thejefflarson Pull Request Security Gate (A01:2025–A10:2025 + LLM01:2025–LLM10:2025)
What this checks
Fast single-pass review of the files in a pull request diff. Reports
only Critical and High severity findings. Designed for CI: completes
in seconds-to-minutes on haiku, well under the time a developer
expects to wait on a PR check. This is the per-PR gate. For nightly
or deep audits, reach for security-review or contract-review.
Vulnerable patterns
Soundcheck's per-category skills (injection, csrf, ssrf,
broken-access-control, authentication-failures,
hardcoded-secrets, path-traversal, prompt-injection, …)
auto-invoke when their description matches code in the diff. Examples
this gate is expected to catch on a routine PR:
- SQL or shell string built from request data —
injection - Path opened from a URL parameter without normalization —
path-traversal - A new credential string-literal in a config file —
hardcoded-secrets - A new state-changing endpoint without
csrfProtection—csrf - A new fetch to a caller-supplied URL with no host allowlist —
ssrf
Procedure
This skill is the coordinator — it does the reading; per-category skills supply the patterns. No subagents.
- [ ] Read every file in the supplied changed-file list once
- [ ] Apply soundcheck's per-category skills as their descriptions match
- [ ] Filter findings to severity Critical or High
- [ ] Emit findings table only (no chains, no design review)
- [ ] Append the machine-readable <soundcheck-findings>[...] block
For each finding, the table row contains:
| Severity | File:Line | Skill | Finding | Fix |
- Severity is Critical or High; drop Medium and Low.
- File:Line must be inside the supplied changed-file list.
- Skill names the auto-invoking soundcheck skill the pattern matched.
- Finding is one short sentence a non-security developer can act on.
- Fix is one short sentence; link or quote the secure pattern from
the relevant skill's
Fix immediatelyblock.
If a pattern requires cross-file tracing or threat modeling to classify, do not emit it — that's mode 2's job. Better a clean miss in CI than a noisy false positive that trains developers to ignore the gate.
Finish with one summary line: N findings (M Critical/High) or
No Critical or High findings. Suggest /security-review for a
deeper manual scan.
Verification
- Every emitted finding is severity Critical or High
- Every emitted finding's file:line is inside the changed-file list
- No Agent tool invocations (single-pass, no subagent dispatch)
- Response ends with a one-line summary
- Machine-readable
<soundcheck-findings>block appended