security: fix shell injection risk in _goal_coverage

[shigel]

Summary

autoforge/harness/run_goal.py:129 passed test_cmd (read from autoforge.yml) to subprocess.run(..., shell=True). If the YAML config comes from an untrusted source, this allows command injection.

Fix

# Before
result = subprocess.run(test_cmd, shell=True, timeout=300)

# After
import shlex
result = subprocess.run(shlex.split(test_cmd), timeout=300)

shlex.split handles quoted arguments correctly. This preserves the existing behavior for normal commands like pytest -v while eliminating the shell injection vector.

How this was detected

This issue was detected automatically by the Survey2Ship s2s scan command:

$ s2s scan --repo-root /path/to/autoforge
Scanned 74 files in 0.1s
High: 4, Medium: 2, Low: 0
  [MEDIUM] autoforge/harness/run_goal.py:125 — Unsafe pattern: shell_injection_risk

Survey2Ship is a research-driven autonomous improvement loop. This finding
is based on claim-0003 (arxiv:2604.07223)
which identifies agent execution traces as an under-explored vulnerability surface.

Verification

After the fix:

$ s2s scan --repo-root /path/to/autoforge
High: 4, Medium: 1, Low: 0   (was Medium: 2)

The shell_injection_risk finding is gone. The remaining 4 HIGH findings are in tests/core/test_hooks.py and appear to be intentional fake credentials for testing.

---

Generated by Survey2Ship SecurityScanner