sofagent-loop turns one prompt into a gated agent pipeline that writes code, audits changes, reviews results, and returns control to a human.
Install & Download
1. ClawHub CLI
The fastest way to install a skill directly from the registry.
npx clawhub@latest install sofagent-loop
2. Manual Installation
Copy the skill folder to one of these locations
~/.openclaw/skills/ <project>/skills/ Priority: Workspace > Local > Bundled
3. Prompt Installation
Copy this prompt to OpenClaw to install it automatically.
Help me install sofagent-loop using Clawhub. If Clawhub is not installed, install it first (npm i -g clawhub).
Prefer to download?
Get the raw skill files in a ZIP archive.
What is sofagent-loop?
sofagent-loop is a self-iterative development skill for Openclaw Skills that turns a single prompt into a repeatable engineering loop: code, audit, review, and human confirmation.
It connects the agent definitions in agents/, the orchestration contract in LOOP/LOOP.md, and the verification artifacts in docs/verification/ so sofagent can operate as an agent-driven development system instead of a one-shot assistant. For teams using Openclaw Skills, it is the cleanest way to run auditable, multi-agent coding cycles with minimal manual coordination.
sofagent-loop Use Cases
- You want Openclaw Skills to drive repeated code changes with human oversight.
- You need an automated loop for feature work, bug fixes, refactors, or compliance updates.
- You want multiple agents to collaborate on code generation, auditing, and review.
- You need a traceable development process with git commits, audit reports, and review history.
- You are adapting sofagent to your own repository by customizing the agent definitions under
agents/. - You want a disciplined workflow where the human only confirms the final review output before the next round.
How sofagent-loop Works
- Install and activate the skill, then load the orchestration guide in
LOOP/LOOP.mdand the agent definitions inagents/. - Submit a task and let the loop route it to
minimal-change-engineerfor the smallest safe code change. - The coding agent makes the change, commits it to git, and prepares the change set for verification.
sofagent-auditchecks the implementation and records audit findings.code-reviewerproduces a severity-tagged review report, and any blocking issues are handled in the next iteration.- A human confirms the review output, approves the round, and decides whether the loop continues.
- In the outer loop,
forward-deployed-engineercan analyzethink.md, improve agent definitions, triggercompliance-auditor, and refresh verification documents for the next cycle.
sofagent-loop Setup
- Install sofagent first, because sofagent-loop depends on the core agent, audit, and orchestration stack.
- Install the skill from ClawHub/SkillHub or run the platform installer manually.
- Ensure your OpenClaw environment supports
session.spawnfor sub-agent startup. - If you are adapting the skill to your own project, customize the agent files under
agents/before running the loop. - Review
quick-start.mdto validate the first end-to-end run.
# Install from ClawHub / SkillHub
clawhub skill install KongFangXun/sofagent-loop
# Manual install for supported platforms
bash LOOP/loop-install.sh --platform openclaw
bash LOOP/loop-install.sh --platform workbuddy
- On OpenClaw, the skill can load automatically when a development-loop scenario is detected.
- On WorkBuddy, trigger it with
@skill:sofagent-loop. - To verify the installation, run the quick start flow and confirm the loop reaches audit and review.
sofagent-loop Data Schema & Taxonomy
The skill separates orchestration, agent behavior, and verification evidence so Openclaw Skills can reason about each stage of the loop.
| Path / Field | Type | Purpose |
|---|---|---|
agents/ |
Agent definitions | Stores role-specific behavior for minimal-change-engineer, sofagent-audit, code-reviewer, and related agents. |
LOOP/LOOP.md |
Workflow spec | The authoritative loop contract that defines task flow, gates, and iteration rules. |
docs/verification/ |
Verification artifacts | Holds audit documents, review evidence, and compliance outputs. |
think.md |
Reflection log | Captures task-level and outer-loop reflections used to refine the system. |
history.jsonl |
Append-only log | Stores machine-readable audit history for traceability and trend analysis. |
quick-start.md |
Onboarding guide | Provides the fastest path to a successful first run. |
Metadata taxonomy used by the skill:
- Identity:
name,slug,displayName,version. - Discovery:
tags,triggers,scenarios. - Guardrails:
not_whento avoid unnecessary activation. - Runtime outputs: code changes, audit results, review reports, human confirmation, and historical logs.
- Platform mapping: OpenClaw, WorkBuddy, and manual one-key activation paths.
sofagent-loop Advanced Features
- Multi-agent orchestration across coding, auditing, reviewing, and human approval.
- Hard-gated verification flow that should not skip
sofagent-audit; bypass attempts such as--no-verifyare detectable. - Outer-loop automation that uses
think.mdtrends to improve agent definitions and compliance coverage. - Project-specific customization by editing
agents/without changing the core loop contract. - Persistent traceability through git commits,
history.jsonl, and structured review reports. - Platform-aware activation paths for OpenClaw, WorkBuddy, and manual trigger commands.
- Ideal for Openclaw Skills users who want repeatable, auditable self-iteration instead of ad hoc prompting.
SKILL.md
Loading
Related Openclaw Skills
Feishu File Operations
A utility skill for automating Feishu file uploads, message delivery, and collaborative cloud document creation.
Gmail Lead Desk
Gmail Lead Desk turns Gmail into an AISA-powered sales and support inbox workflow for Openclaw Skills, with OAuth connect, unread lead triage, thread summaries, safe draft replies, and archiving.
OSINT Personal Profiler
A legal OSINT profiling skill that turns minimal public identifiers into a confidence-scored, 12-dimensional person profile.
KOL Creator Discovery
Openclaw Skills turns one creator profile into a verified, outreach-ready KOL prospect list with similarity ranking, email enrichment, and evidence-backed metadata.
CVE Alerts
CVE Alerts is an Openclaw Skills workflow that monitors CVE feeds, matches them to your stack, and alerts only on net-new risks.
Biaoshu-Bailian (ZCM Bidding Assistant)
A specialized AI skill that calls the ZCM Open API to automate tender file analysis, bid document generation, and compliance auditing.
Featured*
