技能 编写

Use this skill when creating or updating skills for the onscreen chat agent.

Skill File Layout

• Onscreen chat-agent skills live inside browser modules under mod/<author>/<repo>/ext/skills/....
• Repo-owned first-party shared onscreen skills should normally live under app/L0/_all/mod/_core/skillset/ext/skills/....
• Module-specific skills that describe one module's private contracts may live under that owning module's ext/skills/... tree.
• Group-scoped or admin-only onscreen skills may live under readable customware roots such as app/L0/_admin/mod/_core/<module>/ext/skills/....
• A skill file is always named SKILL.md.
• The skill id is the path relative to ext/skills/ with the trailing /SKILL.md removed.

Examples:

• mod/_core/skillset/ext/skills/browser-control/SKILL.md -> browser-control
• mod/_core/skillset/ext/skills/development/SKILL.md -> development
• mod/_core/skillset/ext/skills/development/modules-routing/SKILL.md -> development/modules-routing

Catalog Rules

• The onscreen prompt catalog lists only top-level skills from ext/skills/*/SKILL.md.
• Nested skills are not listed by default.
• Both the catalog and explicit space.skills.load(...) calls evaluate the current document's <x-context> tags before a skill is eligible.
• metadata.when may be true or a { tags: [...] } condition; metadata.when.tags requires all listed tags before the skill becomes catalog-loadable.
• metadata.placement accepts system, transient, or history; ordinary skills default missing or invalid placement to history, but auto-loaded skills may not resolve to history, so missing or invalid placement and explicit history all fall back to system unless the skill explicitly sets transient.
• Only top-level ext/skills/*/SKILL.md skills can auto-load through prompt discovery. Nested skills stay explicit-load-only routing targets even if they define metadata.loaded.
• Auto-loaded skills appear after the catalog in the auto loaded prompt block when their effective placement is system, or in the transient channel when they explicitly set metadata.placement: transient.
• Routing skills should tell the agent which deeper skill ids to load next.
• space.skills.load("<path>") loads the full skill file on demand and applies the same placement rule.
• history placement keeps the loaded skill body in ordinary execution-output history.
• system placement stores the loaded skill in the current chat's runtime system-skill registry and execution reports skill loaded to system message.
• transient placement stores the loaded skill in the current chat's runtime transient-skill registry and execution reports skill loaded to transient area.
• Auto-loaded skills cannot resolve to history; if they also set metadata.loaded, then metadata.placement: history is treated as system.
• A plain top-level await space.skills.load("<path>") is enough to apply the placement; use return only if you also want the execution result value explicitly.

Conflict Rules

• Skill ids must be unique across readable mods.
• If a skill is visible only to a narrower audience, still give it a top-level id that will not collide with shared _all skills that those users can also read.
• Conflicting ids are omitted from the catalog.
• Loading a conflicting id fails with an ambiguity error.

Skill Content Rules

• Start with frontmatter containing name, description, and optional runtime-owned metadata.
• Framework bootstrap already exports exactly one runtime context: data-runtime="browser" in normal web sessions or data-runtime="app" in the packaged desktop runtime, plus the derived tag runtime-browser or runtime-app; packaged app routes derive that from the desktop bridge before falling back to launcher runtime info or frontend config.
• Import /mod/_core/framework/js/context.js when code needs to read the current live <x-context> set directly; use getTags(...) for tag collection, getAttributeValues("data-runtime", ...) for raw runtime reads, and the other helpers for broader context inspection.
• Use <x-context> tags in mounted DOM when a module needs to expose additional live skill-filter state such as onscreen, admin, route:spaces, or space:open.
• Use metadata.when.tags when the skill should exist only in those live contexts.
• Use metadata.loaded only when the skill should be auto-injected without an explicit space.skills.load(...) call.
• Use metadata.placement: system when the skill body is durable instruction, metadata.placement: transient when it should live in the mutable transient block, and let the default history placement stand only for ordinary non-auto-loaded skills that should behave like normal conversation context.
• Prompt-facing skill text is token-budgeted. Keep wording terse, avoid unnecessary markdown or filler, and measure before or after changes with the local tokenizer when you edit auto-loaded or catalog-facing skill text.
• When a skill needs reusable browser logic, prefer a small JS helper stored inside that skill's own folder and imported from a stable /mod/<author>/<repo>/ext/skills/... path instead of pasting a long inline script into SKILL.md.
• Keep the top-level router skill directive and concise.
• Keep nested skills focused on one stable area.
• Prefer exact file paths, runtime names, and examples over vague guidance.
• If a skill subtree becomes complex, add an AGENTS.md file inside that subtree and keep it current.

Maintenance Rules

• When a mirrored source contract changes, update the affected skill files in the same session.
• When a stable feature or workflow changes, update the relevant docs under /mod/_core/documentation/docs/ and the documentation skill at /mod/_core/documentation/ext/skills/documentation/SKILL.md in the same session.
• Do not let skill guidance drift away from the owning AGENTS.md files.
• For the development super-skill specifically, keep /mod/_core/skillset/ext/skills/development/AGENTS.md current whenever the framework, router, API, layer, or auth contracts it mirrors change.