Risk Assess

/risk-assess — Interactive Risk Assessment

Assess the vibe-coding risk level of the current repository by scanning code patterns, detecting modules, and interactively confirming uncertain dimensions with the user.

Reference model: Read .claude/skills/shared/risk-model.md before proceeding. It contains the dimension definitions, scoring tables, grep patterns, module detection strategy, mitigation signals, and the required output format. Do NOT duplicate those patterns inline — always read them from the shared file at runtime.

Step 1 — Module Detection

Detect the modules (independently assessable units) in this repository.

1a. Check Workspace Configs (confidence 0.9)

Look for these files in the repo root and parse the relevant field:

Resolve any glob patterns to actual directories.

1b. Conventional Directories (confidence 0.6-0.8)

If no workspace config is found, check for these directory patterns:

• packages/*/package.json — JS/TS monorepo
• apps/*/ with a build config — application packages
• services/*/Dockerfile — microservices
• frontend/ + backend/ — client/server split
• src/client/ + src/server/ — co-located client/server
• docker-compose.yml with multiple build: entries — multi-service

1c. Fallback

If neither workspace configs nor conventional patterns are found, treat the entire repository as a single module.

1d. Present and Confirm

Present the discovered modules to the user:

Detected modules:
  1. {module-name} ({path}) — detected via {method}
  2. ...

Does this look correct? Should I add, remove, or rename any modules?

Wait for user confirmation before proceeding.

Step 2 — Auto-Scan per Module

For each confirmed module, run automated detection for every dimension. Use the grep patterns defined in .claude/skills/shared/risk-model.md.

2a. Language Detection

Count files by extension within the module directory. Use the extension-to-score mapping from the shared risk model. The module's language score is the maximum score across all detected languages (weighted by file count — if only 1-2 files of a high-score language exist among hundreds of low-score files, note this).

Report:

Language scan for {module}:
  .ts/.tsx: 42 files (score 1)
  .js/.jsx: 8 files  (score 2)
  → Auto-detected score: 2 (Dynamically typed) — JS files present

2b. Code Type Detection

Search for patterns in the shared risk model, starting from the highest score (4) and working down. Stop at the first match level that has significant hits.

Report each match with the file and line:

Code Type scan for {module}:
  Auth/Security patterns (score 4):
    src/auth/login.ts:15 — matches "authenticate"
    src/middleware/csrf.ts:3 — matches "csrf"
  API/DB patterns (score 3):
    src/routes/users.ts:8 — matches "app.get("
  → Auto-detected score: 4 (Auth/Security/Crypto)

2c. Data Sensitivity Detection

Search for data sensitivity patterns from the shared risk model, starting from score 4 (PHI/PCI) down to score 2 (General PII).

Report matches with evidence:

Data Sensitivity scan for {module}:
  PHI/PCI patterns (score 4): no matches
  Sensitive PII patterns (score 3): no matches
  General PII patterns (score 2):
    src/models/user.ts:5 — matches "email"
    src/models/user.ts:7 — matches "phone_number"
  → Auto-detected score: 2 (General PII)

2d. Deployment Hints

Search for deployment/regulatory patterns from the shared risk model. Also check for:

• Dockerfile, docker-compose.yml — containerized deployment
• .github/workflows/, Jenkinsfile, .gitlab-ci.yml — CI/CD presence
• kubernetes/, k8s/, helm/ — orchestrated deployment

This dimension has low auto-detection confidence (0.2-0.5). Note findings but flag that user confirmation is required.

2e. Blast Radius Hints

Blast radius is nearly impossible to auto-detect. Note any hints:

• Number of downstream dependents (if library)
• Presence of health/safety keywords
• Scale indicators (load balancer configs, horizontal scaling)

Flag that user confirmation is required.

2f. LLM Runtime Integration Detection

Grep for LLM SDK imports and agentic patterns using the patterns in the LLM Runtime Integration section of .claude/skills/shared/risk-model.md.

Report findings with evidence:

LLM Runtime scan for {module}:
  LLM SDK imports: 3 files
    src/chat/client.ts:1 — imports "openai"
    src/summarize.ts:2 — imports "@anthropic-ai/sdk"
  Generative patterns (L2+): 2 files
    src/chat/client.ts:14 — matches "messages.create"
  Tool use / agentic patterns (L3+): 0 matches
  Code-execution sandbox (L4): 0 matches
  → Auto-detected hint: L2 (Generative)

If no LLM imports are found:

LLM Runtime scan for {module}:
  No LLM SDK imports found.
  → Auto-detected hint: L0 (No LLM)

Auto-detection produces only a hint — the user must always confirm the level explicitly in Step 3. A library import alone is ambiguous (could be test code, data-science notebooks, build tooling, or production).

Step 3 — Interactive Confirmation

Process modules one at a time. For each module, present the auto-scan results and ask the user to confirm or adjust.

3a. High-Confidence Dimensions

For dimensions with high confidence (language, codeType when score >= 3):

Language: Auto-detected score 2 (Dynamically typed)
  Evidence: 42 .ts files (score 1), 8 .js files (score 2)
  → Accept score 2? [Y/n]

3b. Low-Confidence Dimensions (ALWAYS ask)

For deployment and blastRadius (and any dimension with low confidence), present a multiple-choice question:

Deployment context for {module}:
  Auto-detected hints: Dockerfile found, GitHub Actions CI
  No regulatory keywords detected.

  What best describes the deployment context?
  [0] Personal / Prototype — local tools, learning projects
  [1] Internal tool — company-internal dashboards
  [2] Public-facing app — SaaS, public APIs, mobile apps
  [3] Regulated system — HIPAA, PCI-DSS, SOC2, GDPR-critical
  [4] Safety-critical — avionics, medical devices, automotive

  Suggested: [2] (containerized with CI/CD, no regulatory signals)

Blast Radius for {module}:
  What is the worst realistic impact of a bug in this module?
  [0] Cosmetic / Tech debt — UI glitches, code smell
  [1] Performance / DoS — slowdowns, service unavailability
  [2] Data loss (recoverable) — lost data restorable from backups
  [3] Systemic breach — unrecoverable data exposure
  [4] Safety (life & limb) — physical harm, loss of life

3c. Data Sensitivity Confirmation

If data sensitivity was auto-detected at score >= 2, confirm with the user:

Data Sensitivity: Auto-detected score 2 (General PII)
  Evidence: email, phone_number fields in user model
  Could there be higher-sensitivity data not detected by pattern scan?
  [Keep 2] / [Upgrade to 3: Sensitive PII] / [Upgrade to 4: PHI/PCI]

3d. LLM Runtime Integration Confirmation (ALWAYS ask)

Auto-detection produces only a hint — always ask the user to confirm, even if L0 was detected (the user may know about planned future LLM use).

Present the hint with evidence, then ask:

LLM Runtime Integration for {module}:
  Auto-detected hint: L2 (Generative)
  Evidence: openai + @anthropic-ai/sdk imported, messages.create pattern found in src/chat/client.ts

  How does this module use LLMs at runtime?
  [0] No LLM — classical software, no LLM at runtime
  [1] Classify — passive use: sentiment, intent, embeddings
  [2] Generate — generative output: chat, summaries
  [3] Tool Use — function calling, LLM triggers actions
  [4] Agentic — autonomous loops, code execution, self-modification

  Suggested: [2]

Remind the user of the tier implications before they answer:

Note: L3 forces at least Tier 3, L4 forces at least Tier 4, regardless of the code dimensions. A coding agent that could run rm -rf is safety-critical by definition.

Step 4 — Tier Calculation and Output

4a. Calculate Tier

base     = max(codeType, language, deployment, data, blastRadius)
baseTier = base <= 1 ? 1 : base <= 2 ? 2 : base <= 3 ? 3 : 4
floor    = llmRuntimeLevel >= 4 ? 4 : llmRuntimeLevel >= 3 ? 3 : 1
tier     = max(baseTier, floor)

Present the result. If the LLM Runtime modifier lifted the tier above the base, make that explicit:

{module} Risk Assessment:
  Code Type:        3 (API / DB Queries)
  Language:         2 (Dynamically typed)
  Deployment:       2 (Public-facing app)
  Data Sensitivity: 2 (General PII)
  Blast Radius:     1 (Performance / DoS)
  LLM Runtime:      L0 (No LLM)

  → Tier 3 — determined by Code Type = 3

Example with modifier lift:

{module} Risk Assessment:
  Code Type:        2 (Business Logic)
  Language:         2 (Dynamically typed)
  Deployment:       2 (Public-facing app)
  Data Sensitivity: 1 (Internal business data)
  Blast Radius:     2 (Data loss recoverable)
  LLM Runtime:      L4 (Agentic)

  Base Tier: 2 — Moderate
  → Tier 4 — lifted from Tier 2 by LLM Runtime L4 (Agentic)

4b. Scan Existing Mitigations

Before writing, scan for mitigation signals as listed in the shared risk model. Check for config files and CI workflow steps that indicate existing mitigations.

4c. Check for Existing Assessment

Before writing to CLAUDE.md:

• Check if CLAUDE.md already contains a ## Risk Radar Assessment section
• If it does, ask the user: "CLAUDE.md already contains a risk assessment. Overwrite it?"
• If the user declines, skip writing

4d. ADR Generation

For all tiers, offer to generate an Architecture Decision Record:

Generate an Architecture Decision Record (ADR nach Nygard)?
  [y/N]

If the user accepts:

1. Detect next ADR number:

   • Check if docs/adr/ exists; if not, create it
   • Find the highest existing ADR number (docs/adr/NNN-*.md)
   • Use next number (zero-padded to 3 digits)

2. Write ADR nach Nygard to docs/adr/NNN-risk-classification-{project}.md:

   • Title: # {NNN}. Risk Classification — {project/module}
   • Date: Current date (YYYY-MM-DD)
   • Status: Proposed
   • Context: Summary of dimension scores with reasoning for each module
   • Decision: Tier classification result, determining dimension
   • Consequences: Positive (security baseline), Negative (CI overhead, workflow changes)

   Use "ADR nach Nygard" as the semantic anchor — no need for a custom template. The LLM knows the format: Title, Status, Context, Decision, Consequences.

3. Arc42 integration (if applicable):

   • Check if docs/arc42/ exists

   • If yes, check if docs/arc42/chapters/09_architecture_decisions.adoc exists

   • If yes, append a reference to the ADR:

     === ADR-NNN: Risk Classification — {project}
     
     See link:../../adr/NNN-risk-classification-{project}.md[ADR-NNN] for vibe-coding risk assessment.
     
     **Status:** Proposed | **Date:** YYYY-MM-DD | **Tier:** {N}
     

4. Reference in CLAUDE.md:

   • Add a line to the Risk Radar Assessment header:

     _Architecture Decision: See [ADR-NNN](docs/adr/NNN-risk-classification-{project}.md)_
     

4e. Write to CLAUDE.md

Use the exact output format from .claude/skills/shared/risk-model.md under "CLAUDE.md Output Format". Write:

1. The assessment header with timestamp (and ADR reference if generated)
2. Per-module dimension table with scores, levels, and evidence
3. Tier result with determining dimension
4. Per-module mitigation status table

Insert or replace the ## Risk Radar Assessment section in CLAUDE.md. Preserve all other existing content in CLAUDE.md.

Important Guidelines

• One module at a time: Complete the full assess-confirm cycle for each module before moving to the next. Do not overwhelm the user with all modules at once.
• Show evidence: Always explain WHY a score was auto-detected. Include file paths and matched patterns.
• Respect the shared model: Read patterns from .claude/skills/shared/risk-model.md at runtime. Do not hardcode pattern lists.
• Tiers are cumulative: When listing mitigations, include all tiers up to and including the assessed tier.
• Be conservative: When uncertain, suggest the higher (more cautious) score and let the user downgrade if appropriate.
• Timestamp: Use the current date in YYYY-MM-DD format in the output header.