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。
开发与测试

instrument-feature-flag

Gate application code behind an existing Harness FME feature flag. Detect the app's language, framework and any existing flag wrapper first; ask only what that can't resolve (target app, subject identity, gating point, treatment/config/fallback names, extension needs). Prefer an existing wrapper; otherwise propose a thin, customer-extensible facade or hook — never a new framework or second SDK client — approved before editing. Map named treatments to branches with a safe, non-boolean-coerced fallback, then add tests at the wrapper and call-site boundary. New flags delegate to create-feature-flag with the one environment actually selected; missing targeting delegates to update-flag-targeting under fresh approval. Not for flag creation alone, metrics (choose-metric, create-metric, instrument-metric), or first-time onboarding (feature-flag-onboarding). Trigger phrases: gate this code behind a flag, add flag check, wire up getTreatment, add SDK evaluation call.

Harness · 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 harness/harness-skills --skill instrument-feature-flag

技能说明


name: instrument-feature-flag description: >- Gate application code behind an existing Harness FME feature flag. Detect the app's language, framework and any existing flag wrapper first; ask only what that can't resolve (target app, subject identity, gating point, treatment/config/fallback names, extension needs). Prefer an existing wrapper; otherwise propose a thin, customer-extensible facade or hook — never a new framework or second SDK client — approved before editing. Map named treatments to branches with a safe, non-boolean-coerced fallback, then add tests at the wrapper and call-site boundary. New flags delegate to create-feature-flag with the one environment actually selected; missing targeting delegates to update-flag-targeting under fresh approval. Not for flag creation alone, metrics (choose-metric, create-metric, instrument-metric), or first-time onboarding (feature-flag-onboarding). Trigger phrases: gate this code behind a flag, add flag check, wire up getTreatment, add SDK evaluation call. metadata: author: Harness version: 1.0.1 mcp-server: harness-mcp license: Apache-2.0 compatibility: Requires the Harness MCP server or the Harness CLI for flag reads, plus local application code access to make edits; guidance-only (no edits) if application access is unavailable

Instrument Feature Flag

Skill identifier: instrument-feature-flag. Gate application code behind an existing FME flag: inspect the real app first, prefer or design a small customer-extensible wrapper, get it approved, implement the minimal edits, and test at the wrapper and call-site boundary. Related: /create-feature-flag (new flag), /update-flag-targeting (missing/insufficient targeting), /feature-flag-onboarding (first-time setup, which delegates here for code).

Tools

Works through the Harness MCP server or the Harness CLI; names are from tool-map.md. This skill only reads Harness flag state for the selected environment — flag/definition writes are delegated to /create-feature-flag and /update-flag-targeting.

OperationMCPCLI
List flagsharness_list · fme_feature_flag · size: 50 · compact: falseharness list feature_flag --json --limit 50
Get flagharness_get · fme_feature_flag · params.feature_flag_nameharness get feature_flag <name> --json
List environmentsharness_list · fme_environment · compact: falseharness list fme_environment --json
List traffic typesharness_list · fme_traffic_type · compact: falseharness list traffic_type --json
Get definitionharness_get · fme_feature_flag_definition · params: { feature_flag_name, environment_id }harness get feature_flag:definition <flag> --env <env-id> --json

Instructions

Phase 1: Scope

Follow scope-establishment.md. Carry confirmed account/org/project, transport and application target from the request/caller; ask only for missing context before API calls. Keep Harness scope separate from the application repo/runtime — never derive one from the other. See mcp-setup.md for a supported CLI or MCP choice. List environments and require one explicitly confirmed environment before any flag/definition handoff. Resolve its ID and isProduction; if none is selected or the choice is ambiguous, ask and STOP—never inherit an all-environments default. This selection does not authorize a write.

Phase 2: Detect the codebase before integration questions

Inspect the real application: language/framework/runtime and package manager (lockfiles/manifests), which app/package in a monorepo is plausible, the installed FME/OpenFeature/Classic-FF SDK and its version, and any existing flag wrapper/facade (grep flag-adjacent helper names, not only raw SDK calls). See sdk-patterns.md for detection signals and language-guidance.md for the native-FME/OpenFeature/Classic-FF contract table — FME native treatments are names, not booleans, and batch/config/async evaluation calls have their own result shapes; never assume a signature shared across them. Stop: Classic FF-only repo with no matching FME flag; or an OpenFeature provider whose configured target (FME project/environment) is unverified — confirm it before treating getBooleanValue/etc. as evidence about this flag. Summarize detected facts in 3–5 bullets before asking integration questions. With no application access, ask for the missing non-secret context and provide guidance only: no invented file locations, edits or test success. An absent SDK is a setup case, not an unknown-signature exception: propose a runtime-appropriate package from verified official docs, and include installation/version changes in the approval plan.

Phase 3: Ask only what inspection didn't resolve

Ask only for facts Phase 2 left open, and never re-ask a fact already confirmed by inspection or already supplied by a calling skill (e.g. /feature-flag-onboarding):

  1. Target app/package, if a monorepo has more than one plausible target.
  2. Evaluation subject key source and traffic type, if more than one identity candidate exists (user vs. account vs. anonymous).
  3. Exact gating point — candidate file:line if more than one placement is plausible.
  4. Named treatments/config this call site must branch on, and the safe fallback for control/errors.
  5. Extension needs — will the app or its customers add treatments/branches later without touching this call site again?

Phase 4: Resolve the flag definition

Get flag using the confirmed name and scope. A 404 is not proof a new flag is needed: recheck the identifier, scope and access per tool-map.md. Only after resolving

阅读完整来源 ↗