dsh-cost-governor

The built-in harness token meter tells you how many tokens a request used, but it cannot answer the more practical questions: how much does this actually cost in dollars, and what happens when you blow past your budget? dsh-cost-governor exists for exactly that gap. It reuses the harness native disjoint token buckets (input, output, cache-read, cache-write, reasoning), converts them into currency via a configurable price catalog, and then layers a full budget-governance workflow on top: soft warn, hard ceiling, block or steer to a cheaper model.

Under the hood the plugin ships a multi-provider price catalog covering DeepSeek, OpenAI, Anthropic, Google, xAI, Qwen, GLM and Moonshot, with separate cache and reasoning rates that are fully overridable. Budgets can be set per day, week, month or unlimited; a soft warnRatio (default 80%) triggers threshold events and webhook notifications (Slack-style, Feishu, DingTalk), while a hardRatio ceiling can notify-only, block new requests, or steer traffic to a cheaper model. All accounting rides on the official sessionProjections seam, so numbers stay correct through compaction and cold reads, and editing the price table re-prices history instantly without re-folding the log. The Settings dashboard adds KPI cards, an animated budget bar, a per-model spend table, a daily spend chart, a savings tip, and CSV export; a compact HUD ring lives in the sidebar rail.

It is aimed at teams and individual developers who burn a meaningful number of tokens across several providers and want an actionable budget keeper rather than a passive counter: someone who needs to see month-end cost, get warned at 80%, and have the system automatically pull the reins when the ceiling is hit.