@chrisdudek/yg 5.0.0-alpha.6 → 5.0.0

package/README.md

@@ -2,7 +2,7 @@
 
 **Your agent will ignore CLAUDE.md. Yggdrasil makes sure it doesn't.**
 
-Architecture rules your agent can't ignore. You write them in plain Markdown; a reviewer verifies every change and feeds violations back into the agent's loop — before it moves on. Works with Claude Code, Cursor, Copilot, Codex, Cline, and more. The reviewer runs against your code, not your diffs. The feedback is specific, and the agent has to fix before moving on.
+Architecture rules your agent can't ignore. Before it edits a file, your agent gets the few rules that apply — and writes to them. After, every change is checked before it moves on: by a script that runs locally for free, or by an LLM reviewer. A rule written as a script *runs* — your agent can't quietly optimize it away the way it drops a line in CLAUDE.md. Works with Claude Code, Cursor, Copilot, Codex, Cline, and more. Checks run against your code, not your diffs; the feedback is specific; the agent has to fix before it can move on.
 
 See the [main README](https://github.com/krzysztofdudek/Yggdrasil#readme) for documentation, or visit
 [krzysztofdudek.github.io/Yggdrasil](https://krzysztofdudek.github.io/Yggdrasil/).

package/dist/bin.js

@@ -13453,7 +13453,10 @@ var OllamaProvider = class {
       messages: [{ role: "user", content: prompt }],
       stream: false,
       think: false,
-      options: { temperature: this.temperature, num_predict: 500 },
+      // num_predict: -1 → generate until the model stops; no cap, so the verdict
+      // is never truncated (a cut-off JSON would otherwise fail to parse and waste
+      // a re-verification).
+      options: { temperature: this.temperature, num_predict: -1 },
       format: "json"
     };
     try {
@@ -13571,7 +13574,11 @@ var AnthropicProvider = class {
         body: JSON.stringify({
           model: this.model,
           messages: [{ role: "user", content: prompt }],
-          max_tokens: 500,
+          // Never cap the verdict. The Anthropic API requires a max_tokens ceiling
+          // (there is no "unlimited"), so set it far above any reviewer reply — a
+          // verdict is hundreds of tokens; this is many times that — so it is never
+          // truncated. A truncated reply still fails closed when parsed.
+          max_tokens: 8192,
           temperature: this.temperature
         })
       }, "anthropic");

package/package.json

@@ -1,7 +1,7 @@
 {
   "name": "@chrisdudek/yg",
-  "version": "5.0.0-alpha.6",
-  "description": "Architecture rules your coding agent can't ignore. Written in Markdown, verified on every change, enforced in the agent's loop — not after on a PR. Works with Claude Code, Cursor, Copilot, Codex, Cline.",
+  "version": "5.0.0",
+  "description": "Architecture rules your AI coding agent can't ignore. It gets the rules for a file before it edits, and every change is checked — by a free local script or an LLM reviewer — before it moves on. Works with Claude Code, Cursor, Copilot, Codex, Cline.",
   "type": "module",
   "bin": {
     "yg": "dist/bin.js"