--- title: Sessions and traces description: Every conversation is saved with every step; reopen one later, and read the trace to see exactly what the agent did and how long it took. --- # Sessions and traces ## Sessions Each conversation is a **session**, saved as it goes — every message, tool call, result, error and denial. - `/sessions` lists them; `/resume ` reopens one and continues where you left off. In the web UI they are the first section of the left panel. - `/new` starts a fresh one in the same project. - Resuming a session from a previous run tells the agent what changed since (for example, servers it had started are no longer running). - `/delete ` removes one. ## Traces Next to each session there is a **trace**: a plain-text log, one line per step, with elapsed time and tokens — the fastest way to see where the agent did well or badly. ``` 12:03:41 step 3 read src/api/auth.ts 41 lines 0.1s 12:03:44 step 4 edit src/components/LoginForm.tsx +6 −1 2.8s 12:03:44 ask bash npm test → allowed 12:03:52 step 5 bash npm test exit 0 7.9s 12:03:55 done 5 steps · in 11,802 · out 640 · 38s ``` `/trace` in the terminal (or `/trace 2` for the one before), `≡` next to a session in the web UI, or `tail -f ~/lampson/.lampson/ws//.lampson/trace/.log`. ## Tokens and budget `/tokens` shows what this session has used. Each turn has a budget (400k tokens of cost by default); at 80% the agent is nudged to wrap up, at 95% it must answer without more tools. Prompt caching is on for providers that support it, so long sessions stay cheap. ## Long outputs A tool result over 10k characters is saved whole to a file and the model sees the beginning, the end and the path — it can read the rest in pages. Old results are folded into one-line receipts as the conversation grows, and the context is compacted when it gets too long. You do not have to manage any of this; it is why a session can run for hours.