Flow Steward Agent Plugin

Download the complete plugin ZIP, or read the packaged README. The archive extracts to one installable flow-steward/ root with both manifests and six canonical skills (plugin version 1.1.0).

The workflow skill uses your already configured authenticated Flow Steward MCP connection. The other skills work from a terminal with git, the GitHub CLI gh or a browser and, for development, a Flow Steward application checkout.

Skill	Audience and use	Page
flow-steward-workflow-authoring	End users: reuse/compose local workflows, inspect runs, errors and approvals, request operational reports, export or import a workflow.	Tools And MCP
flow-steward-report-issue	End users and their agents: report a bug or request a feature on GitHub, with the redacted System Diagnostics report attached.	Report a Bug
flow-steward-publish-workflow	Workflow authors: share a workflow template in the public workflow catalog.	Publish a Workflow Template
flow-steward-extension-authoring	Developers: build extensions against current manifests, SDK and supported CLI lifecycle.	Extension Authoring Skill
flow-steward-publish-extension	Developers: publish an extension and its later versions to the public extension catalog.	Publish an Extension
flow-steward-contribute-code	Contributors: fix a bug or propose an improvement in the core repository, from the issue to a reviewed pull request.	Contributing Code

The extracted flow-steward/ directory is the actual plugin root. Root plugin.json targets Agent Plugins 1.0.0 for Codex and Cursor; .claude-plugin/plugin.json provides Claude Code metadata. Both discover the same skills/ sources. This skills-only package contains no endpoint, token, server process, hooks or dependencies. Configure MCP separately using your installation's connection instructions. The workflow, issue and workflow publishing skills work without a platform checkout; extension work and code contributions use the target release's contracts and the supported Development environment.

Install and verify

Review the package before installing. Extract an archive to obtain flow-steward/plugin.json, flow-steward/.claude-plugin/plugin.json and the six flow-steward/skills/<name>/SKILL.md files. Use the extracted flow-steward directory as the plugin root. Do not load only .claude-plugin/ or install the whole application repository as a plugin. Inspect an existing installation before replacing files; keep a rollback copy of the version you use.

Codex

Put the extracted directory at plugins/flow-steward inside a local marketplace root of your choosing. Create or extend .agents/plugins/marketplace.json there with this entry; preserve other entries:

{
  "name": "flow-steward-local",
  "plugins": [{
    "name": "flow-steward",
    "source": {"source": "local", "path": "./plugins/flow-steward"},
    "policy": {"installation": "AVAILABLE", "authentication": "ON_INSTALL"},
    "category": "Productivity"
  }]
}

From that marketplace root, inspect your installed CLI's help, then:

codex plugin marketplace add .
codex plugin list --marketplace flow-steward-local --available --json
codex plugin add flow-steward@flow-steward-local

Restart the client and verify the six skills in its available skills listing. The local marketplace is private; this package is not listed in a public marketplace. Codex packaging documentation explains local sources and cache behavior.

Cursor

Copy the extracted root to ~/.cursor/plugins/local/flow-steward, reload the window and inspect Customize for the six skills. Local imports must be allowed by your organization. A marketplace installation of the same name takes precedence; symlinks outside the local plugins directory are skipped. See Cursor local plugins.

Claude Code

For a session-local load, inspect claude --help and run:

claude --plugin-dir /absolute/path/to/flow-steward

Verify the skills in the session's skill listing, for example /flow-steward:flow-steward-workflow-authoring, /flow-steward:flow-steward-report-issue and /flow-steward:flow-steward-contribute-code. claude plugin validate /absolute/path/to/flow-steward checks the manifest; it does not prove tool connectivity or successful use. Persistent installation uses a configured Claude marketplace and its normal plugin lifecycle. See Claude plugin reference.

Optional individual skill installation

Copy any complete leaf under skills/ into one native skills directory: Codex .agents/skills/ or ~/.agents/skills/; Cursor .agents/skills/, .cursor/skills/ or their user equivalents; Claude .claude/skills/ or ~/.claude/skills/. Use the same canonical bytes and keep the leaf name. Standalone names are $flow-steward-workflow-authoring / $flow-steward-extension-authoring in Codex and the corresponding / names in Claude; the other skills follow the same pattern with their leaf names. For plugin installs use the client's discovered names; Claude adds its plugin namespace. Avoid duplicate standalone and plugin installs. Native paths are documented for Codex, Cursor and Claude.

Use and update

After client discovery, confirm the configured MCP connects and tools/list returns the expected token-scoped tools. Try a requested scoped read, such as listing workflows or inspecting a known job; document permission errors as errors. For example: "Find my project's existing workflow for weekly summaries", "Show failed runs and pending approvals for this workflow", or "Report activity for this project for the last seven days". Anyone can ask "Report this error as a Flow Steward bug" or "Publish this workflow to the Flow Steward catalog". Developers can ask to build, validate and publish an extension, or to fix a core bug and open a pull request.

On update, compare package version and hashes, replace the reviewed plugin root at the same source location, refresh/restart the client and verify skill discovery again. Codex loads a cached installed copy; confirm the refreshed install's bytes before claiming an update took effect. Session-local Claude loads use the new root on the next session. Cursor local copies need a reload. Marketplace-managed installs use that client's marketplace update lifecycle. Never overwrite MCP credentials while updating skills.

Choose a task

Users can ask the workflow skill to reuse a project’s existing workflow, inspect failed runs or pending approvals, or request an operational activity report. Specify the project/workflow and period; the agent reports scope, observed results and limits. It needs no platform source checkout. See Tools And MCP and MCP Reference for pagination, UTC periods and authorization. Basic MCP calls work without either skill.

Example requests and results

Reuse an existing workflow

In project Supplier Operations, find an existing workflow for weekly supplier updates. Show its required inputs and observed readiness before I choose a run.

Expect an accessible workflow’s identity, inputs and any readiness gaps, or a clear no-match result. Finding a workflow does not execute it.

Compose a supplier-attachment workflow

In project Supplier Operations, use my sample supplier attachment to design a workflow that extracts product updates for Shopify. Inspect the available workflow examples, schemas and component contracts, validate the draft, and save it disabled if my existing permissions allow.

Expect a draft based on the sample and accessible contracts, validation results, and either the saved disabled workflow identity or a clear dependency/permission limitation. This example depends on your installed integrations; it does not establish that an attachment reader or Shopify operation is available. Validation is separate from execution; enablement and scheduling require their own supported, authorized path. See workflow authoring.

Inspect failed runs and pending approvals

For project Supplier Operations and workflow supplier-product-updates, show failed runs and pending approvals. Explain the observed errors and gate context; leave approval decisions and recovery actions for me.

Expect scoped job IDs, statuses, errors or approval context, with coverage limits stated. See run inspection for complete-page and bounded-triage behavior.

Request an operational activity report

Report activity for project Supplier Operations and workflow supplier-product-updates for October 1–7, 2026, inclusive UTC calendar days. Show the exact scope, period, coverage and observed outcomes; do not persist it.

Expect an operational activity report for that stated window, or a permission or coverage limitation. These results describe activity rather than complete financial reporting. See report periods and authority for UTC ranges, local-day requests and generation permissions.

Developers use Extension Authoring Skill with the target application checkout, current manifests/SDK and the supported Development CLI to create, validate, test and package an extension.

Report a bug, publish, contribute

Report a bug

The supplier import job failed with "connector timeout". Report it as a Flow Steward bug.

The agent searches existing issues, asks you to download the System Diagnostics report (System → Monitoring → Download diagnostics, or ./flow-steward reports diagnostics) and to glance through it, collects the version, steps, expected and actual result and the job id, and shows you the draft. It files the issue only after you agree, and never pastes secrets or customer data. See Report a Bug or Request a Feature.

Publish a workflow template

Publish my workflow "Import supplier stock" to the Flow Steward workflow catalog.

The agent checks the name, description and version, exports the workflow with its child workflows, reviews the file with you for private data, and prepares the catalog pull request. See Publish a Workflow Template.

Publish an extension

Publish version 1.2.0 of my extension com.acme.crm-sync to the catalog.

The agent checks the manifest listing fields and icon.png, runs the gate's extension checks, confirms wheels for Linux AMD64 and ARM64, and walks through the release tag and, for a first listing, the catalog pull request. See Publish an Extension.

Contribute a fix

Fix issue #123 in Flow Steward and open a pull request.

The agent works in the Development environment, writes the failing regression test first, runs the lint.sh tiers, and opens a pull request with the evidence after you approve. See Contributing Code.

Individual skill downloads

The optional individual ZIPs copy the same canonical leaves; they are not separate instruction editions: workflow authoring, report an issue, publish a workflow, extension authoring, publish an extension and contribute code. Installing a skill does not grant permissions. Draft save needs an admin user with admin:all; machine tokens remain operators. Report generation needs reports.write even with persist=false. Enablement, execution and approval remain deliberate authorized actions.