<aside> 🛡️

Sovereign engineering is what happens when you treat your AI coding agent like production infrastructure: governed, secure, modular, and measurably reliable.

This manifesto documents a practical setup that prevents context drift, eliminates brittle imperative habits, and turns “AI coding” into a repeatable engineering system.

</aside>

<aside> 🧠

Goal

Persistent, context-aware coding.

</aside>

<aside> ⚙️

Mechanism

A modular instruction stack + functional patterns.

</aside>

<aside> 🏛️

Non‑negotiables

Privacy, PoLP, and post‑task memory.

</aside>

Why this exists: context drift is the real enemy

AI coding fails in predictable ways:

• Context drift: The agent forgets architectural decisions and re-invents patterns every session.
• Imperative relapse: Under time pressure, it falls back to try/catch soup, long methods, and leaky exception contracts.
• Governance gaps: Infra gets created without PoLP, without tags, without documentation, and without an audit trail.

Sovereign Engineering treats the agent as a first-class engineer inside a governed system. The system forces alignment before code and closes the loop after code.

Runtime Configuration: ~/.claude/settings.json

<aside> ⚙️

This file is the runtime control plane for Claude Code: it enables plugins, sets environment flags, defines guardrail permissions, and wires lifecycle hooks that keep memory in sync.

</aside>

<aside> 📝

Note: This settings.json is subjective to the specific setup and use-cases covered in this article (memory workflows, tool-guardrails, and multi-agent experimentation). Your optimal configuration can be very different depending on your risk tolerance, local environment, and the plugins/hooks you rely on.

</aside>

What your settings.json is doing

<aside> 🧪

1) Environment flags

• CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1: Enables experimental agent teams / multi-agent workflows.
• CLAUDE_MEM_MODE=code--zh: Selects the memory mode used by claude-mem (tuned for code-centric capture). </aside>

<aside> 🧩

2) Plugin enablement

• enabledPlugins.claude-mem@thedotmack=true: Turns on the claude-mem plugin so memory can be injected/captured automatically.
• extraKnownMarketplaces.thedotmack: Pins the plugin source to GitHub (thedotmack/claude-mem) so Claude knows where to resolve it. </aside>

<aside> 🔒

3) Guardrail permissions (deny list)

You explicitly deny reading sensitive locations and dangerous shell commands.

• Secrets/identity stores:
  • Read(…/.ssh/**), Read(…/.aws/**), Read(…/.gnupg/**)
  • Read(…/**/.env*) (blocks common dotenv secret sprawl)
• Noise + supply-chain risk:
  • Read(…/node_modules/**), Read(…/vendor/**)
• Destructive commands:
  • Bash(rm -rf /), Bash(rm -rf ~)
  • Bash(git push --force)
  • Bash(curl *|bash), Bash(wget |bash)

This enforces PoLP + blast-radius reduction for an agent that can use tools.

</aside>

<aside> 🔁

4) Lifecycle hooks (memory automation)

You’ve wired hooks so memory is injected at the start, captured after tool use, and finalized on stop:

• SessionStart (matcher: startup|resume|clear|compact)
  • Runs npx claude-mem inject (timeout 60s)
  • Purpose: hydrate context at the start of a session so the agent begins aligned.
• PostToolUse (matcher: .)
  • Runs npx claude-mem capture (timeout 120s, async: true)
  • Purpose: capture new learnings/decisions after tool actions without blocking your flow.
• Stop (matcher: .*)
  • Runs npx claude-mem stop (timeout 120s)
  • Purpose: graceful shutdown / finalization of memory workflow. </aside>

~/.claude/settings.json