Criticality tiered reporting
Skill selamy-labs/agent-skills/skills/criticality-tiered-reporting
Use when deciding what a service or worker should report and how loudly — alerts, progress updates, logs, notifications. Tier output by criticality so the important signal is not buried in routine chatter.From its SKILL.md
npx -y skills add selamy-labs/agent-skills --skill criticality-tiered-reportingAssembled from the repository path, not quoted from the project. Check it against their README if it does not work.
One thing to look at
- 0 stars0 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.
SKILL.md
3.3 KB, 688 tokens by cl100k_base, as published. Nobody here has run it
Criticality Tiered Reporting
A reporter that emits at one volume is wrong at every volume. Page on everything and the operator mutes you; report nothing and a real failure passes unnoticed. The fix is to tier the signal by how much the consumer needs to act on it, then route each tier to a channel matched to that urgency.
The tiers
Classify every reportable event into one of three tiers and route accordingly:
- Down or blocking — the work has stopped, is stuck, or needs a human decision to proceed. This is the only tier that may interrupt a person: page, alert, or message a channel someone watches. It must be actionable — what broke, what is blocked, and the first step to unblock.
- Routine progress — the work is proceeding normally: started, milestone reached, finished. Summarize, do not interrupt. A periodic digest, a status surface someone can pull, or a single completion message — not a ping per step.
- Noise — heartbeats, retried-and-recovered transients, per-iteration detail. Stay silent on the active channels; write it to logs or metrics for later inspection, not to anyone's attention.
Route to the consumer's need, not your chattiness
The tier is defined by what the consumer must do, not by how much is happening internally. A thousand successful iterations are still tier "routine" (one summary). One stuck queue is tier "blocking" (one page). Map each tier to a channel whose interruption level matches: interrupting channel for blocking, pull/digest surface for progress, log/metric sink for noise.
Design rules
- Escalate, don't duplicate: a blocking event pages once; it does not also spray progress updates. Lower tiers never borrow the higher tier's channel.
- Actionable at the top: every interrupting message names the problem and a next step. "Something went wrong" is not a tier-one message.
- Recovered transients are noise: a retry that succeeded is not a page; count it as a metric. Only a transient that exhausts retries escalates.
- Make routine pullable: progress should be queryable on demand so the consumer checks when they want, instead of being pushed each step.
- Silence is a valid output: when nothing crosses a tier threshold, emit nothing to attention channels. No-news is good news, by design.
Anti-patterns
- alerting on every event so real failures drown in green noise (alert fatigue)
- emitting a notification per step of a long routine job
- a "failure" page that contains no cause and no next action
- paging on a transient that already recovered on retry
- silence by omission: a worker that simply never reports, so a real stall is invisible until someone notices the absence of output
Done means
Events are classified into down/blocking, routine progress, and noise; each tier routes to a channel matched to its urgency; interrupting messages are actionable; routine is summarized or made pullable; and noise stays out of attention channels while remaining inspectable in logs or metrics.
What ships with it
Read from the repository
Just SKILL.md. No reference files, no scripts.