Create compact handoff documents for another agent, session, or human to continue work. Use when a task is long-running, context is about to compact, work must pause, an audit needs resumable state, or the user asks for a handoff, continuation note, checkpoint, or next-session brief.
---
name: handoff
description: Create compact handoff documents for another agent, session, or human to continue work. Use when a task is long-running, context is about to compact, work must pause, an audit needs resumable state, or the user asks for a handoff, continuation note, checkpoint, or next-session brief.
license: MIT
metadata:
abvx_status: experimental
abvx_origin: adapted
---
# Handoff
Capture operational state, not a transcript. The next agent needs facts, paths, decisions, and next actions.
## When To Use
Use for:
- long-running audits;
- multi-repo publication work;
- paused implementation;
- debugging sessions with active hypotheses;
- research with source ledgers;
- job/client-search workflows;
- context compaction or thread transfer.
## Handoff Contents
Include:
- objective and current status;
- repos, branches, important paths, URLs, and local servers;
- decisions made and why;
- files changed, commits, PRs, or releases;
- commands already run and their result;
- active hypotheses, blockers, or pending checks;
- suggested skills for the next session;
- exact next steps in priority order.
## Closeout Receipt
When the handoff follows completed project work, include:
```text
changed: files, commits, PRs, or artifacts that matter
verified: checks and results, or why verification was unavailable
skipped_with_reason: docs, memory, cleanup, or rules intentionally not updated
follow_up: bounded next actions only
```
Link to existing evidence instead of repeating it. Separate local checks from
push, CI, deployment, and public availability. Do not turn closeout into cleanup
authority: retain temporary files and old plans unless removal was approved.
Memory changes require an explicit user request; a handoff is not that request.
## Compression Rules
- Link to artifacts instead of copying them.
- Do not duplicate full diffs, logs, PRDs, plans, or source docs.
- Preserve concrete dates and versions when relevant.
- Redact secrets, credentials, private contacts, and unnecessary personal data.
- Separate confirmed facts from assumptions.
## Output Location
If creating a file, write it outside the project unless the user asks for an in-repo artifact. Use the OS temp directory or a clear `handoff/` location approved by the repo conventions.
## Final Response
Tell the user where the handoff lives, what it covers, and what the next agent should do first.
When Claude will use it
Use when a task is long-running, context is about to compact, work must pause, an audit needs resumable state, or the user asks for a handoff, continuation note, checkpoint, or next-session brief.
---
name: agent-friction-ledger
description: Capture recurring agent-development friction as a local, privacy-safe Markdown report. Use only when the user explicitly requests a friction review of blockers, misleading errors, missing documentation, incompatible tools, or manual workarounds that should improve a doc, tool, skill, process, or evaluation.
---
---
name: agent-learning-layer-triage
description: Route reusable agent learning after an error, success pattern, review finding, or repeated workflow. Use when deciding whether an improvement belongs in the model prompt, memory/context note, durable docs, SKILL.md, checklist, script/tool, eval, golden fixture, or should be rejected as overfit.
---
---
name: agent-operations-contract
description: Define and review concrete agent capability cards, asynchronous operation receipts, decision receipts with revalidation, isolated and trust-graded memory scopes, and evidence-backed provider or tool registry entries. Use when a project introduces named agents, scheduled or long-running work, durable decisions or memory, provider routing, MCP tools, or external-action approval boundaries.
---
---
name: agent-tool-contract-review
description: Review tools, MCP functions, CLI commands, SET inputs, and AGENTS.md generator outputs as agent-facing contracts. Use when a tool description, schema, workflow input, or generated instruction may be ambiguous, unsafe, noisy, or hard for agents to call correctly.
---