@pushpalsdev/cli 1.1.34 → 1.1.36

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@pushpalsdev/cli",
-  "version": "1.1.34",
+  "version": "1.1.36",
   "description": "PushPals terminal CLI for LocalBuddy -> RemoteBuddy orchestration",
   "license": "MIT",
   "repository": {

package/runtime/sandbox/apps/workerpals/src/backends/openai_codex/openai_codex_executor.py

@@ -116,6 +116,9 @@ _WEB_REVIEW_NO_EDIT_WATCHDOG_S = 240
 _BACKGROUND_NO_EDIT_WATCHDOG_S = 120
 _NO_EDIT_RECOVERY_WATCHDOG_S = 90
 _DEFAULT_NO_EDIT_RECHECK_S = 120
+_DEFAULT_NO_EDIT_COMMAND_GRACE_S = 240
+_DEFAULT_STARTUP_STALL_WATCHDOG_S = 210
+_RECOVERY_STARTUP_STALL_WATCHDOG_S = 150
 _DEFAULT_ROLLOUT_WATCHDOG_S = 300
 _SMALL_TASK_ROLLOUT_WATCHDOG_S = 240
 _NARROW_TEST_TASK_ROLLOUT_WATCHDOG_S = 150
@@ -755,6 +758,65 @@ def _resolve_no_edit_recheck_seconds(communicate_timeout_s: Optional[int]) -> in
     return max(1, min(_DEFAULT_NO_EDIT_RECHECK_S, upper))
 
 
+def _resolve_no_edit_command_grace_seconds(communicate_timeout_s: Optional[int]) -> Optional[int]:
+    if not communicate_timeout_s:
+        return None
+
+    raw = os.environ.get("WORKERPALS_OPENAI_CODEX_NO_EDIT_COMMAND_GRACE_S", "").strip()
+    if raw:
+        if raw == "0":
+            return None
+        parsed = _to_positive_int(raw)
+        if parsed is None:
+            log.info(
+                "Invalid WORKERPALS_OPENAI_CODEX_NO_EDIT_COMMAND_GRACE_S="
+                f"{raw!r}; using default command-progress grace."
+            )
+        else:
+            return max(1, min(parsed, max(1, communicate_timeout_s - 1)))
+
+    upper = max(1, communicate_timeout_s - 1)
+    return max(1, min(_DEFAULT_NO_EDIT_COMMAND_GRACE_S, upper))
+
+
+def _resolve_startup_stall_watchdog_seconds(
+    communicate_timeout_s: Optional[int],
+    recovery_attempt: int = 0,
+) -> Optional[int]:
+    if not communicate_timeout_s:
+        return None
+
+    raw = os.environ.get("WORKERPALS_OPENAI_CODEX_STARTUP_STALL_WATCHDOG_S", "").strip()
+    if raw:
+        if raw == "0":
+            return None
+        parsed = _to_positive_int(raw)
+        if parsed is None:
+            log.info(
+                "Invalid WORKERPALS_OPENAI_CODEX_STARTUP_STALL_WATCHDOG_S="
+                f"{raw!r}; using default startup-stall watchdog."
+            )
+        else:
+            return max(1, min(parsed, max(1, communicate_timeout_s - 1)))
+
+    default_s = (
+        _RECOVERY_STARTUP_STALL_WATCHDOG_S
+        if recovery_attempt > 0
+        else _DEFAULT_STARTUP_STALL_WATCHDOG_S
+    )
+    floor_s = 60
+    return max(floor_s, min(default_s, max(floor_s, communicate_timeout_s - 1)))
+
+
+def _startup_stall_recovery_model(current_model: str) -> str:
+    normalized = str(current_model or "").strip()
+    if not normalized:
+        return LEGACY_CODEX_MODEL_FALLBACK
+    if normalized.lower() == LEGACY_CODEX_MODEL_FALLBACK.lower():
+        return normalized
+    return LEGACY_CODEX_MODEL_FALLBACK
+
+
 def _looks_like_web_review_prompt(prompt: str) -> bool:
     text = str(prompt or "").lower()
     return "repo-native web review" in text or "web review path" in text
@@ -1299,10 +1361,94 @@ def _empty_codex_trace() -> Dict[str, Any]:
         "prompt_tokens": 0,
         "completion_tokens": 0,
         "total_tokens": 0,
+        "active_command_ids": [],
+        "command_event_count": 0,
+        "last_command_activity_at": None,
+        "last_command_summary": "",
     }
 
 
-def _record_live_codex_stdout_line(line: str, use_json: bool, trace: Dict[str, Any]) -> None:
+def _looks_like_codex_command_item(value: Any) -> bool:
+    if not isinstance(value, dict):
+        return False
+    type_text = " ".join(
+        str(value.get(key) or "")
+        for key in ("type", "item_type", "kind", "name", "tool_name")
+    ).lower()
+    if any(marker in type_text for marker in ("command_execution", "exec_command", "shell_command")):
+        return True
+    return any(key in value for key in ("command", "cmd", "exit_code", "aggregated_output"))
+
+
+def _record_codex_command_activity(
+    parsed: Dict[str, Any],
+    event_type: str,
+    trace: Dict[str, Any],
+    now: float,
+) -> None:
+    item = parsed.get("item")
+    command_source: Any = item if _looks_like_codex_command_item(item) else parsed
+    if not _looks_like_codex_command_item(command_source):
+        return
+
+    command_text = ""
+    if isinstance(command_source, dict):
+        for key in ("command", "cmd", "name"):
+            raw = command_source.get(key)
+            if isinstance(raw, str) and raw.strip():
+                command_text = _truncate_inline(raw.strip(), 160)
+                break
+    command_id = ""
+    if isinstance(command_source, dict):
+        command_id = str(
+            command_source.get("id")
+            or command_source.get("call_id")
+            or command_source.get("item_id")
+            or command_text
+            or "command"
+        ).strip()
+    command_id = command_id or "command"
+
+    active = trace.setdefault("active_command_ids", [])
+    if not isinstance(active, list):
+        active = []
+        trace["active_command_ids"] = active
+
+    status_text = ""
+    if isinstance(command_source, dict):
+        status_text = " ".join(
+            str(command_source.get(key) or "")
+            for key in ("status", "state", "outcome")
+        ).lower()
+    event_lower = event_type.lower()
+    completed = (
+        "completed" in event_lower
+        or "failed" in event_lower
+        or "error" in event_lower
+        or any(marker in status_text for marker in ("completed", "failed", "cancelled", "canceled", "exited"))
+    )
+    started = (
+        "started" in event_lower
+        or "updated" in event_lower
+        or any(marker in status_text for marker in ("running", "in_progress", "started"))
+    )
+
+    if completed:
+        trace["active_command_ids"] = [item for item in active if str(item) != command_id]
+    elif started and command_id not in active:
+        active.append(command_id)
+
+    trace["command_event_count"] = to_int(trace.get("command_event_count"), 0) + 1
+    trace["last_command_activity_at"] = float(now)
+    trace["last_command_summary"] = command_text or event_type
+
+
+def _record_live_codex_stdout_line(
+    line: str,
+    use_json: bool,
+    trace: Dict[str, Any],
+    now: Optional[float] = None,
+) -> None:
     stripped = line.strip()
     if not stripped:
         return
@@ -1329,6 +1475,7 @@ def _record_live_codex_stdout_line(line: str, use_json: bool, trace: Dict[str, A
             return
 
         if isinstance(parsed, dict):
+            observed_at = float(now if now is not None else time.monotonic())
             usage = _extract_usage_counts(parsed)
             if usage is not None:
                 trace["prompt_tokens"] = max(
@@ -1345,6 +1492,7 @@ def _record_live_codex_stdout_line(line: str, use_json: bool, trace: Dict[str, A
                 .strip()
                 or "event"
             )
+            _record_codex_command_activity(parsed, event_type, trace, observed_at)
             event_type_counts[event_type] = to_int(event_type_counts.get(event_type), 0) + 1
             summary = _summarize_json_event(parsed)
             # Reasoning can arrive under generic event types (for example item.updated).
@@ -1409,10 +1557,13 @@ def _finalize_codex_stdout_trace(trace: Dict[str, Any], use_json: bool) -> Dict[
     prompt_tokens = to_int(trace.get("prompt_tokens"), 0)
     completion_tokens = to_int(trace.get("completion_tokens"), 0)
     total_tokens = to_int(trace.get("total_tokens"), 0)
+    command_event_count = to_int(trace.get("command_event_count"), 0)
     if reasoning_events > 0:
         log.info(f"[codex] Reasoning-like event(s): {reasoning_events}")
     elif use_json and valid_json > 0:
         log.info("[codex] No reasoning-like events observed in this run.")
+    if command_event_count > 0:
+        log.info(f"[codex] Command execution event(s): {command_event_count}")
     if total_tokens > 0:
         log.info(
             f"[codex] Usage observed: prompt={prompt_tokens} completion={completion_tokens} total={total_tokens}"
@@ -1433,6 +1584,7 @@ def _finalize_codex_stdout_trace(trace: Dict[str, Any], use_json: bool) -> Dict[
         "prompt_tokens": prompt_tokens,
         "completion_tokens": completion_tokens,
         "total_tokens": total_tokens,
+        "command_event_count": command_event_count,
     }
 
 
@@ -2259,8 +2411,14 @@ def _run_codex_task(
                         if not line:
                             continue
                         with trace_lock:
-                            last_activity_at["ts"] = time.monotonic()
-                            _record_live_codex_stdout_line(line, use_json, stdout_trace_state)
+                            observed_at = time.monotonic()
+                            last_activity_at["ts"] = observed_at
+                            _record_live_codex_stdout_line(
+                                line,
+                                use_json,
+                                stdout_trace_state,
+                                observed_at,
+                            )
                 except Exception:
                     pass
                 finally:
@@ -2337,6 +2495,16 @@ def _run_codex_task(
                 else None
             )
             no_edit_recheck_s = _resolve_no_edit_recheck_seconds(communicate_timeout_s)
+            no_edit_command_grace_s = _resolve_no_edit_command_grace_seconds(communicate_timeout_s)
+            startup_stall_watchdog_s = _resolve_startup_stall_watchdog_seconds(
+                communicate_timeout_s,
+                recovery_attempt=startup_stall_recovery_attempt,
+            )
+            startup_stall_deadline = (
+                started_at + float(startup_stall_watchdog_s)
+                if startup_stall_watchdog_s is not None
+                else None
+            )
             rollout_watchdog_s = (
                 _resolve_rollout_watchdog_seconds(
                     prompt,
@@ -2351,6 +2519,11 @@ def _run_codex_task(
                 if no_edit_watchdog_s is not None
                 else None
             )
+            no_edit_command_grace_cap_deadline = (
+                started_at + float(no_edit_watchdog_s + no_edit_command_grace_s)
+                if no_edit_watchdog_s is not None and no_edit_command_grace_s is not None
+                else None
+            )
             rollout_deadline = (
                 started_at + float(rollout_watchdog_s)
                 if rollout_watchdog_s is not None
@@ -2364,9 +2537,93 @@ def _run_codex_task(
                     _terminate_active_child()
                     break
 
+                if startup_stall_deadline is not None and now >= startup_stall_deadline:
+                    with trace_lock:
+                        live_trace = dict(stdout_trace_state)
+                        summaries = stdout_trace_state.get("summaries")
+                        if isinstance(summaries, list):
+                            live_trace["summaries"] = list(summaries)
+                    if _codex_trace_is_startup_stall(live_trace):
+                        changed_paths, _, effective_paths = _codex_changed_paths(repo, baseline_snapshot)
+                        if not effective_paths:
+                            no_edit_artifact_only_paths = _describe_non_publishable_paths(
+                                changed_paths,
+                                baseline_snapshot,
+                            )
+                            no_edit_watchdog_fired = True
+                            elapsed_s = int(max(0.0, now - started_at))
+                            log.info(
+                                f"Startup-stall watchdog fired after {elapsed_s}s with no assistant/tool progress."
+                            )
+                            _terminate_active_child()
+                            break
+                    startup_stall_deadline = None
+
                 if no_edit_deadline is not None and now >= no_edit_deadline:
                     changed_paths, _, effective_paths = _codex_changed_paths(repo, baseline_snapshot)
                     if not effective_paths:
+                        with trace_lock:
+                            live_trace = dict(stdout_trace_state)
+                            summaries = stdout_trace_state.get("summaries")
+                            if isinstance(summaries, list):
+                                live_trace["summaries"] = list(summaries)
+                        startup_only = _codex_trace_is_startup_stall(live_trace)
+                        if (
+                            startup_only
+                            and startup_stall_deadline is not None
+                            and now < startup_stall_deadline
+                        ):
+                            no_edit_deadline = startup_stall_deadline
+                            remaining_s = int(max(1.0, startup_stall_deadline - now))
+                            log.info(
+                                "No-edit watchdog observed only Codex startup events; "
+                                f"allowing {remaining_s}s for first assistant/tool progress "
+                                "before startup-stall recovery."
+                            )
+                            continue
+                        command_event_count = to_int(live_trace.get("command_event_count"), 0)
+                        active_commands_raw = live_trace.get("active_command_ids")
+                        active_command_count = (
+                            len(active_commands_raw)
+                            if isinstance(active_commands_raw, list)
+                            else 0
+                        )
+                        last_command_activity_at = 0.0
+                        try:
+                            last_command_activity_at = float(
+                                live_trace.get("last_command_activity_at") or 0.0
+                            )
+                        except Exception:
+                            last_command_activity_at = 0.0
+                        if command_event_count > 0 and no_edit_command_grace_s is not None:
+                            command_grace_deadline = 0.0
+                            if active_command_count > 0:
+                                # Do not kill while Codex is actively running a tool command; poll
+                                # again soon, but keep the total grace bounded by the hard cap below.
+                                command_grace_deadline = now + min(60.0, float(no_edit_command_grace_s))
+                            elif last_command_activity_at > 0:
+                                command_grace_deadline = last_command_activity_at + float(
+                                    no_edit_command_grace_s
+                                )
+                            if no_edit_command_grace_cap_deadline is not None:
+                                command_grace_deadline = min(
+                                    command_grace_deadline,
+                                    no_edit_command_grace_cap_deadline,
+                                )
+                            if command_grace_deadline > now:
+                                no_edit_deadline = command_grace_deadline
+                                remaining_s = int(max(1.0, command_grace_deadline - now))
+                                command_detail = (
+                                    f"{active_command_count} active command(s)"
+                                    if active_command_count > 0
+                                    else "recent command completion"
+                                )
+                                log.info(
+                                    "No-edit watchdog observed Codex tool progress "
+                                    f"({command_detail}); allowing {remaining_s}s for a "
+                                    "publishable patch before recovery."
+                                )
+                                continue
                         no_edit_artifact_only_paths = _describe_non_publishable_paths(
                             changed_paths,
                             baseline_snapshot,
@@ -2377,9 +2634,15 @@ def _run_codex_task(
                             if no_edit_artifact_only_paths
                             else ""
                         )
-                        log.info(
-                            f"No-edit watchdog fired after {int(no_edit_watchdog_s or 0)}s with no publishable file changes.{artifact_detail} Retrying with patch-first guidance."
-                        )
+                        if startup_only:
+                            elapsed_s = int(max(0.0, now - started_at))
+                            log.info(
+                                f"Startup-stall watchdog fired after {elapsed_s}s with no assistant/tool progress."
+                            )
+                        else:
+                            log.info(
+                                f"No-edit watchdog fired after {int(no_edit_watchdog_s or 0)}s with no publishable file changes.{artifact_detail} Retrying with patch-first guidance."
+                            )
                         _terminate_active_child()
                         break
                     no_edit_deadline = now + float(no_edit_recheck_s)
@@ -2550,9 +2813,15 @@ def _run_codex_task(
                     *supplemental_guidance,
                     _build_startup_stall_recovery_guidance(trace_excerpt),
                 ]
+                recovery_model = _startup_stall_recovery_model(model)
+                recovery_detail = (
+                    f" using fallback model {recovery_model!r}"
+                    if recovery_model and recovery_model != model
+                    else ""
+                )
                 log.warning(
                     "Codex emitted only startup events before the no-edit watchdog; "
-                    "restarting Codex once before classifying the job terminally."
+                    f"restarting Codex once{recovery_detail} before classifying the job terminally."
                 )
                 retry_result = _run_codex_task(
                     repo,
@@ -2563,7 +2832,7 @@ def _run_codex_task(
                     startup_stall_recovery_attempt=startup_stall_recovery_attempt + 1,
                     no_edit_recovery_attempt=no_edit_recovery_attempt,
                     rollout_recovery_attempt=rollout_recovery_attempt,
-                    model_override=model_override,
+                    model_override=recovery_model or model_override,
                     baseline_changes=baseline_snapshot,
                 )
                 retry_result["usage"] = _merge_usage_records(usage, retry_result.get("usage"))

package/runtime/sandbox/apps/workerpals/src/backends/openai_codex/test_openai_codex_runtime_config.py

@@ -49,6 +49,7 @@ from openai_codex_executor import (
     _resolve_codex_command_prefix,
     _resolve_no_edit_watchdog_seconds,
     _resolve_rollout_watchdog_seconds,
+    _resolve_startup_stall_watchdog_seconds,
     _unwrap_shell_wrapper_command,
     _usage_from_trace_or_estimate,
 )
@@ -372,6 +373,63 @@ class OpenAICodexRuntimeConfigTests(unittest.TestCase):
             self.assertEqual(task.repo, str(repo.resolve()))
             self.assertEqual(task.instruction, "Make one small publishable change")
 
+    def test_parse_payload_accepts_positional_payload_file_path(self) -> None:
+        with tempfile.TemporaryDirectory(prefix="pushpals-payload-file-positional-") as temp_dir:
+            repo = Path(temp_dir) / "repo"
+            repo.mkdir(parents=True, exist_ok=True)
+            payload = {
+                "kind": "task.execute",
+                "repo": str(repo),
+                "params": {"instruction": "Recover from a direct-worker payload handoff"},
+            }
+            encoded = base64.b64encode(json.dumps(payload).encode("utf-8")).decode("ascii")
+            payload_file = Path(temp_dir) / "payload.b64"
+            payload_file.write_text(encoded, encoding="utf-8")
+
+            task = parse_task_execute_payload(
+                ["executor", str(payload_file)],
+                logger=Logger("[test]"),
+            )
+
+            self.assertEqual(task.kind, "task.execute")
+            self.assertEqual(task.repo, str(repo.resolve()))
+            self.assertEqual(task.instruction, "Recover from a direct-worker payload handoff")
+
+    def test_parse_payload_accepts_unpadded_base64_payload(self) -> None:
+        with tempfile.TemporaryDirectory(prefix="pushpals-payload-unpadded-") as temp_dir:
+            repo = Path(temp_dir) / "repo"
+            repo.mkdir(parents=True, exist_ok=True)
+            payload = {
+                "kind": "task.execute",
+                "repo": str(repo),
+                "params": {"instruction": "Accept wrapper-normalized payload padding"},
+            }
+            encoded = base64.b64encode(json.dumps(payload).encode("utf-8")).decode("ascii")
+            unpadded = encoded.rstrip("=")
+
+            task = parse_task_execute_payload(["executor", unpadded], logger=Logger("[test]"))
+
+            self.assertEqual(task.kind, "task.execute")
+            self.assertEqual(task.repo, str(repo.resolve()))
+            self.assertEqual(task.instruction, "Accept wrapper-normalized payload padding")
+
+    def test_parse_payload_accepts_raw_json_payload(self) -> None:
+        with tempfile.TemporaryDirectory(prefix="pushpals-payload-raw-json-") as temp_dir:
+            repo = Path(temp_dir) / "repo"
+            repo.mkdir(parents=True, exist_ok=True)
+            payload = {
+                "kind": "task.execute",
+                "repo": str(repo),
+                "params": {"instruction": "Accept raw JSON from a recovery wrapper"},
+            }
+            raw_json = json.dumps(payload)
+
+            task = parse_task_execute_payload(["executor", raw_json], logger=Logger("[test]"))
+
+            self.assertEqual(task.kind, "task.execute")
+            self.assertEqual(task.repo, str(repo.resolve()))
+            self.assertEqual(task.instruction, "Accept raw JSON from a recovery wrapper")
+
     def test_parse_payload_prefers_helper_tests_for_visual_derivation_tasks(self) -> None:
         with tempfile.TemporaryDirectory(prefix="pushpals-visual-guidance-") as temp_dir:
             repo = Path(temp_dir) / "repo"
@@ -1091,13 +1149,16 @@ class OpenAICodexRuntimeConfigTests(unittest.TestCase):
                         "",
                         "argv = sys.argv[1:]",
                         "last_message_path = None",
+                        "model = ''",
                         "for index, arg in enumerate(argv):",
                         "    if arg == '--output-last-message' and index + 1 < len(argv):",
                         "        last_message_path = argv[index + 1]",
+                        "    if arg == '-m' and index + 1 < len(argv):",
+                        "        model = argv[index + 1]",
                         "        break",
                         "",
                         "prompt = sys.stdin.read()",
-                        "if 'Codex startup-stall recovery' in prompt:",
+                        "if 'Codex startup-stall recovery' in prompt and model == 'gpt-5.4':",
                         "    Path('src').mkdir(exist_ok=True)",
                         "    Path('src/startup-stall-recovered.txt').write_text('patched after restart\\n', encoding='utf-8')",
                         "    if last_message_path:",
@@ -1119,7 +1180,8 @@ class OpenAICodexRuntimeConfigTests(unittest.TestCase):
                 "OPENAI_API_KEY": "pushpals-startup-stall-test-key",
                 "WORKERPALS_OPENAI_CODEX_JSON": "true",
                 "WORKERPALS_OPENAI_CODEX_TIMEOUT_S": "20",
-                "WORKERPALS_OPENAI_CODEX_NO_EDIT_WATCHDOG_S": "1",
+                "WORKERPALS_OPENAI_CODEX_NO_EDIT_WATCHDOG_S": "0",
+                "WORKERPALS_OPENAI_CODEX_STARTUP_STALL_WATCHDOG_S": "1",
                 "WORKERPALS_OPENAI_CODEX_PROGRESS_LOG_INTERVAL_S": "1",
             }
             with mock.patch.dict(os.environ, env_overrides, clear=False):
@@ -1189,6 +1251,7 @@ class OpenAICodexRuntimeConfigTests(unittest.TestCase):
                 "WORKERPALS_OPENAI_CODEX_JSON": "true",
                 "WORKERPALS_OPENAI_CODEX_TIMEOUT_S": "20",
                 "WORKERPALS_OPENAI_CODEX_NO_EDIT_WATCHDOG_S": "1",
+                "WORKERPALS_OPENAI_CODEX_STARTUP_STALL_WATCHDOG_S": "1",
                 "WORKERPALS_OPENAI_CODEX_PROGRESS_LOG_INTERVAL_S": "1",
             }
             with mock.patch.dict(os.environ, env_overrides, clear=False):
@@ -1284,6 +1347,90 @@ class OpenAICodexRuntimeConfigTests(unittest.TestCase):
         self.assertIn("Patched immediately after no-edit recovery", str(result.get("stdout") or ""))
         self.assertIn("src/", str(result.get("stdout") or ""))
 
+    def test_run_codex_task_no_edit_watchdog_allows_command_backed_discovery(self) -> None:
+        with tempfile.TemporaryDirectory(prefix="pushpals-codex-no-edit-command-grace-") as temp_dir:
+            repo = Path(temp_dir) / "repo"
+            repo.mkdir(parents=True, exist_ok=True)
+            (repo / "README.md").write_text("# command grace repo\n", encoding="utf-8")
+            subprocess.run(["git", "init"], cwd=repo, check=True, capture_output=True, text=True)
+            subprocess.run(
+                ["git", "config", "user.name", "PushPals Test"],
+                cwd=repo,
+                check=True,
+                capture_output=True,
+                text=True,
+            )
+            subprocess.run(
+                ["git", "config", "user.email", "pushpals-tests@example.com"],
+                cwd=repo,
+                check=True,
+                capture_output=True,
+                text=True,
+            )
+            subprocess.run(["git", "add", "README.md"], cwd=repo, check=True, capture_output=True, text=True)
+            subprocess.run(
+                ["git", "commit", "-m", "chore: seed command grace repo"],
+                cwd=repo,
+                check=True,
+                capture_output=True,
+                text=True,
+            )
+
+            stub_path = Path(temp_dir) / "fake_codex_no_edit_command_grace.py"
+            stub_path.write_text(
+                "\n".join(
+                    [
+                        "from pathlib import Path",
+                        "import json",
+                        "import sys",
+                        "import time",
+                        "",
+                        "argv = sys.argv[1:]",
+                        "last_message_path = None",
+                        "for index, arg in enumerate(argv):",
+                        "    if arg == '--output-last-message' and index + 1 < len(argv):",
+                        "        last_message_path = argv[index + 1]",
+                        "        break",
+                        "",
+                        "sys.stdin.read()",
+                        "print(json.dumps({'type': 'thread.started'}), flush=True)",
+                        "print(json.dumps({'type': 'turn.started'}), flush=True)",
+                        "print(json.dumps({'type': 'item.started', 'item': {'id': 'cmd-read-target', 'type': 'command_execution', 'command': 'sed -n 1,120p README.md', 'status': 'in_progress'}}), flush=True)",
+                        "time.sleep(1.4)",
+                        "print(json.dumps({'type': 'item.completed', 'item': {'id': 'cmd-read-target', 'type': 'command_execution', 'command': 'sed -n 1,120p README.md', 'status': 'completed', 'exit_code': 0, 'aggregated_output': '# command grace repo'}}), flush=True)",
+                        "time.sleep(1.6)",
+                        "Path('src').mkdir(exist_ok=True)",
+                        "Path('src/command-grace.txt').write_text('patched after command-backed discovery\\n', encoding='utf-8')",
+                        "if last_message_path:",
+                        "    Path(last_message_path).write_text('Patched after command-backed discovery.', encoding='utf-8')",
+                        "print(json.dumps({'type': 'item.completed', 'item': {'type': 'message', 'text': 'Patched after command-backed discovery.'}}), flush=True)",
+                    ]
+                ),
+                encoding="utf-8",
+            )
+
+            env_overrides = {
+                "PUSHPALS_OPENAI_CODEX_BIN_JSON": json.dumps([sys.executable, str(stub_path)]),
+                "PUSHPALS_OPENAI_CODEX_AUTH_MODE": "api_key",
+                "OPENAI_API_KEY": "pushpals-no-edit-command-grace-test-key",
+                "WORKERPALS_OPENAI_CODEX_JSON": "true",
+                "WORKERPALS_OPENAI_CODEX_TIMEOUT_S": "20",
+                "WORKERPALS_OPENAI_CODEX_NO_EDIT_WATCHDOG_S": "1",
+                "WORKERPALS_OPENAI_CODEX_NO_EDIT_COMMAND_GRACE_S": "5",
+                "WORKERPALS_OPENAI_CODEX_PROGRESS_LOG_INTERVAL_S": "1",
+            }
+            with mock.patch.dict(os.environ, env_overrides, clear=False):
+                result = _run_codex_task(
+                    str(repo),
+                    "Add one focused contract assertion after inspecting the hinted test.",
+                    [],
+                )
+
+        self.assertTrue(result.get("ok"), result)
+        self.assertEqual(result.get("exitCode"), 0)
+        self.assertIn("Patched after command-backed discovery", str(result.get("stdout") or ""))
+        self.assertIn("src/", str(result.get("stdout") or ""))
+
     def test_run_codex_task_recovery_attempt_is_still_guarded_by_no_edit_watchdog(self) -> None:
         with tempfile.TemporaryDirectory(prefix="pushpals-codex-no-edit-watchdog-fail-") as temp_dir:
             repo = Path(temp_dir) / "repo"
@@ -1587,6 +1734,31 @@ class OpenAICodexRuntimeConfigTests(unittest.TestCase):
 
         self.assertEqual(watchdog_s, 180)
 
+    def test_startup_stall_watchdog_allows_slower_first_response_than_no_edit_watchdog(self) -> None:
+        with mock.patch.dict(
+            os.environ,
+            {"WORKERPALS_OPENAI_CODEX_STARTUP_STALL_WATCHDOG_S": ""},
+            clear=False,
+        ):
+            watchdog_s = _resolve_startup_stall_watchdog_seconds(1200)
+            recovery_watchdog_s = _resolve_startup_stall_watchdog_seconds(
+                1200,
+                recovery_attempt=1,
+            )
+
+        self.assertEqual(watchdog_s, 210)
+        self.assertEqual(recovery_watchdog_s, 150)
+
+    def test_explicit_startup_stall_watchdog_override_is_bounded(self) -> None:
+        with mock.patch.dict(
+            os.environ,
+            {"WORKERPALS_OPENAI_CODEX_STARTUP_STALL_WATCHDOG_S": "500"},
+            clear=False,
+        ):
+            watchdog_s = _resolve_startup_stall_watchdog_seconds(120)
+
+        self.assertEqual(watchdog_s, 119)
+
     def test_narrow_contract_regression_with_required_e2e_uses_fast_no_edit_watchdog(self) -> None:
         prompt = (
             "Harden the opportunity graph contract around autonomous delivery-loop failure signals. "

package/runtime/sandbox/apps/workerpals/src/backends/shared/executor_base.py

@@ -155,14 +155,39 @@ def fail(summary: str, stderr: Optional[str] = None, exit_code: int = 1) -> int:
     return exit_code
 
 
-def decode_payload(raw: str) -> Dict[str, Any]:
-    decoded = base64.b64decode(raw).decode("utf-8")
-    payload = json.loads(decoded)
+def _parse_payload_json(raw: str) -> Dict[str, Any]:
+    payload = json.loads(raw)
     if not isinstance(payload, dict):
         raise ValueError("payload must be a JSON object")
     return payload
 
 
+def decode_payload(raw: str) -> Dict[str, Any]:
+    stripped = str(raw or "").strip()
+    if not stripped:
+        raise ValueError("empty job payload")
+
+    # Direct workers normally receive a file-backed base64 payload, but this
+    # parser intentionally accepts the safe adjacent encodings too. That keeps
+    # executor startup resilient if an outer wrapper normalizes padding, uses
+    # url-safe base64, or hands through raw JSON during recovery.
+    if stripped.startswith("{"):
+        return _parse_payload_json(stripped)
+
+    compact = "".join(stripped.split())
+    padded = compact + ("=" * ((4 - len(compact) % 4) % 4))
+    decode_errors: List[str] = []
+    for decoder in (base64.b64decode, base64.urlsafe_b64decode):
+        try:
+            decoded = decoder(padded).decode("utf-8")
+            return _parse_payload_json(decoded)
+        except Exception as exc:
+            decode_errors.append(str(exc))
+
+    detail = "; ".join(error for error in decode_errors if error) or "unknown decode error"
+    raise ValueError(f"invalid base64/JSON job payload: {detail}")
+
+
 def read_encoded_payload_arg(argv: List[str]) -> str:
     if len(argv) < 2:
         raise ValueError("missing base64 job payload")
@@ -174,6 +199,13 @@ def read_encoded_payload_arg(argv: List[str]) -> str:
         return path.read_text(encoding="utf-8").strip()
     if mode == "--payload-stdin":
         return sys.stdin.read().strip()
+    if len(mode) < 4096:
+        try:
+            path = Path(mode).expanduser()
+            if path.is_file():
+                return path.read_text(encoding="utf-8").strip()
+        except OSError:
+            pass
     return mode
 
 

package/runtime/sandbox/apps/workerpals/src/docker_executor.ts

@@ -1918,8 +1918,6 @@ export class DockerExecutor {
 
   private matchesRetryablePattern(text: string): boolean {
     const transientPatterns: RegExp[] = [
-      /\bstalled before first response\b/i,
-      /\bstartup stall\b/i,
       /warm .*runtime/i,
       /failed to start warm container/i,
       /docker execution error/i,