Add `mlx-stack integrate` command for AI agent harness onboarding

[weklund-agent]

Summary

Users who already have an AI agent harness installed (Hermes Agent, Claude Code, Cursor, etc.) currently have to manually edit config files to point at mlx-stack's endpoint. This is error-prone and requires knowing the config schema of both tools. A single command like mlx-stack integrate hermes would eliminate this friction.

Motivation

Setting up Hermes Agent to use mlx-stack today requires manually:

1. Editing ~/.hermes/config.yaml to set provider: "custom", base_url: "http://localhost:4000/v1", and model: "standard"
2. Setting context_length explicitly (auto-detection probes from 2M down, wasting time)
3. Routing auxiliary tasks (compression, vision) to the fast tier via separate config keys
4. Knowing which mlx-stack tier names map to which harness concepts

This is a common onboarding path — users discover mlx-stack because they want local inference for an agent they already use. Making this a one-command experience would significantly reduce setup friction.

Proposed UX

# Auto-detect and configure a known harness
mlx-stack integrate hermes

# List supported harnesses
mlx-stack integrate --list

# Dry-run: show what would change without writing
mlx-stack integrate hermes --dry-run

# Undo: restore the harness config to its previous state
mlx-stack integrate hermes --remove

Example output

$ mlx-stack integrate hermes

  Detected Hermes Agent v0.7.0 at ~/.hermes/

  Applying configuration:
    model.provider:       custom
    model.base_url:       http://localhost:4000/v1
    model.default:        standard (Qwen3-Coder-30B-A3B)
    model.context_length: 32768
    compression.summary_model:    fast (Qwen3-8B)
    compression.summary_provider: main

  ✓ Updated ~/.hermes/config.yaml
  ✓ Backed up original to ~/.hermes/config.yaml.bak

  Run `hermes` to start using local models.

Suggested initial harness support

Harness	Config Location	Key Mappings
Hermes Agent	~/.hermes/config.yaml	model.default → standard tier, compression → fast tier
Claude Code	~/.claude/settings.json	API proxy / MCP server config
Cursor	.cursor/mcp.json or settings	model endpoint override
Continue.dev	~/.continue/config.json	models[].apiBase
Open Interpreter	env vars	OPENAI_BASE_URL + OPENAI_API_KEY

Each integration would need to:

• Detect if the harness is installed (check config dir / binary)
• Read the current stack tiers and map them to harness-specific roles
• Back up the existing config before modifying
• Support --remove to cleanly revert

Additional considerations

• Tier-to-role mapping: Different harnesses have different concepts (Hermes has main + auxiliary, Claude Code has a single model). The integration should map tiers intelligently based on the harness's capabilities.
• Stack changes: If a user runs mlx-stack setup --add/--remove after integrating, the harness config could become stale. Consider a post-setup hook or a warning.
• Plugin system: If the number of supported harnesses grows, this could be plugin-based — each harness integration is a small module that knows how to read/write that harness's config.