Files

m3tm3re f20f5223d5 Create agents.json with 6 agent definitions (Wave 1, Task 1)

- Added all 6 agents: chiron, chiron-forge, hermes, athena, apollo, calliope
- Primary agents (2): chiron (Plan Mode), chiron-forge (Build Mode)
- Subagents (4): hermes (communications), athena (work knowledge), apollo (private knowledge), calliope (writing)
- All agents use model: zai-coding-plan/glm-4.7
- Prompt references use file pattern: {file:./prompts/<name>.txt}
- Permission structure: primaries have external_directory rules, subagents have simple question: allow
- Verified with Python JSON validation (6 agents, correct names)
- Documented patterns and learnings in notepad

2026-02-03 20:14:34 +01:00

2.2 KiB

Raw Blame History

Learnings - Chiron Agent Framework

Wave 1, Task 1: Create agents.json with 6 agent definitions

Agent Structure Pattern

Required fields per agent:

description: Clear purpose statement
mode: "primary" for orchestrators, "subagent" for specialists
model: "zai-coding-plan/glm-4.7" (consistent across all agents)
prompt: File reference pattern {file:./prompts/<name>.txt}
permission: Either explicit permissions or simple "question": "allow"

Primary vs Subagent Modes

Primary agents (2): chiron, chiron-forge

Can be invoked directly by user
Orchestrate and delegate work
Higher permission levels (external_directory rules)

Subagents (4): hermes, athena, apollo, calliope

Invoked by primary agents via Task tool
Specialized single-purpose workflows
Simpler permission structure (question: "allow")

Permission Patterns

Primary agents: Complex permission structure

"permission": {
  "external_directory": {
    "~/p/**": "allow",
    "*": "ask"
  }
}

Subagents: Simple permission structure

"permission": {
  "question": "allow"
}

Agent Domains

chiron: Plan Mode - Read-only analysis and planning
chiron-forge: Build Mode - Full execution with safety prompts
hermes: Work communication (Basecamp, Outlook, Teams)
athena: Work knowledge (Outline wiki, documentation)
apollo: Private knowledge (Obsidian vault, personal notes)
calliope: Writing (documentation, reports, prose)

Verification Commands

Agent count:

python3 -c "import json; data = json.load(open('agents/agents.json')); print(len(data))"
# Expected output: 6

Agent names:

python3 -c "import json; data = json.load(open('agents/agents.json')); print(sorted(data.keys()))"
# Expected output: ['apollo', 'athena', 'calliope', 'chiron', 'chiron-forge', 'hermes']

Key Takeaways

Prompt files use file references, not inline content (Wave 2 will create these)
Model is consistent across all agents for predictable behavior
Permission structure matches agent capability level (more complex for primaries)
Mode determines how agent can be invoked (direct vs delegated)

2.2 KiB Raw Blame History