Adapters per stack

What spur can record for each stack, what each tracer actually emits, and what was verified. Only Python is traced in-process. For the others spur attaches a call tree the program's own tracer wrote (--calls) to the root step, and hashes the files you declare (--reads/--writes): declared I/O, marked "io": "declared" in the record, not observed I/O.

stack	calls	file I/O	status
Python ≥ 3.12	sys.monitoring, every function under the project root	sys.addaudithook: open, os.rename, os.remove, spawns	built; acceptance test passes on 3.12, 3.13, 3.14, 3.15
Python 3.11	none (no sys.monitoring; the page says so)	as above	built; calls unavailable by design
Rust	tracing spans written by tracing-chrome	declared only	importer built, tested on real tracing-chrome 0.7.2 output
Node	node --cpu-prof V8 profile, sampled	declared only	importer built, tested on a live node --cpu-prof run (Node 26)
Go	none	declared only	not built — see below

Python

Nothing to wire: spur_trace.py run -- <command> prepends scripts/spur_boot/ to PYTHONPATH and sets SPUR_TRACE_DIR, so site imports spur's sitecustomize in the root process and in every Python child that inherits the environment. A project's own sitecustomize, if any, still runs: spur's chains to it.

Audit events confirmed by running them on 3.12, 3.13 and 3.15 (the CPython docs tool does not carry the audit table): os.replace raises os.rename, os.unlink raises os.remove, os.open raises open with mode None and the flags, and imports raise open for each .py and __pycache__/*.pyc — which is why code files are listed apart from data and never enter the chain.

Not observable: python -I, -E or -S, a child given its own env= without PYTHONPATH, and os._exit (no atexit: the step's record is missing, so it shows as an untraced step).

Rust: tracing-chrome

tracing-chrome is a tracing-subscriber layer that writes Chrome trace events (docs.rs, 0.7.2). ChromeLayerBuilder::new().build() returns the layer and a FlushGuard; the file is complete only when the guard is dropped. Only functions with a span appear — #[tracing::instrument] or a manual span — so the tree is the instrumented subset, not every call.

let (chrome_layer, _guard) = ChromeLayerBuilder::new().file("trace.json").build();
tracing_subscriber::registry().with(chrome_layer).init();

python3 spur_trace.py run --calls chrome:trace.json --reads 'data/*' --writes 'out/*' -- cargo run --release

The real output it was tested on is scripts/fixtures/adapters/rust/tracing-chrome-0.7.2.json, from main.rs and Cargo.toml beside it: B/E pairs carrying .file and .line. The importer also reads X complete events and the {"traceEvents": [...]} object form.

Node: --cpu-prof

node --cpu-prof starts the V8 CPU profiler and writes the profile before exit, into --cpu-prof-dir (default the working directory) under --cpu-prof-name (default CPU.<date>.<time>.<pid>.<tid>.<seq>.cpuprofile). This is a V8 .cpuprofile (nodes, samples, timeDeltas), not Chrome trace events, and it is sampled: the tree's counts are samples, not calls, and a short function can be missing. Only frames whose script URL is under the project root are kept.

python3 spur_trace.py run --calls cpuprofile:prof -- node --cpu-prof --cpu-prof-dir=prof app.js

Go: not built

runtime/trace (trace.Start/trace.Stop, regions via trace.StartRegion or trace.WithRegion) writes Go's own binary execution trace, read by go tool trace. It is not the Chrome trace format, and its format is documented as unstable, so spur has no importer. A Go step can still be traced for declared I/O and the chain check; its call tree stays empty and the page says so.

File I/O outside Python

The issue that proposed spur names strace for Rust and Go. That is Linux-only and is not built; on macOS the equivalent needs System Integrity Protection relaxed. Use --reads and --writes: the runner hashes the declared reads before the command starts and the declared writes after it exits.