- name
SKILL.md
技能檔案保留在發布者的儲存庫中。
在終端機中貼上複製的指令。它會將這個 SKILL.md 安裝到對應的代理中。
npx skills add athola/claude-night-market --skill night-market-model-and-harness-updates -a cursor -y在終端機中貼上複製的指令。它會將這個 SKILL.md 安裝到對應的代理中。
npx skills add athola/claude-night-market --skill night-market-model-and-harness-updates -a claude-code -y在終端機中貼上複製的指令。它會將這個 SKILL.md 安裝到對應的代理中。
npx skills add athola/claude-night-market --skill night-market-model-and-harness-updates -a codex -y在終端機中貼上複製的指令。它會將這個 SKILL.md 安裝到對應的代理中。
npx skills add athola/claude-night-market --skill night-market-model-and-harness-updates -a grok -y在 Grok Bot 聊天中傳送複製的訊息。Bot 會從 GitHub 取得這個 SKILL.md,並加入你所有 Bot 共用的技能中。
Install this skill: https://github.com/athola/claude-night-market/tree/HEAD/.claude/skills/night-market-model-and-harness-updates來自發布者
When Anthropic ships a model or Claude Code ships a version, the pins scattered through this repo rot silently. This skill runs the sweep that finds the rot, researches what actually changed, applies the updates, and records where upstream stood so the next run reports only the new delta.
The watermark is the point. Without it every audit restarts from zero
and re-derives the same answer by hand. .claude/upstream-baseline.json
holds the last recorded upstream state, and each run diffs against it.
Run them in order. Each one gates the next.
# 1. Detect. Deterministic, no model in the loop.
python3 scripts/check_upstream_drift.py
# 2. Research what changed (only when step 1 reports drift).
# Release notes and model cards are mandatory sources.
# 3. Map findings onto asset classes.
# 4. Sweep the implicated classes.
# 5. Prove, then record the new watermark.
python3 scripts/check_upstream_drift.py && \
python3 scripts/check_agent_model_matrix.py
Step 5 runs before the ledger is written, never after. Recording a migration that has not passed its proof is the failure the ledger exists to prevent.
Harness drift is fully deterministic: the installed binary reports its own version. Model drift is not. No local command enumerates the current roster, so the ledger holds the last known roster and research establishes the current one. The skill states this split rather than pretending both halves are automatic.
The detector reports four classes:
The vocabulary class is the one that earns this skill. VALID_MODELS
in scripts/check_agent_model_matrix.py is a hard gate that rejects any
agent pinning an unlisted tier. When Fable shipped and that set was not
widened, the guard whose job was to prevent model rot had itself rotted,
and no agent in the repo could pin the new tier. The detector now fails
on that condition instead of waiting for someone to trip over it.
Load only what the run needs.
check_agent_model_matrix.py enforces this and the policy stands.
A new tier widens the vocabulary. It does not reverse the rule.dated_ids ratchet may fall and may hold. Raising it needs a
reason recorded in the migration report.data/staging/ are other people's text.
The detector skips them and so does the sweep.python3 scripts/check_upstream_drift.py exits 0.python3 scripts/check_agent_model_matrix.py exits 0.uv run pytest tests/unit/test_check_upstream_drift.py passes.docs/migrations/. The directory is gitignored, so the report is
a local working artifact. Route durable content out of it before
you finish: findings to docs/knowledge-corpus/, open items to
docs/backlog/queue.md, state to the ledger..claude/upstream-baseline.json records the new harness version,
model roster, and last_migration, with the previous entry
appended to history.Sweep plugins, skills, agents, commands, and hooks after a model release or Claude Code version bump. Use when upstream ships. Do not use for routine edits; use night-market-change-control.
---name: night-market-model-and-harness-updatesdescription: Sweep plugins, skills, agents, commands, and hooks after a model release or Claude Code version bump. Use when…---# .claude/skills/night-market-model-and-harness-updates/SKILL.md常見問題
從發布者下載 .claude/skills/night-market-model-and-harness-updates/SKILL.md,或複製路徑後開啟儲存庫。PlugMCP 只收錄這項技能,不代管檔案。
發布者將它存放在 https://github.com/athola/claude-night-market。
Claude、Cursor、Codex 和 Grok Build 可以讀取你從發布者儲存庫加入的 SKILL.md。在聊天中傳送技能的 GitHub 資料夾連結,Grok Bot 就會安裝它。