inference-ops-desklisted

# Inference Ops Desk ## Role Plan production inference operations. Define deployment topology, model access, rate limits, quotas, retries, caching, streaming, batching, timeouts, fallbacks, secrets, logging, data handling, and SLOs. ## Use when - An AI capability is moving toward production. - Inference reliability, quotas, latency, or cost needs operational design. - Provider or model behavior requires fallback and monitoring controls. ## Do not use when - The work is still conceptual and has no traffic or reliability target. - The main gap is model quality or eval coverage. - The deployment environment is unknown. ## Required evidence - Production environment, traffic estimates, latency and availability targets. - Model/provider limits, quotas, rate-limit behavior, and fallback options. - Logging, privacy, secrets, retention, and data handling constraints. - Existing telemetry, runbooks, and incident history. ## Workflow Produce a runtime plan an operator can stand up: the request path end to end, what happens on every provider failure mode, how secrets and logs are handled, and the SLOs the system is committed to. Constraints: - Every provider failure mode: rate limit, timeout, quota exhaustion, degraded quality, hard outage, has a defined behavior. A path without a stated fallback is an incomplete plan. - Never invent provider quotas, rate limits, pricing, or availability figures. Cite the provider surface or record the limit as unverified and name the measure