Fourteen roles repeated the same two lines to say they use the gateway: a
`shared` flag gated on group_names, and an `enabled` flag conjoining that gate
with LITELLM_CHAT_MODEL_SERVED. Both are now single-sourced.
`shared` becomes literal true. The flag means the provider is a shared instance
rather than one deployed per consumer, and a central gateway is always that;
the dynamic form only said "shared if deployed", which `enabled` already
decides. The 50 shared pins under `litellm:` leave meta/variants.yml with it.
The tree had already noticed: 20 of those 50 lines carried
`# nocheck: variants-static-override`, two of them with copy-pasted text about
container_backup. This narrows e5de4e085a for one service key rather than
reversing it; every other dynamic flag keeps the group_names form.
`enabled` becomes `{{ LITELLM_USABLE }}`, defined in group_vars/all/16_ai.yml.
That shorthand is only possible because is_explicit_truth learns about it.
The resolver reads a dependency out of the flag TEXT — "in group_names" in
value — so a constant hides the role id and the provider silently leaves the
static closure of every role referencing it. Measured: the old expression
resolves True, `{{ LITELLM_USABLE }}` resolves False. GROUP_GATED_CONSTANTS
names the constants whose own definition carries the gate, and the docstring
states the obligation: a SPOT of this kind that is not listed there drops its
provider without a word.
The wider fix, accepting any Jinja as dynamic, was measured and rejected: 35
other flags carry a Jinja value without the marker (MCP sidecars, javascript
and SSO gates, recaptcha), and they would all enter closures they are not in
today.
A literal shared flag is no longer dynamic, so test_dynamic_flags wants the
`# nocheck: dynamic-flag` its own message documents for legitimate exceptions;
where a marker already existed the rule list is extended rather than a second
comment added.
Behaviour is unchanged where it counts. The closure keeps litellm with the same
app counts (wordpress 26, openwebui 23, hermes 23, mediawiki 22, flowise 23),
unrelated Jinja still resolves False, and the truth table of the new enabled
form matches the old one in all four served/deployed combinations. Three roles
resolve without litellm, and `git show HEAD:` proves that predates this change:
zammad pins enabled false in variant 0 at line 39, nextcloud at line 97.
Six unit tests cover the resolver, including one asserting that every
registered constant is actually defined in group_vars, so an entry cannot
become a dead promise.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The lookup writes overrides.get(key, effective_default), so a service that
repeats what it would have inherited changes nothing and only hides the day the
default moves. Forty seven such lines stood in the tree.
Almost all of them came from the migration that moved the literal blocks into
meta/services.yml. The values were carried over verbatim so the move could be
proven lossless, which was right for the proof and left the no-ops behind.
The defaults they are measured against are the effective ones, not the base
class. Curl probes every minute, Connect grants a fifteen minute start window,
and a prefix raises timeout, retries and start_period further. Comparing against
Probe alone would have called out every service that simply agrees with its
flavor.
Eight services lose interval: 10s as well and fall back to their flavor's
cadence, thirty seconds, or a minute for decidim's curl. Since start_interval
exists, interval no longer touches convergence at all: it is steady state load
for the life of the deployment. That takes the five databases, matomo, snipe-it
and decidim from 360 probes an hour each to 120. It buys that with slower
failure detection, 180s to 540s for the databases, which is acceptable because
an unhealthy container fails the deploy whichever number stands there.
Twelve deviations that do cost something now carry their reason. Where the
mechanism is known it is named: ejabberdctl and rabbitmq-diagnostics start an
Erlang CLI node, mmctl spawns a binary and talks to the local socket, php artisan
boots Laravel on every probe, keycloak's management endpoint answers from the
JVM.
Two say instead that the value was carried over and never measured. './fider
ping' and pgrep both fit the five second default and no reason for ten is
visible, so the comment records that there is no evidence rather than inventing
one. They mark where a measurement belongs.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
42ebd9ea71 gave the healthcheck SPOT start_interval, the cadence docker uses
while start_period is still running. A block a template writes out itself
bypasses the SPOT and never gained it. Measured across the repository: 46 such
blocks, 40 of them already carrying start_period, not one carrying
start_interval.
Without it a probe fires only every interval, so a service ready after two
seconds is reported healthy at the next tick and a stack converges no faster
than its coarsest quantisation. That lands in sys-svc-compose's swarm converge
wait, 1551.7s per CI job over 38 stack deploys.
Twenty five of the blocks move into meta/services.yml and render through
container_healthcheck. Each keeps an explicit test rather than picking a flavor:
a flavor would replace the command instead of carrying it across.
The move is lossless by construction. probes.py block() writes
overrides.get(key, default) for interval, timeout, retries, start_period and
start_interval, so a declaration repeating the literal values renders the same
block plus the missing cadence. Checked per block against the old values: the
only difference is start_interval 5s, nothing changed, nothing lost.
Twenty one stay literal. Seventeen interpolate a template variable, whose value
resolves in a scope meta/services.yml does not share, so moving them would be a
rewrite rather than a transcription. Four interpolate nothing and still cannot
move: svc-db-memcached and svc-db-redis probe every second in their swarm
service partial, where a 5s start cadence would report healthy later than the
block already does, and the sys-svc-rdbms and keycloak sso_proxy partials render
for the consuming role's application_id, so their declaration has no single
meta/services.yml to live in. Each carries a nocheck naming its reason.
All twenty one gain start_interval. Three also gain start_period, without which
the key does nothing at all, since it only applies while the start period runs.
That widens their failure tolerance rather than narrowing it: bluesky from 90 to
120 seconds, magento from 50 to 80, the mariadb partial from 180 to 210.
pixelfed spelled its probe as a bare string and now spells it CMD-SHELL. Docker
reads both as a shell command; only the YAML shape changes.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The volume mounted at /mattermost/plugins was an NFS export shared by every
replica, so a redeploy had one task remove the tree while the outgoing task
still held the plugin binary open. Measured on task .1.ixph4yrvwfbo7mgghcbdsyreh:
13:41:15.658 [outgoing] plugin exited mattermost-ai
13:41:17.280 [incoming] Removing local installation of managed plugin before sync
13:41:17.304 [incoming] Failed to remove ... unlinkat plugins/mattermost-ai/server/dist: directory not empty
13:41:24.227 [incoming] Failed to install prepackaged plugin ... destination already exists
14:28:23.936 [incoming] Access to route for non-existent plugin mattermost-ai
NFS replaces a still-open file with a .nfsXXXX entry on unlink, so unlinkat on
the parent returns ENOTEMPTY. The removal fails, the directory survives, the
reinstall cannot move into it, and that node answers 404 on the plugin routes
for the rest of its life while the service VIP keeps sending it traffic. On a
local filesystem unlinking an open file is legal, so the failure cannot arise.
Four of six replicas in the same run had no overlapping predecessor and came up
clean.
becd8f1577 fixed the sibling volume and deliberately left this one shared, on
the grounds that uploaded plugin archives live in it. That reason does not
hold. Upstream server/channels/app/plugin.go lists bundles from
fileStorePluginFolder, reads them with fileReader and calls installPluginLocally
after removePluginLocally, which is the code path that emits the log line above;
the local directory is the copy and the file store is the original. In this
deployment the file store holds no plugin at all: every replica logs "No plugins
to sync from the file store" and the bundle comes from
/mattermost/prepackaged_plugins inside the image.
So plugins and client-plugins are the two halves of one per-node extraction. The
activation error names bundle_path plugins/mattermost-ai while failing to copy
into client/plugins, and splitting their lifetimes is what left one half node
local and the other shared. They now share a lifetime.
Two consequences worth stating. The flag only takes effect on a volume that does
not exist yet, because docker does not recreate a named volume when driver_opts
change and stack deploy --prune prunes services; a cluster already running the
NFS stack needs the volume removed per node. And the worker copies leave the
export's backup path for svc-bkp-volume-2-local, which runs manager-only, so
they are covered by nothing. The content is extracted per node from the image
on every boot, which is what the nfs-false-data-loss exemption argues.
The README promised persistence through both plugin volumes. That has been
untrue for client-plugins since becd8f1577, so it now says what the volumes are
and where an uploaded bundle actually persists.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
confluence 10.2.16 to 10.2.18, erpnext v16.34.1 to v16.34.2, espocrm 10.0.7 to 10.0.8, hermes v2026.7.20 to v2026.9.7, mattermost 11.10.1 to 11.11.0, nextcloud context_agent 2.7.0 to 2.8.0, seaweedfs 4.45 to 4.46, coturn 4.17.2 to 4.18.0. The automated updater opens one PR per role and had not reached these eight; playwright and elasticsearch were already current in the tree.
audit.md follows rather than leads: the MCP report carries a version column and test_mcp_audit_report compares the committed file against a fresh generation, so hermes and mattermost move there too. Regenerated with audit.gen.py as that test instructs. Note that make docs does not run this generator, so the report only tracks reality when somebody calls it by hand.
Verified by the docker-version check itself: eleven Outdated warnings before, zero after.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
svc-ai-litellm has two variants: 0 with both backends on, 1 with both off. The planner assigns a dependency round_index when it has that many variants and falls back to 0 otherwise (inventory/planner.py:111), so round 1 is the one round in which the gateway deliberately serves nothing - a topology that is green in its own job.
Commit e0048f157b gated 14 consumers on LITELLM_CHAT_MODEL_SERVED so a consumer switches itself off when no model is served. Exactly two of them pinned litellm on in round 1, and a variant's literal beats the base Jinja gate, so mattermost rendered LITELLM_SERVICE_ENABLED=true against an empty model_list. The Agents plugin then failed client-side with 'bifrost error: model is required' 47ms after the prompt, never reaching the gateway.
Moving both pins to variant 2 lands them where the fallback supplies svc-ai-litellm=v0 with svc-ai-ollama=v0. Measured with the planner: round 2 pulls both. Both polarities stay pinned, so the variant-coverage lint still holds. The contract assert now also refuses an empty model, which the stored-vs-intended comparison read as converged.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The gateway publishes only models it has a backend for and renders model_list: [] when neither ollama nor lmstudio is deployed and no OpenRouter key is set. LITELLM_CHAT_MODEL fell back to openrouter/auto without checking that key, so consumers were configured for a model the gateway does not carry and every prompt failed with a 400.
Observed in the Mattermost#1 CI job: litellm.log carries 12 occurrences of "Invalid model name passed in model=openrouter/auto" next to 400 Bad Request, and tests/test-litellm-agent.js failed on every attempt while the rest of the suite passed.
Make the fallback conditional on the key so the name is empty when nothing is served, promote LITELLM_CHAT_MODEL_SERVED from the role vars to group_vars where consumers can see it (it was the third copy of this logic and the only correct one), and switch the 14 consumers from "litellm is deployed" to "litellm serves a model". The Playwright service flag reads the same lookup, so the spec skips together with the Ansible gate. The bool filter carries weight: the folded scalar yields the string False, which Jinja treats as truthy.
A new unit test renders both expressions against Jinja and pins the invariant that a model name exists exactly when a model is served, including the case where ollama preloads embedding models only.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
supported_version and minimum_version sat in 24 roles and were read by no template, no task and no lookup. Their only consumer was the audit report version column, which turned false the moment an automated image bump ran: qdrant reported a verified version nobody had verified, two releases long. Both leave MCP_KEYS, so a role that declares them again fails the schema lint instead of carrying a field with no effect.
Requirement 025 mandated minimum_version as a MUST and 035 carried it in an example; both are corrected here, otherwise the repository would contradict its own code.
The gate then surfaced five dead docs.mattermost.com links. Mattermost moved to /administration-guide/ and /deployment-guide/, and every replacement was probed for 200 rather than guessed. The MCP page is gone with no successor, so mattermost now points at mcpserver/README.md in the Agents repository, the source rather than a rendered mirror.
Both concerns share three files - mattermost's meta/mcp.yml, requirement 025, and the generated audit.md, which derives from a source of each - so splitting them would commit an artifact that does not match its sources.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
An MCP client spec signs in through the client's own OIDC button, which
redirects to Keycloak. Every provider put its MCP pairing in the lean variant,
and that variant pinned sso off - so the round that exists to prove the pairing
was the one round without an identity provider. Run 33574056480 shows what that
costs: openwebui's MCP specs reached _shared.js:131 and died on "openwebui OIDC
sign-in button must be visible on the auth page".
The rule now demands two things of the variant that carries the consumers: the
provider pins sso on, and each admitted consumer, in the variant that round
actually pulls in, does not pin it off. The second half matters because a round
deploys every dependency at its own index and falls back to variant 0 only when
the dependency has fewer, so the client shipping alongside is that one and not
the client's variant 0. A role that declares no sso, or pins it to a literal
false because upstream speaks no OIDC, is exempt from a flag it cannot have.
Twenty roles moved to satisfy it. Ten already had a second sso: false pin and
only needed the flip. Six could host the pairing in variant 0, which already
has sso on, and dropped back to two variants. Six kept a third lean variant to
carry the false polarity variants-coverage wants. moodle and mattermost took
swaps instead: moodle's variant 1 is the mandated LDAP-only branch, and
mattermost was already at the three-variant cap with sso on in its variant 2.
nextcloud also turns context_agent on in that round, the ExApp that serves the
endpoint - with it off the round proved nothing it was built to prove.
The rule moves to tests/lint/ansible/services/ because it reads meta YAML and
executes nothing, which is where integration.md says such a check belongs, and
its suppression key joins the catalog it was never listed in.
nextcloud_exapp_reregister becomes nextcloud_exapp_renew: mcp-secret-masking
matches `register:` unanchored, so the substring in the old name read as a task
that prints a secret. register.yml already masks the two tasks that carry one.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The mcp-consumer-variant guard demanded three things of a provider's MCP
round: every admitted consumer enabled, each of them pinned `services: null`
so it dragged no dependencies along, and no other service left enabled. Two
of those were the rule legislating a round's contents rather than checking
that the pairing gets proven at all.
Keep the first and drop the rest. What else the round carries is the role's
own call, so the tor/litellm exemption, the datastore exemption and the
`allowed` set they fed have no reader left and go with them.
Strip the 63 `services: null` pins the rule forced across 20 variants, all
written bare as `services:`. The topic-emptying capability behind them
stays: apply_topic still normalises None to {} for the inventory payload,
and the unit tests still cover it. Only the uses and the mandate are gone.
That leaves overridden_providers scanning a repository with nothing left to
find, so its three assertions passed over an empty dict and would have
stayed green however the function broke. Rebuild them on a synthetic role
tree, which also covers the case the repository scan never held: a variant
naming a plain flag is not an override.
Removing the pins changes what deploys. Measured across all 20 variants: 18
identical, prometheus#1 grows 23 to 31 roles, moodle#2 grows 31 to 32 and
crosses the memory budget until the following commit shrinks it back.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
A round deploys a provider once, so the role that pulls it in is the only place that can say how lean it has to be. A variant entry may now carry any application config topic under a service it depends on: a mapping deep-merges, so naming one key changes that key alone, and null empties the topic. Empty was the case that needed a marker of its own, because merging an empty mapping is a no-op and would leave exactly the services the round was trying to shed.
The topic vocabulary and the merge semantics live in utils.roles.applications.topics, which is also where resources.py now gets them rather than spelling the same null-or-merge out a second time. Collecting who demands what stays in the inventory tree, refuses two roles demanding different values for one provider in the same round, and skips a role's own entity entry so the existing services.<own>.users declarations are not read as overrides.
The CI set-cover learns the same fact. Its greedy cover retires a role's variant-0 row when an earlier row of another role embeds it, on the stated premise that embedding pulls the provider's variant 0. A variant that dictates that provider's config breaks the premise from the other side: whatever ran, it was not the provider as the provider declares it. Such a pair now covers nothing, so the retired row keeps its job, its artefact and its verdict.
Measured on the current tree: 20 override pairs, 41 covered rows before and after, so nothing is revoked today. The rule bites the moment an override lands in a row that covers.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The contract check ran as an Ansible task in the middle of every provider deploy, and the two client checks ran the same way. A test that is part of the deploy cannot be run without deploying, cannot be re-run on its own, and reports its verdict as a deploy failure. Twenty-seven task files lose 491 lines to this.
It is now what it always was. Each provider ships files/test/test.sh, templates/test.env.j2 and meta/tests.yml, and test-e2e-cli stages one shared harness beside it: run.sh drives contract.sh in a throwaway container on the provider's own network, then isolation.sh for adapter-fronted providers. The read probe each provider offers moved from the calling task into meta/mcp.yml next to the allowlist it must satisfy, and the schema now rejects a read probe naming a tool outside that allowlist.
The bearer the test presents is the one a client resolves through mcp.credential, via the new mcp_client_bearer lookup over resolve_credential, which moved to utils.roles.applications.mcp so it is the same function mcp_servers renders every client connection with. The probe presented whatever the calling role passed, which for the fifteen adapter-fronted providers is a different secret than any client holds.
Sidecar isolation is read from the engine rather than measured inside the container, so it no longer needs a delegation to the node holding the task: attachment is compared by network name, which drops the gwbridge special case, and a swarm worker says it cannot measure instead of skipping silently.
Retrying is bounded by wall clock against MCP_DEADLINE_SECONDS, which scales through the same lookup('timeout') as the harness budget it has to stay under. A flag that is neither true nor false aborts rather than reading as off.
mcp-provider-probe becomes mcp-provider-cli-test: an enabled provider must ship the test rather than include the task. mcp-probe-bearer becomes mcp-adapter-bearer, tying ADAPTER_BEARER to the declared credential key now that no probe include carries it. A new integration test renders the shared env template for every provider through Ansible's own templar, so a later change to it breaks visibly instead of poisoning twenty deploys quietly.
Hermes and OpenClaw keep their client-side proof as CLI tests of their own.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Thirteen of sixteen MCP sidecar entries carried no enabled at all, so
collect_role_resources fell back to _looks_like_container and billed every
one of them as running. Three roles already gated theirs on mcp.enabled;
the rest now do the same, gitea twice because it is the only role with a
vendor container behind the adapter.
The fifteen adapter roles then read services.<key>.enabled instead of
mcp.enabled, so the compose block and the task gates follow the services
entry rather than running beside it on a second path. The six plugin and
native roles keep reading the topic: they ship no sidecar entry to point at.
The three-level resolution is not new. KEYCLOAK_LOGOUT_PANEL_ENABLED has
the same shape and depth and gates a task the same way, and the same value
reaches JAVASCRIPT_SERVICE_ENABLED, where service-gating.js throws on
anything but "true" or "false" — an unresolved chain would have been
killing keycloak's Playwright run since 575a5a87ac.
This does not change what the resource model counts: _is_enabled reads the
unrendered Jinja and still treats it as on. What it buys is one switch
instead of two, and a declaration that matches it.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
83c1c729de gave every provider a group_names-gated services entry per MCP
client, including the clients it refuses with mcp_consumer: false. The gate
is what makes the inventory resolver plan provider and client into one round,
so 19 of the repo's 82 co-deploy pairs started a client for a pairing the
provider had already declined. flowise is a live MCP client of qdrant,
homeassistant and prometheus; the other 17 providers were bringing up its
node stack for nothing.
The entries now read enabled: false, shared: false, mcp_consumer: false, and
the variants that pinned them to true are gone. Setting those variants to
false instead is not an option: test_variants_no_static_override forbids any
variant override of a literal base value in either direction.
bond leaves with them, and a new lint holds that line: collect_role_resources
skips a literally disabled entry before it reads any resource key, so a bond
there describes a cost nobody pays. Thirteen entries unrelated to MCP carried
one and are cleared too.
test_mcp_consumer_services_entry demanded the group_names shape for every
client entry, refused ones included, and now demands the opposite there.
dynamic-flag keeps its rule and takes its own documented per-key exemption.
make docs redraws twenty graphs: the edges go, the nodes stay as
'openwebui ❌' and 'flowise ❌', so the refusal is visible rather than read as
an integration.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
enabled and shared were hand-maintained in 24 roles and wrong in both
directions. Both now resolve from the mcp_consumer edges already declared
in meta/services.yml: a server takes the consumers it admits, a client
takes the servers that admit it, and the pair renders as "'peer' in
group_names or ...", or False when no peer is admitted at all. setdefault
keeps meta/mcp.yml and meta/variants.yml as the override channel, so
pinning either key still wins.
The resolved block replaces the mapping instead of being written into it.
_build_role_base_config parks the YAML cache's own object under mcp, so an
in-place setdefault injected the derived keys into every later raw read and
made test_mcp_schema pass alone but fail in a directory run.
The derivation keys on peer presence, never on a peer's own value: a
re-entrant read under _RENDER_GUARD returns the unrendered dict.
Four lints now ask the resolver instead of reading meta/mcp.yml, which they
could not do while the key was hand-written. That is what exposed the eleven
missing hostile-path specs the previous commit closed.
Carries one change that belongs to the WordPress probe but shares a file
with this one: the tool allowlist moves to the hyphenated names the adapter
actually serves.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
allowed_consumers was twenty hand-kept lists of which fourteen were identical, so every new client meant an edit in every provider. A client now declares services.<own-entity>.mcp_consumer: true once and a provider refuses one with mcp_consumer: false, which is also what tells a deliberate exclusion apart from a forgotten entry.
The group_names-gated consumer entries give the inventory resolver the co-deploy edge meta/mcp.yml never had: one of sixty-nine matrix rounds carried an MCP client before, twenty-three role/round pairs do now. bond 0.5 keeps the clients out of round 0 so the host memory budget still holds.
BREAKING CHANGE: meta/mcp.yml no longer accepts allowed_consumers; declare services.<entity>.mcp_consumer on the client instead.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
svc-ai-mcp-adapter declares its own probe - http_status on /health,
rendered through container_healthcheck - but the four roles that run it
write their sidecar block inline and dropped both halves. Without a
probe docker calls the container up while the process lives, which says
nothing about whether it answers.
CI named only baserow, because the rule skips a service the inventory
does not deploy and the other three have MCP off there. They carry the
same defect, so all four get the declaration.
The check cannot run locally at all: it sweeps every declared image
through the GHCR mirror, whose prefix only CI writes, so it skips here
by design. make quality-high was green and could not have been anything
else.
e3e81e816f made services.yml the place a probe is declared; the sidecars
were added afterwards and never picked it up.
f5272fa85f derived the 404 from mcp.endpoint.path, which for an
adapter-fronted provider names the sidecar. Mattermost declares /mcp
there and answers on /plugins/mattermost-ai/mcp-server/mcp, so the proxy
blocked a path the application never served while the native surface
stayed reachable. Baserow passed only because its two paths coincide -
one role agreeing proved nothing about the other four.
That surface is what the adapter exists to narrow: it serves four tools
out of fourteen, and the native path reaches all fourteen.
adapter.upstream_path joins upstream_network and upstream_transport in
meta, and the four roles whose upstream lives on their own vhost declare
it there. Gitea does not: its upstream is a separate sidecar on its own
network, so it has nothing to hide on the application's vhost, and the
absence says that.
The roles read the value back through lookup('config', ...) rather than
keeping a second copy, so the path is stated once.
The template now separates the two roles the old variable conflated:
endpoint.path is what gets served, and only where the surface is public
and enabled; every declared path that differs from it gets a 404. Three
tests pin the distinction rather than the outcome - mattermost's two
paths must differ, gitea must declare none, and a passthrough provider
must say where its upstream lives.
Eleven roles still declared their credentials in meta/schema.yml while the
rest of the repository had moved to meta/secrets.yml. The name is not
cosmetic: the inventory resolves credentials at that path, as
d86913225b and c541e7c22f record, so a role left on the old name drops out
of that resolution.
All eleven are pure renames with no content change.
Also correct the four remaining places that still named the old path in
prose: the credentials reset CLI help, two test docstrings and the espocrm
README. None of them resolved a file, but they misdescribe where credentials
live, which is exactly what sent an earlier reader to the wrong location.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
834 commits from the Tor onion-service branch against 299 of ours, from merge
base 7b646bdece: 127 conflicted files, 3506 files changed.
The two sides are largely orthogonal. Theirs adds svc-net-tor and a `tor` service
across the role tree, plus two sweeping refactors: shipped scripts sorted by
language (files/ruby/, files/python/, files/shell/, scripts/lint/*.sh) and the
unit suite moved under tests/unit/python/. Ours is requirement 036, the native AI
gateway, plus the MCP role work. Most conflicts were both sides appending to the
same service map, resolved as unions.
Resolutions that were not unions:
Credentials moved namespace. Theirs renamed meta/schema.yml to meta/secrets.yml
and re-addressed credentials as secrets.credentials.<name>. Git carried our
litellm_api_key blocks through the rename, but every consumer we added still read
'credentials.<name>', which no longer resolves. All 45 call sites are migrated.
plugins/filter/container_volumes.py is deleted by theirs and relocated to
plugins/lookup/container_volumes.py, where it carries the same tmpfs defect our
f85996b43 had just fixed: a hard-coded {"size": ...} that silently drops a per
mount `mode`. The fix is re-applied in the relocated file, as that commit's merge
note anticipated.
compose.yml keeps only INFINITO_PLAYWRIGHT_REPORTS_BASE_DIR. Theirs also forwards
INFINITO_PLAYWRIGHT_STAGE_BASE_DIR; that line is deliberately dropped, per the
analysis recorded in 45804e9ad -- STAGE has a drift-proof role-vars fallback,
while forwarding it promotes a setdefault-pinnable .env value above that SPOT.
tasks/stages/01_constructor.yml takes their include, ./tasks/utils/setup/group.yml,
because their refactor removed tasks/groups/*-roles.yml, leaving our include
target nonexistent. Our task name is kept, since svc-virt is in the loop.
Makefile swarm-zombie is a union: their validate plus teardown, then our
lab_subnet cleanup, which their side dropped while leaving lab_subnet.sh as the
only orphaned reference in the file.
01_bootstrap.sh keeps our deterministic --ip pinning, which distinguishes an
already-attached node from a real failure, over their `|| true`.
Their language-sorted layout orphaned references our side still held: zammad's
two scripts/apply/*.rb, jenkins' java-with-ca-trust.sh, and the notify topics
their handlers renamed to hyphenated form. All corrected.
roles/sys-svc-compose/files/shell/swarm/stack_ready.sh gains our latched-update
block: paused, rollback_paused, rollback_started and rollback_completed exit 2
rather than polling a state that cannot resolve without another deploy.
config_from_run is unioned to (title, logged, jobs); the production caller in
trigger/__main__.py now passes jobs, without which our suite suppression would
have been inert.
READMEs are regenerated rather than hand-merged: 15 LiteLLM rows and 91 Tor rows
now derive from the merged metadata, which is itself evidence the meta merge is
coherent.
Verified: 5161 unit tests pass. The single failure, test_query.py TestSortSpec,
also fails on fork/feature/svc-net-tor at b51d2f9919, measured in a baseline
worktree -- so this merge introduces no new unit-test failure. Two modules are
excluded from that run because `baudolo`, a dependency their branch introduces,
is not installed on this host.
Not verified: no deploy has run against the merged tree, and make quality has not
been run end to end here.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The previous contract could not fail. Its guard and its closing assertion both
keyed on apiURL -- the one field the merge itself writes -- so once any run had
set the URL, every later run skipped the merge on a false_condition and still
reported all assertions passed. The live row shows what that hides: an apiKey of
`sk-` followed by 48 literal `a` characters, which the gateway answers 401 to,
and defaultModel `llama3:latest` while the round's gateway serves only
`smollm2:135m`. The merge task is `skipping` in every deploy log on this host.
The read now compares the whole desired service and bot with jsonb equality and
answers `converged` or `stale`, so a drifted key or model is drift the guard can
see. jsonb equality is key-order independent, which keeps the check from
depending on how Jinja happened to serialise the dict. Both SQL templates render
from the same two role vars, so read and write cannot disagree about what
correct means. The `jsonb_typeof(...) = 'array'` guards stay: before the first
write those keys exist as JSON null, coalesce does not fire on JSON null, and
jsonb_array_elements then raises "cannot extract elements from a scalar".
A row edit cannot create the bot account, which is why the Playwright spec 404s
on user `agent`. In mattermost-plugin-agents v2.5.1 the config lives in an
in-memory atomic pointer; bots.EnsureBots() is the only path that calls
pluginAPI.Bot.Create, and it runs from OnActivate and from the config update
listener. Writing straight into agents_confighistory reaches neither. The role's
only reload lever was `notify: container recreate`, which never fired -- the
container is unchanged with RestartCount 0 while the config row was written
inside it. The plugin is now cycled explicitly through mmctl, and the enable step
moves into tasks/utils/agents/enable.yml so the MCP path and the AI path share
one definition.
The closing gate is deliberately not another read of what we wrote: it polls the
bots-to-users join until the agent account exists. That table is written by the
plugin and never by this role, so it cannot pass by echoing the deploy's own
work. The reload fires on stale config OR a missing bot, which covers the state
this stack is actually in, where the config matched but no account existed.
Verified: the stale row, the absent bot account and the always-skipping merge are
all measured on the live stack; the EnsureBots path was read from upstream v2.5.1
source rather than inferred; every changed task file parses.
Not verified: none of this has executed yet. Specifically open are whether
`MATTERMOST_LITELLM_SERVICE | to_json` renders nested Jinja inside a dict var,
whether the converged/stale SQL is valid against live Postgres, and whether a
disable/enable cycle triggers EnsureBots on this version.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
unsafe-inline re-admits the inline execution an XSS needs and unsafe-eval re-admits string-to-code, so a CSP relaxation must be verified against the running app rather than copied from another role. The new lint requires a nocheck marker plus a Reason on the flag's own line, so a marker cannot drift to the neighbouring directive.
The 197 pre-existing grants across 70 roles are checked out with a TODO: they were never verified and still have to be.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
72fd9c870 required the onion in every variant but exempted variant 1, on the
argument that the minimal round is the one place the stack runs without it. That
argument does not hold: the onion is part of the deployment under test, not a
feature bolted onto the full shape. A round that skips it is not a smaller
deployment, it is a different one.
So the exemption is gone and 95 variant-1 entries pin the onion on. Two
exclusions remain, both read from data rather than a comment: a role with no tor
block has no onion to pin, and a role whose meta/services.yml pins tor.enabled to
a literal false has already declared it has none.
That put the rule in conflict with test_coverage, which requires both polarities
of every dynamic flag: with tor on everywhere, the false side has no variant left
to live in. 94 roles take the exemption that contract already provides for
exactly this - "flags whose group-derived value is a hard invariant that variants
must never override", the same one 4e6ef2542 used for the seaweedfs frontend.
web-app-nextcloud needs none: its variant 8 pairs the clearnet-only PeerTube and
keeps the false polarity.
Measured, because the declaration is not what CI runs. assign_tor gives the onion
to every second capable entry with the run number as the parity switch, so more
capable variants do not mean proportionally more onion jobs: across five roles
and 21 entries the count goes 8 -> 10 on an even run and 9 on an odd one. The
clearnet path also survives - checkmk reports v1=false on run 0 and v1=true on
run 1, so both states are still covered across two consecutive runs. What the
94 markers give up is a declarative guarantee in variants.yml, not the coverage
itself.
The real cost is weight, not job count. web-app-nextcloud variant 1 goes from
3.09GB reservation / 7.27GB limit to 18.26GB / 43.01GB once svc-net-tor and its
closure join, so the cheap round is no longer cheap. Every variant stays under
the 32/64GB budget, and the same effect applies to the other 94 unmeasured.
Verified: tests/lint 336 passed / 307 subtests, tests/integration/roles/meta/
variants plus the role-meta lints 39 passed. The rule was checked in both
directions - it reported the variant-1 offenders before the flips and none after.
Not measured: what the heavier rounds do to wall-clock, against swarm jobs capped
at 355 minutes and one that has hit that ceiling before.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
7a70d27b7 made a variant that switches SSO on also switch the onion on, because
that pairing is where the platform breaks in ways clearnet never shows. The same
argument reaches further than SSO: a Secure cookie is dropped, an https identity
provider refuses to embed, and crypto.subtle is undefined for any feature on a
non-secure origin, not only for a login. A variant that never runs the onion is
not testing the deployment users get.
So the rule now covers every variant of a role that declares a tor block.
Variant 1 stays exempt - it is the minimal, partner-free round every role
declares and the one place the stack is exercised without the onion at all - and
a role whose meta/services.yml pins tor.enabled to a literal false is skipped
outright, because that declaration already says the role has no onion and needs
no comment repeating it.
That found 137 variants. 101 were variant 1. Of the rest, 28 are the LDAP-only
auth-isolation rounds, which gain real coverage from running LDAP over the onion
and are flipped on here. Four cannot: svc-net-tor is the provider itself, and
nextcloud's variant 8 pairs the clearnet-only PeerTube. Both carry a marker
naming the reason; opentalk and coturn need none, since their services.yml pins
the literal false the rule now reads.
Flipping them exposed a contract that contradicted itself. test_auth_isolation
and test_non_baseline_explicit_disables demanded `shared: false` beside
`enabled: false`, while variants-static-override flagged that same line as a
static override - which is why all 133 of them carried a nocheck marker. The
demand only makes sense where `shared` is dynamic: a literal base value cannot
leak through deep-merge, so restating it protects nothing. Both contracts now
ask for it only in that case, and the 133 lines are gone.
The result is a stricter rule with fewer lines: 189 insertions against 193
deletions.
Verified: tests/lint 336 passed / 307 subtests, tests/integration/roles 107
passed / 610 subtests. The rule is checked in both directions - it reported 137
offenders before the resolution and none after.
Not measured: 28 additional onion rounds cost matrix wall-clock, against swarm
jobs already capped at 355 minutes and one that has hit that ceiling before.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Every variants.yml header states the rule the bond feeds: variant 0 carries each
tightly-coupled (bond>=1, same-host) service, while loosely-coupled (bond<1,
separate-host) partners are excluded from the closure and distributed across
later variants to keep every host under budget. tor was declared bond: 1 in all
101 consumers, so it had to sit in every variant-0 closure.
It does not belong there. The onion is a separate-host provider reached over the
node's hidden service, the same shape as the bridged partners that already carry
0.5, and f1dff50fb established both the reasoning and the precedent when it
dropped xwiki and bigbluebutton from 1 to 0.5: bridged partners with their own
role and their own spec, whose bond declaration was the outlier rather than
their placement.
This also unblocks the variant-level tor contract. Requiring a non-baseline
variant to pin the onion on while tor still counted as tightly coupled would
have pulled variant 0's closure requirements along with it.
Verified: tests/lint 336 passed / 307 subtests and tests/integration/roles 107
passed / 610 subtests, including test_resource_budget over every variant's
deduplicated footprint and the bond-aware test_order closure.
Not measured, and the honest cost: f1dff50fb's own caveat applies unchanged
here. Whether redistributing a service out of the baseline closure buys back
wall-clock is not established; role-runtimes has the per-role seconds and no run
has been made with this shape.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The adapter was the only image in the tree pinned by a hand-maintained digest,
and the pin was never recomputed by anything. It drifted twice: run 31374959721
lost seven of eight jobs to mirror/docker.io/python@sha256:6771159c... not
found, and the renewal to the newest upstream digest earlier today would have
reproduced it, because builds resolve FROM against the GHCR mirror while the
mirror is copied by tag and lags upstream by design.
The base is now FROM <image>:<version>, taken from the sidecar service in
meta/services.yml - the only file iter_role_images scans, so the tag the build
asks for is by construction the tag the mirror was told to copy. The version
moves from the moving 3.13-slim to 3.13.15-slim, which upstream stops rewriting
once published, and which the existing version drift job renews like every other
image.
With the build reading services.yml, the adapter block's digest, image and
version had no consumer left. They are removed together with the schema check
for the digest form, the warn-only drift test, and the manifest_digest helper
that existed only for it. What remains in the block describes the adapter rather
than the image it is built from.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
A newly created swarm service starts every replica at the same instant -
update_config.Parallelism governs updates, not the initial scale-up - and each
process unpacks the prepackaged plugin into /mattermost/client/plugins, which
was an NFS export shared by all of them. The loser of that race aborts with
'unable to copy webapp bundle directory: chmod .../main.js: no such file or
directory', never activates the plugin, and answers 404 on its routes for the
rest of its life while the VIP keeps sending it a share of the traffic.
Measured on the live cluster: task IPs .30 and .32 answer 401 on the plugin
route, .31 answers 404; the shared directory holds the hashed bundle and the
LICENSE file but no main.js; byte-identical nanosecond mtimes on two nodes
prove one filesystem.
client/plugins is regenerated from the plugin at every activation, so it is
per-node state. plugins/ stays shared because uploaded plugin archives live
there, and sharing the generated bundle could never substitute for activation
anyway - a replica without the plugin in its process answers 404 regardless of
what sits in the directory.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Reverts the twelve pins to 9662417a, the digest the GHCR mirror currently
carries for python:3.13-slim. The previous commit renewed them to the newest
Docker Hub digest, which is the wrong registry: builds resolve FROM against the
mirror whenever a mirror prefix is configured, and the mirror is copied by tag,
so it holds whatever the tag pointed at when the sync last ran. A pin ahead of
the mirror resolves to nothing there while a local build keeps working - run
31374959721 lost seven of its eight jobs to exactly that, as
mirror/docker.io/python@sha256:6771159c...: not found.
The external check measured Docker Hub and would therefore have kept advising
the change that breaks the build. It now resolves the source the build uses via
mirror_image, asks whether the pin exists there at all, and only falls back to
upstream when no mirror is configured.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Every one of the eleven declared tools carried the placeholder schema
{"properties": {}, "type": "object"}. The adapter answers tools/list purely
from this contract, so each client saw parameterless tools: nothing could know
that read_post needs a post_id, and the call failed upstream instead of being
formed correctly. The schemas here were measured against the running upstream;
list_agents keeps an empty property set because upstream declares none.
Both pins of the file follow, as the adapter rehashes the contract at startup.
Add a lint binding the declared surface to the shipped one. Only the allowlist
is an equality: writer_allowlist is a superset by construction, because
passthrough.py rejects a whole contract as soon as one tool declares
mutating: true while mutations are off, and upstream_serves is an observation
that may legitimately exceed what the platform allows.
The suppression lookup both MCP lints need moves into suppress.py rather than
being written twice.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Three fields pinned artefacts and nothing recomputed any of them.
The adapter base digest was stale for all twelve provider roles: the tag
python:3.13-slim moved to ffb752e1 while every role still pinned 9662417a.
The GHCR mirror copies by tag, so the pinned manifest is one it was never
asked to copy - a mirrored build resolves FROM to nothing while a direct
build keeps working.
adapter.specification_sha256 matched no revision of the file it pins in ten
of twelve roles; three of the old values were byte-identical to the same
role's tools.schema_sha256, so they had been copied rather than computed.
The field is enforced nowhere at runtime, which is exactly why it needs a
check.
Add a hard lint for both offline pins, computing the tool digest with the
adapter's own policy.schema_digest so it cannot drift from the code that
enforces it in production, and a warn-only external check that reports a
base digest the upstream tag no longer resolves to.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
tools.allowlist meant two things at once: the surface a consumer may reach,
which is a decision of ours, and the surface the upstream offers, which the
vendor changes without asking. Conflating them made a version bump read as a
contract violation, and left no place to record what was actually measured.
Split them. allowlist stays policy; upstream_serves records the observation at
supported_version. State the second only where it DIFFERS, because a provider
that serves exactly what is allowed would otherwise carry the same list twice
and the copies would drift apart. Absence means identical.
The difference between the two measures unenforced exposure: mattermost 103,
gitlab 2, and zero wherever the fields collapse. A non-empty difference is what
tells us a provider needs a gateway in front of it.
Require a non-empty allowlist for an enabled, non-adapter, server-capable
provider. An empty one used to make the probe's comparison vacuous, which read
as a pass while every tool the upstream serves stayed reachable. Nextcloud was
in exactly that state.
Three exemptions, all semantic: an adapter authors its tools here against a
REST upstream, so there is nothing to observe; a provider declared
enabled: false is never admitted; and upstream_serves: dynamic covers an
upstream composing its surface at runtime, which nextcloud does by proxying
MCP servers an administrator configures.
Reject an allowlist naming a tool the upstream does not serve. That found
mattermost's add_user_to_channel, gone since 11.10.0 renamed it
add_channel_member, which would have been advertised and never callable.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Being an MCP client was enough to join a provider's overlay, so a single
admission anywhere reached every provider that opened one. Attachment now
requires both that the role speaks as a client and that the provider
lists it in mcp.allowed_consumers.
consumer.kind accepts a list so one overlay can carry both its
service-flag consumers and its MCP clients, which is what the vector
store needs.
The block sat under services but never behaved like a service. A service has
exactly one providing role, declared with provides, so the registry can map the
key to that role and consumers bind their flag to its presence. No role declares
provides: mcp. There are seventeen providers and the pairing is done by
allowed_consumers inside the block itself.
Every rule that enforces the service model therefore needed an exemption for it:
a shared-provider suppression in fourteen roles, a playwright-service-flag
suppression, and two freshly added exemptions in the dynamic-flag and
variants-static-override checks. Four independent rules, four escape hatches,
one cause. They are all gone now, together with a bond key that only the
resource budgeter reads and only for entries under services, and a wordpress
addon bridge that pointed at a service block that no longer exists.
The block moves to meta/mcp.yml. The applications loader maps meta/<name>.yml to
applications.<app>.<name> generically, so the move needs no wiring; the file
takes the same shape as meta/rbac.yml. Every config path follows, the variant
overrides are lifted out of their services map, and the vocabulary module leaves
the services package because it no longer describes one. roles_with_service
gains an explicit topic parameter rather than an implicit fallback, so nothing
silently resolves against the wrong container.
Files that are otherwise unrelated to this move are included where the committed
tree would otherwise keep pointing at a path that no longer resolves: thirteen
vars files, eleven playwright env templates and one task file.
Four roles declared auth_subject: user while handing the client an
account-bound credential. Nextcloud and Home Assistant mint against the
administrator account, Mattermost issues the administrator's personal access
token, and Baserow creates a dedicated mcp@<domain> account with its own
workspace. The declarations now name what the minting tasks do; Baserow's
README additionally claimed the administrator owns its endpoint, which it
never did.
The tools declarations follow one convention instead of two: the field records
the deployment's intent and the README names the missing enforcement, which is
what GitLab already did and what Baserow and Mattermost now do.
The discovery lookup carries auth_subject through so clients can see the
subject. No consumer branches on it.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Eight server roles declare an `mcp` RBAC role, so the platform mints
roles/<app>/mcp and the administrator joins it. Open WebUI now honours those
groups instead of flattening them to user/admin, and each MCP tool server is
restricted to the group of the role that serves it.
The grant cannot be rendered: Open WebUI matches it against a group id that
insert_new_group mints as a fresh UUID, so tasks/01_mcp.yml resolves the group
after the container is up and writes the grant through the admin API, the way
every instance-minted credential in this repository is provisioned.
Entries render disabled. An empty access_grants is not 'nobody', Open WebUI
reads it as every administrator, and ENABLE_PERSISTENT_CONFIG=false makes the
env authoritative again on every start. Enabling only together with the grant
keeps the fallback at no access rather than too much.
The provisioning account is the role-internal openwebui-api-bot, never a
platform identity: OAUTH_MERGE_ACCOUNTS_BY_EMAIL ships disabled and an OIDC
login whose email a local account already holds is refused, so seeding a
person's address would lock them out.
Known limits are documented rather than claimed away: a bare restart leaves MCP
off until the next deploy, a user's roles list carries no application context so
roles: [mcp] grants every deployed mcp group at once, and Open WebUI drops a
stale membership only while the groups claim is non-empty.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
