@pushpalsdev/cli 1.0.18 → 1.0.20

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

@@ -0,0 +1,1259 @@
+#!/usr/bin/env python3
+"""PushPals OpenAI Codex backend wrapper.
+
+Runs `codex exec` in non-interactive mode and emits one structured result line
+that the TypeScript host parses.
+"""
+
+from __future__ import annotations
+
+import json
+import os
+import re
+from shutil import which
+import shlex
+import signal
+import subprocess
+import sys
+import tempfile
+import threading
+import time
+import traceback
+from dataclasses import dataclass
+from pathlib import Path
+from typing import Any, Dict, List, Optional
+
+_SHARED_DIR = Path(__file__).resolve().parents[1] / "shared"
+if str(_SHARED_DIR) not in sys.path:
+    sys.path.insert(0, str(_SHARED_DIR))
+
+from executor_base import (
+    Logger,
+    SettingsResolver,
+    build_settings_resolver,
+    emit,
+    log_git_status,
+    looks_local_base_url,
+    parse_task_execute_payload,
+    resolve_llm_config,
+    summarize_git_changes,
+    to_int,
+    to_single_line,
+)
+
+LOG_PREFIX = "[OpenAICodexExecutor]"
+DEFAULT_CODEX_MODEL = "gpt-5-codex"
+_ACTIVE_CHILD: Optional[subprocess.Popen[str]] = None
+_INTERRUPTED_SIGNAL: Optional[int] = None
+log = Logger(LOG_PREFIX)
+
+_PROMPT_TEMPLATE_CACHE: Dict[str, str] = {}
+_PROMPT_TOKEN_REGEX = re.compile(r"\{\{\s*([a-zA-Z0-9_]+)\s*\}\}")
+_TASK_SYSTEM_PROMPT_PATH = "workerpals/openai_codex_task_execute_system_prompt.md"
+_DEFAULT_TASK_SYSTEM_PROMPT_PATH = "workerpals/openai_codex_default_system_prompt.md"
+_MANDATORY_RUNTIME_POLICY_APPENDIX_PATH = "workerpals/openai_codex_runtime_policy_appendix.md"
+_INSTRUCTION_WRAPPER_PROMPT_PATH = "workerpals/openai_codex_instruction_wrapper.md"
+_SUPPLEMENTAL_GUIDANCE_SECTION_PATH = "workerpals/openai_codex_supplemental_guidance_section.md"
+_CODEX_WORKAROUND_PATTERNS = (
+    re.compile(
+        r"\bcodex cli\b.{0,120}\b(isn't|is not|not)\b.{0,120}\bavailable\b.{0,120}\b(so|therefore|instead|fallback|workaround|without|using)\b",
+        re.IGNORECASE,
+    ),
+    re.compile(r"\bwithout requiring\b.{0,120}\bcodex\b", re.IGNORECASE),
+    re.compile(r"\bavoid(?:ing)?\b.{0,120}\bcodex\b.{0,120}\bcall", re.IGNORECASE),
+    re.compile(r"\b(fell back|fallback|worked around|workaround|bypass(?:ed)?|switched to)\b.{0,120}\bcodex\b", re.IGNORECASE),
+)
+_CODEX_WORKAROUND_NEGATION_HINTS = (
+    "do not",
+    "don't",
+    "never",
+    "must not",
+    "fail loudly",
+    "hard-fail",
+    "hard fail",
+    "explicit failure",
+    "codex cli is required infrastructure",
+)
+
+_VALID_APPROVAL_POLICIES = {"untrusted", "on-failure", "on-request", "never"}
+_VALID_SANDBOX_POLICIES = {"read-only", "workspace-write", "danger-full-access"}
+_VALID_COLORS = {"always", "never", "auto"}
+_VALID_AUTH_MODES = {"auto", "api_key", "chatgpt"}
+_VALID_REASONING_EFFORTS = {"low", "medium", "high"}
+
+
+@dataclass(frozen=True)
+class OpenAICodexRuntimeConfig:
+    codex_bin_json: str
+    codex_bin: str
+    auth_mode: str
+    base_url_override: str
+    timeout_seconds_override: int
+    timeout_ms_top_level: int
+    timeout_ms_llm_codex: int
+    timeout_ms_backend: int
+    progress_log_interval_s: int
+    reasoning_effort: str
+    approval_policy: str
+    sandbox: str
+    color: str
+    json_output: bool
+
+    @classmethod
+    def from_sources(cls, settings: Optional[SettingsResolver] = None) -> "OpenAICodexRuntimeConfig":
+        cfg = settings or build_settings_resolver()
+        return cls(
+            codex_bin_json=cfg.get_str(
+                env_names=("PUSHPALS_OPENAI_CODEX_BIN_JSON",),
+                config_paths=("workerpals.llm.codex_bin_json", "workerpals.openai_codex.bin_json"),
+                default="",
+            ),
+            codex_bin=cfg.get_str(
+                env_names=("PUSHPALS_OPENAI_CODEX_BIN",),
+                config_paths=("workerpals.llm.codex_bin", "workerpals.openai_codex.bin"),
+                default="",
+            ),
+            auth_mode=cfg.get_str(
+                env_names=("PUSHPALS_OPENAI_CODEX_AUTH_MODE",),
+                config_paths=("workerpals.llm.codex_auth_mode", "workerpals.openai_codex.auth_mode"),
+                default="auto",
+            ),
+            base_url_override=cfg.get_str(
+                env_names=("PUSHPALS_OPENAI_CODEX_BASE_URL",),
+                config_paths=("workerpals.llm.codex_base_url", "workerpals.openai_codex.base_url"),
+                default="",
+            ),
+            timeout_seconds_override=cfg.get_int(
+                env_names=("WORKERPALS_OPENAI_CODEX_TIMEOUT_S",),
+                config_paths=("workerpals.openai_codex.timeout_s",),
+                default=0,
+            ),
+            timeout_ms_top_level=cfg.get_int(
+                env_names=("WORKERPALS_OPENAI_CODEX_TIMEOUT_MS",),
+                config_paths=("workerpals.openai_codex_timeout_ms",),
+                default=0,
+            ),
+            timeout_ms_llm_codex=cfg.get_int(
+                env_names=("WORKERPALS_LLM_CODEX_TIMEOUT_MS",),
+                config_paths=("workerpals.llm.codex_timeout_ms",),
+                default=0,
+            ),
+            timeout_ms_backend=cfg.get_int(
+                env_names=("WORKERPALS_OPENAI_CODEX_BACKEND_TIMEOUT_MS",),
+                config_paths=("workerpals.openai_codex.timeout_ms",),
+                default=0,
+            ),
+            progress_log_interval_s=cfg.get_int(
+                env_names=("WORKERPALS_OPENAI_CODEX_PROGRESS_LOG_INTERVAL_S",),
+                config_paths=("workerpals.openai_codex.progress_log_interval_s",),
+                default=30,
+            ),
+            reasoning_effort=cfg.get_str(
+                env_names=("WORKERPALS_LLM_REASONING_EFFORT", "WORKERPALS_OPENAI_CODEX_REASONING_EFFORT"),
+                config_paths=("workerpals.llm.reasoning_effort", "workerpals.openai_codex.reasoning_effort"),
+                default="high",
+            ),
+            approval_policy=cfg.get_str(
+                env_names=("WORKERPALS_OPENAI_CODEX_APPROVAL_POLICY",),
+                config_paths=("workerpals.openai_codex.approval_policy",),
+                default="never",
+            ),
+            sandbox=cfg.get_str(
+                env_names=("WORKERPALS_OPENAI_CODEX_SANDBOX",),
+                config_paths=("workerpals.openai_codex.sandbox",),
+                default="workspace-write",
+            ),
+            color=cfg.get_str(
+                env_names=("WORKERPALS_OPENAI_CODEX_COLOR",),
+                config_paths=("workerpals.openai_codex.color",),
+                default="never",
+            ),
+            json_output=cfg.get_bool(
+                env_names=("WORKERPALS_OPENAI_CODEX_JSON",),
+                config_paths=("workerpals.openai_codex.json",),
+                default=False,
+            ),
+        )
+
+
+def shutil_which(binary: str) -> str:
+    return which(binary) or ""
+
+
+def _truncate(text: str, max_chars: int = 4000) -> str:
+    value = str(text or "")
+    if len(value) <= max_chars:
+        return value
+    return value[: max(1, max_chars - 15)] + "\n...[truncated]"
+
+
+def _repo_root_for_prompt_loading() -> Path:
+    current = Path(__file__).resolve()
+    for parent in current.parents:
+        if (parent / "prompts").is_dir():
+            return parent
+    # Fallback to historical layout depth if prompts/ cannot be discovered.
+    return current.parents[5]
+
+
+def _resolve_prompt_file(relative_path: str) -> Path:
+    return _repo_root_for_prompt_loading() / "prompts" / relative_path
+
+
+def _load_prompt_template(
+    relative_path: str, replacements: Optional[Dict[str, str]] = None
+) -> str:
+    prompt_path = _resolve_prompt_file(relative_path)
+    cache_key = str(prompt_path)
+    cached = _PROMPT_TEMPLATE_CACHE.get(cache_key)
+    if cached is not None:
+        template = cached
+    else:
+        try:
+            template = prompt_path.read_text(encoding="utf-8").strip()
+        except Exception:
+            template = ""
+        _PROMPT_TEMPLATE_CACHE[cache_key] = template
+    if not replacements:
+        return template
+
+    def _replace(match: re.Match[str]) -> str:
+        key = match.group(1)
+        value = replacements.get(key)
+        if value is None:
+            raise KeyError(f"Missing prompt replacement '{{{{{key}}}}}' for {prompt_path}")
+        return value
+
+    return _PROMPT_TOKEN_REGEX.sub(_replace, template)
+
+
+def _to_positive_int(raw: str) -> Optional[int]:
+    try:
+        parsed = int(raw)
+    except Exception:
+        return None
+    return parsed if parsed > 0 else None
+
+
+def _normalize_choice(
+    value: str,
+    valid: set[str],
+    default: str,
+    *,
+    env_name: str,
+) -> str:
+    normalized = value.strip().lower()
+    if normalized in valid:
+        return normalized
+    if normalized:
+        log.info(
+            f"Invalid {env_name}={value!r}; using default {default!r}. "
+            f"Allowed: {', '.join(sorted(valid))}."
+        )
+    return default
+
+
+def _is_git_repo(repo: str) -> bool:
+    try:
+        proc = subprocess.run(
+            ["git", "rev-parse", "--is-inside-work-tree"],
+            cwd=repo,
+            capture_output=True,
+            text=True,
+            timeout=10,
+            check=False,
+        )
+        if proc.returncode != 0:
+            return False
+        return (proc.stdout or "").strip().lower() == "true"
+    except Exception:
+        return False
+
+
+def _resolve_codex_command_prefix(config: OpenAICodexRuntimeConfig) -> List[str]:
+    override_json = config.codex_bin_json
+    if override_json:
+        try:
+            parsed = json.loads(override_json)
+            if isinstance(parsed, list):
+                parts = [str(p).strip() for p in parsed if str(p).strip()]
+                if parts:
+                    return parts
+        except Exception:
+            log.info(
+                "Invalid PUSHPALS_OPENAI_CODEX_BIN_JSON; expected JSON array of command segments."
+            )
+
+    override = config.codex_bin
+    if override:
+        try:
+            parts = [p for p in shlex.split(override) if p.strip()]
+        except Exception:
+            log.info(
+                "Invalid PUSHPALS_OPENAI_CODEX_BIN value; expected a command string parseable by shlex."
+            )
+            return []
+        return parts
+
+    # Prefer bunx to avoid requiring a separate node runtime in the container.
+    if shutil_which("bunx"):
+        return ["bunx", "--yes", "@openai/codex"]
+    if shutil_which("codex"):
+        return ["codex"]
+    return []
+
+
+def _resolve_communicate_timeout_seconds(config: OpenAICodexRuntimeConfig) -> Optional[int]:
+    explicit_s = _to_positive_int(str(config.timeout_seconds_override))
+    if explicit_s is not None:
+        return explicit_s
+    # Top-level execution budget (e.g. openai_codex_timeout_ms = 7200000 in [workerpals])
+    # takes precedence over the more granular LLM/CLI-level timeout settings.
+    top_level_ms = config.timeout_ms_top_level
+    if top_level_ms > 0:
+        return max(1, top_level_ms // 1000)
+    timeout_ms = config.timeout_ms_llm_codex
+    if timeout_ms <= 0:
+        timeout_ms = config.timeout_ms_backend
+    if timeout_ms <= 0:
+        return None
+    return max(1, timeout_ms // 1000)
+
+
+def _resolve_reasoning_effort(config: OpenAICodexRuntimeConfig) -> str:
+    raw = config.reasoning_effort
+    normalized = str(raw).strip().lower()
+    if normalized in _VALID_REASONING_EFFORTS:
+        return normalized
+    log.info(
+        "Invalid workerpals.openai_codex.reasoning_effort="
+        f"{raw!r}; using default 'high'. Allowed: low, medium, high."
+    )
+    return "high"
+
+
+def _resolve_progress_log_interval_seconds(config: OpenAICodexRuntimeConfig) -> int:
+    interval = to_int(config.progress_log_interval_s, 30)
+    # Avoid noisy logs (<30s) and stale logs (>120s).
+    return max(30, min(120, interval))
+
+
+def _normalize_auth_mode(raw: str) -> str:
+    lowered = (raw or "").strip().lower()
+    aliases = {
+        "apikey": "api_key",
+        "api": "api_key",
+        "api-key": "api_key",
+        "chatgpt_login": "chatgpt",
+        "chatgpt-pro": "chatgpt",
+        "subscription": "chatgpt",
+    }
+    normalized = aliases.get(lowered, lowered)
+    if normalized in _VALID_AUTH_MODES:
+        return normalized
+    if lowered:
+        log.info(
+            f"Invalid PUSHPALS_OPENAI_CODEX_AUTH_MODE={raw!r}; using default 'auto'. "
+            f"Allowed: {', '.join(sorted(_VALID_AUTH_MODES))}."
+        )
+    return "auto"
+
+
+def _run_codex_login_status(codex_cmd_prefix: List[str], repo: str, env: Dict[str, str]) -> Dict[str, Any]:
+    try:
+        proc = subprocess.run(
+            [*codex_cmd_prefix, "login", "status"],
+            cwd=repo,
+            env=env,
+            capture_output=True,
+            text=True,
+            encoding="utf-8",
+            errors="replace",
+            timeout=25,
+            check=False,
+        )
+        return {
+            "ok": proc.returncode == 0,
+            "exitCode": int(proc.returncode),
+            "stdout": proc.stdout or "",
+            "stderr": proc.stderr or "",
+        }
+    except Exception as exc:
+        return {
+            "ok": False,
+            "exitCode": 1,
+            "stdout": "",
+            "stderr": f"Failed to run `codex login status`: {exc}",
+        }
+
+
+def _terminate_active_child() -> None:
+    global _ACTIVE_CHILD
+    proc = _ACTIVE_CHILD
+    if proc is None or proc.poll() is not None:
+        return
+    try:
+        proc.terminate()
+    except Exception:
+        pass
+    try:
+        proc.wait(timeout=3)
+    except Exception:
+        try:
+            proc.kill()
+        except Exception:
+            pass
+
+
+def _truncate_inline(text: str, max_chars: int = 180) -> str:
+    value = " ".join(str(text or "").split())
+    if len(value) <= max_chars:
+        return value
+    return value[: max(1, max_chars - 3)] + "..."
+
+
+def _contains_reasoning_marker(value: str) -> bool:
+    lowered = str(value or "").strip().lower()
+    if not lowered:
+        return False
+    return "reasoning" in lowered or "thinking" in lowered
+
+
+def _event_contains_reasoning(value: Any) -> bool:
+    max_nodes = 256
+    visited = 0
+    stack: List[Any] = [value]
+    while stack and visited < max_nodes:
+        current = stack.pop()
+        visited += 1
+        if isinstance(current, str):
+            if _contains_reasoning_marker(current):
+                return True
+            continue
+        if isinstance(current, list):
+            for item in reversed(current[:80]):
+                if isinstance(item, (dict, list, str)):
+                    stack.append(item)
+            continue
+        if not isinstance(current, dict):
+            continue
+
+        for raw_key, nested in current.items():
+            key = str(raw_key or "")
+            key_lower = key.lower()
+            if _contains_reasoning_marker(key_lower):
+                return True
+            if key_lower in ("type", "kind", "event", "item_type", "role", "channel"):
+                if isinstance(nested, str) and _contains_reasoning_marker(nested):
+                    return True
+            if isinstance(nested, (dict, list, str)):
+                stack.append(nested)
+
+    return False
+
+
+def _collect_text_fragments(value: Any, out: List[str]) -> None:
+    if isinstance(value, str):
+        text = _truncate_inline(value, 220)
+        if text:
+            out.append(text)
+        return
+    if isinstance(value, list):
+        for item in value:
+            _collect_text_fragments(item, out)
+        return
+    if isinstance(value, dict):
+        matched_key = False
+        for raw_key, nested in value.items():
+            key = str(raw_key or "").lower()
+            if key.endswith("_text") or key.endswith("_message"):
+                matched_key = True
+                _collect_text_fragments(nested, out)
+                continue
+            if (
+                key in ("text", "content", "summary", "message", "error", "reason", "delta", "output", "item")
+                or _contains_reasoning_marker(key)
+            ):
+                matched_key = True
+                _collect_text_fragments(nested, out)
+        if not matched_key:
+            # Fallback: recurse into nested containers so unknown payload shapes still surface text.
+            for nested in value.values():
+                if isinstance(nested, (dict, list)):
+                    _collect_text_fragments(nested, out)
+        return
+
+
+def _summarize_json_event(obj: Dict[str, Any]) -> str:
+    event_type = str(obj.get("type") or obj.get("event") or obj.get("kind") or "event").strip()
+    if not event_type:
+        event_type = "event"
+    # Skip noisy streaming deltas unless they contain meaningful text fragments.
+    delta_like = event_type.endswith(".delta") or event_type.endswith("_delta")
+    # Reasoning/thinking events are always surfaced — they show the model's reasoning process.
+    reasoning_like = _contains_reasoning_marker(event_type) or _event_contains_reasoning(obj)
+
+    tool_name = ""
+    for key in ("tool_name", "tool", "name"):
+        raw = obj.get(key)
+        if isinstance(raw, str) and raw.strip():
+            tool_name = raw.strip()
+            break
+        if isinstance(raw, dict):
+            nested = raw.get("name")
+            if isinstance(nested, str) and nested.strip():
+                tool_name = nested.strip()
+                break
+    # For Codex CLI item.* events, the tool/function name is nested under obj["item"]["name"].
+    if not tool_name and isinstance(obj.get("item"), dict):
+        nested = obj["item"].get("name")
+        if isinstance(nested, str) and nested.strip():
+            tool_name = nested.strip()
+
+    fragments: List[str] = []
+    # "item" covers Codex CLI's item.started/updated/completed events where reasoning and
+    # tool call content is nested under the item object.
+    # "output" covers turn.completed and similar events that carry output arrays.
+    # "delta" covers reasoning delta events (response.reasoning_summary_text.delta).
+    extract_keys = ["message", "text", "summary", "content", "output_text", "error", "item", "output", "delta"]
+    for key in extract_keys:
+        if key in obj:
+            _collect_text_fragments(obj.get(key), fragments)
+    deduped: List[str] = []
+    seen: set[str] = set()
+    for frag in fragments:
+        if frag in seen:
+            continue
+        seen.add(frag)
+        deduped.append(frag)
+    text_part = deduped[0] if deduped else ""
+
+    # Suppress noisy deltas, but always surface reasoning events even if text is empty.
+    if delta_like and not text_part and not reasoning_like:
+        return ""
+
+    parts = [event_type]
+    if tool_name:
+        parts.append(f"tool={tool_name}")
+    if text_part:
+        parts.append(text_part)
+    elif reasoning_like:
+        parts.append("reasoning update")
+    return " | ".join(parts)
+
+
+def _format_codex_trace_excerpt(trace: Dict[str, Any], max_items: int = 20) -> str:
+    summaries = trace.get("summaries")
+    if isinstance(summaries, list):
+        items = [str(item).strip() for item in summaries if str(item).strip()]
+        if items:
+            shown = items[:max_items]
+            lines = [f"- {item}" for item in shown]
+            omitted = len(items) - len(shown)
+            if omitted > 0:
+                lines.append(f"- ... ({omitted} more event(s) omitted)")
+            return "Codex event trace:\n" + "\n".join(lines)
+
+    event_counts = trace.get("event_type_counts")
+    if isinstance(event_counts, dict):
+        pairs = [
+            (str(key).strip() or "event", to_int(value, 0))
+            for key, value in event_counts.items()
+            if to_int(value, 0) > 0
+        ]
+        if pairs:
+            pairs.sort(key=lambda item: item[1], reverse=True)
+            listed = ", ".join(f"{name}={count}" for name, count in pairs[:8])
+            return f"Codex event types: {listed}"
+
+    return ""
+
+
+def _empty_codex_trace() -> Dict[str, Any]:
+    return {
+        "line_count": 0,
+        "valid_json": 0,
+        "invalid_json": 0,
+        "summaries": [],
+        "event_type_counts": {},
+        "live_logged": 0,
+        "live_omitted": 0,
+        "raw_logged": 0,
+        "raw_omitted": 0,
+        "reasoning_events": 0,
+    }
+
+
+def _record_live_codex_stdout_line(line: str, use_json: bool, trace: Dict[str, Any]) -> None:
+    stripped = line.strip()
+    if not stripped:
+        return
+
+    trace["line_count"] = to_int(trace.get("line_count"), 0) + 1
+    summaries = trace.setdefault("summaries", [])
+    event_type_counts = trace.setdefault("event_type_counts", {})
+    max_recorded_summaries = 500
+    max_live_logged = 300
+    max_raw_logged = 5
+
+    if use_json:
+        try:
+            parsed = json.loads(stripped)
+            trace["valid_json"] = to_int(trace.get("valid_json"), 0) + 1
+        except Exception:
+            trace["invalid_json"] = to_int(trace.get("invalid_json"), 0) + 1
+            raw_logged = to_int(trace.get("raw_logged"), 0)
+            if raw_logged < max_raw_logged:
+                log.info(f"[codex/raw] {_truncate_inline(stripped, 220)}")
+                trace["raw_logged"] = raw_logged + 1
+            else:
+                trace["raw_omitted"] = to_int(trace.get("raw_omitted"), 0) + 1
+            return
+
+        if isinstance(parsed, dict):
+            event_type = (
+                str(parsed.get("type") or parsed.get("event") or parsed.get("kind") or "event")
+                .strip()
+                or "event"
+            )
+            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).
+            priority = _event_contains_reasoning(parsed)
+            if priority:
+                trace["reasoning_events"] = to_int(trace.get("reasoning_events"), 0) + 1
+            if summary:
+                if len(summaries) < max_recorded_summaries:
+                    summaries.append(summary)
+                live_logged = to_int(trace.get("live_logged"), 0)
+                if live_logged < max_live_logged or priority:
+                    log.info(f"[codex] {summary}")
+                    trace["live_logged"] = live_logged + 1
+                else:
+                    trace["live_omitted"] = to_int(trace.get("live_omitted"), 0) + 1
+        return
+
+    summary = _truncate_inline(stripped, 220)
+    if summary:
+        if len(summaries) < max_recorded_summaries:
+            summaries.append(summary)
+        live_logged = to_int(trace.get("live_logged"), 0)
+        if live_logged < max_live_logged:
+            log.info(f"[codex] {summary}")
+            trace["live_logged"] = live_logged + 1
+        else:
+            trace["live_omitted"] = to_int(trace.get("live_omitted"), 0) + 1
+
+
+def _finalize_codex_stdout_trace(trace: Dict[str, Any], use_json: bool) -> Dict[str, Any]:
+    line_count = to_int(trace.get("line_count"), 0)
+    valid_json = to_int(trace.get("valid_json"), 0)
+    invalid_json = to_int(trace.get("invalid_json"), 0)
+    summaries = trace.get("summaries")
+    if not isinstance(summaries, list):
+        summaries = []
+    else:
+        summaries = [str(item).strip() for item in summaries if str(item).strip()]
+    event_type_counts_raw = trace.get("event_type_counts")
+    event_type_counts: Dict[str, int] = {}
+    if isinstance(event_type_counts_raw, dict):
+        for key, value in event_type_counts_raw.items():
+            name = str(key).strip() or "event"
+            count = to_int(value, 0)
+            if count > 0:
+                event_type_counts[name] = count
+
+    if use_json:
+        log.info(
+            f"Codex JSON stream captured ({line_count} line(s), valid_json={valid_json}, invalid={invalid_json})."
+        )
+    else:
+        log.info(f"Codex stdout captured ({line_count} non-empty line(s)).")
+
+    live_omitted = to_int(trace.get("live_omitted"), 0)
+    if live_omitted > 0:
+        log.info(f"[codex] ... {live_omitted} additional event(s) omitted.")
+    raw_omitted = to_int(trace.get("raw_omitted"), 0)
+    if raw_omitted > 0:
+        log.info(f"[codex/raw] ... {raw_omitted} additional line(s) omitted.")
+    reasoning_events = to_int(trace.get("reasoning_events"), 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 not summaries and event_type_counts:
+        ranked = sorted(event_type_counts.items(), key=lambda item: item[1], reverse=True)
+        top = ", ".join(f"{name}={count}" for name, count in ranked[:8])
+        log.info(f"[codex] Event types: {top}")
+
+    return {
+        "line_count": line_count,
+        "valid_json": valid_json,
+        "invalid_json": invalid_json,
+        "summaries": summaries,
+        "event_type_counts": event_type_counts,
+        "reasoning_events": reasoning_events,
+    }
+
+
+def _log_stderr(stderr: str) -> None:
+    lines = [line.strip() for line in stderr.splitlines() if line.strip()]
+    if not lines:
+        return
+    max_lines = 20
+    for line in lines[:max_lines]:
+        log.info(f"[stderr] {line}")
+    if len(lines) > max_lines:
+        log.info(f"[stderr] ... {len(lines) - max_lines} additional line(s) omitted.")
+
+
+def _safe_model_for_codex(raw_model: str, base_url: str) -> str:
+    model = str(raw_model or "").strip()
+    if not model:
+        return DEFAULT_CODEX_MODEL
+    if "/" not in model:
+        return model
+    provider, bare = model.split("/", 1)
+    provider = provider.strip().lower()
+    bare = bare.strip()
+    if provider == "openai" and bare:
+        return bare
+    if looks_local_base_url(base_url) and bare:
+        return bare
+    return DEFAULT_CODEX_MODEL
+
+
+def _build_instruction(instruction: str, supplemental_guidance: List[str]) -> str:
+    system_prompt = (_load_prompt_template(_TASK_SYSTEM_PROMPT_PATH) or "").strip()
+    if not system_prompt:
+        system_prompt = (_load_prompt_template(_DEFAULT_TASK_SYSTEM_PROMPT_PATH) or "").strip()
+    if not system_prompt:
+        raise RuntimeError(
+            "Missing required OpenAI Codex system prompt template. "
+            f"Expected one of: {_TASK_SYSTEM_PROMPT_PATH}, {_DEFAULT_TASK_SYSTEM_PROMPT_PATH}"
+        )
+
+    runtime_policy_appendix = (
+        _load_prompt_template(_MANDATORY_RUNTIME_POLICY_APPENDIX_PATH) or ""
+    ).strip()
+    if not runtime_policy_appendix:
+        raise RuntimeError(
+            "Missing required OpenAI Codex runtime policy appendix template. "
+            f"Expected: {_MANDATORY_RUNTIME_POLICY_APPENDIX_PATH}"
+        )
+    if runtime_policy_appendix.lower() not in system_prompt.lower():
+        system_prompt = f"{system_prompt}\n\n{runtime_policy_appendix}".strip()
+
+    supplemental_section = ""
+    filtered_guidance = [str(item).strip() for item in supplemental_guidance if str(item).strip()]
+    if filtered_guidance:
+        supplemental_section_template = _load_prompt_template(_SUPPLEMENTAL_GUIDANCE_SECTION_PATH)
+        if not supplemental_section_template.strip():
+            raise RuntimeError(
+                "Missing required OpenAI Codex supplemental guidance section template. "
+                f"Expected: {_SUPPLEMENTAL_GUIDANCE_SECTION_PATH}"
+            )
+        supplemental_section = "\n\n" + _load_prompt_template(
+            _SUPPLEMENTAL_GUIDANCE_SECTION_PATH,
+            {"guidance_lines": "\n".join(filtered_guidance)},
+        ).strip()
+
+    wrapped = _load_prompt_template(
+        _INSTRUCTION_WRAPPER_PROMPT_PATH,
+        {
+            "system_prompt": system_prompt,
+            "instruction": instruction,
+            "supplemental_section": supplemental_section,
+        },
+    )
+    if not wrapped.strip():
+        raise RuntimeError(
+            "Missing required OpenAI Codex instruction wrapper template. "
+            f"Expected: {_INSTRUCTION_WRAPPER_PROMPT_PATH}"
+        )
+    return wrapped.strip()
+
+
+def _detect_codex_workaround_signal(*texts: str) -> Optional[str]:
+    for text in texts:
+        source = str(text or "")
+        if not source:
+            continue
+        for pattern in _CODEX_WORKAROUND_PATTERNS:
+            for match in pattern.finditer(source):
+                snippet = match.group(0).strip()
+                if not snippet:
+                    continue
+                lowered = snippet.lower()
+                if any(hint in lowered for hint in _CODEX_WORKAROUND_NEGATION_HINTS):
+                    continue
+                return snippet
+    return None
+
+
+def _read_text_if_exists(path: Path) -> str:
+    try:
+        if not path.exists():
+            return ""
+        return path.read_text(encoding="utf-8", errors="replace").strip()
+    except Exception:
+        return ""
+
+
+def _terminate_child(signum: int, _frame: Any) -> None:
+    global _INTERRUPTED_SIGNAL
+    _INTERRUPTED_SIGNAL = signum
+    _terminate_active_child()
+
+
+def _install_signal_handlers() -> None:
+    for sig in (signal.SIGTERM, signal.SIGINT):
+        try:
+            signal.signal(sig, _terminate_child)
+        except Exception:
+            pass
+
+
+def _run_codex_task(
+    repo: str,
+    instruction: str,
+    supplemental_guidance: List[str],
+) -> Dict[str, Any]:
+    global _ACTIVE_CHILD, _INTERRUPTED_SIGNAL
+    _INTERRUPTED_SIGNAL = None
+    _install_signal_handlers()
+
+    if not _is_git_repo(repo):
+        return {
+            "ok": False,
+            "summary": "openai_codex requires a git repository",
+            "stderr": (
+                f"Refusing to run codex in a non-git directory: {repo}. "
+                "Validate repo/worktree setup before dispatching this backend."
+            ),
+            "exitCode": 2,
+        }
+
+    runtime_config = OpenAICodexRuntimeConfig.from_sources()
+    codex_cmd_prefix = _resolve_codex_command_prefix(runtime_config)
+    if not codex_cmd_prefix:
+        return {
+            "ok": False,
+            "summary": "openai_codex CLI is not installed",
+            "stderr": (
+                "Could not find a runnable Codex command. "
+                "Expected one of: `bunx --yes @openai/codex` or `codex` in PATH. "
+                "You can also set PUSHPALS_OPENAI_CODEX_BIN explicitly."
+            ),
+            "exitCode": 3,
+        }
+
+    configured_model, api_key, base_url = resolve_llm_config(DEFAULT_CODEX_MODEL, logger=log)
+    auth_mode_raw = runtime_config.auth_mode
+    auth_mode_configured = _normalize_auth_mode(auth_mode_raw)
+    model = _safe_model_for_codex(configured_model, base_url)
+    approval = _normalize_choice(
+        runtime_config.approval_policy,
+        _VALID_APPROVAL_POLICIES,
+        "never",
+        env_name="workerpals.openai_codex.approval_policy",
+    )
+    sandbox = _normalize_choice(
+        runtime_config.sandbox,
+        _VALID_SANDBOX_POLICIES,
+        "workspace-write",
+        env_name="workerpals.openai_codex.sandbox",
+    )
+    color = _normalize_choice(
+        runtime_config.color,
+        _VALID_COLORS,
+        "never",
+        env_name="workerpals.openai_codex.color",
+    )
+    # JSON event output is noisy by default; prefer plain text + output-last-message.
+    use_json = runtime_config.json_output
+    reasoning_effort = _resolve_reasoning_effort(runtime_config)
+    communicate_timeout_s = _resolve_communicate_timeout_seconds(runtime_config)
+    prompt = _build_instruction(instruction, supplemental_guidance)
+    baseline_changes = summarize_git_changes(repo)
+
+    with tempfile.TemporaryDirectory(prefix="pushpals-codex-") as tmp_dir:
+        last_message_path = Path(tmp_dir) / "codex-last-message.txt"
+        cmd: List[str] = [
+            *codex_cmd_prefix,
+            "-c",
+            f'model_reasoning_effort="{reasoning_effort}"',
+            "-a",
+            approval,
+            "exec",
+            "-s",
+            sandbox,
+            "--color",
+            color,
+            "--output-last-message",
+            str(last_message_path),
+        ]
+        if use_json:
+            cmd.append("--json")
+        if model:
+            cmd.extend(["-m", model])
+        cmd.append("-")
+
+        env = os.environ.copy()
+        env["PYTHONIOENCODING"] = "utf-8"
+        env["PUSHPALS_REPO_PATH"] = repo
+        env["PUSHPALS_ASSIGNED_REPO_ROOT"] = repo
+        existing_openai_key = (env.get("OPENAI_API_KEY") or "").strip()
+        llm_key = api_key.strip()
+        if llm_key.lower() == "lmstudio":
+            llm_key = ""
+        auth_mode = auth_mode_configured
+        if auth_mode == "auto":
+            auth_mode = "api_key" if (llm_key or existing_openai_key) else "chatgpt"
+        log.info(f"Codex auth mode: {auth_mode} (configured={auth_mode_configured})")
+
+        existing_openai_base = (
+            env.get("OPENAI_BASE_URL", "").strip() or env.get("OPENAI_API_BASE", "").strip()
+        )
+        override_base = runtime_config.base_url_override
+        effective_base = ""
+
+        if auth_mode == "chatgpt":
+            if llm_key or existing_openai_key:
+                log.info(
+                    "ChatGPT auth mode selected; ignoring OPENAI_API_KEY to use Codex CLI login credentials."
+                )
+            if override_base or existing_openai_base or base_url:
+                log.info("ChatGPT auth mode selected; ignoring OPENAI_BASE_URL/OPENAI_API_BASE overrides.")
+            env.pop("OPENAI_API_KEY", None)
+            env.pop("OPENAI_BASE_URL", None)
+            env.pop("OPENAI_API_BASE", None)
+            login_status = _run_codex_login_status(codex_cmd_prefix, repo, env)
+            if not login_status.get("ok"):
+                detail = (
+                    str(login_status.get("stderr") or "").strip()
+                    or str(login_status.get("stdout") or "").strip()
+                    or "codex login status returned non-zero"
+                )
+                return {
+                    "ok": False,
+                    "summary": "openai_codex chatgpt auth is not ready",
+                    "stdout": _truncate(str(login_status.get("stdout") or "")),
+                    "stderr": _truncate(
+                        "Codex CLI is not logged in for ChatGPT subscription mode. "
+                        "Run `bunx --yes @openai/codex login` on the host (no global install needed), "
+                        "complete browser sign-in, then retry.\n"
+                        f"Details: {detail}"
+                    ),
+                    "exitCode": int(login_status.get("exitCode") or 1),
+                }
+        else:
+            final_key = llm_key or existing_openai_key
+            if not final_key:
+                return {
+                    "ok": False,
+                    "summary": "openai_codex api_key auth requires OPENAI_API_KEY",
+                    "stderr": (
+                        "API-key auth mode selected, but no API key is available. "
+                        "Set OPENAI_API_KEY (or WORKERPALS_LLM_API_KEY), "
+                        "or set PUSHPALS_OPENAI_CODEX_AUTH_MODE=chatgpt."
+                    ),
+                    "exitCode": 2,
+                }
+            env["OPENAI_API_KEY"] = final_key
+            effective_base = override_base or base_url
+            if (
+                not override_base
+                and not existing_openai_base
+                and looks_local_base_url(base_url)
+                and (env.get("OPENAI_API_KEY") or "").strip()
+            ):
+                # If an OpenAI key exists but base URL came from local worker LLM config,
+                # prefer Codex/OpenAI defaults unless explicitly overridden.
+                log.info(
+                    "Detected local worker LLM endpoint with OPENAI_API_KEY present; "
+                    "using Codex default OpenAI endpoint (set PUSHPALS_OPENAI_CODEX_BASE_URL "
+                    "to force local)."
+                )
+                effective_base = ""
+            if effective_base:
+                env["OPENAI_BASE_URL"] = effective_base
+                env["OPENAI_API_BASE"] = effective_base
+            else:
+                env.pop("OPENAI_BASE_URL", None)
+                env.pop("OPENAI_API_BASE", None)
+
+        log.info(f"Starting codex exec in {repo}")
+        log.debug(f"Codex command: {' '.join(codex_cmd_prefix)}")
+        log.debug(f"Model: {model}")
+        base_for_log = (
+            env.get("OPENAI_BASE_URL", "").strip()
+            or env.get("OPENAI_API_BASE", "").strip()
+            or "<default>"
+        )
+        log.debug(f"Base URL: {base_for_log}")
+        if communicate_timeout_s:
+            log.debug(f"communicate timeout: {communicate_timeout_s}s")
+
+        proc = subprocess.Popen(
+            cmd,
+            cwd=repo,
+            env=env,
+            stdout=subprocess.PIPE,
+            stderr=subprocess.PIPE,
+            stdin=subprocess.PIPE,
+            text=True,
+            encoding="utf-8",
+            errors="replace",
+        )
+        _ACTIVE_CHILD = proc
+        started_at = time.monotonic()
+        progress_interval_s = _resolve_progress_log_interval_seconds(runtime_config)
+
+        stdout_chunks: List[str] = []
+        stderr_chunks: List[str] = []
+        stdout_trace_state = _empty_codex_trace()
+        trace_lock = threading.Lock()
+        last_activity_at = {"ts": started_at}
+
+        def _drain_stdout() -> None:
+            stream = proc.stdout
+            if stream is None:
+                return
+            try:
+                for chunk in iter(stream.readline, ""):
+                    if chunk == "":
+                        break
+                    stdout_chunks.append(chunk)
+                    line = chunk.strip()
+                    if not line:
+                        continue
+                    with trace_lock:
+                        last_activity_at["ts"] = time.monotonic()
+                        _record_live_codex_stdout_line(line, use_json, stdout_trace_state)
+            except Exception:
+                pass
+            finally:
+                try:
+                    stream.close()
+                except Exception:
+                    pass
+
+        def _drain_stderr() -> None:
+            stream = proc.stderr
+            if stream is None:
+                return
+            try:
+                for chunk in iter(stream.readline, ""):
+                    if chunk == "":
+                        break
+                    stderr_chunks.append(chunk)
+            except Exception:
+                pass
+            finally:
+                try:
+                    stream.close()
+                except Exception:
+                    pass
+
+        stdout_thread = threading.Thread(target=_drain_stdout, daemon=True)
+        stderr_thread = threading.Thread(target=_drain_stderr, daemon=True)
+        stdout_thread.start()
+        stderr_thread.start()
+
+        if proc.stdin is not None:
+            try:
+                proc.stdin.write(prompt)
+                proc.stdin.close()
+            except Exception:
+                pass
+
+        deadline = (
+            started_at + float(communicate_timeout_s)
+            if communicate_timeout_s and communicate_timeout_s > 0
+            else None
+        )
+        next_progress_at = started_at + float(progress_interval_s)
+        timed_out = False
+
+        while proc.poll() is None:
+            now = time.monotonic()
+            if deadline is not None and now >= deadline:
+                timed_out = True
+                _terminate_active_child()
+                break
+
+            if now >= next_progress_at:
+                elapsed = int(max(0.0, now - started_at))
+                with trace_lock:
+                    last_event = float(last_activity_at.get("ts", started_at))
+                    valid_json = to_int(stdout_trace_state.get("valid_json"), 0)
+                    total_lines = to_int(stdout_trace_state.get("line_count"), 0)
+                idle_for = int(max(0.0, now - last_event))
+                if use_json:
+                    log.info(
+                        f"codex exec still running ({elapsed}s elapsed, json_events={valid_json}, idle={idle_for}s)"
+                    )
+                else:
+                    log.info(
+                        f"codex exec still running ({elapsed}s elapsed, stdout_lines={total_lines}, idle={idle_for}s)"
+                    )
+                next_progress_at = now + float(progress_interval_s)
+
+            time.sleep(1.0)
+
+        try:
+            proc.wait(timeout=5)
+        except Exception:
+            try:
+                proc.kill()
+                proc.wait(timeout=5)
+            except Exception:
+                pass
+
+        stdout_thread.join(timeout=2)
+        stderr_thread.join(timeout=2)
+
+        return_code = proc.returncode
+        _ACTIVE_CHILD = None
+        elapsed_total = int(max(0.0, time.monotonic() - started_at))
+        log.info(f"codex exec finished in {elapsed_total}s")
+
+        stdout = "".join(stdout_chunks)
+        stderr = "".join(stderr_chunks)
+        stdout_trace = _finalize_codex_stdout_trace(stdout_trace_state, use_json)
+        trace_excerpt = _format_codex_trace_excerpt(stdout_trace)
+        _log_stderr(stderr)
+
+        if timed_out:
+            detail = (
+                f"codex exec timed out after {communicate_timeout_s}s"
+                if communicate_timeout_s
+                else "codex exec timed out"
+            )
+            if trace_excerpt:
+                detail = f"{detail}\n{trace_excerpt}"
+            return {
+                "ok": False,
+                "summary": "openai_codex execution timed out",
+                "stdout": _truncate(stdout),
+                "stderr": _truncate(f"{detail}\n{stderr}".strip()),
+                "exitCode": 124,
+            }
+
+        last_message = _read_text_if_exists(last_message_path)
+        log_git_status(repo, log)
+
+        if _INTERRUPTED_SIGNAL is not None:
+            return {
+                "ok": False,
+                "summary": f"openai_codex interrupted by signal {_INTERRUPTED_SIGNAL}",
+                "stdout": _truncate(stdout),
+                "stderr": _truncate(stderr),
+                "exitCode": 128 + int(_INTERRUPTED_SIGNAL),
+            }
+
+        if return_code is None:
+            return {
+                "ok": False,
+                "summary": "openai_codex execution ended without a process return code",
+                "stdout": _truncate(stdout),
+                "stderr": _truncate(stderr),
+                "exitCode": 1,
+            }
+
+        exit_code = int(return_code)
+
+        if exit_code != 0:
+            detail = stderr.strip() or stdout.strip() or "codex exec exited with a non-zero status"
+            if last_message:
+                detail = f"{detail}\nLast assistant message:\n{last_message}"
+            if trace_excerpt:
+                detail = f"{detail}\n{trace_excerpt}"
+            return {
+                "ok": False,
+                "summary": f"openai_codex execution failed (exit {exit_code})",
+                "stdout": _truncate(stdout),
+                "stderr": _truncate(detail),
+                "exitCode": exit_code,
+            }
+
+        policy_signal = _detect_codex_workaround_signal(last_message)
+        if not policy_signal and not last_message.strip():
+            # Fallback only when the CLI did not emit a final assistant message.
+            policy_signal = _detect_codex_workaround_signal(stdout)
+        if policy_signal:
+            detail = (
+                "Codex CLI is mandatory in this backend, but worker output suggests a workaround "
+                f"instead of hard-failing: {policy_signal!r}. "
+                "Return an explicit failure if Codex auth/execution is unavailable."
+            )
+            if last_message:
+                detail = f"{detail}\nLast assistant message:\n{last_message}"
+            if trace_excerpt:
+                detail = f"{detail}\n{trace_excerpt}"
+            return {
+                "ok": False,
+                "summary": "openai_codex policy violation: Codex CLI workaround detected",
+                "stdout": _truncate(stdout),
+                "stderr": _truncate(detail),
+                "exitCode": 5,
+            }
+
+        changed_paths = summarize_git_changes(repo)
+        delta = [p for p in changed_paths if p not in baseline_changes]
+        effective = delta if delta else changed_paths
+        stdout_parts: List[str] = []
+        if last_message:
+            stdout_parts.append(last_message)
+        elif trace_excerpt:
+            stdout_parts.append(trace_excerpt)
+        if effective:
+            listed = "\n".join(f"- {path}" for path in effective[:40])
+            if len(effective) > 40:
+                listed += "\n- ..."
+            stdout_parts.append(f"Changed files:\n{listed}")
+            return {
+                "ok": True,
+                "summary": f"Executed task and modified {len(effective)} file(s)",
+                "stdout": "\n\n".join(stdout_parts),
+                "stderr": "",
+                "exitCode": 0,
+            }
+
+        if not stdout_parts:
+            stdout_parts.append("No modified files were detected after execution.")
+        return {
+            "ok": True,
+            "summary": "Executed task via openai_codex (no file changes detected)",
+            "stdout": "\n\n".join(stdout_parts),
+            "stderr": "",
+            "exitCode": 0,
+        }
+
+
+def main() -> int:
+    try:
+        task = parse_task_execute_payload(sys.argv, logger=log)
+        result = _run_codex_task(
+            task.repo,
+            task.instruction,
+            task.supplemental_guidance,
+        )
+    except Exception as exc:
+        result = {
+            "ok": False,
+            "summary": "openai_codex wrapper crashed while executing task.execute",
+            "stdout": "",
+            "stderr": traceback.format_exc(),
+            "exitCode": 1,
+            "error": to_single_line(exc, 300),
+        }
+
+    emit(result)
+    return 0 if bool(result.get("ok")) else to_int(result.get("exitCode"), 1)
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())