claude-evolve 1.11.17 → 1.11.18

package/bin/claude-evolve-check

@@ -171,7 +171,7 @@ SCRIPT
 exec gemini -y -m gemini-5-flash -p "$1"
 SCRIPT
       ;;
-    qwen)
+    qwen-openrouter)
       cat > "$test_script" << 'SCRIPT'
 #!/usr/bin/env bash
 exec opencode -m openrouter/qwen/qwen3.5-plus-02-15 run "$1"

package/lib/ai-cli.sh

@@ -252,7 +252,7 @@ $prompt"
       ai_output=$(opencode -m openrouter/moonshotai/kimi-k2.5 run "$prompt" 2>&1)
       local ai_exit_code=$?
       ;;
-    qwen)
+    qwen-openrouter)
       local ai_output
       # Qwen latest - Alibaba's flagship model (currently qwen3.5-plus)
       # Linear attention + sparse MoE, strong multimodal capabilities
@@ -288,8 +288,8 @@ $prompt"
     echo "[AI] Raw output from $model_name:" >&2
     echo "----------------------------------------" >&2
     if [[ ${#ai_output} -gt 2000 ]]; then
-      echo "$ai_output" | head -50 >&2
-      echo "... (truncated from ${#ai_output} characters to first 50 lines) ..." >&2
+      echo "... (truncated from ${#ai_output} characters to last 50 lines) ..." >&2
+      echo "$ai_output" | tail -50 >&2
     else
       echo "$ai_output" >&2
     fi

package/lib/config.sh

@@ -68,7 +68,7 @@ DEFAULT_LLM_RUN_FALLBACK="haiku glm-5-zai gemini-5-flash gpt-5.3-codex-spark"
 #
 # Ideate: Agentic models that can edit files for ideation
 # All CLI tools (opencode, claude, gemini, kimi) are agentic and can edit files
-DEFAULT_LLM_IDEATE="opus-think glm-5-zai gemini-pro kimi-coder gpt-5.2 gpt-5.3-codex"
+DEFAULT_LLM_IDEATE="opus-think glm-5-zai gemini-pro kimi-coder gpt-5.2 gpt-5.3-codex qwen-openrouter"
 DEFAULT_LLM_IDEATE_FALLBACK="haiku glm-5-zai gemini-5-flash gpt-5.3-codex-spark"
 
 # Load configuration from a YAML file and update variables

package/lib/evolve_run.py

@@ -71,7 +71,11 @@ class WorkerPool:
         try:
             # Don't capture output - let it stream directly to terminal
             # This provides real-time visibility into which models are being used
-            proc = subprocess.Popen(cmd)
+            # AIDEV-NOTE: Explicitly pass stdin=DEVNULL so workers don't inherit
+            # a closed/bad stdin FD from parent (e.g. when run via nohup or after
+            # terminal disconnect). Without this, Python workers crash at startup
+            # with "OSError: [Errno 9] Bad file descriptor" on sys stream init.
+            proc = subprocess.Popen(cmd, stdin=subprocess.DEVNULL)
             self.workers[proc.pid] = proc
             log(f"Spawned worker {proc.pid}")
             return proc.pid

package/lib/evolve_worker.py

@@ -205,9 +205,15 @@ CRITICAL: If you do not know how to implement what was asked for, or if the requ
                     preview = output[-300:] if output else "(empty)"
                     log(f"Bandit model {selected_model} completed but didn't modify file ({len(output)} chars), trying fallback...")
                     log(f"AI output preview: {preview}")
+                    # AIDEV-NOTE: Report no-modification as failure to bandit
+                    self.bandit.update(selected_model, child_score=None, parent_score=self._parent_score)
+                    log(f"Bandit update: {selected_model} no file modification")
 
             except AIError as e:
                 log(f"Bandit model {selected_model} failed: {e}, trying fallback...")
+                # AIDEV-NOTE: Report AI-level failure to bandit so it learns to avoid broken models
+                self.bandit.update(selected_model, child_score=None, parent_score=self._parent_score)
+                log(f"Bandit update: {selected_model} AI call failed")
 
         # Fallback to round-based retry with all models
         try:
@@ -229,7 +235,7 @@ CRITICAL: If you do not know how to implement what was asked for, or if the requ
                 return True, model
             else:
                 # AIDEV-NOTE: Log output so we can diagnose why file wasn't modified
-                preview = output[:300] if output else "(empty)"
+                preview = output[-300:] if output else "(empty)"
                 log(f"AI completed but did not modify file ({len(output)} chars)")
                 log(f"AI output preview: {preview}")
                 return False, model

package/lib/llm_bandit.py

@@ -86,6 +86,12 @@ class LLMBandit:
         # Baseline score for normalizing improvements
         self._baseline_score: float = 0.0
 
+        # AIDEV-NOTE: Decay counter - only apply decay every N updates to prevent
+        # aggressive memory loss. With decay_factor=0.95 applied every update,
+        # n_completed floors at 1 after ~50 updates and the bandit can't learn.
+        self._updates_since_decay: int = 0
+        self._decay_interval: int = 50  # Apply decay every 50 updates
+
         # Load existing state if available
         if state_file and Path(state_file).exists():
             self.load()
@@ -215,13 +221,23 @@ class LLMBandit:
         return improvement
 
     def _apply_decay(self) -> None:
-        """Apply decay to reduce influence of old observations."""
+        """Apply decay to reduce influence of old observations.
+
+        AIDEV-NOTE: Only applies every _decay_interval updates to prevent
+        aggressive memory loss. The int() truncation on n_completed was
+        destroying the bandit's ability to learn from failures.
+        """
+        self._updates_since_decay += 1
+        if self._updates_since_decay < self._decay_interval:
+            return
+
+        self._updates_since_decay = 0
         for stats in self.models.values():
-            # Decay both counts and totals proportionally
+            # Decay totals to reduce influence of old observations
             stats.total_improvement *= self.decay_factor
-            # Don't decay counts below a small floor to preserve some memory
-            if stats.n_completed > 1:
-                stats.n_completed = max(1, int(stats.n_completed * self.decay_factor))
+            # Decay counts but preserve enough memory to differentiate models
+            if stats.n_completed > 2:
+                stats.n_completed = max(2, int(stats.n_completed * self.decay_factor))
 
     def save(self) -> None:
         """Persist state to file."""

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "claude-evolve",
-  "version": "1.11.17",
+  "version": "1.11.18",
   "bin": {
     "claude-evolve": "bin/claude-evolve",
     "claude-evolve-main": "bin/claude-evolve-main",