dsh-analytics

When a DeepSeek Harness-driven agent fires off model calls and tool chains in rapid succession, token spend quickly becomes an opaque black box: you cannot tell which session was the most expensive, which tool keeps re-sending stale context, how much the cache actually saved, or whether the month-end budget will be breached. dsh-analytics is a FinOps plugin built to close that gap. It aggregates scattered usage signals from session events into a local SQLite ledger, prices each row against a time-aware pricing table (peak/off-peak, cache hit vs. miss), and layers on daily/monthly budget limits with end-of-month projection so every model call becomes queryable, attributable, and optimizable.

Once data is at rest, the plugin exposes three consumption paths: agents can invoke the analytics_query tool to pull metrics for any time window; JSON API routes (/api/analytics/overview, sessions, models, tools, pricing, budget) feed external systems; and the most intuitive surface is the zero-build browser dashboard at /analytics, featuring KPI cards, per-turn session waterfalls, reasoning-efficiency comparisons, a parent/child agent cost tree, rule-generated optimization suggestions, and a Context Inflation panel that quantifies wasted tokens and avoidable cost from tool results re-sent in later requests. In-shell integration adds a sidebar Token Analytics entry, a per-session Analytics overlay in the conversation header, and a token·cost·duration line under each completed turn, so you never have to leave the workbench to check the numbers.

Everything stays local: data never leaves the machine and the plugin never calls back to the provider API. It is built for individual developers and engineering teams already running DeepSeek Harness who want to fold agent operating costs into a daily FinOps rhythm—whether that means holding a monthly budget, benchmarking the cost-effectiveness of different reasoning-effort levels, or pinpointing redundant tool-call overhead.