/harden

Harden Codebase Command

Active security hardening for the existing codebase, paired with a report and concrete proposals you can apply. Complements the Claude Code built-in `/security-review` (which scans pending diff) by scanning the whole repository against citation-backed checks from NIST SSDF, CWE Top 25, and language-specific frontier practices for Python and Rust.

Usage

# Default: full codebase, report + interactive proposals
/harden

# Narrow to a focus area
/harden --focus python
/harden --focus rust
/harden --focus deps
/harden --focus secrets
/harden --focus ci
/harden --focus hooks
/harden --focus all

# Read-only report; no proposals applied
/harden --report-only

# Apply approved proposals without prompting per-file
/harden --auto-apply minor

# File GitHub issues for findings above a severity
/harden --file-issues high

# Dry-run: produce the report without writing anything
/harden --dry-run

# Restrict to one tier of audit depth
/harden --tier 1   # git-history triage only
/harden --tier 2   # targeted areas
/harden --tier 3   # full codebase

What It Does

1. **Discovery.** Detects languages, build files, CI workflows, hooks, and secret-bearing config in the repo. 2. **Citation-backed scan.** Runs the `pensive:harden` skill, which composes existing skills:

• `pensive:rust-review` for deep Rust audits
• `leyline:supply-chain-advisory` for dependency posture
• `leyline:authentication-patterns` for auth/credential review
• `leyline:content-sanitization` for input handling
• `abstract:hook-authoring` for hook-security checks
• `pensive:safety-critical-patterns` for NASA Power-of-10
• `pensive:tiered-audit` for tier discipline

3. **Findings classification.** Each finding ships with a severity, a CWE/NIST citation, the detection signal in code, and a concrete proposed remediation. 4. **Active hardening proposals.** For approved findings, the command proposes a specific diff or config change with a blast-radius assessment (per `pensive:blast-radius`). 5. **Approval gate.** You decide per-finding: apply, file as issue, defer to backlog, or reject. 6. **Validation.** After applying, re-runs gates (tests, lint, type-check) to confirm no behavior change. 7. **Report.** Markdown report saved to `reviews/harden-<date>.md`; summary posted to GitHub Discussions via `abstract:post_review_insights` when run on a PR.

Scope

• **Python** (frontier 2025-2026): unsafe-deserialization

alternatives, tarfile member filter (PEP 706), pip-audit / osv-scanner, typosquatting and dependency confusion defenses, PEP 740 sigstore attestations, async TOCTOU patterns, MCP server hardening, prompt-injection patterns in LLM SDK clients.

• **Rust** (frontier 2025-2026): `cargo-audit` / `cargo-deny` /

`cargo-vet` / `cargo-supply-chain`, `#![forbid(unsafe_code)]` coverage, `subtle` and `zeroize` for sensitive-data paths, `loom` model-checking opportunities, RustSec advisories.

• **Cross-cutting**: SLSA build-level posture, SBOM generation

(CycloneDX/SPDX), gitleaks/trufflehog secret scanning, pre-commit security hooks, container hardening when Dockerfiles are present.

• **CI/CD**: GitHub Actions pinned-by-SHA discipline, OIDC

publishing, restricted token scopes.

• **Hooks**: Claude Code hook-event security per

`abstract:hook-authoring`.

Out of Scope

• Pending-diff review (use Claude Code built-in `/security-review`).
• Architecture-level threat modeling (use a war-room session

via `attune:war-room`).

• Cryptographic protocol review (specialist work; the skill flags

but does not propose crypto fixes).

Output

• `reviews/harden-<date>.md`: full markdown report with

findings table, citation column, and per-finding proposal.

• Optional GitHub issues for findings above the chosen severity.
• Applied diffs as discrete commits per finding (one finding =

one commit, so revert is per-finding).

See Also

• `Skill(pensive:harden)`: the skill this command invokes
• `pensive:rust-review`, `pensive:bug-review`,

`pensive:safety-critical-patterns` (composed)

• `leyline:supply-chain-advisory`,

`leyline:authentication-patterns`, `leyline:content-sanitization` (composed)

• Built-in `/security-review`: pending-diff scanner (different

scope)