Insights

sessions:insights

You are the conductor over the local analytics engines. There is no separate /trends or /perf plugin command — mix lives under agents insights mix. This skill decides which engines to run, stitches their output, and returns a short list of evidence-backed actions.

Never upload raw transcripts. Everything below stays on-machine unless the user explicitly asked for a shareable HTML artifact.

Engines (orchestrate; do not invent a parallel pipeline)

Related (only if the ask needs them; not the default path):

• agents cost — what it cost
• agents output — what shipped
• agents view — live quota

If $ARGUMENTS already names one engine (mix, perf, stats, insights alone), run that engine and still synthesize actions — do not dump a raw table as the whole answer.

Default recipe

Forward user flags (--since, --agent, --project, --json, --narrative, …) to the engines that accept them. Defaults when unset: last 30 days, all harnesses in scope.

Run in parallel when the shell allows:

agents insights --since 30d --json
agents insights mix --days 30 --json    # or: agents insights mix harness-mix --json
agents insights perf --days 30 --json            # plus: agents insights perf friction --days 30 --json when relevant
agents sessions stats --since 30d --json
agents sessions stats --zero --since 30d --json   # dead weight

Adjust windows from $ARGUMENTS (7d, 90d, --all). Prefer --json for the agent pass; render a human table yourself. Use --narrative on agents insights only when the user explicitly wants coaching prose from aggregate counts.

First-run / low coverage:

• agents insights caches facets; --refresh only if numbers look stale.
• agents sessions stats may need agents sessions backfill resources once when coverage is low — run it, then re-query; say so in the recap.

What to return

Lead with actions, not a dashboard dump. Each action needs a quoted signal (a number, a top row, a friction label) from the CLI output — no paraphrase without evidence.

Shape:

1. Window + scope — since, agents, projects actually queried.
2. Top findings (3–7 bullets) — each cites an engine + a concrete figure.
3. Actions (ranked) — what to change: drop a dead skill, fix a hot guard, rebalance harness mix, investigate a slow hook path, reclaim unused plugins.
4. Gaps — harnesses with no stats signal, missing backfill, unattributed buckets.

Skip vanity metrics. "13 minutes" not "12m 49s". If grandmother can't parse it, rewrite.

Guardrails

• Local only by default. No shipping transcript bodies to a model or network path for this skill except the optional agents insights --narrative path the user opted into.
• Explicit invocations only for sessions stats — auto-loaded skills read as 0; say that when interpreting dead weight.
• Do not invent agents sessions insights unless agents sessions insights --help proves it exists on this install; the conductor is this skill, the engines are the four CLIs above.
• Multi-harness: run engines that accept --agent once per harness of interest when the user cares about parity; otherwise one pass with defaults and call out unattributed:* buckets.