Archive original DeepSWE GPT xhigh v1 code and historical results

[gwh6669999]

Publish the original DeepSWE GPT xhigh v1 source snapshot and historical 113-task summary under benchmark/deepswe-gptxhigh-v1/. This PR replaces #4466 and now excludes the later v1-revised execution package entirely.

本 PR 仅提交旧 v1 的原始代码与结果，不包含正在运行的新 benchmark，也不把后续执行逻辑修复混入历史快照。原始 14 个文件（包括 README）内容及权限完全不变。

The archive matches source subtree 98c262a:benchmark/deepswe-five-arm exactly, with Git tree 1bc5d2b3b74761a97d34ba3f3612e977fd610340. Only the containing directory name changes. Companion documentation lists provenance, external runtime prerequisites, known original defects, and evidence limits. benchmark/check_deepswe_v1.py verifies exact archive bytes/modes and Python/Bash syntax without running a benchmark.

Validation:

• Exact original tree identity verified.
• 11 Python files, 2 embedded Python programs, and 2 Bash scripts passed syntax checks.
• All five original arm imports and factory dispatch passed locally with Python 3.12.13 / datacurve-pier 0.3.1 and external experiment modules. No agents were instantiated.
• Whitespace and credential/private-path checks passed.

Execution boundary: this original export is not a standalone runnable package. It depends on external launch scripts, task manifests/definitions, the original plain runner, container/network setup, and provider configuration. Known original admission, retry, and launcher defects are deliberately preserved. We do not claim an end-to-end run, reproducibility from this archive alone, or validation of historical scores. The companion note retains the current SSH Goal / Codex CLI withdrawal caveat.

The revised package cited by the previous Copilot review has been removed from the final diff. Its former code remains in prior commit history for separate work; no additional execution changes are proposed here.

[Copilot]

🟡 Changes recommended

Unresolved execution, admission, network-isolation, retry, and result-attribution findings remain.

Get a fresh assessment by requesting another Copilot review.

Pull request overview

Preserves the original DeepSWE v1 snapshot while isolating later execution fixes in a revised package.

Changes:

• Documents provenance and result-attribution boundaries.
• Preserves the historical snapshot unchanged.
• Adds revised admission, delivery, retry, networking, and launcher logic.
• Adds offline execution-contract tests; no benchmark rerun is claimed.

File summaries

File	Summary
benchmark/deepswe-gptxhigh-versions.md	Documents provenance and version boundaries.
benchmark/deepswe-gptxhigh-v1/workspace_delivery.py	Preserved original delivery logic.
benchmark/deepswe-gptxhigh-v1/run_loopx_rerun_54_20260908.sh	Preserved original launcher.
benchmark/deepswe-gptxhigh-v1/run_five_arms_remaining59_20260910.sh	Preserved original launcher.
benchmark/deepswe-gptxhigh-v1/README.md	Preserved original documentation.
benchmark/deepswe-gptxhigh-v1/preflight_loopx_rerun.py	Preserved original preflight.
benchmark/deepswe-gptxhigh-v1/pier_cn.py	Preserved original launcher support.
benchmark/deepswe-gptxhigh-v1/loopx_wen_native_runner.py	Preserved original native runner.
benchmark/deepswe-gptxhigh-v1/loopx_heartbeat_supervisor.py	Preserved original heartbeat support.
benchmark/deepswe-gptxhigh-v1/loopx_codex_cli_runner.py	Preserved original CLI runner.
benchmark/deepswe-gptxhigh-v1/goal_claude.py	Preserved original adapter.
benchmark/deepswe-gptxhigh-v1/codex_nosandbox_wrapper.py	Preserved original wrapper.
benchmark/deepswe-gptxhigh-v1-revised/workspace_delivery.py	Validates linked-worktree delivery.
benchmark/deepswe-gptxhigh-v1-revised/tests/test_execution_contracts.py	Covers revised execution contracts offline.
benchmark/deepswe-gptxhigh-v1-revised/run_loopx_rerun_54_20260908.sh	Runs the revised 54-task launcher.
benchmark/deepswe-gptxhigh-v1-revised/run_five_arms_remaining59_20260910.sh	Runs the revised five-arm experiment.
benchmark/deepswe-gptxhigh-v1-revised/README.md	Documents the revised package scope.
benchmark/deepswe-gptxhigh-v1-revised/preflight_loopx_rerun.py	Validates admission and task manifests.
benchmark/deepswe-gptxhigh-v1-revised/plain_appserver_runner.py	Provides plain app-server execution.
benchmark/deepswe-gptxhigh-v1-revised/pier_cn.py	Applies harness and network configuration.
benchmark/deepswe-gptxhigh-v1-revised/loopx_wen_native_runner.py	Handles native execution and retries.
benchmark/deepswe-gptxhigh-v1-revised/loopx_heartbeat_supervisor.py	Provides heartbeat-based continuation.
benchmark/deepswe-gptxhigh-v1-revised/loopx_codex_cli_runner.py	Runs CLI segments and delivery checks.
benchmark/deepswe-gptxhigh-v1-revised/codex_nosandbox_wrapper.py	Provides the revised Codex wrapper.

Review details

Suppressed comments (7)

benchmark/deepswe-gptxhigh-v1-revised/codex_nosandbox_wrapper.py:6

• The revised package is documented as having no benchmark results, but this module's top-level docstring states that a run scored f2p 31/35 without identifying the harness/version or labeling it as historical. Because the statement lives inside v1-revised, it can be read as a result attributable to this package and undermines the version/result boundary. Remove the score or explicitly identify its original source and non-attribution status.

LoopX's codex-cli host is the path that works: driven through it, a Turn loop
ran four times on one task, committed a 22 KB patch and scored f2p 31/35.  Its
one problem is the sandbox — it always passes `--sandbox <mode>` (or

benchmark/deepswe-gptxhigh-v1-revised/loopx_codex_cli_runner.py:346

• A nonzero turn run-once result only stops here when its JSON payload is empty. Structured host_failure records for permanent errors such as authentication or quota failures still reach _retry_delay, increment the streak, and launch another model turn even when host_failure.retryable is false. Please fail closed for a missing/false retryable marker and only continue this loop for explicitly retryable host failures.

        if completed.returncode != 0 and not payload:
            break

benchmark/deepswe-gptxhigh-v1-revised/loopx_codex_cli_runner.py:320

• This runner normalizes delivery after every segment, but it never releases the canonical copy after a linked-worktree patch is recovered. If segment 1 recovers patch A into the canonical checkout and the resumed agent adds B in the linked worktree, the next call sees canonical A and linked A+B and returns ambiguous, preventing the multi-segment Codex CLI arm from ever reaching valid delivery. Carry the selected worktree forward or reset/release the canonical duplicate between segments, as the heartbeat supervisor does.

        delivery = normalize_delivery(project, base_sha)
        write_receipt(delivery_path, delivery)
        todo_state = _todo_state(args, project)

benchmark/deepswe-gptxhigh-v1-revised/loopx_codex_cli_runner.py:304

• This turn run-once subprocess is not wrapped for subprocess.TimeoutExpired. When a segment exceeds the configured timeout, the exception escapes run() before the delivery receipt and structured failed result are written, so the launcher exits with a traceback instead of recording a controlled failed segment. Catch the timeout, record it as a failed segment, and finalize or retry through the same bounded path as other failures.

        completed = subprocess.run(
            command,
            cwd=project,
            env=env,
            capture_output=True,
            text=True,
            timeout=segment_timeout + 240,
            check=False,
        )

benchmark/deepswe-gptxhigh-v1-revised/loopx_wen_native_runner.py:142

• The app-server v2 codexErrorInfo discriminator is camelCase here. Values such as serverOverloaded and rateLimitExceeded are therefore classified as non-retryable because this set only contains snake_case aliases, so _terminal_error aborts a transient provider failure instead of taking the retry path. The canonical Codex classifier handles these variants; normalize or add the aliases before deciding retryability.

    retryable = code in {
        "rate_limit_exceeded", "overloaded", "server_error", "request_timeout",
        "temporarily_unavailable", "httpConnectionFailed", "responseStreamConnectionFailed",
        "responseStreamDisconnected", "responseTooManyFailedAttempts",
    }

benchmark/deepswe-gptxhigh-v1-revised/loopx_wen_native_runner.py:163

• The native app-server error contract also emits a standalone method: "error" notification with params.error (for example, usage-limit failures), but this terminal predicate only recognizes turn/completed and legacy event_msg events. Such a failure is ignored by _terminal_error; if the following completion has no embedded error, the runner loses the structured classification and may treat the turn as normal or wait until timeout instead of applying the retry/abort decision. Handle the typed error notification in the terminal/error path.

    terminal = method == "turn/completed" or (
        event_type == "event_msg" and payload_type in {"task_complete", "task_completed", "turn_completed"}
    )
    if not terminal:
        return None
    turn = params.get("turn") if isinstance(params.get("turn"), Mapping) else {}
    for container in (payload, turn, params):
        error = container.get("error") if isinstance(container, Mapping) else None
        if error is None:
            continue
        return classify_model_error(error)
    return None

benchmark/deepswe-gptxhigh-v1-revised/preflight_loopx_rerun.py:133

• tomllib accepts special floating-point values such as nan and inf; both pass timeout <= 0 here. A malformed task can therefore contribute a non-finite agent timeout, and minimum_agent_timeout/the admission comparison can accept or mis-handle it instead of rejecting the task definition. Require a finite positive timeout before appending it to the manifest.

        if re.fullmatch(r"[0-9a-fA-F]{7,40}", base) is None or timeout <= 0:
            missing.append(task)

• Files reviewed: 29/29 changed files
• Comments generated: 2
• Review effort level: Lite

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[gwh6669999]

Superseded by #4498, submitted from a clean branch containing only the original v1 archive and its verification documentation. 已由干净分支重新提交为 #4498；原始 v1 内容和结果不变，不包含修复版。