fix: revert -p removal — Claude -p mode has full tool access

Claude -p (print mode) is non-interactive but retains full tool access
(Edit, Write, Bash, etc.) with --dangerously-skip-permissions. Removing
-p caused Claude to enter interactive mode which requires a TTY and
produces zero output when run as a subprocess with piped I/O.

Now delivers prompt via stdin for both Claude and Codex in agentic mode.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

cross_eval/agent.py

@@ -28,6 +28,12 @@ _NO_CHANGE_ACK_MARKERS = (
     "no modifications were necessary",
     "no update was necessary",
     "already satisfied",
+    "no changes needed",
+    "no fixes needed",
+    "everything is correct",
+    "code is correct as-is",
+    "already correct",
+    "no action required",
 )
 _CHANGE_CLAIM_MARKERS = (
     "summary of all changes made",
@@ -59,6 +65,14 @@ _CHANGE_CLAIM_MARKERS = (
     "wrote the code",
     "refactored",
     "i refactored",
+    "completed all the changes",
+    "finished implementing",
+    "all tasks completed",
+    "done with the implementation",
+    "successfully implemented",
+    "completed the implementation",
+    "all changes have been made",
+    "changes are complete",
 )
 
 
@@ -419,11 +433,10 @@ def invoke_agent_agentic(
     if agent.reasoning_effort and _supports_reasoning_effort(agent.command):
         cmd.extend(["-c", f'model_reasoning_effort="{agent.reasoning_effort}"'])
 
-    # Strip print-mode flags and stdin sentinel from args for agentic mode.
-    # -p / --print makes Claude a one-shot text completer that cannot use tools
-    # or modify files, which defeats the entire purpose of agentic execution.
-    _STRIP_FOR_AGENTIC = {"-", "-p", "--print"}
-    args = [a for a in agent.args if a not in _STRIP_FOR_AGENTIC]
+    # Strip stdin sentinel ("-") from args for agentic mode.
+    # Keep -p/--print: Claude -p mode still has full tool access (Edit, Write,
+    # Bash, etc.) and is the correct mode for non-interactive subprocess use.
+    args = [a for a in agent.args if a != "-"]
     cmd.extend(args)
 
     # System prompt via flag if supported
@@ -441,16 +454,9 @@ def invoke_agent_agentic(
         else:
             input_data = prompt
     else:
-        # claude: use positional arg with a pointer to the task file
-        # (avoids OS arg length limits for large prompts).
-        # Send empty stdin so Claude receives EOF and exits after processing
-        # instead of hanging in interactive mode waiting for more input.
-        input_data = ""
-        cmd.append(
-            f"Read the task file at {task_file} and execute all instructions in it. "
-            f"Work only inside the current directory and do not modify files "
-            f"outside it."
-        )
+        # claude -p: deliver prompt via stdin (same as codex).
+        # -p mode is non-interactive and reads from stdin, then exits.
+        input_data = prompt
 
     cmd_preview = " ".join(cmd[:6])
     logger.debug(