跳至正文
OpenAgentSkill
浏览已开放 · 目录记录日期 2026-11-21 · 自动更新和账户服务暂时关闭。
设计与 UI

tool-renderer-contract

Design and review structured tool-result rendering. Use when rendering tool calls, tool messages, structured tool outputs, status cards, fallback JSON/text displays, parser boundaries, display status mappings, or unknown tool results.

HsienW · 已有静态检查记录

安装与使用

运行命令前先阅读来源与依赖说明。复制命令不会自动安装或执行。

npx skills add HsienW/ai-agent-engineering-playbook --skill tool-renderer-contract

技能说明


name: tool-renderer-contract description: Design and review structured tool-result rendering. Use when rendering tool calls, tool messages, structured tool outputs, status cards, fallback JSON/text displays, parser boundaries, display status mappings, or unknown tool results.

Tool Renderer Contract

Skill Interface

  • Name: tool-renderer-contract.
  • Description: Design and review structured tool-result rendering for tool calls, tool messages, structured tool outputs, status cards, fallback JSON or text displays, parser boundaries, display status mappings, and unknown tool results.
  • Parameters: Tool result schema, raw message content, parser boundary, display status mapping, known specialized renderers, fallback rendering behavior, sensitive field policy, and verification cases.
  • Instructions: Use this skill when rendering tool output in UI. Parse raw output as untrusted data, validate minimum structured fields, map status through a closed display union, provide safe fallbacks for unknown tools, and verify malformed, future-version, and sensitive-data cases.

Tool rendering should be schema-driven and resilient to unknown tools. The UI must not trust raw tool output as display-ready data.

Layers

  1. Message container: owns message grouping and association between tool call metadata and tool output.
  2. Generic tool renderer: parses common metadata, displays status, and chooses a specialized renderer when one is available.
  3. Specialized renderer: renders a known tool result schema.
  4. Fallback renderer: safely displays unknown JSON, text, or parse failures.

Structured Result Shape

Use a generic shape that can be specialized:

type ToolResult<TStatus extends string, TData = unknown> = {
  schemaVersion: string;
  tool: string;
  status: TStatus;
  data?: TData;
  summary?: string;
  errorCode?: string;
};

The frontend parser should accept unknown or raw message content, validate the minimum fields, and return a typed result or undefined.

Status Mapping

Map tool status through a closed display status union:

type ToolDisplayStatus =
  | 'running'
  | 'success'
  | 'needs_input'
  | 'not_found'
  | 'denied'
  | 'error'
  | 'timeout'
  | 'cancelled'
  | 'unknown';

Use a single Record<ToolDisplayStatus, DisplayConfig> for labels, icons, colors, and ARIA text. Do not scatter status rendering across component branches.

Safety Rules

  • Treat tool output as untrusted.
  • Never render raw HTML from tool output unless sanitized by a trusted policy.
  • Never expose stack traces, credentials, provider secrets, or raw internal payloads.
  • Do not infer status from user-visible text.
  • Do not hardcode one tool result as the only rendering path.
  • Unknown tools must have a safe fallback.

Parser Rules

  • Validate schemaVersion, tool, and status.
  • Preserve unknown fields for diagnostics only when safe.
  • Be forward compatible with unknown minor schema versions.
  • Provide unknown-status fallback.
  • Keep type assertions behind runtime checks.

Verification

Test known structured results, unknown tool results, malformed JSON, plain text output, unknown status, future schema versions, error envelope rendering, and sensitive field redaction.