Round 2 review (claude-bot, codex-bot — both raised this, independently).
The `return` added in round 1 aborted all of reconcile_pr, not just the
label edit. Everything below it is independent of the state:* taxonomy:
`merge-next` clearing and the stale sweep both stopped running. So a
cold-start repo left `merge-next` claiming "merge this one next" on a PR
the board had moved to the agent — the original false-invitation bug,
reintroduced inside the very fix meant to survive a cold start.
It was also a regression against main, not just a missed improvement:
the old code failed the `gh issue edit`, logged, and fell through to both
blocks. Round 1 turned a per-edit failure into a per-PR abort.
Now a `skip_edit` flag skips only the edit and control reaches the rest.
Also from review: drop the dead `"$desired"` term from the filter loop
(it was appended and then unconditionally continued past), and turn
`[ -n "$missing" ] && log` into a proper `elif` rather than an
&&-as-statement under `set -e`.
Adds the first four fixtures that exercise reconcile_pr itself, stubbing
run/gh to probe a cold-start repo against a bootstrapped one. Everything
before this tested pure functions, which is exactly why a per-PR return
got through: nothing could see it.
Fixtures 68 -> 72.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Round 1 review (claude-bot, codex-bot — both raised 1 and 2).
1. `gh issue edit --add-label` rejects the WHOLE call on one unknown
label name, and the blocker:* labels are created only by the
dispatch-only bootstrap. So the first sweep after this lands would
have converged NOTHING on exactly the PRs this change exists to heal,
surfacing only as a WARNING in a cron log. Batching state and blockers
into one edit for anti-flicker is what widened that blast radius.
Every label about to be ADDED is now filtered against the repo's real
label set, read once per sweep; removals need no filter because they
are built from has_label. An unreadable label set does not filter, so
a failed read cannot silently strip the board.
2. blocker:unrequested fired only on MISSING, so a round whose approvals
all staled behind a push — with nothing re-requested — carried no
blocker at all, though the agent owes exactly the same ask. Now
MISSING or STALE: both mean this head has no verdict from that
reviewer.
3. LABELS.md: restore the substantive "Leaves when" text for
state:addressing, and widen the blocker:unrequested row to name both
shapes now that (2) changes what the label means.
Fixtures 66 -> 68.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
When `gh pr view` failed, the fallback left the `statusCheckRollup` key
absent, and `(.statusCheckRollup // [])` collapsed that into the same
NONE as a PR with genuinely no checks. NONE blocks nothing, so an API
hiccup presented the PR as mergeable by a human — an unknown certified
as green, the shape #87 exists to stop, in the one place it never looked.
checks_state now returns UNREADABLE for an absent key versus NONE for a
present-but-empty array, and the sweep leaves an UNREADABLE PR alone
rather than recomputing on facts it did not read. Not a blocker on
purpose: blocking would flap the board on one bad call.
Fixtures 64 -> 66.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The CI wiring is independently valuable from the two-axis refactor and
was missing its own entry: rig's label state machine gates every PR and
its fixtures had never executed in CI, including for #88 which merged
today reporting 51 passing fixtures.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Retires `state:needs-rebase` in favour of two independent axes: `state:*`
(whose ball, exactly one) and `blocker:*` (what is in the way, additive).
One rule joins them: `state:needs-human` requires zero blockers.
The single-label design projected independent facts — mergeability, check
status, review round — onto one totally-ordered value, so one always won
and the rest vanished. Every precedence bug this machine has had lived on
that ordering. Blockers are a set, so there is no precedence between them
to get wrong.
`state:bots-reviewing` tightens to mean strictly "a request is live"; a
ready PR nobody was asked to review is now `state:addressing` +
`blocker:unrequested`. The reconciler strips the retired
`state:needs-rebase` on sight via a RETIRED array.
Also wires test/labels-reconcile.sh into CI, where it had never run.
Fixtures 51 -> 64.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The dating expression took the newest stamp each run carries. That reads
`completedAt` for a finished run and `startedAt` for a live one, so the
comparison comes down to "when this one ended" against "when that one
began" — which is not an ordering on runs at all.
A run cancelled by the concurrency group does not stop the instant its
replacement starts: the runner has to receive the signal and wind down. So
`predecessor.completedAt > successor.startedAt` is the ordinary case, not a
corner. On the box#137 tip that motivated the supersede rule the window was
13s wide — the superseding run started 15:19:38, the run it cancelled did
not finish until 15:19:51 — and for that whole window the dying predecessor
out-dated its own live replacement, so `last` discarded the replacement and
judged the corpse.
Both round-3 failure modes came back inside that window, narrowed rather
than closed: a draining CANCELLED predecessor reported FAILURE and sent the
agent to fix nothing, and a draining SUCCESS predecessor reported SUCCESS —
mergeable, all bots approve, state:needs-human — over a tree whose merge
button branch protection had already disabled. #136 again, one field over.
Dated by `first` of the preference-ordered stamps rather than `max` of
them: start time if the run recorded one, falling back only if it did not.
The sentinel filtering is unchanged, and finished runs still date by
completion when that is all they carry, so the supersede rule keeps the
case it exists for.
Found independently by claude-bot-andresmgsl and codex-bot-andresmgsl.
No existing fixture could express it — `run_()` carries no startedAt, so
every supersede fixture spaced the predecessor's completion safely before
the successor's start, the same blind spot as round 3 one field over. New
`drained_()` helper pins both directions; fixtures 48 -> 51 (with the
reverse-direction in-flight fixture ported from cast#128).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The supersede collapse added in the previous commit dated each run by
`.completedAt // .startedAt // .createdAt`. A run still in flight has no
completion, but `gh` does not omit the field: its Go struct marshals the
zero time as the string "0001-01-01T00:00:00Z", and jq's `//` falls
through null and false only. The sentinel was therefore taken as the sort
key, and it sorts before every real timestamp — so the LIVE re-run became
the oldest entry in its context, `last` discarded it, and the run it
superseded was judged instead.
That restored #136 through the fix for it: a green context with a
replacement mid-flight reported SUCCESS, so a PR read mergeable, green,
all bots approve — state:needs-human — while branch protection had the
merge button disabled. It also narrowed rather than removed the flap the
supersede rule exists to prevent: between "run A cancelled by the
concurrency group" and "run B finishes", the PR reported FAILURE and the
agent was sent to fix something that was not broken.
Runs are now dated by the newest timestamp they actually carry, with both
spellings of absent discarded (null, and the zero sentinel). Entries that
carry no usable timestamp sort LAST rather than first: something we cannot
date is most likely the thing just created, and treating it as newest
keeps an undateable in-flight run from being discarded in favour of a
stale success. Every ambiguity resolves toward "not settled".
Found independently by claude-bot-andresmgsl and codex-bot-andresmgsl.
The fixtures could not have caught it: `run_()` always emits a real
completedAt, so every supersede fixture was a race between two finished
runs, and the bug lived in the one shape the helper could not express.
New `inflight_()` helper covers it; fixtures 44 -> 48.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Round 2 review found two ways the "a human could merge this right now"
invariant still leaked, both of which let state:needs-human land on a PR
the button would refuse.
The check-rollup classifier enumerated the outcomes that block and
defaulted the rest to SUCCESS, so ERROR, CANCELLED and STALE fell through
into green. Inverted to an allow-list of the outcomes that DON'T block
(SUCCESS, NEUTRAL, SKIPPED, plus the pending set); everything else,
including an outcome neither enum has today, blocks. The rollup mixes
CheckRun.conclusion with StatusContext.state and an outcome the list
forgets is one we cannot certify as mergeable — a false FAILURE parks the
PR on the agent, a false SUCCESS invites a bad merge. The classifier also
moved out of main() into checks_state(), which is why no fixture caught
this: it was inline in the fetch loop and the jq itself was untestable.
Once CANCELLED blocks, superseded runs must be dropped first — a re-run
does not evict the run it replaced, and judging every entry would strand
every re-run PR in needs-rebase. Each context now collapses to its newest
entry, keyed on workflow + job name because a bare job name is only
unique within its workflow.
decide_state() returned from inside the bot loop on the first MISSING, so
a STALE belonging to a later bot in BOTS was never read: a round that was
both unfinished and staled came out needs-human with nothing bound to the
head. The whole round is now collected before any precedence is applied,
STALE ahead of MISSING. The MISSING-yields-to-an-explicit-human-request
rule is untouched.
Fixtures 29 -> 44, pinning the whole check-outcome enum, the supersede
rule in both orders, and the mixed round at both ends of BOTS.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Ported from heavy-duty/box#137 (heavy-duty/box#136) so the three repos'
reconcilers stay byte-identical. The state machine here was byte-identical to
box's before this change, and remains so after -- only the scope:* taxonomy
differs, correctly.
decide_state() derived state from three inputs -- draft flag, requested
reviewers, submitted reviews -- and read NOTHING about mergeability or checks.
With the `if requested "$HUMAN"` short-circuit at the top of its precedence,
the label was sticky: once the maintainer was requested, a PR read
state:needs-human through conflicts, through red CI, through a force-push that
staled every approval.
This repo paid for it directly. During the ten-PR batch merged today, every
merge re-conflicted the PRs below it through CHANGELOG.md, and each kept its
state:needs-human label throughout -- inviting merges that could not happen.
It was caught only by opening them one at a time, which is the work the label
exists to save.
The rule the label now keeps: state:needs-human means a human could merge this
RIGHT NOW, so anything making that false outranks the request that put it
there.
CONFLICTING or failing checks -> state:needs-rebase (new; the agent's to fix)
approvals staled by a push -> state:addressing (nobody reviewed this tree)
An UNFINISHED round still yields to an explicit human request -- MISSING
(nobody has reviewed yet) is a different fact from STALE (everyone reviewed
something else). UNKNOWN mergeability is NOT treated as unmergeable: GitHub
reports it for about a minute after every merge, and flapping every open PR
through needs-rebase on each merge would be worse than the bug. A failed read
degrades to the same "do not know" value.
Also adds merge-next: queue order is intent, so the reconciler never sets it,
only CLEARS it once the PR stops being mergeable-by-a-human.
Fixtures 19 -> 29, including that UNKNOWN does not trigger needs-rebase and a
draft outranks a conflict. No live dry-run evidence here -- this repo has no
open PRs right now -- so the fixtures and box's live dry-run are the proof.
Closes#87
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
A rig-managed machine recorded nothing about its own provenance. The entire
durable output of a bootstrap run was one line in /etc/rig/role, and that line
says what the box IS, never what built it. VERSION was read in exactly one
place (bin/rig:9, for --version) and reports the currently INSTALLED tree, not
the one that ran; there was no timestamp anywhere in the codebase.
bootstrap now stamps a second file beside the marker: schema=1, a birth pair
(bootstrapped_by/_at, pinned forever) and a latest pair (converged_by/_at).
key=value, one per line, 0644 — the one file that must stay readable on the
most broken machine in the fleet, where there is no YAML parser and no jq.
`rig manifest [<key>]` reads it back.
Only DECIDED facts go in, which is what keeps bootstrap.sh:3's convergence
contract intact: bootstrapped_* is first-write-wins, and converged_* updates
only when the version actually differs — it is the time the converging version
last changed, not the time of the last run. The renderer is pure, so a re-run
by the same rig is byte-identical no matter where the clock is, and the
cmp-guard stays silent. OBSERVED facts (cores, RAM, disk, kernel) stay out:
they go stale on their own and belong to `rig platform` (#64).
/etc/rig/role is untouched.
Closes#61
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The previous commit's test and README described a freshly bootstrapped
machine as lacking converged_* and rendering 'not recorded'. That pins the
wrong contract: #61 states plainly that "on a fresh machine both pairs are
written with equal values", and rule 2 only suppresses converged_* churn on
a later same-version re-run.
So two identical lines mean bootstrapped-and-never-re-converged, and a
manifest missing the pair is partial or hand-edited rather than fresh --
no writer produces it. The graceful degradation stays for that case, still
without backfilling from birth, which would invent a convergence that never
happened.
The reader itself was already correct; this fixes the fixture and the prose
that claimed otherwise.
Found in review of #74.
Refs #64
The reader asked for `version` and `bootstrapped`. #61 specifies `schema`,
`bootstrapped_by`/`bootstrapped_at` and `converged_by`/`converged_at` — so
no writer would ever have produced the keys being read, and the day #61
landed this command would have rendered 'unknown' with the timestamp
omitted, forever, with nothing to say why.
Keyed to #61's spelling, with fixtures carrying that schema verbatim so
the contract is pinned rather than assumed. Birth and latest are reported
separately and neither is inferred from the other: under #61 rule 2
converged_* is written only when the version differs, so its absence is a
legitimate state on a freshly bootstrapped box, printed 'not recorded'
rather than backfilled from birth. A manifest whose schema this rig does
not know is named as such instead of being half-read in silence.
Found in review of #74.
Refs #64
rig read no hardware at all. The single exception was `uname -m` in
runner-install.sh, used to pick a runner tarball and then discarded — so
"is this the 32GB one, or the M900?" was a question you answered by
logging in and running free -h, nproc, df -h and uname -r by hand, four
commands deep, on a machine you were already unsure about.
`rig platform` prints hostname, OS, kernel, CPU, memory, disk and
virtualization, then a provenance block: which rig, when, and the role
marker's traits.
It COMPUTES rather than stores, and that is the design rather than an
implementation detail. Specs change without rig doing anything — RAM
added, root disk resized, the unattended-upgrades bootstrap itself
enables patching the kernel — so a stored spec is stale the moment the
machine changes, and refreshing one on every run would collide with
bootstrap's "safe to re-run; a second run changes nothing" contract.
Nothing is written, so nothing can go stale.
The corollary is deliberate: reading only /proc, uname, /etc/os-release,
df and systemd-detect-virt means no root, no network, and it runs on a
pristine Debian box rig has never bootstrapped — useful for deciding
what to converge a machine into, not only for auditing it afterwards.
That also makes it the rare rig command the harness can RUN for real
rather than grep: the tests assert the answer describes the actual test
machine (kernel and hostname compared against independently computed
values), and assert it writes nothing.
Both known traps are handled explicitly. /etc/os-release is sourced in a
SUBSHELL — it defines VERSION, NAME and ID and would otherwise clobber
same-named script variables, the form every other site in this tree uses
and test/cli.sh already greps for. systemd-detect-virt exits non-zero on
bare metal while printing 'none', a normal answer that set -e would
otherwise turn into a failed run, so it is wrapped in `|| true`.
Provenance is read, never written, and degrades per file.
/etc/rig/manifest is #61 and does not exist yet, so that line reads
'not bootstrapped' on every machine today; the command ships complete
without it and neither blocks the other.
Named `platform` and not `status`: `users status` and `runner status`
cross-check recorded against live state and print DRIFT, and a command
that records nothing cannot drift, so calling it status would borrow a
promise it structurally cannot make. It also leaves `rig status` free
for the machine-wide roll-up it will eventually want to be.
Refs #64
A users file naming zero users is a valid instruction to revoke every
operator on the box, and it is indistinguishable from the file a stray '>'
produces. The per-user warnings apply already emitted arrive after the
decision and scale wrong: twenty operators is twenty lines of scrollback,
so the signal was loudest exactly where it read as noise.
The /etc/rig/users ledger draws the line apply needs. An empty file against
an empty ledger is an unambiguous no-op; against a populated one it closes
every named door. Only the second now stops, states how many operators are
at risk, and requires explicit consent: --yes, RIG_YES=1 (the
installer-family variable bin/rig's uninstall_confirm already reads), or a
y on a TTY. Without a terminal and without consent it exits 2 in that same
refusal's words, rather than assume a yes it cannot ask for or hang on a
prompt nothing can answer.
A confirmation, not bootstrap's flat refusal of the same file (#57/#59):
bootstrap asserts who lives on a box, apply converges, and converging to
zero stays a legitimate de-provisioning. Ledger entries already marked
revoked do not count toward the number, so a second identical run stays the
silent no-op convergence promises.
Mass revocation below the empty-file bright line is deliberately still
ungated — that needs a threshold someone has to justify.
Refs #65
uninstall_confirm() read the operator's answer unguarded:
read -r reply
case "$reply" in y|Y|yes|YES|Yes) return 0 ;; *) die "aborted." ;; esac
bin/rig runs under `set -euo pipefail`, and both call sites (the
single-version and the --all confirms) invoke the function as a plain
statement — nothing suppresses errexit. Ctrl-D makes `read` return
non-zero, so the shell died AT THE READ and the case on the next line
was never evaluated: `die "aborted."` could not fire. The operator saw
the question, pressed Ctrl-D, and got nothing — no message, exit 1, at
exactly the moment the tool had asked whether to delete their install.
It failed closed, so nothing was ever wrongly removed; the damage was
that rig went silent at the one moment silence is unreadable.
The fix is `read -r reply || reply=""` — commands/db.sh:152's spelling
for the identical [y/N] confirm one file away. Empty routes through the
existing `*)` arm, so EOF aborts through the same path a bare Enter
already does: exactly one "aborted." message, no second die to keep in
sync.
test/cli.sh gains the first drills of the interactive path, which was
structurally untested (every existing uninstall check goes through
--force or RIG_YES, which is why this survived): `y` and Ctrl-D driven
through a real pty via util-linux `script`, guarded by a command -v
skip. They assert the MESSAGE, never the exit code — the unfixed code
also exits 1, so an exit-code assertion is green against the bug.
Mutation-verified: with `|| reply=""` reverted, 403 passed / 1 failed,
the single failure being `output missing 'aborted.'`; restored, 404
passed / 0 failed.
Refs #68
CI's shellcheck step set `globstar` and globbed `bin/* **/*.sh`. Globs do
not match dot-prefixed names without `dotglob`, so `**/` never descended
into `.github/` and two tracked scripts were linted by nothing:
`.github/scripts/labels-reconcile.sh` and `.github/scripts/release-lib.sh`.
release-lib.sh is the one that matters: it holds `changelog_section`, which
release.yml sources to build the published release body and which
test/release.sh's `changelog_armed` guard calls to decide whether main is
armed. The script that decides both what ships and whether the changelog is
safe was the script CI never read.
Measured rather than assumed: `dotglob` adds exactly those two files to
rig's line and nothing else, and `**` descending into `.git/` matches no
`.sh` on a checkout. Both files already pass `shellcheck -x`, so this
closes a hole in the net rather than fixing a defect behind it.
Paired with a class check — `comm` against `git ls-files '*.sh'` — that
fails the step naming any tracked script outside the globbed set, so the
gap cannot reopen quietly. It also covers an escape `dotglob` does not:
`globstar` declines to traverse symlinked directories.
Refs #70
Caught in review. root_door_of matched unanchored substrings, so any value
that EXTENDS a real one resolved as that value: `root-door=closedish` read as
`closed` and PASSED close-root's gate -- the one arm in this repo that
authorizes an irreversible act -- and `class=humanoid` did the same through
the compat arm. Both contradicted the function's own header, which promises a
value outside the set resolves empty and fails closed.
Only reachable by hand-editing a marker, so it was never a live incident. It
gets fixed anyway because this is the single function every consumer trusts --
close-root's gate, apply's root-SSH note, and bootstrap-tenant's machine
guard all ask it -- and a resolver that is nearly right about a root door is
the wrong kind of nearly.
The marker is one line of space-separated key=value fields (bootstrap writes
it with a single printf), so padding both ends and matching on field
boundaries is exact rather than heuristic. Whitespace is normalised first so a
hand-edit using tabs still reads correctly -- anchoring must not trade one
silent misread for another.
BOTH vocabularies are anchored. Fixing only the current spelling would have
left the hole open on every box bootstrapped before #77, which is precisely
the population the compat arm exists to serve.
Tests pin the resolver and the end-to-end refusal, since the resolver
returning "" is only safe because consumers treat it as one. Reverting the
anchoring turns the suite red (447/4); restoring it returns 451/0. The
original compat proof still holds: removing the class= arm gives 441/10.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The trait was named for who lives on a box; what it decides is whether
root SSH stays open as the control plane's automation door. Those are
different questions, and `dev-server` proved it: an unattended VM-host
appliance nobody lives on, correctly class=human because its root door
must close. After #76 gave `-server` the job of naming the machine
family, that box carried a suffix saying server and a trait saying
human. `dev-server --root-door closed` says what is true, once.
Unlike #76's role rename this field is read back on live machines, so
the compat read is mandatory rather than courteous: one resolver,
root_door_of, reads both vocabularies and every consumer goes through
it — close-root's gate, apply's note, and bootstrap-tenant's
machine-marker guard, which used the presence of `class=` as its "is
this a real fleet machine?" test and would otherwise have let a tenant
converge clobber a live box. New markers are written as `root-door=`
only. Markers carrying both fields in disagreement, or neither, fail
closed with a re-run-bootstrap repair.
Fixture markers are kept deliberately at the retired spelling (the
convention #76's pre-rename-cp fixture established) and pinned at both
consumers; deleting the compat arm turns ten checks red.
Closes#77
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The other half of #76. claude -> claude-box, codex -> codex-box, grok ->
grok-box, staging -> staging-box, so a role name always says which family it
belongs to: -server builds a fleet machine, -box converges a guest a box
minted. With both halves in, the two families can no longer collide on a
word the way `staging` did.
The role carries the suffix; nothing inside the guest does. A tenant user is
the account the box SEED created (BOX_USER) and each agent CLI reads its own
dotdir, so claude-box still converges the `claude` user and still writes
~/.claude/CLAUDE.md. Every rename here is a $ROLE comparison or a case arm --
no CLI binary name, no dotdir path, and no account moved. README's tenant
table now shows role and user in adjacent columns, because that distinction
stopped being cosmetic the moment they differed.
Hard cut, no aliases. The old names are refused as unknown at BOTH
entrypoints -- `rig bootstrap <name>` and bootstrap-tenant.sh directly -- and
the suite asserts each of the four at each, because bootstrap.sh keeps its
own dispatch list and a name could survive in one and not the other. An alias
left in for a single tenant is the shape that survives review: the taxonomy
reads complete while one old name still quietly converges.
The consequence is cross-repo. A seed carrying BOX_BOOTSTRAP_ROLE="claude"
now fails its own mint-time bootstrap, so heavy-duty/box#123 updates the
seeds and must land after this.
Closes#76 (tenant half)
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Two review findings from the bot round on this stack.
BLOCKING (codex-bot, claude-bot -- both, independently). bootstrap-tenant.sh
emits the staging guest's tailnet-join next step at the end of a converge
("box shell -> sudo rig bootstrap workload"), repeats it in usage, and two of
its refusals recite the old machine-role list. Fixed here rather than on the
stacked tenant PR because THIS is the branch that removes the `workload` role
-- shipping it alone would print a next step naming a role that no longer
exists.
None of those four sites is code that ACCEPTS a role, which is why the rename
missed them, and is also what makes them the worse failure. A stale flag dies
immediately with a usage error. A stale next-step is copy-pasted by a human
onto a DIFFERENT box, minutes after the run that printed it reported success,
and dies there with no thread back to the cause.
So test/cli.sh sweeps every shipped script under bin/ and commands/ for
`rig bootstrap <pre-#76 name>` rather than pinning the four known sites: the
next instance of this class will be somewhere else. Proven non-vacuous --
reintroducing the bare `workload` next-step turns the suite red (412/1),
restoring it turns it green (413/0).
NON-BLOCKING (claude-bot). The migration story was documented and untested:
every marker fixture was renamed alongside the code, so nothing asserted what
a real pre-rename box does. A `role=control-plane` fixture now pins both
halves of the promise -- such a box WARNS on the coolify verbs (its marker no
longer names a role that exists) and is never REFUSED. Both halves matter: a
rename that turned this into a refusal would break the exact boxes the
CHANGELOG promises keep working, on the command that installs the control
plane.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
rig builds two kinds of thing on opposite sides of a trust boundary --
tailnet machines it converges, and guests a box mints -- and both families
lived in one flat namespace with nothing in a role name saying which you
meant. `staging` is where that stopped being cosmetic: the word names the
metal that hosts guests and the guests on it, only one could have it, and
#31 gave it to the guests. The VM-host shape was left nameless, spelled
`custom --class server --host yes --join authkey`, which is what every
refusal recited at an operator who had confused the two.
The suffix now names the family: control-plane-server, workload-server,
runner-server, dev-server, plus the restored staging-server (class=server
host=yes join=authkey). host=yes already installs the box CLI and runs box's
setup-host, so staging-server is a table row, not new machinery. It stays
OUT of the tag:server allow-list deliberately -- a host is never managed by
the control plane, its guests are -- so its key is minted tag:local.
custom and workstation keep bare names as the rule, not an exception to it:
custom presets nothing and can be any shape including a guest, so a family
claim is one it cannot make; a workstation is somebody's own device, joined
by interactive login, user-owned and untagged, never tailnet-managed.
Hard cut, no aliases -- old names are refused as unknown. Two consequences
this reaches beyond the CLI surface. TS_HOSTNAME defaults to the role name,
so a box taking the default now comes up control-plane-server. And the two
coolify commands match the ROLE NAME in /etc/rig/role, not the traits, so
they now look for role=control-plane-server; a pre-rename control plane
takes their warning branch, which is advisory and never a gate, so the run
proceeds and the message names the repair.
dev-server is class=human, which reads like a contradiction and is not: the
suffix names the family, the class names the root-SSH door policy. The two
axes share the word "server", which is a real wart -- #77 renames the class
trait to what it controls, kept separate because it reaches markers on live
machines that guard root SSH.
Tests cover both directions of the cut: every new name resolves, every old
name is refused as unknown, and the two deliberately-bare roles are proven
NOT to have been swept up -- the inverse error, which would otherwise only
surface at somebody's laptop.
Closes#76 (machine-role half)
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Bump VERSION 0.1.1-dev -> 0.2.0 and stamp CHANGELOG.md's Unreleased
section as `## 0.2.0 — 2026-07-19`. Minor, not patch: the section carries
an explicit `### Changed` / BREAKING entry (#51 — `rig bootstrap` takes
the users file, and requires it) alongside an `### Added` (#49), and a
patch release would misrepresent a documented breaking change.
Re-armed in the same PR per CONTRIBUTING step 1 (#66): a fresh, empty
`## Unreleased` sits immediately above the stamped section, so a PR
authored before this release and merged after it files its entry under
Unreleased rather than silently inside the release that already shipped.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
An empty, comments-only or whitespace-only users file is not a parse error,
so it walked straight through the requirement #51 built: pre-flight passed,
apply converged nothing, and the box came up root-only — the exact outcome
--no-users exists to make explicit, reached by the flag added to guarantee
the opposite. `--users ./empty` and `--no-users` produced the identical box
and only one of them said so.
Catch the zero-user parse in bootstrap's pre-flight, where the file is
already parsed for validation and before apt, the hostname change, or a
spent pre-auth key. The refusal names --no-users: the root-only box is
reachable, it just has to be asked for out loud.
Deliberately narrow. This is bootstrap's contract, not the parser's and not
apply's: zero users is a legal file, and a standalone `rig users apply`
against an emptied file is a real de-provisioning operation that must stay
possible. Negative-grep tests pin both.
Closes#57
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
`rig bootstrap` already knew everything else about what a box is — class,
host, join, hostname — and wrote /etc/rig/role to say so. The users file was
the last piece of that answer it did not take, so bring-up was two commands
and the second one was the forgettable one.
--users <path> now runs the `users apply` convergence as bootstrap's final
phase: after the traits, after the verified tailnet join, after the role
marker (apply reads that marker), and after the host=yes box install (so
box-role users find the incus group box's own setup-host built). One
command, and the box has its people on it.
BREAKING: --users is required on every machine role, with --no-users as the
explicit opt-out. Omitting both is a usage error naming both flags; passing
both is a usage error too. class=server is required as well: a machine
nobody logs into routinely is exactly where shared-root access rots, and
per-human accounts keep attribution intact for the times someone does go in.
The file is never persisted — passed per invocation, read once through
apply, copied nowhere. `--users -` is refused: bootstrap's stdin belongs to
the pre-auth key prompt. The box TENANT roles take neither flag; a guest is
minted non-interactively, never joins the tailnet, and has no SSH door of
its own.
rig still never installs Incus and never calls `box setup-host` itself. The
host=yes box-role precondition refuses early only where the outcome is
already proven (RIG_SKIP_BOX_INSTALL=1); every other way that step can fail
lands in `users apply`'s existing refusal, unchanged.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Role `box` resolved to exactly one action, `usermod -aG incus`. That is
the socket — step 1 of the five `box grant` performs. Without the other
four (the user-<uid> project, its narrowing to boxnet and only boxnet,
the snapshot and backup allowances clone and `box export` ride, and the
shipped box-net profile installed into that project) the user's first
`box new` refuses for want of a box-net profile, so apply's promise —
the users file is the fleet's source of truth — was not kept for this
role. Worse, until an admin arrived by hand the user held an `incus`
membership with no converged project, and incus-user would lazily hand
them a stock unhardened NAT bridge: a state box's own contract forbids.
On host=yes apply now calls `box grant <user>` per box-role user. rig
calls box's grant rather than reimplementing four fifths of it — the
"rig never installs Incus" boundary is about installation, not
invocation, and grant is already script-callable: idempotent,
root-or-sudo, stdin-pinned, with its own run-as-the-user touch.
Three decisions the code carries in comment form:
- Ordering. The call sits after `useradd` (grant opens with a getent
passwd and refuses an unknown account) and after the other groups, so
a user whose grant fails still lands with everything rig owns outright.
- Failure granularity, split the way the host= guard beside it already
splits. A missing box CLI on host=yes dies, like the missing incus
group: a broken VM host, not a per-user accident. A per-user grant
failure warns and continues — one box-role user somewhere in the fleet
must not stop apply everywhere VMs don't live. host=no and marker-less
boxes keep their existing skip-with-warning untouched.
- The group ADD is deferred to grant, while `incus` stays in the wanted
set so the exact-convergence loop never strips a box-role user's
socket. Grant's rollback only reaches a membership that run added, so
rig opening the socket first would leave a failed grant unable to
close it. And grant is the authority on whether the group belongs at
all: for an incus-admin member it deliberately does not add `incus`.
An incus-admin member is warned, never fatal: box grant refuses them
today, which heavy-duty/box#99 fixes box-side with no rig change needed.
Closes#49
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
`users apply` converged group `incus` with a bare `gpasswd -d`, the same
move it makes for `rig-admin` and `rig`. Those two are rig's. `incus` is
box's, and `box revoke` does strictly more with it: it says out loud that
supplementary groups are read AT LOGIN, so a session the dropped operator
already holds keeps the Incus socket until that session dies, and it hands
over `loginctl terminate-user <user>` as the remedy.
rig logged "removed <user> from incus" and moved on. An operator who
dropped someone from the users file and watched apply succeed believed the
VM access was gone — and was wrong for as long as that user held a session.
Both removal paths — the per-user convergence loop and the dropped-user
sweep — now route the incus group through one `drop_incus` helper that
calls `box revoke`, keeping a single owner for the group. Never `--purge`:
that deletes the user's boxes, images and project, and destroying someone's
running machines is not a convergence step; it stays an explicit admin act.
The exit code is not trusted (the #12 lesson bootstrap already applies to
box's installer): a revoke that returns 0 with the membership still
standing has not closed the socket, so the effective state is checked and
rig falls back to removing the group itself — as it also does on a host
where box is not installed. Every fallback path carries the session warning
in rig's own voice, because the silence was the bug. The absent-group case
needs no new guard: `id -nG` cannot report a group that does not exist, so
the existing `in_group` test at both call sites is already false on a
host=no box or one where `box setup-host` never ran.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Both reviewers caught it: #41 merged after the changelog convention
landed (#40) and skipped its entry; the release PR is the last gate
before the section becomes the permanent release body.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
rig's first release: VERSION leaves 0.1.0-dev and the Unreleased section
is stamped with today's date. The surface was drilled twice on 2026-07-19
(offline suites, tenant guests on real incus via box, db round-trip, a
real GitHub runner lifecycle, coolify install) and the ceremony itself
rehearsed end to end on a scratch fork — including the guard that made
this very PR possible (#44).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
users apply consulted the host= trait only when group incus was ABSENT,
so a host=no or marker-less box that nonetheless carried the group handed
box-role users a bare `usermod -aG incus` — the socket with no tier, which
incus-user answers by lazily building an unhardened project under whoever
opens it.
The marker now decides in both directions through one pure gate,
assert_marker_hosts_vms, so the verdict is identical whether or not the
group exists. The marker wins over the machine deliberately — it is the
box's declared identity and every other host= decision already treats it
as authoritative — but not silently: when the group exists and the trait
disagrees, the skip names the contradiction and rig bootstrap as the fix.
Closes#58
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The cast-twin blocker (claude-bot): a declared permissions: block zeroes
unspecified scopes, so the label read and the bump fallback's gh pr
create could only 403 — every genuine ceremony would end red at the
label check. pull-requests: write added, consumers named. CONTRIBUTING
step 3 and the changelog entry now tell the shipped story: push-to-main
door, event.before interlock, self-re-arm, manual-path bump stays the
operator's.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The rig twin of heavy-duty/box#96, from the release-ceremony retro: the
tag was a separate, manual, silent-when-forgotten step, and a forgotten
tag produces no red X — the worst failure shape. The ship decision
already lives in the release PR; merging it is "ship". After that,
tagging is transcription, and transcription belongs to machines.
release.yml now also fires on pull_request closed into main, gated on
merged AND the `release` label. The job asserts in order, each fail-loud
and creating nothing: VERSION at the merge commit is non--dev; VERSION
changed in THIS PR (base vs merge — the interlock that fails a
mislabeled ordinary PR); the changelog section for that version extracts
non-empty via the existing changelog_section from release-lib.sh; and no
tag or release exists yet. Then, in the same job, it API-creates the tag
at the merge commit and publishes the release with the extracted notes.
Same-job is load-bearing: a GITHUB_TOKEN-created tag does not fire the
tag-push trigger, so the publish must live next to the tag and the
fallback job cannot double-publish; the nothing-exists assert covers a
manual race. The tag-push path survives verbatim as the documented
manual fallback and backfill, and CONTRIBUTING's Releasing section now
reads merge-is-ship with the manual tag as fallback.
test/release.sh pins the merge path in the house grep-pin style: the
merged+labeled gate, the four asserts, the same-job tag+publish (awk
from release-on-merge: to EOF), the asserts-precede-the-tag ordering,
and the surviving tag-push trigger.
Fixes#47
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
test/release.sh demanded a literal '## Unreleased' heading extracting
non-empty with '#32' in it — all false by construction on the very tree
the release PR produces, so the first real 'release: 0.1.0' PR turned CI
red and the ceremony blocked itself. Both fork rehearsals missed it: they
tag a branch, which runs release.yml and never ci.yml. The guard now
asserts its actual purpose — the TOP section, whatever its name, extracts
non-empty via the exact function release.yml runs — and passes on both
legitimate tree states (verified on main's shape and on a stamped copy).
Fixes#44
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
A bare 'read -rsp' with no tty exits non-zero and set -e ends the script
with no output at all: the release drill watched 'rig runner remove' exit 1
in complete silence, and a guest bootstrap stop mid-log the same way. Every
prompt now checks for a tty first and dies naming the variable that
unblocks an unattended run; every read is || die-guarded so EOF at a real
prompt also gets a last word. The no-bare-read test swept up runner
repoint's two prompts, which the issue had not counted.
Fixes#42
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The PR-then-tag half of box#83's flow, rig's side. CHANGELOG.md starts
with an Unreleased section (feature PRs land their entry as part of the
PR — box's convention, now written into CONTRIBUTING alongside the
release ritual). On a tag push, release.yml asserts the bare tag equals
the tree's own VERSION — a mismatch fails loudly and creates nothing —
then creates the GitHub release with that version's changelog section as
the body, extracted by changelog_section in .github/scripts/release-lib.sh:
one function, sourced by the workflow and driven by test/release.sh
against fixtures and the shipped CHANGELOG.md itself. No assets — for a
pure-bash tree, the tag's source tarball IS the package.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
