Huo15 Autoresearch Loop
SuspiciousAudited by ClawScan on May 10, 2026.
Overview
This skill is an autonomous code-modification loop that can run arbitrary shell commands, commit changes, and reset/delete workspace files without clearly enforcing its advertised scope.
Install only if you intentionally want an autonomous agent to edit a Git worktree. Run it in a disposable branch or sandbox with clean backups, review every generated change before relying on commits, and avoid passing verification commands or environment variables that could execute unsafe shell actions.
Findings (4)
Artifact-based informational review of SKILL.md, metadata, install specs, static scan signals, and capability signals. ClawScan does not execute the skill or run runtime probes.
A failed or mistaken iteration could discard local work or delete untracked files, and a successful iteration could commit unintended changes.
The loop automatically stages and commits all changes on success, and hard-resets plus deletes untracked files on failure. These actions are broad and can affect the whole repository rather than a clearly enforced file scope.
git add -A
git commit -m "autoresearch iter $iteration: $goal" ...
git reset --hard HEAD ...
git clean -fdOnly use in a clean, backed-up Git repository. Require per-iteration approval, enforce scope_globs before git add/reset/clean, and avoid running it where untracked or unsaved work matters.
If the environment variable is wrong, injected, or overbroad, the skill can execute arbitrary local commands under the user's account.
The script shell-evaluates an environment-provided task. That is dynamic command execution with no visible allowlist or containment in the artifact.
if [[ -n "$CLAUDE_TASK" ]]; then
# 子 agent 模式：执行修改
eval "$CLAUDE_TASK" || modify_result=$?Replace eval with explicit, constrained commands or a reviewed tool interface, and require clear user confirmation before executing generated shell commands.
Users have less context for where the scripts came from and what local tools must exist before use.
The submitted package includes runnable shell scripts, but the source/provenance is not identified and there is no install specification documenting runtime dependencies.
Source: unknown Homepage: none Install specifications: No install spec — this is an instruction-only skill. Code file presence: 3 code file(s)
Review the included scripts manually, run them only in a sandboxed project, and prefer a version with clear source repository, pinned version, and declared runtime requirements.
Project goals, commands, and change summaries may remain on disk after the loop finishes.
The skill persists goal, verification command, iteration counts, and history to local files for later status/summary use.
"state_file": "~/.openclaw/tmp/autoresearch-loop-state.json", "log_file": "~/.openclaw/tmp/autoresearch-loop.log"
Do not put secrets in goals or verification commands, and delete the state/log files when they are no longer needed.