Orchestration & Rules

Operational governance layer: multi-agent coordination, IDE conventions, repository context, task directives, and hardcoded safety gates.

AGENTS 🤖 Ego Playbook

Codebase architecture, multi-agent coordination rules, and mandatory boot/memory sequences.

Agent Types

Coordination Rules

coordination:
  max_concurrent_children: 3
  max_spawn_depth: 1  # No nested delegation
  orchestrator_enabled: false
  context_isolation: true  # Subagents know nothing of parent
  verification_required: true  # All external side-effects must return handles

Mandatory Boot Sequence

graph TD
    A[Gateway Start] --> B[Load config.yaml]
    B --> C[Source ~/.hermes/.env]
    C --> D[Load pinned skills]
    D --> E[Inject MEMORY]
    E --> F[Inject USER profile]
    F --> G[Verify IDENTITY_HASH]
    G --> H[Ready]

Memory Injection Protocol

Every agent turn receives:

1. MEMORY (semantic base) — full
2. USER profile — full
3. IDENTITY_HASH markers — full
4. Recent FEEDBACK-LOG — last 5 entries
5. Active PLAN — current goal state

CLAUDE 📜 Cascading Conventions

IDE/directory-scoped guidelines for build steps, testing, and linting (loads local > global).

Hierarchy

Global (~/.claude/CLAUDE.md)
    │
    ├─ Project (~/.hermes/hermes-agent/CLAUDE.md)
    │
    ├─ Project (~/hugo-llm-wiki/CLAUDE.md)
    │
    ├─ Project (~/svelte-flexiboards-dashboard/CLAUDE.md)
    │
    └─ Project (~/.hermes/plugins/omp-suite/CLAUDE.md)

Convention Categories

Loading Protocol

1. Search upward from working directory for CLAUDE.md
2. Merge: global → project → subdirectory (last wins)
3. Apply to all terminal/execute_code operations in that context

CONTEXT 🌐 Background Scope

General repository background and project scope discovered upward from working directory.

Discovery Process

# On session start or workdir change
find . -maxdepth 3 -name "README*" -o -name "AGENTS.md" -o -name "CLAUDE.md" -o -name ".cursorrules" | head -20

Context Package

guidelines ⚖️ Task Directives

Technical instructions nested in .junie/ defining exactly how the agent updates checklists.

Structure

.junie/
├── guidelines.md          # Master directives
├── checklist-updates.md   # How to modify tasks
├── verification.md        # Verification standards
└── handoff.md             # Subagent handoff protocol

Core Directives

# .junie/guidelines.md
 
## Checklist Updates
- Only ONE item `in_progress` at a time
- Mark `completed` IMMEDIATELY when done
- If failed → `cancelled` + add revised item
- Never batch-complete without verification
 
## Verification Standards
- External writes: return handle (URL, path, HTTP status)
- Verify handle before reporting success
- Internal state: read back to confirm
 
## Subagent Handoff
- Pass ALL context: paths, errors, constraints
- Specify language/tone if non-English
- Require verifiable handles in summary

OPS 🔧 Protocol Gates

Hardcoded operational boundaries, credentials, and safety validation checks for execution.

Hard Boundaries

Credential Handling

# NEVER in execute_code heredocs
# ALWAYS via Python env=dict
import os
env = {k: v for k, v in (line.split('=', 1) for line in open('.env'))}
subprocess.run(cmd, env={**os.environ, **env})

Safety Validation Checklist

• No secrets in terminal output
• Cron jobs source .env
• Subagents verify external writes
• Memory writes are declarative facts
• Skill patches address root cause
• Deploy verified via live URL check

Cross-References

• Identity Anchors → SOUL constraints bind all AGENTS
• Mutable Memory → SESSION_LOG feeds AGENTS boot
• Capabilities → SKILL executes within OPS gates
• Autonomic → BOOT runs AGENTS boot sequence
• Spec-Driven → guidelines drive tasks execution

Orchestration & Rules is the constitutional layer — it governs HOW the system operates, not WHAT it knows.