fix(osint-autopilot): close remaining domain-arg traversal at all three entry points

[elementalsouls]

Follow-up to #14 — stacked on top of it (branched from that PR's head, so the diff here includes it). Merge #14 first and this rebases clean; or merge this alone and #14 is subsumed.

#14's guard is the right idea in the right place (before ENG= is built, before mkdir -p), but it leaves the hole open in two ways and covers one of three entry points.

1. . and .. still pass

^[A-Za-z0-9.-]+$ matches them:

$ bash recon_pipeline.sh ..
[*] engagement dir: /Users/me/Research/engagements/..

mkdir -p "$ENG"/{evidence/{...},assets,findings,notes,buckets} then scatters the whole evidence tree into ~/Research. A leading - also passes (-rf, --help), reaching whois "$D" (recon_pipeline.sh:26) and gau --subs --threads 5 "$D" (:82) as an option rather than an operand.

2. Two of three entry points had no guard at all

Both Python scripts build the same path from the same unvalidated sys.argv[1]:

• findings_gen.py:7 — ENG = os.path.expanduser(f"~/Research/engagements/{D}")
• build_xlsx.py:10 — identical, plus OUT = f"{ENG}/{D}-osint-consolidated.xlsx"

Verified by removing the new guard and re-running the self-check:

FAIL: findings_gen.py did not reject '../../tmp/pwned'
  (got: WROTE /tmp/sandbox/Research/engagements/../../tmp/pwned/findings/findings.csv)

Fix

Full dotted-hostname match, applied at all three entry points:

[A-Za-z0-9]([A-Za-z0-9-]*[A-Za-z0-9])?(\.[A-Za-z0-9]([A-Za-z0-9-]*[A-Za-z0-9])?)+

Requires at least one label separator, so ./../nodot are out; forbids leading/trailing - per label, so option injection is out. example..com, example.com., and example.com;id are rejected too.

Check

scripts/test_domain_guard.sh — no network, $HOME sandboxed to a temp dir:

$ bash skills/osint-autopilot/scripts/test_domain_guard.sh
PASS: domain guard holds on all three entry points

It asserts on the guard's own invalid domain string rather than a non-zero exit, so a script failing for an unrelated reason can't make it pass vacuously — confirmed by deleting a guard and watching it fail (output above). It also asserts the sandbox is empty afterward, and that a real domain still passes.

Note, not fixed here

.github/workflows/lint.yml pins shellcheck to scandir: ./scripts, so the scripts under skills/*/scripts/ — including the one this PR adds — are never linted. Fixed separately in #17. (An earlier revision of this description said that path doesn't exist; it does, it just holds one file.)

🤖 Generated with Claude Code