Create the release-notes authoring agent skill

Context

Leaf of epic #14483 — the operator-directed centerpiece (2026-07-02): "we must create a new agent skill this time. it requires heavy memory mining. it must meet the quality bar of previous release notes." No release-notes skill exists; the nearest siblings (blog-post, guide-authoring, update-roadmap) cover adjacent surfaces but not this protocol. v13.0's process reached the bar through an unformalized 7-leaf iteration cluster (#12695→#12924) and was lost.

Sweep: epic-reshape under live operator direction; no competing skill ticket (skills tree checked this session; latest-open sweep 14:42Z).

The Problem

Every release re-derives the notes process from scratch, and single-pass derivation measurably fails: the v13.1 iteration-1 draft carried a War-Story environment error (LOCAL incident narrated under cloud framing), one-lens narration, and release-downplaying scope framing — precisely the misses a codified protocol with per-claim V-B-A prevents.

The Fix

Author .agents/skills/release-notes/ under create-skill + turn-memory-pre-flight discipline (Progressive Disclosure: thin SKILL.md router + workflow payload). The payload codifies:

1. Scope derivation — analyzeClosedSinceRelease.mjs (cutoff = previous release commit) + release project-board views + epic closures + the explicit caveat that the ticket record LAGS shipped reality (v13.1: est. 300–500 unfiled changes); no single tracker view is the window's truth.
2. Heavy Memory-Core mining per narrative arc — query_summaries/query_raw_memories sweeps over the window per candidate hero chapter; sessions and summaries ground the story, ticket titles do not.
3. Per-claim V-B-A — every factual claim carries a source anchor; environment claims (local vs cloud), scope claims (what a release "is"), and stability claims verified against primary sources before assertion.
4. The quality-bar contract distilled from the v13.0.0 notes (resources/content/release-notes/chunk-2/v13.0.0.md): narrative hero chapters, the "in 2 minutes" section, TL;DR discipline, honest bounds, upgrade-path framing, War-Story arc (Symptom → Investigation → Culprit → Fix) — and the operator directives as named rules: do NOT downplay a release, multiple iterations ARE needed (draft merges early with an explicit iteration marker; mining-driven refinement leaves follow until cut-readiness).
5. Publish-flow mechanics — flat-root staging lifecycle (publish.mjs:83-226: required pre-flight → atomic-hash append → GitHub Release parse (frontmatter/H1 contract) → self-removal post-release → chunk-N re-materialization), the orphan-guard staging window (#14484-refined), and the human-only cut gate.

Acceptance Criteria

• .agents/skills/release-notes/SKILL.md (thin router, trigger description) + references/ workflow payload exist and pass the create-skill shape bar
• The five protocol areas above are each covered with concrete tool commands / file anchors, not prose gestures
• Substrate-mutation slot-rationale included in the PR body per turn-memory-pre-flight (new skill-loaded substrate: disposition + trigger-frequency rationale)
• The skill's quality-bar section cites the v13.0.0 notes as the measurable precedent and encodes the iteration cadence
• PR Resolves this leaf, Refs #14483

Out of Scope

• Authoring the v13.1 iterations themselves (sibling leaves under #14483).
• Modifying blog-post / guide-authoring / update-roadmap (cross-reference only).

Related

Related: #14483 (parent epic), #12695 (v13.0 lineage), #10321 (release-cut process sibling)

Origin Session ID: 8cf234b7-e698-47ca-99e2-bf865196b6aa

Retrieval Hint: release-notes agent skill protocol memory mining quality bar iteration cadence publish flow