dsh-conversation-stats

You run dozens of conversations in DSH every day, yet the dialogue list only shows titles and timestamps. Digging through zstd-compressed logs to find out which session burned the most tokens, how well the cache performed, or which step took the longest is tedious and error-prone. dsh-conversation-stats puts all of that on a single screen: a new Conversation Stats tab appears in the DSH Web GUI, aggregating every historical session key metrics with no extra service or resident process. Data survives restarts as-is.\n\nThe experience centers on an overview table and a per-session detail view. The overview lists each session in reverse chronological order with turns, steps, model-call count, input/output/cache tokens, and cache-hit rate, while summary chips at the top give you an at-a-glance picture of total spend. Clicking any row expands the full breakdown: per-call model details (timestamp, model, finish reason, token split), tool-call statistics (count, total duration, failures), model distribution, and LLM vs. tool latency. Data auto-refreshes every 60 seconds or on demand. The table also includes a Delete column (with a red confirmation dialog) for permanently removing a session directory, and a Load column that restores an archived session back into the DSH dialogue list.\n\nUnder the hood, the plugin reads session.jsonl.zstd files directly from ~/.dsh/sessions, decompresses each zstd frame natively, parses the newline-delimited JSON event stream, and collapses statistics using the same logic as the built-in sessionStats projection. No API server, no daemon. It is well suited for heavy multi-turn users who track token cost and cache efficiency, or for teams that accumulate large backlogs of sessions and need regular cleanup and audit trails.