plat 576

← brain / skills

PLAT-576: Brain owns company skills; MCP get_skill replaces CLI skill install

Field	Value
State	fixed on main
Priority	P2
Product	brain
Area	skills
Summary	Brain becomes the owner of company-wide skills: published into Brain folders, used by every product and by MCP clients through get_skill

What happened

Fix

Left

Why

Owner, 2026-10-06. Today the only shared skills are the installation's flat skills/ library: everyone sees all of it, anyone whose workflow installs a skill changes it for everyone, and there are no versions or backup. Code projects keep private skills/ folders that cannot be shared. list_skills lists that library; search_skills searches the public internet registry. The global skills library was removed by PLAT-581 (2026-10-06): skills now belong only to each workflow, Crew and Code project. Owner: global skills are replaced by Brain. Until this is built, a skill can only be shared by installing it into each workspace. Workflow learnings (learnings/_global) and step-specific skills are out of scope.

Design (agreed in discussion, not built)

• Storage: a Brain skill is a standard skill package (SKILL.md, references/, scripts/) in a Brain folder, for example Company/Skills/<name>/ or RTS/Skills/<name>/. It gets Brain's folder roles, versions, history and Git backup.
• Access: folder Reader can use a skill; Editor can publish or update. Open question: should a skill with scripts need an Owner to publish (scripts run on colleagues' machines and in agents' sandboxes)?
• Publishing: an MCP tool (for example publish_skill) uploads a package into a folder where the caller is Editor. Re-publishing makes a new version; author and date are recorded. The app can do the same.
• Use on the platform (Code, Crew, Goals): skill search and listing include Brain skills the person or project can read; attaching one projects it read-only into the CLI's skills folder with the platform's ownership marker (PLAT-568 keeps those folders writable for the person's own skills).
• Use from MCP clients: get_skill(name) returns a skill's files; the client's agent writes them into its own skills folder (.claude/skills, .agents/skills). This replaces agentworks skills install (the AgentWorks skill itself becomes get_skill("agentworks")), so skills need no CLI. search_skills searches company (Brain) skills first, then the public registry.
• Shown to the agent: each skill's author and version; agents say which version they used.

Open

• Whether Brain entries can hold a skill's non-markdown files (scripts, assets) as they are; check before building.
• The old root skills/ store stays on disk only as migration input (PLAT-581); importing its skills into a Company/Skills Brain folder is a candidate first step. list_skills / install_skill are now workspace-scoped; Brain skills add a company source to search and attach.
• Owner rule for scripts (above).
• Add the Relays section to the AgentWorks skill (agent_go/pkg/agentworksclient/skills/agentworks/SKILL.md has none) when this ships.

Built (2026-10-06, owner: "lets do full")

• Brain stores any file type first (PLAT-588), so a skill's scripts and assets are stored as they are.
• New Brain tool knowledgebase_skills (pkg/knowledgebase/skills.go), on the ordinary entry operations so roles, versions, request idempotency and Git backup apply:
  • list (optional folder_path scope, query): every folder holding a SKILL.md the caller can read, with name/description from its front matter, version, updated by/at.
  • get (folder_path): every file of the skill (content, or content_base64 for binary) plus install instructions.
  • publish (folder_path, files[{path, content|content_base64}], request_id): creates missing subfolders, creates/updates files, removes files the new version dropped (a publish replaces the package). Editor on the folder; a skill with scripts (scripts/ or a script extension) needs Owner. Each step has a request ID derived from the outer one, so a retried publish replays.
• Everywhere Brain tools go: external MCP (catalog, instructions now "six tools"), platform agents via ConnectionToolDefinitions (publish only where the project may write), Goals builder/run tool lists, the knowledgebase feature, external builder, step test mode (list/get run, publish is stubbed), and the step execution policy that removes Brain tools.
• Platform agents: search_skills lists company skills from Brain first, then the public registry; install_skill(source="brain:<folder>") copies a Brain skill into the current workflow/Crew/Code skills/ through the caller's Brain access and the project's Brain mode (cmd/server/brain_skills.go). Re-install to update.
• MCP clients: get returns the files and the client writes them under its own skills folder; the AgentWorks skill (agentworksclient/skills/agentworks/SKILL.md) and Brain's MCP prompt say how. No CLI needed.
• Tests: TestBrainSkillsPublishGetListAndRoles (real service: Editor vs Owner for scripts, binary asset round trip, list, republish replaces, no grant sees nothing) and TestBrainSkillSearchAndInstallIntoWorkspace (server path: search finds it, install sends every file byte for byte to the workspace import). Tests that pinned five Brain tools now pin six.

Left

• Not deployed; not tried live with a real agent or MCP client.
• No app UI for skills (the Brain tab shows them as folders/files).
• Old root skills/ library content is not imported into Brain automatically; an Owner can publish the ones worth keeping.

Top-level Skills folder (2026-10-06)

Owner: "yes. top level skills". brain_skills publish without folder_path now publishes to Skills/<name> (the SKILL.md front-matter name, kebab-cased), creating the top-level Skills folder if needed; naming a folder still works. The tool says to publish only when the person asks to share a skill company-wide. Why: an RTS Crew put a skill in RTS/Latency/skills because that was the only Brain content it saw. Not deployed. The existing RTS/Latency/skills copy on RTS is unchanged.