dsh-fuse

In multi-agent pipelines every LLM call silently burns budget, yet most teams lack a local guard that can reject a request before it ever hits the wire. dsh-fuse fills that gap: at the agent pre-step stage it measures tokens and cost for each call, and the moment the cumulative spend in a window exceeds the configured limit the local fuse trips an instant rejection with zero network dependency. Price resolution draws from both a pricing registry and the gateway, and models that no source can price are flagged as unpriced rather than silently zeroed, so the fuse always fires when it should.

The plugin also bridges the control plane and the data plane. On boot it pulls org-level budgets and policies from the panel, pushing edits to every machine in real time; every 60 seconds it ships a batch of usage events and fuse-cut records, with session IDs SHA-256 hashed on the client so raw content never leaves the machine. On the panel side, graduated 50 / 80 / 95% alerts and anomaly spend detection can be routed to Telegram or a webhook. For the agent itself, the plugin registers a read-only dsh_budget_status tool that lets the model inspect current quota, spend, and headroom mid-session, so it can adapt by choosing a cheaper route or stopping gracefully instead of retrying blindly.

It is aimed at teams running multiple agents, organisations that need unified inference budget control in CI or headless containers, and developers who want to bolt a financial circuit breaker onto their DeepSeek Harness stack without touching business code.