Extensions review

After an agent loads a skill, incident.io assesses the load once the run is scored: did the agent follow the skill, did following it help, and a short written account of what happened. This skill turns a window of those assessments into a story people read — the loads where an extension let an agent do something it could not have done alone, and the incidents where no extension helped — so the team sees what their content earns and what to write next.

The model that shapes everything here: an assessment is a claim about a run; the run is the evidence. The written account routinely understates or inverts what the run shows, so nothing is published from the assessment alone — every published item is read at its source first. And assessment lags the run it judges, so a recurring review windows on when loads were assessed, not when they happened: that way each load is seen exactly once, and none are missed for being assessed late.

The two legs

• Usage — pull the window's assessed loads, shortlist the few that clear the bar, verify each against the run it happened in, and write them as standalone items. → references/usage.md
• Gaps — the same window's incidents, checked for coverage: which ones no skill load touched, whether an extension could have advanced them, and a brief for each gap worth filling. → references/gaps.md

Run both for a full review. A request for one ("what got used", "what are we missing") runs that leg alone: do not read the other leg's reference, run its preflight, or name its capabilities in the report — a usage-only review that fails its preflight names the three census surfaces and nothing more. Both legs end in one report — references/report.md owns its shape and how it is delivered, including "post to a channel" requests.

Ground rules

• Read before you publish. A load may only appear in the report after its source — the investigation or the conversation — has been read and supports it. Where the source is unreadable from this session, the item is either dropped or written as the assessment's claim, marked as such — never asserted as fact.
• Platform rows are not extensions. A load row with no plugin_id belongs to the platform's own agent machinery. The per-plugin pull never sees them; drop them whenever the global stream is read instead. The managed incident.io plugin's rows (which carry a plugin_id) count.
• Show a miss when the window has one. A review that only shows wins reads as marketing and stops being read. The best miss names the line of content that caused it.
• A quiet window is a finding. When nothing clears the bar, say so in one line with the load count — never inflate a routine load into a highlight.
• This skill reads and reports. It edits no content and dismisses no feedback. Gap briefs are proposals for the skills that act on them.

What this skill is not for

• Estate health — sync failures, funnels, feedback issues worth acting on, convention drift — is the doctor skill. Doctor aggregates and diagnoses; this skill narrates a window. Where this review notices a health problem (a plugin that stopped syncing, a skill whose loads keep hurting), note it in one line and route it to doctor rather than absorbing the diagnosis.
• Editing a skill from what a review found is the skill-authoring skill's improve job; writing a missing runbook is the runbooks skill's write job. Gap briefs hand over there.
• A skill's cumulative track record — rates, open issues, examples — is extension_skill_feedback_list and the dashboard's extensions pages, not a window review.