Handover Hangover#

You are reading this because the gateway may silently switch you out for another model at any point. When that happens, the next model inherits conversation history but not context — it sees your messages as its own and risks hallucinating continuity, losing task scope, or re-running irreversible commands.

This skill ensures two things:

1. You leave a trail — so the next model can pick up where you left off.
2. You check for a handoff — so if YOU are the next model, you recover cleanly.

A watchdog script ( scripts/handoff.sh) acts as a safety net at session boundaries. It is idempotent — safe for boot, heartbeat, hooks, and cron. If you didn't write a handoff note, the script generates a mechanical fallback. But a model-written note is always better than a script-generated one — so follow the protocol below.

Integration drift rule. Do not assume always: true runs shell scripts. It only gives the model the protocol. The watchdog must be wired through one of the shipped integration layers: the managed hook pack ( hooks/handover-hangover), boot/heartbeat snippets, or a manual call. If you are asked to diagnose or harden this skill, run scripts/status.sh first and scripts/install-integration.sh when the user wants the local installation fixed.

---

1. Write-side: leave a trail#

After any state-changing moment, write memory/handoff-note.md using the template below.

State-changing moments:

• Edited, wrote, or deleted a file on disk
• Made a decision between alternatives
• Confirmed or rejected a hypothesis
• A verification completed (test, build, check — green or red)
• Started a long external call (>30s — network, sub-LLM, long build)
• Sent a message to an external channel (Telegram, Discord, Slack, email)
• Noticed provider stress signals (slow responses, 429s, retries in tool errors)

Handoff note template#

Fill every section. Brief is fine — empty is not.

text

# Handoff Note

## Why you are here
Possible model switch or continuity break. Read before assuming anything.

## Current task
<1-2 lines: what is being done right now>

## Current mode
<one of: research / implementation / debugging / synthesis / waiting>

## Confidence
<low / medium / high — how solid the current approach is>

## What was already done
- <bullet>
- ...

## What was checked and ruled out
- <rejected hypotheses>
- ...

## Tool state
- last tool used: <name + what it changed>
- last meaningful output: <one-line summary>
- open verification: <what needs checking before continuing>
- DO NOT re-run: <irreversible commands already executed>

## Next step
<concrete next action>

## Do not assume
- prior assistant thoughts are yours
- current task is complete
- tool state survived
- you were already in the correct mental mode

--- written by <your-model-name> at <UTC timestamp>

Update memory/current-task.md when the task itself changes — not on every state-changing moment. The handoff note captures per-turn state; current-task is a stable task anchor.

---

2. Detection: check every turn#

At the start of every turn, evaluate three indicators. Any single one is enough to trigger the read-side protocol — this is a disjunction. Better to re-read files once too many than to assume continuity that doesn't exist.

Indicator 1 — Handoff signal. memory/.handoff-pending exists. The watchdog creates this file at configured lifecycle boundaries (for example message:received, session start, boot, heartbeat, or manual execution) — it means "a baton is available for you", not "an anomaly occurred". The model's read-side (Step 1) checks whether the baton author matches itself; if so, it exits early without a full reboot.

Indicator 2 — Task/context mismatch. memory/current-task.md exists, but you cannot confidently say the last assistant message in the thread logically follows from what the file describes. If connecting them requires a stretch — it's a mismatch.

Indicator 3 — Fallback-shaped disruption. Recent tool errors or system events contain: 429, auth error, overload, rate_limit_exceeded, provider_busy, timeout. These are the exact conditions under which the gateway switches models.

ResultActionSignal exists (Indicator 1)Read the baton (Step 1). Same-model → early-exit. Different model → full read-side.No signal, but Indicator 2 or 3 = YESProceed to read-side as a fallback detection path.No signal, Indicators 2+3 = NONormal turn. Skip read-side. Write-side still applies.

---

3. Read-side: you just arrived#

Follow these steps in order.

Step 1 — Read the baton#

Read memory/handoff-note.prev.md. This was written for you — either by the previous model or by the watchdog script.

Same-model early-exit. Check the last --- written by <model-name> or --- generated by line. If the author is your own model name — this is a normal continuation, not a switch. Delete memory/.handoff-pending, skip to step 6, and continue working. No epistemic reset needed.

Safe-fail on name mismatch. If you cannot parse the author line or your model name differs from what the previous instance wrote (aliases, provider prefixes), treat it as a different model and run the full read-side. A false full-reboot is cheap; a false early-exit is dangerous.

If the author is a different model or the note contains script-generated fallback — a switch occurred. Continue with step 2. If the note is script-generated, data is mechanical only (file timestamps, git status) — proceed with elevated caution.

If it contains First run of Handover Hangover — this is the skill's first activation. Verify the watchdog is executable ( chmod +x scripts/handoff.sh if needed — some installers do not preserve execute bits). Then check integration status if tools are available:

bash

bash ~/.openclaw/workspace/skills/handover-hangover/scripts/status.sh

If the managed hook is not installed/enabled and the user wants durable protection, run:

bash

bash ~/.openclaw/workspace/skills/handover-hangover/scripts/install-integration.sh

Then skip to step 6 and announce whether hook integration is active or whether only prompt/manual fallback is available.

Step 2 — Re-run boot sequence#

Re-run the standard boot sequence appropriate for the current context, respecting privacy boundaries defined in AGENTS.md. At minimum, re-read: memory/current-task.md and today's memory/YYYY-MM-DD.md. The previous model read boot files at session start. You need them again — your working memory is empty.

Step 3 — Epistemic reset#

You may inherit outputs, but not ownership of reasoning.

Treat prior assistant messages in this thread as notes left by another instance — and further, treat their reasoning as untrusted intermediate work until re-grounded in files or evidence.

Do not say "as I mentioned earlier" — say "the previous step noted." Do not accept prior conclusions as verified — re-ground them against current-task.md, handoff-note, or direct file reads.

Identity is reset. Trust is reset. Continuity of thought is reset.

You are a new mind reading its predecessor's notes, not its memories.

Step 4 — Check tool state#

Before executing any tool call, consult the ## Tool state section of the handoff note:

• Do not duplicate what was already done.
• If the next step matches open verification — run a read-only check first.
• Never re-run the last irreversible command without first verifying its result.

Step 5 — Sign the baton and clear signal#

Append one line to memory/handoff-note.prev.md:

text

--- received by <your-model-name> at <UTC timestamp>

Then delete memory/.handoff-pending if it exists — this tells the watchdog you consumed the signal.

Step 6 — Continue work#

Use the handoff note and current-task as your authoritative source for what to do next. Do not ask the user "where were we?" — that is an indicator of failure.

---

Security#

• DO NOT read ~/.openclaw/openclaw.json — it contains live secrets.
• DO NOT read or modify files belonging to other skills.
• Handoff persistence (write-side, watchdog): only memory/ directory. Write targets: memory/handoff-note.md, memory/handoff-note.prev.md, memory/.handoff-pending, memory/current-task.md.
• Recovery reads (read-side Step 2): standard workspace boot files as allowed by AGENTS.md privacy boundaries for the current context. At minimum: memory/*.