workflow-pack

A curated, npx-installable pack of Claude Code dynamic workflows — install one, run its /slash-command, and get a multi-agent run that fans out and synthesizes a result.

Dynamic workflows are a research-preview feature in Claude Code v2.1.154+. A workflow is a JavaScript orchestration script that the runtime executes in the background, spawning up to 16 concurrent subagents. Saved workflows live in ~/.claude/workflows/ (user) or .claude/workflows/ (project) and surface as /slash-commands. See the docs.

The workflows

| Command | What it does | |---|---| | /codebase-survey [path] | Map a codebase's modules in parallel → one architecture overview. | | /bug-sweep [path] | Sweep code across 5 bug lenses in parallel, then adversarially verify each finding. | | /compare-options A vs B vs C | One researcher per option, then a scored comparison table. | | /dependency-audit [path] | Audit dependencies in parallel for outdated / deprecated / vulnerable packages. | | /test-gap-finder [path] | Find under-tested code and propose concrete missing test cases. | | /release-notes [git range] | Group recent commits by area, summarize in parallel, compile release notes. | | /doc-audit [path] | Check docs against the code to find stale, wrong, and missing documentation. | | /api-surface-map [path] | Discover endpoints, document each group in parallel, compile an API reference. | | /scaffold-feature <description> | Plan a feature, generate each file in parallel, then review. | | /jira-digest [JQL] | Pull Jira issues, analyze each group in parallel, produce a sprint/standup digest. |

Each one fans out into a multi-agent tree, so the run is genuinely parallel — and looks great when recorded (see Get reports).

Install (npm)

Interactive — pick what you want with arrow keys + space:

npx workflow-pack            # opens the picker: select workflows, choose user/project scope

Or install directly without the picker:

npx workflow-pack list                 # see what's available
npx workflow-pack install bug-sweep    # install one
npx workflow-pack install --all        # install all of them
npx workflow-pack uninstall bug-sweep  # remove one

• Installs into ~/.claude/workflows/ by default. Add --project to install into ./.claude/workflows/ (shared via your repo).
• Idempotent and reversible: re-running install is safe; uninstall only removes a file it recognizes (use --force to overwrite or remove a file you've since edited).
• After installing, reload Claude Code / start a new session to pick up the new /slash-commands.

Install (Claude Code marketplace plugin)

/plugin marketplace add samarthpatel24/workflow-pack
/plugin install workflow-pack@workflow-pack

Then run the picker command inside Claude Code:

/workflow-pack

It lists the workflows, lets you include/exclude the ones you want, asks for user or project scope, and installs just those — the same selection experience as the npm npx workflow-pack setup.

Run one

/bug-sweep src/api
/compare-options Postgres vs MySQL vs SQLite
/scaffold-feature add a /health endpoint returning JSON status

Claude Code asks to approve the workflow before it runs, then the run starts in the background. Watch it with /workflows. Input is passed via the args global — pass a path, a git range, a list, or a description as shown in the table.

Get reports of these runs

Pair workflow-pack with workflow-replay (the Dynamic Workflow Trace Recorder) to turn any of these runs into a shareable report.html:

1. Install the recorder's Stop hook (see its README).
2. Run a packed workflow, e.g. /codebase-survey.
3. Open the generated report.html — a rendered view of the multi-agent run.

The wiring is documentation-only: these workflows run fine on their own; the recorder is optional.

Notes

• Research preview: the feature is pinned to Claude Code ≥ v2.1.154 and the saved-workflow format may change. The workflows are plain .js files (each begins with export const meta = {...}) — easy to read and edit.
• Workflows can route stages to different models and spawn up to 16 concurrent / 1,000 total agents per run. A run uses more tokens than a single conversation — try a narrow scope first.

License

MIT