feat: Hermes Agent harness — non-interactive CLI + messaging gateway modes

[c-h-]

Summary

Add Hermes Agent (NousResearch) as a new agentctl harness with two distinct adapters:

1. hermes adapter — non-interactive CLI mode via hermes chat -q
2. hermes-gateway adapter — messaging gateway mode (persistent agent with API)

These are meaningfully different execution models and warrant separate adapters.

Why

Hermes Agent is a hybrid CLI + platform agent from NousResearch (MIT, launched Feb 2026). It has unique capabilities:

• Auto-managed persistent memory (MEMORY.md + USER.md)
• Autonomous skill creation at runtime
• Multi-model runtime switching (live, not just config)
• 6 terminal backends including serverless (Daytona, Modal)
• FTS5 full-text session search across all sessions

Adapter 1: hermes (CLI mode)

Non-interactive coding agent, analogous to Claude Code --print or Slate -q.

agentctl launch hermes --worktree ~/personal/orgloop --branch fix/foo -p "fix the bug"

Implementation:

• Launch: hermes chat -q <prompt> (non-interactive)
• Need to test: streaming output format, exit codes, session ID scheme
• Persistent memory carries across sessions (unique vs other harnesses)
• Skills accumulate over time (agent gets better at repeated task types)
• hermes chat --resume <id> for session continuation

Key unknowns (test during implementation):

• Does hermes chat -q produce streaming JSONL or plain text?
• Exit code conventions (0 = success, non-zero = error?)
• Can we pass workspace directory? (check --workspace or --cwd flag)
• Does it support --dangerously-set-permissions equivalent?

Adapter 2: hermes-gateway (messaging gateway mode)

Long-running agent with an API endpoint, analogous to an OpenClaw actor.

agentctl launch hermes-gateway --port 7878 --workspace ~/personal/orgloop

Implementation:

• Launch: start Hermes messaging gateway (persistent process)
• Send tasks via API (HTTP POST to gateway endpoint)
• Receive results via webhook callback or polling
• Agent maintains memory/skills across tasks
• Multiple tasks can be sent to the same running agent

This mode is more like a managed service than an ephemeral CLI. agentctl would:

• Start/stop the gateway process
• Track which gateway is serving which workspace
• Route tasks to the right gateway instance

Key unknowns:

• Gateway API format (REST? WebSocket? gRPC?)
• How to send tasks programmatically
• Callback/webhook support for task completion
• Can multiple workspaces share one gateway?

Shared concerns

• Installation: pip install hermes-agent (Python 3.11+) — first Python harness
• Config: Hermes uses its own config format (not slate.json style)
• Memory persistence: Hermes auto-manages MEMORY.md — this may conflict with worktree isolation (agent remembers things from other tasks). Could be a feature or a bug depending on use case.

Implementation plan

1. Install Hermes, test hermes chat -q output format and behavior
2. Build src/adapters/hermes.ts (CLI mode first — simpler, known pattern)
3. Test with real tasks in worktrees
4. Investigate gateway API, build src/adapters/hermes-gateway.ts if API is suitable
5. Document the two modes and when to use each

References

• NousResearch Hermes Agent (search for official repo)
• Evaluation: Obsidian Personal/research/hermes-agent-evaluation.md
• Related: feat: Slate harness — swarm-native coding agent via CLI + ACP #140 (Slate harness)