| name | agent-handoff-receiver |
| description | Read and operationalize handoff artifacts created for a future AI agent, including operational handoff md files and compact conversation-context md files. Use when asked to take over, resume, continue from handoff docs, read AgentHandoff outputs, verify current state before acting, or prepare a safe next-step plan from prior agent documentation. |
Agent Handoff Receiver
Overview
Use this skill when inheriting work from another agent. The goal is to become operationally safe before acting: understand the objective, recover recent user intent, verify live state, and identify the next low-risk step.
For a reusable checklist and response skeleton, read references/receiver-checklist.md when taking over a substantial project.
Core Workflow
- Locate the artifacts the user provided or named. Prefer an operational handoff first, then a conversation context file, then linked project docs.
- Read for hazards before plans: running jobs, services not to kill, GPUs/nodes to avoid, files not to overwrite, network/storage constraints, and destructive commands to avoid.
- Extract the current objective, newest user request, latest successful artifact, latest failure, and immediate next action.
- Verify live state before acting. Treat handoffs as potentially stale; check processes, jobs, logs, GPUs, branches, and file existence with targeted commands.
- Reconcile conflicts in this order: newest user message, current live state, operational handoff, conversation context, older summaries.
- Produce or internalize a ready-to-work brief: objective, hazards, current state, verification result, next command or edit, and open questions.
- Continue the work only after the safety-critical facts are verified. Ask the user only for ambiguity that cannot be resolved locally and would create meaningful risk.
Reading Order
Use this order unless the user gives a different one:
- Operational handoff md: current executable state and next steps.
- Conversation context md: compressed earlier history and full recent dialogue.
- Linked docs/scripts/configs: only the files needed for the next action.
- Logs: use
tail, rg, or bounded reads. Do not load huge logs wholesale.
- Live environment: verify running services and resources before changing anything.
Conflict Handling
When artifacts disagree, state the conflict briefly and follow the most recent reliable source. Do not silently choose a risky interpretation.
Use these labels:
confirmed: verified from live state or direct file/log evidence.stale: artifact says it, but live state has changed.unknown: not yet verified.assumption: reasonable but not proven.
Do Not
- Do not blindly rerun destructive commands, cleanup scripts, or broad process kills.
- Do not kill unknown jobs or services because an old handoff says they are stale.
- Do not read entire huge logs when targeted search is enough.
- Do not overwrite user changes or generated artifacts without checking.
- Do not treat compressed conversation summaries as exact quotes.
- Do not proceed with cluster or production changes until the relevant node/container/path is verified.
