claude-evolve 1.12.0 → 1.14.0

package/lib/ai-cli.sh

@@ -53,7 +53,7 @@ EOF
 call_ai_model_configured() {
   local model_name="$1"
   local prompt="$2"
-  local codex_gpt_model="${CODEX_GPT_MODEL:-${CODEX_GPT5_MODEL:-gpt-5.2}}"
+  local codex_gpt_model="${CODEX_GPT_MODEL:-${CODEX_GPT5_MODEL:-gpt-5.5}}"
   
   # Record start time
   local start_time=$(date +%s)
@@ -96,12 +96,12 @@ $prompt"
       ;;
     opus-openrouter)
       local ai_output
-      ai_output=$(opencode -m openrouter/anthropic/claude-opus-4.1 run "$prompt" 2>&1)
+      ai_output=$(opencode -m openrouter/anthropic/claude-opus-4.7 run "$prompt" 2>&1)
       local ai_exit_code=$?
       ;;
     cursor-sonnet)
       local ai_output
-      ai_output=$(cursor-agent sonnet-4.5 -p "$prompt" 2>&1)
+      ai_output=$(cursor-agent sonnet-4.6 -p "$prompt" 2>&1)
       local ai_exit_code=$?
       ;;
     cursor-opus)
@@ -123,7 +123,7 @@ $prompt"
     codex-think)
       local ai_output
       # High reasoning - for ideation tasks requiring deep thinking
-      ai_output=$(codex exec -m gpt-5.4 -c model_reasoning_effort="high" --dangerously-bypass-approvals-and-sandbox "$prompt" 2>&1)
+      ai_output=$(codex exec -m gpt-5.5 -c model_reasoning_effort="high" --dangerously-bypass-approvals-and-sandbox "$prompt" 2>&1)
       local ai_exit_code=$?
       ;;
     codex-coding)
@@ -135,7 +135,7 @@ $prompt"
     codex-spark)
       local ai_output
       # Cheap/fast lightweight fallback
-      ai_output=$(codex exec -m gpt-5.1-codex-mini --dangerously-bypass-approvals-and-sandbox "$prompt" 2>&1)
+      ai_output=$(codex exec -m gpt-5.4-mini --dangerously-bypass-approvals-and-sandbox "$prompt" 2>&1)
       local ai_exit_code=$?
       ;;
     # --- Gemini (subscription) ---
@@ -258,7 +258,7 @@ $prompt"
       ;;
     ollama-qwen)
       local ai_output
-      ai_output=$(codex exec --dangerously-bypass-approvals-and-sandbox --skip-git-repo-check --oss --local-provider=ollama -m qwen3.5:cloud "$prompt" 2>&1)
+      ai_output=$(codex exec --dangerously-bypass-approvals-and-sandbox --skip-git-repo-check --oss --local-provider=ollama -m qwen3.6:cloud "$prompt" 2>&1)
       local ai_exit_code=$?
       ;;
     # --- Local inference ---
@@ -365,19 +365,21 @@ get_models_for_command() {
   echo "$model_list"
 }
 
-# Get fallback models for a specific command (run or ideate)
-# Usage: get_fallback_models_for_command <command>
-# Returns: Space-separated list of fallback model names
-get_fallback_models_for_command() {
+# Get escalation models for a specific command (run or ideate)
+# AIDEV-NOTE: Escalation models are big/commercial models used only when
+# cheap primary models produce code with syntax or validation errors.
+# Usage: get_escalation_models_for_command <command>
+# Returns: Space-separated list of escalation model names
+get_escalation_models_for_command() {
   local command="$1"
   local model_list=""
 
   case "$command" in
     run)
-      model_list="$LLM_RUN_FALLBACK"
+      model_list="$LLM_RUN_ESCALATION"
       ;;
     ideate)
-      model_list="$LLM_IDEATE_FALLBACK"
+      model_list="$LLM_IDEATE_ESCALATION"
       ;;
     *)
       echo "[ERROR] Unknown command: $command" >&2

package/lib/ai_cli.py

@@ -202,22 +202,26 @@ def get_models_for_command(command: str) -> List[str]:
     return model_list.split()
 
 
-def get_fallback_models_for_command(command: str) -> List[str]:
+def get_escalation_models_for_command(command: str) -> List[str]:
     """
-    Get the list of fallback models for a command.
+    Get the list of escalation models for a command.
+
+    AIDEV-NOTE: Escalation models are big/commercial models (sonnet, codex, gemini)
+    used only when cheap primary models produce code with syntax or validation errors.
+    They are NOT for API-down situations — if primary is down, fail fast.
 
     Args:
         command: Either "run" or "ideate"
 
     Returns:
-        List of fallback model names
+        List of escalation model names
     """
     bash_script = f'''
         source "{SCRIPT_DIR}/config.sh"
         load_config
         case "$1" in
-            run) echo "$LLM_RUN_FALLBACK" ;;
-            ideate) echo "$LLM_IDEATE_FALLBACK" ;;
+            run) echo "$LLM_RUN_ESCALATION" ;;
+            ideate) echo "$LLM_IDEATE_ESCALATION" ;;
         esac
     '''
 
@@ -423,34 +427,30 @@ def call_ai_with_backoff(
     env_vars: Optional[dict] = None,
     max_rounds: int = 10,
     initial_wait: int = 60,
-    max_wait: int = 600,
-    use_fallback: bool = True
+    max_wait: int = 600
 ) -> Tuple[str, str]:
     """
-    Call AI with tiered fallback and round-based retries with exponential backoff.
+    Call AI with round-based retries and exponential backoff.
 
-    AIDEV-NOTE: Tiered fallback system:
-    1. First tries all primary models with backoff
-    2. If primary exhausted and use_fallback=True, tries fallback models
-    3. Fallback models are cheaper/simpler backups (haiku, flash, etc.)
+    AIDEV-NOTE: No fallback tier — if all primary models are down, fail fast
+    so monitoring catches it. Escalation to big models is handled separately
+    by evolve_worker when code quality issues are detected.
 
     Args:
         prompt: The prompt to send
         command: "run" or "ideate" - determines model pool
         working_dir: Directory for file operations
         env_vars: Additional environment variables
-        max_rounds: Maximum number of full rounds per tier
+        max_rounds: Maximum number of full rounds
         initial_wait: Initial wait time in seconds after first failed round
         max_wait: Maximum wait time in seconds between rounds
-        use_fallback: Whether to try fallback tier if primary fails
 
     Returns:
         Tuple of (output, model_name)
 
     Raises:
-        AIError: If all tiers exhausted without success
+        AIError: If all models exhausted without success
     """
-    # Try primary tier first
     primary_models = get_models_for_command(command)
     if not primary_models:
         raise AIError(f"No primary models configured for command: {command}")
@@ -463,28 +463,58 @@ def call_ai_with_backoff(
     if output is not None:
         return output, model_name
 
-    # Primary exhausted - try fallback if enabled
-    if use_fallback:
-        fallback_models = get_fallback_models_for_command(command)
-        if fallback_models:
-            _log(f"Primary tier exhausted, trying {len(fallback_models)} fallback models...")
+    # All models exhausted — fail fast, let monitoring catch it
+    error_summary = "; ".join(f"{m}: {e[:50]}" for m, e in list(primary_errors.items())[:3])
+    raise AIError(f"All primary models exhausted ({max_rounds} rounds). Last errors: {error_summary}")
 
-            output, model_name, fallback_errors = _try_models_with_backoff(
-                prompt, fallback_models, "Fallback",
-                working_dir, env_vars, max_rounds, initial_wait, max_wait
-            )
 
-            if output is not None:
-                return output, model_name
+def call_ai_escalation(
+    prompt: str,
+    command: str = "run",
+    working_dir: Optional[str] = None,
+    env_vars: Optional[dict] = None
+) -> Tuple[str, str]:
+    """
+    Try each escalation model once (no backoff). Used when cheap models produced
+    bad code (syntax/validation errors) and we need a smarter model to fix it.
 
-            # Both tiers exhausted
-            all_errors = {**primary_errors, **fallback_errors}
-            error_summary = "; ".join(f"{m}: {e[:50]}" for m, e in list(all_errors.items())[:3])
-            raise AIError(f"All tiers exhausted ({max_rounds} rounds each). Last errors: {error_summary}")
+    AIDEV-NOTE: Quality-triggered escalation. Each model gets one shot.
+    If all escalation models fail (rate limits etc), raise AIError — the candidate
+    will be marked failed-ai-retry so it can be retried when limits reset.
 
-    # Primary exhausted, no fallback
-    error_summary = "; ".join(f"{m}: {e[:50]}" for m, e in list(primary_errors.items())[:3])
-    raise AIError(f"Primary tier exhausted ({max_rounds} rounds). Last errors: {error_summary}")
+    Args:
+        prompt: The prompt to send (should include error context)
+        command: "run" or "ideate" - determines escalation model pool
+        working_dir: Directory for file operations
+        env_vars: Additional environment variables
+
+    Returns:
+        Tuple of (output, model_name)
+
+    Raises:
+        AIError: If no escalation models available or all failed
+    """
+    escalation_models = get_escalation_models_for_command(command)
+    if not escalation_models:
+        raise AIError(f"No escalation models configured for command: {command}")
+
+    # Shuffle so we don't always burn the same model first
+    models = escalation_models.copy()
+    random.shuffle(models)
+
+    last_errors = {}
+    for model in models:
+        try:
+            _log(f"Escalation: trying {model}...")
+            output, model_name = call_ai_model(prompt, model, working_dir, env_vars)
+            _log(f"Escalation succeeded with {model}")
+            return output, model_name
+        except AIError as e:
+            _log(f"Escalation {model} failed: {str(e)[:60]}...")
+            last_errors[model] = str(e)
+
+    error_summary = "; ".join(f"{m}: {e[:50]}" for m, e in last_errors.items())
+    raise AIError(f"All escalation models failed. Errors: {error_summary}")
 
 
 def call_ai_for_file_edit(

package/lib/config.sh

@@ -57,20 +57,22 @@ DEFAULT_MEMORY_LIMIT_MB=12288
 # Workers will exit after processing this many candidates to pick up library updates
 DEFAULT_WORKER_MAX_CANDIDATES=3
 
-# Default LLM CLI configuration - tiered fallback system
-# Primary: Strong models used in normal operation
-# Fallback: Cheap/backup models used only when primary tier exhausted
+# Default LLM CLI configuration - quality-triggered escalation system
+# AIDEV-NOTE: Two-tier design for coding (run):
+#   Primary: Cheap/open models handle normal code generation
+#   Escalation: Big commercial models only activated on syntax/validation failure
+# Ideation keeps its own primary list (thinking models for creative work)
+# No fallback tier — if all models are down, fail fast so monitoring catches it.
 #
-# Run: Subscription-based agentic models for code generation
-# All CLI tools (opencode, claude, gemini, kimi) are agentic and can edit files
-# Ollama cloud models are flat-rate (subscription), so prefer them over per-token OpenRouter
-DEFAULT_LLM_RUN="gemini-pro gemini-pro ollama-glm ollama-glm ollama-qwen ollama-qwen ollama-minimax ollama-minimax ollama-gemma ollama-gemma kimi-coder kimi-coder codex-coding codex-coding glm-zai qwen-coder minimax sonnet"
-DEFAULT_LLM_RUN_FALLBACK="haiku ollama-glm ollama-gemma ollama-minimax ollama-qwen glm-zai gemini-cheap codex-spark qwen"
+# Run: Cheap/open models for code generation (flat-rate or low-cost)
+DEFAULT_LLM_RUN="ollama-glm ollama-glm ollama-qwen ollama-qwen ollama-minimax ollama-minimax ollama-gemma ollama-gemma kimi-coder kimi-coder glm-zai qwen-coder minimax"
+# Run escalation: Big models activated only when cheap models produce bad code
+DEFAULT_LLM_RUN_ESCALATION="sonnet codex-coding gemini-pro"
 #
-# Ideate: Agentic models that can edit files for ideation
-# All CLI tools (opencode, claude, gemini, kimi) are agentic and can edit files
+# Ideate: Strong models for creative ideation
 DEFAULT_LLM_IDEATE="opus-think ollama-glm ollama-glm gemini-pro ollama-qwen ollama-minimax ollama-gemma kimi-coder gpt codex-think glm-zai qwen-coder minimax qwen"
-DEFAULT_LLM_IDEATE_FALLBACK="haiku ollama-glm ollama-gemma ollama-minimax ollama-qwen glm-zai gemini-cheap codex-spark qwen"
+# Ideate escalation: not currently used but available for future use
+DEFAULT_LLM_IDEATE_ESCALATION=""
 
 # Load configuration from a YAML file and update variables
 _load_yaml_config() {
@@ -147,12 +149,16 @@ _load_yaml_config() {
         lock_timeout) LOCK_TIMEOUT="$value" ;;
       esac
     elif [[ $in_llm_cli_section == true ]]; then
-      if [[ $key == "run" || $key == "ideate" || $key == "run_fallback" || $key == "ideate_fallback" ]]; then
+      if [[ $key == "run" || $key == "ideate" || $key == "run_escalation" || $key == "ideate_escalation" || $key == "run_fallback" || $key == "ideate_fallback" ]]; then
         case $key in
           run) LLM_RUN="$value" ;;
-          run_fallback) LLM_RUN_FALLBACK="$value" ;;
+          run_escalation) LLM_RUN_ESCALATION="$value" ;;
+          # Legacy fallback keys map to escalation for backward compatibility
+          run_fallback) LLM_RUN_ESCALATION="$value" ;;
           ideate) LLM_IDEATE="$value" ;;
-          ideate_fallback) LLM_IDEATE_FALLBACK="$value" ;;
+          ideate_escalation) LLM_IDEATE_ESCALATION="$value" ;;
+          # Legacy fallback keys map to escalation for backward compatibility
+          ideate_fallback) LLM_IDEATE_ESCALATION="$value" ;;
         esac
       else
         value=$(echo "$value" | sed "s/^'//;s/'$//")
@@ -223,9 +229,9 @@ load_config() {
   WORKER_MAX_CANDIDATES="$DEFAULT_WORKER_MAX_CANDIDATES"
   
   LLM_RUN="$DEFAULT_LLM_RUN"
-  LLM_RUN_FALLBACK="$DEFAULT_LLM_RUN_FALLBACK"
+  LLM_RUN_ESCALATION="$DEFAULT_LLM_RUN_ESCALATION"
   LLM_IDEATE="$DEFAULT_LLM_IDEATE"
-  LLM_IDEATE_FALLBACK="$DEFAULT_LLM_IDEATE_FALLBACK"
+  LLM_IDEATE_ESCALATION="$DEFAULT_LLM_IDEATE_ESCALATION"
 
   # Determine local config file path relative to EVOLUTION_DIR
   local local_config_file="$EVOLUTION_DIR/config.yaml"
@@ -318,7 +324,7 @@ show_config() {
   echo "  Memory limit: ${MEMORY_LIMIT_MB}MB"
   echo "  Worker max candidates: $WORKER_MAX_CANDIDATES"
   echo "  LLM for run: $LLM_RUN"
-  echo "  LLM for run (fallback): $LLM_RUN_FALLBACK"
+  echo "  LLM for run (escalation): $LLM_RUN_ESCALATION"
   echo "  LLM for ideate: $LLM_IDEATE"
-  echo "  LLM for ideate (fallback): $LLM_IDEATE_FALLBACK"
+  echo "  LLM for ideate (escalation): $LLM_IDEATE_ESCALATION"
 }

package/lib/evolve_worker.py

@@ -35,8 +35,8 @@ set_prefix("WORKER")
 
 from lib.evolution_csv import EvolutionCSV
 from lib.ai_cli import (
-    call_ai_with_backoff, call_ai_model, get_models_for_command,
-    get_git_protection_warning, AIError
+    call_ai_with_backoff, call_ai_escalation, call_ai_model,
+    get_models_for_command, get_git_protection_warning, AIError
 )
 from lib.llm_bandit import LLMBandit
 
@@ -170,11 +170,12 @@ CRITICAL: If you do not know how to implement what was asked for, or if the requ
 
     def _call_ai_with_backoff(self, prompt: str, target_file: Path) -> Tuple[bool, str]:
         """
-        Call AI with bandit-based model selection and fallback.
+        Call AI with bandit-based model selection from the primary (cheap) tier.
 
         AIDEV-NOTE: First tries model selected by UCB bandit.
-        If that fails, falls back to round-robin retry approach.
-        The bandit learns which models produce better algorithm improvements.
+        If that fails, retries with round-robin across all primary models.
+        The bandit learns which cheap models produce better algorithm improvements.
+        Escalation to big models is handled separately on quality failures.
 
         Returns:
             Tuple of (success, model_name)
@@ -203,14 +204,14 @@ CRITICAL: If you do not know how to implement what was asked for, or if the requ
                 else:
                     # AIDEV-NOTE: Log output so we can diagnose why file wasn't modified
                     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"Bandit model {selected_model} completed but didn't modify file ({len(output)} chars), trying other primary models...")
                     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...")
+                log(f"Bandit model {selected_model} failed: {e}, trying other primary models...")
                 # 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")
@@ -244,6 +245,42 @@ CRITICAL: If you do not know how to implement what was asked for, or if the requ
             log_error(f"All AI retries exhausted: {e}")
             return False, ""
 
+    def _call_ai_escalated(self, prompt: str, target_file: Path) -> Tuple[bool, str]:
+        """
+        Call escalation-tier AI (big commercial models) to fix code quality issues.
+
+        AIDEV-NOTE: Quality-triggered escalation. Only called when cheap primary models
+        produced code with syntax or validation errors. Each escalation model gets one
+        shot — no backoff loops. If all escalation models fail, the candidate should be
+        marked failed-ai-retry (API limits) not failed-validation (bad idea).
+
+        Returns:
+            Tuple of (success, model_name)
+        """
+        hash_before = self._file_hash(target_file) if target_file.exists() else None
+
+        try:
+            output, model = call_ai_escalation(
+                prompt,
+                command="run",
+                working_dir=self.config.evolution_dir
+            )
+
+            hash_after = self._file_hash(target_file) if target_file.exists() else None
+
+            if hash_before != hash_after and hash_after is not None:
+                log(f"Escalation AI successfully modified file (model: {model})")
+                return True, model
+            else:
+                preview = output[-300:] if output else "(empty)"
+                log(f"Escalation AI completed but did not modify file ({len(output)} chars)")
+                log(f"AI output preview: {preview}")
+                return False, model
+
+        except AIError as e:
+            log_error(f"All escalation models failed: {e}")
+            return False, ""
+
     def _file_hash(self, path: Path) -> Optional[str]:
         """Get file hash."""
         try:
@@ -550,59 +587,93 @@ python validator.py {target_basename}
                 with EvolutionCSV(self.config.csv_path) as csv:
                     csv.update_candidate_field(candidate.id, 'run-LLM', model)
 
+            # AIDEV-NOTE: Quality-triggered escalation system.
+            # Phase 1: Check syntax from cheap model output
+            # Phase 2: If syntax fails, escalate to big model with error context
+            # Phase 3: Validate, if fails escalate to big model with error context
+            # If escalation models also fail (API limits), mark failed-ai-retry.
+            # If escalation models produce code but it's still bad, mark failed-validation.
+
             # Check syntax
             if not self._check_syntax(target_file):
-                log_error("Syntax error in generated file")
-                target_file.unlink(missing_ok=True)
-                with EvolutionCSV(self.config.csv_path) as csv:
-                    csv.update_candidate_status(candidate.id, 'pending')
-                return 0  # Will retry
+                log("Syntax error from primary model, escalating to big model...")
+                # Get the syntax error details for context
+                syntax_result = subprocess.run(
+                    [self.config.python_cmd, "-m", "py_compile", str(target_file)],
+                    capture_output=True, text=True
+                )
+                syntax_error = syntax_result.stderr.strip()
 
-            # Run validator with retry loop
-            # AIDEV-NOTE: Validator catches structural errors before expensive full evaluation.
-            # If validation fails, we give the AI feedback and ask it to fix the code.
-            validation_passed = False
-            for validation_attempt in range(self.config.max_validation_retries + 1):
-                valid, error_info = self._run_validator(candidate.id)
+                fix_prompt = self._build_fix_prompt(
+                    candidate, target_file.name,
+                    {'error_type': 'syntax', 'error': syntax_error}
+                )
+                success, fix_model = self._call_ai_escalated(fix_prompt, target_file)
 
-                if valid:
-                    validation_passed = True
-                    break
+                if not success:
+                    log_error("Escalation models failed to fix syntax error")
+                    target_file.unlink(missing_ok=True)
+                    with EvolutionCSV(self.config.csv_path) as csv:
+                        csv.update_candidate_status(candidate.id, 'failed-ai-retry')
+                    return 77
 
-                if validation_attempt >= self.config.max_validation_retries:
-                    log_error(f"Validation failed after {self.config.max_validation_retries} fix attempts")
-                    break
+                # Record escalation model
+                if fix_model:
+                    with EvolutionCSV(self.config.csv_path) as csv:
+                        current_llm = csv.get_candidate_info(candidate.id).get('run-LLM', '')
+                        new_llm = f"{current_llm}+ESC:{fix_model}" if current_llm else f"ESC:{fix_model}"
+                        csv.update_candidate_field(candidate.id, 'run-LLM', new_llm)
+
+                # Re-check syntax after escalation fix
+                if not self._check_syntax(target_file):
+                    log_error("Escalation model also produced syntax error — idea too hard")
+                    target_file.unlink(missing_ok=True)
+                    with EvolutionCSV(self.config.csv_path) as csv:
+                        csv.update_candidate_status(candidate.id, 'failed-validation')
+                    return 1
+
+            # Run validator with escalation on failure
+            # AIDEV-NOTE: Validator catches structural errors before expensive full evaluation.
+            # First attempt uses the code as-is from primary model. On failure, escalate once.
+            valid, error_info = self._run_validator(candidate.id)
 
-                # Ask AI to fix the validation error
-                log(f"Validation failed (attempt {validation_attempt + 1}), asking AI to fix...")
+            if not valid:
+                log("Validation failed from primary model, escalating to big model...")
                 fix_prompt = self._build_fix_prompt(candidate, target_file.name, error_info)
-                success, fix_model = self._call_ai_with_backoff(fix_prompt, target_file)
+                success, fix_model = self._call_ai_escalated(fix_prompt, target_file)
 
                 if not success:
-                    log_error("AI failed to fix validation error")
-                    break
+                    log_error("Escalation models failed to fix validation error")
+                    target_file.unlink(missing_ok=True)
+                    with EvolutionCSV(self.config.csv_path) as csv:
+                        csv.update_candidate_status(candidate.id, 'failed-ai-retry')
+                    return 77
 
-                # Record that we used an additional model call for fixing
+                # Record escalation model
                 if fix_model:
                     with EvolutionCSV(self.config.csv_path) as csv:
                         current_llm = csv.get_candidate_info(candidate.id).get('run-LLM', '')
-                        new_llm = f"{current_llm}+{fix_model}" if current_llm else fix_model
+                        new_llm = f"{current_llm}+ESC:{fix_model}" if current_llm else f"ESC:{fix_model}"
                         csv.update_candidate_field(candidate.id, 'run-LLM', new_llm)
 
-                # Re-check syntax after fix
+                # Check syntax after escalation fix (escalation might break it)
                 if not self._check_syntax(target_file):
-                    log_error("Fix introduced syntax error")
-                    # Don't break - try again if we have retries left
+                    log_error("Escalation fix introduced syntax error")
+                    target_file.unlink(missing_ok=True)
+                    with EvolutionCSV(self.config.csv_path) as csv:
+                        csv.update_candidate_status(candidate.id, 'failed-validation')
+                    return 1
 
-            if not validation_passed:
-                # Validation failed after all retries
-                with EvolutionCSV(self.config.csv_path) as csv:
-                    csv.update_candidate_status(candidate.id, 'failed-validation')
-                    # Store the last error for debugging
-                    if error_info:
-                        error_summary = f"{error_info.get('error_type', 'unknown')}: {error_info.get('error', '')[:100]}"
-                        csv.update_candidate_field(candidate.id, 'validation_error', error_summary)
-                return 1
+                # Re-validate after escalation fix
+                valid, error_info = self._run_validator(candidate.id)
+                if not valid:
+                    log_error("Validation still fails after escalation — idea too hard")
+                    with EvolutionCSV(self.config.csv_path) as csv:
+                        csv.update_candidate_status(candidate.id, 'failed-validation')
+                        if error_info:
+                            error_summary = f"{error_info.get('error_type', 'unknown')}: {error_info.get('error', '')[:100]}"
+                            csv.update_candidate_field(candidate.id, 'validation_error', error_summary)
+                    return 1
 
         # Run evaluator
         log("Running evaluator...")

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "claude-evolve",
-  "version": "1.12.0",
+  "version": "1.14.0",
   "bin": {
     "claude-evolve": "bin/claude-evolve",
     "claude-evolve-main": "bin/claude-evolve-main",

package/templates/config.yaml

@@ -86,47 +86,49 @@ parallel:
   # Timeout in seconds when waiting for CSV locks
   lock_timeout: 30
 
-# LLM/AI CLI configuration
+# LLM/AI CLI configuration — quality-triggered escalation system
+# AIDEV-NOTE: Two tiers for coding:
+#   Primary (run): Cheap/open models for normal code generation
+#   Escalation (run_escalation): Big models activated ONLY on syntax/validation failure
+# No fallback tier — if all primary models are down, fail fast so monitoring catches it.
 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., "gemini-pro gemini-pro sonnet" for 2:1 ratio)
+  # Models are selected randomly, repeat for weighted selection (e.g., "model model" = 2x weight)
+  # Commented out because defaults change over time; uncomment to override
 
-  # Default configuration: sonnet at ~11%, rest doubled for cost savings
-  # Commented out because these change over time; uncomment to override
-  #run: gemini-pro gemini-pro glm-5-zai glm-5-zai kimi-coder kimi-coder codex-coding codex-coding sonnet
-  #ideate: opus-think glm-5-zai gemini-pro kimi-coder gpt-5.2 codex-think qwen-openrouter
-  #run_fallback: haiku glm-5-zai gemini-5-flash codex-spark
-  #ideate_fallback: haiku glm-5-zai gemini-5-flash codex-spark
+  # Coding: cheap/open models (flat-rate or low per-token cost)
+  #run: ollama-glm ollama-glm ollama-qwen ollama-qwen ollama-minimax ollama-minimax ollama-gemma ollama-gemma kimi-coder kimi-coder glm-zai qwen-coder minimax
+
+  # Coding escalation: big commercial models, only used when primary produces bad code
+  #run_escalation: sonnet codex-coding gemini-pro
+
+  # Ideation: strong thinking models for creative work
+  #ideate: opus-think ollama-glm ollama-glm gemini-pro ollama-qwen ollama-minimax ollama-gemma kimi-coder gpt codex-think glm-zai qwen-coder minimax qwen
+
+  # Ideation escalation: not currently used
+  #ideate_escalation:
 
   # Available models:
-  # 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):
-  # - codex-think: GPT-5.4 high reasoning effort (ideation)
-  # - codex-coding: GPT-5.4 medium reasoning effort (coding/run)
-  # - codex-spark: GPT-5.1 Codex Mini (lightweight fallback)
-  # - gpt-5.4: GPT-5.4 no reasoning effort override via Codex CLI
-  # - gpt-5.2: GPT-5.2 via Codex CLI
-  # - gpt-5.3-codex: GPT-5.3 Codex (code-specialized) 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)
+  # --- Cheap/open (primary coding tier) ---
+  # Ollama cloud (flat-rate $20/mo subscription):
+  # - ollama-glm: GLM 5.1 via Ollama cloud
+  # - ollama-qwen: Qwen 3.6 via Ollama cloud
+  # - ollama-minimax: MiniMax M2.7 via Ollama cloud
+  # - ollama-gemma: Gemma 4 31B via Ollama cloud
+  #
+  # Low-cost APIs:
+  # - kimi-coder: Kimi for Coding via kimi CLI (subscription)
+  # - glm-zai: GLM 5 via Z.AI agentic mode
+  # - qwen-coder: Qwen 3 Coder via OpenRouter
+  # - minimax: MiniMax M2.7 via OpenRouter
+  # - qwen: Qwen 3.6 Plus via OpenRouter
   #
-  # 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)
+  # --- Big commercial (escalation tier) ---
+  # - sonnet: Claude Sonnet via Claude CLI
+  # - codex-coding: GPT-5.4 medium reasoning via Codex CLI
+  # - gemini-pro: Gemini 3 Pro via Gemini CLI
   #
-  # 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
+  # --- Ideation models ---
+  # - opus-think: Claude Opus with extended thinking
+  # - codex-think: GPT-5.5 high reasoning effort
+  # - gpt: GPT-5.5 via Codex CLI