connection (metadata: fragment:
connection) is disabled, by name (Hermes' skills.disabled), in the profile
of an agent whose credentials hold no such provider, and shown once they
do. So an agent's prompt lists the skills of what its owner connected,
however many connections the platform offers: with a hundred, about 5,000
tokens a chat that no agent pays. Read again whenever a profile's config
is (a connect, a disconnect, and on the models cadence), so it follows a
new skill too. The image's, as the platform rule has it: no platform or
wire change.

This branch has not been deployed

No deployments
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant