lineup-freshness-sweep

Skill: Lineup Freshness Sweep

The cross-tool-model-lineup-2026.md knowledge bank is Tier-4 (fast-churn): model pickers, pricing, and context-window limits change weekly to monthly. A stale lineup produces confidently wrong recommendations. This skill produces a structured refresh checklist that a researcher (or ravenclaude-core/deep-researcher) can execute against primary sources.

When to run this skill

• The knowledge bank's Last-retrieved header is more than 4 weeks old.
• A developer reports that a model described in the lineup is missing from their picker.
• A developer hits an unexpected charge suggesting a retirement redirect is not documented.
• The researcher-reminder.yml weekly sweep fires.

Step 1 — Identify all dated claims in the knowledge bank

Open ../../knowledge/cross-tool-model-lineup-2026.md and catalogue every claim that carries a [verify-at-use — YYYY-MM] marker or a Last-retrieved: tag. Group by ecosystem:

GitHub Copilot section:
  - Model picker contents (all surfaces × all plans)
  - Plan-gated model list (Free sub-list, Pro, Business, Enterprise)
  - Org model rules feature availability

OpenAI Codex section:
  - Default model id
  - Frontier / Pro model id
  - Reasoning-level options (flag names, levels available)
  - Spark / lightweight tier

xAI Grok section:
  - Active model ids (flagship, fast, lightweight)
  - Retirement / redirect table
  - Any pricing-tier markers

Step 2 — Score each claim by churn risk

Risk level	Signal	Refresh priority
High	Model ids, picker contents, pricing tiers	Every 2-4 weeks
Medium	Plan gates, surface availability, org policy features	Every 4-8 weeks
Low	Vendor-neutral methodology, decision-tree logic	Quarterly

Step 3 — Produce the refresh checklist

For each high-priority claim group, output a research task:

[ ] GitHub Copilot model picker — verify against:
    https://docs.github.com/copilot/about-github-copilot/github-copilot-models  [verify-at-use]
    Check: all surfaces, all plans, any new additions or removals
    Last verified: [date from knowledge bank]

[ ] OpenAI Codex default model — verify against:
    https://platform.openai.com/docs/guides/code  [verify-at-use]
    Check: default model id, reasoning-level flag names, any new tiers
    Last verified: [date from knowledge bank]

[ ] Grok active lineup + retirement table — verify against:
    https://docs.x.ai/docs  [verify-at-use]
    Check: active model ids, any new retirements or redirects, billing consequences
    Last verified: [date from knowledge bank]

Step 4 — Execute or delegate

• If ravenclaude-core/deep-researcher is available: delegate the checklist; specify primary sources from Step 3; require citation + retrieval date on every updated entry.
• If executing directly: visit each primary source, update the knowledge bank entry with the new value, update the Last-retrieved: date, and add a [verify-at-use — YYYY-MM] marker with the current month.
• If a claim cannot be verified: mark it [unverified — training knowledge — reverify YYYY-MM] and flag it prominently in the bank.

Step 5 — Update the bank header

After the sweep, update the bank's top-level Last-retrieved: date and add a one-line sweep note:

Last-retrieved: YYYY-MM-DD
Sweep note: full lineup sweep; [N] claims updated; [M] claims unchanged; retrieval source: [primary URLs]

Pitfalls

• Updating the knowledge bank from memory or training data rather than a live primary-source fetch — this is the failure mode the bank exists to prevent.
• Marking a claim as verified without recording the primary source URL and retrieval date.
• Skipping the retirement table for Grok — silent billing redirects are the highest-blast stale-claim type in this plugin.
• Treating a quarterly sweep as sufficient for model id claims — picker contents can change in days.

See also

• ../../knowledge/cross-tool-model-lineup-2026.md — the knowledge bank this skill sweeps
• ../../CLAUDE.md — §7 on the Tier-4 freshness cadence and researcher-reminder integration