ai-augmented-sdlc

name	ai-augmented-sdlc
description	AI-assisted development governance — Copilot custom agents, MCP servers, change control, audit trail — per Hack23 ISMS AI Policy
license	MIT

AI-Augmented SDLC Skill

Context

Applies whenever GitHub Copilot, custom agents, MCP servers, or any LLM-based tool participates in a code, docs, or configuration change.

Codifies the requirements from:

• Information Security Policy — AI-First Operations Governance
• Secure Development Policy — AI-Augmented Development Controls
• AI Policy
• Change Management

Rules

1. AI outputs are proposals, not authority — every AI-generated change needs human review before merge
2. No autonomous deployment — AI may not bypass CI gates, branch protection, signing, or approvals
3. Human accountability — the merging human owns the change regardless of AI assistance
4. Document AI assistance — PR descriptions note the agent/skill used (e.g., "assisted by Copilot frontend-specialist")
5. Security gates unchanged or tightened — AI may never weaken linting, CodeQL, tests, or branch protection
6. Least-privilege tooling — each agent uses tools appropriate to its role; widening is a governed change
7. Curator-agent changes are Normal Changes — edits to .github/agents/*.md, .github/skills/*/SKILL.md, .github/copilot-mcp*.json, and .github/workflows/copilot-setup-steps.yml require CEO or security-owner approval
8. MCP governance — MCP server config uses secrets.COPILOT_MCP_GITHUB_PERSONAL_ACCESS_TOKEN (never hard-coded tokens) and is reviewed like other security-sensitive code
9. Audit trail — every agent action is captured in PRs, commits, CI runs, and issue history
10. Risk assessment for capability expansion — new MCP servers, new integrations, broader tool scopes need a documented risk note
11. Policy-cited commits — security-relevant AI-generated changes cite the applicable ISMS policy
12. No secret leakage through prompts — never paste credentials, tokens, PII, or production data into prompts or shared memory
13. Avoid data-dependency leaks — do not accept AI output that fabricates API signatures, options, URLs, or dependencies

Agent Tiers (per ISP §AI-First Operations)

Tier	Purpose
Curator-agent	Maintains agent + MCP + workflow configuration
Task agents	Product analysis, issue creation, ISMS mapping
Specialist agents	Domain-specific implementation (game / UI / test / security / docs)

Copilot Coding-Agent Tool Catalogue

Tool	Purpose
assign_copilot_to_issue(owner, repo, issue_number, [base_ref], [custom_instructions])	Start autonomous implementation on an issue
create_pull_request_with_copilot(owner, repo, title, body, [base_ref], [custom_agent])	Create a PR where Copilot implements changes
get_copilot_job_status(owner, repo, id)	Track agent job progress / completion

base_ref use cases

Scenario	base_ref
Feature branch	feature/new-auth
Stacked PR	copilot/issue-123
Release branch	release/v2.0

custom_instructions template

- Follow existing patterns in src/components/
- Add tests (≥ 80 % coverage, ≥ 95 % on security paths)
- Cite ISMS: Secure Development Policy §Phase 3 in the PR body
- Do not add new dependencies without `npm audit` + license check

Review Checklist for AI-Assisted Changes

• PR description documents the agent/skill used
• Diff reviewed for fabricated imports, APIs, URLs
• No hardcoded secrets, tokens, or PII introduced
• No weakening of validation, encoding, auth, or authorization
• No new network calls or eval/Function usage without justification
• Tests added; coverage thresholds met
• ISMS policy cited where applicable
• Dependencies audited + licensed; SBOM still valid
• CodeQL / Scorecard / Dependabot remain green

Change Management for Agent / MCP / Workflow Edits

1. Open an issue describing the capability and risk
2. Draft the change on a dedicated branch
3. Reviewer verifies: least-privilege, secret wiring, policy citations, audit trail
4. CEO or designated security owner approves
5. Merge behind branch protection; monitor agent behavior after merge

Anti-Patterns

• ❌ Merging an AI PR without reading the diff
• ❌ Disabling a failing security test to make CI green
• ❌ Adding a new dependency the agent "said was fine"
• ❌ Pasting secrets or production data into a prompt or agent memory
• ❌ Silently widening a tool scope (e.g., tools: ["*"] for a specialist who previously had a narrow set) without review

Remember

AI accelerates delivery; humans remain accountable for security, quality, and compliance. Every change passes the same gates regardless of authorship. Policy citations, audit trail, and least-privilege are non-negotiable.