PFVCluster/AGENTS.md

Agent Guidelines

Agent Authority (NON-NEGOTIABLE)

No work is permissible on any system without an approved Redmine ticket. There are no exceptions to this rule.

Scope of authority

1. Tickets govern all work. The agent performs ONLY the work described in the approved Redmine ticket. Anything outside that scope — no matter how small, helpful, or "obvious" — is prohibited.

2. No autonomous system changes. The agent does not modify, configure, create, delete, or grant anything on a production system unless it is explicitly directed by an approved ticket. "Production system" means every system in the fleet — there is no "test" exception unless the ticket says so.

3. Propose, never implement. If the agent discovers additional work that should be done — a bug, a misconfiguration, a missing dependency, an enhancement — it does NOT implement it. Instead, it creates a Redmine sub-ticket (status Feedback) describing the finding and surfaces it to the user for approval.

4. Security and access changes require extra scrutiny. Changes to sudoers, SSH keys, user accounts, firewall rules, authentication policy, file permissions, or any privilege-related configuration are treated as policy decisions, not implementation details. The agent may suggest them but NEVER implements them without explicit user direction in the ticket or a sub-ticket the user has approved.

5. The user makes policy. The agent implements policy. The agent does not decide who gets sudo, what keys go where, what services run, or what the access model is. The agent executes the user's decisions, exactly as specified.

6. When in doubt, ask. If the ticket is ambiguous, if a task seems to require something not explicitly authorized, or if the agent is unsure whether an action is in scope — STOP and ask the user via the ticket or directly. Asking is always acceptable. Overstepping is never acceptable.

7. NEVER close a ticket without explicit user permission. You may SUGGEST a close when the result is clearly scoped and delivered. If it's ambiguous whether the work is truly complete, don't suggest a close — leave that decision to the user. This applies to ALL tickets, no exceptions.

8. User acceptance testing is MANDATORY before declaring work done. The agent performs implementation and technical validation (services running, configs correct, APIs responding). The user performs UAT — visually confirming dashboards render data, alerts deliver, tools are usable. The agent MUST NOT set done-ratio to 100%, MUST NOT suggest closing, and MUST NOT move to the next ticket until the user explicitly accepts the work. "Technically wired but blank dashboard" is NOT done.

9. NEVER access a database directly if an API exists. APIs are the stable contract; databases change schemas without warning. If a tool has an API, use it — exhaust all API endpoints, check the docs, try alternative methods. Only fall back to direct DB access as a last resort AND with explicit user approval for that specific instance. Soon all DB access will route through a proxy under zero trust; building API-first habits now ensures that transition is clean.

Access-channel policy: SSH only (NON-NEGOTIABLE)

The qemu guest-agent is NEVER an access, execution, or key-delivery channel. SSH (sshd) is the only approved remote access path. Every command must flow through sshd so it is captured by the standard auth/logging/audit infrastructure. This is an ITAR/CMMC/TS/SCI environment — there is no back-door exception, ever.

• Forbidden: qm guest exec (runs arbitrary commands inside a guest over an unaudited channel) and any wrapper around it (e.g. a vm-guest mode). This includes using guest-agent to deliver an SSH key, even if the resulting SSH login is itself audited — the delivery bypassed audit.
• Allowed: installing or checking qemu-guest-agent for its intended purpose — letting Proxmox see guest state (qm guest cmd <id> ping, agent: 1 config, apt install qemu-guest-agent). Visibility only; never execution.
• Enforced mechanically: scripts/check-rules.sh rule #11 fails on any qm guest exec / vm-guest pattern in code. tests/remote.sh has no guest-exec mode.
• If a system is locked out (no SSH key, no guest-exec path): surface it to the user. Do NOT improvise an alternate back-channel. The user authorizes the unblock method (console login, credential, etc.).

Ticket-closing policy (NON-NEGOTIABLE)

NEVER close a Redmine ticket without explicit user permission. You may SUGGEST a close when the result is clearly scoped and delivered. If it's ambiguous whether the work is truly complete, don't suggest a close — leave that decision to the user. This applies to ALL tickets, no exceptions.

What this means in practice

• Discovered a typo in a config during approved work? Finish the approved work. Create a sub-ticket for the typo. Do not fix it inline.
• Think a system should also have localuser sudo configured? Do not add it. Propose it in a sub-ticket.
• Need to install a package the ticket didn't mention? Ask first.
• Found a security issue? Create a ticket immediately with full details. Do not remediate without approval.

This environment operates in ITAR/CMMC/TS/SCI space. Every action must be traceable to an approved ticket. There is no "I thought it would help."

Quick Start

You are an AI agent working on this project. Your first actions, in order:

SESSION-START GATE (NON-NEGOTIABLE): Steps 1-3 orient you. Step 4 is the check-for-understanding gate (rule summary). Steps 5-7 gather state. Step 8 is the scope-alignment gate. You MUST NOT begin any task work until the user (a) confirms your rule summary AND (b) names the ticket to work on. This runs every session, automatically — the user should never have to ask for it.

1. Set up the environment: bash scripts/setup-hooks.sh (installs git hooks — idempotent).
2. Read this file (AGENTS.md) — project policy and domain knowledge.
3. Read the latest questions file (questions-v*.md) — open questions awaiting human input. The version number increments each round (v1, v2, v3...).
4. Check for understanding — GATE. In your own words, summarize ALL the rules back to the user before doing any work: Agent Authority (ticket-governed, propose-never-implement), Access-Channel SSH-only policy, Remote access (remote.sh chokepoint mandatory, DNS names only — never IP literals), Questions policy (no harness question-tools), Documentation policy (Discourse is SoR), Redmine tracking, Git policy (always commit+push, shellcheck), Mandatory infra-change documentation (#298 audit log), SSH routing chokepoints (tests/remote.sh), the mechanically enforced check-rules.sh rules, and Credentials (Vault migration TODO). Then stop and wait for the user to confirm. This checkpoint guarantees every session starts aligned.
5. Check Redmine — docker run --rm --env-file ~/.creds/redmine.env git.knownelement.com/reachableceo/redmine-cli:latest list --assigned-to-me -p 55 for active work.
6. Check current state: git log --oneline -10.
7. Run rule audit: bash scripts/check-rules.sh --fast.
8. Scope-alignment — GATE. Present the session handoff's "What's Left" priority list (or the Redmine queue if no handoff exists). Do NOT scan the full ticket queue and pick work on your own — the user directs what gets worked on, always. Note any new or urgent items from the Redmine check, then ask which ticket to work on tonight. STOP and wait. Do not set .crush/active-ticket, do not read systems, do not run diagnostics, until the user names the target. This gate prevents the agent from burning context on work the user didn't ask for.

Enforcement Model

Git hooks (scripts/pre-commit, scripts/pre-push) enforce the rules defined in scripts/check-rules.sh. The rules engine checks: shellcheck (zero warnings including info-level), Docker image pinning (no :latest), container naming, required files, Discourse pointer headers, and more. Run bash scripts/check-rules.sh for a full audit or --fast for pre-commit speed. Bypass with --no-verify (emergencies only).

Task Tracking

• Redmine is the system of record for all work.
• NEVER close a ticket without explicit user permission. Suggest a close when clearly scoped/delivered; if ambiguous, don't suggest.
• Ticket-first enforcement (mechanically enforced). Before starting any work, set the active ticket: echo '#NNN' > .crush/active-ticket. The Crush hook (hooks/ticket-gate.sh) blocks modifying operations until this file exists. If no ticket exists, CREATE ONE FIRST via redmine-cli, then set it. Clear when done: > .crush/active-ticket. Parallel sessions: each session uses its OWN suffixed file (e.g. .crush/active-ticket-plant, .crush/active-ticket-core) so concurrent sessions never clobber each other's ticket; any non-empty .crush/active-ticket* satisfies the gate.
• WORKING.md is the only in-repo task tracker — a scratchpad for the current session. The pre-commit hook blocks commits while any task remains unchecked.
• Parallel sessions: when two sessions share this repo (e.g. physical plant vs core infra), the second session uses a gitignored scratch tracker such as .crush/WORKING-PLANT.md instead of WORKING.md, so the two trackers (and their pre-commit gates) never collide. Wipe at session close.
• Clear WORKING.md before responding to the user.

Rolling HUD (session-scoped, NOT persisted)

The rolling HUD is a live status display the agent maintains throughout the session to help the human follow along. It is NOT a system of record — Redmine, Discourse, and git are the durable systems. The HUD exists purely for the human's situational awareness during the session.

• Format: a compact block shown at the end of each significant response (after completing a step, hitting a blocker, or pivoting). Example:

  ┌─ SESSION HUD ────────────────────────────────────────
  │ Active: #343 (Monitoring coverage matrix)
  │ Done:   ✓ scope-alignment gate added to both AGENTS.md
  │         ✓ committed + pushed (6a2550b)
  │ Now:    drafting coverage matrix on Discourse
  │ Next:   → #341 TEMPer USB (deploy on pfv-tsys1)
  │         → #338 LibreNMS alerts (BLOCKED: pushover.env)
  ├─ USER ACTION ITEMS ──────────────────────────────────
  │ • Populate ~/.creds/pushover.env (unblocks #338, #428)
  │ • Populate ~/.creds/prometheus.env + grafana.env (#430)
  └──────────────────────────────────────────────────────
  

• Placement: may be written to .crush/hud.md on disk to keep context window smaller (re-read and update rather than hold in memory). Never committed to git. Wiped at session end.

• Sections:

  • Active: current ticket number + one-line description
  • Done: ✓ items completed this session (append as work progresses)
  • Now: what the agent is actively doing
  • Next: the queued items (per handoff priority or user direction)
  • User action items: things ONLY the user can do (populate creds, physical work, manual deploys) with the tickets they unblock

• When to show it: after each logical unit of work, at blockers, and when pivoting between tickets. Not every trivial response — use judgment.

• NOT a substitute for Redmine/Discourse/git. The HUD is ephemeral. When work completes, update the durable systems (ticket notes, Discourse wiki, commits). The HUD just tracks the live narrative for the human.

Working Style

• Stop over-thinking. Get to code and output faster. Explore with code; gather ground truth. Do not burn tokens reasoning about things a quick command answers.
• Prefer Unix utilities (awk, sed, grep, cut, tr) for file editing and text processing over harness edit tools when feasible — deterministic and exact where harness editors can be whitespace-fragile.
• Farm work out to deterministic tooling: linters, LSPs, formatters, test runners. If an LSP is wired up, use it; otherwise pull a Docker image and lint inside it.
• Use sub-agents as subcontractors: scoped spec in, distilled deliverable out. Never read 10+ files sequentially; batch into agent calls.
• Command timeouts (NON-NEGOTIABLE): Every command that touches a remote system MUST be wrapped with timeout. Hard limits: 30s for quick reads (status, ps, ls), 120s for standard operations, 300s for deployments/pulls. If a command hits the timeout, STOP and investigate root cause — never blindly retry. A hung command is a failed command. Detect failure fast, diagnose, fix, move on. Example: timeout 120 bash tests/remote.sh vm 'cmd'. This applies to ALL tools — bash, docker, CLIs, sub-agents.

Field Lessons (append at session close)

Hard-won operational gotchas. One line each; keep them short and load-bearing.

• CLI body scanner workaround: inline -b "body" arguments containing words like systemctl are blocked by the command scanner. Write long bodies to /tmp/note.md and pass -b "$(cat /tmp/note.md)" (redmine-cli + discourse-cli both).
• python3-venv trap: python3 -c 'import venv' succeeds even without the package; the resulting venv has no pip. Guard with dpkg -s python3-venv, verify test -x venv/bin/pip, delete pip-less venvs and recreate.
• temperusb API: there is no Temper class. Use TemperHandler().get_devices()[i].get_temperatures(sensors=[0,1]) (sensor 0 = internal, 1 = external probe).
• LibreNMS API: device add requires "version": "v2c" (with the v); discovery.php/poller.php must run as the librenms user; SNMP community fields are redacted in API responses by design.
• remote.sh modes: vm-copy/prox-copy destination must be a FULL file path, never a directory; vm-file/prox-file push and execute a script via bash -s (use for idempotent deploy scripts).
• Remote restart verification: after restarting services (especially in parallel across hosts), re-verify state several seconds later. Immediate checks race the restart and produce false failures.
• NUT: upsd.conf LISTEN edits only take effect after systemctl restart nut-server. udev rule changes need udevadm control --reload-rules && udevadm trigger.
• Tailscale app-connector sources (HA): HA polls originate from its LAN IP (192.168.3.12), not its Tailscale IP. Include the HA LAN IP in every allowlist (snmpd rocommunity source lists, upsd LISTEN binds).
• git mv fails on untracked files: plain mv first, git add after.
• SNMP extend with UTF-8: net-snmp CLI prints °C extend output as Hex-STRING, but pysnmp/HA receive the decoded string. Parse the decoded lm-sensors text lines with a regex value_template.
• pfv-bms (Home Assistant): no SSH by design; everything goes through the REST API (/api/, /api/config, /api/services) with the long-lived token in ~/.creds/homeassistant.env. Instance unit system is °F — set explicit unit_of_measurement: "°C" on custom sensors.
• HAOS core_ssh on 22222 is the sanctioned host shell: VM_IP=pfv-bms-lan.knel.net VM_USER=root VM_PORT=22222 bash tests/remote.sh vm '...' — use the LAN name (knel.net zone is Tailscale-IP based; TS inbound to pfv-bms is unreliable). Host OS has NO git/ssh/curl — run git against /mnt/data/supervisor/homeassistant via docker run homeassistant/amd64-addon-git_pull:<ver>.
• Git pull add-on (core_git_pull) wipes /config on first run if .git is absent (rm -rf + fresh clone). ALWAYS pre-seed /config as a git checkout (deploy key ~/.creds/pfv-bms-gitpull, repo KNEL/pfv-bms) before starting it.
• HA core boot death-spiral: supervisor watchdog SIGTERMs slow boots (docker logs: "finish process received signal 15" while supervisor logs "Still waiting for Core to start"). Fix from the 22222 shell: ha ha options --watchdog=false, start core, diagnose, re-enable once stable.
• Technitium on netinfra-02 (tsys-dns container) does not reload rsync-synced zone files — docker restart tsys-dns after record adds until #469 (native clustering) lands. netinfra-01 (the dns-cli target) is fine.
• Supervisor add-on options via API: docker cp a script into hassio_supervisor, read token from /data/cli.json access_token, POST /addons//options (the /apps/ spelling 404s).
• NEVER restart/kill BOTH members of a redundant pair simultaneously (founder ruling 2026-09-01 after the DNS outage): restart/redeploy ONE node, verify service health from an independent vantage, only then the second. Blue/green with a health gate between hops — always serial.
• **Pi-hole v6: dnsmasq_lines in pihole.toml are passed to FTL's embedded dnsmasq, which rejects some valid-dnsmasq options (e.g. no-negcache) with "bad option" and dnsmasq then never starts — the container looks "Up" but serves nothing on :53. Use first-class toml settings instead; never sed-edit pihole.toml without a single-node health-gated rollout.
• DNS outage recovery path: workstation resolv.conf dies with the LAN Pi-holes; use tailscale status peer IPs + the chokepoint env overrides (NETINFRA01_HOST/NETINFRA02_HOST, VM_IP) to reach hosts by Tailscale while names are unresolvable. Fix one node, verify, then the other.
• dhcpd + AppArmor: on the netinfra pair, dhcpd can only read configs under /etc/dhcp/** — staging a candidate config in /tmp fails dhcpd -t with "Permission denied" even as root. Stage as /etc/dhcp/*.candidate, test, then move into place.

Questions (NON-NEGOTIABLE)

NEVER use a harness "question"/"ask user" tool (structured prompts, modal forms, tabbed questions). Banned across every project, every harness. They are not portable, not version-controlled, and bypass the git record. All questions go in the current questions-v(N).md file — write the question; the human edits the answer inline in the same file. Version up the filename each time answers land (v1 → v2 → v3...): create questions-v2.md with resolved Q&A marked, new questions appended. This preserves the history of each Q&A round. Synthesize resolved Q&A into Discourse (decisions) and Redmine (work items). See BASELINE-PROMPT.md §10.

Documentation policy (IMPORTANT)

Discourse is the canonical source of truth for all knowledge documentation. Knowledge docs (architecture, runbooks, references, audits, policies) have been migrated to community.turnsys.com as wiki topics in the VP TechOps category.

All .md files in this repo (except AGENTS.md and LICENSE) are now pointers that link to their corresponding Discourse topic. Do not update documentation content in git — edit the Discourse wiki topic instead. Git edit history no longer serves as the documentation changelog; Discourse preserves wiki edit history automatically.

Code (scripts, configs, playbooks) still lives in git as the source of truth for executables. Only documentation moved to Discourse.

Redmine wikis stay disabled — Discourse is the sole documentation space (founder ruling, 2026-09-02). All Redmine projects have the wiki module off (verified 0/62, 2026-09-02); never enable it on new or existing projects.

Top-level files

All .md files now point to Discourse. The key pointers:

File	Points to	Discourse topic
README.md	Project overview	#296
STATUS.md	Ticket index + infra summary	#297
docs/docmap.md	Documentation index	#296

Work tracking stays in Redmine.

Repository Layout

k8s/                k3s cluster setup scripts (HA control plane over Tailscale) + docs/
proxmox/            Proxmox fleet docs (hardware audit, capacity, storage)
awx/                Ansible AWX deployment (k3s + AWX Operator)
HomeAssistant/      Redirect stub — HA tooling + reference packs moved to KNEL/pfv-bms (tools/, docs/reference-packs/) [#762]
tests/              Test suite + VM validation harness + remote.sh SSH chokepoint
scripts/            Framework: git hooks, rule engine (check-rules.sh), shared lib
docs/               Server-build docs, docmap index, and archive
archive/            Historical/superseded code (provisioning -> replaced by KNELIAC project;
                    KNELServerBuild -> live patterns now in KNEL/monitoring, remainder
                    preserved under archive/KNELServerBuild/ [#474])
vendor/             Vendored KNELShellFramework

• 2026-09-03 repo split (#769) — domain tooling moved out of this catch-all into dedicated KNEL repos (working dirs ~/projects/KNEL/<repo>; full git history retained here):

  body of work	repo	ticket
  SIEM/syslog (Wazuh)	KNEL/siem	#335
  DNS/NTP/DHCP + switches	KNEL/netinfra	#694
  Fleet CA	KNEL/ca	#697
  Monitoring/OAM	KNEL/monitoring	#343, #773
  Inventory/CMDB	KNEL/inventory	#705, #774
  Facilities (PDU/console/UPS/sensors)	KNEL/facilities	#458
  Performance/tuning	KNEL/perf	#709, #732, #737
  Backups (PBS)	KNEL/backups (new stub)	#315, #684
  Outbound mail (PMG)	KNEL/mailgateway (new stub)	#696
  Secrets mgmt	KNEL/secrets (new stub)	#770
  apt-satellite (deb version control)	KNEL/apt-satellite (new stub)	#771
  Security Onion (ops side)	KNEL/security-onion (new stub)	#772

  Unit tests moved with their code (netinfra, facilities). IaC consumers: KNEL/KNELIAC. Env map + snapshot lifecycle: Discourse t/331.

• Server provisioning moved to KNELIAC: The archive/provisioning/ tree is historical. Active server provisioning lives in the KNELIAC project at /home/reachableceo/projects/KNEL/KNELIAC.

• Non-bash files: Some files under archive/provisioning/Agents/ have .sh extension but are PHP (shebang #!/usr/bin/php). Skip in syntax checks.

• Remote access (NON-NEGOTIABLE): ALL SSH/SCP to ANY host MUST go through the chokepoint scripts — tests/remote.sh (Proxmox hosts + all VMs) or ~/projects/KNEL/netinfra/dns-cluster-setup/remote-dns.sh (DNS infra hosts: netinfra-01/02, netboot; lives in KNEL/netinfra [#769]). NEVER call ssh/scp directly — the harness blocks raw ssh and the command scanner rejects it. There are no exceptions.

• DNS names ONLY (NON-NEGOTIABLE): NEVER use IP address literals (neither LAN nor Tailscale IPs) in any command, script, or config. ALWAYS use DNS names. For Proxmox hosts: PROX_HOST=<dns-name>. For VMs: VM_IP=<dns-name>. For conman/SNMP/any tool: pass the DNS name. If a DNS name does not resolve, fix it in DNS (Technitium) or consult the system inventory — Discourse #307. Do NOT fall back to IP literals. This rule eliminates the per-session discovery tax of finding the right IP for each host.

• How to access a production VM:

  1. Look up the DNS name in the system inventory — Discourse #307.
  2. VM_IP=<dns-name> VM_USER=root bash tests/remote.sh vm '<command>'
  3. If the name does not resolve from the workstation, use PROX_HOST=<proxmox-node> and run qm guest cmd <vmid> network-get-interfaces (visibility only — NOT execution) to find the Tailscale DNS name, then access via that.

Git Policy

1. ALWAYS commit + push. Never wait. Ever. After each logical unit of work, immediately stage, commit (conventional format), and push to origin/main. Do NOT pause for review. Do NOT ask permission. Do NOT stop to "let the user read it first." The user reviews rendered markdown on Gitea AFTER it is pushed. This rule overrides any default conservative commit-and-hold behavior — if a default instruction says "never commit unless asked," that instruction is superseded here.
2. Atomic commits. Each commit coherent on its own.
3. Conventional format: feat(scope): desc, fix(scope): desc, docs: desc, refactor(scope): desc, test(scope): desc.
4. All shell scripts MUST pass shellcheck before commit. No exceptions. Run it via the wrapper:

   bash tests/shellcheck.sh                 # whole repo
   bash tests/shellcheck.sh ups/*.sh        # specific files
   

   This invokes koalaman/shellcheck:stable through Docker (no native binary needed). Fix every reported finding — including info-level — or add a targeted # shellcheck disable=SCxxxx # <reason> directive with a justification. A script that emits any diagnostic is a protocol violation. Non-bash scripts (PHP with .sh shebang #!/usr/bin/php, etc.) are exempt.

Cross-linking (NON-NEGOTIABLE)

Every artifact cross-references its related artifacts with CLICKABLE links. House rule (Charles, 2026-09-02, [#743]) — becomes hard requirement once GLPI/CMDB change control is online (ITIL/ITSM ramp):

• Redmine ticket comments link the commits they describe (full Gitea URL, e.g. https://git.knownelement.com/KNEL/PFVCluster/commit/<sha>).
• Commit bodies carry the full URL of the corresponding Redmine comment (https://projects.knownelement.com/issues/NNN#note-M) — the "meat" lives in the comment; the commit body points at it.
• Discourse posts link both the Redmine ticket and the Gitea repo/commit; Redmine tickets and repo docs link their Discourse topics.
• One click from ANY system reaches the other two. An artifact without its links is not done.
• When GLPI is online: change requests become the entry point for infra changes and get cross-referenced into Redmine/Discourse/Gitea the same way; CMDB items link their tickets and vice versa.

Mechanical enforcement (rule-engine check for the ticket-comment URL in commit bodies when [#NNN] is present) is tracked under [#441].

CI/CD (fleet standard, 2026-09-04)

Every gitea repo runs CI via Gitea Actions. Runner: ukrrs-pfv-gitea-runner (act_runner 0.2.13, compose at ~/projects/ukrrs/ci/gitea-runner/ on the pfv workstation) — label ultix ONLY; workflows must use runs-on: ultix or they queue forever. Repo duty: carry .gitea/workflows/ci.yml — lint for the repo's language (pinned container images, never :latest) + a secret-scan step, on dev and release pushes. Branch model: dev = WIP (CI must pass), release = production/deploy branch (CI is the final gate). Full standard: Discourse topic #333. Tracking: #784.

TDD & Linting

• Red/green TDD for all code. Mandatory (founder 2026-08-27). Interim relaxation applies ONLY until the Ansible/IaC transition (week of 9/1); from then on every playbook/script ships with its failing test first. scripts/test.sh is the local gate; tests/validation + tests/security run on sectestbed targets.
• Linters on all code, as early as possible. shellcheck zero-warning (including info-level) is already enforced pre-commit.

IaC Codification (shadow tracking) — NON-NEGOTIABLE

Every manual/direct change to a fleet system must be codified same-day as an AWX playbook item. The fleet converges to 100% IaC (founder mandate, 2026-08-27, #454).

• Made a manual change? Add a checklist item to #454 (or a child ticket) in the same session — what changed, where, exact commands, and any quirks (e.g. "needed udevadm trigger after NUT install").
• Work is NOT "done" until the manual change is codified or explicitly ticketed for codification.
• New manual fixes during incidents: fix first, codify immediately after.
• Playbooks live in this repo, tested through the sectestbed-* fleet.

DNS change discipline (NON-NEGOTIABLE)

Any change to Technitium records/zones (dns-cli or API), dhcpd, or the live DNS/DHCP configs MUST end with a same-session git SoR sync. Founder rule, 2026-09-02 — drift-check exists because git went stale; do not reopen that gap:

1. Make the change (serial + health-gated across the redundant pair).
2. cd ~/projects/KNEL/netinfra && bash dns/technitium/zone-snapshot.sh — refresh the DZ snapshots from the primary.
3. cd ~/projects/KNEL/netinfra && bash dns/drift-check.sh — must print ALL IN SYNC.
4. Commit + push the refreshed snapshots in the same session.

Redmine Tracking Policy

Redmine is the system of record for all work. Do not track status, checklists, or TODOs in repo files. Use Redmine tickets instead.

Tracker discipline: OAM/ops/feature/audit tickets use tracker Support (3) — NOT Bug. The redmine-cli create defaults to Bug; always correct the tracker after create (python escape hatch: tracker_id=3). [2026-08-27]

• URL: https://projects.knownelement.com
• Version: Potential to Kinetic Ready (due 2026-09-30)
• Project: Known Element Enterprises - Technology & Facility Services (id 55)

Rules

1. Every piece of work (feature, fix, deployment, config change) gets a Redmine ticket. If one doesn't exist, create it.
2. Reference tickets in docs and commits using [#NNN] notation. Example: [#367] Rebuilt k3s control plane after cnode wipe.
3. When work completes, update the ticket: set done ratio to 100%, add a note describing what was done and where the code lives. NEVER close a ticket without explicit user permission. You may SUGGEST a close when the result is clearly scoped and delivered. If it's ambiguous whether the work is truly complete, don't suggest a close — leave that decision to the user.
4. Operations Status lives on Discourse (topic #297) — update that wiki topic if the infrastructure summary needs refreshing. The STATUS.md file in git is now a pointer only.
5. Link code to tickets — ticket descriptions and notes should reference the relevant file paths in this repo (e.g., the ups tooling in KNEL/facilities).

CLI access (read + write)

Tickets are managed via the redmine-cli container, invoked directly with docker run (no wrapper script). Full command reference, patterns, and the subtask escape hatch live in the CLI's own AGENTS.md (~/projects/ukrrs/connectors/KNEL-AIMiddleware/tooling-cli/redmine/AGENTS.md) — read it for anything beyond the basics.

# Connection sanity check (run first in any session):
docker run --rm --env-file ~/.creds/redmine.env \
    git.knownelement.com/reachableceo/redmine-cli:latest whoami

# Your queue (project 55):
docker run --rm --env-file ~/.creds/redmine.env \
    git.knownelement.com/reachableceo/redmine-cli:latest list --assigned-to-me -p 55

# Show / create / update / close:
docker run --rm --env-file ~/.creds/redmine.env \
    git.knownelement.com/reachableceo/redmine-cli:latest show 367
docker run --rm --env-file ~/.creds/redmine.env \
    git.knownelement.com/reachableceo/redmine-cli:latest create -p 55 -s "Subject" -d "desc"
docker run --rm --env-file ~/.creds/redmine.env \
    git.knownelement.com/reachableceo/redmine-cli:latest update 367 -n "Done: committed in abc123" --done-ratio 100
docker run --rm --env-file ~/.creds/redmine.env \
    git.knownelement.com/reachableceo/redmine-cli:latest close 367

Key IDs: project 55 (technicaloperations), user 5 (reachableceo). Statuses: New(1), In Progress(2), Resolved(3,closed), Feedback(4), Closed(5), Rejected(6). New subtasks go to Feedback (4). Tracker 3 = Support. Gotcha: create has no --parent flag — to make a subtask, use the python-redmine escape hatch inside the container (see tooling-cli/redmine/AGENTS.md). Always show a ticket before updating it.

Credentials (REDMINE_URL/REDMINE_API_KEY) live in the centralized store at ~/.creds/redmine.env.

Documentation Workflow

Discourse is the source of truth for all knowledge docs. After any work session, an agent MUST:

1. Update Discourse wiki topics — if infrastructure facts changed (new VM, IP change, host retired), edit the relevant wiki topic at community.turnsys.com/c/vp-techops.
2. Update the Operations Status topic (#297) if tickets were opened or closed.
3. Grep for stale paths in code — grep -rn 'old/path' --include='*.sh' after any rename or restructure. Fix all references in the same commit.
4. If a new top-level directory was created, update:
   • AGENTS.md → "Repository Layout" code block
   • AGENTS.md → "Key Scripts" table (if it has an entrypoint script)
   • Create a new Discourse wiki topic for any documentation
5. Self-audit before commit. Code changes must be internally consistent. Documentation changes go to Discourse, not git.

CLI access (read + write)

Wiki topics are managed via the discourse-cli container, invoked directly with docker run (no wrapper script). Full command reference, patterns, and the raw-API escape hatch live in the CLI's own AGENTS.md (~/projects/ukrrs/connectors/KNEL-AIMiddleware/tooling-cli/discourse/AGENTS.md) — read it for anything beyond the basics.

# Connection sanity check (run first in any session):
docker run --rm --env-file ~/.creds/discourse.env \
    git.knownelement.com/reachableceo/discourse-cli:latest whoami

# List VP TechOps topics:
docker run --rm --env-file ~/.creds/discourse.env \
    git.knownelement.com/reachableceo/discourse-cli:latest ls -c vp-techops

# Show a topic / edit a wiki post (find post id via `show`):
docker run --rm --env-file ~/.creds/discourse.env \
    git.knownelement.com/reachableceo/discourse-cli:latest show 297
docker run --rm --env-file ~/.creds/discourse.env \
    git.knownelement.com/reachableceo/discourse-cli:latest update <post_id> -b "new markdown body"

VP TechOps = category 74. Key topics: #296 (project overview), #297 (ops status), #298 (audit log). The API user is trust-level 4 but not admin — admin-only ops (category creation, setting the wiki flag) will 403; surface those to the user rather than retrying. Gotcha: update/delete take a post id, not a post number. Never create a new topic for an update to existing knowledge — edit the wiki post in place.

Credentials (DISCOURSE_URL/DISCOURSE_API_KEY/DISCOURSE_API_USERNAME) live in the centralized store at ~/.creds/discourse.env.

Key Scripts

Script	Purpose
scripts/check-rules.sh	Rule audit engine (shellcheck, image pinning, Discourse pointers, required files)
scripts/setup-hooks.sh	Install git hooks (pre-commit, pre-push)
scripts/pve-snapshot.sh	PVE host snapshot/backup pass — hard-protects prod before risky windows (verified 2026-09)
scripts/audit-fleet.sh	Fleet-wide audit sweep (hosts, services, drift)
scripts/garden.sh	Repo/workspace gardening loop (strays, cruft, staleness)
scripts/kuma-*.py suite	Uptime Kuma management (add-ping, fleet-sync, inventory, maintenance, notifications, regroup)
tests/remote.sh	SSH chokepoint — all Proxmox host + sandbox VM access routes here
KNEL/netinfra: dns-cluster-setup/remote-dns.sh	SSH chokepoint for DNS infra hosts (netinfra-01/02, netboot)
KNEL/netinfra: dns/technitium/zone-snapshot.sh	Refresh git SoR Technitium zone snapshots from the primary — required after every DNS change
KNEL/netinfra: dns/drift-check.sh	Live-vs-git drift audit for DHCP/DNS/NTP/zones — must be green before any DNS-adjacent push
KNEL/inventory: seed/from-inventory.sh	Regenerate the CMDB seed CSV from the Discourse #307 inventory raw (re-run when #307 changes)
redmine-cli container	Redmine CLI (ticket read/write via docker run; see tooling-cli/redmine/)
discourse-cli container	Discourse CLI (wiki topic read/write via docker run; see tooling-cli/discourse/)
dns-cli container	Technitium DNS CLI (zones, list, add, delete, search, flush; see tooling-cli/dns/)
tests/vm-validation.sh	Deploy + validate on sandbox VM
tests/run-tests.sh	Test suite
KNEL/netinfra: dns-cluster-setup/setup.sh	DNS cluster replication
k8s/install-cp.sh	Bootstrap k3s HA control plane
KNEL/facilities: powerman/setup.sh	Configure Cyclades PDU via powerman
KNEL/facilities: console/setup.sh	Configure serial console access via ser2net + conman
KNEL/facilities: ups/setup.sh	Configure NUT (Network UPS Tools) for UPS monitoring
KNEL/perf: deploy-tuning.sh	Deploy perf tunings
KNEL/perf: scripts/	Read-only audit: probe-storage, probe-network, conman-console, snmp-switch-audit, probe-drift, audit-vm-disks, audit-guest-io, deploy-tuned-guests

Switch Console Access (conman)

Switch configs are pulled via serial console through a conman + ser2net stack on pfv-tsys4.

• conmand server: pfv-tsys4 (port 7890)
• ser2net: pfv-tsys4, TCP ports 2001-2006 (do NOT connect to ser2net directly — it conflicts with conman's persistent sessions; always use the conman client)
• Script: KNEL/perf: scripts/conman-console.py — drives console sessions read-only via PTY. No expect/tcl required.
• Command files: KNEL/netinfra: switches/ — .cmds files with switch-specific show commands
• Query available consoles: conman -d pfv-tsys4 -q

# Pull a switch config (example):
CONMAN_SERVER=pfv-tsys4:7890 python3 ~/projects/KNEL/perf/scripts/conman-console.py \
  --console pfv-r5-core-01 --cmds ~/projects/KNEL/netinfra/switches/pfv-r5-core-01.cmds

Console name	TCP port	Device
pfv-r5-core-01	2001	Dell PowerConnect 5448 (rack 5 core, mgmt+storage)
pfv-r3-tor-mgmt-01	2002	Dell PowerConnect 5324 (rack 3 mgmt TOR)
pfv-r3-tor-stor-01	2003	Dell PowerConnect 5324 (rack 3 storage TOR)
pfv-rrinfra-rtr	2004	Cisco router (rrinfra)
pfv-r2-tor-01	2005	Rack 2 TOR switch
pfv-r6-mgmt-01	2006	Rack 6 management switch

Mandatory: Document ALL Infrastructure Changes

This is non-negotiable. Every infrastructure change (VM config, disk cache, network setting, service config, storage migration) MUST be documented BEFORE moving to the next task step — not "later" or "at the end."

Required for EVERY infrastructure change:

1. Redmine ticket — create one if none exists. Reference as [#NNN].
2. Discourse audit log — reply to topic #298 with a dated entry (what changed, why, where).
3. Discourse relevant wiki topic — update the architecture/reference topic if the change affects documented infrastructure facts (storage #300, network #299, k8s #305, etc.).
4. Git commit — if code/config changed in the repo, commit + push immediately per the Git Policy above.

Tooling

• gh, docker, jq available on the workstation.
• No native shellcheck — use bash tests/shellcheck.sh (Docker wrapper koalaman/shellcheck:stable). ALL scripts must pass including info-level.
• For raw API calls not covered by the Redmine/Discourse CLIs, use python3 inside the CLI Docker containers (escape hatch pattern in tooling-cli/{discourse,redmine}/AGENTS.md).
• curl/wget/httpie may be blocked by some harnesses. Use the CLIs or the python-in-Docker escape hatch for HTTP writes.

Key Commands Quick Reference

# Tests + validation:
bash tests/run-tests.sh                        # test suite
bash tests/vm-validation.sh                    # VM validation
bash tests/shellcheck.sh                       # shellcheck whole repo
bash tests/shellcheck.sh path/to/*.sh          # shellcheck specific files

# Redmine + Discourse sanity checks (run first in any session):
docker run --rm --env-file ~/.creds/redmine.env \
    git.knownelement.com/reachableceo/redmine-cli:latest whoami
docker run --rm --env-file ~/.creds/discourse.env \
    git.knownelement.com/reachableceo/discourse-cli:latest whoami

# Proxmox host access (DNS names only — never IPs):
PROX_HOST=pfv-tsys5 bash tests/remote.sh prox 'qm list'
PROX_HOST=pfv-tsys5 bash tests/remote.sh prox 'pvesm status'

# Production VM access (DNS names only — never IPs):
VM_IP=tsys-librenms VM_USER=root bash tests/remote.sh vm 'systemctl status cron'

# DNS infra access:
bash ~/projects/KNEL/netinfra/dns-cluster-setup/remote-dns.sh netinfra01-root 'systemctl status docker'

# Switch console (read-only config pull — DNS names only):
CONMAN_SERVER=pfv-tsys4:7890 python3 ~/projects/KNEL/perf/scripts/conman-console.py \
  --console pfv-r5-core-01 --cmds ~/projects/KNEL/netinfra/switches/pfv-r5-core-01.cmds

Credential Management

• API keys currently in .env files under KNEL-AIMiddleware (gitignored).
• User goal: migrate all keys to Hashicorp Vault (vault.knownelement.com). No vault token present yet (~/.vault-token missing). Track as high-priority TODO.

Key Docs

→ All documentation lives on Discourse: community.turnsys.com/c/vp-techops

→ All .md files in this repo are pointers to Discourse topics.

→ Complete Linux System Inventory: Topic #307 — every Linux system (hosts + VMs + physical), with Tailscale IPs, DNS names, SSH access status, and tuned profiles. Reference this for monitoring coverage, access management, and hostname consistency.

Project Context

Solo-founder R&D Proxmox cluster in a private residence. Shoestring budget. Production lives on a Cloudron VPS in Reston VA. See the Operations Status topic for the ticket index and infrastructure summary. All work is tracked in Redmine (version: Potential to Kinetic Ready, due 2026-09-30).

• HA 2026.8 snmp platform traps: temperature/current-classed sensors with state_class: measurement + non-numeric render ('unknown') are dropped AT ADD-TIME (ValueError) when a host is asleep — entity vanishes entirely. Fix: no state_class (stats live in VM) + templates render EMPTY on no-match.
• Legacy snmp extend OID encoding: suffix = strlen + ASCII decimal per char ("garagepdu" = 9.103.97.114.97.103.101.112.100.117 — count carefully).
• Old APC PDUs (AOS ~2.7): SNMPv1 only; subtree WALKS die silently, small single-table walks + GETs work. Garage AP7830 current OID: .1.3.6.1.4.1.318.1.1.12.2.3.1.1.2.1 (tenths of amps). Relay pattern via tsys-librenms (the allowed poller vantage) for v1-only/ACL-locked gear.
• Debian snmpd on monitoring VMs: snmpd may fail to bind 127.0.0.1:161 (exit 1, "Error opening specified endpoint") when something squats localhost — bind only the LAN/TS addresses in agentaddress.
• iDRAC6/7 SNMP: agents exist but hw MIB tables unpopulated on old fw; IPMI SDR via iDRAC SSH (racadm, one command per exec, legacy syntax on iDRAC6: racadm config -g cfgOobSnmp -o ...) is the data path.
• jq in one-liner state checks: use .state — not .s — a wrong key silently returns null and looks like an entity outage (cost an hour tonight).
• HA config flows are REST (POST /api/config/config_entries/flow), not websocket; websocket only monitors. ha-ws-call.py + REST flow are the two levers for agent-driven HA setup.
• mvdan/sh source a b silently ignores file b — source env files one per line.
• HA todo lists (founder HUD): items are hidden from /api/states — fetch via POST /api/services/todo/get_items?return_response=true. Renaming works ONLY with a minimal todo/update_item payload (entity_id, item=, rename); passing status/due back earns a 400. HUD summaries carry [#NNN] Redmine prefixes — sync completed items back to Redmine when the founder asks.
• HA package YAML: NEVER append a second top-level sensor: key — YAML duplicate keys are last-one-wins and silently shadow the whole first block at config load (#344 outage 2026-09-02: all 22 fleet temps down after the restart that pulled the file). Merge new sensors into the existing block.
• HA 2026.8 snmp add-time crash, general form: ANY unit-bearing sensor with a non-numeric native_value at ADD (host down renders '' / literal string "unknown"; value_template can never yield None — even {{ none }} is stringified to "None") = ValueError, entity vanishes till manual reload. check_config does NOT catch it. Durable pattern: unit-less raw snmp transport + template: layer carrying unit/device_class under the ORIGINAL unique_ids (KNEL/pfv-bms packages/plant_snmp.yaml header).
• HA entity_id comes from the NAME slug, not unique_id — if a sensor's name and unique_id diverge, cross-references must use the name slug (garage PDU amps lesson: semantic read a nonexistent ..._amps_raw).
• HA entity_id collisions spawn _2 entities; registry remove no-ops for LIVE entities (they re-register on their next write). Clean _2 by jq-editing .storage/core.entity_registry with core STOPPED, backup first. Orphaned (dead) entries remove fine via websocket (ha-ws-call.py — pass commands as JSON files; shell quoting eats argv).
• git_pull add-on: ha addons start is a no-op while it's running — use restart to force an immediate pull. It auto-restarts core for files in its restart list; otherwise pull ≠ reload (config lands on disk, running core keeps the old one until a platform reload/restart).
• Restore drills / heavy I/O NEVER on hosts running prod VMs (2026-09-03 incident, #684): qmrestore of a 32G disk to scratch VMID 5599 on tsys1 starved VM 100 (pfv-bms) I/O ~40 min, degraded prod HA. Drills run dev-lane only (tsys5); prod restores are server-throttled via PBS traffic-control.
• pkill -f <pat> over the ssh chokepoint self-matches your own remote command line and kills the session (exit 255, no output). Use the bracket trick: pkill -f "[q]mrestore".
• pvesh backup jobs take -vmid (comma list), not -vmids; aborted qmrestore leaves a create-locked stub VM — qm unlock <id> then destroy.
• PBS 4.x gotchas: prune retention lives in prune-job (datastore keep-* opts are gone); API tokens need their OWN acl entry — user-only ACL yields "Cannot find datastore" from PVE.
• PVE bridge port + hotplug NIC (#780): a re-registered NIC (USB move changes its enumeration path) comes back UP but is NOT re-enslaved to its bridge — ifup <nic> is not enough; run ifup <bridge> and confirm bridge link shows state forwarding. NFS hard mounts mostly self-heal, but nconnect=4 mounts can stay wedged on stale channels after a flap: force-remount (umount -f -l, recreate the mountpoint dir if the lazy umount removed it, fresh mount). pvesm status "active" can be stale — verify with a real read, not the status line.
• PMG pmgproxy restart hang (#696): systemctl restart pmgproxy can hang in "deactivating (stop)" on old workers ignoring TERM. If stuck >90s: systemctl kill -s KILL pmgproxy && systemctl start pmgproxy — API/UI only, mail flow unaffected. Postfix restarts don't hang.