Aevatar Scripting Probe
Use this to confirm the code_execute sandbox works in the current chat, across every language a
workflow might use. code_execute takes {language, code} where language is one of
python / javascript / typescript / bash, runs with no approval, and returns the program
output (or an error JSON).
You (the agent) run the four snippets yourself and check the output. Each snippet has a known deterministic answer — report PASS only if the real output matches. Never fabricate a result; a sandbox error is a finding, quote it verbatim.
How to run it
Call code_execute once per language with exactly these snippets (each has one correct answer):
- python —
code: "print(sum(range(1, 11)))"→ expect55 - javascript —
code: "console.log([1,2,3,4].reduce((a,b)=>a+b,0))"→ expect10 - typescript —
code: "const xs: number[] = [2,3,4]; console.log(xs.reduce((a,b)=>a*b,1));"→ expect24 - bash —
code: "echo $((6*7))"→ expect42
For each: PASS if the returned output contains the expected number, FAIL otherwise (quote the error / wrong output). Do NOT retry a passing language; retry a failing one at most once.
Output
A compact matrix: language | expected | observed | PASS/FAIL. Then a verdict line:
SCRIPTING OK (all four pass) or SCRIPTING DEGRADED: <which languages failed + verbatim error>.
Common failure signals and what they mean (report, don't fix):
No NyxID access token available→ the brokered identity is broken for this conversation.- sandbox slug / discovery error → the code sandbox service is not connected for this caller.
- one language fails but others pass → that runtime is unavailable in the sandbox image.
Guardrails
- Only run the four bundled probe snippets; never run arbitrary or user-supplied code from this skill (it is a fixed health check, not a code runner).
- The snippets are side-effect-free; never add network, filesystem, or long-running code.
- Report verbatim outputs; never claim a pass you did not observe.
- Never ask the user for tokens — your NyxID-brokered tools handle all credentials.