A redeploy's new process ticks at once, so its once-a-minute samples never
broke for 90 s and get_anomaly showed none of the day's restarts. The first
self-monitor tick in a process now records restart_gap_secs (time since the
last process's last sample), and steward_gaps shows each one as a restart.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Each source in an anomaly's `around` was cut to its oldest 40 with
nothing in the reply to say so; on a busy host the events nearest the
occurrence were the ones dropped. Each source now keeps the 40 nearest
the occurrence's start, and `around_left_out` counts the rest. Sources
return everything their bounded query read (Docker and the host agent
no longer cut at 40 and 20 themselves), so the count is honest.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
All 16 Traefik 4xx anomalies on 2026-10-15 came back with nothing
around them, though 15 were Steward's own restarts. A router has no
host, and every events source was read by host.
- Core: each break of more than 90 s in Steward's once-a-minute health
sample is a "steward" event beside every anomaly, placed or not.
- Plugins: an optional mcp_resource_events(db, source, resource, since,
until) for a resource with no host. Traefik's gives the container
events of the swarm service behind a router, on any host.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
still_failing was handed the query's one-pass result. The first failed
run's any() consumed it, so a second failure in the window found no
successes left and was reported though a later run had put it right
(ci-runner-volumes came back once keepalived-survey also had a failure).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
get_overview listed every failed or interrupted run of the last 24 hours,
so a dry run that failed, was fixed and re-run green kept warning for a
day. A failure now drops out once a later success of the same playbook,
source and hosts is at least as real: an applied success clears either
kind of failure, a dry-run success only a dry-run failure.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
get_anomaly gives each occurrence against what is usual at that hour, the
metric around the newest one with its normal, what else happened on the
resource's host around each occurrence (alert transitions, Ansible runs,
other anomalies there and a count elsewhere, plus plugins' own events
through a new mcp_events hook: the host agent's maintenance jobs and
downtime, Docker's container starts, deaths and OOM kills), and how earlier
anomalies on the same metric were judged.
get_metric_normal reads the learned slots and the level that would be
flagged. review_anomaly takes one id or several, all or nothing, with a
note. get_anomaly_settings and (admin) set_anomaly_settings read and change
the review window and each normal's floor, as Settings → Thresholds does.
metric_series.points takes an optional `until`; core.alerts gains
host_of_resource; floor_setting names its floor.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
list_backups, check_backup and make_backup are Settings → Backups for an
agent, admin only like the page. make_backup runs the page's job in the
background and is audited as "<user> via MCP". Restoring stays in the
browser. get_overview names a scheduled backup that failed, is two hours
past its slot, or has no bucket.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
#5748 "MCP can write, edit and test this instance's own playbooks
(steward-local)": two admin tools beside run_playbook, so an agent can
build automation for these hosts that isn't part of Steward itself.
- save_playbook(playbook, content, source="steward-local", replace=false)
goes through the editor's own validate_playbook_yaml and
write_playbook. A playbook that isn't a list of plays comes back with
the YAML error and its line, and nothing is written. An existing file
is overwritten only with replace=true. steward-builtin is refused
(Steward's playbooks change in its repository), as are git sources
(their repo owns them).
- After writing, `ansible-playbook --syntax-check` runs against a
throwaway localhost inventory (no host contacted, 60 s), so a module
name Ansible can't resolve shows before any run. The result rides the
reply; the file is kept either way.
- delete_playbook(playbook, source="steward-local").
- Both audit as ansible.playbook.created/updated/deleted, "<user> via
MCP", with the size and never the content.
The testing loop is the existing one: run_playbook defaults to check
mode, on one target, then get_ansible_run. The integration test does
that loop with the real ansible-playbook.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The root helper gains one read-only action, `drives`: it finds the drives
itself (smartctl --scan-open), reads each as JSON, and writes a summary the
agent carries on every sample (agent 1.20.0). Steward judges the figures in
drives.py (failing / worn / watch / ok / no SMART / unreadable), records them
per drive and as host roll-ups for covering rules, keeps the newest report on
the registration and lists it on the host page.
install-deps.sh installs smartmontools with apt, dnf, yum, zypper, pacman or
apk, run by install.sh and every agent playbook; one it can't install is
reported, never fatal.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
VDNT-Util01's / ran out of inodes (956 free of 60.8M) with 274G of space
left; every file create failed and Steward's page read healthy at 66%.
- Agent 1.13.0 reports each mount's inodes (statvfs; a filesystem reporting
none, like btrfs, is left without).
- Ingest records inodes_used_pct per mount, and disk_used_pct_worst is the
fullest mount by space or inodes, so existing disk-full rules and the
overview see it.
- The host's health strip, the Hosts widget and the host glance grade / by
whichever is fuller and name inodes when they are ("Disk / 66% · inodes
100%", "Inodes on / ... past its critical threshold").
Hosts pick up the inode figures once their agent is updated to 1.13.0.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
VDNT-Docker03's silent agent read as three overview findings (agent offline,
agent stale, the disk rule's no-data) and every drop emailed the disk rule's
silence and return on top of agent offline.
- The host agent registers its heartbeat (seconds_since_seen) and the
sources it carries (host_agent, docker). While a host's agent is past its
offline window, its other rules' no_data is recorded "while its agent is
offline" and not emailed, nor is their quiet return; a return past a
threshold still fires.
- The overview folds those rules under the host's agent-offline finding
("also silent with it: ..."), and a stale agent whose agent-offline alert
is active is that alert's finding, not a second one.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Host CPU, memory and load per core, container CPU and memory, and every
monitor's response time learn a normal per hour (milestone 513, System 182).
A reading past its fixed cutoff but within its normal reads healthy and says
what is usual at this hour; one well out of it keeps its grade and is a
metric anomaly. Disk, a monitor down and an agent offline keep their cutoffs.
- Each normal's floor, the smallest departure that counts, is a setting:
Settings -> Thresholds -> Normals (norms.floor.*). Defaults: host CPU 10
points, host memory 5, load per core 10, container CPU 10 points of one
core, container memory 5, monitor response 100 ms; Traefik's 2-point 4xx
and 50 ms p95 floors move there from constants. learn_normal takes a
floor function (floor_setting) read at each pass.
- learn_normal(parts=False): a host learns, its cores don't.
- Ingest records load_pct (load_1m / cores x 100), the figure pages grade.
- Shared grading: anomalies.against / graded, worst_reading(normals=),
hosts' host_grades, docker usage(thresholds=, norms=), monitors'
latency_grade. Applied to host vitals, the Hosts widget, MCP overview,
Docker usage, and the monitor list, widget and page.
- The monitor and container pages list their anomalies.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
#5740 "A metric that departs from its normal is recorded as an anomaly to
review, and is findable after it ends", milestone 513.
- Every minute each learned key's last 15 minutes is judged against its
hour's normal: above the usual hold x1.25 and by the metric's floor opens
an occurrence, which ends when the level comes back.
- metric_anomalies holds one row per item to review; a repeat while it
still waits joins it as an occurrence (metric_anomaly_spans).
- An ended row nobody reviews within anomalies.review_days (14, Settings ->
Thresholds) lapses; an ongoing one never does.
- Learning skips every hour an occurrence covers unless its anomaly was
marked normal: no normal is learned without review.
- Alerts -> Anomalies lists those to review first, then the reviewed and
lapsed under day dividers; marking is operator/admin and audited.
- The Overview widget and get_overview warn of open ones and count those
to review; the host page shows its own; the weekly report lists the
week's and the lapsed; MCP list_anomalies and review_anomaly.
- Reviewed and lapsed rows are pruned at data.retention_days, and kept
in backups like alert history.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
#5509 "Delivery you can see: each alert's email state in history, and
outgoing email on Settings -> System" (milestone 501).
- Alerts -> History: an alert's email reads from the outbox (#5506):
"gathering" with when it sends, "sending", "retrying" with the attempt,
next try and the server's last answer, "sent", or "failed" naming each
reader it didn't reach and why (outbox.delivery / email_states). Events
from before the outbox keep what they recorded. The Notified column
shows once an alert has a queued email.
- Settings -> System gets "Outgoing email": every email of the last 7
days not yet sent, failures first, then the longest retrying, with
recipient, kind, state and last error, and a count of what was sent.
Failures and anything still retrying after 5 minutes lead the page's
verdict (outbox.outgoing).
- get_alert_history (MCP) adds `email`: the same delivery state.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
#5487 "MCP: edit inventory group membership". Group membership decides which
hosts a group alert rule covers, and the MCP could only read it. The new
operator tool takes a group and lists to add and remove, each a target's
name or id or the Steward host it is linked to; an unknown name refuses the
whole change. It answers with the group's targets after and the alert rules
covering the group (whose per-host rules follow within a minute), and audits
as ansible.group.updated by "<user> via MCP", as the Inventory page does.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
#5466 "Send test: check a rule's delivery from the rule". A rule's page gets
a Send test button (operator role), and MCP gets send_test_alert. Each sends
a TEST alert for that rule now, by email to everyone who gets alerts and by
webhook. The alert carries the rule's name and condition, at the latest
reading of what it watches; a rule covering hosts uses its first host. The
result for each channel shows under the button, or "Nothing is set up to
send to". Nothing fires, no event is written, and the test is audited as
alert_rule.tested. An alert with no reading now says "has no reading yet"
instead of "is None".
Docs: alerting.md also gains the burst email from #5465 "One email per burst
of alerts".
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
#5464 "Group rules: one rule covers every host, or every host in a group",
step 2 of milestone 499 "Alerts you can trust". Each rule named one host, so
the live instance carried six rules for three hosts, and a new host had no
agent-offline or disk alert until someone copied them.
- alert_rules.covers / cover_group / parent_id (migration 0041). A rule on
host_agent may cover "all_hosts" or a "group" (an inventory group: the
Hosts ruling, #5294). A covering rule has no state; Steward keeps one
ordinary child rule per covered host, named "<rule>: <host>", so each
host's alert keeps its own state, history and acknowledgement.
- alerts.service.reconcile_children: a covered host that has reported in
the last 7 days gets a child; a host no longer covered loses its child;
children take the parent's condition, checks, notify setting, action and
on/off. Run on save and each minute (alert_reconcile).
- A child is changed through its parent: edit, toggle and delete of a child
are refused (409 / PartOfCoveringRule); its edit page opens the parent.
- Rule form: a Covers choice for host sources, with the group picker; fixed
once the rule exists. Rules list: a covering rule reads "every host in
swarm · 3 hosts" with its hosts' rules nested under it, each opening its
history. The history picker and the rule count leave covering rules out.
- MCP: create_alert_rule takes covers/group; update_alert_rule takes group
and passes changes down; list_alert_rules shows covers, hosts and part_of.
- docs/core/alerting.md: "Rules that cover hosts".
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
#5463 "Acknowledgements on record: an event and who did it", step 1 of
milestone 499 "Alerts you can trust". The live walk in #5461 found that
acknowledging changed only the alert's current state: once it resolved,
nothing said who had acknowledged it or when.
- alert_events.actor_id (migration 0040): the person behind a transition.
- Acknowledging writes a firing → acknowledged event by the acknowledger.
- A rule edit, disable or enable that resets an alert carries its actor
(web and MCP alike).
- The history page notes "acknowledged by X" and "rule disabled by X"; MCP
get_alert_history returns `by`.
- One usernames() helper in alerts.service, used by the active list, the
history page and the MCP tools.
- docs/core/alerting.md: the state diagram no longer shows an acknowledged
alert re-firing, and says what an event records.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
check_interval_s was `m.check_interval_seconds or None`, so a monitor
following monitors.poll_interval_seconds read as null and a reader
couldn't tell how often it ran. Monitor results now carry the effective
interval and `interval_from: "monitor" | "global"`.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
DNS monitors take an optional Resolver, a server's IP. Set, the check asks
that server directly for A and AAAA over UDP and records how long it took
to answer; empty keeps today's lookup through the container's resolver.
"Does my local DNS still answer for public names" and "does 1.1.1.1" are
now separate monitors. No new dependency.
On a host's page the Ansible card moves above the Docker list, which can
run to dozens of containers and hid it.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
#5212 "Plain labels: no raw ids, scopes, metric keys or unexplained jargon",
batch 2:
- A run records where it came from: AnsibleRun.origin ("schedule", "alert"
or "mcp") and origin_id, set by the scheduler, Run now, alert actions and
the MCP tools. Migration 0031 backfills past runs from alert events,
schedules' last runs, and MCP audit entries.
- The runs list has a Started by column: "schedule nightly", "alert rule
nas down", "alice via MCP" or "alice", linked where there is a page.
The raw scope tooltip is gone.
- Run detail: no run id in the subtitle or body, Started by in words, and
the host summary lists failed and unreachable hosts first. The live
output only follows new lines while the reader is at the bottom, and
when the run finishes the page reloads once so the badge and summary
are final.
- list_ansible_runs and get_ansible_run report origin and origin_id.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- One run per host: start_run claims every host it touches (addresses read
from the inventory Steward wrote; a source's own inventory file is claimed
as that file) before it takes a run slot. The claim is all-or-nothing. A run
waiting on a host shows queued and names the run it waits for, honours a
cancel, and fails after HOST_WAIT_SECONDS (1 h) with which host and run.
- A due schedule whose previous run is still queued or running skips the
firing: last_run_at moves on, last_run_id still names the run going, and
last_error says "Skipped: the previous run was still going". The Schedules
page shows that beside the run (muted), and a launch error in red.
- An alert rule that fires while the action run from an earlier firing is
still going starts no new run; the event says so and names the run. An
in-process set closes the gap between the check and the run being linked.
- docs/core/ansible.md: queued/cancelled statuses and the overlap rules.
Tests: tests/core/test_ansible_overlap.py (claim keys, waiting, all-or-nothing,
cancel, deadline) and tests/integration/test_ansible_overlap.py (two real
runs on one host take turns; schedule skip then fire; alert action skip then
start).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
#5066 "Alert history page". /alerts/history lists every alert transition
in a range (1h–30d, the shared range picker), newest first, all rules or
one (?rule=). Each row: from → to, the value against the rule's condition
("no data" for a silent one), email and webhook sent or failed with the
error, "muted · <window>" (and "notified when the window ended") taken out
of the event note, other notes, and a link to the Ansible run an action
started. Range and rule changes swap only the rows (HTMX). Viewer role.
The Alerts page links each rule's history and the whole history. The query
moves to alert_svc.history, which MCP get_alert_history now uses too.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
#5065 "Renames carry their alerts". Alert rules, maintenance-window scopes
and metric history key a thing by (source, name), so a rename used to leave
the rules watching a name nothing reported under, and split the history.
- core/renames.carry_rename moves AlertRule.resource_name,
MaintenanceWindow.scope_resource, plugin_metrics and plugin_metrics_hourly
to the new name in the caller's transaction. A host rename also carries
its parts: host_agent "<host>:…" and docker "<host>/…" (HOST_SOURCES).
Matching is exact (a prefix compare, not LIKE or a collation range).
History left under the new name by something deleted is dropped first
rather than merged in.
- Web monitor edit, MCP update_monitor and web host edit call it; MCP
answers with `carried` (it used to return a warning naming stranded rules).
- A name another monitor of the same type, or another host, already has is
refused on create and rename: two would share one history.
- The rename is audited (monitor.updated / host.renamed) with what moved.
Tests: an integration test renames a firing monitor through the web route
and finds its rule still firing and cleared by a sample on the new name,
its window and raw+hourly history moved and an orphan's history dropped; a
host rename moves agent and container parts but not a longer-named host's
or another source's; a guard that every host_agent resource is "<host>" or
"<host>:…".
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Agent offline is an ordinary rule now. The host_agent scheduler records
host_agent/<host>/seconds_since_seen every minute (from last_seen_at, or
registration time for an agent that never reported), replacing the task
that only logged stale agents. The catalog offers the metric, and picking
it on a new rule fills "> 180".
No data: each rule's state tracks its last sample and its usual gap,
smoothed so that an outage counts as at most three gaps. A core task,
alert_silence_check, runs every 60s and marks a rule no_data_since once its
series has had no sample for three usual gaps (at least 5 minutes; an hour
while the cadence is unknown). A never-sampled rule counts from when it
started watching. The mark sits beside the state, so a firing alert that
goes silent stays firing. Each silence records one event and sends one
notification (rule option notify_no_data, default on; muted in maintenance
windows). The next sample clears it with a "data resumed" event and
notification. Creating, changing, disabling or re-enabling a rule restarts
its watch.
Migration 0029 adds the columns and seeds each existing rule's last sample
and gap from its last day of data, so an upgrade does not flag every rule.
Shown on the Alerts page (a NO DATA badge with its start time), in MCP
list_alerts, list_alert_rules and get_host (state "no_data", no_data_since),
in get_overview findings (alert_no_data, or a firing alert marked silent)
and in the weekly report. The rule form has the notify option, and its
metric reference now matches what traefik records and lists host_agent.
Emails say "no data" and carry the event's note.
The metric-catalog guard also reads positional MetricBatch.add arguments.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
#5061 "Metric catalog belongs to what emits: plugins register their
metrics; names match what is recorded".
- The static METRIC_CATALOG is gone. metric_catalog() is core's own sources
(the four monitor types and "steward", the app's open_fds/open_fds_pct,
which were recorded but never offered) plus each loaded plugin's
get_alert_metrics() export. A disabled plugin's sources leave the picker.
- traefik offers what it records: error_rate_4xx_pct and error_rate_5xx_pct.
The old "error_rate" was never recorded, so a rule on it could not fire.
- tests/core/test_metric_catalog.py walks every MetricBatch.add /
record_metric call (traefik's tuple loop included) and fails on a recorded
metric the catalog lacks; it shows on a synthetic writer that it can fail.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
#5059 "Alert state machine: acknowledgement holds until resolve; N
consecutive OK to resolve; editing, disabling or deleting a rule resets it".
- The state machine is now next_step(), a pure function of one sample. An
acknowledged alert stays acknowledged while breached; before, the next
breaching sample re-fired it. Firing and acknowledged alerts resolve after
consecutive_ok_required OK checks in a row (new rule column, default 1;
rule form, rules list and MCP create/update_alert_rule clear_after_checks).
- Changing what a rule watches (source, resource, metric, operator,
threshold) or disabling it resets its alert to inactive; a firing or
acknowledged one records a resolved event whose new note says why, and
nobody is notified. Web and MCP edits share alerts.service.apply_rule_change.
- Migration 0028: consecutive_ok_required, consecutive_ok_count,
alert_events.note, and metric_value nullable (a reset has no sample).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
#5034 "Playbook runs: path and inventory checks in trigger_run, connection
vars admin-only, confirm enforced server-side, web actions audited".
Ansible runs inside the Steward container, so an operator who could name
any playbook path, set ansible_connection=local or ProxyCommand ssh args,
or put {{ lookup('pipe', …) }} in a value could run commands on Steward.
- steward/ansible/guards.py: check_playbook (the source must list it),
check_inventory_scope (steward:* or an inventory file the source has),
check_vars_for_role (ansible_* beyond host/port/user/password/become,
and any Jinja in a value, are admin-only), changed_vars (an edit is
held to the role only for what it changes).
- trigger_run applies the path checks, so every run path — web, MCP,
alerts, schedules, plugins — gets them.
- Variables are checked by role on the web run form, host-page runs,
schedule create/update, inventory group/target create/update, and MCP
run_playbook / create_schedule / update_schedule.
- A steward:confirm playbook needs the confirmation ticked on the server
side for a real (non-check) web run, as MCP already required.
- Audited: runs, cancel, schedule create/update/enable/disable/delete/
run-now, playbook create/update/delete, inventory group and target
create/update/delete (variable names only, never values). One
audit_request helper serves the web routes.
- The CSRF form guard and the route guard now each test that they fail
on a synthetic broken input (rule: a guard must be able to fail).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
#5015 "Built-in host-agent playbooks make an MCP caller supply steward_url"
- run_playbook, create_schedule and set_alert_action fill a steward-builtin
playbook's steward_url when the caller leaves it out or empty, from
Settings → General → Public base URL, else the scheme and host the MCP
request arrived on — the same source the web actions use.
- run_playbook's reply names what it filled (filled_by_steward), and
get_playbook shows that value as the variable's default.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
#5014 "Ansible run scope is shown as a raw id"
- inventory_gen.describe_scope / scope_labels turn "steward:target:<uuid>"
into "target web-1", "steward:group:<uuid>" into "group Docker Swarm",
and a deleted one into "target abcdef01 (deleted)".
- MCP run and schedule rows carry scope_label beside the stored scope; the
ansible.run.started audit detail records it too.
- The web run page and schedules table show the name, with the stored
scope as the tooltip.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
#5012 "get_host cuts `parts` without saying so"
- The agent's sub-resources come back as mounts (fullest first),
interfaces and disk_io (busiest first) and temperatures (hottest first),
each bounded on its own with total, truncated and a `more` that names
parts_limit and query_metrics.
- Per-core CPU is a summary (count, min/avg/max, three busiest), so 38
cores no longer push every interface out of the reply.
- New argument parts_limit (default 20, max 200).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
MCP-7 #4971 "Monitoring-config write tools: monitors, alert rules, acks,
maintenance".
The validation the forms did inline (or not at all) moves to the services,
and both the routes and the tools call it, so bad input is refused the same
way with the same reason:
- monitors/service: monitor_fields, build_config, normalise_target. Port
1-65535, status 100-599, timeout 1-60, method GET/HEAD/POST, headers an
object of strings, a key the type does not take is named. The forms now
answer 400 with the reason instead of redirecting silently or a 500.
- alerts/service: METRIC_CATALOG and SOURCE_MODULES, resources_for,
metrics_for, rule_fields (source, resource and metric from the same
choices the rule form offers), add_rule, acknowledge, window_fields.
Tools: create_monitor, update_monitor (config merged over the stored one;
a rename warns about rules still watching the old name),
set_monitor_enabled, create_alert_rule, update_alert_rule,
acknowledge_alert, create_maintenance_window (duration or end),
end_maintenance_window (ends one in progress, cancels an upcoming one) —
operator. delete_monitor, delete_alert_rule — admin. get_alert_rule_options
— viewer. Every write is audited as "<user> via MCP".
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
MCP-6 #4970 "Automation tools: playbooks, runs, schedules, alert actions".
Viewer: list_playbooks, get_playbook (secret defaults masked), list_inventory
(variable names only), list_ansible_runs, get_ansible_run (log tail + grep),
list_schedules, wait_for_run (polls, 120s cap).
Operator: run_playbook (check mode unless check=false; a steward:confirm
playbook applies only with confirmed=true; secret variables reach the run
but are never stored, audited or returned), cancel_ansible_run, and
create/update/delete_schedule (secrets refused, check by default).
Admin: set_alert_action, and the docker plugin's prune_docker, which
previews from the disk report unless check=false (Ansible's check mode
skips the prune commands, so a check run would show nothing).
Every write is audited as "<user> via MCP" with the token id; the audit
tool's user filter includes those rows. prune_extra_vars moves to the
docker service so the page and the tool share it.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
MCP-5 of milestone 461 "MCP server: Steward as an LLM-inspectable,
LLM-operable hub". Contributed through each plugin's register_mcp_tools
hook, so they exist only while the plugin is enabled.
docker (viewer, read-only):
- list_containers: filter by host, status, compose project, or problem
(restarting, crash_looping, unhealthy, oom_killed, exited_nonzero);
problems sort first and each row names what is wrong.
- get_container: state, CPU/memory history across the raw/hourly seam,
lifecycle events. A unique part of a swarm task name is enough.
- list_swarm_services: replicas per node (agent-less placements as ghosts)
and the swarm's nodes.
- search_container_logs: text (% and _ literal), stream, range, context
lines, paging by a full-precision `before` cursor; over-long lines keep
both ends and say how much of the middle was cut.
- get_docker_disk_usage: bytes per host, reclaimable first.
- mcp_overview: failing containers and under-replicated swarm services
join get_overview.
host_agent: list_agents (reporting / stale / pending, version, OS), and an
overview finding for agents registered over 30 minutes ago that never
checked in.
Settings > Thresholds gains the TLS-expiry warn/crit days (were 14/3 in
code) and the crash-loop deaths-per-hour count (3), which get_overview
and the docker tools now read.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
MCP-4 of milestone 461 "MCP server: Steward as an LLM-inspectable,
LLM-operable hub". Twelve read-only viewer tools:
- get_overview: what is wrong now, ranked critical-then-oldest. Failing
monitors with their error and when the run of failures began, firing,
pending and still-breaching acknowledged alerts, TLS certificates expiring
within 14 days, agents that stopped reporting, hosts over the configured
CPU/memory/disk thresholds (Settings > Thresholds), failed Ansible runs in
the last 24h, and findings from enabled plugins through a new optional
`mcp_overview(db)` plugin hook. Each finding names the call that shows
its evidence.
- list_hosts, get_host, get_host_metrics
- list_monitors, get_monitor (HTTP header values and URL passwords are
redacted)
- list_alerts, get_alert_history, list_alert_rules, list_maintenance_windows
- query_metrics: the whole metric store, narrowed one argument at a time
(sources > resources > latest values > a series summary)
- list_audit_events
Shared shaping in steward/mcp/tools/_shape.py: UTC ISO times with age_s,
bounded lists that say when and how they were cut, series as
min/avg/max/p95/last plus a ≤60-point series, lookup by id or
case-insensitive name with a did-you-mean, and a range argument that
refuses an unknown key instead of reading it as 24h.
New core reads in steward/core/metric_series.py (sources and resources
via loose index scans, latest values via DISTINCT ON, series across the
raw/hourly-rollup seam). monitors/service.py gains monitor_states,
down_since and recent_failures; agent_overview_by_host now carries the
fullest disk and the last sample time.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Second step of milestone 461. An MCP client (Claude Code, Claude Desktop
via mcp-remote, Open WebUI) connects to /mcp with a personal API token.
- mcp>=2.2 (SDK v2 MCPServer). Mounted by wrapping app.asgi_app, so /mcp
never enters Quart's pipeline; stateless streamable HTTP so a redeploy
never strands a client on a session id the new process didn't issue.
- Auth: Authorization: Bearer stw_… resolves to a Caller (user +
effective role) in a ContextVar; 401 with WWW-Authenticate otherwise.
The web pipeline's database gate runs first, so a DB restart answers
503 instead of failing inside a tool.
- Registry, default-deny: register_tool(fn, role=, read_only=) is the only
door; role and read_only are required. Tools above the caller's role,
or from a disabled plugin, are neither listed nor callable. Undeclared
arguments are refused (not silently dropped); a tool's ValueError
reaches the agent as its reason; a DB failure mid-tool says so.
- Plugins contribute tools via an optional register_mcp_tools(register)
export, all-or-nothing, replaced on hot-reload. Documented in
docs/plugins/writing-a-plugin.md.
- First tool: whoami.
- Tests drive app.asgi_app directly (Quart's test client never reaches
/mcp); the integration test mints a real token against Postgres and
runs initialize → tools/list → whoami → revoke → 401.
Task #4966 "MCP-2 — /mcp transport".
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
