Port RLM SWE taskset to v1

[rasdani]

Summary

• add a reusable v1 SWETaskset adapter over the existing composable SWE taskset backends
• rewire rlm_swe_v1 to compose vf.SWETaskset with the existing v1 vf.RLM harness
• expose SWETaskset / SWETasksetConfig from the v1 and top-level APIs
• add the RLM SWE edit skill package and update focused tests/docs

Notes

This keeps the existing SWE backend implementations in place and adapts them into the v1 taskset/runtime shape. It does not run a live prime eval run; that should be the final smoke before publishing broader env changes.

Validation

• uv run pre-commit install
• uv run ruff check --fix verifiers/v1/packages/tasksets/swe.py verifiers/v1/packages/tasksets/skills/edit/src/edit/edit.py
• uv run pytest tests/test_v1_rlm_swe.py
• uv run pre-commit run semgrep-v1-policy --files verifiers/v1/packages/tasksets/swe.py verifiers/v1/packages/tasksets/skills/edit/src/edit/edit.py
• uv run pre-commit run --hook-stage pre-push ty --files verifiers/v1/packages/tasksets/swe.py
• pre-commit hooks on commit
• pre-push hooks on push

Note

Port RLM SWE taskset to use shared vf.SWETaskset adapter

• Replaces the environment-local R2ESWETaskset and all r2e-specific helpers in rlm_swe_v1.py with the new vf.SWETaskset adapter.
• SWETaskset bridges legacy SWE backends into the v1 runtime: it constructs the backend via make_swe_taskset, produces v1 task rows, maps sandbox specs, adapts reward signals, and registers setup_swe_sandbox/cleanup_swe_state handlers.
• SWETasksetConfig supports multiple backends (r2e, swebench, multiswe, openswe, and others) via a unified task_type field, replacing the old r2e-only config.
• Adds a new edit skill package under verifiers/v1/packages/tasksets/skills/edit/ for safe single-occurrence string replacement in files.
• SWETaskset and SWETasksetConfig are exported from the top-level verifiers package.
• Behavioral Change: setup/cleanup handler names changed from setup_r2e_sandbox/cleanup_r2e_state to setup_swe_sandbox/cleanup_swe_state.

📊 Macroscope summarized 28edca4. 10 files reviewed, 3 issues evaluated, 0 issues filtered, 2 comments posted

🗂️ Filtered Issues

[macroscopeapp]

🟢 Low tasksets/swe.py:210

timeout_minutes = int(sandbox_data.get("timeout_minutes") or 60) silently replaces an explicit 0 with 60, so a user who sets timeout_minutes: 0 (e.g., to disable timeout) gets 60 minutes instead. Consider using sandbox_data.get("timeout_minutes", 60) to preserve 0 as a valid value.

-            timeout_minutes = int(sandbox_data.get("timeout_minutes") or 60)
+            timeout_minutes = int(sandbox_data.get("timeout_minutes", 60))

🚀 Reply "fix it for me" or copy this AI Prompt for your agent:

In file verifiers/v1/packages/tasksets/swe.py around line 210:

`timeout_minutes = int(sandbox_data.get("timeout_minutes") or 60)` silently replaces an explicit `0` with `60`, so a user who sets `timeout_minutes: 0` (e.g., to disable timeout) gets 60 minutes instead. Consider using `sandbox_data.get("timeout_minutes", 60)` to preserve `0` as a valid value.

Evidence trail:
verifiers/v1/packages/tasksets/swe.py line 210 at REVIEWED_COMMIT: `timeout_minutes = int(sandbox_data.get("timeout_minutes") or 60)` — Python `or` treats `0` as falsy, so `0 or 60` evaluates to `60`.

[macroscopeapp]

🟢 Low edit/edit.py:30

The write_text() call on line 30 is unguarded, so write failures (permission denied, disk full, read-only filesystem) raise unhandled exceptions instead of returning error message strings like the rest of the function. This breaks the established error-handling pattern where all I/O errors return formatted error strings.

-    filepath.write_text(content.replace(old_str, new_str, 1))
-    return f"Edited {path}"
+    try:
+        filepath.write_text(content.replace(old_str, new_str, 1))
+    except Exception as exc:
+        return f"Error writing {path}: {exc}"
+    return f"Edited {path}"

🚀 Reply "fix it for me" or copy this AI Prompt for your agent:

In file verifiers/v1/packages/tasksets/skills/edit/src/edit/edit.py around lines 30-31:

The `write_text()` call on line 30 is unguarded, so write failures (permission denied, disk full, read-only filesystem) raise unhandled exceptions instead of returning error message strings like the rest of the function. This breaks the established error-handling pattern where all I/O errors return formatted error strings.

Evidence trail:
File: verifiers/v1/packages/tasksets/skills/edit/src/edit/edit.py (lines 1-31 at REVIEWED_COMMIT). Line 19-22: read_text() wrapped in try/except returning error string. Line 30: write_text() has no try/except guard. The function's pattern (lines 17-18, 19-22, 24-28) is to return error strings for all failure cases.