nc://custodian

Custodian: watches the log files you name for known failure patterns, keeps an issue ledger, and gives the fix for each. A Nodecode rewrite of indigokarasu/hermes-custodian-plugin by indigokarasu.

Failures repeat. A job that was disabled after three failures, a rate limit that hit because four jobs started on the same minute, an expired token, a full disk: each one has a log line that names it and a fix that worked. Custodian is that knowledge as a table of 35 fingerprints, so the model reads the log once and knows what it is looking at, instead of starting from the message in front of it.

Get it

Tell the agent "add the custodian cell", or run nodecode add nc://custodian.

Set it up

Name what to watch. In the cell's sources setting (/settings, or "custodian" in ~/.nodecode/config.jsonc) list the log files to read, one path each — a directory means every .log and .jsonl in it:

{
  "custodian": {
    "sources": ["~/.nodecode/logs/gateway.log", "~/.local/state/my-service/"],
    "max_age_hours": 24,
    "tail_lines": 500
  }
}

No key and no account: a scan reads the files you named and writes only under ~/.nodecode/custodian/. A cell with an empty sources refuses to start and says so, rather than running mute.

Use it

The model gets three tools:

The operator types /custodian status, /custodian scan, /custodian issues, /custodian issues resolve ID.

Three things keep it quiet enough to leave on. A finding carries the line that matched, not the top of the file, so it is something to read. A line that merely quotes a pattern — an earlier report about a finding — is not a finding, so a repaired issue does not report itself forever. And a match older than max_age_hours is dropped, with a timestamp-less continuation line judged by the last dated line above it, so a three-day-old traceback does not read as live.

A finding is evidence, not a diagnosis: read the line, decide whether the fix applies, apply it, resolve the issue, scan again. Nothing repairs anything by itself.