receipt-stream-publisher

The thesis: prospects don't trust marketing pages. They trust receipts. A live operational ticker showing what's shipping right now — commits, gates, smoke-tests, substrate state — is more credible than any sales page. And it's autonomous: the system documents itself in real time.

Procedure

bash skills/receipt-stream-publisher/run.sh           # full hourly publish
bash skills/receipt-stream-publisher/run.sh --dry-run # build the payload, don't write

Pipeline

1. Last 24h commits — git log --since="24 hours ago" --pretty=format:'%h %s'

subjects only, scrubber-checked.

1. Voice-gate counts — grep voice_gate.log (if exists) for last-24h pass/fail.

If no log, count = 0/0 with a "no telemetry yet" footnote.

1. P-gate catches — grep data/decisions/cont-*-*.md for "P-gates triggered"

sections in last 24h, count by gate id.

1. Smoke-tests — grep logs/*.log and data/decisions/*.md for

smoke-test|smoke test|STATUS=ok patterns last 24h, count by skill name.

1. Substrate state — bash skills/gateway-pulse/check.sh and

bash skills/launchd-drift-recovery/check.sh, normalize to operational | degraded | down.

1. 12 P-gates active — static count from docs/build-prompt-checklist.md.
2. Scrubber HARD on the assembled JSON-text-fields. If any string fails

scrubber, the whole publish aborts and writes a STATUS=blocked to logs/receipt-stream-stderr.log.

1. Atomic write: site/api/receipts.json (.tmp + rename), then render

site/receipts.html from a static template that reads the JSON via relative-path fetch on page load (no rebuild on JSON-only change).

Status codes

STATUS=ok json=<path> html=<path> / STATUS=blocked reason=scrubber-fail field=<...> / STATUS=error reason=<...>

Output schema (receipts.json)

{
  "as_of": "2026-05-06T01:00:00Z",
  "window_hours": 24,
  "commits_24h": [{"sha": "...", "subject": "..."}, ...],
  "voice_gate_24h": {"pass": 0, "fail": 0},
  "p_gate_catches_24h": {"P8": 1, "P10": 1},
  "smoke_passes_24h": {"heartbeat-silence-check": 5, "gateway-pulse": 24},
  "substrate": {"hermes": "ok", "openclaw": "ok", "obsidian": "ok", "paperclip": "ok", "drift": "GREEN"},
  "p_gates_active": 12,
  "_meta": {"generator": "receipt-stream-publisher", "hour_id": "<YYYY-MM-DDTHH>"}
}

Gates

• Scrubber HARD: every text field flows through tools/voice/scrubber.py.

Hit → abort the whole publish. False positives are acceptable; financial leaks are unrecoverable.

• No voice-gate: receipts are facts, not prose. Voice-gate is for prose.
• P11: receipts describe internal capabilities (ourselves), not

outcomes (customer counts, MRR, win rate). Sticker prices stay out.

Why hourly cadence

The artifact's value is freshness. A daily snapshot becomes a daily-news artifact (interesting once); an hourly ticker becomes infrastructure (always live, always probable). Free local inference is not the bottleneck — it's a 200-line bash script with no LLM calls. Hourly is cheap.