| name | heterogeneous-stack-validation |
| description | Validate a mixed local-plus-remote AI stack with privacy-safe evidence, endpoint health checks, inference smokes, and repo lockstep gates. Use when a request spans multiple local services, remote hosts, and heterogeneous inference lanes that must be tested end to end without launching restricted avatar endpoints. |
Heterogeneous Stack Validation
Use this skill when one workflow must reconcile:
- the active workstation stack
- remote host endpoints
- mixed inference lanes
- policy/docs alignment
- privacy/public-shape gates
Workflow
- Confirm the control authority and startup contract for the current host.
- Probe local health lanes first:
skill-arbiter
- the approved private workstation control repos listed in the active machine roster
- the approved private inference and bridge repos listed in the active machine roster
- the authoritative model lanes on
127.0.0.1:9000 and 127.0.0.1:2337
- Probe remote host runtime contracts and record exact failures.
- Run real inference smokes on every lane that is meant to be live.
- Reconcile docs and open-diff reports with the observed runtime truth.
- Run privacy/public-shape gates before closing the pass.
- Record the outcome back into Skill Arbiter collaboration and skill-game lanes.
- If a remote host package is stale, repair the authoritative repo/config/install sources first and treat GitHub sync plus reinstall as the preferred fix path.
Required Evidence
- local health checks with exact status codes
- remote host health checks with exact status codes
- at least one successful real inference request for each live lane
- explicit contract-failure notes for any broken lane
- repo test/build output for each touched repo
- privacy/public-shape gate results for public repos
Guardrails
- Do not launch local avatar endpoints when the task forbids it.
- Treat already-running remote avatar surfaces as read-only probe targets only.
- Keep host-specific secrets, usernames, and absolute paths out of repo-tracked docs.
- Prefer loopback and LAN-local validation over cloud checks.
- Fail closed on missing health or inference contracts; do not call a lane healthy because a parent app window exists.
- Treat
http://127.0.0.1:1234/v1 and the LM Studio loaded-models panel as non-authoritative operator surfaces when present.
Canonical Lanes
- Local workstation services:
127.0.0.1:17665
127.0.0.1:9000
127.0.0.1:2337
127.0.0.1:11420
127.0.0.1:8800
127.0.0.1:18789
127.0.0.1:8890
127.0.0.1:8895
- Remote hosts:
<remote-host-a>
<remote-host-b>
Output
Produce:
- a concise runtime truth table
- repo validation results
- remaining blockers
- follow-up skill work worth creating, upgrading, or consolidating
References
references/validation-contract.md
Loopback
If the stack cannot be validated cleanly:
- capture the failing endpoint or repo gate with exact evidence
- repair the smallest authoritative layer first
- rerun only the affected validation slice
- record the repaired result in Skill Arbiter collaboration state