What
load_all_traces() currently loads ALL traces into memory on every invocation. --limit N on agentdiff list only truncates display after a full load. Push --limit and --since filters into the storage layer so only needed traces are fetched.
Why
At 100k+ traces (enterprise scale, 1+ year of data), agentdiff list and agentdiff stats will take seconds and consume hundreds of MB. Every subcommand hits this path.
How
With per-UUID file storage:
--limit N: list only the most recent N files from the trace tree (sort by UUID timestamp prefix)--since <date>: read only shards after that date prefix- Change
load_all_traces() signature to accept optional limit/since params and short-circuit early
Effort
~45 min. Depends on per-UUID file storage refactor (if not yet done, can be approximated with early-exit on sorted JSONL).
What
load_all_traces()currently loads ALL traces into memory on every invocation.--limit Nonagentdiff listonly truncates display after a full load. Push--limitand--sincefilters into the storage layer so only needed traces are fetched.Why
At 100k+ traces (enterprise scale, 1+ year of data),
agentdiff listandagentdiff statswill take seconds and consume hundreds of MB. Every subcommand hits this path.How
With per-UUID file storage:
--limit N: list only the most recent N files from the trace tree (sort by UUID timestamp prefix)--since <date>: read only shards after that date prefixload_all_traces()signature to accept optional limit/since params and short-circuit earlyEffort
~45 min. Depends on per-UUID file storage refactor (if not yet done, can be approximated with early-exit on sorted JSONL).