Skill: Hunt a Prod Error

An opaque error has two halves. Local logs say where (route, status, timing, burst shape). Sentry says what (the exception and stack). You need both before naming a cause; either alone produces a guess.

1. Local logs: where and how often

Desktop server log (OTLP JSON, one record per request):

~/Library/Application Support/com.differentai.openwork/logs/openwork-server.log{,.1}

Engine log: ~/.local/share/opencode/log/opencode.log.

Tally 5xx by route, then compare against the same route family that did not fail. A failure rate on one route and zero on its siblings isolates the code path; a burst of failures in the same millisecond points at a shared event (GC, rollover, socket), not per-request data.

rg '"status":5' ~/Library/Application\ Support/com.differentai.openwork/logs/openwork-server.log* \
  | python3 -c 'import sys,json,re,collections; c=collections.Counter()
for l in sys.stdin:
  a=json.loads(l)["attributes"]; c[(a["method"], re.sub(r"(ses|ws|msg)_[A-Za-z0-9]+", r"\1_*", a["path"]), a.get("error.code"))]+=1
[print(v,k) for k,v in c.most_common()]'

Know the blind spot: the OpenCode proxy catch blocks in apps/server/src/server.ts forward the real exception to Sentry only; the local record carries the synthetic Unexpected server error and no cause.

2. Sentry: what threw

Sentry is connected through OpenWork Connect (search_capabilities → search_issues, get_sentry_resource). Org different-ai-inc, project desktop-app, region https://us.sentry.io.

Filter by the tags the server sets: surface:server and the templated route (for example route:"/workspace/:id/opencode/*"), plus release. Read the top issue's stack, firstSeen, users, and event count. The exact exception message is the string a regression spec must reproduce.

3. Name the regression