zames_pro 2.82.105 → 2.82.106

package/CHANGELOG.md

@@ -7,6 +7,12 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ## [Unreleased]
 
+## [2.82.106] - 2026-10-24
+
+### Fixed
+
+- **goal:** a tool call mentioning VERDICT: MET does not clear the goal
+
 ## [2.82.105] - 2026-10-24
 
 ### Fixed

package/dist/goal.js

@@ -76,12 +76,44 @@ export function buildContinuationDirective(reason) {
 function argumentReason(reason) {
     return String(reason ?? '').trim().length > 0;
 }
+/**
+ * The text a verdict may be read from. A plain answer is returned as-is; a
+ * TOOL CALL (JSON or DSML/XML) yields only its `respond` message, because the
+ * call's ARGUMENTS are data, not a judgement — a Bash/Write argument that
+ * happens to contain `VERDICT: MET` must not clear the goal. Any other call
+ * yields '' (fail closed: the caller treats it as "continue").
+ */
+function toolCallVerdictText(raw) {
+    const hasCallMarkers = /"tool"\s*:/.test(raw) ||
+        /<[^>]*\b(?:invoke|parameter)\b/i.test(raw) ||
+        /DSML/i.test(raw);
+    if (!hasCallMarkers)
+        return raw;
+    // `respond` carries the operator-facing text; unwrap its message only.
+    const respond = /"tool"\s*:\s*"respond"[\s\S]*?"message"\s*:\s*"([\s\S]*?)"\s*\}\s*\}/i.exec(raw);
+    if (respond)
+        return respond[1];
+    const dsml = /invoke[^>]*name\s*=\s*"respond"[^>]*>([\s\S]*?)(?:<\/[^>]*invoke|$)/i.exec(raw);
+    if (dsml) {
+        const msg = /parameter[^>]*name\s*=\s*"message"[^>]*>([\s\S]*?)</i.exec(dsml[1]);
+        if (msg)
+            return msg[1];
+    }
+    return '';
+}
 // Parse the evaluator's reply into a verdict. Tolerant of markdown emphasis
 // and casing (```VErdict: MET``). Returns null when the verdict line is
 // missing, so the caller can fail closed (treat an unparsable answer as
 // "continue", never as "done").
 export function parseGoalVerdict(text) {
-    const t = String(text ?? '');
+    const rawText = String(text ?? '');
+    // A TOOL CALL is not a verdict. The evaluator runs as a normal agent turn,
+    // so in reasoning mode it sometimes answers with a tool-call JSON/DSML whose
+    // ARGUMENTS merely MENTION `VERDICT: MET` (a Bash grep, a Write of a summary,
+    // a `respond` echo). Reading that as the verdict CLEARS the goal — an
+    // unfinished task was silently accepted as done. A call is recognized by its
+    // own markers, and the verdict is taken from its `respond` message only.
+    const t = toolCallVerdictText(rawText);
     // The label may be bolded/italicised and the colon may carry surrounding
     // spaces (`**VERDICT:** MET`) — a markdown-model answer did exactly that and
     // the strict `VERDICT:` matcher returned null, so the goal never parsed and

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "zames_pro",
-  "version": "2.82.105",
+  "version": "2.82.106",
   "description": "Terminal coding agent that drives a real browser to chat with DeepSeek, Qwen and other web LLMs — no API key required. Reads and edits files, runs commands, commits to git, like Claude Code / Codex CLI.",
   "type": "module",
   "bin": {