Local Compute Usage

Local Compute Usage

Use this skill when work must stay on local compute across service and hardware surfaces.

Workflow

1. Run local preflight to validate workspace, VS Code CLI availability, endpoint host locality, hardware readiness commands, and live loopback accounting where available.
2. Pin runtime URLs and guardrail env vars to loopback/local values and normalize any legacy Documents\GitHub repo aliases to G:\GitHub.
3. Route execution through existing local service and hardware skills (especially MemryX shim lanes).
4. Fail closed if non-local hosts, missing live stack evidence, or hardware-check failures are detected when the lane requires them.

Local Preflight

Run from the active workspace root:

python3 "$CODEX_HOME/skills/local-compute-usage/scripts/local_compute_preflight.py" \
  --workspace-root . \
  --require-vscode \
  --url http://127.0.0.1:9000/health \
  --url http://127.0.0.1:10000 \
  --probe-http \
  --strict-probe \
  --hardware-check "acclBench --hello" \
  --stack-health-url http://127.0.0.1:9000/health \
  --stack-summary-url http://127.0.0.1:9000/api/accounting/summary \
  --json-out /tmp/local-compute-preflight.json

For non-MemryX hardware lanes, replace --hardware-check with lane-appropriate local hardware commands.

MemryX Shim-First Routing

When the user is actively working on MemryX shim flows:

1. Use $repo-b-control-center-ops for service startup/restart and endpoint readiness.
2. Use $repo-b-hardware-first for real-hardware diagnostics and deterministic probe evidence.
3. Use $memryx-official-hardware-check when official service checks fail (127.0.0.1:10000 path).
4. Use $repo-b-thin-waist-routing and $repo-b-mcp-comfy-bridge for API/MCP routing diagnostics.
5. Use $repo-b-local-bridge-orchestrator for read-only bridge orchestration.

Local Environment Defaults (PowerShell)

$env:REPO_B_SIDECAR_URL = "http://127.0.0.1:9000"
$env:REPO_B_FORCE_LOCAL_ONLY = "1"
$env:REPO_B_CONTINUE_BRIDGE_ENABLED = "1"
$env:REPO_B_CONTINUE_MODE = "read_only"
$env:REPO_B_CONTINUE_ALLOWED_ROOTS = "G:\GitHub\<PRIVATE_REPO_B>"

Guardrails

1. Do not route required service traffic to non-local hosts unless the user explicitly authorizes it.
2. Do not skip required hardware checks for hardware-backed workflows.
3. Prefer local scripts, local MCP adapters, local service endpoints, and local hardware probes over cloud-hosted alternatives.
4. Treat hostname drift or hardware-check failure as a policy violation and stop before mutation steps.
5. If the stack publishes positive displacement preview or healthy TPK, use that as evidence for stronger local-first routing rather than an optional hint.
6. Treat direct LM Studio http://127.0.0.1:1234/v1 as the authoritative model lane, http://127.0.0.1:2337/v1 as an explicitly selected hosted alternative, and http://127.0.0.1:9000 only as non-model MX3 support.

Scope Boundary

Use this skill when the task is primarily about enforcing local compute usage policy and local service/hardware routing.

Do not use this skill for:

1. Deep repo-specific debugging that already has a dedicated skill lane.
2. Remote cluster/cloud execution plans.

For those lanes, route through $skill-hub and the most specific matching skill.

References

• references/local-first-checklist.md
• scripts/local_compute_preflight.py

Loopback

If this lane is unresolved, blocked, or ambiguous:

1. Capture current evidence and failure context.
2. Route back through $skill-hub for chain recalculation.
3. Resume only after the updated chain returns a deterministic next step.