fix(harness): command arms send the task instruction as written, no dokimi wrapper

dokimi/changelog.d/20261104_120000_napetrov_raw-task-prompt.md

@@ -0,0 +1,8 @@
+<!--
+Copyright (C) 2026 Intel Corporation
+SPDX-License-Identifier: Apache-2.0
+-->
+
+### Changed
+
+- **Command arms (`command:` and `--command-template`) send the task instruction as written** — the `Solve this coding task. Work only inside the task environment…` and `# Task: <name>` lines that dokimi put around `{prompt}` are gone, as a user's agent never sees them. `{prompt}` and `{prompt_quoted}` now expand to the bare `instruction.md`, and `{prompt_escaped}` to the same text with `{` and `}` doubled. Command-arm runs measured before this change are not comparable in their prompt text.

dokimi/dokimi/harnesses/command.py

@@ -386,11 +386,13 @@ def _render_template(template: str, context: dict[str, Any]) -> str:
 
 
 def _task_prompt(task: TaskSpec) -> str:
-    return (
-        "Solve this coding task. Work only inside the task environment. "
-        "When finished, leave the files ready for the verifier.\n\n"
-        f"# Task: {task.name}\n\n{task.instruction}"
-    )
+    """The task instruction as the user wrote it: no wrapper, the way a real harness sends it.
+
+    A wrapper ("Solve this coding task…", "# Task: <name>") is text dokimi adds that a
+    user's agent never sees, so it can move the score of a command arm for no reason
+    the task measures.
+    """
+    return task.instruction
 
 
 # Windows reports a crash or Ctrl-C as an NTSTATUS error code (0xC0000005, 0xC000013A, 0xC0000409).

dokimi/tests/test_command_prompt.py

@@ -0,0 +1,29 @@
+# Copyright (C) 2026 Intel Corporation
+# SPDX-License-Identifier: Apache-2.0
+
+"""A command arm's prompt is the task instruction as written, with no dokimi wrapper."""
+
+from __future__ import annotations
+
+from dokimi.harnesses.command import CommandHarness
+from dokimi.harnesses.tasks import TaskSpec
+
+
+def test_command_prompt_is_the_raw_instruction(tmp_path):
+    task = TaskSpec(name="toy", path=tmp_path / "toy", instruction="Fix the off-by-one in sum().")
+    harness = CommandHarness("fake", "echo {prompt_quoted}")
+    argv = harness.build_command(task, model=None, output_dir=tmp_path / "out")
+    assert argv == ["echo", "Fix the off-by-one in sum()."]
+
+
+def test_every_prompt_placeholder_keeps_a_multiline_instruction_as_written(tmp_path):
+    instruction = "Fix {x}.\nSecond line.\n"
+    task = TaskSpec(name="toy", path=tmp_path / "toy", instruction=instruction)
+
+    def argv(template):
+        harness = CommandHarness("fake", template)
+        return harness.build_command(task, model=None, output_dir=tmp_path / "out")
+
+    assert argv("echo '{prompt}'") == ["echo", instruction]
+    assert argv("echo {prompt_quoted}") == ["echo", instruction]
+    assert argv("echo '{prompt_escaped}'") == ["echo", "Fix {{x}}.\nSecond line.\n"]