This is a follow up from an earlier discussion on agent sandboxes https://news.ycombinator.com/item?id=49605644 which happened prior to the release of Muse so this time I'm sharing my notes on it.
Muse has a lot of cool things visible through the VM worth talking about (safety classifiers, cloud-hypervisor instead of firecracker, eager compaction, and an interesting rescue system, memory in postgres).
Just started using it today so pretty excited to put it on some harder problems and see how it goes :).
I saw `ads_mcp` in the skills list; I wonder what this does? Are skill files readable? Not to jump to conclusions, but a LUKS volume aint gonna keep your data private when the back door is wide open.
Not entirely sure but I know there is a scoping system so the base skills are in opt/hatch/skills/ in the container but some skills are in /opt/hatch/skills-scoped/ in the vm. so all the skills in that container are easily readable but for the other ones i assume they are early access things that only mount for specific channels (user groups)
# skill-scopes.conf -- GENERATED from extensions_scoped.toml. DO NOT EDIT.
#
# extensions_scoped.toml is the single source of truth: each [[skills]] entry
# lists the skill's `scopes`, and this file is derived from them -- one
# `<scope> <directory> ...` line per scope, both sorted for a stable render.
#
# The tokens after the scope are skill DIRECTORY names, not SKILL.md names:
# launch-daemon.sh resolves each against /opt/hatch/skills-scoped/<directory>.
#
# launch-daemon.sh reads this at cell boot and reveals only the gated skills
# whose scope matches the VM's JARVIS_CD_CHANNEL. An unset/unknown channel or a
# missing line reveals nothing (fail-closed). Edit the manifest, not this file.
hatch-e2e ads_mcp health nutrition
hatch-evals ads_mcp health nutrition
internal-test ads_mcp audio_notes_read documents end-call health nutrition polymarket price-tracker
prod granola notion whatsapp
This is a follow up from an earlier discussion on agent sandboxes https://news.ycombinator.com/item?id=49605644 which happened prior to the release of Muse so this time I'm sharing my notes on it.
Muse has a lot of cool things visible through the VM worth talking about (safety classifiers, cloud-hypervisor instead of firecracker, eager compaction, and an interesting rescue system, memory in postgres).
Just started using it today so pretty excited to put it on some harder problems and see how it goes :).
Fascinating; thank you.
I saw `ads_mcp` in the skills list; I wonder what this does? Are skill files readable? Not to jump to conclusions, but a LUKS volume aint gonna keep your data private when the back door is wide open.
Not entirely sure but I know there is a scoping system so the base skills are in opt/hatch/skills/ in the container but some skills are in /opt/hatch/skills-scoped/ in the vm. so all the skills in that container are easily readable but for the other ones i assume they are early access things that only mount for specific channels (user groups)
The channels concept is visible via a conf file:
# skill-scopes.conf -- GENERATED from extensions_scoped.toml. DO NOT EDIT. # # extensions_scoped.toml is the single source of truth: each [[skills]] entry # lists the skill's `scopes`, and this file is derived from them -- one # `<scope> <directory> ...` line per scope, both sorted for a stable render. # # The tokens after the scope are skill DIRECTORY names, not SKILL.md names: # launch-daemon.sh resolves each against /opt/hatch/skills-scoped/<directory>. # # launch-daemon.sh reads this at cell boot and reveals only the gated skills # whose scope matches the VM's JARVIS_CD_CHANNEL. An unset/unknown channel or a # missing line reveals nothing (fail-closed). Edit the manifest, not this file. hatch-e2e ads_mcp health nutrition hatch-evals ads_mcp health nutrition internal-test ads_mcp audio_notes_read documents end-call health nutrition polymarket price-tracker prod granola notion whatsapp