Hold stops an AI agent after an unanticipated failure, reports what happened, and waits for the user to choose how to proceed.
Install & Download
1. ClawHub CLI
The fastest way to install a skill directly from the registry.
npx clawhub@latest install hold
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 hold 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 Hold?
Hold is a stop-and-recover gate for AI agent workflows. It activates only when a failure or deviation was not identified in preflight’s residual risk list, preventing agents from improvising fixes or blindly retrying systemic problems. This focused approach helps make Openclaw Skills more reliable without interrupting work for routine or anticipated errors.
When triggered, Hold classifies the issue as a new failure class, a materialised residual risk, or an invalidated plan. It then provides an evidence-based report and exactly three recovery options—fix and restart, continue with caution, or abort—and waits for the user’s decision.
Hold Use Cases
- An agent encounters an unexpected error during a planned workflow.
- A step deviates from the plan in a way preflight did not flag.
- A plan’s underlying assumption becomes false, making the remaining steps unreliable.
- A failure appears systemic and retrying or improvising a workaround could cause further side effects.
- A user needs a clear record of completed work, in-progress actions, and changes before choosing how to recover.
How Hold Works
- Check the trigger: Confirm that a failure or deviation occurred and that it is absent from preflight’s residual risk list. Routine anticipated errors and preflight test failures are outside Hold’s scope; transient errors should be retried once before escalation.
- Stop immediately: Do not retry, apply a workaround, or continue to another step.
- Classify the failure: Identify it as a new failure class, a materialised residual risk, or an invalidated plan.
- Report the evidence and state: Provide the exact error message, affected plan step, reason the issue was not anticipated, completed work, in-progress work, and side effects.
- Present exactly three options: Offer fix and restart, continue with caution, and abort, including relevant risks, restart point, estimated additional cost, and cleanup details.
- Wait for the user: Do not select or execute an option until the user replies with “1”, “2”, or “3”. If they choose option 1, run preflight again on the fixed path before restarting.
Hold Setup
Hold is a behavioral skill: its SKILL.md defines when to stop, how to classify failures, and the exact report and decision flow. The provided file does not specify an installation command, dependencies, or configuration variables. Add the skill to the location supported by your Openclaw Skills environment, then make its trigger conditions and user-confirmation requirement available to the agent.
# No installation command is specified in the provided SKILL.md.
# Place the skill file in your environment's supported skills directory.
Before using it, ensure the workflow has a preflight step that records residual risks. Without that list, the agent cannot determine whether a failure is genuinely unanticipated.
Hold Data Schema & Taxonomy
Hold does not define persistent storage, generated files, or machine-readable metadata beyond the skill’s frontmatter and its required incident report.
| Element | Format or contents |
|---|---|
| Skill metadata | YAML frontmatter: name, description, and license (MIT) |
| Trigger context | Failure or plan deviation checked against preflight’s residual risk list |
| Classification | new class, materialised risk, or invalidated plan |
| Incident report | Markdown code block containing the affected step, exact error, explanation, current state, and recovery options |
| Current state | Completed work, in-progress work, and side effects written, sent, or changed |
| User decision | One of three numbered choices: fix and restart, continue with caution, or abort |
| Persistent data | Not specified by the skill |
Hold Advanced Features
- Preflight-aware gating: Distinguishes unanticipated failures from risks already recorded during preflight.
- Three-way failure taxonomy: Separates new failure classes, materialised residual risks, and invalidated plans to clarify recovery decisions.
- Evidence-based incident reporting: Requires the exact error message and a concrete account of execution state and side effects.
- User-controlled recovery: Always presents exactly three options and pauses until the user chooses.
- Safe restart discipline: Requires a fresh preflight on the fixed path before restarting after a fix.
- No autonomous remediation: Prohibits blind retries, improvised workarounds, and resuming without user approval.
SKILL.md
Loading
Related Openclaw Skills
Scope
Scope catches AI agent actions that cross the boundaries agreed during reconnaissance and pauses for your decision.
Shell Chat
Shell Chat lets AI agents exchange real-time direct messages, group chats, photos, and files with people and other agents.
TikTok Comment Consumer Insights
Retrieve public TikTok search results, creator posts, and video comments as structured JSON to support consumer insights and content research.
Blave Quant
Blave Quant connects AI agents to global market data and exchange workflows, with explicit confirmation required for every trading action.
Recon
Recon inspects a repository, surfaces risks and unknowns, and requires user confirmation of task-specific rules before coding begins.
Finance Theory Master
Finance Theory Master helps assess which financial theories fit a scenario by testing assumptions, weighing evidence, and explaining limitations.
Featured*
