claude-evolve 1.11.13 → 1.11.14

package/lib/ai-cli.sh

@@ -1,13 +1,11 @@
 #!/bin/bash
 # Centralized AI CLI invocation library for claude-evolve
 #
-# AIDEV-NOTE: All timeout commands use -k flag to ensure process termination
-# The -k flag sends SIGKILL if the process doesn't respond to SIGTERM within
-# the grace period (30 seconds). This prevents AI CLI processes from hanging
-# indefinitely when they ignore the initial SIGTERM signal.
-# Example: timeout -k 30 600 means:
-#   - Wait 600 seconds, then send SIGTERM
-#   - If still running after 30 more seconds, send SIGKILL (force kill)
+# AIDEV-NOTE: Timeouts are now handled by the Python caller (ai_cli.py), not by
+# bash timeout commands. This allows for better control and monitoring of AI CLI
+# processes from the Python layer, including graceful timeout handling and
+# proper error recovery. The bash functions here focus on clean command execution
+# without timeout wrapping.
 
 # Source config to get LLM_CLI array and model lists
 # This will be sourced after config.sh in the main scripts
@@ -64,183 +62,183 @@ call_ai_model_configured() {
   case "$model_name" in
     opus)
       local ai_output
-      ai_output=$(timeout -k 30 300 claude --dangerously-skip-permissions --mcp-config '' --model opus -p "$prompt" 2>&1)
+      ai_output=$(claude --dangerously-skip-permissions --mcp-config '' --model opus -p "$prompt" 2>&1)
       local ai_exit_code=$?
       ;;
     sonnet)
       local ai_output
-      ai_output=$(timeout -k 30 300 claude --dangerously-skip-permissions --mcp-config '' --model sonnet -p "$prompt" 2>&1)
+      ai_output=$(claude --dangerously-skip-permissions --mcp-config '' --model sonnet -p "$prompt" 2>&1)
       local ai_exit_code=$?
       ;;
     sonnet-think)
       local ai_output
       # Use extended thinking with sonnet 4.5 - prepend ultrathink instruction
-      # AIDEV-NOTE: Extended thinking needs 30 min timeout - can take long for complex ideation
+      # AIDEV-NOTE: Extended thinking can take long for complex ideation
       local think_prompt="ultrathink
 
 $prompt"
-      ai_output=$(timeout -k 30 1800 claude --dangerously-skip-permissions --mcp-config '' --model sonnet -p "$think_prompt" 2>&1)
+      ai_output=$(claude --dangerously-skip-permissions --mcp-config '' --model sonnet -p "$think_prompt" 2>&1)
       local ai_exit_code=$?
       ;;
     opus-think)
       local ai_output
       # Use extended thinking with opus - prepend ultrathink instruction
-      # AIDEV-NOTE: Extended thinking needs 30 min timeout - can take long for complex ideation
+      # AIDEV-NOTE: Extended thinking can take long for complex ideation
       local think_prompt="ultrathink
 
 $prompt"
-      ai_output=$(timeout -k 30 1800 claude --dangerously-skip-permissions --mcp-config '' --model opus -p "$think_prompt" 2>&1)
+      ai_output=$(claude --dangerously-skip-permissions --mcp-config '' --model opus -p "$think_prompt" 2>&1)
       local ai_exit_code=$?
       ;;
     haiku)
       local ai_output
-      ai_output=$(timeout -k 30 300 claude --dangerously-skip-permissions --mcp-config '' --model haiku -p "$prompt" 2>&1)
+      ai_output=$(claude --dangerously-skip-permissions --mcp-config '' --model haiku -p "$prompt" 2>&1)
       local ai_exit_code=$?
       ;;
     gpt5high)
       local ai_output
-      ai_output=$(timeout -k 30 600 codex exec -m "$codex_gpt5_model" -c model_reasoning_effort="high" --dangerously-bypass-approvals-and-sandbox "$prompt" 2>&1)
+      ai_output=$(codex exec -m "$codex_gpt5_model" -c model_reasoning_effort="high" --dangerously-bypass-approvals-and-sandbox "$prompt" 2>&1)
       local ai_exit_code=$?
       ;;
     gpt5)
       local ai_output
-      ai_output=$(timeout -k 30 600 codex exec -m "$codex_gpt5_model" --dangerously-bypass-approvals-and-sandbox "$prompt" 2>&1)
+      ai_output=$(codex exec -m "$codex_gpt5_model" --dangerously-bypass-approvals-and-sandbox "$prompt" 2>&1)
       local ai_exit_code=$?
       ;;
     o3high)
       local ai_output
-      ai_output=$(timeout -k 30 600 codex exec -m o3-mini -c model_reasoning_effort="high" --dangerously-bypass-approvals-and-sandbox "$prompt" 2>&1)
+      ai_output=$(codex exec -m o3-mini -c model_reasoning_effort="high" --dangerously-bypass-approvals-and-sandbox "$prompt" 2>&1)
       local ai_exit_code=$?
       ;;
     gemini-pro)
       local ai_output
-      # Gemini needs longer timeout as it streams output while working (20 minutes)
-      ai_output=$(timeout -k 30 1800 gemini -y -m gemini-3-pro-preview -p "$prompt" 2>&1)
+      # Gemini streams output while working
+      ai_output=$(gemini -y -m gemini-3-pro-preview -p "$prompt" 2>&1)
       local ai_exit_code=$?
       ;;
     gemini-flash)
       local ai_output
-      # Gemini needs longer timeout as it streams output while working (20 minutes)
-      ai_output=$(timeout -k 30 1200 gemini -y -m gemini-2.5-flash -p "$prompt" 2>&1)
+      # Gemini streams output while working
+      ai_output=$(gemini -y -m gemini-2.5-flash -p "$prompt" 2>&1)
       local ai_exit_code=$?
       ;;
     gemini-3-pro-preview)
       local ai_output
-      # Gemini v3 Pro Preview via OpenRouter (30 minute timeout) - EXPENSIVE
-      ai_output=$(timeout -k 30 1800 opencode -m openrouter/google/gemini-3-pro-preview run "$prompt" 2>&1)
+      # Gemini v3 Pro Preview via OpenRouter - EXPENSIVE
+      ai_output=$(opencode -m openrouter/google/gemini-3-pro-preview run "$prompt" 2>&1)
       local ai_exit_code=$?
       ;;
     gemini-3-flash)
       local ai_output
       # Gemini 3 Flash - fast, cheap, strong thinker
-      ai_output=$(timeout -k 30 600 opencode -m openrouter/google/gemini-3-flash-preview run "$prompt" 2>&1)
+      ai_output=$(opencode -m openrouter/google/gemini-3-flash-preview run "$prompt" 2>&1)
       local ai_exit_code=$?
       ;;
     cursor-sonnet)
       local ai_output
-      ai_output=$(timeout -k 30 600 cursor-agent sonnet-4.5 -p "$prompt" 2>&1)
+      ai_output=$(cursor-agent sonnet-4.5 -p "$prompt" 2>&1)
       local ai_exit_code=$?
       ;;
     cursor-opus)
       local ai_output
-      ai_output=$(timeout -k 30 600 cursor-agent opus -p "$prompt" 2>&1)
+      ai_output=$(cursor-agent opus -p "$prompt" 2>&1)
       local ai_exit_code=$?
       ;;
     glm-openrouter)
       local ai_output
-      ai_output=$(timeout -k 30 600 opencode -m openrouter/z-ai/glm-4.7 run "$prompt" 2>&1)
+      ai_output=$(opencode -m openrouter/z-ai/glm-4.7 run "$prompt" 2>&1)
       local ai_exit_code=$?
       ;;
     glm-5)
       local ai_output
       # GLM-5: 744B MoE model, very cheap ($0.80/$2.56 per 1M tokens), 200K context
       # Released Feb 2026 - scores 77.8% SWE-bench, MIT license
-      ai_output=$(timeout -k 30 600 opencode -m openrouter/z-ai/glm-5 run "$prompt" 2>&1)
+      ai_output=$(opencode -m openrouter/z-ai/glm-5 run "$prompt" 2>&1)
       local ai_exit_code=$?
       ;;
     glm-zai)
       # GLM 4.7 via Z.AI agentic mode -- can be slow sometimes
       local ai_output
-      ai_output=$(timeout -k 30 1800 opencode -m zai-coding-plan/glm-4.7 run "$prompt" 2>&1)
+      ai_output=$(opencode -m zai-coding-plan/glm-4.7 run "$prompt" 2>&1)
       local ai_exit_code=$?
       ;;
     glm-5-zai)
       # GLM-5 via Z.AI agentic mode - supports file editing for ideation
       # 744B MoE, strong reasoning, can edit files
       local ai_output
-      ai_output=$(timeout -k 30 1800 opencode -m zai-coding-plan/glm-5 run "$prompt" 2>&1)
+      ai_output=$(opencode -m zai-coding-plan/glm-5 run "$prompt" 2>&1)
       local ai_exit_code=$?
       ;;
     deepseek-openrouter)
       local ai_output
-      ai_output=$(timeout -k 30 600 opencode -m openrouter/deepseek/deepseek-v3.2 run "$prompt" 2>&1)
+      ai_output=$(opencode -m openrouter/deepseek/deepseek-v3.2 run "$prompt" 2>&1)
       local ai_exit_code=$?
       ;;
     grok-code-fast-openrouter)
       local ai_output
-      ai_output=$(timeout -k 30 600 opencode -m openrouter/x-ai/grok-code-fast-1 run "$prompt" 2>&1)
+      ai_output=$(opencode -m openrouter/x-ai/grok-code-fast-1 run "$prompt" 2>&1)
       local ai_exit_code=$?
       ;;
     grok-4-openrouter)
       local ai_output
       # EXPENSIVE - consider grok-4.1-fast instead
-      ai_output=$(timeout -k 30 600 opencode -m openrouter/x-ai/grok-4 run "$prompt" 2>&1)
+      ai_output=$(opencode -m openrouter/x-ai/grok-4 run "$prompt" 2>&1)
       local ai_exit_code=$?
       ;;
     grok-4.1-fast)
       local ai_output
       # Grok 4.1 Fast - close to Grok 4 quality, much cheaper
-      ai_output=$(timeout -k 30 600 opencode -m openrouter/x-ai/grok-4.1-fast run "$prompt" 2>&1)
+      ai_output=$(opencode -m openrouter/x-ai/grok-4.1-fast run "$prompt" 2>&1)
       local ai_exit_code=$?
       ;;
     opus-openrouter)
       local ai_output
-      ai_output=$(timeout -k 30 600 opencode -m openrouter/anthropic/claude-opus-4.1 run "$prompt" 2>&1)
+      ai_output=$(opencode -m openrouter/anthropic/claude-opus-4.1 run "$prompt" 2>&1)
       local ai_exit_code=$?
       ;;
     kimi-k2-openrouter)
       local ai_output
       # Kimi K2 Thinking via OpenRouter (no separate auth needed)
-      ai_output=$(timeout -k 30 600 opencode -m openrouter/moonshotai/kimi-k2-thinking run "$prompt" 2>&1)
+      ai_output=$(opencode -m openrouter/moonshotai/kimi-k2-thinking run "$prompt" 2>&1)
       local ai_exit_code=$?
       ;;
     kimi-k2-think-moonshot)
       local ai_output
       # Use kimi CLI directly (assumes kimi is installed and configured)
-      ai_output=$(timeout -k 30 600 kimi --print -c "$prompt" 2>&1)
+      ai_output=$(kimi --print -c "$prompt" 2>&1)
       local ai_exit_code=$?
       ;;
     kimi-coder)
       local ai_output
       # Kimi for Coding model via kimi CLI (fast coding-focused model)
       # Use --print to see agent actions while still allowing file modifications
-      ai_output=$(timeout -k 30 600 kimi --print -y -m kimi-for-coding -c "$prompt" 2>&1)
+      ai_output=$(kimi --print -y -m kimi-for-coding -c "$prompt" 2>&1)
       local ai_exit_code=$?
       ;;
     kimi-k2.5)
       local ai_output
       # Kimi K2.5 - Moonshot's most powerful model (Jan 2025)
       # Native multimodal agentic model, stronger than GLM-4.7
-      ai_output=$(timeout -k 30 600 opencode -m openrouter/moonshotai/kimi-k2.5 run "$prompt" 2>&1)
+      ai_output=$(opencode -m openrouter/moonshotai/kimi-k2.5 run "$prompt" 2>&1)
       local ai_exit_code=$?
       ;;
     qwen)
       local ai_output
       # Qwen latest - Alibaba's flagship model (currently qwen3.5-plus)
       # Linear attention + sparse MoE, strong multimodal capabilities
-      ai_output=$(timeout -k 30 600 opencode -m openrouter/qwen/qwen3.5-plus-02-15 run "$prompt" 2>&1)
+      ai_output=$(opencode -m openrouter/qwen/qwen3.5-plus-02-15 run "$prompt" 2>&1)
       local ai_exit_code=$?
       ;;
     codex-oss-local)
       # Codex-OSS via Codex CLI with Ollama backend
       local ai_output
-      ai_output=$(timeout -k 30 2400 codex exec --dangerously-bypass-approvals-and-sandbox --skip-git-repo-check --oss --local-provider=ollama "$prompt" 2>&1)
+      ai_output=$(codex exec --dangerously-bypass-approvals-and-sandbox --skip-git-repo-check --oss --local-provider=ollama "$prompt" 2>&1)
       local ai_exit_code=$?
       ;;
     deepseek-v3-llamacloud)
       # Deepseek via Codex CLI with Ollama cloud backend
       local ai_output
-      ai_output=$(timeout -k 30 600 codex exec --dangerously-bypass-approvals-and-sandbox --skip-git-repo-check --oss -m deepseek-v3.1:671b-cloud "$prompt" 2>&1)
+      ai_output=$(codex exec --dangerously-bypass-approvals-and-sandbox --skip-git-repo-check --oss -m deepseek-v3.1:671b-cloud "$prompt" 2>&1)
       local ai_exit_code=$?
       ;;
   esac

package/lib/ai_cli.py

@@ -1,11 +1,14 @@
 #!/usr/bin/env python3
 """
 Python wrapper around ai-cli.sh for AI model invocation.
-AIDEV-NOTE: This keeps ai-cli.sh as the source of truth for model configs and timeouts.
+AIDEV-NOTE: ai-cli.sh handles model commands, Python handles timeouts.
+Timeouts were moved from bash to Python because the bash `timeout` command
+causes claude CLI and gemini CLI to hang in nested subprocess contexts.
 """
 
 import os
 import random
+import signal
 import subprocess
 import sys
 import tempfile
@@ -234,6 +237,29 @@ def get_fallback_models_for_command(command: str) -> List[str]:
     return model_list.split()
 
 
+# AIDEV-NOTE: Model timeout lookup. Timeouts are handled here in Python, not
+# in bash, because the bash `timeout` command causes claude CLI (and sometimes
+# gemini CLI) to hang when called from nested subprocess contexts.
+MODEL_TIMEOUTS = {
+    # Claude models - 5 minutes for standard, 30 minutes for thinking
+    'opus': 300, 'sonnet': 300, 'haiku': 300,
+    '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,
+    # 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)
+    'codex-oss-local': 2400,
+}
+DEFAULT_MODEL_TIMEOUT = 600  # 10 minutes for everything else
+
+
+def get_model_timeout(model_name: str) -> int:
+    """Get the timeout in seconds for a given model."""
+    return MODEL_TIMEOUTS.get(model_name, DEFAULT_MODEL_TIMEOUT)
+
+
 def call_ai_model(
     prompt: str,
     model_name: str,
@@ -243,6 +269,10 @@ def call_ai_model(
     """
     Call a specific AI model.
 
+    AIDEV-NOTE: Timeouts are handled here in Python using subprocess.Popen +
+    process group kill. The bash timeout command was removed because it causes
+    claude CLI and gemini CLI to hang in nested subprocess contexts.
+
     Args:
         prompt: The prompt to send to the AI
         model_name: The specific model to use
@@ -268,18 +298,34 @@ def call_ai_model(
     if env_vars:
         env.update(env_vars)
 
+    timeout_secs = get_model_timeout(model_name)
+
     try:
-        result = subprocess.run(
+        # Use Popen with process group so we can kill all children on timeout
+        proc = subprocess.Popen(
             ["bash", "-c", bash_script, "bash", model_name, prompt],
-            capture_output=True,
+            stdout=subprocess.PIPE,
+            stderr=subprocess.PIPE,
             text=True,
             cwd=working_dir,
-            env=env
+            env=env,
+            start_new_session=True,  # Creates new process group
         )
 
-        output = result.stdout
-        stderr = result.stderr
-        exit_code = result.returncode
+        try:
+            output, stderr = proc.communicate(timeout=timeout_secs)
+            exit_code = proc.returncode
+        except subprocess.TimeoutExpired:
+            # Kill the entire process group
+            try:
+                os.killpg(os.getpgid(proc.pid), signal.SIGTERM)
+                time.sleep(2)
+                os.killpg(os.getpgid(proc.pid), signal.SIGKILL)
+            except (ProcessLookupError, OSError):
+                pass
+            proc.kill()
+            proc.wait()
+            raise TimeoutError(f"AI call timed out (model: {model_name})")
 
         # Print stderr (contains debug info)
         if stderr:
@@ -299,7 +345,9 @@ def call_ai_model(
 
         return output, model_name
 
-    except subprocess.SubprocessError as e:
+    except (subprocess.SubprocessError, OSError) as e:
+        if isinstance(e, subprocess.TimeoutExpired):
+            raise TimeoutError(f"AI call timed out (model: {model_name})")
         raise AIError(f"Failed to call AI: {e}")
 
 

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "claude-evolve",
-  "version": "1.11.13",
+  "version": "1.11.14",
   "bin": {
     "claude-evolve": "bin/claude-evolve",
     "claude-evolve-main": "bin/claude-evolve-main",