Trace the impact of a change in Prolaborate

Goal

Find what is connected to an element, one hop at a time, and report what is affected together with exactly how far the search went. This is for an architect or change manager assessing what a change to an element might touch. You end with a bounded impact answer stating the seed element, the depth, the relationship types, the counts, whether a cap or filter shaped the result and the identity it was read as, and optionally a saved Analyzer view.

Where it runs

This scenario needs browser control. It uses no MCP server operation.

How it works

A change manager at Westbrook Bank plans to retire the Payments Gateway and asks what breaks. Prolaborate answers one hop per read, so the assistant does not promise a complete blast radius. It starts from the element, expands outward one hop at a time, keeps the result readable, and reports what it found along with what it did not reach. If you want a visual, it can build an Analyzer view in Prolaborate on request.

Before you start

• Browser control (a built-in browser or Chrome) and a signed-in Prolaborate session. No server is needed.
• The element exists in an active Prolaborate repository and you can see it.
• If an Analyzer view will be saved, you agree to create an object on the Prolaborate tenant. There is no undo, and baselines do not cover Prolaborate objects.

Steps

1. You Sign in to Prolaborate in the browser.
2. AI Power Tools prol-start-here runs the preflight: identity, version and repositories. You confirm the repository.
3. AI prol-impact-analysis decides what is being promised. One hop is complete. "What breaks" is a bounded traversal. "Everything" is pushed back with a bounded offer. You agree the question, the depth (two hops by default) and the relationship types to follow.
4. AI prol-impact-analysis identifies the seed element by selecting it in the Repository Browser tree (the folder icon in the left rail). Search on an empty canvas returns "No results found" for everything and is not used to find the seed. You say which, if several elements share a name.
5. AI Power Tools prol-impact-analysis seeds the traversal with the element's traceability: its connections grouped by connector type, with the related elements under each.
6. AI Power Tools prol-impact-analysis expands one hop at a time, taking a whole frontier of elements per call. It de-duplicates elements, caps depth and node count, counts what a filter discarded, and records why it stopped. You agree before it expands to a third hop or beyond; the assistant shows the counts first.
7. AI prol-impact-analysis chooses which relationship field to follow: whether the follow list matches the connector's stereotype (for example Uses) or its base type (for example Association), and says which. You confirm. Following Uses answers a dependency question; following Aggregation answers a containment question.
8. AI Power Tools prol-impact-analysis resolves context: the packages of the affected elements.
9. AI prol-impact-analysis reports the result with its bounds: the seed, the depth, the relationship types, the elements and relationships found (at least N if a cap was hit), whether a cap was hit, what the filter discarded, and the identity.
10. AI prol-impact-analysis offers to save an Analyzer view, driving Prolaborate's own interface: create a view, set the title, place the seed from the Repository Browser, select the node, tick neighbors with the group-level checkbox, add the selection, set the layout and save. You agree to create the view; the title is set before the first save.
11. AI prol-impact-analysis keeps the view readable: it reads live node and connector counts after each expansion, toggles element-type and connector-type chips off, and chooses a directional layout (top-down, down-top, right-left, left-right, radial in, radial out) over the forced graph. You say when to stop expanding while the view is legible.
12. AI prol-impact-analysis cleans up only what this session created: it deletes a scratch view only if this session created it, then re-reads the list of views. You ask for the deletion. The assistant deletes nothing it did not create. If a view has to be left behind, it tells you which one, and why it is still there.

Step 3. The assistant never describes a capped traversal as a complete blast radius. The skill's wording is "three hops along Uses and Aggregation, 61 elements, stopped at depth 3". The relationship total is a floor, because an edge between two nodes first reached at the last depth is never fetched.

Steps 5 and 6. Every read is one hop. The assistant does not create, update or delete through the diagram interface that serves both Analyzer views and Prolaborate diagrams. It creates views by driving Prolaborate's own interface (step 10).

Step 6. A newly reached element arrives with all its edges to elements already in the set. In one test, two elements added at hop two took the graph from 26 nodes and 25 connectors to 28 nodes and 63 connectors. The budget is edges, not nodes.

Step 7. A filter written against Association and applied to data keyed by Uses silently matches nothing. That looks like "nothing is affected". A non-zero discarded count with an empty graph means the filter, not the model.

Step 9. An empty or small result has four possible causes: genuinely nothing, a cap, a filter that matched nothing, or a permissions boundary. The first three are instrumented. A permissions-trimmed read looks the same as a sparse one, which is why the identity is reported beside the result.

Step 10. Pressing Save on a new view creates it at once under the title Analyzer View, with no dialog. The "check all" control in the traceability panel did not tick anything when tried, so the group-level checkbox is used.

Prompts to use

Opening:

Follow-ups that drive each phase:

• "What directly depends on it? Is that complete?" (step 3, one hop)
• "Follow only the Uses relationships, and tell me whether you matched on stereotype or base type." (step 7)
• "Go to a third hop, but show me the node and connector counts first." (step 6)
• "How are the Customer Portal and the Core Banking System connected?" (steps 5 and 6; expect a path, or "no path found within the cap")
• "Save that as an Analyzer view called 'Payments Gateway retirement impact', laid out as a dependency chain." (steps 10 and 11)
• "Is that view private?" (expect: the assistant does not say whether a saved view is private or shared)

Recommended models

Ratings are guidance based on the kind of reasoning the scenario needs, not measured benchmarks.

Checking the result

• Correct. The one-hop neighbors match the Traceability panel in the Analyzer for the same element. The relationship field matched (stereotype or base type) was stated. Relationships are reported as "at least N" when a cap was hit. Red flag: a "complete" claim from a traversal that stopped at a cap.
• Complete, as far as it claims. The report states the depth reached, the relationship types followed, why the search stopped, what the filter discarded, and the identity. It does not claim a complete blast radius. A "not connected" answer is qualified as "within the cap".
• Nothing unintended changed. No model element or connector was changed. No Analyzer view exists unless you asked for one, and none is named Analyzer View. No view the session did not create was edited or deleted. After a delete, the list of views was re-read.
• Readable. If a view was built, node and connector counts were reported, and the assistant stopped expanding while the view was legible.

Variations

• One hop only. Steps 5 and 9, stated as complete.
• Two hops along dependencies. The default: Uses only, depth two, with the counts.
• How are X and Y connected? Traverse from one toward the other and report the path and its length, or "no path found within the cap".
• Read an existing view. List the saved views and describe one without editing it.
• Impact against the model rather than Prolaborate. For the same question against the EA repository itself, see Trace dependencies, which uses the model's own traversal operations.

Limitations

• Every multi-hop result is composed one hop at a time with the assistant's own cap. "Everything affected" is not a supportable claim.
• Edges grow faster than nodes. A graph can become unreadable at 28 nodes. Hop three is usually the one that ruins the picture.
• Search filters the canvas, not the repository. An empty canvas returns "No results found" for everything.
• The assistant does not say a saved view is private or shared.
• A saved view is a tenant object with no undo. Create one only on request. Delete only what this session created.
• Super Admin only. A restricted user may legitimately see a smaller graph. The identity is reported with the result.