impact-analysis

name	impact-analysis
description	Analyze what would break if a symbol changes: finds callers, groups by risk level, assesses impact. Use when the user asks about blast radius, who uses a symbol, or is planning a refactor.
user-invocable	true
arguments	<symbol, file path, or change target>
allowed-tools	mcp__julie__fast_search, mcp__julie__fast_refs, mcp__julie__deep_dive, mcp__julie__get_context, mcp__julie__call_path, mcp__julie__blast_radius, mcp__julie__spillover_get, mcp__julie__manage_workspace

Impact Analysis

Analyze the impact of changing a symbol by finding all references and assessing risk. Use this BEFORE modifying widely-used symbols.

Process

Step 1: Resolve the change target

If the user gives you a symbol name, resolve the definition first so you know which file or symbol they mean:

fast_search(query="<symbol_name>")
deep_dive(symbol="<symbol_name>", context_file="<partial_file_path>", depth="overview")

Use context_file when the name is ambiguous. If the target is described conceptually rather than named exactly, try fast_search(query="<concept>", backend="semantic") or backend="hybrid" to find candidate symbols. Semantic/hybrid fast_search results are symbol-only; use explicit lexical for file paths and pure lexical comparison. Default fast_search may show labeled semantic fallback candidates only after an identifier-like unscoped lexical zero-hit. blast_radius(symbol_ids=[...]) is the tightest seed mode, but only use it when another Julie result already gave you concrete symbol IDs. If all you have is a definition file, use file_paths=[...].

Step 2: One-shot impact via blast_radius

blast_radius(file_paths=["<definition_file>"], max_depth=2, include_tests=true)

blast_radius is the primary entry point for impact analysis. One call returns ranked impacted symbols with why-reasons, likely tests, and (for revision-range seeds) deleted files. It walks the reference graph deterministically, so you don't have to chain get_context → fast_refs → deep_dive to build the same picture.

You can seed it three ways:

• file_paths=["src/foo.rs"] — default when you know the changed file but not a symbol ID
• symbol_ids=["<id>"] — tighter impact when another Julie result already gave you concrete symbol IDs
• from_revision=<number>, to_revision=<number> — advanced mode using Julie's canonical revision numbers, not Git refs or SHAs

If the impact list is large, the first page includes spillover_handle=br_xxx. Hold onto it for Step 4.

Step 3: Drill down into high-risk callers

For any impacted symbol you need to understand in depth (unfamiliar caller, ambiguous usage, high centrality), use the targeted tools:

fast_refs(symbol="<caller>", include_definition=true, limit=100)
deep_dive(symbol="<caller>", depth="context")

If deep_dive returns the wrong symbol (common names like new, result, config), use context_file to disambiguate:

deep_dive(symbol="<caller>", context_file="<partial_file_path>")

Step 4: Page long impact lists

If Step 2 returned spillover_handle=br_xxx, fetch the rest without rerunning the walk:

spillover_get(spillover_handle="br_xxx")

Keep paging until the handle stops appearing.

Reviewing what changed since a revision

For "what's the blast radius of everything that changed in a recorded Julie revision range?" use revision-range seeds:

blast_radius(from_revision=101, to_revision=108, include_tests=true)

This walks from every symbol touched in that range. Deleted files get reported in a separate section of the output because they have no symbols left to walk from. Do not pass branch names, SHAs, or tags here, the tool only accepts canonical numeric revisions.

Step 5: Sample Deep Dives on High-Risk Callers

For each high-risk file surfaced by blast_radius, deep_dive on the calling function to understand HOW the symbol is used:

• Is it called with specific arguments?
• Does the caller depend on the return type?
• Is it used in error handling paths?

If you need the shortest route from one surfaced caller into a downstream sink or shared dependency, use call_path after blast_radius:

call_path(from="<impacted_symbol>", to="<downstream_symbol>")

Categorizing Callers by Risk

blast_radius gives you ranked impact with why-reasons, but you still want to sort callers into tiers for the final report:

High Risk — Changes here could cause cascading failures:

• Entry points and core orchestration files (e.g., main entry, request handlers, routers)
• Files with 10+ references to this symbol
• Files that re-export or wrap this symbol

Medium Risk — Changes need careful testing:

• Feature implementation files
• Data access / storage modules
• Files with 3-9 references

Low Risk — Changes are isolated:

• Test files (paths containing test, tests, spec, __tests__, or test annotations)
• Files with 1-2 references

Step 6: Report

Impact Analysis: <change_target>
Definition: <file>:<line> (<kind>)
Centrality: <high/medium/low>

Total: <N> references across <M> files

High Risk (<count> files):
  src/handler.rs — 15 refs
    Callers: process_request, handle_error, validate_input
    Usage: Core request pipeline, changes here affect all tool calls

  src/database/queries.rs — 12 refs
    Callers: fetch_symbols, update_index
    Usage: Database layer, type changes would require migration

Medium Risk (<count> files):
  src/tools/search.rs — 5 refs
  src/tools/navigation.rs — 3 refs

Low Risk (<count> files):
  src/tests/search_tests.rs — 8 refs (test code)
  src/tests/handler_tests.rs — 2 refs (test code)

Likely Tests:
  src/tests/handler_tests.rs::test_process_request
  src/tests/integration/pipeline.rs::test_end_to_end

Recommendation:
  <1-2 sentences on how to approach this change safely>

Important Notes

• Always check test coverage — high-risk changes with no test references are especially dangerous. blast_radius surfaces likely tests via the include_tests=true flag, so use it.
• Type changes cascade — if the symbol is a type/struct, any field change affects all users
• Interface/trait changes are widest — changing an interface method, trait, or abstract class affects all implementors
• Cross-workspace: Call manage_workspace(operation="open", path="<path>") first, then pass the returned workspace_id to all tool calls