| name | agent-bench-bootstrap |
| description | Bootstrap or refresh OpenClaw bench agents from a provided 7-file operating pack. Use when an agent needs to create, upgrade, or replicate another agent on the same or a separate OpenClaw installation using files like SOUL.md, IDENTITY.md, USER.md, AGENTS.md, MEMORY.md, HEARTBEAT.md, and TOOLS.md. Triggers include requests to generate a new agent, apply a reference bootstrap pack, clone a proven agent pattern onto another install, or validate that a bench-agent rollout preserved the required operating contract. |
Agent Bench Bootstrap
Bootstrap a target OpenClaw agent from a supplied operating pack. Treat the source files as the authority, keep the rollout reversible, and validate the result before declaring success.
Quick workflow
- Read
references/file-contract.md to confirm the required file set and optional outputs.
- Read
references/playbook.md and follow the rollout sequence.
- If a target path is provided, inspect the live target before writing.
- If the source pack is incomplete or ambiguous, stop and ask only for the missing blocker.
- If repeated validation is needed, run
scripts/validate_agent_bootstrap.py <target-dir>.
Operating rules
- Prefer applying a provided 7-file pack over rewriting the agent from scratch.
- Preserve continuity when the source pack represents an existing persona family. Do not accidentally create a disconnected persona unless the user explicitly wants that.
- Treat overwrite of live agent files as reversible but sensitive. Create a dated backup before replacement unless the user explicitly says not to.
- Preserve installation-specific details only when they do not conflict with the new operating model.
- Do not invent current facts, tool policies, or org details that are not present in the provided files.
- Keep communication concise and diff-oriented.
- Validate before finishing. At minimum check required files, missing files, and forbidden punctuation or style constraints that the source pack imposes.
Default rollout sequence
1. Intake
Confirm:
- target agent id or folder name
- source directory containing the 7 files
- whether this is a net-new agent or an update to an existing one
- whether optional
README.md and IMPLEMENTATION-PLAN.md should also be generated
2. Inspect
Inspect both sides before writing:
- source pack contents
- target directory existence and current files
- any local-only notes worth preserving
3. Backup
If the target already exists, create a dated backup copy before overwrite.
4. Apply
Copy the source files into the target agent directory:
SOUL.md
IDENTITY.md
USER.md
AGENTS.md
MEMORY.md
HEARTBEAT.md
TOOLS.md
Generate optional README.md and IMPLEMENTATION-PLAN.md only if the user wants them or the installation pattern clearly benefits from them.
5. Validate
Run the validator script or equivalent checks. Confirm:
- all required files exist
- no unexpected missing files remain
- no forbidden em dashes if the pack forbids them
- the target agent identity matches the requested role
- the result preserved continuity and specialization
6. Summarize
Return:
- outcome: created, updated, or blocked
- files applied
- local details preserved, if any
- validation result
- next action for the operator, if any
When to stop and ask
Ask one concise question only when a non-retrievable blocker exists, such as:
- source file set is missing one of the seven required files
- target path or agent id is unknown
- the user has not approved overwrite of a live custom agent and the change would replace nontrivial local work
Output shape
Use a short implementation summary with:
- recommendation
- what changed
- what was preserved
- validation result
- blocker, if any