SKILL.md
Verify a Bun runtime change
Build and drive the debug binary directly — never bun test, never import-and-call.
Build
bun bd --version # builds ./build/debug/bun-debug and prints its version
Drive
For any JS-visible change, run the debug binary with -e and observe stdout:
bun bd -e '<repro>' # builds, then runs; sets BUN_DEBUG_QUIET_LOGS for you
For worker/subprocess-shaped changes, spawn a subprocess (still -e) so worker teardown / event-loop-idle paths are exercised. Cross-check against node -e '<same repro>' for Node-compat changes.
Gotchas
- A
src/js/edit can silently not reach the binary.**bundle-modules
regenerates build/<cfg>/codegen/InternalModuleRegistryConstants.h, but the C++ TU that embeds it is not always recompiled, so the build succeeds while the binary still runs the OLD JS. Gate on the binary, not the build: ask the binary you just built — bun bd -e 'console.log(<Class>.toString().includes("<new-id>"))' (or run ./build/<cfg>/bun directly). Plain bun is the system Bun on $PATH and never has your edit, so it answers about the wrong binary. If false, touch src/jsc/bindings/InternalModuleRegistry.cpp and rebuild.
- Prefix every
bun bdwithPATH="$HOME/.cargo/bin:$PATH"— Homebrew'srust
formula shadows the pinned nightly, and bun bd dies with the option 'Z' is only accepted on the nightly compiler. bun bd re-runs cargo on every invocation, so this is needed for follow-up runs too, not just the first build.
node:clusterchanges can't be driven with-e:cluster.fork()re-execsargv[1], so workers need a real file on disk. Write a scratch script and run./build/debug/bun-debug <file>.- Only one
bun bdper worktree at a time — a second one blocks on the build lock and looks like a runtime hang. Build once, then drive./build/debug/bun-debugdirectly undertimeout. BUNDEBUGQUIET_LOGS=1suppresses debug-build log spam.- Debug builds print
[cachefs]/[sys]lines to stdout; filter them before diffing
output against node.
- MessagePort's
.on/.offare added by requiringworker_threads— plainnew MessageChannel()ports only haveaddEventListeneruntil then. - The debug+asan build is 10-100× slower than release; large-allocation stress tests can time out locally while passing in CI.