nc://error-ledger

Error Ledger: the provider calls that failed, by model, provider, cause and day. A Nodecode rewrite of olric3015/hermes-plugins by olric3015.

Get it

Tell the agent "add the error-ledger cell", or run nodecode add nc://error-ledger.

Set it up

Nothing is required: the cell reads what the log already holds and starts on its defaults. Two settings change the report.

{
  "error-ledger": {
    "window_hours": 24,  // how far back the bare report looks
    "recent_rows": 10    // how many of the newest failures `recent' lists
  }
}

Use it

/error-ledger             last 24 hours, by model
/error-ledger session     the most recent session that had a failed call
/error-ledger all         everything in the log
/error-ledger 7d          the last 7 days (any number of hours or days: 6h, 30d)
/error-ledger providers   add to any of the above to group by provider
/error-ledger reasons     add to any of the above to group by cause
/error-ledger days        add to any of the above to group by day, newest first
/error-ledger recent      add to any of the above to list the newest 10 failures
/error-ledger recent 25   the newest 25 instead (1 to 100)
/error-ledger clear       hide what the log holds so far from the report

/error-ledger all answers a table:

Provider errors, all recorded:

model       errors  moved      top reason  top status
the-model        2      1  rate_limit (1)     429 (1)

One row per failed attempt the log recorded; the retry after it may have succeeded.

• errors counts failed attempts, not failed turns: an attempt Nodecode retried twice before the round moved to another model is two rows.
• moved counts the failures that cost the turn its model — the ones Nodecode moved off.
• top reason is the failure's own classification (rate_limit, overloaded, timeout, ...) when the round moved, or the budget the failure spent (provider_request, provider_stream, provider_contract) when it was retried in place; grouped by reasons, this column names the model it happened on most instead.
• top status is the most common HTTP status; a failure with none (a timeout, an in-stream cut) leaves it empty.

/error-ledger recent says when rather than how often: the newest failures of the chosen window, newest first, so you can tell whether a provider is still failing or stopped an hour ago.

The model can ask the same question through the error_ledger tool (window, by, recent), and eval reaches everything: (error-ledger:report :window :day :by :provider), (error-ledger:failures :window :all) for the rows as data, and (error-ledger:clear).

One honest difference from the original: nothing is recorded twice. The original kept its own file and its own hook; Nodecode's log already holds a turn.provider_retry fact for every failed attempt it will retry and a turn.provider_fallback fact when the round moves to another model, so this cell folds those and writes nothing — clear only hides what is recorded so far, since the log is the organism's own record and this cell does not destroy it.