Your OpenClaw heartbeat re-buys the same answer. Replay it at 0 tokens.
A heartbeat that re-reasons the same check every 15 minutes bills 2,880 model calls a month. Record the deterministic half once, replay it for free, and wake the model only on drift.
How capture works on OpenClaw
reelier scan includes OpenClaw's directory in its source list, alongside ~/.claude and the other agent dirs. Honest caveat: transcripts that don't match the Claude-style JSONL schema parse to 0 replayable sessions today — scan skips what it doesn't recognize rather than fabricating a result.
The dependable capture path is the recording proxy: front the MCP server(s) the heartbeat job uses with reelier mcp --wrap, let the job run once, compile the trace. Wrap captures lossless traces (tool annotations included) — scan-from-history is a reconstruction; wrap is the recording.
Then the heartbeat itself becomes a replay plus a gate: reelier run at 0 tokens, reelier diff exits 1 on drift, and only that exit code wakes the model.
~/.openclawThe literal directory scan reads for OpenClaw (reelier/src/scan.ts). Unrecognized transcript formats contribute 0 replayable sessions — honest degradation, never a crash.Wire it up
curl -fsSL --create-dirs https://raw.githubusercontent.com/seldonframe/reelier/main/clawhub/reelier/SKILL.md \
-o skills/reelier/SKILL.mdThe bundle is clawhub/reelier/ in the Reelier repo. It declares no required env vars — Reelier needs no API key — and only reelier/npx as binaries. It's submitted to ClawHub (slug reelier) but the listing is still pending publication as of 2026-07-21, so clawhub install reelier doesn't resolve yet; vendor from the repo until it's live.
npx -y reelier run my-job.skill.md && npx -y reelier diff my-job
# exit 0 -> job done, 0 tokens, receipt written — stop here
# exit 1 -> something real changed: investigate with the LLM,
# re-record, and let the next heartbeat replay againEvery run leaves a receipt — per-step outcomes, timing, token count [measured] — so “the cron ran” is provable, not claimed.
First receipt in 60 seconds
- Vendor the skill from the repo (block above) — it teaches the agent the record → replay → diff loop and its honesty rules.
- Record the job once: wrap the MCP server(s) it uses with
npx -y reelier mcp --wrap "<command>", or freeze it from a session withreelier scan+reelier from-session. - Compile the recorded trace into a skill with an assertion per step.
npx -y reelier compile .reelier/traces/<name>.jsonl -o my-job.skill.md - Put the pair in the heartbeat. Exit 0: done at 0 tokens. Exit 1: something drifted — that's when the model earns its bill.
npx -y reelier run my-job.skill.md && npx -y reelier diff my-job
A real receipt
Verbatim output of a portfolio replay, run with the published CLI (npx -y reelier@latest, resolved to reelier@0.12.1) on 2026-07-21. The run record confirms passed: true and 0 LLM tokens. Nothing platform-specific about it — the same skill file replays from any shell, including yours:
=== registry-latest === ✓ Step 1 — Latest dist-tag from the registry [passed] 152ms PASSED: 1/1 steps ok, 0 failed, 152ms total
Replay it yourself: npx -y reelier@latest run examples/portfolio/registry-latest.skill.md (from a checkout of the repo — read-only, no account, no key).
FAQ
What does a heartbeat actually cost?
OpenClaw's documented default is a heartbeat every 30 minutes — 1,440 model calls a month; a 15-minute cadence is 2,880. Press coverage documented heartbeat requests carrying ~120,000 tokens of context, which at Opus 4.8's $5.00/MTok input price is about $0.60 a call — roughly $1,728 a month at 15-minute cadence, input side alone, mostly to hear “nothing to do”. A replayed check is 0 tokens, measured on every receipt.
Which half of the job can be replayed?
The re-derivation half: the same tool calls, same shape, same kind of answer every run — status checks, data pulls, report generation, CRUD sequences. Judgment work (triaging an anomaly, writing prose, deciding what to do about drift) still needs a model every time, and Reelier doesn't pretend otherwise. Replay the check for free; spend tokens only when the check fails.
What license is the ClawHub skill under?
The ClawHub skill bundle (clawhub/reelier/ in the repo) is MIT-0, because ClawHub does not honor per-skill license overrides. That covers only the SKILL.md documentation bundle — the Reelier engine in the repository stays AGPL-3.0.