feat(mosaic): manifest-owned upgrade guard so updates never wipe operator config (#791) (#802)
All checks were successful
ci/woodpecker/push/publish Pipeline was successful
ci/woodpecker/push/ci Pipeline was successful

This commit was merged in pull request #802.
This commit is contained in:
2026-07-16 23:01:26 +00:00
parent 8536454257
commit 32a0ffba13
18 changed files with 2711 additions and 124 deletions

View File

@@ -0,0 +1,85 @@
# Mosaic framework path-ownership manifest — SSOT for the updater.
#
# This single file is the source of truth consumed by BOTH the bash installer
# (packages/mosaic/framework/install.sh) and the TypeScript config adapter
# (packages/mosaic/src/config/file-adapter.ts). A parity test asserts both
# paths resolve the same ownership from this file, so the two can never drift
# (the failure mode that #631 patched by hand in two places).
#
# Format: one glob per line, relative to the mosaic home (~/.config/mosaic).
# - Lines starting with '#' and blank lines are ignored.
# - '[framework]' / '[operator]' switch the active section.
# - '**' matches any depth; '*' matches within a single path segment.
#
# Ownership resolution for a path P (deny-wins / fail-safe):
# 1. P matches an [operator] glob -> operator-owned.
# 2. else P matches a [framework] glob -> framework-owned.
# 3. else (matches neither) -> OPERATOR-OWNED BY DEFAULT.
#
# Rule 3 is the root-cause fix for #791: a path the manifest authors never
# anticipated is protected because UNKNOWN defaults to operator. The updater
# may only ever create/overwrite framework-owned paths, and may only prune a
# framework-owned path that lives inside a shipped framework subtree and is
# absent from the current framework source (a genuinely retired file).
# Operator-owned and unknown paths are structurally unreachable by pruning.
[framework]
# Top-level framework contract files (also reconciled from defaults/ on upgrade).
CONSTITUTION.md
AGENTS.md
STANDARDS.md
# Shipped framework subtrees — pruning is scoped to these roots.
adapters/**
constitution/**
CONTRIBUTING.md
defaults/**
examples/**
guides/**
install.sh
install.ps1
LICENSE
profiles/**
runtime/**
systemd/**
templates/**
tools/**
# Fleet: only the framework-seeded fleet subtrees are framework-owned.
fleet/README.md
fleet/examples/**
fleet/profiles/**
fleet/roles/**
fleet/roster.schema.json
fleet/services/**
# The manifest itself is framework-owned.
framework-manifest.txt
[operator]
# Identity / user-seeded contract files — generated by the wizard or seeded
# once from defaults/, then owned by the operator. Never overwritten on upgrade.
SOUL.md
USER.md
TOOLS.md
# Local overlays (tighten-only) authored by the operator.
*.local.md
# Operator-owned trees the updater must never write over or prune.
agents/**
policy/**
memory/**
sources/**
credentials/**
# Secret-bearing operator file INSIDE the framework-owned tools/ subtree.
# Listed explicitly so the deny-wins rule carves it out of tools/**.
tools/_lib/credentials.json
# Operator-owned fleet state (roster SSOT, per-agent env, heartbeats, backlog,
# persona overrides). Losing these silently downgrades a running fleet (#791).
fleet/roster.yaml
fleet/roster.json
fleet/agents/**
# Runtime state, incl. the #797 Runtime Session Ledger at fleet/run/sessions/
# (events.ndjson journal + ledger.json projection). This carve-out is the
# mechanism that makes the ledger upgrade-safe: an upgrade that wiped it would
# defeat its reason to exist. The HARD GATE (test-upgrade-manifest-guard.sh)
# proves a populated ledger survives byte-identical + mtime-unchanged.
fleet/run/**
fleet/backlog/**
fleet/roles.local/**

View File

@@ -1,5 +1,10 @@
#!/usr/bin/env bash
set -euo pipefail
# -E (errtrace): the ERR trap must propagate INTO functions and command
# substitutions. Without it the `trap restore_snapshot ERR` set below is dead
# code for any failure inside sync_framework_keep() (its whole body runs in a
# function) — a mid-sync failure would abort with a half-written target and NO
# rollback (#791 B1). Keep -E first so every later function inherits the trap.
set -Eeuo pipefail
# ─── Mosaic Framework Installer ──────────────────────────────────────────────
#
@@ -19,32 +24,19 @@ SOURCE_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
TARGET_DIR="${MOSAIC_HOME:-$HOME/.config/mosaic}"
INSTALL_MODE="${MOSAIC_INSTALL_MODE:-prompt}"
# Files/dirs protected from rsync --delete during sync. NOTE: framework-owned
# entries (CONSTITUTION/AGENTS/STANDARDS) ARE re-applied afterward by
# reconcile_framework_files (overwrite + backup-once); the rest stay user-owned.
# User-created content in these paths survives rsync --delete.
#
# fleet/* — the framework SEEDS fleet/examples, fleet/roles, fleet/profiles, and
# fleet/roster.schema.json (synced normally — every fleet/roles/*.md role contract
# and fleet/profiles/*.yaml system-type profile lands automatically via this sync,
# so no per-file entry is needed; exact preserved roster paths are anchored to
# the top level only and do NOT shadow fleet/profiles/*.yaml). The user's
# own fleet files MUST
# survive `mosaic update` (which runs this sync automatically): the active
# rosters (`fleet/roster.yaml` and `fleet/roster.json`), per-agent env
# (`fleet/agents/`), heartbeat run dir (`fleet/run/`), and the Mosaic-native
# backlog-of-record store (`fleet/backlog/` — embedded PGlite data dir; see
# packages/mosaic/src/commands/fleet-backlog.ts). Without these, an update
# wipes the operator's fleet AND their backlog. Glob entries are honored by
# both the rsync path (`--exclude`) and the glob-aware cp fallback below.
#
# fleet/roles.local — the persona OVERRIDE layer (H4). Baseline personas in
# fleet/roles/ are reseeded normally on every update (delivering new baseline
# personas), so any local edit there would be clobbered. User customizations
# and user-ADDED personas instead live in fleet/roles.local/ and MUST survive
# `mosaic update` — they win over the baseline on merge (AC-NS-7; see
# packages/mosaic/src/commands/fleet-personas.ts).
PRESERVE_PATHS=("CONSTITUTION.md" "AGENTS.md" "SOUL.md" "USER.md" "TOOLS.md" "STANDARDS.md" "memory" "sources" "credentials" "fleet/roster.yaml" "fleet/roster.json" "fleet/agents" "fleet/run" "fleet/backlog" "fleet/roles.local")
# Shared framework path-ownership manifest reader (#791). Parity with
# packages/mosaic/src/framework/manifest.ts — both consume framework-manifest.txt.
# Sourcing does not run its CLI dispatch (guarded by BASH_SOURCE==$0).
# shellcheck source=tools/_lib/manifest.sh
source "$SOURCE_DIR/tools/_lib/manifest.sh"
# Which paths a keep-mode upgrade may touch is no longer a hand-maintained
# denylist. It is derived from the shared framework-manifest.txt (#791): the
# updater only ever creates/overwrites framework-owned paths and only prunes a
# retired framework file inside a shipped framework subtree. Everything else —
# every operator file, and every path the manifest never anticipated — is
# operator-owned by default (fail-safe) and is never written or deleted. See
# sync_framework_keep() below and packages/mosaic/src/framework/manifest.ts.
# Framework-owned contract files: re-copied from defaults/ on every upgrade (the
# user must not edit them; a divergent copy is backed up once before overwrite).
@@ -75,14 +67,45 @@ step() { echo -e "\n${BOLD}$1${RESET}"; }
SNAPSHOT_DIR=""
make_snapshot() {
is_existing_install || return 0
# mktemp -d creates the dir 0700 — the snapshot (which mirrors operator config,
# possibly including secrets) is never world-readable.
SNAPSHOT_DIR="$(mktemp -d "${TMPDIR:-/tmp}/mosaic-snapshot-XXXXXX")"
cp -a "$TARGET_DIR/." "$SNAPSHOT_DIR/" 2>/dev/null || true
# The snapshot MUST be complete: restore rebuilds the target from it, so a
# partial capture (unreadable file, disk-full, I/O error) would silently
# discard whatever it missed. If cp -a cannot copy the whole tree, abort NOW —
# before the restore trap is armed and before anything is mutated. Fail closed
# rather than proceed with a snapshot we cannot trust (#791 blocker-2).
if ! cp -a "$TARGET_DIR/." "$SNAPSHOT_DIR/"; then
fail "Could not capture a complete pre-upgrade snapshot of $TARGET_DIR — aborting before any changes were made (fail-closed)."
rm -rf "$SNAPSHOT_DIR"; SNAPSHOT_DIR=""
exit 1
fi
}
restore_snapshot() {
# Disarm the trap first: restore runs under `set -e`, and a non-zero step
# inside it must not re-enter this handler (errtrace makes ERR fire in
# functions now). One restore attempt, then let the script exit non-zero.
trap - ERR INT TERM
[[ -n "$SNAPSHOT_DIR" && -d "$SNAPSHOT_DIR" ]] || return 0
fail "Install interrupted/failed — restoring previous state from snapshot"
rm -rf "$TARGET_DIR"; mkdir -p "$TARGET_DIR"
cp -a "$SNAPSHOT_DIR/." "$TARGET_DIR/" 2>/dev/null || true
# Reset the target before rebuilding from the snapshot — but CHECK it. Under
# `set -e` (trap already disarmed) a bare `rm -rf; mkdir -p` that fails would
# exit the whole script immediately, after `rm` may have deleted part of the
# target, WITHOUT ever printing the recovery pointer below — the operator would
# be left with a half-removed target and no idea the snapshot survives in /tmp.
# Test the reset explicitly (like the cp -a below), and on failure keep the
# snapshot and tell the operator where it is (#791 blocker-D2).
if ! rm -rf "$TARGET_DIR" || ! mkdir -p "$TARGET_DIR"; then
fail "Snapshot restore could not reset $TARGET_DIR. Your previous configuration is preserved at: $SNAPSHOT_DIR — copy it back into $TARGET_DIR manually."
return 1
fi
# Surface an incomplete restore instead of swallowing it: the snapshot is the
# last good copy, so if cp cannot fully rebuild the target we must NOT delete
# the snapshot — point the operator at it for manual recovery (#791 blocker-2).
if ! cp -a "$SNAPSHOT_DIR/." "$TARGET_DIR/"; then
fail "Snapshot restore did not complete cleanly. Your previous configuration is preserved at: $SNAPSHOT_DIR — copy it back into $TARGET_DIR manually."
return 1
fi
}
cleanup_snapshot() { [[ -n "$SNAPSHOT_DIR" && -d "$SNAPSHOT_DIR" ]] && rm -rf "$SNAPSHOT_DIR"; SNAPSHOT_DIR=""; }
@@ -184,63 +207,105 @@ sync_framework() {
return
fi
if command -v rsync >/dev/null 2>&1; then
local rsync_args=(-a --delete --exclude ".git" --exclude ".framework-version" --exclude "*.pre-constitution.bak")
if [[ "$INSTALL_MODE" == "keep" ]]; then
# Anchor to the transfer root (leading /) so we preserve the TOP-LEVEL
# ~/.config/mosaic/<file> without also excluding defaults/<file> from sync
# (reconcile_framework_files needs the freshly-synced defaults/ copies).
for path in "${PRESERVE_PATHS[@]}"; do
rsync_args+=(--exclude "/$path")
done
fi
rsync "${rsync_args[@]}" "$SOURCE_DIR/" "$TARGET_DIR/"
if [[ "$INSTALL_MODE" == "keep" ]]; then
# The `mosaic update` path. Manifest-driven, never-deleting-outside-framework:
# operator config is structurally protected (#791). No rsync --delete here.
# The manifest is already loaded+validated in main() BEFORE the snapshot/trap
# (a fail-closed manifest must abort without ever restoring over operator
# files — see the pre-flight in main, #791 blocker-1).
sync_framework_keep
return
fi
# Fallback: cp-based sync. Exact top-level preserved paths mirror the
# root-anchored rsync excludes above.
local preserve_tmp=""
if [[ "$INSTALL_MODE" == "keep" ]]; then
preserve_tmp="$(mktemp -d "${TMPDIR:-/tmp}/mosaic-preserve-XXXXXX")"
local match rel
for path in "${PRESERVE_PATHS[@]}"; do
# Unquoted $path lets the glob expand against TARGET_DIR; nullglob makes a
# non-matching pattern vanish instead of staying literal.
shopt -s nullglob
for match in "$TARGET_DIR/"$path; do
[[ -e "$match" ]] || continue
rel="${match#"$TARGET_DIR/"}"
mkdir -p "$preserve_tmp/$(dirname "$rel")"
cp -R "$match" "$preserve_tmp/$rel"
done
shopt -u nullglob
done
fi
# overwrite mode — a full replace, chosen only for a fresh install or when the
# operator explicitly asks to replace everything. No operator state to protect.
sync_framework_overwrite
}
find "$TARGET_DIR" -mindepth 1 -maxdepth 1 ! -name ".git" ! -name ".framework-version" ! -name "*.pre-constitution.bak" -exec rm -rf {} +
# Enumerate a NUL-delimited file list via `find` into the temp file $1, failing
# CLOSED if find errors. We capture to a checked file instead of consuming
# `< <(find …)` directly because a process substitution discards the producer's
# exit status: an EACCES/I/O failure partway through a scan would truncate the
# list yet leave the reading `while` loop exiting 0, so a partial upgrade would
# commit and report success and the ERR/restore trap would never fire. Running
# find to completion first, then checking its status, turns that silent
# truncation into a fail-closed abort that the restore trap can act on (#791
# blocker-D1). $1 after the shift is the scan root — named in the error.
_scan_or_die() {
local out="$1"; shift
if ! find "$@" -print0 > "$out"; then
fail "Could not enumerate framework files under '$1' — aborting before committing an incomplete sync (fail-closed)."
return 1 # D1-GUARD
fi
}
# Keep-mode sync: create/refresh framework-owned files and prune only retired
# framework files inside shipped framework subtrees. Operator-owned and unknown
# paths (fail-safe default) are never written and never deleted — the #791 HARD
# GATE. Single code path (no rsync) so it is byte-for-byte parity-testable.
sync_framework_keep() {
local src="$SOURCE_DIR" dst="$TARGET_DIR" abs rel root list
# 1) Overlay copy — every framework-owned source file, refreshed only when its
# bytes changed (no mtime churn on unchanged files, never on operator files).
# The source scan is captured fail-closed (#791 blocker-D1): a find failure
# aborts the sync (→ ERR trap → restore) rather than silently truncating it.
list="$(mktemp)"
_scan_or_die "$list" "$src" -type f || { rm -f "$list"; return 1; }
while IFS= read -r -d '' abs; do
rel="${abs#"$src"/}"
case "$rel" in
.git|.git/*|.framework-version|*.pre-constitution.bak) continue ;;
esac
manifest_is_framework "$rel" || continue
if [[ -f "$dst/$rel" ]] && cmp -s "$abs" "$dst/$rel"; then continue; fi
[[ "$rel" == */* ]] && mkdir -p "$dst/${rel%/*}"
cp "$abs" "$dst/$rel"
done < "$list"
rm -f "$list"
# 2) Scoped prune — within each shipped framework subtree root, remove
# framework-owned target files the current source no longer ships. Operator
# carve-outs (e.g. tools/_lib/credentials.json) resolve to operator and are
# skipped; unknown paths resolve to operator too — both are unreachable here.
# Each subtree scan is captured fail-closed for the same reason as the copy.
while IFS= read -r root; do
[[ -n "$root" && -d "$dst/$root" ]] || continue
list="$(mktemp)"
_scan_or_die "$list" "$dst/$root" -type f || { rm -f "$list"; return 1; }
while IFS= read -r -d '' abs; do
rel="${abs#"$dst"/}"
case "$rel" in *.pre-constitution.bak) continue ;; esac
[[ -f "$src/$rel" ]] && continue # still shipped
manifest_is_framework "$rel" || continue
rm -f "$abs"
done < "$list"
rm -f "$list"
# Drop framework dirs left empty by the prune (never touches a dir that still
# holds an operator file — those are never emptied). A genuine find failure
# (unreadable dir) is surfaced as a warning rather than silently swallowed;
# the "directory not empty" races we tolerate are ignored via -delete's own
# rc, not by hiding stderr — so a real error is still visible to the operator.
if ! find "$dst/$root" -type d -empty -delete 2>/dev/null; then
warn "prune: could not fully sweep empty framework dirs under $root (left as-is)"
fi
done < <(manifest_subtree_roots)
}
# Overwrite-mode sync: full replace. Only reached for a fresh install or an
# explicit operator "replace everything" choice, so nothing is preserved.
sync_framework_overwrite() {
if command -v rsync >/dev/null 2>&1; then
rsync -a --delete \
--exclude ".git" --exclude ".framework-version" --exclude "*.pre-constitution.bak" \
"$SOURCE_DIR/" "$TARGET_DIR/"
return
fi
find "$TARGET_DIR" -mindepth 1 -maxdepth 1 \
! -name ".git" ! -name ".framework-version" ! -name "*.pre-constitution.bak" \
-exec rm -rf {} +
cp -R "$SOURCE_DIR"/. "$TARGET_DIR"/
rm -rf "$TARGET_DIR/.git"
if [[ -n "$preserve_tmp" ]]; then
# Restore by re-globbing the SAME patterns against preserve_tmp, so each
# preserved item is restored at its own relative path (e.g. only
# fleet/roster.yaml is replaced — the freshly-synced fleet/examples stays).
for path in "${PRESERVE_PATHS[@]}"; do
shopt -s nullglob
for match in "$preserve_tmp/"$path; do
[[ -e "$match" ]] || continue
rel="${match#"$preserve_tmp/"}"
rm -rf "$TARGET_DIR/$rel"
mkdir -p "$TARGET_DIR/$(dirname "$rel")"
cp -R "$match" "$TARGET_DIR/$rel"
done
shopt -u nullglob
done
rm -rf "$preserve_tmp"
fi
}
# ═══════════════════════════════════════════════════════════════════════════════
@@ -311,9 +376,23 @@ else
ok "Install mode: overwrite"
fi
# Pre-flight (keep mode): load + validate the framework manifest BEFORE taking a
# snapshot or arming the restore trap. A fail-closed manifest (missing / empty /
# malformed) must abort here WITHOUT deleting or restoring over operator files —
# the snapshot/restore path exists only for a genuine mid-sync mutation failure,
# not for a validation failure that has touched nothing yet (#791 blocker-1).
if [[ "$INSTALL_MODE" == "keep" ]]; then
manifest_load
fi
# Snapshot before any destructive file operation; restore on interrupt/failure.
# The trap MUST exit after restoring: a bash INT/TERM handler that merely returns
# does NOT terminate the script — execution would resume past the interrupt,
# clear the snapshot, and report success, leaving a partial post-interrupt update
# (#791 blocker-A). `restore_snapshot; exit 1` guarantees a non-zero exit for
# both the errtrace (ERR) and signal (INT/TERM) paths.
make_snapshot
trap 'restore_snapshot' ERR INT TERM
trap 'restore_snapshot; exit 1' ERR INT TERM
sync_framework

View File

@@ -0,0 +1,253 @@
#!/usr/bin/env bash
# Shared bash reader for framework-manifest.txt (#791).
#
# This is the bash half of the SSOT ownership resolver; the TypeScript half is
# packages/mosaic/src/framework/manifest.ts. BOTH read the same
# framework-manifest.txt and MUST resolve identical ownership for any path — the
# parity test (manifest-parity.spec.ts) invokes this file's `resolve` CLI and
# compares it against the TS resolver, so the two can never drift (the #631
# two-copies failure class this closes).
#
# Ownership resolution (deny-wins / fail-safe):
# 1. operator glob matches -> operator
# 2. else framework glob -> framework
# 3. else -> operator (UNKNOWN defaults to operator, #791)
#
# Globs are compiled once at load into exact-prefix checks or POSIX EREs, so the
# hot resolver (manifest_is_framework) forks no subprocesses — the installer
# calls it once per file across the whole tree.
#
# Usage as a library (source it, then):
# manifest_load [manifest-file] # populates + compiles the manifest
# manifest_is_framework <rel-path> # rc 0 = framework-owned, rc 1 = operator
# manifest_resolve <rel-path> # echoes: framework | operator
# manifest_subtree_roots # echoes shipped framework `dir/**` roots
#
# Usage as a CLI (parity harness):
# bash manifest.sh resolve <rel-path>
# bash manifest.sh subtree-roots
# bash manifest.sh classify # reads paths on stdin -> "<own>\t<path>"
MANIFEST_FRAMEWORK=()
MANIFEST_OPERATOR=()
# Compiled forms (parallel arrays). _*_KIND[i] is "exact" or "re".
_MF_KIND=(); _MF_EXACT=(); _MF_RE=()
_MO_KIND=(); _MO_EXACT=(); _MO_RE=()
_MF_ROOTS=()
_manifest_default_root() { cd "$(dirname "${BASH_SOURCE[0]}")/../.." && pwd; }
# Normalize a path/glob: backslashes -> slashes, strip leading ./ and /, strip
# trailing / (mirrors normalizeRel in manifest.ts).
_manifest_norm() {
local p="$1"
p="${p//\\//}"
p="${p#./}"
while [[ "$p" == /* ]]; do p="${p#/}"; done
while [[ "$p" == */ ]]; do p="${p%/}"; done
printf '%s' "$p"
}
# Translate a normalized glob into a POSIX ERE body (mirrors globToRegExpBody).
_manifest_glob_to_ere() {
local pattern; pattern="$(_manifest_norm "$1")"
local out="" c n i len=${#pattern} trailing
for (( i = 0; i < len; i++ )); do
c="${pattern:i:1}"
if [[ "$c" == "*" ]]; then
n="${pattern:i+1:1}"
if [[ "$n" == "*" ]]; then
i=$((i + 1))
trailing=0
if [[ "${pattern:i+1:1}" == "/" ]]; then i=$((i + 1)); trailing=1; fi
if [[ "$out" == */ ]]; then
out="${out%/}(/.*)?"
elif [[ "$trailing" -eq 1 ]]; then
out="$out(.*/)?"
else
out="$out.*"
fi
else
out="$out[^/]*"
fi
else
case "$c" in
.|+|\?|^|\$|\{|\}|\(|\)|\||\[|\]|\\) out="$out\\$c" ;;
*) out="$out$c" ;;
esac
fi
done
printf '%s' "$out"
}
# Compile one raw glob into (kind, exact, re) appended to the given section.
# $1 = raw glob, $2 = section letter (F|O).
_manifest_compile_one() {
local norm; norm="$(_manifest_norm "$1")"
[[ -n "$norm" ]] || return 0
if [[ "$norm" == *"*"* ]]; then
local re="^$(_manifest_glob_to_ere "$norm")\$"
if [[ "$2" == F ]]; then
_MF_KIND+=(re); _MF_EXACT+=(""); _MF_RE+=("$re")
else
_MO_KIND+=(re); _MO_EXACT+=(""); _MO_RE+=("$re")
fi
else
if [[ "$2" == F ]]; then
_MF_KIND+=(exact); _MF_EXACT+=("$norm"); _MF_RE+=("")
else
_MO_KIND+=(exact); _MO_EXACT+=("$norm"); _MO_RE+=("")
fi
fi
[[ "$2" == F && "$norm" == */"**" ]] && _MF_ROOTS+=("${norm%/**}")
return 0
}
_manifest_compile() {
_MF_KIND=(); _MF_EXACT=(); _MF_RE=(); _MF_ROOTS=()
_MO_KIND=(); _MO_EXACT=(); _MO_RE=()
local g
for g in "${MANIFEST_FRAMEWORK[@]:-}"; do [[ -n "$g" ]] && _manifest_compile_one "$g" F; done
for g in "${MANIFEST_OPERATOR[@]:-}"; do [[ -n "$g" ]] && _manifest_compile_one "$g" O; done
# Explicit success: an empty operator array makes the final `[[ -n "" ]] && …`
# short-circuit to rc 1, which would otherwise become this function's (and
# manifest_load's) return code — a spurious failure (#791 B2). Never rely on
# the last loop's exit status here.
return 0
}
# Load + compile the manifest. Rejects a malformed file the same way
# parseManifest() does (entry before a section header / unknown header).
manifest_load() {
local file="${1:-}"
[[ -n "$file" ]] || file="$(_manifest_default_root)/framework-manifest.txt"
# Fail CLOSED on a missing/unreadable manifest. Without this, `done < "$file"`
# aborts on a raw redirection error with no explanation; downstream that reads
# as "no framework paths" and an upgrade could no-op silently (#791 B2/B3).
if [[ ! -r "$file" ]]; then
echo "manifest: cannot read manifest file: $file — refusing to sync (fail-closed)." >&2
return 1
fi
MANIFEST_FRAMEWORK=()
MANIFEST_OPERATOR=()
local section="" line
while IFS= read -r line || [[ -n "$line" ]]; do
line="${line#"${line%%[![:space:]]*}"}" # ltrim
line="${line%"${line##*[![:space:]]}"}" # rtrim
[[ -z "$line" || "${line:0:1}" == "#" ]] && continue
case "$line" in
"[framework]") section=framework; continue ;;
"[operator]") section=operator; continue ;;
"["*) echo "manifest: unknown section header: $line" >&2; return 1 ;;
esac
if [[ -z "$section" ]]; then
echo "manifest: entry before any [section] header: $line" >&2
return 1
fi
if [[ "$section" == framework ]]; then
MANIFEST_FRAMEWORK+=("$line")
else
MANIFEST_OPERATOR+=("$line")
fi
done < "$file"
# An empty or comment-only manifest defines NO framework-owned paths. Treating
# that as valid would make every path resolve operator and an upgrade prune
# nothing / write nothing — a silent no-op indistinguishable from success.
# Fail loud instead, mirroring parseManifest()'s throw in manifest.ts (#791 B2).
if [[ ${#MANIFEST_FRAMEWORK[@]} -eq 0 ]]; then
echo "manifest: no [framework] entries in $file — refusing to sync (empty or malformed manifest)." >&2
return 1
fi
# An entry like `/` or `./` normalizes to nothing and compiles to a glob that
# matches no path — so a manifest whose only [framework] entries are degenerate
# passes the count guard above but leaves the framework matcher empty: every
# path resolves operator, the exact silent no-op we fail closed against. Require
# at least one entry with a real (non-slash, non-dot) character. Mirrors
# parseManifest()'s `isUsableFrameworkGlob` `/[^/.]/` test in manifest.ts (#791 blocker-B).
local _g _usable=0
for _g in "${MANIFEST_FRAMEWORK[@]:-}"; do
if [[ "$(_manifest_norm "$_g")" =~ [^/.] ]]; then _usable=1; break; fi
done
if [[ "$_usable" -eq 0 ]]; then
echo "manifest: no usable [framework] entries in $file (every entry is empty or a bare dot segment) — refusing to sync (malformed manifest)." >&2
return 1
fi
_manifest_compile
return 0
}
# Fork-free: does $1 (a mosaic-home-relative path) match an operator glob?
_mo_matches() {
local path="$1" i n=${#_MO_KIND[@]} re pat
for (( i = 0; i < n; i++ )); do
if [[ "${_MO_KIND[i]}" == exact ]]; then
pat="${_MO_EXACT[i]}"
[[ "$path" == "$pat" || "$path" == "$pat/"* ]] && return 0
else
re="${_MO_RE[i]}"
[[ "$path" =~ $re ]] && return 0
fi
done
return 1
}
# Fork-free: does $1 match a framework glob?
_mf_matches() {
local path="$1" i n=${#_MF_KIND[@]} re pat
for (( i = 0; i < n; i++ )); do
if [[ "${_MF_KIND[i]}" == exact ]]; then
pat="${_MF_EXACT[i]}"
[[ "$path" == "$pat" || "$path" == "$pat/"* ]] && return 0
else
re="${_MF_RE[i]}"
[[ "$path" =~ $re ]] && return 0
fi
done
return 1
}
# The installer hot path — no subshell. rc 0 = framework-owned, rc 1 = operator
# (deny-wins / fail-safe). Assumes an already-clean POSIX relative path.
manifest_is_framework() {
_mo_matches "$1" && return 1
_mf_matches "$1" && return 0
return 1
}
# Echo the ownership of a path: framework | operator. Normalizes first, so it is
# safe for CLI / test callers passing unnormalized input.
manifest_resolve() {
local path; path="$(_manifest_norm "$1")"
if manifest_is_framework "$path"; then echo framework; else echo operator; fi
}
# Echo each shipped framework subtree root (a `dir/**` entry, without the /**).
manifest_subtree_roots() {
local r
for r in "${_MF_ROOTS[@]:-}"; do [[ -n "$r" ]] && printf '%s\n' "$r"; done
}
# CLI dispatch — only when executed directly, never when sourced.
if [[ "${BASH_SOURCE[0]}" == "${0}" ]]; then
set -o pipefail
# Propagate a fail-closed manifest_load (missing/empty/malformed) as a non-zero
# exit instead of continuing to resolve against empty compiled arrays — that is
# what lets the parity test assert bash and TS reject the same bad inputs (#791 B2).
manifest_load "${MANIFEST_FILE:-}" || exit 1
cmd="${1:-}"
case "$cmd" in
resolve) manifest_resolve "${2:?path required}" ;;
subtree-roots) manifest_subtree_roots ;;
classify)
while IFS= read -r p; do
[[ -z "$p" ]] && continue
printf '%s\t%s\n' "$(manifest_resolve "$p")" "$p"
done
;;
*)
echo "usage: manifest.sh {resolve <path>|subtree-roots|classify}" >&2
exit 2
;;
esac
fi

View File

@@ -61,8 +61,10 @@ MOSAIC_HOME="$T5" MOSAIC_INSTALL_MODE=bogus MOSAIC_SYNC_ONLY=1 bash "$INSTALL" >
chk "F5 failure: invalid mode rejected (nonzero exit)" "[ $rc -ne 0 ]"
chk "F5 failure: SOUL + credentials intact" "grep -q orig '$T5/SOUL.md' && grep -q keepme '$T5/credentials/c.json'"
# F6 — keep-mode re-seed (the `mosaic update` path) MUST preserve only the
# exact user-owned roster paths while refreshing framework-owned schema/examples.
# F6 — keep-mode re-seed (the `mosaic update` path) MUST preserve ALL user-owned
# fleet state — including an unanticipated file the manifest never names, which
# resolves to operator-owned by the #791 fail-safe — while refreshing the
# framework-owned schema/examples.
T6=$(mktemp -d); mkdir -p "$T6/fleet/examples" "$T6/fleet/run" "$T6/fleet/agents"
printf '# persona\n' > "$T6/SOUL.md" # makes it a recognized existing install (→ keep mode)
printf 'version: 1\nagents:\n - name: coder0\n' > "$T6/fleet/roster.yaml"
@@ -75,13 +77,14 @@ printf '{"stale":true}\n' > "$T6/fleet/roster.schema.json"
E6=$(mktemp -d)
cp "$T6/fleet/roster.yaml" "$E6/roster-yaml.expected"
cp "$T6/fleet/roster.json" "$E6/roster-json.expected"
cp "$T6/fleet/my-fleet.yaml" "$E6/my-fleet.expected"
cp "$T6/fleet/run/coder0.hb" "$E6/run.expected"
cp "$T6/fleet/agents/coder0.env" "$E6/agent.expected"
echo 3 > "$T6/.framework-version"
run "$T6" keep
chk "F6 reseed: exact roster.yaml bytes survive keep-mode sync" "cmp -s '$T6/fleet/roster.yaml' '$E6/roster-yaml.expected'"
chk "F6 reseed: exact roster.json bytes survive keep-mode sync" "cmp -s '$T6/fleet/roster.json' '$E6/roster-json.expected'"
chk "F6 reseed: unrelated fleet YAML is not preserved" "[ ! -f '$T6/fleet/my-fleet.yaml' ]"
chk "F6 reseed: unanticipated operator fleet file survives (fail-safe, #791)" "cmp -s '$T6/fleet/my-fleet.yaml' '$E6/my-fleet.expected'"
chk "F6 reseed: per-agent env bytes survive" "cmp -s '$T6/fleet/agents/coder0.env' '$E6/agent.expected'"
chk "F6 reseed: heartbeat bytes survive" "cmp -s '$T6/fleet/run/coder0.hb' '$E6/run.expected'"
chk "F6 reseed: framework examples are refreshed" "grep -q orchestrator '$T6/fleet/examples/general.yaml'"

View File

@@ -0,0 +1,224 @@
#!/usr/bin/env bash
# test-upgrade-manifest-guard.sh — the #791 HARD GATE.
#
# Proves that a keep-mode framework upgrade (the `mosaic update` path:
# install.sh with MOSAIC_INSTALL_MODE=keep MOSAIC_SYNC_ONLY=1) touches NO path
# outside the framework-owned manifest. Every operator-owned sentinel — including
# a deliberately UNANTICIPATED one the manifest never names — must survive
# byte-identical with an unchanged mtime (not even rewritten). Framework files
# must still update, and a retired framework file inside a shipped subtree must
# still be pruned. No operator secret value may appear in installer output.
#
# Keep mode is a SINGLE code path (sync_framework_keep, a manifest-driven cp
# overlay + scoped prune — no rsync). The matrix still runs twice, once with
# rsync on PATH and once with it hidden, to prove the keep path is genuinely
# rsync-independent: it must obey the manifest identically whether or not rsync
# happens to be installed (rsync --delete is only ever used by overwrite mode,
# which has no operator state to protect).
#
# It also runs a fail-closed matrix (#791 B2/B3): an empty, operator-only,
# malformed, or missing manifest must ABORT the upgrade loudly and leave every
# operator path untouched — never silently no-op to "complete".
#
# Usage: bash test-upgrade-manifest-guard.sh
set -uo pipefail
FW="$(cd "$(dirname "${BASH_SOURCE[0]}")/../../.." && pwd)" # packages/mosaic/framework
INSTALL="$FW/install.sh"
pass=0; fail=0
chk() { if eval "$2"; then echo "$1"; pass=$((pass + 1)); else echo "$1"; fail=$((fail + 1)); fi; }
SECRET='SUPER-SECRET-TOKEN-do-not-log-3f9a'
# Seed a throwaway MOSAIC_HOME with an operator sentinel per ownership class.
seed_home() {
local H="$1"
mkdir -p "$H/agents" "$H/policy" "$H/memory" "$H/tools/_lib" \
"$H/fleet/agents" "$H/fleet/run/sessions" "$H/harvester" \
"$H/unknown-operator-dir" "$H/guides"
printf '# persona\n' > "$H/SOUL.md" # marks a recognized existing install → keep mode
printf 'MODEL=opus\n' > "$H/agents/coder0.conf"
printf '# operator policy\n' > "$H/policy/custom.md"
printf '# soul overlay\n' > "$H/SOUL.local.md"
printf '# operator memory\n' > "$H/memory/note.md"
printf 'TOKEN=%s\n' "$SECRET" > "$H/tools/_lib/credentials.json"
printf 'MOSAIC_AGENT_NAME=coder0\n' > "$H/fleet/agents/coder0.env"
printf 'version: 2\nagents:\n - name: coder0\n' > "$H/fleet/roster.yaml"
printf '# harvester SOP\n' > "$H/harvester/sop.md"
printf 'operator data the manifest never anticipated\n' > "$H/unknown-operator-dir/x"
printf 'version: 1\nagents:\n - name: mine\n' > "$H/fleet/my-fleet.yaml"
# #797 Runtime Session Ledger (Mos-elevated to a #791 PR1 merge-blocker): a
# populated ledger under fleet/run/sessions/ must survive the upgrade — a
# runtime ledger an upgrade rsync can wipe is worthless. Seed it exactly as
# #797 writes it: a non-empty append journal + a non-empty compacted
# projection, files 0600 under a 0700 dir.
printf '%s\n%s\n%s\n' \
'{"seq":1,"kind":"session.spawn","node":"sess-42","generation":7}' \
'{"seq":2,"kind":"lease.grant","node":"sess-42","lease":"web1"}' \
'{"seq":3,"kind":"dispatch.create","from":"sess-42","to":"disp-9"}' \
> "$H/fleet/run/sessions/events.ndjson"
printf '%s\n' \
'{"generation":7,"nodes":[{"id":"sess-42","kind":"session"}],"edges":[{"from":"sess-42","to":"disp-9","kind":"dispatch"}]}' \
> "$H/fleet/run/sessions/ledger.json"
chmod 0700 "$H/fleet/run" "$H/fleet/run/sessions"
chmod 0600 "$H/fleet/run/sessions/events.ndjson" "$H/fleet/run/sessions/ledger.json"
# A retired framework file inside a shipped subtree (absent from source) — must be pruned.
printf '# retired guide\n' > "$H/guides/RETIRED-OLD-GUIDE.md"
echo 3 > "$H/.framework-version"
}
OPERATOR_SENTINELS=(
"agents/coder0.conf"
"policy/custom.md"
"SOUL.local.md"
"memory/note.md"
"tools/_lib/credentials.json"
"fleet/agents/coder0.env"
"fleet/roster.yaml"
"harvester/sop.md"
"unknown-operator-dir/x"
"fleet/my-fleet.yaml"
# #797 Runtime Session Ledger — populated journal + projection must survive.
"fleet/run/sessions/events.ndjson"
"fleet/run/sessions/ledger.json"
)
run_matrix() {
local label="$1"; shift # extra env / PATH override applied to the run
local H E OUT rel before_hash after_hash before_mt after_mt
H=$(mktemp -d); E=$(mktemp -d); OUT=$(mktemp)
seed_home "$H"
# Snapshot hash + mtime of every operator sentinel before the upgrade.
for rel in "${OPERATOR_SENTINELS[@]}"; do
sha256sum "$H/$rel" | awk '{print $1}' > "$E/$(echo "$rel" | tr / _).hash"
stat -c %Y "$H/$rel" > "$E/$(echo "$rel" | tr / _).mt"
done
# Snapshot the ledger directory permission bits (#797 assert: perms unchanged).
local before_dirperm after_dirperm
before_dirperm=$(stat -c %a "$H/fleet/run/sessions")
# The upgrade under test (keep + sync-only = the `mosaic update` reseed path).
MOSAIC_HOME="$H" MOSAIC_INSTALL_MODE=keep MOSAIC_SYNC_ONLY=1 "$@" bash "$INSTALL" >"$OUT" 2>&1
# HARD GATE: every operator sentinel survives byte-identical AND mtime-unchanged.
for rel in "${OPERATOR_SENTINELS[@]}"; do
before_hash=$(cat "$E/$(echo "$rel" | tr / _).hash")
before_mt=$(cat "$E/$(echo "$rel" | tr / _).mt")
after_hash=$(sha256sum "$H/$rel" 2>/dev/null | awk '{print $1}')
after_mt=$(stat -c %Y "$H/$rel" 2>/dev/null || echo MISSING)
chk "[$label] operator sentinel survives byte-identical: $rel" \
"[ -n '$after_hash' ] && [ '$before_hash' = '$after_hash' ]"
chk "[$label] operator sentinel not rewritten (mtime unchanged): $rel" \
"[ '$before_mt' = '$after_mt' ]"
done
# #797 assert 7: the ledger directory's permission bits are unchanged.
after_dirperm=$(stat -c %a "$H/fleet/run/sessions" 2>/dev/null || echo MISSING)
chk "[$label] ledger dir perms unchanged (#797): $before_dirperm" \
"[ '$before_dirperm' = '$after_dirperm' ]"
# Positive controls / negative controls — prove the test discriminates: the
# upgrade DOES write and prune framework-owned paths, so the operator sentinels
# (incl. the #797 ledger) survive because of the manifest, not because the
# upgrade is a no-op.
chk "[$label] positive control: framework file present after upgrade (guides synced)" \
"[ -f '$H/guides/E2E-DELIVERY.md' ]"
chk "[$label] negative control: retired framework file inside a subtree IS pruned" \
"[ ! -f '$H/guides/RETIRED-OLD-GUIDE.md' ]"
chk "[$label] manifest itself is installed" "[ -f '$H/framework-manifest.txt' ]"
# Secret-safety: the operator secret value never appears in installer output.
chk "[$label] operator secret value absent from installer stdout/stderr" \
"! grep -q '$SECRET' '$OUT'"
rm -rf "$H" "$E" "$OUT"
}
# Fail-closed matrix (#791 B2/B3 + blocker-1): run install.sh from a COPY of the
# framework so the shipped manifest can be corrupted. Every corruption must abort
# the upgrade non-zero with a manifest error, leaving all operator sentinels
# byte-identical AND on the SAME inode. The inode check is the load-bearing part:
# manifest validation is hoisted BEFORE make_snapshot/the restore trap, so a bad
# manifest must abort without ever snapshotting, deleting, and restoring the
# target. Were validation still armed under the ERR trap, restore_snapshot would
# rm -rf + rebuild the target — same bytes but a NEW inode (broken hard links,
# changed ctime), which a content-only hash would miss (#791 blocker-1).
run_failclosed() {
local label="$1" mutate="$2"
local SRC H E OUT rc rel before_hash after_hash before_ino after_ino key
SRC=$(mktemp -d); H=$(mktemp -d); E=$(mktemp -d); OUT=$(mktemp)
cp -a "$FW/." "$SRC/"
case "$mutate" in
empty) : > "$SRC/framework-manifest.txt" ;;
operator-only) printf '[operator]\nSOUL.md\n*.local.md\n' > "$SRC/framework-manifest.txt" ;;
malformed) printf 'stray.md\n[framework]\nguides/**\n' > "$SRC/framework-manifest.txt" ;;
degenerate) printf '[framework]\n/\n./\n[operator]\nSOUL.md\n' > "$SRC/framework-manifest.txt" ;;
missing) rm -f "$SRC/framework-manifest.txt" ;;
esac
seed_home "$H"
for rel in "${OPERATOR_SENTINELS[@]}"; do
key=$(echo "$rel" | tr / _)
sha256sum "$H/$rel" | awk '{print $1}' > "$E/$key.hash"
stat -c '%i' "$H/$rel" > "$E/$key.ino"
done
MOSAIC_HOME="$H" MOSAIC_INSTALL_MODE=keep MOSAIC_SYNC_ONLY=1 bash "$SRC/install.sh" >"$OUT" 2>&1
rc=$?
chk "[fail-closed:$label] upgrade aborts non-zero" "[ '$rc' -ne 0 ]"
chk "[fail-closed:$label] refuses loudly with a manifest error" \
"grep -qi 'manifest' '$OUT'"
for rel in "${OPERATOR_SENTINELS[@]}"; do
key=$(echo "$rel" | tr / _)
before_hash=$(cat "$E/$key.hash")
after_hash=$(sha256sum "$H/$rel" 2>/dev/null | awk '{print $1}')
chk "[fail-closed:$label] operator sentinel untouched: $rel" \
"[ -n '$after_hash' ] && [ '$before_hash' = '$after_hash' ]"
before_ino=$(cat "$E/$key.ino")
after_ino=$(stat -c '%i' "$H/$rel" 2>/dev/null)
chk "[fail-closed:$label] operator sentinel not deleted/recreated (inode stable): $rel" \
"[ -n '$after_ino' ] && [ '$before_ino' = '$after_ino' ]"
done
chk "[fail-closed:$label] operator secret value absent from output" \
"! grep -q '$SECRET' '$OUT'"
chmod -R u+w "$SRC" "$H" 2>/dev/null || true
rm -rf "$SRC" "$H" "$E" "$OUT"
}
echo "#791 upgrade manifest guard (HARD GATE):"
# 1) rsync path (if available on this host).
if command -v rsync >/dev/null 2>&1; then
run_matrix "rsync"
else
echo " · rsync not installed — skipping rsync-path matrix"
fi
# 2) rsync-absent path — hide rsync behind a scratch PATH. Keep mode never calls
# rsync, so this must resolve identically to run (1); it proves the keep path
# does not silently depend on rsync being installed. (Provide the coreutils the
# installer needs on the stripped PATH.)
FBIN=$(mktemp -d)
for t in bash cp find mktemp rm mkdir chmod cmp sed grep cat dirname basename stat sha256sum awk tr; do
p=$(command -v "$t" 2>/dev/null) && ln -s "$p" "$FBIN/$t"
done
run_matrix "rsync-absent" env "PATH=$FBIN"
rm -rf "$FBIN"
# 3) fail-closed matrix (#791 B2/B3) — corrupt the shipped manifest four ways.
run_failclosed "empty-manifest" empty
run_failclosed "operator-only" operator-only
run_failclosed "malformed-manifest" malformed
run_failclosed "degenerate-framework" degenerate
run_failclosed "missing-manifest" missing
echo
echo "RESULT: $pass passed, $fail failed"
[ "$fail" -eq 0 ]

View File

@@ -0,0 +1,319 @@
#!/usr/bin/env bash
# test-upgrade-rollback.sh — the #791 B1 regression gate.
#
# A keep-mode upgrade takes a pre-update snapshot and installs an ERR/INT/TERM
# trap that restores it if the sync aborts midway (install.sh: make_snapshot +
# `trap restore_snapshot`). That trap is only reached if `set -E` (errtrace) is
# active — otherwise a failure INSIDE sync_framework_keep() (which runs entirely
# in a function) never fires the trap, and the upgrade aborts leaving a
# half-written target with NO rollback. This test proves:
#
# Part A (the gate): the shipped installer rolls back a mid-sync failure —
# the restore message fires, the corrupted file is put
# back, AND the whole target is byte-identical to its
# pre-upgrade state.
# Part B (the control): the SAME installer with `-E` stripped does NOT roll back
# (dead trap) — the mid-sync corruption survives, proving
# errtrace is load-bearing. If anyone removes `set -E`,
# Part A goes red.
#
# The mid-sync failure is injected with a PATH-shadowing `cp` shim rather than
# file permissions. The earlier 0400/EACCES approach was NOT portable: Woodpecker
# runs steps as root (node:24-alpine has no USER directive), and root overwrites a
# 0400 file, so the failure never fired and this gate silently passed (#791
# blocker-3). The shim fails deterministically for one framework-owned
# destination regardless of uid, and — like a real interrupted cp (disk-full
# mid-write) — leaves a partially-written target behind, so rollback has real
# damage to undo and the control has real damage to expose.
#
# Usage: bash test-upgrade-rollback.sh
set -uo pipefail
FW="$(cd "$(dirname "${BASH_SOURCE[0]}")/../../.." && pwd)" # packages/mosaic/framework
INSTALL="$FW/install.sh"
ORIG_PATH="$PATH"
# The `-E`-stripped control installer must live INSIDE $FW: install.sh derives
# SOURCE_DIR from its own path and `source`s $SOURCE_DIR/tools/_lib/manifest.sh,
# so a copy anywhere else aborts at the source line before ever reaching the sync
# loop — which would make the control a false negative. A root dotfile is
# operator-owned (unknown→operator), so the sync loop skips it. Clean up on exit.
STRIPPED="$FW/.install-rollback-control.tmp.sh"
NOEXIT="$FW/.install-noexit-control.tmp.sh"
D1CTRL="$FW/.install-d1guard-control.tmp.sh"
D2CTRL="$FW/.install-d2guard-control.tmp.sh"
rm -f "$STRIPPED" "$NOEXIT" "$D1CTRL" "$D2CTRL"
trap 'rm -f "$STRIPPED" "$NOEXIT" "$D1CTRL" "$D2CTRL"' EXIT
pass=0; fail=0
chk() { if eval "$2"; then echo "$1"; pass=$((pass + 1)); else echo "$1"; fail=$((fail + 1)); fi; }
SECRET='SUPER-SECRET-TOKEN-do-not-log-b1'
# A framework-owned file the shim fails the copy of. The seeded target holds GOOD
# bytes; source ships different bytes, so sync_framework_keep() attempts the copy
# and the shim intercepts it. Root-level framework files sort before guides/, so
# several framework files are already refreshed when the copy reaches this one.
POISON_REL='guides/E2E-DELIVERY.md'
GOOD='GOOD-REFERENCE-CONTENT-pre-upgrade-b1'
GARBAGE='PARTIAL-WRITE-GARBAGE-mid-sync-b1'
# A `cp` shim: for the poisoned destination, simulate an interrupted copy — write
# partial garbage to the target, then fail — otherwise delegate to the real cp
# (resolved via the ORIGINAL PATH so make_snapshot/restore still work).
make_cp_shim() {
local dir="$1"
cat > "$dir/cp" <<SHIM
#!/usr/bin/env bash
dest="\${@: -1}"
case "\$dest" in
*/$POISON_REL)
printf '%s' '$GARBAGE' > "\$dest" 2>/dev/null || true
exit 1 ;;
esac
exec env PATH="$ORIG_PATH" cp "\$@"
SHIM
chmod +x "$dir/cp"
}
# A `find` shim that fails every enumeration scan (`-print0`) as if it hit an
# EACCES/I/O error partway — it emits the real (here: complete) list first, then
# exits non-zero, exactly the class of failure a `< <(find …)` process
# substitution silently swallows. All non-`-print0` finds (e.g. the -delete
# sweep) delegate to the real find on the original PATH. Used to prove #791
# blocker-D1: the shipped installer must honor find's exit status and roll back.
make_find_fail_shim() {
local dir="$1"
cat > "$dir/find" <<SHIM
#!/usr/bin/env bash
for a in "\$@"; do
if [ "\$a" = "-print0" ]; then
env PATH="$ORIG_PATH" find "\$@" # emit the real list…
exit 1 # …then fail as if the scan hit EACCES
fi
done
exec env PATH="$ORIG_PATH" find "\$@"
SHIM
chmod +x "$dir/find"
}
# An `rm` shim that fails ONLY `rm -rf <FAIL_RM_TARGET>` (the restore's target
# reset) and delegates every other rm to the real one. Used to prove #791
# blocker-D2: when the target reset inside restore_snapshot fails, the installer
# must emit the manual-recovery pointer (snapshot path) instead of exiting
# silently under `set -e`. FAIL_RM_TARGET is exported into the installer env.
make_rm_fail_shim() {
local dir="$1"
cat > "$dir/rm" <<'SHIM'
#!/usr/bin/env bash
last="${@: -1}"
if [ -n "${FAIL_RM_TARGET:-}" ] && [ "$last" = "$FAIL_RM_TARGET" ]; then
exit 1
fi
exec env PATH="$ORIG_PATH_FOR_RM" rm "$@"
SHIM
chmod +x "$dir/rm"
}
seed_home() {
local H="$1"
mkdir -p "$H/agents" "$H/tools/_lib" "$H/memory" "$H/guides"
printf '# persona\n' > "$H/SOUL.md" # recognized install → keep mode + snapshot
printf 'MODEL=opus\n' > "$H/agents/coder0.conf"
printf '# operator memory\n' > "$H/memory/note.md"
printf 'TOKEN=%s\n' "$SECRET" > "$H/tools/_lib/credentials.json"
echo 3 > "$H/.framework-version"
# Good pre-upgrade bytes; source ships different bytes, so cp is attempted.
printf '%s' "$GOOD" > "$H/$POISON_REL"
}
# Run one keep-mode upgrade against $1=installer, seeding a fresh home and a
# byte-for-byte reference of the pre-upgrade state, with the cp shim first on
# PATH. Echoes: "<exit>\t<out>\t<ref>\t<home>".
run_upgrade() {
local installer="$1" shim_maker="${2:-make_cp_shim}" H REF OUT SHIM rc
H=$(mktemp -d); REF=$(mktemp -d); OUT=$(mktemp); SHIM=$(mktemp -d)
seed_home "$H"
env PATH="$ORIG_PATH" cp -a "$H/." "$REF/" # pre-upgrade reference (real cp)
"$shim_maker" "$SHIM"
set +e
PATH="$SHIM:$ORIG_PATH" \
MOSAIC_HOME="$H" MOSAIC_INSTALL_MODE=keep MOSAIC_SYNC_ONLY=1 bash "$installer" >"$OUT" 2>&1
rc=$?
set -e 2>/dev/null || true
rm -rf "$SHIM"
printf '%s\t%s\t%s\t%s\n' "$rc" "$OUT" "$REF" "$H"
}
echo "#791 upgrade rollback (B1 regression gate):"
# ── Part A: the shipped installer must roll back a mid-sync failure ───────────
IFS=$'\t' read -r rcA OUTA REFA HA < <(run_upgrade "$INSTALL")
chk "[shipped] upgrade aborts non-zero on the injected mid-sync failure" \
"[ '$rcA' -ne 0 ]"
chk "[shipped] restore_snapshot fires (rollback message present)" \
"grep -q 'restoring previous state from snapshot' '$OUTA'"
chk "[shipped] the corrupted file is restored to its pre-upgrade bytes" \
"[ \"\$(cat '$HA/$POISON_REL')\" = '$GOOD' ]"
chk "[shipped] target rolled back byte-identical to pre-upgrade state" \
"diff -r '$REFA' '$HA' >/dev/null 2>&1"
chk "[shipped] operator secret value absent from installer output" \
"! grep -q '$SECRET' '$OUTA'"
# ── Part B: control — strip `-E`, the trap is dead, no rollback happens ───────
# Proves errtrace is what makes the trap reachable. If `set -E` is ever removed
# from install.sh, Part A's rollback assertions fail exactly like this control.
sed 's/^set -Eeuo pipefail/set -euo pipefail/' "$INSTALL" > "$STRIPPED"
chk "[control] the -E strip actually changed the installer" \
"! cmp -s '$INSTALL' '$STRIPPED'"
IFS=$'\t' read -r rcB OUTB REFB HB < <(run_upgrade "$STRIPPED")
chk "[control] without -E the upgrade still aborts non-zero" \
"[ '$rcB' -ne 0 ]"
# The load-bearing, deterministic proof of B1: without errtrace the ERR trap
# never fires for a failure inside sync_framework_keep(), so no rollback runs.
chk "[control] without -E the rollback message does NOT fire (dead trap)" \
"! grep -q 'restoring previous state from snapshot' '$OUTB'"
chk "[control] without -E the mid-sync corruption survives (no rollback)" \
"[ \"\$(cat '$HB/$POISON_REL')\" = '$GARBAGE' ]"
# ── Part C: an INT/TERM interrupt must terminate, not resume (blocker-A) ──────
# A bash signal trap that merely returns lets the script continue past the
# interrupt — restoring the snapshot, then resuming the sync and reporting
# success. We inject a SIGTERM mid-sync with a cp that SUCCEEDS (so set -e never
# fires and ONLY the signal path governs), and assert the shipped installer
# restores AND exits without reporting success. The control strips `exit 1` from
# the trap and shows the buggy resume-to-success.
make_term_shim() {
local dir="$1"
cat > "$dir/cp" <<SHIM
#!/usr/bin/env bash
dest="\${@: -1}"
case "\$dest" in
*/$POISON_REL)
kill -TERM "\$PPID" 2>/dev/null # signal install.sh; the copy still succeeds
exec env PATH="$ORIG_PATH" cp "\$@" ;;
esac
exec env PATH="$ORIG_PATH" cp "\$@"
SHIM
chmod +x "$dir/cp"
}
# Run one keep-mode upgrade with the SIGTERM shim. Echoes "<exit>\t<out>\t<home>".
run_signal_upgrade() {
local installer="$1" H OUT SHIM rc
H=$(mktemp -d); OUT=$(mktemp); SHIM=$(mktemp -d)
seed_home "$H"
make_term_shim "$SHIM"
set +e
PATH="$SHIM:$ORIG_PATH" \
MOSAIC_HOME="$H" MOSAIC_INSTALL_MODE=keep MOSAIC_SYNC_ONLY=1 bash "$installer" >"$OUT" 2>&1
rc=$?
set -e 2>/dev/null || true
rm -rf "$SHIM"
printf '%s\t%s\t%s\n' "$rc" "$OUT" "$H"
}
IFS=$'\t' read -r rcC OUTC HC < <(run_signal_upgrade "$INSTALL")
chk "[signal] SIGTERM mid-sync aborts non-zero (trap exits, does not resume)" \
"[ '$rcC' -ne 0 ]"
chk "[signal] restore_snapshot fires on the interrupt" \
"grep -q 'restoring previous state from snapshot' '$OUTC'"
chk "[signal] does NOT resume to report sync success after the interrupt" \
"! grep -q 'file phase complete' '$OUTC'"
# Control: strip `exit 1` from the signal trap → the handler returns, the script
# resumes past the interrupt and wrongly reports success. In $FW so SOURCE_DIR resolves.
sed "s/trap 'restore_snapshot; exit 1' ERR INT TERM/trap 'restore_snapshot' ERR INT TERM/" \
"$INSTALL" > "$NOEXIT"
chk "[control] the exit-strip actually changed the installer" \
"! cmp -s '$INSTALL' '$NOEXIT'"
IFS=$'\t' read -r _rcD OUTD HD < <(run_signal_upgrade "$NOEXIT")
chk "[control] without 'exit 1' the trap resumes and reports sync success (the bug)" \
"grep -q 'file phase complete' '$OUTD'"
# ── Part D: a failed source/prune `find` scan must abort + roll back (D1) ─────
# A `< <(find …)` process substitution discards find's exit status, so an
# EACCES/I/O failure mid-scan would truncate the file list yet leave the reading
# loop exiting 0 — a partial upgrade committed and reported as success, with the
# ERR/restore trap never firing. The shipped installer captures the scan into a
# checked temp file (_scan_or_die) and aborts on failure. We inject a `find` that
# fails every `-print0` scan and assert the shipped installer rolls back.
IFS=$'\t' read -r rcE OUTE REFE HE < <(run_upgrade "$INSTALL" make_find_fail_shim)
chk "[find-fail] a failing framework scan aborts the upgrade non-zero" \
"[ '$rcE' -ne 0 ]"
chk "[find-fail] restore_snapshot fires on the aborted scan" \
"grep -q 'restoring previous state from snapshot' '$OUTE'"
chk "[find-fail] the abort is a fail-closed enumeration error (not a silent truncation)" \
"grep -q 'Could not enumerate framework files' '$OUTE'"
chk "[find-fail] target rolled back byte-identical to pre-upgrade state" \
"diff -r '$REFE' '$HE' >/dev/null 2>&1"
# Control: neuter the D1 guard (turn its `return 1` into a no-op) so a find
# failure is swallowed exactly as `< <(find …)` would — the scan appears to
# succeed and the upgrade reports completion with NO rollback.
sed 's/return 1 # D1-GUARD/: # D1-GUARD-DISABLED/' "$INSTALL" > "$D1CTRL"
chk "[control] the D1-guard strip actually changed the installer" \
"! cmp -s '$INSTALL' '$D1CTRL'"
IFS=$'\t' read -r _rcF OUTF REFF HF < <(run_upgrade "$D1CTRL" make_find_fail_shim)
chk "[control] with the D1 guard disabled the find failure is swallowed (no rollback)" \
"! grep -q 'restoring previous state from snapshot' '$OUTF'"
chk "[control] with the D1 guard disabled the upgrade wrongly reports success" \
"grep -q 'file phase complete' '$OUTF'"
# ── Part E: a failed target reset inside restore must not exit silently (D2) ──
# restore_snapshot resets the target (`rm -rf; mkdir -p`) before rebuilding from
# the snapshot. Under `set -e` (trap disarmed) a bare reset that fails would exit
# the whole script immediately — after `rm` may have deleted part of the target —
# WITHOUT printing where the snapshot lives. We trigger a rollback (cp poison) AND
# fail the target reset (rm shim), then assert the shipped installer emits the
# manual-recovery pointer and preserves the snapshot.
run_rmfail_upgrade() {
local installer="$1" H OUT SHIM rc
H=$(mktemp -d); OUT=$(mktemp); SHIM=$(mktemp -d)
seed_home "$H"
make_cp_shim "$SHIM" # poison cp → triggers the abort + restore
make_rm_fail_shim "$SHIM" # rm -rf <H> fails → exercises the D2 reset guard
set +e
PATH="$SHIM:$ORIG_PATH" ORIG_PATH_FOR_RM="$ORIG_PATH" FAIL_RM_TARGET="$H" \
MOSAIC_HOME="$H" MOSAIC_INSTALL_MODE=keep MOSAIC_SYNC_ONLY=1 bash "$installer" >"$OUT" 2>&1
rc=$?
set -e 2>/dev/null || true
rm -rf "$SHIM"
printf '%s\t%s\t%s\n' "$rc" "$OUT" "$H"
}
IFS=$'\t' read -r rcG OUTG HG < <(run_rmfail_upgrade "$INSTALL")
chk "[reset-fail] a failed target reset still aborts non-zero" \
"[ '$rcG' -ne 0 ]"
chk "[reset-fail] the manual-recovery pointer is emitted (not a silent set -e exit)" \
"grep -q 'Snapshot restore could not reset' '$OUTG'"
chk "[reset-fail] the recovery message points at a preserved snapshot dir" \
"grep -q 'preserved at: .*mosaic-snapshot' '$OUTG'"
SNAP_E="$(grep -o '/[^ ]*mosaic-snapshot[^ ]*' "$OUTG" | head -1)"
chk "[reset-fail] the named snapshot directory actually survives for recovery" \
"[ -n '$SNAP_E' ] && [ -d '$SNAP_E' ]"
chk "[reset-fail] operator secret value never appears in installer output" \
"! grep -q '$SECRET' '$OUTG'"
# Control: delete the D2 recovery line so a failed reset returns non-zero with NO
# operator pointer — the observable defect (half-reset target, snapshot orphaned
# in /tmp with no path told to the operator). Proves the message is load-bearing.
sed '/Snapshot restore could not reset/d' "$INSTALL" > "$D2CTRL"
chk "[control] the D2-recovery strip actually changed the installer" \
"! cmp -s '$INSTALL' '$D2CTRL'"
IFS=$'\t' read -r _rcH OUTH HH < <(run_rmfail_upgrade "$D2CTRL")
chk "[control] without the D2 recovery line the operator gets no snapshot pointer" \
"! grep -q 'Snapshot restore could not reset' '$OUTH'"
[ -n "${SNAP_E:-}" ] && rm -rf "$SNAP_E"
# Reap any snapshot the reset-fail runs left in /tmp (reset failed → never cleaned).
grep -o '/[^ ]*mosaic-snapshot[^ ]*' "$OUTH" 2>/dev/null | head -1 | while read -r s; do rm -rf "$s"; done
# Cleanup ($STRIPPED / $NOEXIT / $D1CTRL / $D2CTRL are also removed by the EXIT trap).
for d in "$HA" "$REFA" "$HB" "$REFB" "$HC" "$HD" "$HE" "$REFE" "$HF" "$REFF" "$HG" "$HH"; do rm -rf "$d"; done
rm -f "$OUTA" "$OUTB" "$OUTC" "$OUTD" "$OUTE" "$OUTF" "$OUTG" "$OUTH" \
"$STRIPPED" "$NOEXIT" "$D1CTRL" "$D2CTRL"
echo
echo "RESULT: $pass passed, $fail failed"
[ "$fail" -eq 0 ]