Skip to main content

v0.12.0: the spend you can act on

avuru obs team
Maintainers

v0.12.0 turns AI spend from a number on a screen into something that reaches you — and opens the estate, read-only and on your switch, to an agent that can investigate it.

  • A threshold on spend. Nobody watches a screen. Monthly budgets in tokens or in money, for one calling service or across the estate, now fire through the alerting channels an install already has. A money budget on an install that declared no prices is refused at startup, naming what is missing — rather than sitting quietly under every threshold forever — and a partly priced scope says its figure is a floor.

  • One rate table, written down once. Model prices and compute rates used to be declared twice, in two formats, with two currency fields — one of them needing a pod restart. They now resolve through a single table, editable in Settings → Rates and applied without a redeploy. Chart-declared values keep working and are shown read-only; UI entries overlay them, each row saying which it is. The budget evaluator and the screens read the same resolver, so a budget cannot be measured against a price other than the one on display.

  • Tool calls are no longer counted as model calls. The AI module tested that gen_ai.operation.name was present and never read its value, so on an agent workload every tool execution counted as a call to a model: call counts inflated, latency ranked a database lookup against a completion, the model resolved to nothing, and that empty row counted as a model of its own. Four numbers were wrong together; they are right now, and tools get a table of their own — calls, failures, p95 and the services invoking each one, with no token or cost columns, because a tool execution spends neither.

  • An agent turn, drawn as the shape it is. A turn is a graph, not a list: model calls and tool executions as distinct node kinds, weighted by time spent inside each call rather than span duration. Offered only on traces that actually hold a turn.

  • An agent can read this estate. A new MCP module serves a Model Context Protocol server at POST /mcp — six read-only tools over the traces, logs, error issues and health you already store, authenticated with personal API tokens that resolve their owner's live permissions. It adds no collection, no schema change and no container. It is off by default, because what an agent reads leaves your cluster for the model provider you chose, and every tool call is logged with the token owner, the tool, its arguments and the row count.

  • Both collector images below their advisories. The gateway distro pins x/crypto, x/text and grpc above the versions its collector line resolves, and the node agent — which needs receivers the minimal distro does not carry — moved to the collector line that scans clearest. This is a deployment concern, not just a report: a registry with a block-on-critical policy stops serving a flagged image, and that surfaces as a rollout timing out on the first node it touches.

Upgrade with the Helm chart as usual; nothing in this release changes collection, and the MCP module stays off unless you enable it.