Function run_patch

harness/patch.py:30–171

Generate and verify a patch for a crash. Returns (diff_bytes, verdict, agent_result). diff_bytes is None if the agent never emitted a readable . verdict is None if no diff was produced. Writes patch.diff + patch_result.json to out_dir on every iteration (last one wins).

(
    crash: CrashArtifact,
    target: TargetConfig,
    model: str,
    out_dir: Path,
    report_text: str | None = None,
    max_iterations: int = DEFAULT_MAX_ITERATIONS,
    max_turns: int = PATCH_MAX_TURNS,
    container_name: str = "patch_target",
    run_reattack: bool = True,
    run_style: bool = False,
    agent_env: dict[str, str] | None = None,
    transcript_path: str | None = None,
    progress_prefix: str | None = None,
    system_prompt: str | None = None,
)

async def run_patch(
    crash: CrashArtifact,
    target: TargetConfig,
    model: str,
    out_dir: Path,
    report_text: str | None = None,
    max_iterations: int = DEFAULT_MAX_ITERATIONS,
    max_turns: int = PATCH_MAX_TURNS,
    container_name: str = "patch_target",
    run_reattack: bool = True,
    run_style: bool = False,
    agent_env: dict[str, str] | None = None,
    transcript_path: str | None = None,
    progress_prefix: str | None = None,
    system_prompt: str | None = None,
) -> tuple[bytes | None, PatchVerdict | None, AgentResult]:
    """Generate and verify a patch for a crash.

    Returns (diff_bytes, verdict, agent_result). diff_bytes is None if the
    agent never emitted a readable <patch_path>. verdict is None if no diff
    was produced. Writes patch.diff + patch_result.json to out_dir on every
    iteration (last one wins).
    """
    if not target.build_command:
        raise ValueError(f"target {target.name!r} has no build_command")
    if crash.poc_path not in crash.reproduction_command:
        raise ValueError(
            f"poc_path {crash.poc_path!r} not in reproduction_command "
            f"{crash.reproduction_command!r}"
        )

    out_dir = Path(out_dir)
    out_dir.mkdir(parents=True, exist_ok=True)

    diff: bytes | None = None
    verdict: PatchVerdict | None = None
    result = AgentResult()
    retry_evidence: tuple[str, str] | None = None
    iterations = 0
    timings: dict[str, float] = {}

    with sandbox.agent_container(
        target.image_tag, container_name, agent_env,
        memory=target.memory_limit, shm_size=target.shm_size,
    ) as container:
        await asyncio.to_thread(
            docker_ops.write_file, container, "/tmp/poc.bin", crash.poc_bytes
        )
        adapted_cmd = crash.reproduction_command.replace(crash.poc_path, "/tmp/poc.bin")
        # Ensure source_root is a git repo with a baseline commit so the
        # agent's `git diff` is deterministic. Gitignore the built binary so
        # the diff is source-only (otherwise the rebuilt binary lands in the
        # diff and grade's `git apply` rejects it).
        binary_rel = os.path.relpath(target.binary_path, target.source_root)
        ignore = f"printf '%s\\n' '{binary_rel}' '*.o' >> .gitignore && "
        await asyncio.to_thread(
            docker_ops.exec_sh,
            container,