Skills Digest — the monthly "what skills do we have, and what changed" email

Turns .claude/skills/ into a skimmable monthly digest: a 📋 What changed since last digest band up top (✨ added · ✏️ updated · 🗑️ removed, computed from git history), then every skill grouped by job (Build / Plan / Commit gates / Review / Deploy / Verify / Meta) with a trigger badge (auto / on ask / in flow) and its one-line description. The point is awareness: catch sprawl, spot a skill that quietly rotted, notice overlap — once a month, without digging.

Single source of truth — no second list to drift. The skill set, the grouping, and the trigger map all come from scripts/gen-skills-doc.mjs (the same module that generates writeups/architecture/skills-and-triggers.html). gather.mjs imports its META / GROUPS / discover() exports rather than re-listing skills, so the digest and the doc can never disagree. A skill missing from META lands under "⚠️ Uncategorised" — same contract as the doc.

Interactively this is render-only. A by-hand run makes the digest (HTML + plain text + subject under writeups/reports/) and shows you where it landed. The scheduled job is what delivers — by email via Resend (mirrors epic-digest; email-only, no standing-issue rail, to avoid the duplicate GitHub-notification mail).

Pipeline (mirrors epic-digest / housekeeping)

gather.mjs ──> skills.data.json ──> render.mjs ──> skills-digest-<date>.{html,txt,subject.txt}
   │                                    │
   reuses gen-skills-doc.mjs            email-safe HTML (inline styles, table layout) + text mirror
   (META + discover) + git delta

By hand

# 1) gather — enumerate skills + compute the since-last-digest delta from git
DIGEST_SINCE=2026-06-01 node .claude/skills/skills-digest/gather.mjs > skills.data.json

# 2) render — email-safe HTML + plain text + subject
node .claude/skills/skills-digest/render.mjs skills.data.json --out-dir writeups/reports

render.mjs also runs straight off the committed example.data.json (offline, no git needed):

node .claude/skills/skills-digest/render.mjs .claude/skills/skills-digest/example.data.json --out-dir /tmp/sd

The "since" window

DIGEST_SINCE (YYYY-MM-DD) is the previous digest's send date — the delta window starts there. Unset → defaults to one month ago. The scheduled workflow passes the exact previous send date.

The delta is computed by comparing the skill set at the since-date commit against HEAD:

Bucket	Meaning
✨ Added	exists at HEAD, absent at the since baseline
✏️ Updated	exists in both, but a file under its skill dir changed in the window
🗑️ Removed	existed at the baseline, gone at HEAD

⚠️ Full history required. The delta needs the since-date commit to be reachable. CI must check out with fetch-depth: 0 — a shallow clone (the default) can't reach a month-ago commit, and gather.mjs will fall back to firstRun: true (no delta) rather than guess.

🔀 The flow view (#843)

The digest also carries a trigger→skill flow — the skill cards answer "how does each skill fire" (auto / on ask / in flow); the flow answers "when — which event sets off which skills, in order" (on commit · on a PR / CI · on a schedule · when a bug is reported · at epic close · …).

It's driven by the exported FLOWS map in scripts/gen-skills-doc.mjs — the same source the skills-and-triggers.html doc page renders, so the email and the page can never disagree. gather.mjs doesn't touch it; render.mjs imports FLOWS and draws an email-safe lane table. Adding a lane or moving a skill between lanes is a one-edit change in that map, and gen-skills-doc.mjs hard-fails (CI --check) if a lane references a skill that no longer exists — the diagram can't silently drift from the real skill set.

Chose inline HTML/SVG over a committed PNG/Excalidraw on purpose: our surfaces are the email and the doc page, where inline markup renders more reliably (incl. mobile), stays crisp and selectable, and regenerates from FLOWS every time. A committed binary would drift with no CI to catch it. (Excalidraw-PNG — the ticket-diagram skill — remains the right tool for per-epic GitHub-mobile status diagrams, a different surface.)

🧮 Budget annotations (loop-station, #1028 A)

When a committed loop-station context-budget record exists (writeups/loop-station/history.jsonl, epic #926), gather.mjs joins each skill's measured token cost and redundancy band onto the digest, and render.mjs surfaces:

• a header line — total harness context budget (tokens, of which in skills), corpus redundancy %, and the overall scorecard band;
• a per-skill token badge — grey when in-band, amber/red when the scorecard flags it oversized;
• an ↔ overlap badge on skills that appear in loop-station's top restated pairs, plus a "most-overlapping skills" callout (e.g. a11y ↔ frontend-review 44.6%) — the merge/trim candidates.

This is the digest ↔ observatory wiring: loop-station is the producer (it computes tokens/redundancy); the digest only reads the latest record and renders — it never recomputes. Everything degrades to a no-op when the record is absent (no header line, no badges; the digest is unchanged). loop-station's own fetch-depth/commit-back concerns are its business; the digest just reads the committed JSONL.

🪦 Dead weight — freshness × usage (#1100 WS5)

The "🪦 Dead weight" band joins two signals to surface retire candidates — a knowledge skill that is both stale and unused:

• Freshness from scripts/skill-freshness.mjs (computeSkillFreshness()): which provenance- stamped skills have a vanished citation or a verified: stamp aged past 90d.
• Usage from the committed loop-station usage rollup (writeups/loop-station/usage.jsonl, #1064): per-skill CI invocation counts, gated by usageCoverage() (advisor.mjs).

A skill that is stale (stamp overdue or a vanished citation) and has 0 CI invocations over the covered window becomes a retire candidate. It never judges "unused" blind: with no rollup (or thin coverage) usageCoverage() returns judged: false, so the band renders freshness-only and states the verdict is withheld. Counts are pipeline scope — interactive session usage isn't measured yet (#1067), so a 0 means "never fired in CI", not "provably dead".

Like the budget band, the digest is a reader/renderer (freshness + usage are the producers) and degrades to a no-op — a green "✅ all N knowledge skills fresh" line — when nothing is stale.

Cadence + delivery (config-as-data)

Declared in .github/digests.yml, exactly like the other digests:

skills-digest:
  schedule: monthly:1        # first of the month, UTC (monthly:<dom> added in #839)
  deliver: [email]           # email-only (Resend); no-op when RESEND_API_KEY unset
  to: [you@example.com]

.claude/skills/housekeeping/schedule.mjs gates a daily workflow cron — it sends only on the configured day-of-month (clamped to the month's last day, so monthly:31 still fires in February). Delivered by .github/workflows/skills-digest.yml.

When you add / rename / remove a skill

The digest follows automatically (it reads the live tree). Two housekeeping steps keep it tidy:

1. Add the skill to the META map in scripts/gen-skills-doc.mjs (group + triggers) — otherwise it shows under "⚠️ Uncategorised" in both the doc and this digest.
2. Run node scripts/gen-skills-doc.mjs so skills-and-triggers.html stays current (CI enforces this via skills-doc.yml; sync-docs does it before /commit).