Ultrawork — Parallel AI Task Execution
AdvisoryAudited by Static analysis on May 3, 2026.
Overview
No suspicious patterns detected.
Findings (0)
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.
Several agents could make project changes at the same time, consume model resources, or act on overly broad task wording before the user has reviewed the plan.
The skill passes arbitrary task text into autonomous subagents in run mode and launches them in the background, without artifact-backed file scopes, per-task approval, or safety limits.
TASKS=("$@") ... openclaw sessions spawn --task "$task_desc" --model "$model" --runtime subagent --mode run --output "$output_file" 2>/dev/null &Use only with explicit, narrow tasks on a clean branch or disposable worktree; require confirmation before edits; avoid suppressing errors; and add file/path scopes or review gates before running subagents.
One bad or overlapping task can interfere with other tasks and leave the repository in a difficult-to-debug state.
The core workflow intentionally starts multiple agents simultaneously, which can create conflicting edits or compound errors if tasks are not truly independent.
4. FIRE ALL AT ONCE - Spawn subagents in parallel - No waiting between spawns
Run parallel tasks in isolated branches/worktrees, enforce file-level ownership, merge changes only after review, and provide an easy rollback path.
In an untrusted or unfamiliar repository, build or test scripts could run commands the user did not expect.
The verification step runs project-defined npm scripts automatically when package.json is present. This is purpose-aligned for coding verification, but npm scripts can execute arbitrary local commands.
if [[ -f "package.json" ]]; then if npm run build > /dev/null 2>&1; then ... if npm test > /dev/null 2>&1; then
Inspect package.json scripts first or run the skill in a sandboxed environment, especially on third-party repositories.
If a spawned agent hangs or keeps working unexpectedly, the user may have to identify and stop it manually.
The skill tracks background subagent processes and waits for them, which is disclosed and purpose-aligned, but it does not show timeouts, cancellation handling, or stuck-process cleanup beyond waiting.
# Save PID echo $! > "$RESULTS_DIR/task-$task_id.pid" ... wait $pid 2>/dev/null || true
Add timeouts, visible process status, cancellation instructions, and cleanup procedures for spawned subagents.
A user or agent may need to create or adapt the script manually, which can introduce differences from the reviewed artifact.
The artifact set does not include an installed executable implementation, while SKILL.md shows a shell script-like implementation path. This is not malicious by itself, but it limits review of the actual runtime packaging.
No install spec — this is an instruction-only skill.
Prefer a packaged, reviewable implementation with declared capabilities, or manually compare any created script against the reviewed SKILL.md before use.