Regex search across a namespace fan-out over pods, merged chronologically
read-only · every pod's tail is fetched with timestamps on and merged oldest → newest
Per-pod match counts “where is this error coming from?” before reading full logs
counts run over the last max-lines-per-pod lines of each pod's default container
Log sources pods · containers · restarts · age — the discovery call before searching
restart counts are summed over the pod's containers — a crash-looping sidecar is triage-relevant
What the four MCP tools are good for all read-only · this page calls the exact same code
This console is a human front-end over the SAME seams, namespace policy, and caps that back the
MCP tools at /mcp — it has no powers the tools don't. Prometheus MCP (the fleet's other
observability server) sees behavior over time: rates, trends, alert state. LogSearch sees
what the applications actually printed: tracebacks, panics, OOM kills, crash loops.
Triage workflow
Sources tab (or list_log_sources): which pods exist, which containers to name, who is restarting
Counts tab (or count_matches): ERROR|Traceback — which pod is loud?
Search tab (or search_logs): the same pattern with provenance, chronologically merged across pods
Agent follow-up: get_pod_logs with previous=true reads the PREVIOUS (crashed) container — the first move on CrashLoopBackOff
LLM-safe shaping why tool answers stay small
Fan-out touches at most LOGSEARCH_MAX_PODS pods, tails at most
LOGSEARCH_MAX_LINES_PER_POD lines each, and returns at most
LOGSEARCH_MAX_TOTAL_LINES merged matches (keeping the MOST RECENT ones, flagged
truncated). One dead pod never sinks a search — it is reported in
errors while the rest complete. This page reuses the exact same caps.