feat: add HermesClaw runtime adapter for Hermes Agent

[willamhou]

What

Add a HermesClaw runtime adapter to support deploying Hermes Agent (by Nous Research) on Kubernetes via k8s4claw.

Why

Hermes Agent is one of the fastest-growing open-source AI agent frameworks (40k+ GitHub stars). It features:

• Self-improving learning loop — generates reusable skills from experience
• Persistent memory — SQLite + FTS5 across sessions
• Multi-platform — Slack, Discord, Telegram, WhatsApp, Signal, Email
• 40+ built-in tools — code execution, web browsing, file management

Currently, Hermes Agent supports Docker and SSH backends but lacks native Kubernetes orchestration. k8s4claw can fill this gap — managing StatefulSets, PVCs (for memory persistence), channel sidecars, and auto-updates.

How

1. Runtime Adapter (internal/runtime/hermesclaw.go)

Implement RuntimeAdapter for Hermes Agent:

• Image: ghcr.io/nousresearch/hermes-agent:latest (or k8s4claw-wrapped version)
• Gateway port: TBD — Hermes uses a REST API gateway, need to check the port
• Probes: HTTP health check at /health or /api/status
• Config mode: DeepMerge — Hermes uses .env + JSON config
• Resources: ~500m CPU / 1Gi RAM (similar to OpenClaw)
• Shutdown: Graceful drain with skill save

2. IPC Bus Bridge

Hermes Agent has a built-in gateway process. Need to determine:

• Does it expose a WebSocket/REST/SSE endpoint for message I/O?
• Can we reuse an existing bridge (WebSocket or SSE)?
• Or do we need a new bridge protocol?

3. Persistence

Critical for Hermes — its value comes from persistent memory:

• Session PVC: SQLite database (~/.hermes/memory.db)
• Workspace PVC: Skills directory (~/.hermes/skills/)
• Both should use reclaimPolicy: Retain by default

4. Channel Integration

Hermes has its own multi-platform gateway (Slack, Discord, Telegram, etc.). Two approaches:

• Option A: Use Hermes's built-in channel support, bypass k8s4claw's channel sidecars
• Option B: Disable Hermes channels, use k8s4claw's IPC Bus + channel sidecars for consistency

Option A is simpler for v1. Option B provides unified channel management across runtimes.

5. Sample CR

apiVersion: claw.prismer.ai/v1alpha1
kind: Claw
metadata:
  name: hermes-assistant
spec:
  runtime: hermesclaw
  config:
    model: "nous-hermes-3"
    learningEnabled: true
  credentials:
    secretRef:
      name: hermes-api-keys  # LLM provider keys
  persistence:
    session:
      enabled: true
      size: 5Gi               # SQLite memory DB
      mountPath: /data/memory
    workspace:
      enabled: true
      size: 20Gi              # Skills + knowledge
      mountPath: /data/skills
  channels:
    - name: slack-team
      mode: bidirectional

Investigation Needed

• Check Hermes Agent's gateway port and API protocol
• Determine if existing bridge (WS/SSE/TCP) works or need a new one
• Decide channel approach: native Hermes channels vs k8s4claw sidecars
• Check Hermes Docker image structure (entrypoint, volumes, env vars)

References

• Hermes Agent GitHub
• Hermes Agent Docs
• Hermes Agent Docker Setup
• Existing runtime adapter pattern