Review findings on #13, round 3. A container pattern is ^<project>- and
then only letters, digits and _.-, a tmp pattern <project>- and then only
letters, digits and _.-*?: no quantifier makes the hyphen optional, no
broken regex is skipped in silence. MAX_AGE_HOURS is 4 to 999 without a
leading zero, so the seconds cannot overflow. A project named in two
files is refused in both. A dry run says it removed nothing.
Review findings on #13. A tmp entry is as old as its ctime: cp -a,
rsync -a and tar -x give a fresh copy an old mtime. A container counts
from the later of its creation and its last start. Project names have no
hyphen, so that "<project>-" is a prefix no other project shares, and a
container pattern has no alternation. The files are read, not sourced.
Failures are collected and make the run exit 1, each project in a
subshell of its own with errexit on.
Review of #13: a name pattern proves no owner on a shared runner, and age
alone did not keep running jobs safe without a measured job limit.
- A project file names PROJECT, CONTAINER_NAME_PATTERN and TMP_PATTERN.
Containers are found by the label org.luki-net.project=$PROJECT, which
the project sets itself, and then matched by name; both patterns must
begin with the project's own name, so ^.*, ?* or another project's
prefix are refused. Each file is read in a subshell of its own.
- Minimum age 4 h: act_runner v0.2.13 ends a job after 3 h (runner.timeout;
no config file on runner-01..03, LoadDefault sets 3 h), and each runner
runs one job at a time. MAX_AGE_HOURS cannot go to 3 or below.
- DRY_RUN: empty or 0 removes, any other value only lists.
- A project whose cleanup fails does not stop the others; the run's exit
code says so.
A cancelled job never reaches its own teardown step, so its test database
container and its working copies stay on the runner. Measured on runner-01
on 07.10.2026: 11 containers, 209 volumes (19 GB) and 4 tree copies of
1.2 GB each from l.kirchner/patchmgr alone (its D64).
install/runner-cleanup.sh removes only what a project registers by name
pattern under /etc/ci-runner-cleanup.d/, and only past MAX_AGE_HOURS
(default 24, at least 2): containers with their anonymous volumes
(rm -f -v), and entries in act-runner's private /tmp. No docker prune --
this runner works for every repository of the instance. Unanchored
container patterns and too wide tmp patterns are skipped; DRY_RUN=1 lists.
runner-install.sh installs it as /usr/local/sbin/ci-runner-cleanup with a
daily timer and patchmgr's pattern file. Only the timer is enabled; the
first run comes with it. tests/test_runner_cleanup.sh runs it against a
stand-in docker in CI.
actions/cache -- and actions/setup-go, which builds on it -- packs its archive
with zstdmt when zstd is present and falls back to single-threaded gzip when
it is not. The archive name says which happened: cache.tzst against cache.tgz.
Measured in l.kirchner/patchmgr, CI run 1957. The cache is GOMODCACHE plus
GOCACHE and runs to 2-5 GB. runner-gpu has zstd and writes cache.tzst;
runner-01, -02 and -03 write cache.tgz, and go test -race (stable) spent 605
seconds packing it for a job with 41 seconds of work. Across all eight jobs of
that run, 2312 of 3146 seconds went into this step.
This does not settle whether the cache is wanted at all -- in host mode with a
persistent home both directories survive between jobs anyway, and patchmgr is
switching it off for that reason. But as long as any repository on the
instance uses it, it should not be compressed on one core.
zstd is a package for the post step, not for jobs, so it sits with the others
rather than in a workflow: this runner installs nothing at job time, by
design.
The script that sets the SSH, PAM, pwquality and auditd baseline on twelve
containers existed only as a root-owned copy on the machines it hardens.
Bring it into the repo so a change reaches one place instead of twelve.
Two substantive changes over the copy that shipped on 2026-07-24:
- Deny forwarding per option instead of via DisableForwarding. Same effect,
but DisableForwarding overrides every other forwarding option and is
invisible in sshd -T, which makes a rejected port-forward read as a
configuration that should work.
- Quote the command substitution in MODDIR (SC2046).
The header and README now record the rollout command, the containers left
unhardened as break-glass foundation, and why host-specific exceptions must
live in a drop-in that sorts after 99-cis-hardening.conf.
Both from the cross-review, both fair.
"Go setzt CGO_ENABLED=0" is right about the effect and wrong about the
mechanism: Go does not set the variable, cgo simply stays off when no C
compiler is found, and go env then reports 0. Reworded.
And protoc on an instance-wide runner ties every repository to the
distribution's version -- 3.21.x on Debian 12. Unlike make and gcc that is a
code generator, so a distro upgrade changes generated code for all users at
once. The comment says so now, and says where a project that needs its own
version should pin it instead of raising it here for everybody.
The runner is host-mode, so there is no image bringing tools along: what is
not on this LXC, no job has. Measured on l.kirchner/patchmgr, a Go project,
where all six CI jobs were assigned and every one of them died in the first
seconds:
make all make: command not found
go test -race go: -race requires cgo; enable cgo by setting CGO_ENABLED=1
make proto sudo: command not found
Four packages, each for a reason:
make the gate commands are make targets
gcc Go turns CGO_ENABLED off when it finds no C compiler,
and the race detector cannot be built without cgo
protobuf-compiler protoc itself
libprotobuf-dev the well-known .proto includes under
/usr/include/google/protobuf; without them protoc fails
even though the binary is there
sudo stays absent on purpose. A workflow must not be able to install anything
on this runner -- what is needed is declared here, in the script, and not in
somebody's pipeline. That also keeps the security note at the top of this file
honest: the LXC owns nothing, and it gains nothing at a workflow's request.
Go is not in the list. Projects fetch it through actions/setup-go, because CI
matrices run more than one version.
Applied to the running LXC (301 on pve-gamer) while the runner was idle, then
verified: make 4.3, gcc 12.2.0, libprotoc 3.21.12, 11 .proto includes present.
The service PATH already contains /usr/bin, so no restart was needed.
- assert encoding BEFORE role/password mutation on re-run, so an old
SQL_ASCII DB aborts with no side effects (codex finding 1)
- ensure_utf8_locale_active honours its locale argument consistently in
match, locale.gen line and export (codex finding 2)
- assert_db_encoding_utf8 uses argv-clean runuser psql with :'db' literal
binding instead of nested su -c shell; docs keep su - postgres -c
(codex finding 3)
A PostgreSQL cluster/database freezes its encoding at initdb / CREATE
DATABASE time; a C (non-UTF-8) locale yields a SQL_ASCII cluster, which
makes psycopg3 return bytes and crashes SQLAlchemy. Harden the installer
and add a reusable pattern for future DB installers:
- ensure_utf8_locale_active: generate AND activate en_US.UTF-8 for the
install process before the server package runs initdb; abort if the
locale is not actually available
- create the database explicitly with TEMPLATE template0 ENCODING 'UTF8'
LC_COLLATE/LC_CTYPE 'en_US.UTF-8' instead of inheriting the cluster
default
- assert_db_encoding_utf8: post-install guard, abort with an actionable
message if pg_encoding_to_char is not UTF8 (catches old SQL_ASCII DBs
on re-run too)
- credentials/README docs use su - postgres -c (minimal LXCs have no sudo)
- merge the two contributing sections into one (PR + cross-review rule,
both real incidents, CI enforcement in present tense - the suite is
live on the homelab runner since K-114/PR #5)
- script catalog: runner is in production (PR #6 merged)
- usage: document input validation behavior (re-prompt on junk bytes,
env values abort when malformed)
- pattern: SSH root login prompt (sshd drop-in) and the locale fix in
setup_base_apt
- repo layout: tests/ added
LXC templates ship without a configured locale, so every apt/perl run
warned 'Setting locale failed'. setup_base_apt now exports C.UTF-8 for
the install run itself, installs the locales package, generates
en_US.UTF-8 and sets it as the system default via update-locale.
prompt_lxc_config asks 'SSH-Root-Login erlauben? [Y/n]' (env-presettable
via SSH_ROOT_LOGIN, validated, normalized to yes|no). The bootstrap passes
the value into the container; configure_ssh_root_login writes
/etc/ssh/sshd_config.d/zz-root-login.conf (yes -> PermitRootLogin yes,
no -> prohibit-password) and reloads sshd.
- CTID prompt re-prompts on invalid interactive input (was: abort)
- env-provided NAMESERVER is validated when non-empty ('' stays inherit)
- prompt_validated handles EOF (no infinite loop, clean abort under -e)
- 10# base forcing in vlan/cidr/ipv4 arithmetic (leading zeros are not
octal errors); is_clean_ascii rejects embedded newline/tab explicitly
(command substitution strips trailing newlines); is_ipv4_list checks
the whole string before word splitting
- nameref guard against reserved variable names in prompt_validated/
require_valid; source-check pattern documented as the repo contract
- 8 new test cases (41 total)
- validation helpers: sanitize_input trims CR/edge whitespace only;
embedded control/non-ASCII bytes FAIL validation and re-prompt with a
hint (2026-06-11 incident: invisible byte in a pasted VLAN tag broke
pct create mid-run) - never silently stripped
- prompt_lxc_config: every prompt validated (uint for CTID/disk/cores/
RAM, VLAN 1-4094, hostname/token formats, IP/CIDR/gateway, DNS list);
env-provided values are sanitized + validated too (abort, no re-prompt
loop in non-interactive use); helpers reusable for app prompts
- tests/test_validation.sh: 34 cases incl. the 2<0x80>0 repro, re-prompt
simulation, BASH_REMATCH clobbering regression (is_cidr), env dry-run
of prompt_lxc_config without PVE/TTY
- tests/check_ct_source.sh: every ct/*.sh must source build.func (bug
shipped twice); negative proof via prepared fixture in the test suite
- .gitea/workflows/ci.yml: bash -n over all scripts, source-check,
validation tests, shellcheck if present (documented skip otherwise)
- README: contributions via PR with cross-review (binding)
- source build.func (script was unrunnable without it)
- validate DB_NAME/DB_USER/DB_PORT/NEXUS_APP_IP before SQL/pg_hba use
- rotate password when role exists but credentials file is missing
webapp-pattern ct/install pair: PGDG repo, database nexus with
least-privilege owner role, pg_hba allowlist restricted to the nexus
app LXC (explicit reject for everything else), pgvector created by the
installer, credentials/DSN summary in /root/nexus-db.credentials.
Idempotent re-runs keep role/db and do not rotate the password.
A reboot before the first deploy would leave the enabled unit in failed
state; the condition keeps it inert until start.sh exists (same guard as
nexus-worker.service).
- new install/nexus-runtime.sh (idempotent, re-runnable on an existing
LXC): uv for the nexus user (manages Python 3.12), tesseract deu+eng,
nexus-worker.service unit (ConditionPathExists guards the skeleton
phase), sudoers extended to cover the worker service
- nexus-install.sh: RUNTIME section now invokes nexus-runtime.sh at the
end of the install (after base sudoers/units, which it extends)
Adds apply_network_profile(), which looks up DNS servers for the entered
VLAN tag in lib/networks.conf and sets --nameserver accordingly — even
when IP is DHCP. Precedence: explicit env NAMESERVER > profile > static
prompt > inherit. Comma-separated DNS is normalised to spaces for pct.
l.kirchner
claude-bot