[FEATURE] Allow the Dynamic Workflows sandbox to execute external code / tools

Preflight Checklist

• [x] I have searched existing requests and this feature hasn't been requested yet
• [x] This is a single feature request (not multiple features)

Problem Statement

There are often very programmatic or deterministic things that need to happen as part of SDLC that are not accessible in the codebase repository itself, e.g updating statuses in project management tools, creating/updating PRs, kicking off builds, etc. Currently the pattern is to "ask" claude to do it, but it ends up being a suggestion and not something that is systematically enforced. I find the hooks are not flexible enough in this usecase. With the introduction of dynamic workflows / ultracode, this seems like a great place to address this issue. It is much better for claude to be able to do this as a harness script instead of agentic behavior, because it is deterministic and more efficient in terms of speed/tokens. However, currently the Workflow execution runtime is completely sandboxed, and you must delegate to an agent to get something done.

Proposed Solution

I would like to be able to inject scripts/npm packages/etc into the Workflow runtime environment that are invokable by the harness itself instead of from an agent. Ensuring that the agent writes these calls into the script seems like a higher probability of success, than instructing dozens/hundreds of agents in the workflow to do it, as one of those is bound to fail and disrupts the entire SDLC for that project.

Alternative Solutions

As mentioned above, you can instruct the agents to do the mandatory guaranteed things you want, but it becomes flaky over many subagent runs.

Priority

Medium - Would be very helpful

Feature Category

Configuration and settings

Use Case Example

I have a custom SDD+PM tool that I use to manage multiple subagents. It is a custom dashboard that let's me see the status of different items being worked on and then gives me a queue of code to review and merge when it's ready. I have a small cli script to manage these things and I instruct claude to call this cli to create tickets and update their progress, and then create the review request when it is complete. This all bundled into a custom SKILL. The skill basically does this, in short:

1. Research the codebase and figure out relevant files and functions.
2. Breakdown the request in small concrete actionable subtasks
3. Use the cli tool to create these subtasks
4. Launch subagents per subtask, and for each subtask use the cli to mark it as complete when the agent for that subtask complete is finished
5. When all agents complete, use the cli to create a review request.

Most of these are static, mandatory requirements, that fail sometimes because subagents don't follow the rules. If I could just create a workflow once, I could ensure that all the "bookkeeping" bits happen automatically.

Additional Context

_No response_