Terminal use without a shell
Summary
MELRA runs a program with its own argument list and no shell in between, confines and bounds it, and checks policy again at the moment it runs.
Process latency
Key results
| Measurement | Result | Source |
|---|---|---|
| Verified executions | 30/30 | core-microbenchmarks.json |
| Process latency, p50 | 48.087 ms | core-microbenchmarks.json |
| Process latency, p95 | 54.492 ms | core-microbenchmarks.json |
MELRA launches an executable with a separate argument vector and shell: false: nothing in between parses the command line. It confines the working directory, filters environment variables, bounds output and time, redacts secret patterns, and supervises background jobs through their status, output, stop and shutdown.
Policy denies shell interpreters and privilege-escalation tools on its own terms, even when a caller adds one of them to the general command allowlist.
The result
The public microbenchmark ran the current Node executable 30 times with a fixed script, and checked both its output and its exit code. All 30 runs verified, with a median process latency of 48.087 ms. This measures the cost of starting a local process and capturing what it does. It does not measure whether arbitrary commands are correct, or safety against a compromised executable.
What we learned
- Keeping the executable apart from its arguments removes a large class of shell-injection behaviour. It does not make an allowed executable harmless.
- Command classification and evidence requirements have to be checked when the command runs, not only when it is planned.
- A timeout needs escalating termination and a truthful failed result.
- Output has to be truncated and redacted before it reaches durable task or receipt storage.
- Installing packages and commands that can reach the network need a higher risk class than read-only inspection.