Add pluggable provider routing for local and hosted backends

[Pro777]

Summary

Downstream integration PR for Spitfire-Cowboy/ml-intern.

This branch does two distinct things:

• pulls in the relevant upstream provider-layer work (#55, #60, #63, #66)
• adds Rowan-specific runtime, policy, and config work on top of that provider layer

This is the fork-side answer to upstream issue huggingface/ml-intern#67.

Review map

1. Provider architecture and upstream-aligned changes

• adapter-based LLM resolution
• config-driven provider catalog
• Bedrock support
• model-switch auth hardening

2. Rowan and local-backend support

• openrouter/<model>
• lm_studio/<model>
• opencode/<model>
• opencode-go/<model>
• mlx/<model>
• ollama/<model>
• bld/<model>
• openai-compat/<model>

3. Rowan operator usability

• local lane reachability hints in /model
• opt-in rowan-mcp wiring from env
• default safe allowlist for Rowan MCP tools
• Rowan preset config profiles
• runtime preset selection via env

Rowan-specific defaults

• mlx/ -> MLX_SERVER_URL default http://localhost:8080/v1
• ollama/ -> OLLAMA_BASE_URL default http://localhost:11434/v1
• bld/ -> BLD_OLLAMA_BASE_URL default http://192.168.50.252:11434/v1

Rowan MCP behavior

• ML_INTERN_ENABLE_ROWAN_MCP=1 auto-injects rowan-mcp if the config does not already define it
• ROWAN_MCP_URL and optional ROWAN_MCP_TOKEN drive that injected server
• empty Rowan auth headers are stripped before MCP client creation
• default Rowan MCP exposure is a safe allowlist, not the full Rowan surface
• policy overrides:
  • ROWAN_MCP_ALLOW_ALL_TOOLS=1
  • ROWAN_MCP_ALLOWED_TOOLS=...
  • ROWAN_MCP_BLOCKED_TOOLS=...

Config profiles

Available presets:

• default
• rowan
• rowan_research_safe
• rowan_ops_full

Selection:

• ML_INTERN_CONFIG_PRESET=<preset>
• ML_INTERN_CONFIG_PATH=/abs/path/to/config.json overrides preset selection

Tests

Ran:

uv run --with pytest pytest tests/ -v --tb=short

Result:

• 29 passed

Upstream split

Reasonable upstream candidates:

• generic OpenAI-compatible routing
• provider registry/catalog shape
• local backend ergonomics

Fork-only pieces:

• Rowan MLX and BLD defaults
• Rowan MCP wiring and policy surface
• Rowan preset configs

[Pro777]

Added a follow-up commit on this branch: local Rowan-style providers now expose reachability metadata in the model catalog, and CLI shows / for local lanes (, , , ).\n\nThis uses a short TTL probe against , can be disabled with , and is covered by tests.

[Pro777]

Follow-up on this branch: local Rowan-style providers now expose reachability metadata in the model catalog, and CLI /model shows lane up or lane down for local lanes (mlx/, ollama/, bld/, lm_studio/).

This uses a short TTL probe against each configured base URL plus /models, can be disabled with ML_INTERN_SKIP_PROVIDER_PROBES=1, and is covered by tests.

[Pro777]

Added Rowan MCP wiring on top of the provider work.

What changed:

• ML_INTERN_ENABLE_ROWAN_MCP=1 now injects a rowan-mcp server automatically when the config does not already define one.
• ROWAN_MCP_URL and optional ROWAN_MCP_TOKEN drive the injected server.
• Empty MCP auth headers are stripped before client creation, so local Rowan setups do not send a junk Authorization: Bearer header.
• Added config and tool-router tests for the new behavior.

This keeps Rowan MCP opt-in, env-driven, and usable for both local and remote/proxied deployments.

[Pro777]

Added default Rowan MCP policy on this branch.

ml-intern no longer exposes the full Rowan MCP surface by default. When rowan-mcp is configured, Rowan-owned tools now go through a safe allowlist focused on read/search/research actions. Queue mutation, notify, and other higher-risk Rowan tools stay hidden unless explicitly enabled.

Policy controls:

• ROWAN_MCP_ALLOW_ALL_TOOLS=1 to disable the safe allowlist
• ROWAN_MCP_ALLOWED_TOOLS=... to add specific tools
• ROWAN_MCP_BLOCKED_TOOLS=... to force-block specific tools

Tests cover the default allowlist and the override behavior.

[Pro777]

Added two Rowan-ready preset configs on this branch.

Profiles:

• configs/rowan_research_safe_config.example.json: mlx/..., lower iteration cap, intended to pair with the default Rowan MCP safe allowlist
• configs/rowan_ops_full_config.example.json: bld/..., longer iteration cap, intended for stronger operator sessions

The existing configs/rowan_agent_config.example.json stays as the backward-compatible Rowan example. Tests now validate that both new preset configs load through the real config loader.

[Pro777]

Added runtime config selection on this branch.

You can now select Rowan profiles without copying files over configs/main_agent_config.json:

• ML_INTERN_CONFIG_PRESET=default
• ML_INTERN_CONFIG_PRESET=rowan
• ML_INTERN_CONFIG_PRESET=rowan_research_safe
• ML_INTERN_CONFIG_PRESET=rowan_ops_full

ML_INTERN_CONFIG_PATH=/abs/path/to/config.json overrides the preset choice.

Both interactive and headless entrypoints now use the shared resolver, and tests cover preset resolution, explicit path override, and bad preset rejection.