[colbymchenry]

codegraph_impact and codegraph impact now stop at a node and edge budget and say the answer was truncated, so one call on a very heavily referenced symbol can't run a process out of memory. This is the impact part of #1583, split out as its own commit.

Root cause

The impact walk had no limit. One call on a hub symbol at a high depth loaded every transitive dependent into the query worker or the MCP process.

Fix (from the PR)

getImpactRadius takes a budget, 10,000 nodes and 50,000 edges by default, merged into the depth-aware walk from #2000. Re-expanding a node from a nearer depth is still allowed; only new nodes and new edges spend the budget. Edge reads are limited in SQL with one row over the budget, so hitting it can be told from an exact fit. codegraph_impact says when the answer was truncated and suggests file or a smaller depth.

What we changed

• codegraph_impact uses the library's budget. The PR gave MCP 1,000 nodes and 5,000 edges. On excalidraw, a mid-sized repo, that truncated 25 of 75 busy-symbol queries, and because the walk is depth-first it dropped direct dependents an answer shows today (ExcalidrawElement at depth 5 kept 147 of its 546). With the library's budget all 75 are unchanged, and the limit still stops the case it is for.
• codegraph impact says so too, in its text and as truncated in --json. Without this the CLI would have stopped at the limit silently.
• Tests for the CLI and the CHANGELOG bullet.

Left as it is

The viewer's blast-radius numbers go through the same default budget and do not say when they stop there. The viewer is not released, and a count near 10,000 is already off any bar it draws.

Validation

• Red on main: 6 dependents past a budget of 3 at node level, 10,051 from the CLI, and no truncation note.
• Before and after: codegraph_impact on the 25 busiest symbols of excalidraw, gin and flask at depth 2, 3 and 5 gives byte-identical text on both arms (225 of 225).
• A synthetic graph with 60 callers per node, depth 3: 219,661 nodes, 174 ms, 76 MB of heap and 155 MB RSS on main; 10,000 nodes, 8 ms, 4 MB and 59 MB here.
• codegraph_explore doesn't use this walk, so its output and budgets are unchanged.
• Full suite on 2da21398, a stack of this commit and its sibling from fix: bound large-repo memory and traversal work #1583 on d494e43 (Vitest 4, CODEGRAPH_KERNEL_EXPECT=1, 4 workers): 842 files, 10,209 passed, 0 failed, 87 skipped. This commit then rebased cleanly onto b6ea617. Main's new commits share src/bin/codegraph.ts and src/mcp/tools.ts with it, with their hunks 723 and 12 lines away from this one's, and nothing this code calls changed. On this head, graph, cli-definition-grouping and the tests of those new commits pass: 7 files, 191 tests.

Lands #1583 (impact part) by @danusha2345.

🤖 Generated with Claude Code