handoff

Handoff Skill

Quick Ref: Create structured handoff for session continuation. Output: .agents/handoff/YYYY-MM-DD-<topic>.md + continuation prompt.

YOU MUST EXECUTE THIS WORKFLOW. Do not just describe it.

Create a handoff document that enables seamless session continuation.

Handoff is the write-side of the self-healing context loop

Handoff is not only an end-of-session ritual — it is the write-side of the handoff → clear → rehydrate loop that keeps a long-running agent (or a peer orchestrator) healthy against context bloat. The natural pre-/clear step is to hand off: capture the working state to durable state first, then clear, then rehydrate from the artifact. Handoff before clear, always — clearing without a current handoff loses the thread.

For that to be safe, the handoff must be complete enough to rehydrate the lane to exactly where it was: the active goal, the claimed bead(s), held file reservations, the peer/comms topology, and a pointer to the working thread. A thin handoff makes clear destructive; a complete one makes the agent disposable-yet-continuous (the system-not-DAG property at the orchestrator level). The structured artifact is ao handoff → .agents/handoff/ (JSON; --no-kill writes without the tmux restart).

Every handoff is a compounding artifact

A handoff is a first-class node in the compounding artifact graph, not a throwaway note. It feeds the llm-wiki / knowledge corpus, gets mined and measured (the artifact already carries consumed / consumed_by hooks), and each cycle should make the next handoff better. Context, code, and markdown are all artifacts to be fed back. Write handoffs dense and honest — they are evidence the flywheel compounds on, not paperwork.

Execution Steps

Given /handoff [topic]:

Step 1: Create Output Directory

mkdir -p .agents/handoff

Step 2: Identify Session Context

If topic provided: Use it as the handoff identifier.

If no topic: Derive from recent activity:

# Recent commits
git log --oneline -5 --format="%s" | head -1

# Check current issue
bd current 2>/dev/null | head -1

# Check ratchet state
ao ratchet status 2>/dev/null | head -3

Use the most descriptive source as the topic slug.

Topic slug format: 2-4 words, lowercase, hyphen-separated (e.g., auth-refactor, api-validation). Fallback: If no good topic found, use session-$(date +%H%M) (e.g., session-1430).

Step 3: Gather Session Accomplishments

Review what was done this session:

# Recent commits this session (last 2 hours)
git log --oneline --since="2 hours ago" 2>/dev/null

# Recent file changes
git diff --stat HEAD~5 2>/dev/null | head -20

# Research produced
ls -lt .agents/research/*.md 2>/dev/null | head -3

# Plans created
ls -lt .agents/plans/*.md 2>/dev/null | head -3

# Issues closed
bd list --status closed --since "2 hours ago" 2>/dev/null | head -5

Step 4: Identify Pause Point

Determine where we stopped:

1. What was the last thing done?
2. What was about to happen next?
3. Were we mid-task or between tasks?
4. Any blockers or decisions pending?

Check for in-progress work:

bd list --status in_progress 2>/dev/null | head -5

Step 5: Identify Key Files to Read

List files the next session should read first:

• Recently modified files (core changes)
• Research/plan artifacts (context)
• Any files mentioned in pending issues

# Recently modified
git diff --name-only HEAD~5 2>/dev/null | head -10

# Key artifacts
ls .agents/research/*.md .agents/plans/*.md 2>/dev/null | tail -5

Step 6: Write Handoff Document

Write to: .agents/handoff/YYYY-MM-DD-<topic-slug>.md (use date +%Y-%m-%d)

# Handoff: <Topic>

**Date:** YYYY-MM-DDTHH:MM:SSZ
**Session:** <brief session description>
**Status:** <Paused mid-task | Between tasks | Blocked on X>

---

## What We Accomplished This Session

### 1. <Accomplishment 1>

<Brief description with file:line citations>

**Files changed:**
- `path/to/file.py` - Description

### 2. <Accomplishment 2>

...

---

## Where We Paused

<Clear description of pause point>

**Last action:** <what was just done>
**Next action:** <what should happen next>
**Blockers (if any):** <anything blocking progress>

---

## Context to Gather for Next Session

1. <Context item 1> - <why needed>
2. <Context item 2> - <why needed>

---

## Questions to Answer

1. <Open question needing decision>
2. <Clarification needed>

---

## Files to Read

Priority files (read first)

path/to/critical-file.py .agents/research/YYYY-MM-DD-topic.md

Secondary files (for context)

path/to/related-file.py

### Step 7: Write Continuation Prompt

**Write to:** `.agents/handoff/YYYY-MM-DD-<topic-slug>-prompt.md` (use `date +%Y-%m-%d`)

```markdown
# Continuation Prompt for New Session

Copy/paste this to start the next session:

---

## Context

<2-3 sentences describing the work and where we paused>

## Read First

1. The handoff doc: `.agents/handoff/YYYY-MM-DD-<topic-slug>.md`
2. <Other critical files>

## What I Need Help With

<Clear statement of what the next session should accomplish>

## Key Files

```

Open Questions

1. <Question 1>
2. <Question 2>

---

<Suggested skill to invoke, e.g., "Use /implement to continue">

### Step 8: Extract Learnings (Optional)

If significant learnings occurred this session, also run post-mortem:

```bash
# Check if post-mortem skill should be invoked
# (if >3 commits or major decisions made)
git log --oneline --since="2 hours ago" 2>/dev/null | wc -l

If ≥3 commits: Suggest running /post-mortem --quick to extract learnings. If <3 commits: Handoff alone is sufficient; learnings are likely minimal.

Step 9: Report to User

Tell the user:

1. Handoff document location
2. Continuation prompt location
3. Summary of what was captured
4. Suggestion: Copy the continuation prompt for next session
5. If learnings detected, suggest /post-mortem --quick

Output completion marker:

<promise>DONE</promise>

If no context to capture (no commits, no changes):

<promise>EMPTY</promise>
Reason: No session activity found to hand off

Example Output

Handoff created:
  .agents/handoff/20260131T143000Z-auth-refactor.md
  .agents/handoff/20260131T143000Z-auth-refactor-prompt.md

Session captured:
- 5 commits, 12 files changed
- Paused: mid-implementation of OAuth flow
- Next: Complete token refresh logic

To continue: Copy the prompt from auth-refactor-prompt.md

<promise>DONE</promise>

Key Rules

• Capture state, not just summary - next session needs to pick up exactly where we left off
• Identify blockers clearly - don't leave the next session guessing
• List files explicitly - paths, not descriptions
• Write the continuation prompt - make resumption effortless
• Cite everything - file:line for all references

Integration with /post-mortem

Handoff captures state for continuation. Post-mortem captures learnings for the flywheel (full knowledge lifecycle).

For a clean session end:

/handoff              # Capture state for continuation
/post-mortem --quick  # Extract learnings for future

Both should be run when ending a productive session.

Without ao CLI

If ao CLI not available:

1. Skip the ao ratchet status check in Step 2
2. Step 8 retro suggestion still works (uses git commit count)
3. All handoff documents are still written to .agents/handoff/
4. Knowledge is captured for future sessions via handoff, just not indexed

---

Examples

Paused Mid-Implementation

User says: /handoff (after working on OAuth flow for 2 hours, need to stop)

What happens:

1. Agent detects recent commits (5 commits in last 2 hours, auth-related)
2. Agent checks in-progress work with bd list (issue #42 still open)
3. Agent identifies pause point: "Completed token generation, about to start refresh logic"
4. Agent lists key files: auth.go, token.go, research doc, plan doc
5. Agent writes handoff document with accomplishments and pause state
6. Agent writes continuation prompt with clear next action
7. Agent checks commits (5) and suggests running /post-mortem --quick to extract learnings

Result: Handoff captures state, continuation prompt ready, post-mortem suggested.

Between Tasks, Clean State

User says: /handoff (just closed issue #40, about to start #41 next session)

What happens:

1. Agent detects 1 commit (closed issue #40), no pending changes
2. Agent identifies pause point: "Between tasks. Last: closed #40 (fixed rate limiting). Next: start #41 (add JWT refresh)"
3. Agent lists files from #40 (middleware.go, config.go)
4. Agent writes handoff with accomplishment summary and next-task preview
5. Agent writes continuation prompt with /implement #41 suggestion
6. Agent skips post-mortem suggestion (<3 commits)

Result: Handoff captures clean boundary, continuation is simple.

Auto-Derived Topic

User says: /handoff (no topic provided, agent derives from commits)

What happens:

1. Agent reads recent commits: "feat: add rate limiting", "fix: token expiry"
2. Agent derives topic slug: "rate-limiting" (from most recent commit)
3. Agent creates handoff files with derived topic in filename
4. Agent reports: "Handoff created: .agents/handoff/20260213T143000Z-rate-limiting.md"

Result: Topic auto-derived from git history, no user input needed.

---

Troubleshooting

Problem	Cause	Solution
"No session activity found to hand off"	No commits, no file changes detected	Expected for idle sessions. Nothing to hand off. Start new work or skip handoff.
Handoff files not written	.agents/handoff/ directory does not exist or not writable	Run mkdir -p .agents/handoff or check directory permissions
Topic slug is generic "session-1430"	No descriptive commits or issues to derive topic from	Provide explicit topic: /handoff auth-refactor for better naming
Continuation prompt missing key context	Recent files or artifacts not listed in handoff	Manually add missing files to handoff document or re-run with explicit topic
Post-mortem suggested but no learnings	Agent sees ≥3 commits and auto-suggests /post-mortem --quick	Run /post-mortem --quick or skip if commits are trivial (agent can't judge learning quality, only commit count)

---

Reference Documents

• references/handoff.feature — Executable spec: topic derivation, evidence-based accomplishments, pause-point + in-progress capture, next-files list, structured doc + continuation prompt (soc-qk4b)

See Also

• skills/post-mortem/SKILL.md — Full validation + knowledge lifecycle (council + extraction + activation + retirement); --quick quick-captures a single learning (folded the retired /retro)