dsh-session-budget-guard

A single prompt in DeepSeek Harness can fan out into three levels of subagents, thousands of model calls, and drain an API balance in twenty minutes. A cumulative budget alone is too slow to matter because the tree is burning several dollars a minute and may never reach a high threshold before the balance is gone. dsh-session-budget-guard puts two independent gates on every session tree: a cumulative-spend cap and a sliding-window burn-rate cap. Whichever trips first acts first, without waiting on the other.

When a limit is crossed, behaviour depends on the configured mode. In notify mode the plugin logs the event and fires a desktop notification, which is ideal for observing real traffic before committing to enforcement. In enforce mode it returns reject at the agent/pre-step hook so the next model request never fires, denies over-quota delegations, workflow calls, and web searches at tools/pre-execute with a model-visible reason so the agent adapts rather than retries blindly, interrupts descendants deepest-first, and cancels the current turn while preserving queued inbox messages. Structural gates such as maximum delegation depth, maximum tree nodes, and per-session search limits keep the tree from growing without bound. Cost data is read from the dsh-cost-meter ledger when available, including user-edited prices; otherwise the guard prices calls itself from DeepSeek peak and off-peak list rates. Lineage is resolved from the first record of each session log, and no message content is ever read.

If you run multi-agent research, batch content generation, or automated workflows in DeepSeek Harness and want an insurance fuse against runaway API spend, this plugin is that fuse. Start in notify mode for a day, read the log, tune the thresholds to your real traffic, then flip on enforce with confidence.