Check whether the super-log hub is running and how much traffic it has seen. Call this first when logs seem missing - it distinguishes "the bench is down" from "the app logged nothing".

Put yourself on the bench's AGENTS blotter: say who you are, which LLM you run on, what you are doing, and how often to expect you. For a long job (an 8-hour build, an overnight proof search), call this on a schedule - every interval_s, default 900 (15 minutes) - AND on events: milestones, errors (level WARN/ERROR), and done. The blotter shows your last status and greys you when you miss your own promised cadence, so a silent agent LOOKS silent. This is the single deliberate write this otherwise read-only server performs, and it can only land on agent.* status topics.

List the alarm gateway's routes: the production alarm door and every provisioned webhook endpoint (capture / relay / forward) with its PUBLIC URL, health light, ping latency and kind. Use this to find the URL to hand a webhook sender (Stripe, GitHub, a partner) - deliveries then arrive as wh. events (tail_logs with topic wh.; fields carry the signature verdict, relay status and up to 32KB of body). Ask stream_guide about "wh" for how to read them.

Summarise which log streams (topics) are active and their level mix. Cheap orientation - call this before tail_logs so you know what topics exist and where the errors are, instead of guessing topic names.

Read recent log events, newest last. ALWAYS narrow with topic and/or level - the firehose can be thousands of events per second. Returns a cursor; pass it back as since to read only what is new since your last call.

Find events matching text across the recent window - use when you know what the message says (an exception, an order id, a URL) but not which stream it is in.

Search the on-disk journal: hours or days of history, not the few minutes the hub keeps in memory. This is the tool for "what happened at 3am" or anything older than the live ring - tail_logs and search_logs cannot see that far back. Needs superlog-journal to have been running at the time.

Block until a matching event appears, then return it. This is the one to use after you trigger something: start the action, then wait for its log line or error rather than sleeping and hoping. Returns promptly on the first match.

Detailed documentation for a bench capability, before working with an unfamiliar topic: what its events and metrics mean, how to read them, and the gotchas (what a stall escalation is, why power says not root, why a diff can be silent). No arguments lists everything; name one entry or a playbook for the detail.