aiwcli 0.9.4 → 0.9.6

package/dist/templates/_shared/hooks/user_prompt_submit.py

@@ -103,7 +103,7 @@ def _update_in_flight_status(context_id: str, hook_input: dict, project_root: Pa
     elif permission_mode != "plan":
         # Any non-plan permission mode transitions pending/planning to implementing
         # This includes "default" (after /clear) and "acceptEdits"/"bypassPermissions"
-        if current_mode in ["pending_implementation", "planning"]:
+        if current_mode in ["pending_implementation", "planning", "none"]:
             update_plan_status(context_id, "implementing", project_root=project_root)
             eprint(f"[user_prompt_submit] Set status to 'implementing' (permission_mode={permission_mode})")
 

package/dist/templates/_shared/lib/base/stop_words.py

@@ -155,4 +155,36 @@ STOP_WORDS = {
     'part', 'parts', 'point', 'points',
     'time', 'times', 'next', 'last', 'end',
     'set', 'list', 'group', 'item', 'items',
+
+    # ========================================================================
+    # PROGRAMMING KEYWORDS (code tokens that leak from pasted code)
+    # ========================================================================
+    'self', 'def', 'return', 'import', 'true', 'false', 'none', 'str',
+    'const', 'async', 'class', 'assert', 'except', 'dict', 'len', 'args',
+    'sys', 'eprint', 'elif', 'lambda', 'yield', 'pass',
+
+    # ========================================================================
+    # GENERIC ADJECTIVES (non-specific modifiers)
+    # ========================================================================
+    'high', 'low', 'important', 'critical', 'optional', 'manual',
+    'real', 'empty', 'stable', 'active', 'proper', 'correct',
+    'basic', 'main', 'primary', 'secondary', 'general', 'overall',
+
+    # ========================================================================
+    # GENERIC TECHNICAL NOUNS (common but non-specific)
+    # ========================================================================
+    'information', 'format', 'status', 'method', 'purpose', 'result',
+    'source', 'value', 'option', 'options', 'feature', 'features', 'issue',
+    'process', 'version', 'mode', 'state',
+
+    # ========================================================================
+    # DOCUMENT/CODE STRUCTURE (generic structural terms)
+    # ========================================================================
+    'section', 'lines', 'line', 'folder', 'directory', 'index',
+    'level', 'block', 'chunk', 'region', 'header', 'footer',
+
+    # ========================================================================
+    # FRAGMENT WORDS (artifacts from contractions/tokenization)
+    # ========================================================================
+    're', 'pl', 'aiw', 've', 'll', 'doesn', 't', 's',
 }

package/dist/templates/_shared/lib/base/utils.py

@@ -138,8 +138,7 @@ def generate_context_id(summary: str, existing_ids: Optional[set] = None) -> str
     """
     Generate a context ID from a summary string.
 
-    Uses AI inference to create a semantic 10-word summary, then slugifies it.
-    Falls back to truncate-and-slugify if inference fails.
+    Filters stop words from the summary and slugifies the result.
 
     Args:
         summary: Context summary text
@@ -151,24 +150,10 @@ def generate_context_id(summary: str, existing_ids: Optional[set] = None) -> str
     if not summary or not summary.strip():
         base_id = "context"
     else:
-        # Try AI-powered semantic summary first
-        base_id = None
-        try:
-            from .inference import generate_semantic_summary
-            semantic = generate_semantic_summary(summary)
-            if semantic:
-                # Slugify the semantic summary (word limit already applied in inference)
-                base_id = sanitize_title(semantic, max_len=100)
-                eprint(f"[utils] Semantic context ID: {base_id}")
-        except Exception as e:
-            eprint(f"[utils] Inference failed, using fallback: {e}")
-
-        # Fallback to old method if inference failed
-        if not base_id:
-            # Fallback: use stop word filter, limit to 12 words
-            from .stop_words import STOP_WORDS
-            words = [w for w in summary.lower().split() if w not in STOP_WORDS and len(w) > 1][:12]
-            base_id = sanitize_title(' '.join(words), max_len=100)
+        # Use stop word filter, limit to 12 words
+        from .stop_words import STOP_WORDS
+        words = [w for w in summary.lower().split() if w not in STOP_WORDS and len(w) > 1][:12]
+        base_id = sanitize_title(' '.join(words), max_len=100)
 
     if not existing_ids:
         return base_id

package/dist/templates/cc-native/{.claude/agents/cc-native → _cc-native/agents}/ACCESSIBILITY-TESTER.md

@@ -7,7 +7,6 @@ enabled: false
 categories:
   - code
   - design
-tools: Read, Grep, Glob, Bash
 ---
 
 ## Role

package/dist/templates/cc-native/_cc-native/agents/CLAUDE.md

@@ -0,0 +1,56 @@
+# CC-Native Plan Review Agents
+
+Agent persona definitions for single-turn plan review.
+
+## System Prompt vs Agent Flag
+
+**Decision:** Use `--system-prompt` with markdown body content instead of `--agent <name>`
+
+**Rationale:**
+- Claude Code's `--agent` flag invokes built-in agents designed for multi-turn agentic workflows with tool access
+- Plan review needs single-turn text analysis: read plan, output structured JSON
+- The `--agent` flag ignores our custom markdown content entirely - it loads Claude Code's built-in agent definitions
+- Using `--system-prompt` lets us inject the full persona (expertise, review approach, output requirements) directly
+- Result: faster execution, no tool overhead, and our rich agent descriptions actually get used
+
+**Constraint:** If you switch back to `--agent`, the elaborate persona content in these markdown files will be ignored. The reviews will use Claude Code's generic agent behavior instead of our specialized reviewers.
+
+## File Structure
+
+Each agent file has:
+- **Frontmatter (YAML):** name, model, focus, categories, enabled
+- **Body (Markdown):** Full persona content → becomes `system_prompt` for `--system-prompt` flag
+
+The `aggregate_agents.py` script extracts both parts. The body becomes `AgentConfig.system_prompt`.
+
+## --setting-sources "" Requirement
+
+**Decision:** Use `--setting-sources ""` to disable user/project settings loading
+
+**Rationale:**
+- Without this flag, Claude Code loads user settings (~43k cached tokens of PAI context)
+- The PAI Algorithm instructions override the agent's system prompt behavior
+- Model tries to follow PAI format instead of calling StructuredOutput directly
+- Result: 6+ turns, 30+ seconds, often no structured output
+
+**Constraint:** If you remove `--setting-sources ""`, agent reviews will be slow and unreliable due to PAI context interference.
+
+## --max-turns 3 Requirement
+
+**Decision:** Use `--max-turns 3` with agent invocations
+
+**Rationale:**
+- `--max-turns 1` is too restrictive - the model needs turn 1 to call StructuredOutput, turn 2 for the tool result
+- `--max-turns 2` works but leaves no buffer for edge cases
+- `--max-turns 3` gives safety margin while still preventing runaway multi-turn behavior
+- With these settings, reviews complete in ~5-10 seconds
+
+**Constraint:** The agent markdown files MUST contain clear instructions to "call StructuredOutput IMMEDIATELY" and "do NOT use any other tools". Without these instructions, the model will try to use its turns for file operations instead of outputting the review.
+
+## enabled: false Convention
+
+**Decision:** Set `enabled: false` in frontmatter for all plan review agents
+
+**Rationale:** The `enabled` field controls Claude Code's auto-suggestion feature (showing agents in command palette). For plan review agents, we don't want them appearing as general-purpose agents - they're invoked programmatically by the hook. Setting `enabled: false` hides them from auto-suggestion while still allowing the hook to use them.
+
+**Constraint:** Don't set `enabled: true` unless you want the agent to appear in Claude Code's agent picker for general use.

package/dist/templates/cc-native/{.claude/agents/cc-native → _cc-native/agents}/CODE-REVIEWER.md

@@ -6,7 +6,6 @@ focus: code quality and security
 enabled: false
 categories:
   - code
-tools: Read, Write, Edit, Bash, Glob, Grep
 ---
 
 ## Role

package/dist/templates/cc-native/{.claude/agents/cc-native → _cc-native/agents}/CONTEXT-EXTRACTOR.md

@@ -9,7 +9,6 @@ categories:
   - infrastructure
   - design
   - research
-tools: ""
 ---
 
 You are a Context Extractor agent that prepares abstracted problem context for the Fresh Perspective agent. Your job is to distill a conversation into its essential elements while **stripping away implementation details** that could bias fresh thinking.

package/dist/templates/cc-native/{.claude/agents/cc-native → _cc-native/agents}/PENETRATION-TESTER.md

@@ -7,7 +7,6 @@ enabled: false
 categories:
   - code
   - infrastructure
-tools: Read, Grep, Glob, Bash
 ---
 
 ## Role

package/dist/templates/cc-native/{.claude/agents/cc-native → _cc-native/agents}/PERFORMANCE-ENGINEER.md

@@ -7,7 +7,6 @@ enabled: false
 categories:
   - code
   - infrastructure
-tools: Read, Write, Edit, Bash, Glob, Grep
 ---
 
 ## Role

package/dist/templates/cc-native/{.claude/agents/cc-native → _cc-native/agents}/PLAN-ORCHESTRATOR.md

@@ -6,7 +6,6 @@ focus: plan complexity analysis and agent routing
 enabled: false
 categories:
   - orchestration
-tools: Read, Glob, Grep
 ---
 
 You are a plan orchestration agent. Your job is to analyze implementation plans and determine:

package/dist/templates/cc-native/{.claude/agents/cc-native → _cc-native/agents}/TRADE-OFF-ILLUMINATOR.md

@@ -12,7 +12,6 @@ categories:
   - research
   - life
   - business
-tools: Read, Glob, Grep
 ---
 
 You are a trade-off illuminator who makes hidden costs explicit. While other agents ask "Is this approach good?", you ask "What are you giving up to get this?" Your focus is exposing the price of every decision—the capabilities sacrificed, the stakeholders who lose, the futures foreclosed.

package/dist/templates/cc-native/_cc-native/hooks/cc-native-plan-review.py

@@ -118,7 +118,6 @@ DEFAULT_ORCHESTRATOR: Dict[str, Any] = {
     "enabled": True,
     "model": "haiku",
     "timeout": 30,
-    "maxTurns": 3,
 }
 
 DEFAULT_AGENT_MODEL: str = "sonnet"
@@ -331,10 +330,9 @@ def load_settings(proj_dir: Path) -> Dict[str, Any]:
         "agentReview": {
             "enabled": True,
             "orchestrator": DEFAULT_ORCHESTRATOR.copy(),
-            "timeout": 120,
+            "timeout": 180,
             "blockOnFail": True,
             "legacyMode": False,
-            "maxTurns": 3,
             "display": DEFAULT_DISPLAY.copy(),
             "agentSelection": DEFAULT_AGENT_SELECTION.copy(),
             "agentDefaults": {"model": DEFAULT_AGENT_MODEL},
@@ -383,9 +381,13 @@ def load_settings(proj_dir: Path) -> Dict[str, Any]:
 
 
 def load_agent_library(proj_dir: Path, settings: Optional[Dict[str, Any]] = None) -> List[AgentConfig]:
-    """Load agent library by auto-detecting from frontmatter."""
-    agents_dir = proj_dir / ".claude" / "agents" / "cc-native"
-    agents_data = aggregate_agents(agents_dir)
+    """Load agent library by auto-detecting from frontmatter.
+
+    Agents are loaded from _cc-native/agents/ directory. The markdown body
+    of each agent file becomes the system_prompt for --system-prompt invocation.
+    """
+    # aggregate_agents now defaults to _cc-native/agents/ relative to the script
+    agents_data = aggregate_agents()
 
     default_model = DEFAULT_AGENT_MODEL
     if settings:
@@ -415,7 +417,7 @@ def load_agent_library(proj_dir: Path, settings: Optional[Dict[str, Any]] = None
             enabled=a.get("enabled", True),
             categories=a.get("categories", ["code"]),
             description=a.get("description", ""),
-            tools=a.get("tools", ""),
+            system_prompt=a.get("system_prompt", ""),
         ))
 
     return agents
@@ -539,7 +541,6 @@ def main() -> int:
         enabled=orch_settings.get("enabled", True) and agent_review_enabled,
         model=orch_settings.get("model", "haiku"),
         timeout=orch_settings.get("timeout", 30),
-        max_turns=orch_settings.get("maxTurns", 3),
     )
 
     eprint(f"[cc-native-plan-review] Codex enabled: {codex_enabled}, Gemini enabled: {gemini_enabled}")
@@ -642,7 +643,6 @@ def main() -> int:
         # PHASE 3: Run selected agents in parallel
         if selected_agents:
             eprint("[cc-native-plan-review] === PHASE 3: Agent Reviews ===")
-            max_turns = agent_settings.get("maxTurns", 3)
             max_parallel = agent_settings.get("maxParallelAgents", 0)  # 0 = unlimited
             num_workers = len(selected_agents) if max_parallel <= 0 else min(max_parallel, len(selected_agents))
             eprint(f"[cc-native-plan-review] Launching {len(selected_agents)} agents in parallel (workers={num_workers})")
@@ -651,13 +651,12 @@ def main() -> int:
             debug_log(context_path, session_id, "hook", "agent_review_start", {
                 "agents": [a.name for a in selected_agents],
                 "timeout": timeout,
-                "max_turns": max_turns,
                 "complexity": detected_complexity,
             })
 
             with ThreadPoolExecutor(max_workers=num_workers) as executor:
                 futures = {
-                    executor.submit(run_agent_review, plan, agent, REVIEW_SCHEMA, timeout, max_turns, context_path, session_id): agent
+                    executor.submit(run_agent_review, plan, agent, REVIEW_SCHEMA, timeout, context_path, session_id): agent
                     for agent in selected_agents
                 }
                 for future in as_completed(futures):

package/dist/templates/cc-native/_cc-native/lib/orchestrator.py

@@ -129,7 +129,7 @@ def run_orchestrator(
     Args:
         plan: The plan content to analyze
         agent_library: List of available agents
-        config: Orchestrator configuration (model, timeout, max_turns)
+        config: Orchestrator configuration (model, timeout)
         settings: Agent review settings (agentSelection, complexityCategories)
 
     Returns:
@@ -154,9 +154,11 @@ def run_orchestrator(
 
     eprint(f"[orchestrator] Found Claude CLI at: {claude_path}")
 
-    # Build agent list for prompt
+    # Build agent list for prompt with rich descriptions
     agent_list = "\n".join([
-        f"- {a.name}: {a.focus} (categories: {', '.join(a.categories)})"
+        f"- {a.name} [{', '.join(a.categories)}]\n"
+        f"  Focus: {a.focus}\n"
+        f"  Expertise: {a.description}"
         for a in agent_library if a.enabled
     ])
     category_list = "/".join(categories)
@@ -164,43 +166,48 @@ def run_orchestrator(
     medium_range = f"{selection.get('medium', {}).get('min', 1)}-{selection.get('medium', {}).get('max', 2)}"
     high_range = f"{selection.get('high', {}).get('min', 2)}-{selection.get('high', {}).get('max', 4)}"
 
-    prompt = f"""IMPORTANT: Analyze this plan and output your decision immediately using StructuredOutput. Do NOT ask questions.
+    # System prompt with orchestrator instructions
+    system_prompt = """You are a plan orchestrator for code review. Your job is to analyze plans and select appropriate reviewer agents.
 
-You are a plan orchestrator. Analyze the plan below and determine:
-1. Complexity level (simple/medium/high)
-2. Category ({category_list})
-3. Which agents (if any) should review this plan
+You MUST call StructuredOutput immediately with your analysis. Do NOT ask questions or use any other tools.
+
+When selecting agents:
+- Match agent expertise to plan requirements
+- Consider what each agent specializes in
+- Only select agents whose categories match the plan category
+- Fewer agents for simple plans, more for complex plans"""
+
+    # User prompt with plan and agent list
+    prompt = f"""Analyze this plan and select appropriate reviewer agents.
 
 Available agents:
 {agent_list}
 
-Rules:
-- simple complexity = {simple_range} agents (CLI review sufficient)
+Selection rules:
+- simple complexity = {simple_range} agents
 - medium complexity = {medium_range} agents
 - high complexity = {high_range} agents
-- Only select agents whose categories match the plan category
+- Only select agents whose categories match the plan category ({category_list})
 - Non-technical plans (life, business) typically need 0 code-focused agents
 
-Analyze and call StructuredOutput with your decision now.
-
 PLAN:
 <<<
 {plan}
 >>>
-"""
+
+Call StructuredOutput now with: complexity, category, selectedAgents, reasoning"""
 
     schema_json = json.dumps(ORCHESTRATOR_SCHEMA, ensure_ascii=False)
 
     cmd_args = [
         claude_path,
         "-p",  # Enable print mode to read prompt from stdin
-        "--agent", "plan-orchestrator",
         "--model", config.model,
-        "--permission-mode", "bypassPermissions",
         "--output-format", "json",
-        "--max-turns", str(config.max_turns),
         "--json-schema", schema_json,
-        "--settings", "{}",
+        "--max-turns", "3",  # Single-turn with buffer for tool call + result
+        "--setting-sources", "",  # Disable PAI context interference
+        "--system-prompt", system_prompt,
     ]
 
     eprint(f"[orchestrator] Running with model: {config.model}, timeout: {config.timeout}s")
@@ -245,6 +252,17 @@ PLAN:
         eprint(f"[orchestrator] stderr: {p.stderr[:300]}")
 
     obj = _parse_claude_output(raw)
+
+    # Debug logging to diagnose empty selectedAgents issue
+    eprint(f"[orchestrator:debug] Raw output length: {len(raw)} chars")
+    if raw:
+        eprint(f"[orchestrator:debug] Raw output (first 500 chars): {raw[:500]}")
+    eprint(f"[orchestrator:debug] Parsed obj: {obj}")
+    if obj:
+        eprint(f"[orchestrator:debug] obj keys: {list(obj.keys())}")
+        eprint(f"[orchestrator:debug] selectedAgents value: {obj.get('selectedAgents', 'MISSING')}")
+        eprint(f"[orchestrator:debug] reasoning value: {obj.get('reasoning', 'MISSING')}")
+
     if not obj:
         eprint("[orchestrator] Failed to parse output, falling back to medium complexity")
         return OrchestratorResult(

package/dist/templates/cc-native/_cc-native/lib/reviewers/agent.py

@@ -81,7 +81,6 @@ def run_agent_review(
     agent: AgentConfig,
     schema: Dict[str, Any],
     timeout: int,
-    max_turns: int = 3,
     context_path: Optional[Path] = None,
     session_name: str = "unknown",
 ) -> ReviewerResult:
@@ -92,7 +91,6 @@ def run_agent_review(
         agent: Agent configuration (name, model, etc.)
         schema: JSON schema for the review output
         timeout: Timeout in seconds
-        max_turns: Maximum agent turns
         context_path: Optional path to context folder for debug logging
         session_name: Session name for debug logging
 
@@ -113,7 +111,8 @@ def run_agent_review(
 
     eprint(f"[{agent.name}] Found Claude CLI at: {claude_path}")
 
-    prompt = f"""{AGENT_REVIEW_PROMPT_PREFIX}
+    # User prompt contains just the task and plan
+    prompt = f"""Review the plan below and provide your assessment using StructuredOutput.
 
 PLAN:
 <<<
@@ -122,19 +121,23 @@ PLAN:
 """
 
     schema_json = json.dumps(schema, ensure_ascii=False)
+
+    # Build command args - use --system-prompt with the markdown body as persona
     cmd_args = [
         claude_path,
         "-p",  # Enable print mode to read prompt from stdin
-        "--agent", agent.name,
         "--model", agent.model,
-        "--permission-mode", "plan",
         "--output-format", "json",
-        "--max-turns", str(max_turns),
         "--json-schema", schema_json,
-        "--settings", "{}",
+        "--max-turns", "3",  # Allow buffer for tool call + result (usually completes in 2)
+        "--setting-sources", "",  # Disable user/project settings to avoid PAI context interference
     ]
 
-    eprint(f"[{agent.name}] Running with model: {agent.model}, timeout: {timeout}s, max-turns: {max_turns}")
+    # Add system prompt if available (the markdown body with full persona)
+    if agent.system_prompt:
+        cmd_args.extend(["--system-prompt", agent.system_prompt])
+
+    eprint(f"[{agent.name}] Running with model: {agent.model}, timeout: {timeout}s")
 
     # Get environment for internal subprocess (bypasses hooks)
     env = get_internal_subprocess_env()
@@ -176,7 +179,6 @@ PLAN:
             "stderr_len": len(err),
             "model": agent.model,
             "timeout": timeout,
-            "max_turns": max_turns,
         })
 
     if raw:

package/dist/templates/cc-native/_cc-native/lib/reviewers/base.py

@@ -39,7 +39,7 @@ class AgentConfig:
     enabled: bool = True
     categories: List[str] = field(default_factory=lambda: ["code"])
     description: str = ""
-    tools: str = ""
+    system_prompt: str = ""  # Markdown body content for --system-prompt
 
 
 @dataclass
@@ -48,7 +48,6 @@ class OrchestratorConfig:
     enabled: bool = True
     model: str = "haiku"
     timeout: int = 30
-    max_turns: int = 3
 
 
 # ---------------------------

package/dist/templates/cc-native/_cc-native/plan-review.config.json

@@ -22,14 +22,12 @@
   },
   "agentReview": {
     "enabled": true,
-    "timeout": 120,
+    "timeout": 180,
     "blockOnFail": true,
-    "maxTurns": 2,
     "orchestrator": {
       "enabled": true,
       "model": "haiku",
-      "timeout": 30,
-      "maxTurns": 6
+      "timeout": 30
     },
     "legacyMode": false,
     "mandatoryAgents": ["handoff-readiness", "clarity-auditor", "skeptic"],

package/dist/templates/cc-native/_cc-native/scripts/aggregate_agents.py

@@ -93,16 +93,31 @@ def extract_frontmatter(content: str) -> dict[str, Any] | None:
         return None
 
 
+def extract_body(content: str) -> str:
+    """Extract markdown body (after frontmatter) from content.
+
+    Args:
+        content: Raw markdown file content
+
+    Returns:
+        Markdown body content (everything after the closing ---)
+    """
+    match = re.match(r'^---\s*\n.*?\n---\s*\n(.*)$', content, re.DOTALL)
+    if match:
+        return match.group(1).strip()
+    return content.strip()
+
+
 def aggregate_agents(agents_dir: Path | None = None) -> list[dict[str, Any]]:
     """Read all agent files and return aggregated metadata.
-    
+
     Scans the agents directory for .md files, extracts frontmatter from each,
     and returns a list of agent configurations.
-    
+
     Args:
         agents_dir: Path to agents directory. If None, uses default
-                   .claude/agents relative to project root.
-                   
+                   _cc-native/agents/ relative to this script.
+
     Returns:
         List of agent configuration dicts with fields:
         - name: Agent identifier (lowercase)
@@ -111,14 +126,14 @@ def aggregate_agents(agents_dir: Path | None = None) -> list[dict[str, Any]]:
         - focus: Brief focus for orchestrator
         - enabled: Whether agent is available
         - categories: Work categories for filtering
-        - tools: Available tools
+        - system_prompt: Full markdown body (persona content)
     """
     if agents_dir is None:
-        # Default to .claude/agents/cc-native relative to this script's location
+        # Default to _cc-native/agents/ relative to this script's location
         # Script is at: _cc-native/scripts/aggregate_agents.py
-        # Agents are at: .claude/agents/cc-native/
+        # Agents are at: _cc-native/agents/
         script_dir = Path(__file__).parent
-        agents_dir = script_dir.parent.parent / ".claude" / "agents" / "cc-native"
+        agents_dir = script_dir.parent / "agents"
 
     agents = []
 
@@ -135,6 +150,8 @@ def aggregate_agents(agents_dir: Path | None = None) -> list[dict[str, Any]]:
                     frontmatter["categories"] = ["code"]
                 elif isinstance(frontmatter["categories"], str):
                     frontmatter["categories"] = [frontmatter["categories"]]
+                # Extract markdown body as system_prompt
+                frontmatter["system_prompt"] = extract_body(content)
                 agents.append(frontmatter)
         except Exception:
             # Skip files that can't be read or parsed

package/oclif.manifest.json

@@ -401,5 +401,5 @@
       ]
     }
   },
-  "version": "0.9.4"
+  "version": "0.9.6"
 }

package/package.json

@@ -1,7 +1,7 @@
 {
   "name": "aiwcli",
   "description": "AI Workflow CLI - Command-line interface for AI-powered workflows",
-  "version": "0.9.4",
+  "version": "0.9.6",
   "author": "jofu-tofu",
   "bin": {
     "aiw": "./bin/run.js"