v1.0.0

First release of dsh-global-rules as its own repository.

Added

• A rule list injected into every request's system prompt (section 100: after the deployment persona, before the
  tool notes), so the agent walks the rules before answering — and so compacting or trimming the conversation cannot
  drop them, which is what happens to instructions that travel as chat messages.
• Three ways to edit the same data: Settings → 全局规则 (order 55), the global_rules tool, and the /rules
  command. The settings page lists every rule with per-rule enable/disable, move up/down and delete, offers a
  one-line-per-rule text editor whose line numbers map onto those states, and previews the exact text the model
  receives.
• Two switches: the master switch (rules stay stored, injection stops) and 「回答里显示规则检查」, off by default —
  rules are still checked, but the answer carries no trace of it.
• Storage at $DSH_HOME/dsh-global-rules/rules.json, written by temp file + rename. Reads tolerate a broken entry
  (only that entry is dropped); writes refuse empty or oversized ones (2000 characters per rule, 200 rules).
• A request fence on the plugin's own HTTP route: cross-site fetches, non-loopback Host headers and foreign
  Origins are refused, and an empty body is a 400 rather than a crash.
• test/host-smoke.mjs — 41 checks over the whole Host half (prompt section, global_rules tool, /rules
  command, the route the browser half uses, and the fence) against a temporary DSH_HOME, so a real rule file is
  never touched. CI runs it on six platforms.

Notes

• The plugin guarantees delivery, not compliance: every enabled rule reaches every request, but whether the model
  follows it is the model's decision.
• Reasoning channels follow user rules in the system prompt noticeably less faithfully than answers do, and a
  session's accumulated reasoning language drags the rest along — a new session is the main lever for controlling
  reasoning language. The measurements, and the two stronger channels if that is not enough, are in the README.