Handoff

Handoff

Overview

Standardize session-to-session continuity when an agent must restart mid-work (plugin updates, context limits, crashes). Replaces ad-hoc memory files with a disciplined, template-driven handoff protocol.

Not the PR handoff. This is session-continuity state, distinct from the .vergil/pr-template.yml PR handoff used by the implement/audit pair.

Two modes:

• /handoff stop — Capture current work state before killing the session.
• /handoff start — Resume from the last handoff file after restarting.

Memory policy exemption. This skill writes to the memory directory as part of its documented workflow. Because the human invokes /handoff explicitly, the write has implicit approval and is exempt from the global memory management policy's approval requirement.

Mode: stop

Write a structured handoff file to the project memory directory:

<memory-dir>/handoff_active.md

Template

The agent fills in the template from conversation context. The agent writes the content — this skill provides the structure.

---
name: active handoff
description: Session handoff — <one-line summary of in-progress work>
type: project
---

## In-progress task

<What the agent was doing when stop was called. Phase, step, branch name,
PR numbers — enough to resume without re-deriving.>

## Completed this session

<Bulleted list of completed items with PR/issue URLs.>

## Remaining work queue

<Numbered list in execution order. Each item is one sentence.>

## Key context

<Anything non-obvious that the next session needs to know: workarounds in
effect, known failures, blocked items, environment state.>

## Restart command

<The exact slash command or instruction to resume.>

Behavior

1. Fill in the template from conversation context.
2. Write the file to <memory-dir>/handoff_active.md.
3. Add a pointer to MEMORY.md (or update the existing pointer): - [Active handoff](handoff_active.md) — <one-line summary>.
4. Confirm the handoff file was written and tell the user it is safe to kill the session.

Constraints

• Only one active handoff at a time. Writing a new one overwrites the previous.
• Do not invent or guess state — if something is unknown, say so in the file.
• All issue and PR references must be full URLs, not short #N references.

Mode: start

When invoked (or when the agent sees handoff_active.md in memory at session start):

1. Read <memory-dir>/handoff_active.md.
2. Summarize it to the user in 3-5 lines: what was in progress, what is next.
3. Ask: "Continue from where the last session left off?"
4. If the user confirms, proceed with the restart command / remaining work queue.
5. After the user confirms and work resumes, rename the handoff file from handoff_active.md to handoff_<date>.md (archived, no longer active). Update the MEMORY.md pointer accordingly.

Auto-detection

If the agent starts a session and finds handoff_active.md in the memory directory, it should proactively mention it: "There is a pending handoff from a previous session. Run /handoff start to resume, or ignore it to start fresh."

Edge cases

• No active handoff on start: Tell the user there is no pending handoff and exit.
• Multiple restarts without completing work: Each /handoff stop overwrites the active file. Old state is lost — this is intentional; only the most recent stop matters.
• User does not run stop before killing: No handoff file is written. The agent has no special recovery. This is the status quo and is acceptable.

What this is NOT

• Not a task tracker. It captures a snapshot for session continuity, not a persistent backlog.
• Not a replacement for issue tracking. Work items belong in GitHub issues.
• Not a memory dump. It captures only what is needed to resume — not everything that happened.