claude-code-source-security-check-loop

name	claude-code-source-security-check-loop
description	Use when the user asks Claude Code to security-check source code, run secure code review, find or fix source security issues, run deterministic SAST or dependency/security scans, or loop fixes until security findings are resolved through the packaged claude-code-source-security-check-loop Riela workflow.

Claude Code Source Security Check Loop

Use this skill when the user asks for a source-code security check with deterministic scanning, harness-style recon, adversarial verification, and fix follow-up.

Workflow

• Package id: claude-code-source-security-check-loop
• Workflow id: claude-code-source-security-check-loop
• Dependency: claude-code-design-and-implement-review-loop

Standard Run

Run from the repository root:

riela package install claude-code-source-security-check-loop
riela package install claude-code-design-and-implement-review-loop
riela workflow run claude-code-source-security-check-loop \
  --variables '{"workflowInput":{"targetPath":".","runNetworkAudits":"false","maxFindings":50,"constraints":["Do not stage, commit, or push unless the user explicitly asks.","Do not revert unrelated dirty worktree changes.","Keep fixes narrowly scoped to verified security findings."]}}' \
  --output jsonl --verbose --no-auto-improve

For registry-backed one-off runs:

riela workflow run claude-code-source-security-check-loop \
  --from-registry \
  --registry default \
  --variables '{"workflowInput":{"targetPath":".","runNetworkAudits":"false","maxFindings":50}}' \
  --output jsonl --verbose --no-auto-improve

Set workflowInput.runNetworkAudits to "true" only when dependency audit commands may use network access. Use JSONL output for this loop so the session id and progress remain visible during long deterministic scans, agent triage, and fix-routing passes.

What The Workflow Checks

The workflow runs separate deterministic vulnerability-check methods before agent triage:

• Secret scan: high-confidence secret patterns and private keys.
• Gitleaks scan: runs gitleaks detect --no-git --redact when the executable is available, otherwise records a missing-tool coverage gap.
• Static source/SAST scan: risky-code heuristics and optional local evidence.
• Dependency audit scan: manifests and optional npm, pnpm, yarn, Bun, Python, Ruby, Rust, and Go audit commands when network audits are allowed.
• Supply-chain/config scan: risky package scripts, build-chain indicators, Docker/container config, CI, and infrastructure-as-code heuristics.
• Harness recon: read-only threat-model and attack-surface focus area partitioning based on Anthropic's defending-code-reference-harness patterns.
• Agent triage: repository-specific exploitability review using deterministic evidence and harness focus areas.
• Adversarial verification: independent grader/judge pass that verifies, deduplicates, reranks, and documents false positives before fix routing.

The design borrows useful patterns from Claude Code and Claude Code security skills: scoped review modes, security-area checklists, automated checks, OWASP framing, and findings with severity, location, and remediation.

Operating Rules

• Do not rely on agent judgment alone; use deterministic method output as routing evidence.
• Keep method outputs separate so a finding can be traced to secrets, gitleaks, static analysis, dependency audit, supply-chain/config, or agent triage.
• Delegate only adversarially verified high and medium fixes through claude-code-design-and-implement-review-loop.
• Rerun all deterministic methods after fixes.
• Accept low findings only as documented residual risk.
• Preserve user constraints about dirty worktrees, staging, commits, and pushes.