heavy-duty/rig
7
0
Fork 1
Code Issues 10 Pull requests Projects Releases 3 Packages Wiki Activity Actions

feat: /etc/rig/manifest — which rig converged this machine, and when

Browse source 
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>
This commit is contained in:
dan-claude-bot 2026-07-20 10:05:13 +00:00
parent d6e8cdda06
commit 5c40e906a1
8 changed files with 644 additions and 0 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

31
CHANGELOG.md
View file

@ -45,6 +45,37 @@ on the way to cutting its first release, and this file starts there.
limits or the host's totals depends on whether `lxcfs` is in play — it is
unverified, so those two lines are unreliable there.
- **`/etc/rig/manifest` records which rig converged a machine, and when**
(#61) — the entire durable output of a bootstrap run was one line in
`/etc/rig/role`, and that line is about what the box *is*, never about what
built it. `VERSION` was read in exactly one place (`bin/rig:9`, for
`--version`) and that reports the *currently installed* tree, not the one
that ran; there was no timestamp anywhere in the codebase. SSH into a
control plane six months on and a machine converged by `0.1.0-dev` was
indistinguishable from one converged by `0.4.0`. Bootstrap — both the
machine roles and the box tenant roles — now stamps a second file beside
the marker: `schema=1`, `bootstrapped_by`/`bootstrapped_at` (the rig that
*first* converged this machine, pinned forever) and
`converged_by`/`converged_at` (the newest rig to have converged it), read
back with a new `rig manifest [<key>]`. `key=value`, one per line, `0644` —
never JSON or YAML, because this is the one file that must stay readable on
the most broken machine in the fleet and a rig-bootstrapped box has no YAML
parser and no `jq`.
Only **decided** facts go in, which is what keeps `bootstrap.sh:3`'s
contract ("a second run changes nothing", enforced by cmp-guards at nine
sites) 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. A naive
timestamp would have made every re-run a diff and had rig report a change it
did not make. **Observed** facts — cores, RAM, disk, kernel — are
deliberately absent: they go stale without rig doing anything, so they
belong to `rig platform` (#64), which computes them fresh and stores
nothing. `/etc/rig/role` is untouched — the marker holds traits and has six
readers; the manifest holds provenance. Readers must ignore keys they do not
know, and the writer preserves lines it does not own, so a manifest written
by a newer rig stays readable to (and survives a rewrite by) an older one.
### Changed
- **BREAKING: `--class human|server` is now `--root-door closed|open`** (#77) —

62
README.md
View file

@ -247,6 +247,13 @@ the marker later (`rig users` keys root policy off `root-door=`). Written
post-join and cmp-guarded, so a marker never describes a box that failed to
become what it claims.
Immediately after it, bootstrap stamps `/etc/rig/manifest` — **provenance**:
which rig converged this box and when (see [`rig
manifest`](#rig-manifest)). Same discipline, same guarantee, and the two files
stay consistent because they land together. The marker says what the box *is*;
the manifest says what *built* it. The tenant roles stamp it too — a
box-minted guest is a machine rig converged.
**`join=login` inverts the tag assertion.** A workstation joins as a
user-owned device: there is no pre-auth key — a set `TS_AUTHKEY` is a loud
usage error (exit 2; unset it, or pass `--join authkey`) — `tailscale up`
@ -1109,6 +1116,61 @@ them back, fresh keys and all. And the sudoers rules land in
bad file under `/etc/sudoers.d` can take down *all* of sudo, locking every
admin out of the very escalation path apply just granted.
### `rig manifest`
```sh
rig manifest # the whole provenance record
rig manifest converged_by # one value, for a shell caller
```
Prints `/etc/rig/manifest` — **which rig converged this machine, and when**.
`rig bootstrap` writes it as its last durable act, beside the role marker;
this command only reads, needs no root (the file is `0644`), and works on a
machine whose rig has since been upgraded or removed.
```
schema=1
bootstrapped_by=0.4.0
bootstrapped_at=2026-07-19T14:24:51Z
converged_by=0.6.0
converged_at=2026-08-02T09:11:03Z
```
Two pairs: **birth** — the rig that *first* converged this machine, pinned
forever — and **latest** — the newest rig to have converged it. On a fresh
machine they are equal. The version recorded is the one that **ran**, captured
at run time; `rig --version` reports the tree installed *now*, which after an
upgrade answers a different question, because a machine outlives the rig that
built it.
Only **decided** facts live here, and that is what keeps bootstrap's
convergence contract intact. `bootstrapped_*` is first-write-wins;
`converged_*` moves **only when the version actually differs** — it records
the time the converging version last changed, not the time of the last run. A
re-run by the same rig therefore renders a byte-identical file and the
cmp-guard stays silent; a re-converge by a *different* rig is a real change and
the guard firing there is correct.
**Observed** facts — cores, RAM, disk, kernel — are deliberately absent: they
go stale on their own (someone adds RAM; unattended-upgrades patches the
kernel), so storing them would either lie or force a rewrite on every run. They
belong to `rig platform`, which computes them fresh and stores nothing.
`key=value`, one per line — never JSON, never YAML. This is the one file that
must stay readable on the most broken machine in the fleet, and a
rig-bootstrapped box has no YAML parser and no `jq`. Readers must ignore keys
they do not know, so `schema=` is bumped only when a key is removed or
repurposed; a manifest written by a newer rig stays readable to an older one,
and the writer preserves lines it does not own rather than eating them. Nothing
here is ever a credential.
Exits 1 when there is no manifest — a machine converged before rig wrote one,
or never converged at all. `RIG_MANIFEST` overrides the path.
The manifest does **not** replace `/etc/rig/role`. The marker holds *traits*
(what this box is) and has six readers; the manifest holds *provenance* (what
built it). Two files, two jobs.
### `rig users status`
```sh

11
bin/rig
View file

@ -91,6 +91,13 @@ commands:
Shut root SSH on a class=human box once an admin key works. Refuses
on class=server — root there is the control plane's automation door —
and while no admin holds a key. Run as root.
manifest [<key>]
Print /etc/rig/manifest — which rig converged this machine and when
(bootstrapped_by/_at pin the FIRST convergence forever; converged_by/_at
name the latest). With a key, print that value alone for shell callers.
Written by 'rig bootstrap' only; read-only here, and no root needed.
Provenance is DECIDED, so it is stored; the machine's specs are
OBSERVED, so they are not — those are 'rig platform'.
versions
List the installed rig versions — install.sh lands each one side by
side at <root>/versions/<v>, a 'current' symlink tracks the default
@ -444,6 +451,10 @@ case "$cmd" in
;;
esac
;;
manifest)
shift
exec "$ROOT/commands/manifest.sh" "$@"
;;
versions)
shift
cmd_versions "$@"

27
commands/bootstrap-tenant.sh
View file

@ -23,6 +23,8 @@ HERE="$(cd "$(dirname "$(readlink -f "${BASH_SOURCE[0]}")")" && pwd)"
. "$HERE/lib/users-config.sh" # read_role_marker / root_door_of
# shellcheck source=SCRIPTDIR/lib/sshd.sh
. "$HERE/lib/sshd.sh" # harden_sshd (the staging-box tenant)
# shellcheck source=SCRIPTDIR/lib/manifest.sh
. "$HERE/lib/manifest.sh" # manifest_stamp — provenance, written beside the marker
log() { printf 'rig-bootstrap: %s\n' "$*"; }
warn() { printf 'rig-bootstrap: WARNING: %s\n' "$*" >&2; }
@ -405,6 +407,31 @@ else
log "machine role marker present (${EXISTING_MARKER}); leaving it alone"
fi
# --- provenance manifest ------------------------------------------------------
# A tenant gets a manifest, in the SAME /etc/rig/manifest, through the same
# writer — and UNCONDITIONALLY, outside the marker gate above (#61's open
# question, answered here).
#
# The reason the marker needs that gate is that it holds TRAITS, and a guest's
# traits and the traits it earns after an operator-run `rig bootstrap workload`
# join are two different, competing statements about one box — so the marker
# has to pick, and it picks the truer one. Provenance has no such conflict.
# "Which rig converged this guest, and when" is a fact whichever bootstrap ran,
# and the two-pair shape composes across them exactly as designed: a staging
# guest later joined as a workload keeps the TENANT bootstrap as its birth —
# that genuinely is when this machine was first converged — and the machine
# bootstrap moves converged_* forward. Skipping the write on a joined guest
# would lose the birth stamp that only this run knows.
#
# One file, not a tenant-shaped second one: the manifest answers a question
# about the MACHINE, and a guest is a machine. The marker already carries
# `tenant=yes` for anyone who needs to know which kind.
if manifest_stamp "$(manifest_running_version "$HERE/..")"; then
log "provenance manifest written: $(manifest_path)"
else
log "provenance manifest already current"
fi
log "done — tenant ${ROLE}, user ${TENANT_USER}"
if [ "$ROLE" = "staging-box" ]; then
log "next (operator-run, holds a credential): box shell → sudo rig bootstrap workload-server --hostname <name> with a tagged pre-auth key"

27
commands/bootstrap.sh
View file

@ -10,6 +10,8 @@ HERE="$(cd "$(dirname "$(readlink -f "${BASH_SOURCE[0]}")")" && pwd)"
. "$HERE/lib/sshd.sh" # harden_sshd — shared with the staging tenant
# shellcheck source=SCRIPTDIR/lib/users-config.sh
. "$HERE/lib/users-config.sh" # parse_users_file — the --users PRE-FLIGHT only
# shellcheck source=SCRIPTDIR/lib/manifest.sh
. "$HERE/lib/manifest.sh" # manifest_stamp — provenance, written beside the marker
# The users lib is sourced for validation, never for convergence: `users apply`
# stays the single owner of what a users file DOES to a box (#51). Bootstrap
# borrows the parser so a typo'd users file is caught in the same breath as a
@ -619,6 +621,31 @@ else
fi
rm -f "$MARKER_TMP"
# --- provenance manifest ------------------------------------------------------
# /etc/rig/manifest records WHICH rig converged this machine and WHEN (#61).
# The marker above says what the box IS; this says what BUILT it — two files,
# two jobs, and the marker is deliberately untouched (it has six readers,
# install.sh:82-90 among them).
#
# Placed HERE, immediately after the marker, so the two agree by construction
# and both inherit the marker's discipline verbatim: written only AFTER the tag
# verification, so neither ever describes a box that failed to become what it
# claims. A manifest that survives a failed run is worse than no manifest — it
# is a confident wrong answer. It deliberately does NOT trail the box install
# and the users phase below: those are the host EXTRA and the operator phase,
# and a box whose people failed to converge was still converged BY this rig at
# this time. Stamping provenance is not a claim that everything after it
# succeeded — the marker beside it makes exactly the same claim, and the two
# landing together is what keeps them readable as one statement.
#
# The version stamped is the one that RAN, captured now — not what `rig
# --version` would answer after a later upgrade.
if manifest_stamp "$(manifest_running_version "$HERE/..")"; then
log "provenance manifest written: $(manifest_path)"
else
log "provenance manifest already current"
fi
# --- box install (host=yes only) -------------------------------------------
# A host=yes box exists to run guest boxes, so bootstrap finishes the job rather
# than printing a to-do: it installs the box CLI globally and lets box's OWN

185
commands/lib/manifest.sh Normal file
View file

@ -0,0 +1,185 @@
#!/usr/bin/env bash
# /etc/rig/manifest — PROVENANCE: which rig converged this machine, and when
# (#61). Sourced by bootstrap.sh, bootstrap-tenant.sh and the test harness.
#
# The file is `key=value`, one per line — not JSON, not YAML. Same constraint
# stated three times in the tree already (lib/users-config.sh:6-12,
# lib/runner-config.sh:6 and :24): a rig-bootstrapped box has no YAML parser
# and no jq, which is why json_field() is grep-and-sed. This is the one file
# that must stay readable on the most broken machine in the fleet, so `read`
# parses it for free.
#
# WHAT GOES IN HERE: facts that are DECIDED. Which rig ran, and when it ran.
# Facts that are OBSERVED — cores, RAM, disk, kernel — belong to `rig platform`
# (#64), which computes them fresh and stores nothing. That split is not
# tidiness: a stored spec goes stale on its own (someone adds RAM; the
# unattended-upgrades bootstrap.sh itself enables patches the kernel), and
# refreshing it on every run collides head-on with bootstrap.sh:3 —
# "Convergent: safe to re-run; a second run changes nothing." Keeping only
# immutable content removes that problem instead of managing it.
#
# NEVER A CREDENTIAL. This file is 0644 and world-readable by design — it is
# an audit record, and an audit record nobody can read is not one. Later
# commands may append their own provenance (runner_installed_at,
# coolify_installed_at, box_version), subject to the same two rules: the EVENT
# of installing something, never its current state, and never a secret. Same
# law runner-install.sh:190 already states for `.rig-labels` — "box-local
# metadata, never a credential."
# The schema version, an INTEGER, independent of the rig versions recorded in
# the file. Bumped only when a key is REMOVED or REPURPOSED — adding a key is
# not a bump, and readers must ignore keys they do not know, so a newer rig's
# manifest stays readable to an older one. No `schema=` line means pre-manifest.
MANIFEST_SCHEMA=1
# The five keys rig's provenance block owns. Everything else in the file is a
# later command's line and is preserved verbatim (see manifest_foreign).
MANIFEST_KEYS='schema bootstrapped_by bootstrapped_at converged_by converged_at'
# manifest_path — where the manifest lives. RIG_MANIFEST overrides it so tests
# point at fixtures (repo precedent: RIG_ROLE_MARKER, bin/rig:148).
manifest_path() {
printf '%s' "${RIG_MANIFEST:-/etc/rig/manifest}"
}
# manifest_value <path> <key> — the value of one key, or nothing when the file
# or the key is absent. First occurrence wins. NO policy here: what a missing
# key MEANS is each caller's call (this reader only reads — repo precedent:
# read_role_marker).
manifest_value() {
[ -r "$1" ] || return 0
local k v
while IFS='=' read -r k v; do
[ "$k" = "$2" ] || continue
printf '%s' "$v"
return 0
done < "$1"
}
# manifest_has <path> <key> — is the key PRESENT, regardless of its value.
# Separate from manifest_value because "absent" and "present but empty" are
# different answers and command substitution collapses both to the empty
# string. Key comparison is a string equality, never a pattern: `rig manifest`
# passes operator input straight in, and a key of `.*` must find nothing rather
# than match the first line.
manifest_has() {
[ -r "$1" ] || return 1
local k
while IFS='=' read -r k _; do
[ "$k" = "$2" ] && return 0
done < "$1"
return 1
}
# manifest_foreign <path> — every line rig's provenance block does NOT own,
# in file order. A newer rig (or a later command) may have written keys this
# one has never heard of; rewriting the file must not eat them, or the schema's
# "readers ignore keys they do not know" promise would only hold for readers
# and not for the writer.
manifest_foreign() {
[ -r "$1" ] || return 0
local line key
while IFS= read -r line; do
[ -n "$line" ] || continue
key="${line%%=*}"
case " $MANIFEST_KEYS " in
*" $key "*) continue ;;
esac
printf '%s\n' "$line"
done < "$1"
}
# manifest_render <path> <version> <now> — the WHOLE convergence contract, as
# a pure text→text function: existing file + running version + a clock reading
# in, the file's desired content out. No side effects, so the harness proves
# the rules non-root against fixtures (repo precedent: parse_users_file,
# assert_marker_human).
#
# Purity is what makes convergence testable rather than asserted. The rendered
# content is a function of (existing file, running version) ALONE — <now> is
# consulted only on the paths that were going to change anyway — so calling
# this twice with two DIFFERENT clock readings must produce byte-identical
# output. test/cli.sh pins exactly that, which is stronger than re-running the
# writer fast enough that the second matches by luck.
#
# Rule 1 — bootstrapped_* is FIRST-WRITE-WINS. Birth is pinned forever. If the
# file already carries a bootstrapped_at, both birth fields are preserved
# verbatim. Regenerating it as now() on every run would make every re-run a
# diff, which is the exact trap that keeping specs out of this file closed.
#
# Rule 2 — converged_* updates ONLY when the version actually differs.
# converged_at is "the time the converging version last changed", NOT the time
# of the last run. If it tracked every run it would be a clock, and a clock in
# a cmp-guarded file makes every re-run a fake change. So: compare the running
# version against the recorded converged_by; equal means the pair is already
# true and is copied through untouched.
#
# Under those two rules a re-run by the SAME rig renders byte-identical content
# and the cmp-guard stays silent, while a re-converge by a DIFFERENT rig
# renders a real diff — and the guard firing there is correct, not spurious. It
# was only ever the clock that was the fake change, never the version.
manifest_render() {
local path="$1" ver="$2" now="$3"
local b_by b_at c_by c_at
b_by="$(manifest_value "$path" bootstrapped_by)"
b_at="$(manifest_value "$path" bootstrapped_at)"
c_by="$(manifest_value "$path" converged_by)"
c_at="$(manifest_value "$path" converged_at)"
# Rule 1. bootstrapped_at is the field that decides, because it is the one
# that can never be reconstructed: a machine's birth version can at least be
# guessed at, its birth INSTANT cannot. So an existing at-stamp pins the
# pair, and a birth-stamp with no birth version records `unknown` rather
# than backfilling today's version as if it had always been there — a
# manifest that lies about which rig built the box is worse than one that
# admits it does not know.
if [ -z "$b_at" ]; then
b_by="$ver"; b_at="$now"
elif [ -z "$b_by" ]; then
b_by=unknown
fi
# Rule 2. The empty-at case is a one-time repair of a truncated file, not a
# clock: once written it satisfies the equality on every later run.
if [ "$c_by" != "$ver" ] || [ -z "$c_at" ]; then
c_by="$ver"; c_at="$now"
fi
printf 'schema=%s\n' "$MANIFEST_SCHEMA"
printf 'bootstrapped_by=%s\n' "$b_by"
printf 'bootstrapped_at=%s\n' "$b_at"
printf 'converged_by=%s\n' "$c_by"
printf 'converged_at=%s\n' "$c_at"
manifest_foreign "$path"
}
# manifest_running_version <rig root> — the version that IS RUNNING, captured
# at run time from the tree's own VERSION file. NOT `rig --version` read back
# later: a machine outlives the rig that built it, so what is installed today
# answers a different question than what converged it. First line only — a
# stray second line would inject a bogus key into a key=value file.
manifest_running_version() {
local v=""
[ -r "$1/VERSION" ] && v="$(head -n1 "$1/VERSION")"
printf '%s' "${v:-unknown}"
}
# manifest_stamp <version> — the writer. Renders, cmp-guards like every file
# rig converges, installs 0644 beside the role marker and the users ledger.
# Returns 0 when it CHANGED the file and 1 when the file was already current,
# so the caller owns the log line (and, under set -e, must call it in an `if`).
manifest_stamp() {
local path tmp rc
path="$(manifest_path)"
tmp="$(mktemp)"
manifest_render "$path" "$1" "$(date -u +%Y-%m-%dT%H:%M:%SZ)" > "$tmp"
if cmp -s "$tmp" "$path" 2>/dev/null; then
rc=1
else
mkdir -p "$(dirname "$path")"
install -m 0644 "$tmp" "$path"
rc=0
fi
rm -f "$tmp"
return "$rc"
}

88
commands/manifest.sh Executable file
View file

@ -0,0 +1,88 @@
#!/usr/bin/env bash
# rig manifest — print this machine's provenance record, /etc/rig/manifest
# (#61). Reads only: this command NEVER writes the file. `rig bootstrap` is its
# single writer, and it stamps it as its own last durable act.
set -euo pipefail
HERE="$(cd "$(dirname "$(readlink -f "${BASH_SOURCE[0]}")")" && pwd)"
# shellcheck source=SCRIPTDIR/lib/manifest.sh
. "$HERE/lib/manifest.sh" # manifest_path / manifest_value
die() { printf 'rig-manifest: ERROR: %s\n' "$1" >&2; exit "${2:-1}"; }
usage() {
cat <<'EOF'
usage: rig manifest [<key>]
Print /etc/rig/manifest — which rig converged this machine, and when. With a
key, print that key's value alone, unquoted and newline-terminated, so a shell
caller does not re-parse the file:
rig manifest schema=1
bootstrapped_by=0.2.0
bootstrapped_at=2026-07-19T14:24:51Z
converged_by=0.4.0
converged_at=2026-08-02T09:11:03Z
rig manifest converged_by 0.4.0
Both pairs are provenance, both immutable in the sense that matters: BIRTH —
the rig that first converged this machine, pinned forever — and LATEST — the
newest rig to have converged it. On a fresh machine the two are equal. Ask the
second pair "is this machine converged by something ancient?"; ask the first
"what built it".
The version recorded is the one that RAN. `rig --version` reports the tree
installed NOW, which after an upgrade is a different question — a machine
outlives the rig that built it.
Reads only. Needs no root (the file is 0644), no network, and works on a
machine whose rig has since been upgraded or removed. Specs — cores, RAM,
disk, kernel — are NOT here: they are observed rather than decided, so they go
stale on their own and belong to `rig platform`, which computes them fresh and
stores nothing.
Exit 1 when there is no manifest — a machine converged before rig wrote one,
or never converged at all.
RIG_MANIFEST override the path (default /etc/rig/manifest)
EOF
}
KEY=""
for a in "$@"; do
case "$a" in
-h|--help) usage; exit 0 ;;
-*)
printf 'rig-manifest: unknown option: %s\n' "$a" >&2
usage >&2
exit 2
;;
*)
if [ -n "$KEY" ]; then
printf 'rig-manifest: manifest takes at most one key\n' >&2
usage >&2
exit 2
fi
KEY="$a"
;;
esac
done
MPATH="$(manifest_path)"
[ -r "$MPATH" ] || die "no manifest at $MPATH — this machine has not been converged by a rig that writes one (rig bootstrap writes it)"
if [ -z "$KEY" ]; then
cat "$MPATH"
exit 0
fi
# A key that is absent and a key whose value is empty are different answers to
# a shell caller, and $(...) collapses both to "". So the ABSENCE is the exit
# code, and only a present key ever prints — a caller reading `rig manifest
# converged_by` into a variable can trust that an empty result it accepted was
# a real empty value, not a missing key.
manifest_has "$MPATH" "$KEY" \
|| die "no such key: $KEY (keys present: $(cut -d= -f1 "$MPATH" | tr '\n' ' '))"
manifest_value "$MPATH" "$KEY"
echo

213
test/cli.sh
View file

@ -1907,6 +1907,219 @@ rm -f "$DUMP_TMP"
check "no main-shell os-release sourcing" 1 "" \
grep -rnE '^[[:space:]]*\.[[:space:]]+/etc/os-release' "$ROOT/commands"
# ---------------------------------------------------------------------------
# /etc/rig/manifest — provenance (#61)
#
# The writer is a pure text→text renderer behind a cmp-guard, for the same
# reason assert_marker_human and parse_users_file are pure: the write itself
# sits behind bootstrap's root check, so every RULE is proven here, non-root,
# against fixtures — and the rules are the whole feature.
# ---------------------------------------------------------------------------
# Every helper sources the lib in a SUBSHELL, so the harness's own $PASS/$FAIL
# and the lib's globals never meet (repo precedent: the bash -c gates above,
# same isolation, one less quoting layer).
render() { # render <path> <version> <now>
( set -euo pipefail
. "$ROOT/commands/lib/manifest.sh"
manifest_render "$1" "$2" "$3" )
}
stamp() { # stamp <path> <version> — the side-effecting writer
( set -euo pipefail
. "$ROOT/commands/lib/manifest.sh"
RIG_MANIFEST="$1" manifest_stamp "$2" )
}
running_version() { # running_version <rig tree root>
( set -euo pipefail
. "$ROOT/commands/lib/manifest.sh"
manifest_running_version "$1" )
}
mode_of() { stat -c %a "$1"; }
mtime_of() { stat -c %Y "$1"; }
text_is() { [ "$(cat "$1")" = "$2" ]; }
MF="$(mktemp -d)"
# -- the shape on a virgin machine ------------------------------------------
check "manifest: a fresh render carries schema=1" 0 "schema=1" \
render "$MF/absent" 0.4.0 2026-07-19T14:24:51Z
check "manifest: a fresh render pins birth to the running version" 0 "bootstrapped_by=0.4.0" \
render "$MF/absent" 0.4.0 2026-07-19T14:24:51Z
check "manifest: a fresh render stamps birth with now" 0 "bootstrapped_at=2026-07-19T14:24:51Z" \
render "$MF/absent" 0.4.0 2026-07-19T14:24:51Z
# Both pairs equal on a fresh machine: mild redundancy, in exchange for a file
# no reader ever has to infer a missing field from.
check "manifest: a fresh render writes latest equal to birth" 0 "converged_by=0.4.0" \
render "$MF/absent" 0.4.0 2026-07-19T14:24:51Z
check "manifest: a fresh render stamps latest with now" 0 "converged_at=2026-07-19T14:24:51Z" \
render "$MF/absent" 0.4.0 2026-07-19T14:24:51Z
# key=value, one per line, nothing else — the shape a machine with no jq and no
# YAML parser can read with `read`. Five lines, no quoting, no nesting.
render_is_flat_kv() { # render_is_flat_kv <path> <version>
local out
out="$(render "$1" "$2" 2026-07-19T14:24:51Z)" || return 1
[ "$(printf '%s\n' "$out" | wc -l)" -eq 5 ] || return 1
! printf '%s\n' "$out" | grep -qvE '^[a-z_]+=[^ ]*$'
}
check "manifest: the render is bare key=value, one per line" 0 "" \
render_is_flat_kv "$MF/absent" 0.4.0
# -- THE CRUX: convergence ---------------------------------------------------
# bootstrap.sh:3 promises "a second run changes nothing", enforced by a
# cmp-guard before every file install. A naive `converged_at=$(now)` would
# break that promise on every single re-run — the file would differ by a
# timestamp, the guard would fire, and rig would report a change it did not
# make. Rules 1 and 2 make the rendered content a function of (existing file,
# running version) alone.
#
# Proven the strong way: render the SAME fixture twice with two clock readings
# a year apart and diff. Byte-identical output means the clock cannot reach the
# file at all — a stronger claim than re-running the writer quickly enough that
# the two stamps happen to match by luck.
clock_cannot_reach() { # clock_cannot_reach <path> <version>
diff <(render "$1" "$2" 2027-01-01T00:00:00Z) <(render "$1" "$2" 2028-06-06T06:06:06Z)
}
reproduces_itself() { # reproduces_itself <path> <version>
diff <(render "$1" "$2" 2029-09-09T09:09:09Z) "$1"
}
BORN="$MF/born"
render "$MF/absent" 0.4.0 2026-07-19T14:24:51Z > "$BORN"
check "manifest: re-render by the SAME rig is byte-identical across a year of clock" 0 "" \
clock_cannot_reach "$BORN" 0.4.0
check "manifest: re-render by the same rig equals the file it read" 0 "" \
reproduces_itself "$BORN" 0.4.0
# The same property through the WRITER, which is what bootstrap actually calls:
# a first stamp writes (exit 0), a second by the same rig reports the file
# already current (exit 1) and touches nothing.
STAMPED="$MF/stamped"
check "manifest: the first stamp writes the file" 0 "" stamp "$STAMPED" 0.4.0
check "manifest: the file landed 0644 — an audit record nobody can read is not one" \
0 "644" mode_of "$STAMPED"
BEFORE="$(cat "$STAMPED")"
check "manifest: a second stamp by the same rig reports already-current" 1 "" stamp "$STAMPED" 0.4.0
check "manifest: a second stamp by the same rig changed no byte" 0 "" \
text_is "$STAMPED" "$BEFORE"
# -- a DIFFERENT rig: a real diff, and only where it belongs ----------------
# The cmp-guard firing here is CORRECT, not spurious. It was only ever the
# clock that was the fake change, never the version.
check "manifest: a re-converge by a newer rig moves converged_by" 0 "converged_by=0.6.0" \
render "$BORN" 0.6.0 2026-08-02T09:11:03Z
check "manifest: a re-converge by a newer rig moves converged_at" 0 "converged_at=2026-08-02T09:11:03Z" \
render "$BORN" 0.6.0 2026-08-02T09:11:03Z
# Rule 1, the load-bearing half: birth is FIRST-WRITE-WINS. bootstrapped_by
# must survive every later convergence — "what built this box" is unanswerable
# by any other means once the run is over.
check "manifest: a re-converge leaves the birth version pinned" 0 "bootstrapped_by=0.4.0" \
render "$BORN" 0.6.0 2026-08-02T09:11:03Z
check "manifest: a re-converge leaves the birth stamp pinned" 0 "bootstrapped_at=2026-07-19T14:24:51Z" \
render "$BORN" 0.6.0 2026-08-02T09:11:03Z
check "manifest: the writer sees a version change as a real change" 0 "" stamp "$STAMPED" 0.6.0
check "manifest: and settles again on the new version" 1 "" stamp "$STAMPED" 0.6.0
# -- a DOWNGRADE is a change too --------------------------------------------
# converged_by is "the rig that last converged this", not "the highest one ever
# seen". Rolling back with `rig use` and re-converging must be recorded, or the
# file would name a version that is no longer what runs here.
check "manifest: re-converging with an OLDER rig is recorded, not ignored" 0 "converged_by=0.1.0" \
render "$BORN" 0.1.0 2026-09-09T09:09:09Z
# -- forward compatibility: unknown keys survive the rewrite ----------------
# The schema promises readers ignore keys they do not know. That promise is
# worthless if the WRITER eats them: a manifest touched by a newer rig, or
# carrying a later command's own provenance line, must come back whole.
FOREIGN="$MF/foreign"
{ cat "$BORN"; printf 'runner_installed_at=2026-07-19T16:10:00Z\n'; printf 'schema_future_key=x\n'; } > "$FOREIGN"
check "manifest: a later command's provenance line survives a rewrite" 0 "runner_installed_at=2026-07-19T16:10:00Z" \
render "$FOREIGN" 0.6.0 2026-08-02T09:11:03Z
check "manifest: a key from a newer schema survives a rewrite" 0 "schema_future_key=x" \
render "$FOREIGN" 0.6.0 2026-08-02T09:11:03Z
# And preserving them must not cost convergence: a file carrying foreign keys is
# still byte-stable under a same-version re-render.
check "manifest: foreign keys do not break convergence" 0 "" \
reproduces_itself "$FOREIGN" 0.4.0
# -- damaged files: repair once, then settle --------------------------------
# A truncated or hand-edited manifest must converge back to a whole one and
# then STAY PUT — a repair that re-fires on every run is a clock by another
# name, and would break convergence exactly where it is hardest to notice.
printf 'bootstrapped_at=2020-01-01T00:00:00Z\n' > "$MF/noby"
check "manifest: a birth stamp with no birth version records unknown, never today's" \
0 "bootstrapped_by=unknown" render "$MF/noby" 0.4.0 2026-07-19T14:24:51Z
check "manifest: ...and still keeps the birth stamp it does have" \
0 "bootstrapped_at=2020-01-01T00:00:00Z" render "$MF/noby" 0.4.0 2026-07-19T14:24:51Z
printf 'schema=1\nbootstrapped_by=0.4.0\nbootstrapped_at=2020-01-01T00:00:00Z\nconverged_by=0.4.0\n' > "$MF/noat"
REPAIRED="$MF/repaired"
render "$MF/noat" 0.4.0 2026-07-19T14:24:51Z > "$REPAIRED"
check "manifest: a converged_by with no converged_at is repaired once" 0 "converged_at=2026-07-19T14:24:51Z" \
cat "$REPAIRED"
check "manifest: the repair settles — it does not re-fire on the next run" 0 "" \
reproduces_itself "$REPAIRED" 0.4.0
# -- the version that RAN, not the one installed now ------------------------
# The whole point: a machine outlives the rig that built it, so this is read
# from the tree at run time and never re-derived from `rig --version` later.
check "manifest: the running version comes from the tree's own VERSION" 0 "$(cat "$ROOT/VERSION")" \
running_version "$ROOT"
check "manifest: a tree with no VERSION records unknown, not an empty key" 0 "unknown" \
running_version "$MF"
# -- the marker is NOT touched ----------------------------------------------
# /etc/rig/role has six readers, install.sh:82-90 among them; the manifest is a
# second file beside it, never a replacement. Assert the writer cannot reach it.
check "manifest: no CODE in the manifest lib reaches the role marker" 1 "" \
grep -nE '^[^#]*(/etc/rig/role|RIG_ROLE_MARKER)' "$ROOT/commands/lib/manifest.sh"
check "bootstrap: the role marker write is still its own cmp-guarded block" 0 "" \
grep -qE '^MARKER=/etc/rig/role$' "$ROOT/commands/bootstrap.sh"
# -- ordering: provenance is written after the tag verification -------------
# The marker's discipline, inherited verbatim — a manifest that survives a run
# which failed to become what it claims is a confident wrong answer.
mfstamp_at="$(grep -nE '^if manifest_stamp ' "$ROOT/commands/bootstrap.sh" | head -n1 | cut -d: -f1)"
verify_at="$(grep -nE '^[[:space:]]*verify_effective_tag back-out$' "$ROOT/commands/bootstrap.sh" | head -n1 | cut -d: -f1)"
check "bootstrap: both ordering anchors were found (guards the greps above)" 0 "" \
test -n "${mfstamp_at:-}" -a -n "${verify_at:-}"
check "bootstrap: the manifest stamp follows the tag verification" 0 "" \
test "${mfstamp_at:-0}" -gt "${verify_at:-999999}"
# Both bootstrap paths stamp it: a box-minted guest is a machine rig converged,
# and "which rig, when" is a fact whichever bootstrap ran.
check "bootstrap-tenant: a tenant gets a manifest too" 0 "" \
grep -q 'manifest_stamp' "$ROOT/commands/bootstrap-tenant.sh"
# -- `rig manifest`, the reader ---------------------------------------------
key_prints_exactly() { # key_prints_exactly <path> <key> <want>
[ "$(RIG_MANIFEST="$1" "$ROOT/commands/manifest.sh" "$2")" = "$3" ]
}
check "manifest: --help exits 0" 0 "usage: rig manifest" "$ROOT/commands/manifest.sh" --help
check "manifest: dispatches through bin/rig" 0 "usage: rig manifest" "$ROOT/bin/rig" manifest --help
check "manifest: rig --help lists the command" 0 "manifest [<key>]" "$ROOT/bin/rig" --help
check "manifest: unknown flag exits 2" 2 "unknown option" \
env RIG_MANIFEST="$STAMPED" "$ROOT/commands/manifest.sh" --nope
check "manifest: two keys is a usage error" 2 "at most one key" \
env RIG_MANIFEST="$STAMPED" "$ROOT/commands/manifest.sh" converged_by schema
check "manifest: an absent manifest exits 1 by name" 1 "no manifest at" \
env RIG_MANIFEST="$MF/absent" "$ROOT/commands/manifest.sh"
check "manifest: bare prints the file" 0 "converged_by=0.6.0" \
env RIG_MANIFEST="$STAMPED" "$ROOT/commands/manifest.sh"
check "manifest: a key prints the value ALONE, for shell callers" 0 "" \
key_prints_exactly "$STAMPED" converged_by 0.6.0
check "manifest: an unknown key exits 1 and names the keys present" 1 "keys present" \
env RIG_MANIFEST="$STAMPED" "$ROOT/commands/manifest.sh" nosuchkey
# Operator input reaches the key lookup, so the lookup is a string equality and
# never a pattern — a key of '.*' must find nothing rather than match line one.
check "manifest: a regex-shaped key matches nothing" 1 "no such key" \
env RIG_MANIFEST="$STAMPED" "$ROOT/commands/manifest.sh" '.*'
# The reader writes NOTHING — bootstrap is the manifest's single writer.
MTIME_BEFORE="$(mtime_of "$STAMPED")"
RIG_MANIFEST="$STAMPED" "$ROOT/commands/manifest.sh" >/dev/null 2>&1 || true
check "manifest: reading it does not write it" 0 "$MTIME_BEFORE" mtime_of "$STAMPED"
# Secrets: this file is 0644 by design, so the rule has to be stated where the
# next command that appends a line will read it (repo precedent:
# runner-install.sh:190's ".rig-labels — box-local metadata, never a credential").
check "manifest: the never-a-credential rule is stated in the writer" 0 "never a credential" \
cat "$ROOT/commands/lib/manifest.sh"
rm -rf "$MF"
# ---------------------------------------------------------------------------
# The versioned install (box#79's layout, ported — #35). RIG_INSTALL_SOURCE
# bypasses the network, so these are REAL runs of install.sh against throwaway