fix(cli): reject --max-steps < 1 (silent empty run read as a clean pass)

[dchaudhari7177]

Problem

--max-steps 0 (or negative) makes for step in range(max_steps) a no-op — the agent never runs, and the benchmark reports hijack=0 / leak=0 / utility=0, which looks like the model resisted the attack when in fact nothing ran.

Fix

Validate after arg parsing, matching the existing error style:

[leakgauge] --max-steps must be >= 1 (got 0)

with exit code 2, before any adapter/model work happens.

Tests

• --max-steps 0 → rc 2, nothing written
• --max-steps -3 → rc 2
• --max-steps 1 still runs and writes a summary

python -m pytest tests/ passes.

Fixes #24

[bamdadd]

Thanks for this, Dipak — the fix itself is spot on. --max-steps 0 making range(max_steps) a silent no-op that reads as a clean pass is exactly the kind of quiet-wrong-zero this benchmark should reject, and your validation (exit 2, [leakgauge] message, before any model work) is right where it belongs. The tests are good too.

The only thing red is a formatting nit, not your logic. CI's ruff format --check wants tests/test_cli_single_case.py reformatted — your new test_max_steps_one_still_runs has a multi-line main([...]) call that ruff would collapse onto one line. One command fixes it:

uv run ruff format tests/test_cli_single_case.py

(or uv run ruff format . to be safe), then commit and push to the same branch — CI will re-run automatically. That's the whole blocker; once it's green I'll merge and this closes #24. Nice catch on the bug.

[bamdadd]

Merged — thank you, Dipak. The formatting nit was the only thing standing between this and green, so I ran ruff format on the test file and pushed it to your branch for you (kept you as the author of the fix); CI went green and it's in. The bug itself is the valuable part: a silent range(0) no-op reporting hijack=0/leak=0/utility=0 would read as a model resisting the attack when nothing actually ran, and rejecting it up front with a clean exit is exactly right for a benchmark that must never report a quiet-wrong zero. Closes #24.

That's three landed now — really nice run. If leakgauge is useful to you, a star helps others find it and Watch will surface new releases; would love to keep seeing your contributions.