Platform skills quick reference

In Chat, type / plus a skill_id (or pick from the list) to load that skill for the current turn. This page remains the tool map.

Trace AI uses skill-based tool gating: only tools listed in loaded skills (plus kernel tools) are callable. The elementary skill is auto-loaded every turn when gating is enabled.

Native tool call shape

Trace AI uses native function calling when skills gating is on. The function name is the tool; arguments are flat JSON matching each tool’s schema, plus optional progress_indicator:

{"skill_id":"integrations","progress_indicator":"Loading integrations skill…"}

Do not wrap args in {"action":"…","parameters":{…}} — that legacy inline shape is only for message-content fallback. The runtime unwraps nested parameters when detected, but flat args are preferred.

Kernel tools (always available when gating is on)

Auto-loaded: elementary

The platform also exposes a native random tool (no extra skill load): it reads OS entropy from /dev/urandom and returns lowercase hex. Use it when ProcessFlow steps or integrations need a true random seed — not Math.random or other pseudo-random helpers. Optional length is the number of bytes to read (default 32, max 4096); the hex string is twice as long.

Common pitfalls: entity_search without query fails immediately. For multi-line text, prefer write_file content_base64 (UTF-8); use content only for short single-line or append logs. Yes/no permission before a move, delete, or other side-effect must be request_human_input — never a Markdown “Please confirm…”.

On-demand: platform-safeguards

Load when backup or support escalation is needed (not auto-loaded).

Router: platform-actions

Load when routing to console CRUD is unclear. Does not replace domain skills.

Job compositions (composition_id)

Source of truth: SKILLS/compositions.json. Pass exactly one of skill_id or composition_id. A composition load unlocks member tools and injects the short job map — not every member SKILL.md. Harness peek (SUGGESTED THIS TURN) is job-composition hints only; elementary/kernel tools are already active. When peek names a specialist slug, use list_subagent_catalog — see Platform specialists.

Domain skills (load via load_platform_skill)

Each domain skill also exposes request_human_input where confirms are required. Full tool lists live in SKILLS/<skill_id>/skill.json in the platform repo.

WebApp create workflow (two skills)

flowchart LR
  A[load webapp-library] --> B[Author page_content]
  B --> C{Review draft?}
  C -->|optional| D[write_file tenant path]
  C --> E[load webapps]
  D --> E
  E --> F[describe_webapp]
  F --> G[create_webapp with webapp_user_id]
  G --> H{Draft OK?}
  H -->|user wants live| I[publish_webapp]

• create_webapp registers an unpublished platform draft (required). Pass webapp_user_id from describe_webapp.webapp_users (role webapp_user). Never omit it and never use the calling tenant_admin. If webapp_users is empty, ask the tenant admin to create a Webapp user in User Management and do not create yet.
• publish_webapp is separate when the user wants the WebApp public.
• Tenant write_file drafts do not replace create_webapp.

Tenant skills

Folders under tenantdata/{tenant_id}/SKILLS/ with SKILL.md and optional skill.json. Author with load_platform_skill({ "skill_id": "tenant-skill-authoring" }). Install from the public Library with model-skill-import (see Model skills library). Run with load_tenant_skill({ "skill_id": "<folder-name>" }) when CHAT_TENANT_SKILLS_ENABLED=1. See Tenant skills user guide.

Typical workflow

flowchart LR
  A[elementary auto-loaded] --> B{Console write?}
  B -->|No| C[entity_search / read_file / websearch]
  B -->|Yes| D[load platform-actions]
  D --> E[load domain skill]
  E --> F[describe_* / list_* / get_*]
  F --> G[request_human_input if risky]
  G --> H[create / update / execute tool]

See also

• Trace AI Agent
• Platform actions
• Trace AI subagents
• Platform specialists
• Platform code generation
• Tenant skills user guide
• Process flows as custom AI agent tools
• Model skills library
• Model Autonomic Agents library