codex-flow 2.1.13__tar.gz → 2.1.14__tar.gz

{codex_flow-2.1.13/codex_flow.egg-info → codex_flow-2.1.14}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: codex-flow
-Version: 2.1.13
+Version: 2.1.14
 Summary: Intelligent, Efficient, Adaptive Multi-Agent Strategy Orchestration for Codex
 Home-page: https://github.com/ParsifalC/codex-flow
 Author: Parsifal
@@ -45,7 +45,7 @@ Dynamic: requires-python
 
 **智能、高效、自适应的 Codex 多 Agent 策略编排引擎**
 
-[![Version](https://img.shields.io/badge/version-2.1.13-blue.svg?style=flat-square)](VERSION)
+[![Version](https://img.shields.io/badge/version-2.1.14-blue.svg?style=flat-square)](VERSION)
 [![Homebrew](https://img.shields.io/badge/Homebrew-ParsifalC%2Ftap-orange.svg?style=flat-square&logo=homebrew)](https://github.com/ParsifalC/homebrew-tap)
 [![PyPI](https://img.shields.io/pypi/v/codex-flow.svg?style=flat-square&logo=pypi)](https://pypi.org/project/codex-flow/)
 [![Platform](https://img.shields.io/badge/platform-macOS%20%7C%20Linux%20%7C%20Windows-brightgreen.svg?style=flat-square)](#-快速安装)
@@ -256,7 +256,7 @@ codex-flow
 
 ```text
 ╭────────────────────────────────────────────────────────────────────╮
-│                  🚀 codex-flow 控制台 (v2.1.13)                  │
+│                  🚀 codex-flow 控制台 (v2.1.14)                  │
 │    FlowPilot 智能编排 · 确定性任务遥测 · 本地 Benchmark 验证     │
 ╰────────────────────────────────────────────────────────────────────╯
   [1] 🪟 macOS 原生悬浮窗 (overlay widget)

{codex_flow-2.1.13 → codex_flow-2.1.14}/README.en.md

@@ -8,7 +8,7 @@
 
 **Intelligent, Efficient, Adaptive Multi-Agent Strategy Orchestration for Codex**
 
-[![Version](https://img.shields.io/badge/version-2.1.13-blue.svg?style=flat-square)](VERSION)
+[![Version](https://img.shields.io/badge/version-2.1.14-blue.svg?style=flat-square)](VERSION)
 [![Homebrew](https://img.shields.io/badge/Homebrew-ParsifalC%2Ftap-orange.svg?style=flat-square&logo=homebrew)](https://github.com/ParsifalC/homebrew-tap)
 [![PyPI](https://img.shields.io/pypi/v/codex-flow.svg?style=flat-square&logo=pypi)](https://pypi.org/project/codex-flow/)
 [![Platform](https://img.shields.io/badge/platform-macOS%20%7C%20Linux%20%7C%20Windows-brightgreen.svg?style=flat-square)](#-quick-start)
@@ -217,7 +217,7 @@ The latest console also integrates Overlay build/start management, effective pol
 
 ```text
 ╭────────────────────────────────────────────────────────────────────╮
-│                  🚀 codex-flow Console (v2.1.13)                   │
+│                  🚀 codex-flow Console (v2.1.14)                   │
 │   FlowPilot orchestration · deterministic telemetry · validation   │
 ╰────────────────────────────────────────────────────────────────────╯
   [1] 🪟 macOS native floating widget (overlay widget)

{codex_flow-2.1.13 → codex_flow-2.1.14}/README.md

@@ -8,7 +8,7 @@
 
 **智能、高效、自适应的 Codex 多 Agent 策略编排引擎**
 
-[![Version](https://img.shields.io/badge/version-2.1.13-blue.svg?style=flat-square)](VERSION)
+[![Version](https://img.shields.io/badge/version-2.1.14-blue.svg?style=flat-square)](VERSION)
 [![Homebrew](https://img.shields.io/badge/Homebrew-ParsifalC%2Ftap-orange.svg?style=flat-square&logo=homebrew)](https://github.com/ParsifalC/homebrew-tap)
 [![PyPI](https://img.shields.io/pypi/v/codex-flow.svg?style=flat-square&logo=pypi)](https://pypi.org/project/codex-flow/)
 [![Platform](https://img.shields.io/badge/platform-macOS%20%7C%20Linux%20%7C%20Windows-brightgreen.svg?style=flat-square)](#-快速安装)
@@ -219,7 +219,7 @@ codex-flow
 
 ```text
 ╭────────────────────────────────────────────────────────────────────╮
-│                  🚀 codex-flow 控制台 (v2.1.13)                  │
+│                  🚀 codex-flow 控制台 (v2.1.14)                  │
 │    FlowPilot 智能编排 · 确定性任务遥测 · 本地 Benchmark 验证     │
 ╰────────────────────────────────────────────────────────────────────╯
   [1] 🪟 macOS 原生悬浮窗 (overlay widget)

codex_flow-2.1.14/VERSION

@@ -0,0 +1 @@
+2.1.14

{codex_flow-2.1.13 → codex_flow-2.1.14}/apps/chatgpt-mcp/server.py

@@ -25,7 +25,7 @@ else:
 
 JSONRPC_VERSION = "2.0"
 SERVER_NAME = "chatgpt-codex-flow"
-SERVER_VERSION = "2.1.13"
+SERVER_VERSION = "2.1.14"
 MAX_REQUEST_BYTES = 1024 * 1024
 
 

{codex_flow-2.1.13 → codex_flow-2.1.14}/apps/macos-overlay/Sources/Controllers/OverlayWindowController.swift

@@ -11,7 +11,13 @@ public enum DockEdge: String, Codable {
 // MARK: - Shared Observable State
 public class OverlayState: ObservableObject {
     @Published public var isExpanded: Bool = false
-    @Published public var isPinned: Bool = false
+    @Published public var isPinned: Bool = false {
+        didSet {
+            if isPinned {
+                windowController?.cancelNotificationAutoCollapseTimer()
+            }
+        }
+    }
     @Published public var isTaskRunning: Bool = false
     @Published public var isDocked: Bool = false
     @Published public var dockEdge: DockEdge = .right
@@ -55,14 +61,24 @@ public class OverlayState: ObservableObject {
         }
     }
 
-    public func expand() {
-        guard !isExpanded else { return }
+    public func expand(notificationTriggered: Bool = false) {
+        if isExpanded {
+            if notificationTriggered {
+                DispatchQueue.main.async {
+                    self.windowController?.scheduleNotificationAutoCollapse()
+                }
+            }
+            return
+        }
         loadMenuData()
         DispatchQueue.main.async {
             self.windowController?.prepareForPresentationChange()
             self.isDocked = false
             self.isExpanded = true
             self.windowController?.updateWindowFrame(animated: true)
+            if notificationTriggered {
+                self.windowController?.scheduleNotificationAutoCollapse()
+            }
         }
     }
 
@@ -354,6 +370,7 @@ class TrackingHostingView<Content: View>: NSHostingView<Content> {
         isDragging = false
         windowController.cancelDwellTimer()
         windowController.cancelTuckTimer()
+        windowController.cancelNotificationAutoCollapseTimer()
         super.mouseDown(with: event)
     }
 
@@ -529,6 +546,8 @@ public class OverlayWindowController: NSObject, NSWindowDelegate {
 
     private var hoverDwellTimer: Timer?
     private var collapseTimer: Timer?
+    private var notificationCollapseTimer: Timer?
+    private let notificationAutoCollapseDuration: TimeInterval = 5.0
     private var tuckTimer: Timer?
     private let edgeTuckIdleInterval: TimeInterval = 30.0
 
@@ -612,6 +631,7 @@ public class OverlayWindowController: NSObject, NSWindowDelegate {
     public func prepareForPresentationChange() {
         cancelDwellTimer()
         cancelTuckTimer()
+        cancelNotificationAutoCollapseTimer()
         collapseTimer?.invalidate()
         collapseTimer = nil
     }
@@ -719,6 +739,9 @@ public class OverlayWindowController: NSObject, NSWindowDelegate {
             return true
         }
         reconcilePointerAfterGeometryIfNeeded()
+        if state.isExpanded && isPointerInAppWindowOrPopover() {
+            cancelNotificationAutoCollapseTimer()
+        }
         return false
     }
 
@@ -813,6 +836,7 @@ public class OverlayWindowController: NSObject, NSWindowDelegate {
     }
 
     public func handleMouseEntered(at point: NSPoint) {
+        cancelNotificationAutoCollapseTimer()
         if isGeometryTransitioning {
             needsPointerReconciliationAfterGeometry = true
             return
@@ -833,6 +857,7 @@ public class OverlayWindowController: NSObject, NSWindowDelegate {
     }
 
     public func handleMouseMoved(at point: NSPoint) {
+        cancelNotificationAutoCollapseTimer()
         if isGeometryTransitioning {
             needsPointerReconciliationAfterGeometry = true
             return
@@ -892,7 +917,41 @@ public class OverlayWindowController: NSObject, NSWindowDelegate {
         }
     }
 
-    private func isPointerInAppWindowOrPopover() -> Bool {
+    public func scheduleNotificationAutoCollapse(duration: TimeInterval? = nil) {
+        cancelNotificationAutoCollapseTimer()
+        let interval = duration ?? notificationAutoCollapseDuration
+        guard state.isExpanded,
+              !state.isPinned,
+              !isInteractingOrDragging else { return }
+
+        if isPointerInAppWindowOrPopover() {
+            return
+        }
+
+        let timer = Timer(timeInterval: interval, repeats: false) { [weak self] _ in
+            guard let self,
+                  self.state.isExpanded,
+                  !self.state.isPinned,
+                  !self.isInteractingOrDragging,
+                  !self.isGeometryTransitioning else { return }
+
+            if self.isPointerInAppWindowOrPopover() {
+                self.cancelNotificationAutoCollapseTimer()
+                return
+            }
+
+            self.state.collapse()
+        }
+        RunLoop.main.add(timer, forMode: .common)
+        notificationCollapseTimer = timer
+    }
+
+    public func cancelNotificationAutoCollapseTimer() {
+        notificationCollapseTimer?.invalidate()
+        notificationCollapseTimer = nil
+    }
+
+    public func isPointerInAppWindowOrPopover() -> Bool {
         let mouseLocation = NSEvent.mouseLocation
         for appWindow in NSApp.windows {
             guard appWindow.isVisible,

{codex_flow-2.1.13 → codex_flow-2.1.14}/apps/macos-overlay/Sources/Services/IPCServer.swift

@@ -196,23 +196,23 @@ public class IPCService {
                        var run = try? JSONDecoder().decode(TaskRun.self, from: data) {
                         TelemetryQueryEngine.shared.enrichRunIfNeeded(&run)
                         state.update(run: run)
-                        state.expand()
+                        state.expand(notificationTriggered: true)
                         return "{\"ok\": true, \"updatedFrom\": \"last.json\"}\n"
                     }
-                    state.expand()
+                    state.expand(notificationTriggered: true)
                     return "{\"ok\": true, \"action\": \"expanded\"}\n"
                 } else {
                     if let fileData = try? Data(contentsOf: URL(fileURLWithPath: payload)),
                        var run = try? JSONDecoder().decode(TaskRun.self, from: fileData) {
                         TelemetryQueryEngine.shared.enrichRunIfNeeded(&run)
                         state.update(run: run)
-                        state.expand()
+                        state.expand(notificationTriggered: true)
                         return "{\"ok\": true, \"updatedFrom\": \"file\"}\n"
                     } else if let json = payload.data(using: .utf8),
                               var run = try? JSONDecoder().decode(TaskRun.self, from: json) {
                         TelemetryQueryEngine.shared.enrichRunIfNeeded(&run)
                         state.update(run: run)
-                        state.expand()
+                        state.expand(notificationTriggered: true)
                         return "{\"ok\": true, \"updatedFrom\": \"json\"}\n"
                     }
                 }

codex_flow-2.1.14/benchmark/prices/gpt-6-astra-2026-09-12.json

@@ -0,0 +1,22 @@
+{
+  "gpt-5.6-luna": {
+    "input": 0.2,
+    "cached_input": 0.02,
+    "output": 1.2
+  },
+  "gpt-5.6-terra": {
+    "input": 2.0,
+    "cached_input": 0.2,
+    "output": 12.0
+  },
+  "gpt-5.6-sol": {
+    "input": 4.0,
+    "cached_input": 0.4,
+    "output": 20.0
+  },
+  "gpt-6-astra": {
+    "input": 10.0,
+    "cached_input": 1.0,
+    "output": 50.0
+  }
+}

{codex_flow-2.1.13 → codex_flow-2.1.14}/benchmark/profiles.json

@@ -31,7 +31,7 @@
       ]
     },
     "full": {
-      "description": "Three repetitions across three high direct models, fixed-high flow, and adaptive-effort flow.",
+      "description": "Three repetitions across three high direct models, fixed-high flow, and adaptive-effort flow. Includes Astra/high parent with the real efficient FlowPilot runtime.",
       "repetitions": 3,
       "matrix": [
         {"id": "luna-direct", "strategy": "direct", "model": "gpt-5.6-luna", "reasoning_effort": "high"},
@@ -56,11 +56,26 @@
             "model": "gpt-5.6-luna",
             "reasoning_effort": {"routine": "high", "complex": "xhigh", "critical": "max"}
           }
+        },
+        {
+          "id": "codex-flow-runtime-astra-high",
+          "strategy": "runtime",
+          "reasoning_policy": "fixed",
+          "profile": "efficient",
+          "routing_mode": "delegate",
+          "parent": {
+            "model": "gpt-6-astra",
+            "reasoning_effort": "high"
+          },
+          "worker": {
+            "model": "gpt-5.6-luna",
+            "reasoning_effort": "high"
+          }
         }
       ]
     },
     "agentic": {
-      "description": "One repetition comparing direct baselines with real FlowPilot runtime execution under efficient and balanced profiles.",
+      "description": "One repetition comparing direct baselines with real FlowPilot runtime execution under efficient and balanced profiles. Includes Astra/high parent with the real efficient FlowPilot runtime.",
       "repetitions": 1,
       "matrix": [
         {"id": "luna-direct", "strategy": "direct", "model": "gpt-5.6-luna", "reasoning_effort": "high"},
@@ -83,6 +98,21 @@
           "routing_mode": "delegate",
           "parent": {"model": "gpt-5.6-sol", "reasoning_effort": "high"},
           "worker": {"model": "gpt-5.6-luna", "reasoning_effort": "high"}
+        },
+        {
+          "id": "codex-flow-runtime-astra-high",
+          "strategy": "runtime",
+          "reasoning_policy": "fixed",
+          "profile": "efficient",
+          "routing_mode": "delegate",
+          "parent": {
+            "model": "gpt-6-astra",
+            "reasoning_effort": "high"
+          },
+          "worker": {
+            "model": "gpt-5.6-luna",
+            "reasoning_effort": "high"
+          }
         }
       ]
     }

{codex_flow-2.1.13 → codex_flow-2.1.14/codex_flow.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: codex-flow
-Version: 2.1.13
+Version: 2.1.14
 Summary: Intelligent, Efficient, Adaptive Multi-Agent Strategy Orchestration for Codex
 Home-page: https://github.com/ParsifalC/codex-flow
 Author: Parsifal
@@ -45,7 +45,7 @@ Dynamic: requires-python
 
 **智能、高效、自适应的 Codex 多 Agent 策略编排引擎**
 
-[![Version](https://img.shields.io/badge/version-2.1.13-blue.svg?style=flat-square)](VERSION)
+[![Version](https://img.shields.io/badge/version-2.1.14-blue.svg?style=flat-square)](VERSION)
 [![Homebrew](https://img.shields.io/badge/Homebrew-ParsifalC%2Ftap-orange.svg?style=flat-square&logo=homebrew)](https://github.com/ParsifalC/homebrew-tap)
 [![PyPI](https://img.shields.io/pypi/v/codex-flow.svg?style=flat-square&logo=pypi)](https://pypi.org/project/codex-flow/)
 [![Platform](https://img.shields.io/badge/platform-macOS%20%7C%20Linux%20%7C%20Windows-brightgreen.svg?style=flat-square)](#-快速安装)
@@ -256,7 +256,7 @@ codex-flow
 
 ```text
 ╭────────────────────────────────────────────────────────────────────╮
-│                  🚀 codex-flow 控制台 (v2.1.13)                  │
+│                  🚀 codex-flow 控制台 (v2.1.14)                  │
 │    FlowPilot 智能编排 · 确定性任务遥测 · 本地 Benchmark 验证     │
 ╰────────────────────────────────────────────────────────────────────╯
   [1] 🪟 macOS 原生悬浮窗 (overlay widget)

{codex_flow-2.1.13 → codex_flow-2.1.14}/codex_flow.egg-info/SOURCES.txt

@@ -55,6 +55,7 @@ benchmark/profiles.json
 benchmark/schema.json
 benchmark/tasks.json
 benchmark/prices/gpt-5.6-2026-08-30.json
+benchmark/prices/gpt-6-astra-2026-09-12.json
 codex_flow.egg-info/PKG-INFO
 codex_flow.egg-info/SOURCES.txt
 codex_flow.egg-info/dependency_links.txt

{codex_flow-2.1.13 → codex_flow-2.1.14}/scripts/analyze-benchmark.py

@@ -176,6 +176,10 @@ def load_jsonl(path: Path) -> list[dict[str, Any]]:
                 for usage in row["model_usage"]:
                     if usage["role"] == "direct":
                         raise ValueError(f"{path}:{n}: {row['strategy']} model_usage cannot use direct role")
+                    # Runtime workers are selected by the planner. The top-level
+                    # worker fields describe configuration, not observed actors.
+                    if row["strategy"] == "runtime" and usage["role"] == "worker":
+                        continue
                     expected_model = row["model"] if usage["role"] == "parent" else row["worker_model"]
                     expected_effort = row["reasoning_effort"] if usage["role"] == "parent" else row["worker_reasoning_effort"]
                     if usage["model"] != expected_model or usage["reasoning_effort"] != expected_effort:

{codex_flow-2.1.13 → codex_flow-2.1.14}/scripts/benchmark-local.py

@@ -11,7 +11,7 @@ from datetime import datetime, timezone
 from pathlib import Path
 
 ROOT = Path(__file__).resolve().parents[1]
-DEFAULT_PRICE = ROOT / "benchmark/prices/gpt-5.6-2026-08-30.json"
+DEFAULT_PRICE = ROOT / "benchmark/prices/gpt-6-astra-2026-09-12.json"
 
 
 def run(cmd: list[str], *, check: bool = True, capture: bool = False) -> subprocess.CompletedProcess[str]:
@@ -35,7 +35,7 @@ def main() -> int:
     ap = argparse.ArgumentParser(
         description="Run the built-in codex-flow benchmark using the local Codex login session."
     )
-    ap.add_argument("profile", nargs="?", choices=["quick", "full"], default="quick")
+    ap.add_argument("profile", nargs="?", choices=["quick", "full", "agentic"], default="quick")
     ap.add_argument("--workspace", default=".codex-flow-benchmark")
     ap.add_argument("--output")
     ap.add_argument("--prices", default=str(DEFAULT_PRICE))
@@ -79,7 +79,7 @@ def main() -> int:
     if disp_output.startswith(home_str):
         disp_output = "~" + disp_output[len(home_str):]
 
-    budget = "~15M tokens (planning + repairs may increase)" if args.profile == "quick" else "high (90 runs, substantial Codex quota)"
+    budget = "~15M tokens (planning + repairs may increase)" if args.profile == "quick" else f"{planned} runs; usage depends on task execution and repairs"
 
     def pad_line(content: str, width: int = 68) -> str:
         pad = max(0, width - len(content))

{codex_flow-2.1.13 → codex_flow-2.1.14}/scripts/render-benchmark-report.py

@@ -71,6 +71,8 @@ def yes_no(value: bool) -> str:
 def composition(row: dict[str, Any]) -> str:
     if row["strategy"] == "direct":
         return f"{row['model']} / {row['reasoning_effort']}"
+    if row["strategy"] == "runtime":
+        return f"{row['model']} / {row['reasoning_effort']} parent → runtime-selected workers"
     policy = row.get("reasoning_policy", "fixed")
     if policy == "adaptive":
         return f"{row['model']} parent → {row['worker_model']} worker / adaptive"
@@ -180,6 +182,22 @@ def main() -> int:
             f"{pct(item_first, len(items))} | {item_repairs} | {item_reviews} | {item_wall:.1f}s | ${item_cost / len(items):.6f} |"
         )
 
+    runtime_actors: dict[tuple[str, str, str, str], int] = defaultdict(int)
+    for row in rows:
+        if row["strategy"] == "runtime":
+            for usage in row["model_usage"]:
+                key = (row["strategy_id"], usage["role"], usage["model"], usage["reasoning_effort"])
+                runtime_actors[key] += usage["calls"]
+    if runtime_actors:
+        lines.extend([
+            "", "## Runtime actor usage", "",
+            "Runtime worker configuration may differ from planner-selected actors. The following model/effort values come from recorded usage attribution.",
+            "", "| Strategy | Role | Model | Effort | Calls |",
+            "| --- | --- | --- | --- | ---: |",
+        ])
+        for (strategy_id, role, model, effort), calls in sorted(runtime_actors.items()):
+            lines.append(f"| {strategy_id} | {role} | {model} | {effort} | {calls} |")
+
     lines.extend([
         "",
         "## Token efficiency",

{codex_flow-2.1.13 → codex_flow-2.1.14}/scripts/run-benchmark.py

@@ -196,6 +196,20 @@ def parse_usage(jsonl: str) -> dict[str, int]:
     return usage
 
 
+def seed_local_login(destination: Path) -> None:
+    """Seed a private run home without importing global instructions/config."""
+    destination.mkdir(parents=True, exist_ok=True)
+    login_home = Path(os.environ.get("CODEX_HOME", str(Path.home() / ".codex")))
+    login_file = login_home / "auth.json"
+    credential = destination / "auth.json"
+    if login_file.is_file() and not credential.exists():
+        with open(credential, "x", encoding="utf-8", opener=lambda path, flags: os.open(path, flags, 0o600)) as sink:
+            sink.write(login_file.read_text(encoding="utf-8"))
+    model_cache = login_home / "models_cache.json"
+    if model_cache.is_file() and not (destination / "models_cache.json").exists():
+        shutil.copyfile(model_cache, destination / "models_cache.json")
+
+
 def run_codex(
     workdir: Path,
     model: str,
@@ -208,16 +222,19 @@ def run_codex(
     output_schema_path: Path | None = None,
     ignore_user_config: bool = True,
     env: dict[str, str] | None = None,
+    writable_dirs: tuple[Path, ...] = (),
 ) -> tuple[int, dict[str, int], str, float, str]:
     cmd = [
-        "codex", "exec", "--ephemeral", "--json",
+        "codex", "exec", "--json",
     ]
     if ignore_user_config:
-        cmd.append("--ignore-user-config")
+        cmd.extend(["--ephemeral", "--ignore-user-config"])
     cmd.extend([
         "--sandbox", sandbox, "--model", model,
         "-c", f'model_reasoning_effort="{effort}"',
     ])
+    for directory in writable_dirs:
+        cmd.extend(["--add-dir", str(directory)])
     if output_schema_path is not None:
         cmd.extend(["--output-schema", str(output_schema_path)])
     if last_message_path is not None:
@@ -225,20 +242,32 @@ def run_codex(
     cmd.extend(["--cd", str(workdir), prompt])
     started = time.monotonic()
     run_env = os.environ.copy()
+    if ignore_user_config:
+        baseline_home = workdir.parent / ".codex-baseline"
+        seed_local_login(baseline_home)
+        run_env["CODEX_HOME"] = str(baseline_home)
     if env:
         run_env.update(env)
     try:
-        proc = subprocess.run(cmd, text=True, stdout=subprocess.PIPE, stderr=subprocess.PIPE, timeout=timeout, env=run_env)
+        proc = subprocess.run(cmd, cwd=workdir, text=True, stdin=subprocess.DEVNULL, stdout=subprocess.PIPE, stderr=subprocess.PIPE, timeout=timeout, env=run_env)
         elapsed = time.monotonic() - started
         last_message = ""
         if last_message_path is not None and last_message_path.exists():
             last_message = last_message_path.read_text()[-MAX_CAPTURE:]
-        return proc.returncode, parse_usage(proc.stdout), (proc.stderr or proc.stdout)[-MAX_CAPTURE:], elapsed, last_message
+        diagnostic = (
+            proc.stderr[-MAX_CAPTURE // 4:] + "\n" + proc.stdout[-3 * MAX_CAPTURE // 4:]
+            if proc.stderr and proc.stdout else (proc.stderr or proc.stdout)[-MAX_CAPTURE:]
+        )
+        return proc.returncode, parse_usage(proc.stdout), diagnostic, elapsed, last_message
     except subprocess.TimeoutExpired as exc:
         elapsed = time.monotonic() - started
-        stdout = exc.stdout if isinstance(exc.stdout, str) else ""
-        stderr = exc.stderr if isinstance(exc.stderr, str) else ""
-        return 124, parse_usage(stdout), (stderr or stdout)[-MAX_CAPTURE:], elapsed, ""
+        stdout = exc.stdout.decode("utf-8", errors="replace") if isinstance(exc.stdout, bytes) else (exc.stdout or "")
+        stderr = exc.stderr.decode("utf-8", errors="replace") if isinstance(exc.stderr, bytes) else (exc.stderr or "")
+        diagnostic = (
+            stderr[-MAX_CAPTURE // 4:] + "\n" + stdout[-3 * MAX_CAPTURE // 4:]
+            if stderr and stdout else (stderr or stdout)[-MAX_CAPTURE:]
+        )
+        return 124, parse_usage(stdout), diagnostic, elapsed, ""
 
 
 def run_verify(workdir: Path, argv: list[str], timeout: int) -> tuple[bool, str, float]:
@@ -335,7 +364,7 @@ def build_row(
         "repetition": repetition,
         "codex_exit_code": codex_exit,
         "verification_excerpt": verification[-2000:] if not passed else "",
-        "diagnostic_excerpt": diagnostic[-2000:] if codex_exit != 0 else "",
+        "diagnostic_excerpt": diagnostic[-MAX_CAPTURE:] if not passed or codex_exit != 0 else "",
     }
 
 
@@ -639,6 +668,11 @@ def execute_runtime(
     # Pre-authorize hooks for headless non-interactive execution
     authorize_hooks(codex_home)
 
+    # Reuse file-backed local login without importing the user's routing or
+    # plugin configuration. Keep the temporary credential private; the run's
+    # TemporaryDirectory removes it together with the isolated runtime.
+    seed_local_login(codex_home)
+
     runtime_env = {
         "CODEX_HOME": str(codex_home),
         "PATH": f"{bin_dir}:{os.environ.get('PATH', '')}",
@@ -653,7 +687,7 @@ def execute_runtime(
 
     codex_exit, usage, diagnostic, elapsed, _ = run_codex(
         workdir, parent_model, parent_effort, runtime_prompt, timeout,
-        ignore_user_config=False, env=runtime_env,
+        ignore_user_config=False, env=runtime_env, writable_dirs=(codex_home,),
     )
     total_wall += elapsed
     cumulative_codex_usage = empty_usage()
@@ -677,7 +711,7 @@ def execute_runtime(
         )
         codex_exit, usage, diagnostic, elapsed, _ = run_codex(
             workdir, parent_model, parent_effort, repair_prompt, timeout,
-            ignore_user_config=False, env=runtime_env,
+            ignore_user_config=False, env=runtime_env, writable_dirs=(codex_home,),
         )
         total_wall += elapsed
         add_usage(cumulative_codex_usage, usage)
@@ -705,15 +739,20 @@ def execute_runtime(
             workers = run_data.get("workers")
             if isinstance(workers, dict):
                 for w_name, w in workers.items():
-                    if "review" in w_name.lower():
+                    if not isinstance(w, dict):
+                        continue
+                    if "review" in str(w.get("agent_type", w_name)).lower():
                         review_cycles += 1
                     if isinstance(w, dict) and isinstance(w.get("usage"), dict):
                         w_u = w["usage"]
                         if any(w_u.get(k, 0) > 0 for k in ("input_tokens", "cached_input_tokens", "output_tokens")):
-                            add_model_usage(usage_by_model, "worker", worker_model, worker_effort, w_u)
+                            add_model_usage(
+                                usage_by_model, "worker", w.get("model") or worker_model,
+                                w.get("reasoning_effort") or worker_effort, w_u,
+                            )
 
     # Fallback to top-level codex exec usage attributed to parent if telemetry was not populated
-    if not usage_by_model:
+    if not any(role == "parent" for role, _, _ in usage_by_model):
         add_model_usage(usage_by_model, "parent", parent_model, parent_effort, cumulative_codex_usage)
 
     return build_row(
@@ -799,7 +838,7 @@ def main() -> int:
                 for repetition in range(1, manifest.get("repetitions", 1) + 1):
                     with tempfile.TemporaryDirectory(prefix="codex-flow-bench-") as tmp:
                         row = execute_run(
-                            task, config, Path(tmp), repetition,
+                            task, config, Path(tmp).resolve(), repetition,
                             manifest.get("timeout_seconds", 1800),
                             task.get("max_repair_cycles", manifest.get("max_repair_cycles", 2)),
                         )

{codex_flow-2.1.13 → codex_flow-2.1.14}/scripts/update-homebrew-formula.py

@@ -16,7 +16,6 @@ FORMULA_PATH = ROOT / "packaging" / "homebrew" / "codex-flow.rb"
 TEMPLATE = """class CodexFlow < Formula
   desc "Intelligent, Efficient, Adaptive Multi-Agent Strategy Orchestration for Codex"
   homepage "https://github.com/ParsifalC/codex-flow"
-  version "{version}"
 
   on_macos do
     if Hardware::CPU.arm?
@@ -42,7 +41,8 @@ TEMPLATE = """class CodexFlow < Formula
     libexec.install Dir["*"]
     bin.write_exec_script libexec/"bin/codex-flow"
     bin.write_exec_script libexec/"bin/codex-flow-mcp"
-    bash_completion.install libexec/"completions/codex-flow.bash" => "codex-flow" if (libexec/"completions/codex-flow.bash").exist?
+    completion = libexec/"completions/codex-flow.bash"
+    bash_completion.install completion => "codex-flow" if completion.exist?
   end
 
   def caveats
@@ -56,7 +56,7 @@ TEMPLATE = """class CodexFlow < Formula
   end
 
   test do
-    assert_match "codex-flow", shell_output("#{{bin}}/codex-flow --help", 0)
+    assert_match "codex-flow", shell_output("#{{bin}}/codex-flow --help")
   end
 end
 """

codex_flow-2.1.13/VERSION

@@ -1 +0,0 @@
-2.1.13