Production Agent Ops — Battle-Tested Architecture Pack#

Free primers included in this bundle:

• Context Death Spiral Prevention — understand compaction before installing
• OpenClaw Bash Safety — understand the security threat model first

These are free on ClawHub. Read them before installing this bundle.

---

What's In This Bundle#

7 production architecture files built for production Claude Code deployments. Each one is a complete SKILL.md with behavioral specification, exact constants, and setup guide.

Phase 1 — COMPACTION_ARCHITECTURE.md#

Token thresholds, 6-condition autocompact gate, 3-strike circuit breaker, post-compaction cleanup, microcompaction, recursion guards, summary structure. Eliminates context death spirals.

Phase 2 — LOOP_TERMINATION_ARCHITECTURE.md#

BudgetTracker state object, 5-condition termination logic, diminishing returns detection, stop hook execution sequence, recursion safety for nested agents. Prevents runaway token burn in autonomous loops.

Phase 3 — SESSION_MEMORY_ARCHITECTURE.md#

Dual memory systems (session + long-term), extraction agent protocol, UUID cursor integrity, forked agent cache-sharing pattern, drain protocol on shutdown. Your agent remembers what matters across sessions.

Phase 4 — BASH_SECURITY_ARCHITECTURE.md#

23 validators across 7 categories: text, structural, encoding, Bash-specific, Zsh-specific, persistence vectors, escalation vectors. Closes the attack surface ClawHavoc exploited.

Phase 5 — AGENT_MEMORY_SCOPING_ARCHITECTURE.md#

Three memory scopes, CLAUDE_CODE_REMOTE_MEMORY_DIR env var, snapshot system, file scanning, cost attribution, tiered model routing for memory operations. Persistent memory that survives WSL2 resets and session restarts.

Phase 6 — AEGIS_COORDINATOR_RESUME_INTEGRITY.md#

Mode mismatch correction on session resume, coordinator operational rules, worker spawning, continue-vs-spawn decision logic, verification standards, failure handling. Multi-agent coordination without losing state on resume.

Phase 7 — AEGIS_FORKED_AGENT_SKILL_ARCHITECTURE.md#

Forked agent pattern, cache-safe params (critical five fields), fork isolation, skill architecture, resolution order, change detection (300ms debounce), skillify workflow. 50-70% token cost reduction on extraction tasks via cache-sharing forks.

---

Setup#

Quick Install#

bash

# Create skills directory
mkdir -p ~/.openclaw/workspace/skills/production-agent-ops

# Copy all 7 SKILL.md files (downloaded from this package)
cp *.md ~/.openclaw/workspace/skills/production-agent-ops/

# Reload OpenClaw
openclaw gateway restart

Recommended Install Order#

Install phases in sequence. Each phase builds on the previous:

1. Phase 4 first (bash security) — harden exec before expanding agent capabilities
2. Phase 1 (compaction) — configure context management
3. Phase 2 (loop termination) — add budget governance to agent loops
4. Phase 3 (session memory) — enable cross-session memory
5. Phase 5 (memory scoping) — set CLAUDE_CODE_REMOTE_MEMORY_DIR for persistence
6. Phase 6 (coordinator) — enable multi-agent coordination
7. Phase 7 (forked agent) — add cache-sharing for cost reduction

Critical Environment Variable#

Phase 5 requires this env var to persist memory across WSL2 resets:

bash

export CLAUDE_CODE_REMOTE_MEMORY_DIR=/home/aegis/.openclaw/workspace/.memory

Add to ~/.bashrc and set in openclaw.json env vars. Without this, memory is WSL2-local and can be lost on reset.

---

What These Constants Are#

These are not theoretical values or documentation approximations. They were extracted from production Claude Code deployments and cross-referenced against production telemetry:

• Token thresholds: p99.99 summary output distribution
• Circuit breaker: failure distribution data from production compaction runs
• Bash validators: actual attack vectors from the ClawHavoc incident (341 skills)
• Memory scoping: CLAUDE_CODE_REMOTE_MEMORY_DIR behavior in Claude Code deployments

If Anthropic changes these values in Claude Code, this package will be updated. Current extraction date: 2026-03-31.

---

Pricing Rationale#

Individual skills: $19 each × 2 = $38 minimum for compaction + bash security alone. This bundle: $69 for all 7 phases.

The two phases not sold individually (loop termination, session memory, memory scoping, coordinator mode, forked agent) are available only in this bundle.

---

Compatibility#

• OpenClaw 2026.3.x and above
• All major models: claude-haiku-4-5, claude-sonnet-4-6, claude-opus-4-6
• WSL2 Ubuntu (primary test environment)
• Native Linux (compatible)
• macOS (compatible, minor path differences noted in Phase 5)