* fix: verify checksums for downloaded ISOs/images, keep sshpass off argv
ensure-base-iso.sh downloaded the PVE install ISO over plain HTTP with no
checksum, caching it on the persistent /opt/pve-integration mount and
booting it as the nested trust root the integration suite relies on.
ensure-cloud-images.sh fetched the Ubuntu cloud image and OVA over HTTPS
but never checked them either. prepare-test-environment.sh and
diagnose-cluster.sh passed the nested root password to sshpass via -p,
putting it in the process table. create-api-token.sh, unused anywhere in
the repo, minted a privsep=0 root token and echoed the secret unmasked.
- ensure-base-iso.sh now downloads from https://enterprise.proxmox.com/iso
and verifies against its SHA256SUMS on every run, including a cache hit.
download.proxmox.com's own TLS cert does not list download.proxmox.com in
its SAN (confirmed with curl/openssl from this environment), so https to
that name fails certificate validation; enterprise.proxmox.com serves the
identical ISO tree over a valid cert. Verification happens before the
downloaded file is moved to its canonical cache path.
- ensure-cloud-images.sh verifies the cloud image and OVA against Ubuntu's
published SHA256SUMS the same way, matching by upstream filename since
the cloud image is cached locally under a different extension (.img
upstream, .qcow2 cached — the bytes are already qcow2-formatted).
- prepare-test-environment.sh and diagnose-cluster.sh now export SSHPASS
and call sshpass -e, keeping the password out of argv/ps. This also fixes
a latent bug: the old unquoted `sshpass -p ${ROOT_PASS}` word-split any
password containing whitespace.
- create-api-token.sh deleted; grep across the repo found no caller.
Reviewers (codex:codex-rescue, correctness-reviewer, security-reviewer) all
independently found the same blocking bug in the first pass: when a cached
file failed verification and the subsequent redownload then failed,
ensure-cloud-images.sh fell through to a "keep the stale copy" branch and
returned that same known-bad file with exit 0 — verification could be
bypassed by inducing one failed redownload. Fixed by deleting the file
immediately on a failed verification, before the redownload is attempted,
so the later "is there a safe stale copy" check can no longer find it.
Added a test case (case 5) that reproduces this exact sequence and
mutation-tested it against the unfixed code. The three reviews also
flagged a real but separate bug already fixed in this same change: `trap
... RETURN` inside a function nested in another function is not scoped to
that function in bash — it re-fires on the OUTER function's return,
referencing an out-of-scope local. Both verify_checksum() helpers now
clean up their temp file explicitly instead of via trap.
Findings not acted on, judged out of scope for this fix:
- SHA256SUMS-fetch failures are treated the same as a checksum mismatch
(delete + fail) rather than left untouched — a transient network blip
destroys a good multi-GB cached ISO. This is the safer failure direction
(never silently trust unverified bytes) and was a deliberate trade-off,
not a defect.
- ensure-cloud-images.sh's 7-day cache window can span an upstream
republish of noble/current, causing a legitimate re-verification churn
(not a security issue, a cache-hit-rate one). Pre-existing cache design,
unrelated to adding verification.
- wait-for-pve.sh (curl -d with the password on argv) and
prepare-test-environment.sh's own positional password argument (from
run-integration.sh) carry the same password-on-argv pattern this issue
targeted in create-api-token.sh, sshpass -p and diagnose-cluster.sh, but
neither script nor run-integration.sh was named in the issue. Left
untouched per scope; worth a follow-up issue.
- GPG/detached-signature verification of the upstream SHA256SUMS was not
added — the new checks defend against cache poisoning and transit
corruption, not a compromised origin. Worth a follow-up issue.
- The two new self-checks (ensure-base-iso.test.sh,
ensure-cloud-images.test.sh) are not wired into
.github/workflows/unit-tests.yml's shell-selfchecks job. That file is
code-owned and out of scope for this change; needs an operator follow-up.
Password rotation (the Testpass123! value from before it moved to a
secret) is unaddressed here per the contract — flagged for the operator.
Mutation-tested: broke the post-download checksum check in
ensure-base-iso.sh, confirmed the affected test cases failed, restored it.
Broke the sshpass -e change back to -p, confirmed the new assertions in
prepare-test-environment.test.sh failed, restored it. Broke the fail-open
fix in ensure-cloud-images.sh, confirmed case 5 failed, restored it.
Closes#149
* fix: also verify the stale-by-age fallback copy in ensure-cloud-images.sh
PR review on #166 (COMMENTED, non-blocking) found the sibling of the
fail-open bug already fixed in this branch: when the cached cloud image
is stale by *age* (>= 7 days) rather than failed verification, the
redownload-failure fallback could hand back that file with exit 0
without ever re-verifying it in this run. A file that failed the
earlier verification is already deleted by the time the fallback runs,
but a stale-by-age file skips verification entirely on the way in.
Fixed by verifying the stale-by-age file at the point of actual
fallback use — after the redownload has failed, not proactively before
it's attempted, so a copy the redownload was about to replace anyway
isn't deleted along a path that would have succeeded. Added two test
cases (6, 7): a still-verifying stale-by-age copy is used as a
fallback; one that no longer verifies is not. Mutation-tested by
reverting to the unfixed fallback and confirming case 7 fails, then
restored.
---------
Co-authored-by: goodolclint-claude[bot] <323206664+goodolclint-claude[bot]@users.noreply.github.com>
Run 173 left node B with healthy corosync (2-member primary component, both
links connected) but no /etc/pve/corosync.conf, no dcdb/status journal lines,
and pvecm status reporting it is not part of a cluster. That file is
database-backed: pmxcfs creates it only when it starts with no config.db and
imports /etc/corosync/corosync.conf, so a surviving standalone config.db would
mean silent local mode.
Capture the package versions, pmxcfs command line, /etc/pve mount, .members,
the config.db and its backup dir, whether the database holds a corosync.conf
row, and the CPG group membership. Read-only; the sqlite3 CLI is not guaranteed
on a PVE node, so fall back to strings.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
"Cluster join aborted!" is PVE's generic wrapper; the reason lives only in
the task log on the joining node. Run 172's log said "An error occurred on
the cluster node: cluster not ready - no quorum?", which is what identified
the race. Capture it so the evidence survives cleanup.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The API reports a joined-but-offline node as online=0 with no further
detail, and the cleanup job destroys the nodes minutes later, so the
reason corosync membership never forms has never reached a log. Read
corosync.conf, corosync-cfgtool, pvecm status and the corosync journal
off both nodes while they are still alive. Best-effort: never fails the
caller.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
goodolclint-claude[bot]