zwarm 3.2.1__py3-none-any.whl → 3.6.0__py3-none-any.whl

zwarm/cli/main.py

@@ -122,16 +122,6 @@ Manage zwarm configurations.
 app.add_typer(configs_app, name="configs")
 
 
-class AdapterType(str, Enum):
-    codex_mcp = "codex_mcp"
-    claude_code = "claude_code"
-
-
-class ModeType(str, Enum):
-    sync = "sync"
-    async_ = "async"
-
-
 @app.command()
 def orchestrate(
     task: Annotated[Optional[str], typer.Option("--task", "-t", help="The task to accomplish")] = None,
@@ -228,6 +218,26 @@ def orchestrate(
         if orchestrator.instance_id and not instance:
             console.print(f"  [dim]Instance: {orchestrator.instance_id[:8]}[/]")
 
+        # Set up step callback for live progress display
+        def step_callback(step_num: int, tool_results: list) -> None:
+            """Print tool calls and results as they happen."""
+            if not tool_results:
+                return
+            for tool_info, result in tool_results:
+                name = tool_info.get("name", "?")
+                # Truncate args for display
+                args_str = str(tool_info.get("args", {}))
+                if len(args_str) > 80:
+                    args_str = args_str[:77] + "..."
+                # Truncate result for display
+                result_str = str(result)
+                if len(result_str) > 100:
+                    result_str = result_str[:97] + "..."
+                console.print(f"[dim]step {step_num}[/] → [cyan]{name}[/]({args_str})")
+                console.print(f"         └ {result_str}")
+
+        orchestrator._step_callback = step_callback
+
         # Run the orchestrator loop
         console.print("[bold]--- Orchestrator running ---[/]\n")
         result = orchestrator.run(task=task)
@@ -289,7 +299,8 @@ def pilot(
     config: Annotated[Optional[Path], typer.Option("--config", "-c", help="Path to config YAML")] = None,
     overrides: Annotated[Optional[list[str]], typer.Option("--set", help="Override config (key=value)")] = None,
     working_dir: Annotated[Path, typer.Option("--working-dir", "-w", help="Working directory")] = Path("."),
-    instance: Annotated[Optional[str], typer.Option("--instance", "-i", help="Instance ID (for isolation)")] = None,
+    resume: Annotated[bool, typer.Option("--resume", help="Resume from previous state")] = False,
+    instance: Annotated[Optional[str], typer.Option("--instance", "-i", help="Instance ID (for isolation/resume)")] = None,
     instance_name: Annotated[Optional[str], typer.Option("--name", "-n", help="Human-readable instance name")] = None,
     model: Annotated[PilotLM, typer.Option("--model", "-m", help="LM to use")] = PilotLM.gpt5_verbose,
 ):
@@ -331,19 +342,30 @@ def pilot(
 
         [dim]# Named instance[/]
         $ zwarm pilot --name my-feature
+
+        [dim]# Resume a previous session[/]
+        $ zwarm pilot --resume --instance abc123
     """
     from zwarm.cli.pilot import run_pilot, build_pilot_orchestrator
 
     # Resolve task (optional for pilot)
     resolved_task = _resolve_task(task, task_file)
 
-    console.print(f"[bold]Starting pilot session...[/]")
+    # Validate resume requirements
+    if resume and not instance:
+        console.print("[red]Error:[/] --resume requires --instance to specify which session to resume")
+        console.print("  [dim]Use 'zwarm instances' to list available instances[/]")
+        raise typer.Exit(1)
+
+    console.print(f"[bold]{'Resuming' if resume else 'Starting'} pilot session...[/]")
     console.print(f"  Working dir: {working_dir.absolute()}")
     console.print(f"  Model: {model.value}")
     if resolved_task:
         console.print(f"  Initial task: {resolved_task[:60]}...")
     if instance:
         console.print(f"  Instance: {instance}" + (f" ({instance_name})" if instance_name else ""))
+    if resume:
+        console.print(f"  [yellow]Resuming from saved state...[/]")
     console.print()
 
     orchestrator = None
@@ -361,6 +383,12 @@ def pilot(
         if orchestrator.instance_id and not instance:
             console.print(f"  [dim]Instance: {orchestrator.instance_id[:8]}[/]")
 
+        # Resume from saved state if requested
+        if resume:
+            orchestrator.load_state()
+            msg_count = len(orchestrator.messages)
+            console.print(f"  [green]✓[/] Resumed with {msg_count} messages")
+
         # Run the pilot REPL
         run_pilot(orchestrator, initial_task=resolved_task)
 
@@ -384,78 +412,68 @@ def pilot(
 @app.command()
 def exec(
     task: Annotated[str, typer.Option("--task", "-t", help="Task to execute")],
-    adapter: Annotated[AdapterType, typer.Option("--adapter", "-a", help="Executor adapter")] = AdapterType.codex_mcp,
-    mode: Annotated[ModeType, typer.Option("--mode", "-m", help="Execution mode")] = ModeType.sync,
     working_dir: Annotated[Path, typer.Option("--working-dir", "-w", help="Working directory")] = Path("."),
     model: Annotated[Optional[str], typer.Option("--model", help="Model override")] = None,
+    wait: Annotated[bool, typer.Option("--wait", help="Wait for completion and show result")] = False,
 ):
     """
-    Run a single executor directly (for testing).
+    Run a single Codex session directly (for testing).
 
-    Useful for testing adapters without the full orchestrator loop.
+    Spawns a session using CodexSessionManager - same as interactive/pilot.
+    Web search is always enabled via .codex/config.toml (set up by `zwarm init`).
 
     [bold]Examples:[/]
-        [dim]# Test Codex[/]
-        $ zwarm exec --task "What is 2+2?"
+        [dim]# Quick test[/]
+        $ zwarm exec --task "What is 2+2?" --wait
 
-        [dim]# Test Claude Code[/]
-        $ zwarm exec -a claude_code --task "List files in current dir"
+        [dim]# Run in background[/]
+        $ zwarm exec --task "Build feature"
 
-        [dim]# Async mode[/]
-        $ zwarm exec --task "Build feature" --mode async
+        [dim]# Web search is always available[/]
+        $ zwarm exec --task "Find latest FastAPI docs" --wait
     """
-    from zwarm.adapters import get_adapter
-
-    console.print(f"[bold]Running executor directly...[/]")
-    console.print(f"  Adapter: [cyan]{adapter.value}[/]")
-    console.print(f"  Mode: {mode.value}")
-    console.print(f"  Task: {task}")
-
-    # Use isolated codex config if available
-    config_path = working_dir / ".zwarm" / "codex.toml"
-    if not config_path.exists():
-        config_path = None
-
-    try:
-        executor = get_adapter(adapter.value, model=model, config_path=config_path)
-    except ValueError as e:
-        console.print(f"[red]Error:[/] {e}")
-        sys.exit(1)
-
-    async def run():
-        try:
-            session = await executor.start_session(
-                task=task,
-                working_dir=working_dir.absolute(),
-                mode=mode.value,
-                model=model,
-            )
-
-            console.print(f"\n[green]Session started:[/] {session.id[:8]}")
-
-            if mode == ModeType.sync:
-                response = session.messages[-1].content if session.messages else "(no response)"
-                console.print(f"\n[bold]Response:[/]\n{response}")
+    import time
+    from zwarm.sessions import CodexSessionManager, SessionStatus
 
-                # Interactive loop for sync mode
-                while True:
-                    try:
-                        user_input = console.input("\n[dim]> (type message or 'exit')[/] ")
-                        if user_input.lower() == "exit" or not user_input:
-                            break
+    console.print(f"[bold]Running Codex session...[/]")
+    console.print(f"  Task: {task[:60]}{'...' if len(task) > 60 else ''}")
+    if model:
+        console.print(f"  Model: {model}")
 
-                        response = await executor.send_message(session, user_input)
-                        console.print(f"\n[bold]Response:[/]\n{response}")
-                    except KeyboardInterrupt:
-                        break
-            else:
-                console.print("[dim]Async mode - session running in background.[/]")
-                console.print("Use 'zwarm status' to check progress.")
+    manager = CodexSessionManager(working_dir / ".zwarm")
+    effective_model = model or "gpt-5.1-codex-mini"
 
-        finally:
-            await executor.cleanup()
+    session = manager.start_session(
+        task=task,
+        working_dir=working_dir.absolute(),
+        model=effective_model,
+    )
 
-    asyncio.run(run())
+    console.print(f"\n[green]Session started:[/] {session.short_id}")
+
+    if wait:
+        console.print("[dim]Waiting for completion...[/]")
+        while True:
+            time.sleep(2)
+            session = manager.get_session(session.id)
+            if session.status != SessionStatus.RUNNING:
+                break
+
+        if session.status == SessionStatus.COMPLETED:
+            console.print(f"\n[green]✓ Completed[/]")
+            # Show last assistant message
+            for msg in reversed(session.messages):
+                if msg.role == "assistant":
+                    console.print(f"\n[bold]Response:[/]\n{msg.content}")
+                    break
+        else:
+            console.print(f"\n[red]Status:[/] {session.status.value}")
+            if session.error:
+                console.print(f"[red]Error:[/] {session.error}")
+    else:
+        console.print("[dim]Running in background. Check with:[/]")
+        console.print(f"  zwarm sessions")
+        console.print(f"  zwarm session show {session.short_id}")
 
 
 @app.command()
@@ -740,12 +758,12 @@ def init(
     [bold]Creates:[/]
         [cyan].zwarm/[/]              State directory for sessions and events
         [cyan].zwarm/config.toml[/]   Runtime settings (weave, adapter, watchers)
-        [cyan].zwarm/codex.toml[/]    Codex CLI settings (model, reasoning effort)
+        [cyan].zwarm/codex.toml[/]    Codex CLI settings (model, web search, etc.)
         [cyan]zwarm.yaml[/]           Project config (optional, with --with-project)
 
     [bold]Configuration relationship:[/]
         config.toml   → Controls zwarm itself (tracing, which watchers run)
-        codex.toml    → Controls the Codex CLI that runs executor sessions
+        codex.toml    → Codex settings, parsed by zwarm and passed via -c overrides
         zwarm.yaml    → Project-specific context injected into orchestrator
 
     [bold]Examples:[/]
@@ -917,6 +935,23 @@ def init(
         codex_toml_path.write_text(codex_content)
         console.print(f"  [green]✓[/] Created .zwarm/codex.toml")
 
+    # Create claude.toml for isolated Claude Code configuration
+    claude_toml_path = state_dir / "claude.toml"
+    write_claude_toml = True
+    if claude_toml_path.exists():
+        if not non_interactive:
+            overwrite_claude = typer.confirm("  .zwarm/claude.toml exists. Overwrite?", default=False)
+            if not overwrite_claude:
+                write_claude_toml = False
+                console.print("  [dim]Skipping claude.toml[/]")
+        else:
+            write_claude_toml = False  # Don't overwrite in non-interactive mode
+
+    if write_claude_toml:
+        claude_content = _generate_claude_toml(model="sonnet")
+        claude_toml_path.write_text(claude_content)
+        console.print(f"  [green]✓[/] Created .zwarm/claude.toml")
+
     # Create zwarm.yaml
     if create_project_config:
         if zwarm_yaml_path.exists() and not non_interactive:
@@ -939,7 +974,7 @@ def init(
     # Explain config files
     console.print("[bold]Configuration files:[/]")
     console.print("  [cyan].zwarm/config.toml[/]  - Runtime settings (Weave tracing, watchers)")
-    console.print("  [cyan].zwarm/codex.toml[/]   - Codex CLI settings (model, reasoning effort)")
+    console.print("  [cyan].zwarm/codex.toml[/]   - Codex CLI settings (model, web search, sandbox)")
     if create_project_config:
         console.print("  [cyan]zwarm.yaml[/]          - Project context and constraints")
     console.print()
@@ -959,40 +994,87 @@ def _generate_config_toml(
     adapter: str = "codex_mcp",
     watchers: list[str] | None = None,
 ) -> str:
-    """Generate config.toml content."""
+    """Generate config.toml content with all options at their defaults."""
     watchers = watchers or []
 
     lines = [
         "# zwarm configuration",
         "# Generated by 'zwarm init'",
+        "# All values shown are defaults - uncomment and modify as needed",
         "",
+        "# ============================================================================",
+        "# Weave Integration (optional tracing/observability)",
+        "# ============================================================================",
         "[weave]",
     ]
 
     if weave_project:
         lines.append(f'project = "{weave_project}"')
     else:
-        lines.append("# project = \"your-entity/your-project\"  # Uncomment to enable Weave tracing")
+        lines.append('# project = "your-entity/your-project"  # Uncomment to enable Weave tracing')
 
     lines.extend([
+        "enabled = true",
         "",
+        "# ============================================================================",
+        "# Orchestrator Settings",
+        "# ============================================================================",
         "[orchestrator]",
-        "max_steps = 50",
+        '# lm = "gpt-5-mini"                    # LLM for orchestrator (gpt-5-mini, gpt-5, claude-sonnet-4)',
+        "max_steps = 50                         # Max steps for orchestrate command",
+        "max_steps_per_turn = 60                # Max steps per turn in pilot mode",
+        "parallel_delegations = 4               # Max concurrent delegations",
+        '# prompt = "path/to/prompt.yaml"       # Custom prompt file (optional)',
+        '# allowed_dirs = ["*"]                 # Directories agent can delegate to (default: working_dir only)',
+        "",
+        "# Context window compaction (prevents overflow on long tasks)",
+        "[orchestrator.compaction]",
+        "enabled = true",
+        "max_tokens = 100000                    # Trigger compaction above this",
+        "threshold_pct = 0.85                   # Compact when at this % of max_tokens",
+        "target_pct = 0.7                       # Target this % after compaction",
+        "keep_first_n = 2                       # Always keep first N messages (system + task)",
+        "keep_last_n = 10                       # Always keep last N messages (recent context)",
         "",
+        "# ============================================================================",
+        "# Executor Settings (codex agent configuration)",
+        "# ============================================================================",
         "[executor]",
-        f'adapter = "{adapter}"',
-        "# model = \"\"  # Optional model override",
+        f'adapter = "{adapter}"                  # codex_mcp | codex_exec | claude_code',
+        '# model = "gpt-5.1-codex-mini"         # Model for delegated sessions (uses codex.toml default if not set)',
+        'sandbox = "workspace-write"            # read-only | workspace-write | danger-full-access',
+        "timeout = 3600                         # Session timeout in seconds",
+        'reasoning_effort = "high"              # low | medium | high',
         "",
+        "# ============================================================================",
+        "# Watchers (automated monitoring and nudges)",
+        "# ============================================================================",
         "[watchers]",
-        f"enabled = {watchers}",
+        f"enabled = {str(bool(watchers)).lower()}",
+        'message_role = "user"                  # Role for nudge messages: user | assistant | system',
         "",
-        "# Watcher-specific configuration",
-        "# [watchers.budget]",
-        "# max_steps = 50",
+        "# Default watchers: progress, budget, delegation_reminder",
+        "# Uncomment below to customize:",
+        "",
+        "# [[watchers.watchers]]",
+        '# name = "progress"',
+        "# enabled = true",
+        "",
+        "# [[watchers.watchers]]",
+        '# name = "budget"',
+        "# enabled = true",
+        "# [watchers.watchers.config]",
+        "# max_sessions = 10",
         "# warn_at_percent = 80",
         "",
-        "# [watchers.pattern]",
-        "# patterns = [\"DROP TABLE\", \"rm -rf\"]",
+        "# [[watchers.watchers]]",
+        '# name = "delegation_reminder"',
+        "# enabled = true",
+        "",
+        "# ============================================================================",
+        "# State Directory",
+        "# ============================================================================",
+        '# state_dir = ".zwarm"                 # Where to store session data',
         "",
     ])
 
@@ -1006,28 +1088,73 @@ def _generate_codex_toml(
     """
     Generate codex.toml for isolated codex configuration.
 
-    This file is used by zwarm instead of ~/.codex/config.toml to ensure
-    consistent behavior across different environments.
+    This file is parsed by zwarm and settings are passed to codex via -c overrides.
+    Each .zwarm directory has its own codex config, independent of ~/.codex/config.toml.
     """
     lines = [
         "# Codex configuration for zwarm",
-        "# This file isolates zwarm's codex settings from your global ~/.codex/config.toml",
+        "# zwarm parses this file and passes settings to codex via -c overrides",
+        "# Each .zwarm dir has its own config, independent of ~/.codex/config.toml",
         "# Generated by 'zwarm init'",
         "",
         "# Model settings",
         f'model = "{model}"',
         f'model_reasoning_effort = "{reasoning_effort}"  # low | medium | high',
         "",
-        "# Approval settings - zwarm manages these automatically",
-        "# disable_response_storage = false",
+        "# DANGER MODE - bypasses all safety controls",
+        "# Set to true to use --dangerously-bypass-approvals-and-sandbox",
+        "full_danger = true",
+        "",
+        "# Web search - enables web_search tool for agents",
+        "[features]",
+        "web_search_request = true",
         "",
-        "# You can override any codex setting here",
+        "# Sandbox settings - network access required for web search",
+        "[sandbox_workspace_write]",
+        "network_access = true",
+        "",
+        "# Approval policy - 'never' means no human approval needed",
+        "# approval_policy = \"never\"",
+        "",
+        "# You can add any codex config key here",
         "# See: https://github.com/openai/codex#configuration",
         "",
     ]
     return "\n".join(lines)
 
 
+def _generate_claude_toml(
+    model: str = "sonnet",
+) -> str:
+    """
+    Generate claude.toml for isolated Claude Code configuration.
+
+    This file is parsed by zwarm and settings are passed to claude via CLI flags.
+    Each .zwarm directory has its own claude config.
+    """
+    lines = [
+        "# Claude Code configuration for zwarm",
+        "# zwarm parses this file and passes settings to claude via CLI flags",
+        "# Each .zwarm dir has its own config",
+        "# Generated by 'zwarm init'",
+        "",
+        "# Model settings",
+        f'model = "{model}"  # sonnet | opus | haiku',
+        "",
+        "# DANGER MODE - bypasses all permission checks",
+        "# Set to true to use --dangerously-skip-permissions",
+        "full_danger = true",
+        "",
+        "# Note: Claude Code uses different CLI flags than Codex",
+        "# Common options:",
+        "#   --model <model>         Model to use (sonnet, opus, haiku)",
+        "#   --add-dir <path>        Additional directories to allow",
+        "#   --allowed-tools <tools> Restrict available tools",
+        "",
+    ]
+    return "\n".join(lines)
+
+
 def _generate_zwarm_yaml(
     description: str = "",
     context: str = "",
@@ -1534,6 +1661,7 @@ def session_start(
     Start a new Codex session in the background.
 
     The session runs independently and you can check on it later.
+    Web search is always enabled via .codex/config.toml (set up by `zwarm init`).
 
     [bold]Examples:[/]
         [dim]# Simple task[/]
@@ -1541,6 +1669,9 @@ def session_start(
 
         [dim]# With specific model[/]
         $ zwarm session start "Refactor the API" --model gpt-5.1-codex-max
+
+        [dim]# Web search is always available[/]
+        $ zwarm session start "Research latest OAuth2 best practices"
     """
     from zwarm.sessions import CodexSessionManager