Shim Pc Control Brain Routing

Shim PC Control Brain Routing

Use this skill when the operator wants:

• all apps to share one brain lane
• VRM Sandbox to act through PC Control governance
• Direct LM Studio http://127.0.0.1:1234/v1 remains the model authority; MemryX http://127.0.0.1:9000 remains typed non-model support.

Non-Negotiable Contract

1. Direct model-plane endpoint stays LM Studio 1234; MX3 9000 is support-only.
2. Do not introduce a new public default port for chat routing.
3. Do not switch workstation mode/profile unless explicitly requested.
4. Normalize repo roots to G:\GitHub; treat legacy Documents\GitHub paths only as aliases to report and correct.
5. Treat the hosted 27B lane on http://127.0.0.1:2337/v1 as authoritative alongside the shim public plane.
6. Treat http://127.0.0.1:1234/v1 as the direct model authority; select :2337/v1 explicitly for hosted execution and never route model traffic to :9000.
7. Do research before edits:
   • PC Control local-agent or status-surface evidence first
   • bounded sub-agents for repo evidence only after first-party evidence is captured

Mandatory Research Workflow

1. Capture live lane state before edits:
   • 127.0.0.1:9000
   • 127.0.0.1:2337
   • 127.0.0.1:1234
   • 127.0.0.1:8890
   • 127.0.0.1:5175
2. Query PC Control local-agent or status evidence first:
   • GET /v1/agent-fabric/contracts/endpoints
   • GET /v1/agent-fabric/models/local
   • POST /v1/agent-fabric/local-agent/chat (research-only prompt)
3. Spawn bounded sub-agents for:
   • PC Control repo endpoint contract and adapter points
   • shim repo route and connector selection path
   • avatar runtime repo OpenAI-compatible request/response expectations
4. Keep local sidecars preferred; if cloud sidecars are unavoidable, keep them lower-reasoning and low-cost.
5. Build a grounded patch plan with file/line references before any implementation.

Canonical Checks

Invoke-RestMethod http://127.0.0.1:9000/v1/models?health=1
Invoke-RestMethod http://127.0.0.1:2337/v1/models
Invoke-RestMethod http://127.0.0.1:9000/api/agent/auth-status
Invoke-RestMethod http://127.0.0.1:8890/v1/agent-fabric/contracts/endpoints
Invoke-RestMethod http://127.0.0.1:8890/v1/agent-fabric/models/local
Invoke-RestMethod http://127.0.0.1:5175/runtime/health

Implementation Rules

1. Keep VRM support/status calls on MX3 support lane (9000) and model calls on direct LM Studio (1234) unless the operator explicitly selects hosted :2337.
2. Route behavior changes behind the shim contract (connector/policy/adapter), not by changing operator-facing endpoints.
3. Keep model authority shim-first:
   • manual loading is owned by shim/control center
   • consumer apps read routed state; they do not self-select alternate brains
4. Verify post-change behavior with real calls, not UI-only assumptions.
5. Do not let MX3 :9000 or other support ports become the model authority source in docs, scripts, or validation examples.

Required Evidence

• pre-change and post-change endpoint status with status codes
• exact files changed for shim, PC Control, and VRM Sandbox
• one successful routed chat request proving one-brain behavior
• explicit note if any lane remains degraded

Scope Boundary

Use this skill for cross-repo routing alignment among:

• the shim repo
• the PC Control repo
• the avatar runtime repo

If the request expands beyond these lanes, route through $skill-hub.

Reference

• references/routing-contract.md

Loopback

If evidence is incomplete or contradictory:

1. stop implementation edits
2. rerun sub-agent + PC Control evidence pass
3. resume only after the direct :1234 model-authority and :9000 support-only contracts are re-confirmed