Fleet logs
Read the runtime log of every selected instance in one pass. Read-only. Parse
[selector] [--tail <n>] [--since <dur>] [--grep <re>] [--errors-only] from "$ARGUMENTS";
defaults --tail 200, no --since, empty selector means every managed instance. The selector grammar and the
instance model live in ${CLAUDE_PLUGIN_ROOT}/skills/fleet-model/SKILL.md.
1. Resolve the targets
python3 "${CLAUDE_PLUGIN_ROOT}/scripts/fleet.py" resolve "<selector>" --json \
| python3 -c 'import json,sys
for i in json.load(sys.stdin)["instances"]:
c, s = i.get("container") or {}, i.get("signals") or {}
print(i["name"], i["state"], c.get("id") or "-", i["project"], s.get("log_age_hours"))'
all is allowed here — this command only reads. An alien row still gets its log read; it just gets no
interpretation, because nothing here knows what it is.
2. Read each instance, redact, then filter
docker logs --timestamps --tail <n> [--since <dur>] <container-id> 2>&1 \
| python3 "${CLAUDE_PLUGIN_ROOT}/scripts/lib/redact.py" \
| grep -Ei '<pattern>'
The order is the point: redact before filtering, never after. A grep in front of the scrubber sends
raw matched lines to a sink that was never cleaned. [scrubbed: N matches] on stderr is the mechanism
working, not a warning; the redactor deliberately over-matches high-entropy strings.
--errors-only→grep -Ei 'error|fatal|panic|exception|unauthor|refused|timeout|ECONN'.--grep <re>→ that expression,-E. Both given:--grepnarrows inside--errors-only.- No container (
down, absent) →docker compose -p <project> logs --no-color --tail <n>through the same redactor. Still nothing → say the instance has no runtime log rather than inventing one.
3. Print
One block per instance, in the resolve order, each opened by a header naming the instance, its state and its log age so an empty block reads as evidence instead of an error:
==> <name> state=<state> log-age=<h>h <==
Hard rules
- Never
cat,tailorgrepa log file on the host, and never read a credential,.env,auth-secretsor auth-profile file to explain a log line. The container's own log stream through the redactor is the whole surface. - Never paste a raw log line into a summary. Quote the redacted line as printed.
- Keep it bounded: past
--tail 2000or a fleet-wide--since 7d, narrow the selector or the window instead — a wall of scrollback buries the one line that mattered. - A crash loop is read here first and restarted never: a restart destroys the lines holding the cause and buys a longer backoff.
Next
A log line that names a symptom but not an id → skill fleet-diagnostics. A green instance whose log has
not moved → /openclaw-ops:status --deep for the HEALTH/LIVENESS split. One CLI call against the same
instance → /openclaw-ops:exec.