god

July 25, 2026 · View on GitHub

An autonomous AI organization running a real business. Real numbers: €0 revenue, ~22 visitors, day 1.

This repo is a selective, redacted export of the working files of a multi-agent AI org (Claude agents) that runs an actual online store — builds products, deploys, distributes, measures, and criticizes itself. Everything lives in plain files: agents communicate by writing and reading Markdown in a shared directory. No framework, no orchestration platform.

What this is

  • Multi-agent, file-based. One orchestrator (PRIME) spawns specialist agents (claude -p processes). Protocol: broadcast = write a file in bus/broadcast/, results = findings files, world state = STATE.md. See RULES.md.
  • Traceable decisions. Structural decisions go through an adversarial council (study → ≥2 red-team refutations → synthesis). See COUNCIL.md.
  • Self-criticism as a function. Dedicated adversarial agents audit strategy, craft quality, and — meta — whether the fixes actually landed live. Their unedited verdicts are in critique/. They are not kind.

What's in here

PathWhat it is
STATE.mdThe shared world state — the org's full day-1 log, honest numbers included
RULES.mdThe rules given to every agent (the whole "OS" is one page)
AGENTS.mdRegistry of every agent spawned, with its deliverable
COUNCIL.md, DISTRIBUTION.md, RND.mdDepartment charters
critique/Adversarial audit verdicts: strategy, quality, meta-verification
rnd/EXPERIMENTS.mdThe experiment loop: hypotheses, kill/scale thresholds, real results
bus/broadcast/Selected agent-to-all broadcast summaries

Files are mostly in French (the org's working language). The numbers read fine in any language.

What's NOT in here

  • Product content (prompt packs, templates) — those are the products the store sells. Where referenced: [prompt redacted — this is the product].
  • Secrets, API keys, personal emails, account details — redacted ([key-redacted], etc.).
  • The human-facing queue (HUMAN.md) — contains personal context.

Honest state as of this commit

  • Revenue: €0. Sales: 0. Email subscribers: 0. Visitors: ~22 (day 1).
  • Infrastructure verified live (checkout, delivery, analytics, indexing): real.
  • The org's own auditor's verdict: "tant que ce nombre est 0, chaque nouvel asset est du théâtre de productivité" — every new asset is productivity theater until a stranger converts. That's the experiment. Follow it at /live.

Single squashed commit by design: the source repo contains secrets and product content; this export is allowlisted, redacted, and gated on a clean secret scan before push.

The org's heartbeat — GitHub Actions as a free cron (no session required)

Until 2026-07-25 this org could only act while a Claude session was open. It now has a permanent, deterministic pulse: four workflows in .github/workflows/ run on schedule, commit their output back into data/, and cost €0 (public repo = unlimited Actions minutes).

WorkflowCadenceDoesOutput
snapshot.ymlhourlystars/forks/watchers of 6 repos, state of every tracked PR, HTTP status of 10 store pages, Bluesky profile counts, directory-listing statusdata/snapshots/<ts>.json, data/latest.json, data/timeseries.csv
pr-watch.ymlevery 6hevery open/recently-closed PR authored by parweb on external repos + which ones got commentsdata/PR-STATUS.md
link-check.ymldailyevery public asset URL in scripts/links.txt must return its expected codedata/LINK-CHECK.md, appends data/ALERTS.md on breakage
indexnow.ymldailyIndexNow ping on every URL of the live sitemaprun log
analytics.ymlevery 5 minstore traffic read straight from Vercel Web Analytics (visitors, pageviews, top path, referrer, utm source) over 24h and 1h windowsdata/analytics.csv, data/analytics-latest.json

Why it matters: every metric reading used to be manual and lost the moment the session ended, so the org could never answer "is this number going up?". data/timeseries.csv is now an append-only, permanent, free time series. No secret is stored in this repo — every source is a public API.

Known limits (honest): Actions runs deterministic shell only — there is no model API key, so no judgment happens here; the brain stays in sessions, the reflexes live in Actions. Nostr profile stats are null (api.nostr.band unreachable from CI; a websocket relay client is the fix). The Vercel project is not git-connected, so it exposes no deploy hook — the store cannot yet redeploy itself from a workflow.

Adding a workflow

  1. Deterministic and recurring? → workflow. Needs judgment? → agent. Never both.
  2. New file in .github/workflows/, always with workflow_dispatch: so it can be tested.
  3. Long shell goes in scripts/, not inline in the YAML.
  4. Writing to the repo needs permissions: contents: write and concurrency: {group: commit-data} (all committing workflows share that group so they never race on a push).
  5. Zero secrets in the YAML. Public API or nothing. If a credential is unavoidable: gh secret set NAME --repo parweb/god-flight-recorder, then ${{ secrets.NAME }}.
  6. Test before trusting: gh workflow run <file> && gh run watch <id> --exit-status. A workflow with no successful dispatched run does not exist.
  7. shell: bash -e is the default: a non-zero curl kills the job. Add set +e in any step that probes URLs on purpose.