macOS alpha, 0.32.0-alpha.1

Skills

Settings reference: behavior, defaults and safeguards for this page of desktop Settings.

Published unchanged from docs/settings/skills.md in Moguri's source repository at commit b686ddc (0.32.0-alpha.1). That repository isn't public yet, so references to files that only exist there appear as plain text with a dotted underline.

Technical reference for Settings / Skills. The app shows a short summary; this page keeps the full detail.

The Skills tab next to Automations shows the same installed skills with On/Off switches, a Marketplace that installs a skill from these sources into one project or every project, and the agent's toolsets. Contract: skills, tools and artifacts (not public yet).

Skill gateway

  • Sources can be a local skill folder, a public GitHub repository or a direct HTTPS ZIP archive.
  • Each search and each read refreshes and validates the source; a lookup error returns no stale result.
  • No scripts run, no skills are silently installed, and skill text cannot grant permissions.
  • Sources belong only to the current profile and are read on the Local control plane, also for Remote sessions.
  • Remote (SSH) sessions load skills with the same approval rules. The first use copies the skill's files to a Moguri folder on the SSH host (<state>/skills/<SHA-256>/); its scripts run there only after normal command approval. Sandbox sessions can't load skills yet.
  • Contract: docs/architecture/skills-and-rewind.md.

Global skill folders

  • ~/.agents/skills and ~/.moguri/skills can be used separately or together.
  • Confirming a folder creates it if it is missing and saves it to the profile's sources.
  • A missing folder offers no skills and never blocks the others.
  • Confirming does not scan the folder at that moment; adding it does not install skills, run instructions or grant model or tool permissions.
  • If loading the presets fails, no directory is scanned.

Install from a source (protocol 114)

  • Install in Customize / Marketplace re-reads the source and installs those exact bytes as a disabled skill in the chosen project or, with All projects, for every project. A changed source needs a fresh search.
  • Installing grants no tools. Reinstalling changed bytes under the same name updates that skill to a new disabled revision; identical bytes and plugin-owned names are refused.
  • TUI: /skills install ID SHA [global]. CLI: moguri skills install ID --digest SHA [--project ID].

Find skills

  • Only saved, enabled sources are read.
  • All downloads are checked again when a selected skill is opened; a changed body needs a fresh search (the results line used to say "This search used fresh source bytes").
  • Previewing a skill is not permission to run its instructions or to access a different execution host.

Skill approvals for this session (was "Skill source reviews")

  • Discovery and source reads are not admission (approval).
  • A hash or repository label is not publisher identity; neither is a name, description or scanner result.
  • Approved (admitted) skill context is limited to reviewed native file tools: never commands, network, credentials or automatic permission.
  • The source and admission transport is Local-only; there is no controller fallback for non-Local sessions.
  • Excluded protected paths cannot be dependencies of the skill.
  • External references in the skill are not fetched, inspected or authorized.
  • Allowed paths are literal paths. Write access ("Read and write files") is offered only when the skill declares native-file access, and stays under current hard policy.
  • Signed publishers: verify against the documented moguri-skill-artifact-v1 signature payload with a separately user-pinned Ed25519 key. Pinning proves control of this reviewed key, not an independently verified real-world organization. HTTPS archives require a pinned signature.
  • Unverified publishers record an actual-user Local review acknowledgement.
  • Saved publisher keys: when the profile has saved keys, Publisher key offers them instead of pasting a key.
  • Indicator findings: the review lists pattern matches (hidden characters, download-and-run, credential paths, instructions to override rules and so on). They are advice, not a safety verdict. A skill from a trusted folder that contains invisible or bidirectional control characters isn't loaded automatically; approve it here.
  • Approvals expire (the UI requests 24 hours; the service allows at most 168).
  • Revocation blocks future consumption and effects. It cannot retract instructions already delivered to a model or undo work already running. A rewind never restores an old source grant.
  • Contract: docs/architecture/skills-and-rewind.md, docs/setup.md (skill admission section).

Source chat

The chat picker was previously labeled "Source chat for skill admission". With no chat selected: choose a source chat to review exact skill admission; no session or workspace is inferred. Save or cancel skill source folder changes before reviewing session admission.

Imported skills

Import standard SKILL.md front matter and instructions. A skill grants no tool permissions.

Built-in slash commands always take precedence. A skill named like one (for example review, or a TUI command such as trace) shows "Rename the skill (the name in its SKILL.md) to use it with /." on its row, after import, in Find skills and in the slash menu. It still works when attached by its exact entry.

Skill changes status

Toggling skills reports "N of M skill changes acknowledged"; existing source reviews and permissions are unchanged. Deleting a skill: sessions that admitted it lose access; no other capability or permission changes.

Delete confirmation

Chats that admitted the skill lose access. Its version history stays in this profile's encrypted event history.

Publisher keys

  • Save publisher key stores an Ed25519 public key (PEM) and a name for this profile. Saving needs the confirmation that you got the key through a channel you trust. A key verifies signatures; it doesn't prove who the publisher is.
  • Block key permanently blocks a key fingerprint, whether saved or pasted. Block this version permanently blocks one exact skill version by its artifact SHA-256 (shown in each review and approval).
  • A block stops new approvals and every later load or effect of anything it covers, in every chat of the profile. It can't recall instructions a model already read. Blocks can't be undone.
  • CLI: moguri skills publishers, moguri skills publisher-save NAME KEY.pem --confirm-review, moguri skills block key|artifact SHA REASON --confirm-review. TUI: /skills publishers, /skills publisher-save confirm-trusted JSON, /skills block key|artifact SHA confirm-block REASON.
  • Contract: docs/architecture/skills-and-rewind.md#source-indicators-and-publisher-trust-protocol-112.

Back to the user guide