Skill visibility — who sees which skill, where

Status: design, 2026-11-21. Extends Composer completion.

1. What is wrong today (measured on the public instance, 2026-11-21)

The / menu is SkillAutocompleteProvider.GetSkills → AiSettingsNodeType.ResolveSkillQueriesOnce → the union of four query rows, deduped by id in SkillNodeType.ProjectSkills:

Row What it reads
platform namespace:Skill nodeType:Skill
user namespace:{user}/Skill nodeType:Skill
space path:{partition of context} scope:descendants nodeType:Skill
node type path:{partition of node type} scope:descendants nodeType:Skill. The composer passes nodeTypePath: null, so this row never resolves.

Nothing else decides visibility: there is no role gate, no "user-facing vs agent-only" flag (AutoMount only steers what the AGENT is advertised) and no notion of a skill that belongs everywhere. That gives four defects.

1.1 Missing: every package skill authored as .md is dropped silently

On the public instance Feedback/Skill/feedback (and every other package .md skill) is stored with content.$type = MarkdownContent. The built-ins (Skill/model, …) carry SkillDefinition. ProjectSkills reads node.ContentAs<SkillDefinition>(), which returns null for a differently-named type, and the loop continues. No log line, no error: the skill is simply not offered. This is the #1984 defect SkillFileParser fixed for one parse path; the path that delivers packages to an instance still lands the catch-all Markdown shape. The package .json skills (DataModelling, Edu, Publish, Video, …) are typed and do appear.

Affected (31): feedback, wrong-answer, schedule, scene, signature, remote, expense, crm, cloud-sync, proton, swiss-transport, quality-time, quality-time-de (both are Pick ACTIONS, so the menu entry and its picker are both dead), email, bug-fix-thread, operation-request, the six Governance skills, the seven Hosting .md skills, governed-mail, providers, build-activities, application, ci-policy, coupon.

1.2 Missing: a skill that belongs everywhere can only live in the platform partition

/feedback is meant to work from any page. A package skill reaches the menu only while the space in view IS that package's partition. The only "everywhere" layer is the platform Skill partition, and the only way into it is the Store's all-or-nothing preInstalled mirror (1.4): a package cannot say "offer THIS skill globally and keep that one in my space".

1.3 Over-exposed: the platform catalogue is shown to everyone

All 28 built-ins go to every signed-in viewer in every context, including operator and developer procedures: driver-release, package-versions, pull-request, recycle-after-deploy, provider-keys, logon-action, ui-extensibility, layout-area, query, code, access, group, activity, thread, build-a-connection. Most of these are instructions for an AGENT, not commands a person types. In the Hosting or Governance spaces the same happens with the agents' operating rules (pr-babysitter, fleet-coordinator, triage, devops, …) once 1.1 is fixed.

1.4 Over-exposed: global mirrors and stale copies

Present on the public instance and in no repository folder that owns them:

2. Design: every skill declares three things

Visibility stops being a side effect of which query row happens to match. Each Skill node declares it (new SkillDefinition members, front matter in the .md):

Field Values Default Meaning
invocation user · agent user when the skill has an action, else user for the platform partition and packages alike agent = never in the / menu; the agent still loads it on demand (load_skill). The equivalent of Claude Code's user-invocable: false.
scope context · app · global global in the platform partition, context in a package context: offered while the space or the node TYPE in view lies in the skill's partition. app: offered everywhere to a viewer who can read the skill node, i.e. who has the app. global: offered everywhere to every signed-in viewer (the skill's partition must be public-read).
requires Read · Create · Admin · GlobalAdmin Read The permission the viewer must hold ON THE CONTEXT (or IsGlobalAdmin). Hides authoring and operator procedures from people who cannot act on them anyway. It is a display filter: the operations the skill drives are still access-checked as always.

2.1 Serving

  1. One catalogue per viewer, cached per (user, context partition, node-type partition, locale): platform row, user row, context row, node-type row (the composer passes the node type in view — fixes the dead row), plus ONE new row partitions:all nodeType:Skill scope:app|global invocation:user, which the store already row-filters by read access. That row is what makes /feedback (scope global) and /schedule (scope app) appear anywhere.
  2. Tolerant projection: ProjectSkills accepts MarkdownContent (body → Instructions) and front-matter fields, and logs a warning naming the path when it has to. Silent-failure rule 1. The delivery path is fixed separately so new nodes land typed.
  3. Gate, then dedupe: drop invocation: agent; drop a skill whose requires the viewer does not hold on the context (one AccessService evaluation per distinct requirement, captured on the request turn); then dedupe by id with precedence user > context > node type > app/global package > platform.
  4. The agent reads the same catalogue (MeshAgentSkillsSource) with step 3's invocation filter skipped and requires kept: what a person sees ⊆ what the agent can load, never the reverse.

2.2 Guard rails

3. Proposed classification of every skill

ctx = scope context, app, global; agent = invocation: agent.

Platform (src/MeshWeaver.AI/Data/Skill)

Skill invocation scope requires
agent, model, harness, effort, clear, navigate user global Read
connect, share-email user global Read (signed in)
markdown, presentation, slide, og-card, space user global Create
code, layout-area, ui-extensibility, query user global Create
access, group user global Admin
provider-keys, logon-action user global GlobalAdmin
activity, thread, build-a-connection, pull-request, recycle-after-deploy, driver-release, package-versions agent — —

Packages

Package Skill invocation scope requires
Feedback feedback, wrong-answer user global Read
Scheduling schedule user app Read
Signature signature user app Read
Expenses expense user app Read
Crm crm user app Read
RemoteControl remote user app Read
WebSearch swiss-transport user app Read
QualityTime / QualityTimeDe quality-time, quality-time-de user app Read
RolePlay scene user ctx Read
CloudSync, Proton cloud-sync, proton user ctx Read
Essentials email user ctx Read
Essentials models user ctx GlobalAdmin
Essentials bug-fix-thread, operation-request agent — —
Governance public-link user ctx Create
Governance tenant-build user ctx GlobalAdmin
Governance devops, long-running-check, propose-activity, review-activity agent — —
Hosting setup-environment user ctx GlobalAdmin
Hosting diagnose-missing-node, fleet-coordinator, live-defect-triage, loki-error-triage, pr-babysitter, pr-review, triage agent — —
BuildServer, Mail, Observability build-activities, governed-mail, debug agent — —
Providers providers user ctx Admin
Store plugin, application user ctx Create
Store coupon user ctx GlobalAdmin
Store ci-policy agent — —
DataModelling DiscoverCube, DomainAttributes, ModelCube, ModelForeignKeys, ShipSampleData user ctx Create
Edu create-course user ctx Create
Edu course-e2e agent — —
Publish multipart-doc, pdf-package user ctx Create
Training tour user ctx Create
Video video, poster user ctx Create

Satellite repos (SocialMedia, Reinsurance, Manufacturing, Education) classify their own skills against the same three fields; the defaults keep them context-scoped until they do.

4. Delivery order

  1. Tolerant projection + node-type row (fixes 1.1 and the dead row; no new fields needed). /feedback appears inside the Feedback space.
  2. The three fields, the app/global row and the gate, with the matrix test (fixes 1.2 and 1.3).
  3. Front matter for every skill in this repo per section 3.
  4. Retire PreInstall.SyncGlobalSkills in favour of the scope: global row, then delete its mirrors and the stale user-home copies.