Local Compute Usage
Use this skill when work must stay on local compute across service and hardware surfaces.
Workflow
- Run local preflight to validate workspace, VS Code CLI availability, endpoint host locality, hardware readiness commands, and live loopback accounting where available.
- Pin runtime URLs and guardrail env vars to loopback/local values and normalize any legacy
Documents\GitHubrepo aliases toG:\GitHub. - Route execution through existing local service and hardware skills (especially MemryX shim lanes).
- 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:
- Use
$repo-b-control-center-opsfor service startup/restart and endpoint readiness. - Use
$repo-b-hardware-firstfor real-hardware diagnostics and deterministic probe evidence. - Use
$memryx-official-hardware-checkwhen official service checks fail (127.0.0.1:10000path). - Use
$repo-b-thin-waist-routingand$repo-b-mcp-comfy-bridgefor API/MCP routing diagnostics. - Use
$repo-b-local-bridge-orchestratorfor 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
- Do not route required service traffic to non-local hosts unless the user explicitly authorizes it.
- Do not skip required hardware checks for hardware-backed workflows.
- Prefer local scripts, local MCP adapters, local service endpoints, and local hardware probes over cloud-hosted alternatives.
- Treat hostname drift or hardware-check failure as a policy violation and stop before mutation steps.
- If the stack publishes positive displacement preview or healthy TPK, use that as evidence for stronger local-first routing rather than an optional hint.
- Treat
http://127.0.0.1:9000/v1andhttp://127.0.0.1:2337/v1as the authoritative model lanes; treathttp://127.0.0.1:1234/v1only as a non-authoritative operator surface.
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:
- Deep repo-specific debugging that already has a dedicated skill lane.
- Remote cluster/cloud execution plans.
For those lanes, route through $skill-hub and the most specific matching skill.
References
references/local-first-checklist.mdscripts/local_compute_preflight.py
Loopback
If this lane is unresolved, blocked, or ambiguous:
- Capture current evidence and failure context.
- Route back through
$skill-hubfor chain recalculation. - Resume only after the updated chain returns a deterministic next step.
Converted and distributed by TomeVault — claim your Tome and manage your conversions.