fix(codex): pipe large prompts via stdin to avoid ARG_MAX

caw/providers/codex.py

@@ -5,6 +5,7 @@
 import atexit
 import json
 import logging
+import os
 import re
 import subprocess
 import threading
@@ -131,6 +132,21 @@ def _cleanup_processes() -> None:
 atexit.register(_cleanup_processes)
 
 
+# -- Prompt argv / stdin ------------------------------------------------------
+
+# Linux ARG_MAX includes the process environment, so keep a conservative
+# threshold well below SC_ARG_MAX. Large prompts (e.g. CodeWiki repository
+# overviews) must use ``codex exec … -`` and stdin.
+_PROMPT_STDIN_THRESHOLD_BYTES = 100_000
+
+
+def _prompt_exceeds_argv_budget(cmd: list[str], prompt: str, *, threshold: int = _PROMPT_STDIN_THRESHOLD_BYTES) -> bool:
+    """Return True when placing ``prompt`` on argv would risk E2BIG."""
+    cmd_bytes = sum(len(part.encode("utf-8", "replace")) + 1 for part in cmd)
+    prompt_bytes = len(prompt.encode("utf-8", "replace"))
+    return (cmd_bytes + prompt_bytes) > threshold
+
+
 # -- MCP config helpers -------------------------------------------------------
 
 
@@ -255,8 +271,13 @@ def send(self, message: str) -> Turn:
 
         cmd += self._mcp_config_args()
 
-        # Prompt as positional arg (last)
-        cmd.append(prompt)
+        # Prompt as positional arg, or stdin ("-") when argv would exceed ARG_MAX.
+        if _prompt_exceeds_argv_budget(cmd, prompt):
+            cmd.append("-")
+            stdin_data = prompt
+        else:
+            cmd.append(prompt)
+            stdin_data = None
 
         # Accumulated state for event processing
         blocks: list[ContentBlock] = []
@@ -267,7 +288,7 @@ def send(self, message: str) -> Turn:
         try:
             proc = subprocess.Popen(
                 cmd,
-                stdin=subprocess.DEVNULL,
+                stdin=subprocess.PIPE if stdin_data is not None else subprocess.DEVNULL,
                 stdout=subprocess.PIPE,
                 stderr=subprocess.PIPE,
                 text=True,
@@ -277,6 +298,9 @@ def send(self, message: str) -> Turn:
 
         _register_process(proc)
         try:
+            if stdin_data is not None and proc.stdin is not None:
+                proc.stdin.write(stdin_data)
+                proc.stdin.close()
             # Stream stdout line by line
             for line in proc.stdout:  # type: ignore[union-attr]
                 line = line.rstrip("\n")

tests/test_codex_large_prompt_stdin.py

@@ -0,0 +1,80 @@
+"""Large prompts must go through stdin to avoid Linux ARG_MAX / E2BIG."""
+
+from __future__ import annotations
+
+import io
+from unittest.mock import MagicMock, patch
+
+from caw.providers.codex import (
+    CodexSession,
+    _PROMPT_STDIN_THRESHOLD_BYTES,
+    _prompt_exceeds_argv_budget,
+)
+
+
+class TestPromptArgvBudget:
+    def test_small_prompt_fits_on_argv(self):
+        assert _prompt_exceeds_argv_budget(["codex", "exec"], "hello") is False
+
+    def test_large_prompt_requires_stdin(self):
+        prompt = "x" * (_PROMPT_STDIN_THRESHOLD_BYTES + 1)
+        assert _prompt_exceeds_argv_budget(["codex", "exec"], prompt) is True
+
+    def test_custom_threshold(self):
+        assert _prompt_exceeds_argv_budget(["codex"], "abcd", threshold=3) is True
+        assert _prompt_exceeds_argv_budget(["codex"], "ab", threshold=100) is False
+
+
+class TestCodexSessionSendStdin:
+    def test_send_pipes_large_prompt_via_stdin(self):
+        session = CodexSession(mcp_servers=[])
+        large = "P" * (_PROMPT_STDIN_THRESHOLD_BYTES + 50)
+
+        stdout = io.StringIO(
+            '{"type":"thread.started","thread_id":"t1"}\n'
+            '{"type":"item.completed","item":{"type":"agent_message","text":"ok"}}\n'
+            '{"type":"turn.completed","usage":{"input_tokens":1,"cached_input_tokens":0,"output_tokens":1}}\n'
+        )
+        proc = MagicMock()
+        proc.stdin = MagicMock()
+        proc.stdout = stdout
+        proc.stderr = io.StringIO("")
+        proc.returncode = 0
+
+        with patch("caw.providers.codex.subprocess.Popen", return_value=proc) as popen:
+            turn = session.send(large)
+
+        cmd = popen.call_args.args[0]
+        kwargs = popen.call_args.kwargs
+        assert cmd[-1] == "-"
+        assert kwargs["stdin"] is not None  # PIPE, not DEVNULL
+        proc.stdin.write.assert_called_once_with(large)
+        proc.stdin.close.assert_called_once()
+        assert "ok" in turn.result
+
+    def test_send_keeps_small_prompt_on_argv(self):
+        from subprocess import DEVNULL
+
+        session = CodexSession(mcp_servers=[])
+        small = "short prompt"
+
+        stdout = io.StringIO(
+            '{"type":"thread.started","thread_id":"t1"}\n'
+            '{"type":"item.completed","item":{"type":"agent_message","text":"ok"}}\n'
+            '{"type":"turn.completed","usage":{"input_tokens":1,"cached_input_tokens":0,"output_tokens":1}}\n'
+        )
+        proc = MagicMock()
+        proc.stdin = MagicMock()
+        proc.stdout = stdout
+        proc.stderr = io.StringIO("")
+        proc.returncode = 0
+
+        with patch("caw.providers.codex.subprocess.Popen", return_value=proc) as popen:
+            session.send(small)
+
+        cmd = popen.call_args.args[0]
+        kwargs = popen.call_args.kwargs
+        assert cmd[-1] == small
+        assert kwargs["stdin"] is DEVNULL
+        proc.stdin.write.assert_not_called()
+        proc.stdin.close.assert_not_called()