agent-prompts

Agent Prompts — Parallel-Terminal Audit

Self-contained prompts to paste into individual Claude Code terminals so each subagent runs fully autonomously, in parallel, and writes its findings to ./audit-reports/.

Two ways to run:

• MASTER (1 terminal, easiest) — paste MASTER.md into a single Claude Code session. Phase 1 runs inline; 14 Phase-2 agents dispatched in parallel via the Agent tool; Phase 3 synthesis. End-to-end, fully autonomous.
• 16 individual terminals (max isolation) — paste each numbered prompt into its own terminal as documented below.

16 prompts — designed to run in 16 individual terminals for max parallelism:

00-orchestrator.md                          ← Phase 3 (last): synthesises the 15 reports
01-threat-modeler.md                        ← Phase 1 (first): drives MAS profile selection
02-secrets-scanner.md                       ┐
03-sbom-vuln.md                             │
04-sast-dast.md                             │
05-supabase-rls.md                          │
06-supabase-auth.md                         │
07-supabase-edge-functions.md               │ Phase 2 (parallel —
08-supabase-postgres.md                     │ run 14 in parallel)
09-supabase-storage-realtime-network.md     │
10-tauri-capabilities.md                    │
11-tauri-ipc.md                             │
12-tauri-config-and-distribution.md         │  (CSP + WebView + updater + binary hardening)
13-mobile-static.md                         │
14-mobile-dynamic.md                        │
15-mobile-platform.md                       ┘  (deeplinks + Keychain/Keystore + cert pinning)

One-time setup (in YOUR APP REPO at ~/desktop/travus)

The audit runs from ~/desktop/travus (your Tauri app repo). audit-skills is installed inside the repo as ./audit/ (gitignored). Reports go to ./audit-reports/ (also gitignored). Every prompt assumes the working directory contains:

• src-tauri/ — Tauri Rust core + capabilities + tauri.conf.json
• supabase/ (or accessible via SUPABASE_DB_URL) — Edge Functions + migrations + tests
• android/ and/or ios/ if mobile is built in this repo (otherwise mobile agents skip)
• package.json — root npm dependencies
• audit/ — cloned audit-skills (gitignored)
• audit-reports/ — created on first run, gitignored

Zero-touch (recommended)

cd ~/desktop/travus
curl -fsSL https://raw.githubusercontent.com/user2343242kdisj/audit-skills-mobile-tauri-supabase/main/install.sh | bash

install.sh clones the repo into ./audit/, installs the exec-agent wrapper, creates ./audit-reports/, and adds both audit/ and audit-reports/ to .gitignore. Secrets live in 1Password and are resolved at runtime via op read — no .audit-env file is created.

Manual

cd ~/desktop/travus

# 1. Clone audit-skills INSIDE your app repo as ./audit/
[ -d ./audit ] || git clone https://github.com/user2343242kdisj/audit-skills-mobile-tauri-supabase.git ./audit

# 2. Audit reports directory + gitignore both audit/ and audit-reports/
mkdir -p audit-reports
grep -qxF "audit-reports/" .gitignore 2>/dev/null || echo "audit-reports/" >> .gitignore
grep -qxF "audit/"         .gitignore 2>/dev/null || echo "audit/"         >> .gitignore

# 3. Verify 1Password CLI is authenticated (agents resolve secrets via `op read`)
op vault list >/dev/null

The shared path is ./audit/templates/agent-prompts/.... If you prefer an env var for scripts:

export AUDIT_SKILLS_PATH="./audit"

Run pattern

Phase 1 — Threat model (sequential, runs FIRST)

# Terminal 1
cd ~/desktop/travus
claude --dangerously-skip-permissions  # auto-approve tool calls
# Then paste the contents of ./audit/templates/agent-prompts/01-threat-modeler.md

Wait for DONE | threat-modeler | … line. Output: audit-reports/01-threat-model.md.

Phase 2 — Domain auditors (parallel)

Open 14 separate terminals (or use tmux/zellij for split-pane). In each:

cd ~/desktop/travus
claude --dangerously-skip-permissions
# Then paste the contents of ./audit/templates/agent-prompts/<NN>-<name>.md

Each writes its report to audit-reports/<NN>-<name>.md.

You can also batch them with claude -p (non-interactive):

# In a launch script, looping over Phase 2 prompts (skip 00 and 01):
cd ~/desktop/travus
op vault list >/dev/null   # unlock 1Password once
for prompt in ./audit/templates/agent-prompts/0[2-9]-*.md \
              ./audit/templates/agent-prompts/1[0-5]-*.md; do
  name=$(basename "$prompt" .md)
  log="audit-reports/$name.log"
  (claude --dangerously-skip-permissions -p "$(cat "$prompt")" > "$log" 2>&1) &
done
wait
echo "All 14 Phase 2 agents complete."

Phase 3 — Orchestrator synthesis (sequential, runs LAST)

cd ~/desktop/travus
claude --dangerously-skip-permissions
# Then paste the contents of ./audit/templates/agent-prompts/00-orchestrator.md

Output: audit-reports/00-FINAL.md.

Autonomy guarantees

Every prompt embeds the same autonomy rules:

1. No questions to user mid-flight. Missing input → write BLOCKED: <reason> and exit cleanly.
2. No destructive operations. Audit only. No DROP, DELETE, force push, rm -rf outside /tmp.
3. Writes restricted to ./audit-reports/. Never modifies app code.
4. Reads from approved paths only: the app repo (including ./audit/), /tmp/.
5. Never pushes to git — the user reviews the report directory and acts on it.
6. Final stdout line is parse-friendly: DONE | <agent-name> | <CRITICAL> CRITICAL | <HIGH> HIGH | <report-path>.

What each agent produces

File	Contents
audit-reports/00-FINAL.md	Aggregated executive report; quotes from subagent reports with attribution
audit-reports/01-threat-model.md	DFD, STRIDE list, attack tree, MAS profile recommendation
audit-reports/02-secrets-scan.md	ggshield + TruffleHog + Gitleaks consolidated findings
audit-reports/03-sbom-vuln.md	CycloneDX SBOMs + Grype/Trivy/cargo-audit findings
audit-reports/04-sast-dast.md	Semgrep + Schemathesis + BOLA harness + ZAP results
audit-reports/05-supabase-rls.md	Splinter (28 rules) + pgTAP coverage + policy review
audit-reports/06-supabase-auth.md	GoTrue version, MFA, JWT migration, CVE-2026-31813 / 2025-48370 / GHSA-3529 checks
audit-reports/07-supabase-edge-functions.md	Deno lint + 13 custom Semgrep rules
audit-reports/08-supabase-postgres.md	Roles, grants, search_path, extensions, FDWs, 7 upstream PG CVEs
audit-reports/09-supabase-storage-realtime-network.md	Bucket + signed URL audit · realtime.messages policies + private channel check · TLS testssl.sh + Network Restrictions
audit-reports/10-tauri-capabilities.md	ACL invariants, high-risk identifiers, runtime add_capability, dep pin (CVE-2026-42184 / 2025-31477 / 2024-35222)
audit-reports/11-tauri-ipc.md	Commands (6-class vuln checklist), custom schemes, isolation pattern, channels, events
audit-reports/12-tauri-config-and-distribution.md	CSP/WebView (sec A) + updater Ed25519 (sec B) + binary hardening + signing (sec C)
audit-reports/13-mobile-static.md	MobSF + jadx + apktool + manifest red-flags + plutil/codesign/otool
audit-reports/14-mobile-dynamic.md	Frida pinning bypass + Burp + Objection memory dump + Drozer scanners
audit-reports/15-mobile-platform.md	Deeplinks/intents (sec A) + Keychain/Keystore (sec B) + cert pinning (sec C)

Skipping irrelevant agents

If your stack doesn't have a layer (e.g. no mobile app, just Tauri + Supabase), skip the mobile prompts. The orchestrator detects missing reports and notes them as N/A rather than failing.

Re-running after fix

After remediation, re-run only the affected agents:

cd ~/desktop/travus
# Example: re-run after fixing CVE-2026-31813
claude --dangerously-skip-permissions -p "$(cat ./audit/templates/agent-prompts/06-supabase-auth.md)"
# Then re-run orchestrator to refresh the synthesis
claude --dangerously-skip-permissions -p "$(cat ./audit/templates/agent-prompts/00-orchestrator.md)"

CI integration

For continuous auditing in CI, see templates/security-workflow.yml. The CI workflow uses the same tools but as deterministic GitHub Actions steps rather than agent-driven prompts. Both have a place:

• Agent prompts → human-in-the-loop deep audit; pre-launch
• CI workflow → mechanical regression gating; every PR