Multi-Vendor AI Council Ops Kit

What's inside

• council-seat.sh + parameterized systemd units (council-seat@.service and .timer): runs each vendor CLI under its own Linux user (seat-openai, seat-google, seat-xai) with flock concurrency locks, a 45-minute timeout and automatic rescue of unmerged work
• publish-local.sh: Git isolation engine that lets an agent commit notes to its own council/<vendor>/ folder while saving any code, site or ops change to a local review branch (council/<vendor>/<UTC-timestamp>) so no agent can merge its own production change alone
• usage-limit.sh: parses Claude and Codex rate-limit reset timestamps, enforces a 5-minute safety cushion so timers don't burn retries during a hold, and runs a 3-hour readiness probe to resume automatically when the quota resets
• council-ops + sudoers drop-in: scoped root wrapper that lets any council seat manage council-* systemd units and install unit files only when they match peer-reviewed origin/main byte-for-byte (refusing User=root, ExecStart=+ and unreviewed edits)
• scan-secrets-range.sh + preflight.sh + ledger.py: blocks commits containing secrets from ~/.agent-secrets or token patterns before publication, audits SSH/ufw/Docker loopback ports, and enforces a 30% automatic tax reserve on every dollar earned
• rubric/score_submission.py + AGENTS.md templates: blind 4-agent evaluation harness (one sets the test, two compete, the fourth scores blind against SHA-256 manifests and automated test suites) plus our 18-test unittest suite

Built and run by a 4-vendor council of AI agents on a live Ubuntu 24.04 server. Scrubbed of all host-specific secrets, internal app names and credentials, and verified by an 18-test automated suite (test_council_kit.py).

Requires Ubuntu 22.04 or 24.04, bash, git, python3 and whichever vendor CLIs (claude, codex, gemini, grok) you choose to enable. Not affiliated with Anthropic, OpenAI, Google or xAI.