nano
forked niri forked from okami.mom/niri
Star 1
Fork 0
Watch

fix(metrics): stop storing full prompt/response bodies in metrics.db #66

Merged callie.on-her.computer
targeting main from deleted fork:fix/metrics-no-full-text
Merged opened by and assigned to callie.on-her.computer
targeting main from deleted fork:fix/metrics-no-full-text

the problem #

metrics.db was multiple GB of body text. every turn stored the entire messages[] array twice - once as a standalone prompt metric, again inside prompt_response (which also stored the full response object). ~400KB/row, kept for a /metrics/:id detail view the dashboard never renders.

the fix #

the list view only ever derives messageCount, lastUserMessage, responsePreview, toolCallCount, model, usage. so:

  • drop the redundant standalone prompt metric (fully subsumed by prompt_response)
  • compute the summary fields at write time via summarizePromptResponse() and persist only those - never raw messages/response
  • remove the dead prompt bucket / list-type / cursor plumbing

each row drops from ~400KB to a few hundred bytes. verified the stored payload contains no messages[] and no full response body; dashboard list fields unchanged; typecheck clean.

note #

separately, the retention prune (pruneOldMetrics, 3-day default) deletes rows but never VACUUMs, so freed pages linger as freelist - that's what let the file sit at 8.2GB even after pruning. worth a periodic VACUUM in the prune path (not in this PR).

Commits 0