claude-evolve 1.11.16 → 1.11.18

package/bin/claude-evolve-check

@@ -141,7 +141,37 @@ SCRIPT
 exec codex exec --dangerously-bypass-approvals-and-sandbox "$1"
 SCRIPT
       ;;
-    qwen)
+    gpt-5-codex)
+      cat > "$test_script" << 'SCRIPT'
+#!/usr/bin/env bash
+exec codex exec -m gpt-5-codex --dangerously-bypass-approvals-and-sandbox "$1"
+SCRIPT
+      ;;
+    gpt-5.2)
+      cat > "$test_script" << 'SCRIPT'
+#!/usr/bin/env bash
+exec codex exec -m gpt-5.2 --dangerously-bypass-approvals-and-sandbox "$1"
+SCRIPT
+      ;;
+    gpt-5.3-codex)
+      cat > "$test_script" << 'SCRIPT'
+#!/usr/bin/env bash
+exec codex exec -m gpt-5.3-codex --dangerously-bypass-approvals-and-sandbox "$1"
+SCRIPT
+      ;;
+    gpt-5.3-codex-spark)
+      cat > "$test_script" << 'SCRIPT'
+#!/usr/bin/env bash
+exec codex exec -m gpt-5.3-codex-spark --dangerously-bypass-approvals-and-sandbox "$1"
+SCRIPT
+      ;;
+    gemini-5-flash)
+      cat > "$test_script" << 'SCRIPT'
+#!/usr/bin/env bash
+exec gemini -y -m gemini-5-flash -p "$1"
+SCRIPT
+      ;;
+    qwen-openrouter)
       cat > "$test_script" << 'SCRIPT'
 #!/usr/bin/env bash
 exec opencode -m openrouter/qwen/qwen3.5-plus-02-15 run "$1"

package/lib/ai-cli.sh

@@ -105,6 +105,30 @@ $prompt"
       ai_output=$(codex exec -m "$codex_gpt5_model" --dangerously-bypass-approvals-and-sandbox "$prompt" 2>&1)
       local ai_exit_code=$?
       ;;
+    gpt-5-codex)
+      local ai_output
+      # GPT-5 Codex - code-specialized variant via Codex CLI
+      ai_output=$(codex exec -m gpt-5-codex --dangerously-bypass-approvals-and-sandbox "$prompt" 2>&1)
+      local ai_exit_code=$?
+      ;;
+    gpt-5.2)
+      local ai_output
+      # GPT-5.2 via Codex CLI
+      ai_output=$(codex exec -m gpt-5.2 --dangerously-bypass-approvals-and-sandbox "$prompt" 2>&1)
+      local ai_exit_code=$?
+      ;;
+    gpt-5.3-codex)
+      local ai_output
+      # GPT-5.3 Codex via Codex CLI
+      ai_output=$(codex exec -m gpt-5.3-codex --dangerously-bypass-approvals-and-sandbox "$prompt" 2>&1)
+      local ai_exit_code=$?
+      ;;
+    gpt-5.3-codex-spark)
+      local ai_output
+      # GPT-5.3 Codex Spark - lightweight fallback via Codex CLI
+      ai_output=$(codex exec -m gpt-5.3-codex-spark --dangerously-bypass-approvals-and-sandbox "$prompt" 2>&1)
+      local ai_exit_code=$?
+      ;;
     o3high)
       local ai_output
       ai_output=$(codex exec -m o3-mini -c model_reasoning_effort="high" --dangerously-bypass-approvals-and-sandbox "$prompt" 2>&1)
@@ -122,6 +146,12 @@ $prompt"
       ai_output=$(gemini -y -m gemini-2.5-flash -p "$prompt" 2>&1)
       local ai_exit_code=$?
       ;;
+    gemini-5-flash)
+      local ai_output
+      # Gemini 5 Flash - cheap fallback model
+      ai_output=$(gemini -y -m gemini-5-flash -p "$prompt" 2>&1)
+      local ai_exit_code=$?
+      ;;
     gemini-3-pro-preview)
       local ai_output
       # Gemini v3 Pro Preview via OpenRouter - EXPENSIVE
@@ -222,7 +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
@@ -258,8 +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/ai_cli.py

@@ -246,7 +246,10 @@ MODEL_TIMEOUTS = {
     'opus-think': 1800, 'sonnet-think': 1800,
     # Gemini - 30 min for pro (streams while working), 20 min for flash
     'gemini-pro': 1800, 'gemini-flash': 1200, 'gemini-3-flash': 600,
-    'gemini-3-pro-preview': 1800,
+    'gemini-3-pro-preview': 1800, 'gemini-5-flash': 600,
+    # Codex/OpenAI models - 10 min standard
+    'gpt-5-codex': 600, 'gpt-5.2': 600, 'gpt-5.3-codex': 600,
+    'gpt-5.3-codex-spark': 300,
     # Z.AI agentic modes - 30 min (can be slow)
     'glm-zai': 1800, 'glm-5-zai': 1800,
     # Codex local - 40 min (local inference can be slow)

package/lib/config.sh

@@ -63,13 +63,13 @@ DEFAULT_WORKER_MAX_CANDIDATES=3
 #
 # Run: Subscription-based agentic models for code generation
 # All CLI tools (opencode, claude, gemini, kimi) are agentic and can edit files
-DEFAULT_LLM_RUN="sonnet glm-5-zai kimi-coder gemini-pro"
-DEFAULT_LLM_RUN_FALLBACK="codex-oss-local"
+DEFAULT_LLM_RUN="gemini-pro gemini-pro glm-5-zai glm-5-zai kimi-coder kimi-coder gpt-5-codex gpt-5-codex sonnet"
+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 sonnet-think glm-5-zai gemini-pro kimi-coder"
-DEFAULT_LLM_IDEATE_FALLBACK="sonnet glm-5-zai"
+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
 _load_yaml_config() {
@@ -318,7 +318,7 @@ show_config() {
   echo "  Worker max candidates: $WORKER_MAX_CANDIDATES"
   echo "  LLM configuration:"
   # Show LLM configurations using dynamic variable names
-  for model in gpt5high o3high codex gemini opus opus_think sonnet sonnet_think cursor_sonnet cursor_opus glm deepseek; do
+  for model in gpt5high o3high gpt_5_codex gpt_5_2 gpt_5_3_codex gpt_5_3_codex_spark codex gemini gemini_5_flash opus opus_think sonnet sonnet_think cursor_sonnet cursor_opus glm deepseek; do
     var_name="LLM_CLI_${model}"
     var_value=$(eval echo "\$$var_name")
     if [[ -n "$var_value" ]]; then

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": "1.11.16",
+  "version": "1.11.18",
   "bin": {
     "claude-evolve": "bin/claude-evolve",
     "claude-evolve-main": "bin/claude-evolve-main",

package/templates/config.yaml

@@ -90,30 +90,41 @@ parallel:
 llm_cli:
   # What to run for each sub-command
   # Models are tried in order, with round-robin distribution across candidates
-  # You can repeat models for weighted selection (e.g., "sonnet sonnet gemini" for 2:1 ratio)
+  # You can repeat models for weighted selection (e.g., "gemini-pro gemini-pro sonnet" for 2:1 ratio)
 
-  # Default configuration: 50/50 split between glm-zai and kimi-coder, commercial ideation
+  # Default configuration: sonnet at ~11%, rest doubled for cost savings
   # Commented out because these change over time; uncomment to override
-  #run: glm-zai kimi-coder glm-zai kimi-coder glm-zai kimi-coder codex-oss-local
-  #ideate: opus-openrouter kimi-k2-openrouter gemini-pro sonnet-think gpt5high grok-4-openrouter deepseek-openrouter glm-zai
+  #run: gemini-pro gemini-pro glm-5-zai glm-5-zai kimi-coder kimi-coder gpt-5-codex gpt-5-codex sonnet
+  #ideate: opus-think glm-5-zai gemini-pro kimi-coder gpt-5.2 gpt-5.3-codex
+  #run_fallback: haiku glm-5-zai gemini-5-flash gpt-5.3-codex-spark
+  #ideate_fallback: haiku glm-5-zai gemini-5-flash gpt-5.3-codex-spark
 
   # Available models:
-  # - sonnet: Claude 3.5 Sonnet via Claude CLI
-  # - sonnet-think: Claude 3.5 Sonnet with extended thinking (ultrathink prompt)
-  # - opus: Claude 3 Opus via Claude CLI
-  # - opus-think: Claude 3 Opus with extended thinking (ultrathink prompt)
-  # - gemini: Gemini via Gemini CLI
-  # - gpt5: GPT-5 via Codex CLI (standard)
+  # Claude (subscription-based, watch usage limits):
+  # - sonnet: Claude Sonnet via Claude CLI
+  # - sonnet-think: Claude Sonnet with extended thinking (ultrathink prompt)
+  # - opus: Claude Opus via Claude CLI
+  # - opus-think: Claude Opus with extended thinking (ultrathink prompt)
+  # - haiku: Claude Haiku via Claude CLI (cheap fallback)
+  #
+  # Codex/OpenAI (subscription-based):
+  # - gpt-5-codex: GPT-5 Codex (code-specialized) via Codex CLI
+  # - gpt-5.2: GPT-5.2 via Codex CLI
+  # - gpt-5.3-codex: GPT-5.3 Codex via Codex CLI
+  # - gpt-5.3-codex-spark: GPT-5.3 Codex Spark (lightweight fallback) via Codex CLI
+  # - gpt5: GPT-5 via Codex CLI (legacy alias)
   # - gpt5high: GPT-5 via Codex CLI (high reasoning)
   # - o3high: O3 via Codex CLI (high reasoning)
-  # - cursor-sonnet: Claude 3.5 Sonnet via Cursor Agent CLI
-  # - cursor-opus: Claude 3 Opus via Cursor Agent CLI
-  # - glm: GLM-4.6 via OpenCode CLI
-  # - grok-code-fast: Grok Code Fast 1 via OpenRouter
-  # - grok-4: Grok 4 via OpenRouter
-  # - opus-openrouter: Claude Opus 4.1 via OpenRouter
-  # - kimi-k2-openrouter: Kimi K2 Thinking via OpenRouter (RECOMMENDED - no separate auth)
-  # - kimi-k2-think-moonshot: Kimi K2 Thinking via Moonshot CLI (requires separate kimi CLI setup)
-  # - kimi-coder: Kimi for Coding model via kimi CLI (fast, good for code generation)
-  # - codex-qwen3: Qwen3-Coder via Codex + Ollama (local, free, RECOMMENDED)
-  # - aider-qwen3: Qwen3-Coder via Aider + Ollama (local, free, experimental)
+  #
+  # Gemini (free tier available):
+  # - gemini-pro: Gemini 3 Pro Preview via Gemini CLI
+  # - gemini-5-flash: Gemini 5 Flash via Gemini CLI (cheap fallback)
+  # - gemini-flash: Gemini 2.5 Flash via Gemini CLI (legacy)
+  #
+  # Other free/cheap models:
+  # - glm-5-zai: GLM-5 via Z.AI agentic mode
+  # - kimi-coder: Kimi for Coding via kimi CLI (fast, good for code gen)
+  # - kimi-k2-openrouter: Kimi K2 Thinking via OpenRouter
+  # - codex-oss-local: Local model via Codex + Ollama
+  # - cursor-sonnet: Claude Sonnet via Cursor Agent CLI
+  # - cursor-opus: Claude Opus via Cursor Agent CLI