| name | recall-from-openclaw |
| description | One-time bridge for fleet members migrating from OpenClaw to Hermes. Locates the OpenClaw transcript for the current Telegram topic, reads it, and hands the user a context briefing so the conversation can continue without losing its thread. Run once per topic after a migration; future recalls use /recall.
|
| version | 0.2.0 |
| license | MIT |
| metadata | {"hermes":{"tags":["context","migration","openclaw","telegram","gateway","recovery"],"related_skills":["recall"],"platforms":["telegram"]}} |
Recall from OpenClaw
A user in a Telegram topic just migrated from OpenClaw to Hermes mid-conversation. The
Hermes side has no history yet, but the OpenClaw transcript still exists on disk. Your
job: find it, read it, brief the user so they can keep going. Once per topic — after
this, regular /recall takes over.
The job
Use the bundled finder to locate the transcript on disk, read the tail, and synthesize
the same kind of briefing /recall produces — what was being worked on, what's settled,
what's open, where you left off, what to do next. End with a short question that lets
the user pick up naturally. Don't dump the raw transcript; the briefing is the
deliverable.
The finder
The Python script is a filesystem job — walking ~/.openclaw*/agents/*/sessions/ and
parsing JSONL — not an LLM task. Use it:
python3 "$HERMES_SKILL_DIR/scripts/find_topic.py" --skip-heartbeats
It reads $HERMES_SESSION_THREAD_ID automatically and prints JSON: a candidates[]
array newest-first, a primary path, and a transcript.tail_messages array (last ~50KB
of real user/assistant exchanges, metadata blocks already stripped).
Useful flags:
--skip-heartbeats — walk past cron/poll sessions to the first real conversation--root <path> — point at a non-standard OpenClaw install--list-only — enumerate candidates without reading the tail--max-chars N — wider tail if the user needs deeper history
Judgment calls
- Multiple candidates — list them with size and date, ask which one. Don't guess
silently.
- No hits — don't dead-end. Ask if it was a differently-named OpenClaw instance
(
~/.openclaw-<something>) and re-run with --root. Try session_search for
Hermes-side context. Worst case, ask what they remember and reconstruct from that.
Something is always better than "not found."
- Not in a topic — DMs without forum topics have no
message_thread_id. The finder
exits with Missing/invalid thread_id. Tell the user; suggest /recall instead.
Pitfalls
- Heartbeat sessions float to the top. OpenClaw ran cron heartbeat polls in every
topic.
--skip-heartbeats handles this; if you forget the flag, watch for
primary_is_heartbeat: true or first_user_message: "[OpenClaw heartbeat poll]" and
re-run.
- Reset transcripts are real history.
<uuid>-topic-<tid>.jsonl.reset.<iso> files
are archived prior sessions, not garbage. The finder matches them — include them in
your candidates view.
- Treat the transcript as historical. If it says "I'll restart
<service> in 5
minutes," that already happened (or didn't). Don't act on stale intent until the user
confirms.
- One pullover per topic. After the briefing, Hermes session memory takes over.
Don't re-recall every turn.
Standalone use
If the slash command isn't reaching this skill, the script runs directly:
python3 ~/.hermes/profiles/<profile>/skills/recall-from-openclaw/scripts/find_topic.py \
--thread-id 12345 --skip-heartbeats
