blast-radius's embedded [deps] section is nondeterministic across identical runs #153

Version: 4.4.2 (Homebrew) · macOS

Summary

Running blast-radius repeatedly on an unchanged repo at a fixed ref produces different function and dependency counts each time. Only the embedded [deps] sub-section varies; the rest of the output is stable.

Repro

Same repo, same ref, no edits between runs:

codemap blast-radius --text --ref <ref> .   # Go repo, 4 consecutive runs

Actual

242 functions · 8 deps
174 functions · 3 deps
191 functions · 7 deps
253 functions · 2 deps

Same on a Swift repo, and the format also shifts between output modes:

--text:      72 functions / 57 / 63
--markdown:  81 / 52 / 81

Scope — this is narrower than it looks, which may help localize it

These are fully deterministic (5/5 identical runs each):

• codemap --deps --json .
• codemap --importers <file> (also verified set-identical to grep -rl, 31/31)

And within blast-radius itself, the importer counters are stable run to run:

files_with_dependents=4
impacted_outside_diff=12/24

So the standalone deps scanner is fine. Only the copy of it embedded in blast-radius races.

Likely cause

Looks like a concurrent scan that's cut off by blast-radius's character budget or a timeout — whichever workers have finished when the budget is hit is what lands in the output. That would explain why counts vary downward-ish and why the standalone command, which has no such budget, is stable.

Why it matters

Anything comparing two blast-radius runs — CI gating, a diff of before/after, an LLM told "here is the dependency context" — sees changes that didn't happen. Nondeterministic output also makes the tool hard to snapshot-test.

Suggested fix

Collect results deterministically before applying the budget: gather the full set, sort by a stable key, then truncate. If the truncation is intentional, report it (showing 174 of N functions) rather than presenting a partial count as the total.