DE//AGENT
Explore

Docs · Project

Maintenance

The weekly cadence that keeps the registry fresh: refresh, rebuild, audit, review oldest first, retire without nostalgia.

The registry stays trustworthy through routine, not heroics. This is the routine.

Weekly cadence

  1. Discover — node pipelines/discover.mjs pulls trusted sources into data/drafts/. Review failures per source; one dead feed never blocks the rest.
  2. Refresh provenance — node pipelines/refresh-provenance.mjs records upstream repo state. Unreachable upstreams are data, not errors.
  3. Rebuild — node pipelines/run.mjs --check then npm run build. Read the step lines: evidence gaps, thin content, drift, and eligibility all report here.
  4. Audit — node pipelines/audit-pages.mjs (registry vs built pages) and node pipelines/audit-links.mjs (links + sitemap coverage). Both exit 1 on drift; fix before deploying.
  5. Review oldest first — the queue shows draft ages. Anything waiting 90+ days gets a decision: promote, re-scope, or delete the draft.

Promotion

Drafts ship as community unless evidence is attached. The gate is node pipelines/promote.mjs --draft <file> [--apply]: unique id/slug, known category, declared license, 200+ character description, and evidence matching any claimed level above community. See Contributing.

Retirement runbook

Retire without nostalgia; never delete history:

  1. Set retired: true on the record (and superseded_by in relationships when a successor exists).
  2. Rebuild and confirm the route is gone: audit-pages shows it missing from dist, audit-links shows no dangling references to it.
  3. Merges go the same way: duplicates.json names the candidates, the curator picks the survivor, moves any unique content, retires the rest.

Retired records stay in curated YAML and the knowledge graph — they leave the sitemap and the routes, not the record.

Backlog discipline

The queue is an editorial backlog, not a publishing queue. Cap promotions per week to what review hours support. If the backlog grows three weeks running, throttle discovery (--limit) before quality slips — narrowing intake beats widening standards.

Deferred to V2 (after first GitHub publish + traction)

Recorded here so deferral is a decision, not drift:

  • Analytics + Search Console. Hooks ship (__DE_OUTBOUND__, opt-in provider config); nothing collects. Enable after the first publish, then let the data pick expansion clusters.
  • Affiliate links. Policy and disclosure ship; zero links live. Review program terms (eligibility, geography, attribution, thresholds) only once content is listed and getting traction.
  • First-party org (de-agent/*). 111 original records point at a planned github.com/de-agent org (8 repos) that does not exist yet; freshness records them unreachable honestly. Decision: create the org and publish the content, then re-run provenance refresh to clear the flags.
  • backend/ dead code and astro check exclusion. Scoped as future work; the exclusion comment in ci.yml tracks the type-debt reason.