Investigating logs

Investigation is a narrowing problem: summarize before you read. One posthog:logs-patterns call compresses millions of lines into at most 200 templates, and one posthog:logs-patterns-diff call answers "what is different about now vs. before" directly. Raw rows (posthog:query-logs) are the last step of an investigation, never the first.

When to use this skill

• "Check the logs" / "is service X healthy?" / "did my deploy (or model bump, config change, migration) break anything?"
• "Why are errors up?" / "explain this spike" / incident triage — "what changed?"
• "What is this service logging?" — orienting in an unfamiliar or noisy stream.
• Finding the log evidence for a failure reported elsewhere (an alert, an error-tracking issue, a user complaint).

When not to use this skill

• Creating or tuning log alerts — that's authoring-log-alerts.
• Analytics over product events, persons, or insights — that's querying-posthog-data.
• HogQL exposes a logs table via posthog:execute-sql, but do not investigate through it: hand-written SQL over logs routinely hits read-byte caps and re-derives what the tools below do in one cheap call. Reserve SQL for the rare case of joining log-derived facts with non-log data.

Tools

Tool	Job
posthog:logs-services-create	Top-25 services with log_count, error_count, error_rate, sparkline. Orientation.
posthog:logs-patterns	Mine one window's message templates, ordered by frequency. "What is this stream saying?"
posthog:logs-patterns-diff	Diff templates between two windows: new / rate-shifted / gone. "What changed?"
posthog:logs-count / posthog:logs-count-ranges	Scalar and time-bucketed counts for a filter. Localize volume before pulling rows.
posthog:logs-sparkline-query	Volume over time broken down by severity or service (the one bucketed view with a breakdown).
posthog:logs-facet-values-create	Distribution of severity/service (or a resource attribute) under a filter.
posthog:logs-attributes-list / posthog:logs-attribute-values-list	Discover attribute keys and values before building filters.
posthog:query-logs	Raw rows. Endpoint of every drill-down, entry point of none.

Each tool's own description documents its parameters and response shape — read it before calling.

Pick the workflow by question shape

"Is it healthy?" — post-deploy / post-change verification

The user changed something (deploy, model bump, config, migration) and wants to know the logs still look right.

1. Pin down the change time and the affected service(s). Ask if the user hasn't said; the diff is meaningless without a boundary.
2. Orient with posthog:logs-services-create: is the service still logging at all, and what is its error_rate now? A service that went silent fails verification just as hard as one that started erroring.
3. posthog:logs-patterns-diff with query.dateRange from the change time to now and baselineDateRange set to a comparable window just before the change, scoped to serviceNames. New error/fatal templates right after a change are the classic regression signature; large rate_ratio shifts on existing error templates are the second thing to check.
4. Check volume continuity with posthog:logs-count-ranges spanning before and after the boundary: a rate discontinuity (crash loop, restart storm, silence) shows up here even when message content looks unchanged.
5. Drill only the suspects: pivot each suspicious pattern to raw lines via its match_regex with posthog:query-logs.

A pass verdict needs all three: no new error templates, no large error rate_ratio shifts, and continuous volume. Say which windows you compared — "healthy" is only as strong as the baseline.

"Explain this spike"

1. Localize it: posthog:logs-count-ranges over the user's window, then recurse into the dense bucket(s) — each bucket's date_from/date_to feeds the next call. Stop after 3–4 levels.
2. Explain it: posthog:logs-patterns-diff with the spike as query.dateRange and the window just before as baselineDateRange. The top new and rate_shift entries are the explanation. Do not mine both windows separately and diff by hand — the diff is one call.

Incident triage — "what broke?"

posthog:logs-patterns-diff first: incident window vs. a known-good window just before (or omit the baseline for same-window-last-week). Suspects are new entries and the biggest rate_ratio shifts; pivot each to raw lines. If the failing service is unknown, find it first with posthog:logs-facet-values-create faceting service_name under severityLevels: ["error", "fatal"].

"What is this stream saying?" — unfamiliar service

posthog:logs-patterns over the last hour, scoped to the service. Scan templates by estimated_count and non-zero error share in severity_counts. Widen the window or add searchTerm only if the answer isn't there.

Known needle — a specific message, attribute, or person

When the target is already precise (an error string, a request id, a distinct_id), skip pattern mining: discover the right keys with posthog:logs-attributes-list / posthog:logs-attribute-values-list, size the result with posthog:logs-count, then pull rows with posthog:query-logs.

Rules that keep investigations honest and cheap

• Scope serviceNames (or a resource-attribute filter) on every call once the target service is known. Unscoped calls scan the whole team's stream and starve the pattern sample budget.
• posthog:query-logs requires an explicit query.dateRange — omitting it is a 400, not a default window.
• Pattern counts are sampled estimates (sampled: true); templates rarer than ~1 in 10,000 rows can be invisible. Absence of a rare template is not evidence it stopped.
• Before trusting a wall of new entries in a diff, check baseline.total_count — a tiny or empty baseline (logging only just started) makes everything look new.
• severityLevels matches the six canonical lowercase buckets against severity_text exactly. Zero rows on a severity filter → check the stored values with posthog:logs-attribute-values-list { key: "severity_text" }.
• Budget: one services call, at most one patterns-diff per window pair, 3–4 count-ranges levels, and query-logs only for confirmed suspects with limit ≤ 100.

Output

Lead with the verdict, then the evidence:

• Verdict: healthy / regressed / inconclusive, with the windows compared.
• Suspects (if any): template, classification (new / rate_shift), estimated counts or rate_ratio, services, and 1–2 sample raw lines.
• What was checked and what wasn't: services covered, windows, and any sampling or baseline caveats that limit confidence.

The user should be able to act on the verdict without re-running the investigation.