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。
部署与云服务

fme-rollout-pipeline

Design and wire Harness FME pipeline stages and native FmeFlag* steps for feature-flag rollout patterns — progressive percentage, environment promotion, approval gates, and downstream gating. Composes FeatureFlag, Approval, and Custom stages rather than a single canned template. Use when asked to build a flag rollout pipeline, promote a flag across environments, add approval before prod, or gate prod until staging is at 100%. Do not use for direct flag CRUD (manage-feature-flags), code cleanup (cleanup-feature-flags), or generic CI/CD without FME steps (create-pipeline). Trigger phrases: FME pipeline, flag rollout pipeline, progressive rollout, environment promotion, approval before prod, feature flag stage, canary rollout, FME rollout.

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 fme-rollout-pipeline

技能说明


name: fme-rollout-pipeline description: >- Design and wire Harness FME pipeline stages and native FmeFlag* steps for feature-flag rollout patterns — progressive percentage, environment promotion, approval gates, and downstream gating. Composes FeatureFlag, Approval, and Custom stages rather than a single canned template. Use when asked to build a flag rollout pipeline, promote a flag across environments, add approval before prod, or gate prod until staging is at 100%. Do not use for direct flag CRUD (manage-feature-flags), code cleanup (cleanup-feature-flags), or generic CI/CD without FME steps (create-pipeline). Trigger phrases: FME pipeline, flag rollout pipeline, progressive rollout, environment promotion, approval before prod, feature flag stage, canary rollout, FME rollout. metadata: author: Harness version: 1.0.0 mcp-server: harness-mcp-v2 license: Apache-2.0 compatibility: Requires Harness MCP v2 server (harness-mcp-v2)

FME Rollout Pipeline

Help users pick and wire Harness FME pipeline steps for rollout use cases. Compose FeatureFlag stages, native FmeFlag* steps, Approval gates, and optional Custom / Wait stages tailored to the user's rollout scenario, rather than outputting a single fixed template.

Direct flag kill/restore/create via MCP is /manage-feature-flags. Post-launch code removal is /cleanup-feature-flags. General CI/CD structure is /create-pipeline.

Instructions

Load references on demand:

  • rollout-patterns.md — pattern catalog and plan checklist
  • fme-pipeline-steps.md — step types and evaluation order
  • pipeline-examples.md — composable YAML snippets

Allowed MCP tools only: harness_list, harness_get, harness_create, harness_update.

Allowed resource types only: pipeline, fme_environment, fme_feature_flag, fme_feature_flag_definition, and project (pre-create existence check only — do not create projects).

Prefer org_id + project_id. Do not require deprecated workspace_id.

Stop before mutating. Do not harness_create / harness_update a pipeline until the user confirms the rollout plan.

No automatic metric rollback. Harness does not auto-rollback a flag on metric regression. Use Approval, Wait, Custom verification, or FmeMetricCheck (fails the step). Rollback is an explicit FmeFlagKill stage or failure path.

Phase 1: Establish scope

Reuse scope-establishment.md. Ask for org_id and project_id if missing.

Restate: Working in org=..., project=...

Phase 2: Choose mode

ModeUser signalOutcome
Design (default if ambiguous)"How should we roll out…?", "What steps do we need?"Pattern + stage plan only; no YAML write
New pipeline"Create a rollout pipeline for…"Plan + YAML draft on first turn; harness_create only after explicit confirm
Update existing"Add FME stages to pipeline X"Fetch pipeline, merge stages, show diff; harness_update only after explicit confirm

Phase 3: Gather rollout intent

Collect (ask only for missing items):

  1. Flag name (feature_flag_name, case-sensitive)
  2. Test treatment and control treatment (or boolean on/off)
  3. Target environments and whether any are held (not released yet)
  4. Pattern — progressive %, env promotion, staging gate, approval, kill-switch rollback (see rollout-patterns.md)
  5. Existing pipeline URL/identifier (update mode)
  6. Approvers for HarnessApproval (user groups, minimum count)
  7. Soak / gate between % increases — Wait, Approval, FmeMetricCheck, or Custom tests (do not invent auto-rollback)

Do not guess environment names, treatments, or approver groups.

Phase 4: Discover FME context

MCP v2 identifier nesting — required; bare top-level FME field names are stripped by the tool schema:

CallWhere identifiers go
harness_list with required list filtersfilters: { feature_flag_name: "..." }
harness_get for fme_feature_flag / fme_feature_flag_definitionparams: { feature_flag_name: "...", environment_id: "..." } (environment_id when per-env)
harness_get / harness_update for pipelinetop-level resource_id
harness_list project verify (pre-create only)top-level resource_type: "project", org_id

If MCP is unavailable, skip discovery, say so, and ask the user for environment names and treatments — do not guess.

Call MCP tool: harness_list
Parameters:
  resource_type: "fme_environment"
  org_id: "<org>"
  project_id: "<project>"

Build a promotion-order proposal: non-production first, production last. Use isProduction when present. Confirm order with the user.

Call MCP tool: harness_get
Parameters:
  resource_type: "fme_feature_flag"
  org_id: "<org>"
  project_id: "<project>"
  params: { feature_flag_name: "<flag_name>" }
Call MCP tool: harness_list
Parameters:
  resource_type: "fme_feature_flag_definition"
  org_id: "<org>"
  project_id: "<project>"
  filters: { feature_flag_name: "<flag_name>" }

Note per environment: isKilled, defaultTreatment, defaultRule (treatment % split), trafficAllocation (experiment participation only), rules, treatment targeting lists. Use this to avoid contradicting live state.

A killed flag cannot take traffic — plan FmeFlagRestore before allocation.

Phase 5: Select pattern and wire stages

Using rollout-patterns.md and fme-pipeline-steps.md:

  1. Map intent → pattern (mix patterns; do not force one canned pipeline)
  2. List stages in execution order with step types per stage
  3. Insert Approval where the user wants human ga
阅读完整来源 ↗