Splunk Search

Use Splunk Search when a user needs read-only Splunk evidence but the result set may be too large for the agent context. The splsearch CLI writes search results to a local SQLite table and returns compact JSON with the table name. Treat the saved table as the artifact, then bring only focused summaries, small samples, or exact evidence back into the conversation.

Prerequisites

The splsearch CLI is installed and available on PATH, or the runtime provides an equivalent splsearch binary for this skill. In this repository, the packaged Go source for that CLI lives in tools/splsearch. From the repository root, build the tool artifact with make -C tools/splsearch build, then put tools/splsearch/bin on PATH before using the skill locally.

The user has read-only access to the target Splunk Enterprise or Splunk Cloud Platform environment and can provide the target URL, time window, and search intent. Do not ask for passwords, cookies, tokens, or raw credential files.

When to Use

Use this skill for Splunk log investigation, production or staging incident triage, saved result-table analysis, and requests that mention splsearch, Splunk search, SPL execution, or querying saved Splunk results.

Do not use this skill for changing Splunk configuration, editing knowledge objects, deleting data, restarting services, creating alerts, modifying indexes, or handling secrets. If the user wants SPL authoring help without running a search or inspecting saved results, answer without invoking this skill.

Workflow Overview

1. Identify the target Splunk URL, time window, impact scope, and strongest available filters.
2. Check cached auth with splsearch auth status --output=json when the target is unclear.
3. Validate a specific target with splsearch auth status --url=<splunk-url> --output=json before deciding whether browser login is needed.
4. Run one bounded table-writing search with a useful time range and result table name.
5. Keep the returned table name and use it for every follow-up command.
6. Review any saved warnings before continuing. If a broad fetch was intentional, record that by accepting the warning.
7. Inspect schema or metadata before writing SQL against the saved table.
8. Use text search, summary, ordered events, or bounded SQL to reduce the saved results locally.
9. Bring only the compact answer set, evidence snippets, and command summary back into chat.
10. Keep the result table while the incident or handoff is active, then drop it when evidence has been captured elsewhere.

Do not shrink the Splunk search just because the agent context is small. Fetch a relevant, bounded slice into SQLite, then reduce locally. Avoid | head 10, tiny limits, or immediate output unless the task truly needs only a tiny sample.

Run table-writing searches serially. Parallel searches can contend for the local result database if they complete at the same time.

Commands

• splsearch auth status --output=json lists cached auth inventory.
• splsearch auth status --url=<splunk-url> --output=json validates one target without forcing a fresh login.
• splsearch search --url=<splunk-url> --query='<SPL>' --earliest=-15m --result-table=<table> runs a bounded Splunk search and saves results locally.
• splsearch search --url=<splunk-url> --query='<SPL>' --earliest=-15m --result-table=<table> --progress=jsonl writes parseable progress events to stderr while keeping stdout as the final JSON result.
• splsearch results-list --limit=20 rediscovers saved tables and follow-up commands.
• splsearch result-info --table=<table> inspects metadata, original SPL, time bounds, row counts, and warnings.
• splsearch result-warnings accept --table=<table> --code=<warning-code> marks a reviewed local warning as intentional.
• splsearch result-schema --table=<table> inspects saved columns before SQL.
• splsearch result-text-search --table=<table> --query='<text>' --limit=20 performs local BM25 text recall over saved rows.
• splsearch result-summary --table=<table> --group-by=<field> --limit=20 builds first-pass aggregates.
• splsearch result-events --table=<table> --field=<field> --value=<value> returns compact ordered event sequences.
• splsearch result-search --table=<table> --query='<SQL SELECT>' --limit=20 runs bounded SQL against the saved table as results.
• splsearch results-drop --table=<table> removes a local saved table after the evidence is no longer needed.

Examples

Investigate recent API errors without dumping raw rows:

splsearch search --url=<splunk-url> --query='index=app_logs component=api severity=ERROR' --earliest=-30m --result-table=app_errors
splsearch result-info --table=app_errors
splsearch result-schema --table=app_errors
splsearch result-summary --table=app_errors --group-by=component --limit=20
splsearch result-events --table=app_errors --request-id=abc-123 --limit=100
splsearch result-search --table=app_errors --query='SELECT component, count(*) AS errors FROM results GROUP BY component ORDER BY errors DESC LIMIT 20' --limit=20

Find a text symptom inside a known saved table:

splsearch result-text-search --table=app_errors --query='timeout retry exhausted' --limit=20

Clean up after the handoff:

splsearch results-drop --table=app_errors

Troubleshooting

If target validation fails with a transport error, inspect the structured error_code, operation, retryable, diagnostic_hint, and message fields. Do not treat DNS, network, or proxy failures as proof that credentials need refreshing.

If browser auth fails, keep the compact stdout JSON in context and inspect error_code, operation, retryable, retryable_after_environment_change, remediation_code, diagnostic_hint, launch_error_summary, message, requested_channel, attempted_channel, and fallback_used. Do not paste diagnostic files or browser logs into chat unless the user explicitly asks.

If a long-running search appears stuck, use progress output rather than guessing. --progress=jsonl keeps stdout machine-readable and puts events such as dispatch, running state, fetch rows, write rows, ETA, and scan counts on stderr. If progress is active but scan counts are huge, narrow the SPL or time window instead of blindly extending --timeout.

If search returns ok:false or table_created:false, stop and report that no saved table exists. Do not continue with invented table names.

If a saved table has active warnings, review them before using the result. Broad or unlimited searches can be valid during an incident, but they must be intentional and named in the evidence.

If a saved table disappears, explain that local result tables are ephemeral and rerun the bounded Splunk search if the evidence is still needed.

Safety

Do not print cookies, auth files, tokens, or ~/.config/splsearch/auth.json. Do not read ~/.config/splsearch/results.sqlite directly unless the CLI cannot answer the question. Do not paste broad raw result sets into chat.