Forge 🔨 Repair-Inspect Loop
ReviewAudited by ClawScan on May 10, 2026.
Overview
Forge is a disclosed code-repair orchestrator, but it can automatically run repair loops, commit code changes, and reuse cross-project repair memory, so users should review it carefully before use.
Before installing, run Forge only in a clean branch or disposable checkout, review the generated task files and diffs, set protected-files.txt for files that must not be touched, avoid automatic commits unless you are comfortable with the workflow, and periodically inspect or clear forge-reflections.jsonl and forge/reflections/patterns.jsonl.
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.
Project code may be changed and committed before the user has reviewed the exact diff.
The skill explicitly instructs automatic repair retries without asking the user and automatic git commits after a pass. This is high-impact local repository mutation without per-change approval in the documented workflow.
1. **FAIL/CONDITIONAL → 自动循环**，不问人 2. **PASS → 自动commit**，然后通知
Use this only on a separate branch or disposable working tree, define protected-files.txt, inspect diffs before accepting commits, and disable or avoid auto-commit if possible.
Details or bad instructions from one project could influence repairs in another project.
The skill creates persistent cross-project memory and injects it into future repair prompts. The artifacts describe decontextualization, but they do not show clear user approval, retention controls, or a robust boundary preventing project-specific or poisoned content from being reused elsewhere.
Universal layer (`forge/reflections/patterns.jsonl`): Abstract lessons stripped of file paths and project context. Auto-extracted from project patterns after each repair. Deduped by pattern_name. Injected into ALL future repairs across any project.
Review and sanitize reflection files, consider disabling the universal reflection layer, and clear cross-project patterns when switching projects or handling sensitive code.
Code context and task details may be shared with spawned agents and then reused in inspection or future repair steps.
The skill is designed to pass generated task files to spawned repair and inspection sessions. This is expected for its orchestrator role, but it means task context and repair results move between agents through files.
sessions_spawn(task=read(task_file), label=label, model=model)
Check generated task files before spawning agents, include only necessary context, and verify that result JSON files came from the expected repair or inspector session.
Users must trust and review the bundled script itself rather than relying on a known upstream source.
The skill includes a runnable helper script but has no declared source homepage or install specification. No remote installer or dependency download is shown, so this is a provenance/review note rather than a direct unsafe behavior.
Source: unknown Homepage: none No install spec — this is an instruction-only skill. Code file presence: scripts/forge.py
Review scripts/forge.py before running it and prefer installing from a known, version-controlled source if available.