The owner's personal workstation setup for macOS and Linux (zsh config, setup and hardening scripts, git and SSH tools, encrypted secrets, Claude Code launchers), plus an older Ansible playbook for remote host operations. Public on GitHub (CLAUDE.md, README.md).
Status: active (last change 2026-11-19). About 117 commits in November 2026 after a quiet summer; the Ansible half last changed 2026-07-31.
- A reproducible workstation: re-runnable macOS and Linux setup that asks every question
first, then runs unattended, with a
DRY_RUN=1mode and a list of what was skipped (7536668, 73510e1, 50d69f2, 6506f10, CLAUDE.md "Local Config Sync"). - Hardening that reports before it changes anything:
harden.shchecks by default and applies only with--apply, mapped to mSCP/CIS (macOS) and CIS/KSPP/ssh-audit (Linux) (30e13a3, 749a788, 80779e5). - Secrets never on disk, argv or screen: the
secrettool (sops + age, per-device keys in the OS keychain) replaces plaintext env files (20eab44, 82022f7, 5a0f636, 0ae1f47). - Claude Code as the only coding harness, on Anthropic or on private model servers through
one registry (
shared/llm/models.conf) andclaude-z, with Paperclip, Hermes and Orca lanes on the same backends (ef8de31, cb3d5c5, 816c26d, 0cf10ea, CLAUDE.md "Claude Code backends"). - Per-project AI policy enforced before any model request (Claude Code hook, orca-lanes, OpenCode plugin), failing closed (71fa985, 85b785c, docs/ai-policy.md).
- Tested and portable: a regression suite in bash and zsh with CI on macOS and Ubuntu, and
dotsto find and document every function (fc0d601, 9f57f07, tests/run.sh). - A company workstation agent stack kept apart from the personal one: scripts that set up company Hermes profiles and a Paperclip team, a read-only monitor, and tools that report a task board from its own records (a05210e, ab1049c, 3e03988, a0ab33f, docs/company-hermes.md).
- Keep
tests/run.shgreen on both OSes and add a test that fails on the broken code with every bug fix (CLAUDE.md "Tests", fc0d601). - Keep the coding-stack launchers working across machines and upgrades: tunnels only to reachable servers, Paperclip surviving Node upgrades, IDs looked up rather than copied (918e46e, 15dd4d0, 11a6bdb).
- Extend the AI policy as new providers are added; every new backend gets a vendor and service entry and its deny rules (85b785c, d0f0df0, docs/ai-policy.md) (inferred).
- Decide the Ansible half's target before any run: the default inventory points at an address no longer on the tailnet (CLAUDE.md warning, verified 2026-11-09).
- Keep
dots --lintempty: every user-facing function gets a one-line description (CLAUDE.md "Finding commands"). - Change the company agent setup in the setup scripts with a test, never by hand on the
device: a refresh reverts hand edits. Keep what the scripts write into each role's
config.yamlbyte for byte unless a requalification is planned (docs/company-hermes.md, a0ab33f).
- Public repo: no
.env,inventory/custom/vars.yaml, keys or secret values; the secret store lives in a private repo or Syncthing folder (CLAUDE.md, .stignore, 8349ee1). - Secrets never in argv, files or terminal output;
rundecrypts into the one command that needs them (20eab44, CLAUDE.md "Secrets"). - Claude Code is the only coding harness; other front ends drive it (CLAUDE.md "Claude Code backends", owner's global instructions).
- Alibaba models and services are denied for IrregularChat and ZenithFlow; Lumo is denied for ZenithFlow (71fa985, 85b785c).
- Hardening and setup default to read-only or dry run; Linux hardening writes only its own drop-ins and skips settings that break Flatpak, snaps, Docker or IPv6 (CLAUDE.md).
- Shell files stay bash-compatible because Linux sources them with bash; no shadowing of real commands except deliberate wrappers (CLAUDE.md "Finding commands").
- Canonical zsh files are in
shared/zsh/;macos/.zshrcis a stale copy (CLAUDE.md). - Ansible:
--checkfirst, never an unfamiliar tag combination against production first, andansible <host> -m pingbefore anything heavy (CLAUDE.md "Safety Notes"). - Downloaded installers are pinned and checksum-verified; Flathub only for verified publishers (CLAUDE.md "Local Config Sync").
- Review-only for the Hermes ship service: bot changes open a merge request and are never merged automatically (ai-coding-env servers/hermes-coding/SHIP.md, e60e44a there).
- Company device details (addresses, ids, private paths, task text) never appear here; they stay in the device's private workspace (docs/company-hermes.md).
- On a work device the main checkout is live: a monitor and a scheduled board intake run from it. Work in a separate worktree; never switch its branch or leave edits in it.
- 2026-11-19: A task board can specify a rough request into one card instead of decomposing it: a private input turns decomposition off in the root configuration only, and a one-minute intake job specifies Triage cards (a0ab33f, e8b5c5e).
- 2026-11-19: Read-only tools report a board's outcomes by who wrote the card and audit a worker's recorded tool calls against its contract (3e03988, fd11992).
- 2026-11-19: A read-only audit of the company agent stack runs every five minutes (ab1049c).
- 2026-11-18: Company Hermes profiles can run as a Paperclip team through request graphs (a05210e).
- 2026-11-16: Add Lumo as a Claude Code and OpenCode backend, denied for ZenithFlow (85b785c).
- 2026-11-12: Per-project AI policy; Alibaba denied for IrregularChat and ZenithFlow (71fa985).
- 2026-11-12: One model registry and
claude-z; drop the Vibe, DeepSeek and Mistral launchers from the aliases (claude-switchkept as legacy) (ef8de31). - 2026-11-10: Encrypted secrets with sops + age and per-device keys; not Secure Enclave keys, because Linux devices could not use them (20eab44).
- 2026-11-09: Regression suite and CI on macOS and Ubuntu (fc0d601).
- 2026-11-09: Hardening scripts read-only by default with explicit
--apply(30e13a3, 749a788). - 2026-07-31: Version-control the Proxmox host's Docker runtime recovery and alerting in
ansible/files/(81d5e3a, c8bca94). - 2026-06-03: Add CLAUDE.md as the Ansible quick reference; remove an env backup (111aacd, ee44408).
- Find out whether the thing already exists before adding it: search the repository and
the decisions above. Lumo, for example, is already a Claude Code and OpenCode backend
(
cc-lumo, docs/claude-lumo.md). - An application for a new machine goes into the question-first setup scripts
(
macos/initial_macos_setup.sh,linux/initial_linux_setup.sh), beside the nearest existing example, and must behave underDRY_RUN=1(CLAUDE.md "Local Config Sync"). A download is pinned and checksum-verified, or comes from Homebrew or a verified Flathub publisher. A "manual install" line with an unpinned link does not meet that. - Every change comes with a test in
tests/run.shthat fails without it; run the whole script before pushing (priority 1). - Open a pull request from a branch named
<feat|fix|docs>/<slug>-YYYYMMDD; nothing is committed tomaindirectly. - When the request does not say which of two readings is meant, take the one the nearest existing example supports, say so in the pull request, and leave the other as a question there rather than stopping.
- What is the future of
ansible/: fix the inventory for the current Proxmox host, or retire it in favour of the ai-coding-env scripts? Several tracked files in this public repo name internal addresses and host paths. - Should the old clutter in
docs/(test_*.txt, demo-file.txt, older guides) be removed? CLAUDE.md calls some of it stale. - Is
claude-switchdue for removal now that no alias uses it?
- CLAUDE.md (main reference: setup, secrets, backends, tests, tools)
- docs/ai-policy.md, docs/orca-lanes.md, docs/claude-lumo.md, docs/claude-opencode.md, docs/opencode-local.md
- ansible/USAGE_GUIDE.md, ansible/ANSIBLE_REVIEW_AND_PLAN.md
- tests/run.sh