#5031 "Forged-request protection: SameSite cookie, per-session CSRF token
on every non-GET, HTMX header". A page on another site could make a
logged-in browser POST to Steward with the session cookie attached.
- steward/auth/csrf.py: a per-session token, checked in one before_request
for every POST/PUT/PATCH/DELETE (header X-CSRF-Token or form field
csrf_token, constant-time compare). The bearer-authenticated host-agent
ingest is exempt; /mcp never reaches Quart.
- Every POST form renders {{ csrf_field() }}; base.html carries the token
in a meta tag and adds it to every HTMX request; the dashboard's fetch()
calls send it.
- The session cookie is SameSite=Lax; login issues a fresh token.
- PUBLIC_BASE_URL is now loaded from Settings at startup (it was only set
when Settings → General was saved).
- tests/test_csrf.py: refusal, acceptance, rotation, the ingest exemption,
and a guard that every POST form in the templates carries the field.
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>
#5010 "Logs of a removed container cannot be reached by name"
- get_container and search_container_logs fall back to the names Steward
still holds logs or events for when no live container matches: the exact
name from logs and events, or a part of it that matches one removed
container's events.
- get_container answers with removed: true, last_seen_at, its history and
events; search_container_logs marks container_removed and notes when it
was last seen, and says when that is older than the range.
- A miss now says the container may have been removed, that Swarm replaces a
dead task under a new name, and where to look instead.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
#5013 "Overview findings for containers and Swarm services carry since=null"
- Container problems are dated from their own evidence: a crash or restart
loop from its first death in the hour, an OOM kill or error exit from the
death that ended it, a failing healthcheck from its last turn to unhealthy.
list_containers and get_container carry problem_since / problem_age_s.
- Swarm services carry trouble_since: a crash loop's first failure in the
hour, a short service's oldest troubled task in each failing slot, a
recovered service's first failure in 24h. list_swarm_services shows it.
- get_overview findings from the docker plugin now have since and age_s,
so they rank by how long they have stood.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Milestone 465 "Swarm task awareness". Fixes three defects from the live MCP
review #4973:
- #5008 "Swarm task history read as live state"
- #5009 "Crash-loop detection is blind to Swarm services"
- #5011 "Swarm task errors are not captured"
Agent and ingest (agent 1.8.0):
- A manager's agent now reports each service's newest tasks: slot, node,
desired and current state, Swarm's Err/Message, exit code and time.
- They are stored in docker_swarm_services.tasks_json (migration
docker_010_swarm_tasks).
Swarm view (pure):
- A task container is "replaced" when Swarm set its task's desired state off
`running`. Without the task on record, a newer or running task in the same
slot decides.
- Replaced containers are no longer replicas or crashes.
- Each service gets, from the task records:
- failures_1h and failures_24h (failed or rejected tasks);
- crash_looping (Settings → Thresholds count);
- last_error, in Swarm's own words.
MCP:
- New findings: swarm_service_crash_looping (critical) and swarm_task_failed
(warning, on a healthy service with a failure in 24h).
- Down and degraded findings name the last task error.
- list_swarm_services carries the failure fields.
- Container rows mark swarm_replaced and give a replaced task no problems.
Web:
- The containers page swarm cards and the Swarm topology page show crash
loops, recent failures and the last task error.
- The dashboard widget flags a crash loop, and its "failed (24h)" count no
longer counts a rolling update's killed task.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
MCP-8 #4972 "Connect card, docs, and end-to-end integration test".
- AI access page: a "Connect a client" card with the /mcp URL (from
Settings → General → Public base URL, else the request host) and recipes
for Claude Code, Claude Desktop (mcp-remote) and other Streamable HTTP
clients, plus the reverse-proxy note. The token-created page shows the
same recipes with the new token filled in, ready to paste.
- docs/core/mcp.md: connecting, the reverse-proxy requirements, the
security model (token ceiling, default deny, check-mode default, secrets,
audit), tools by role, response conventions, adding tools.
docs/architecture.md describes the /mcp dispatch outside the blueprint
pipeline; docs/reference/code-map.md gains an MCP section.
- tests/integration/test_mcp.py: over HTTP, a viewer token lists no write
tools and is refused one by name; an operator token reads get_overview
and its maintenance-window create/end land in the audit log as
"<user> via MCP" with the token id. (Revoked → 401 was already covered.)
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>
MCP-3 of milestone 461 "MCP server: Steward as an LLM-inspectable,
LLM-operable hub". The /mcp tools run outside Quart's request pipeline, so
the queries they share with the pages now live in plain functions that take
an open session:
- steward/hosts/service.py: uptime, status, monitors, agent metrics,
agent overview (stale threshold passed in, not read from current_app),
CPU sparks.
- plugins/docker/service.py: bucket, container_history, host_map,
list_containers, container_events, query_logs.
- steward/alerts/service.py: active_alerts (also used by core/reports),
list_rules, list_maintenance_windows.
- steward/monitors/service.py: monitor_rows (was routes._rows_data).
Routes are repointed and render the same data. One behavior fix rides
along: the container-log search now escapes % and _ so they match
literally instead of acting as LIKE wildcards.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The test client reported http.disconnect as soon as the request body was
read, and the SDK's SSE response cancels its stream on disconnect, so
tools/list came back 200 with an empty body. Model a real client: wait
for the response (with a 15s deadline), then disconnect.
The caller fixture reset its ContextVar from sync teardown, a different
context from the async test that set it; each async test already runs in
its own context copy, so the reset is dropped.
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>
First step of milestone 461 (MCP server). A token is the credential an
MCP client presents as `Authorization: Bearer stw_…` to act as its user.
- api_tokens table (0026): sha256 hash only, display prefix, role
ceiling, expiry, revocation, throttled last_used_at.
- steward/core/api_tokens.py: mint, list, revoke, resolve. Resolution
re-reads the user's current role and acts with the lower of that and
the token's ceiling, so demoting or disabling a user weakens or kills
their tokens with nothing to revoke by hand.
- /ai-access/ (sidebar → Automation → AI access): every user manages
their own tokens, can't grant above their own role, sees the token
once. Session-only: a share-link viewer is refused. Create and revoke
are audited.
- Integration tests drive the routes against real Postgres so their
transactions run for real (#884 escaped CI behind a mocked session).
Task #4965 "MCP-1 — Personal API tokens".
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The admin banner counts every plugin load failure, but settings/plugins.html
only rendered failure reasons inside the plugin card macro -- which iterates
discovered plugins. A plugin that failed BECAUSE it has no directory was
therefore counted in the banner and shown nowhere on the page the banner links
to. That dead end is what the operator hit: "1 plugin failed to load" on the
dashboard, nothing wrong on the details page, and no way to clean it up.
Folds the reason into the "Configured but not installed" rows rather than
adding a competing section, and drives that list from the UNION of stored
plugin.* settings and recorded load failures instead of stored settings alone.
A failure with no stored row -- a plugin enabled by a DEFAULTS entry, which is
exactly what plugin.http was -- would otherwise still be counted and still be
invisible.
Rows carry `removable`: only a plugin with a real stored row can be cleaned up,
so a default-declared one shows the reason and an explanation instead of a
Remove button that would delete nothing while the default reasserts it. Rows
with a reason read as an error (red, "Failed to load"); rows that merely have
leftover config stay a warning. A discovered plugin that failed is excluded --
its own card already shows the reason.
Replaces find_orphaned_plugin_names outright rather than adding a second
overlapping helper; it had one caller.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The operator's 500 was NOT the startup gap I attributed it to. They found it by
logging in and getting Steward's own error page -- which means create_app had
completed and the app was serving, so the gaierror came from a request handler
acquiring a connection, not from boot. Neither prior fix covers that: the
startup retry never runs, and pool_pre_ping only helps once the database is
back, since its replacement connect fails too while the container is gone.
Adds a before_request gate that acquires a pooled connection with a short
bounded retry (~1.75s over 4 tries) so a database restart is ridden out
invisibly, and renders a distinct 503 "database unavailable" page when the
budget is exhausted. The request budget is deliberately far shorter than the
startup one: nobody watches a container boot, but somebody is watching this
page load, and a page that hangs is worse than one that says what is wrong.
/health and static are exempt -- a liveness probe must stay answerable while
the database is down, or a dependency outage triggers a restart loop.
Credential and missing-database errors still propagate rather than being
reported as "unavailable", which would send the operator chasing the wrong
problem entirely.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Operator saw the `http` plugin -- deleted when ping/dns/http were unified into
the Monitor entity -- still reported as enabled, with no way to clear it.
It was never an orphaned row. `plugin.http` was hardcoded in core DEFAULTS, so
deleting the app_settings row did nothing: the default reasserted it on the
next settings load. That is precisely why no cleanup UI could have fixed it.
Rule 22 says the removed subsystem should have taken its setting with it.
Purged the surviving references -- the DEFAULTS key, "http" in
CAPABILITY_PLUGINS, the stale plugin_manager docstring example, and the
capabilities blurb still advertising HTTP/uptime as a bundled capability --
plus a migration dropping any stored plugin.http row. Untouched: `http` as a
MONITOR TYPE (icmp/tcp/dns/http) everywhere it appears, and http_001_initial,
which is kept deliberately so existing DBs resolve the revision graph.
For the general case, cleanup splits by provenance rather than being uniformly
automatic or uniformly manual:
Bundled plugins ship in the image and version atomically with core, so they
cannot be transiently missing -- a plugin.* default with no bundled directory
is unambiguously a bug. Guarded by a unit test that fails CI, which is the
only part safe to automate.
External plugins live in operator-mounted /data/plugins, where absence is
ambiguous: unmounted volume, failed install, mid-upgrade. Auto-deleting their
config would silently destroy unrecoverable credentials on a transient
condition, so Settings now lists them under "Configured but not installed"
with an explicit, confirmed, audit-logged Remove. The section hides entirely
when empty, and the remove route refuses if the plugin is actually installed.
Orphan detection reads STORED rows, never the DEFAULTS-merged view -- offering
to remove a default-only key would be a lie -- and returns names only, since
plugin config can hold credentials and listing orphans never needs their values.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Two connection-recovery gaps, both surfaced by a Postgres restart that left
the app throwing tracebacks while the DB itself was healthy.
pool_pre_ping + pool_recycle on the app engine [#2626]: when the database
restarts, every connection already in the pool is dead at the socket level.
SQLAlchemy only discovered that by failing a real query, so the first
operation after a restart errored out on whatever triggered it. pre_ping
checks liveness on checkout and swaps the dead connection transparently;
pool_recycle caps connection age so a socket stranded by a NAT/conntrack
timeout or a Docker network rebuild is retired on a timer instead.
wait_for_database() gate at startup [#2627]: create_app touches the DB
synchronously (migrations, secret re-encryption, settings load) and assumed
it was both resolvable and accepting connections on the first try. Neither
holds after a host reboot (Docker DNS not yet serving `db` -> gaierror -2)
or an unclean shutdown (Postgres still replaying WAL -> "not yet accepting
connections"). Both are transient, so retry with capped backoff behind one
gate ahead of the first DB touch. Credential and missing-database errors
are classified by SQLSTATE and still fail immediately -- waiting cannot fix
those. Budget is bootstrap-configurable (STEWARD_DB_CONNECT_TIMEOUT /
database.connect_timeout, default 60s) since it governs reaching the DB and
so cannot live in the DB-backed settings.
Tests drive the retry loop off a fake clock, so backoff and timeout
behaviour are deterministic rather than wall-clock dependent.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
test_persist_logs_respects_toggle_and_exclude read via s.execute(SELECT) between
two `async with s.begin()` blocks; that SELECT autobegins a transaction, so the
next begin() raised "A transaction is already begun". Move both reads after the
final begin block (the toggle-off persist stores nothing, so end-state still
reflects the exclude phase), and pass a SimpleNamespace stand-in host to sidestep
ORM attribute-expiry across the commit boundaries.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CAGR73DUowdVFVvYzLXC5C
The user-facing half (rule 27). Per-container log viewer + admin controls, so
container logs are something the operator can actually touch.
Viewer (plugins/docker):
- /container/<host>/<name>/logs full page + /logs/lines HTMX fragment polled
every 5s (near-live follow); stream filter (all/stdout/stderr) + case-insensitive
text search; newest-first so the latest lines survive the poll's scroll reset;
empty/loading states; a "View logs" link from the container detail page.
Jinja auto-escapes log content (no markup injection).
Settings (Thresholds & Retention tab):
- global on/off toggle, comma-separated exclude list, retention days + max MB
per container — DB-backed, no restart. The toggle + exclude are enforced
authoritatively at ingest (_persist_logs drops disabled/excluded lines), since
the push model has no channel to tell an agent to stop.
Tests: route-defined smoke; template-parse covers the new templates; integration
test for the ingest-time toggle + exclude enforcement.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CAGR73DUowdVFVvYzLXC5C
Bound docker_logs growth in the periodic cleanup task (same architecture as the
metrics/events retention). run_docker_retention gains logs_retention_days +
logs_max_bytes_per_container: it prunes lines past the age window, then keeps
only the newest ~cap bytes per (host, container) via a window-function ring
(exclusive-prefix sum, so the newest line always survives even if it alone
exceeds the cap). Containers rotate independently.
- settings DEFAULTS: docker.logs.enabled/exclude/retention_days(3)/
max_bytes_per_container(5MB) — operator preference: ~3 days / ~5 MB
- cleanup.py reads the two windows fresh each run (rule 25, no restart)
- integration rotate test: age prune + per-container byte cap + isolation
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CAGR73DUowdVFVvYzLXC5C
Ingest side of container logs. The host_agent ingest route extracts
sample["docker_logs"] as time-series batches (append every line, each carrying
its own Docker ts with recorded_at as fallback) and hands them to the docker
capability alongside the existing container/swarm/disk data — still under the
begin_nested SAVEPOINT so a logs failure can't sink host metrics.
- ingest.py: _persist_logs + pure _log_rows(batches, host_id) that shapes/filters
records (drops the _steward truncation marker, malformed + lineless records,
normalises unknown stream → stdout); persist_host_docker gains a logs= param
- routes.py: accumulate docker_log_batches, add to the guard + capability call
- unit test for _log_rows shaping/filtering; integration push→store→query test
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CAGR73DUowdVFVvYzLXC5C
Dedicated host-scoped table for pushed container log lines (one row per line),
chained after docker_008. host_id FK CASCADE; twin indexes — (host_id,
container_name, ts) for the viewer, ts alone for the age-cutoff prune — mirroring
docker_events. Integration schema-shape test asserts columns + both indexes.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CAGR73DUowdVFVvYzLXC5C
Agent-side of Docker container-log collection (milestone 79). Each interval the
agent fetches new log lines per running container over the local Docker socket,
demuxes the multiplexed stream, and folds sample["docker_logs"] into the same
push as metrics — no inbound channel needed.
- since-cursor per container kept in rate-state; a container's first interval
seeds from a short tail, then fetches incrementally via ?since=<cursor>
- _docker_request_raw + _demux_docker_logs + RFC3339Nano ts parsing (stdlib,
TTY/unframed fallback); one row per (stream, ts, line)
- per-batch byte cap with a truncation marker; logs are dropped (never buffered)
across a backoff so an outage keeps metrics but not stale logs
- docker_logs_enabled (default on) + docker_log_exclude per-host config keys
- AGENT_VERSION 1.6.0 -> 1.7.0
Server ignores the new sample key until the ingest step lands, so this commit is
CI-safe on its own.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CAGR73DUowdVFVvYzLXC5C
M78 step 2. The /disk page surfaced reclaimable space + stopped counts but
said "prune deferred to a later milestone" — this wires the cleanup.
Adds POST /plugins/docker/disk/<host_id>/prune (admin only) that resolves
the host's linked AnsibleTarget → steward:target:<id> scope and fires the
bundled maintenance/docker_prune.yml through the ansible.run_playbook
capability (audited runner; the collection agent stays read-only). Mirrors
the host_agent update route. `target` ∈ {containers, images, system}:
- containers → docker container prune -f
- images → docker image prune -af (prune_all_images)
- system → docker system prune -f (conservative)
The disk() route now resolves each host group's linked target + capability
availability; disk.html renders three confirm-gated buttons per host, shown
only to admins when Ansible is available and the host has a linked target,
otherwise a clear "link a target" / "runner unavailable" hint (rules 24/26/27).
Tests: the prune_target→extra_vars mapping (the -af-only-for-images rule) is
extracted to a pure helper and unit-tested; disk_prune added to the module
smoke; disk.html covered by the existing template-parse test.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CAGR73DUowdVFVvYzLXC5C
M78 step 1. The bundled maintenance/docker_prune.yml was a single
`docker system prune -f`; the upcoming per-host prune buttons need a
granular split. Add a `prune_target` extra-var:
- containers → docker container prune -f
- images → docker image prune -f (+ -a when prune_all_images)
- system → docker system prune -f (+ -a / --volumes) — the DEFAULT,
so existing manual/scheduled callers (which set no var)
keep today's behavior.
Reuses the existing bundled playbook rather than adding parallel files;
keeps the description/category/confirm meta so it still self-describes as
a confirm-gated maintenance run. Adds unit coverage: the playbook stays
valid YAML, its meta still discovers, and all three targets are guarded.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CAGR73DUowdVFVvYzLXC5C
Guardrails so the fd-leak class of bug (Errno 24 lockup; recent SNMP + UniFi
fixes) surfaces early or can't compound, instead of silently killing the app.
- Self-fd watchdog (steward/core/self_monitor.py): records open_fds and
open_fds_pct (% of soft RLIMIT_NOFILE) as "steward"/"process" metrics each
minute through the normal alert pipeline, so the operator can alert on them
via the existing alert-rules UI. Built-in WARNING floor at 80% gives a
zero-config early signal. Stdlib-only (/proc + resource); degrades to a no-op
off Linux. Registered as a core ScheduledTask in app.py.
- Poll-overlap guard (steward/core/scheduler.py): extract a pure _DueTracker
that skips a tick while a task's prior run is still in flight, so a hung poll
can't stack overlapping runs (which amplify per-poll resource/fd use). A
skipped task isn't penalised — it retries the next tick after it completes.
- fd-stability tests (tests/core/): _DueTracker overlap policy, the watchdog
metric/warning/degradation paths, and a real-fd canary that hammers tcp_check
and asserts /proc/self/fd doesn't grow.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The UniFi poller caches a UnifiClient holding a long-lived httpx.AsyncClient
(a real socket pool). On a login failure or a mid-poll error it set
`_client = None` to force re-auth, but never called close() — orphaning the
connection pool until GC. Under a flapping/unreachable controller that leaks
a file descriptor every failure tick: the same class of bug as the SNMP
engine leak (OSError: [Errno 24] Too many open files).
Add _drop_client() that aclose()s the client (best-effort) before nulling
the cache, and route both discard paths through it. Regression tests cover
both failure paths plus the noop/close-error contracts.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
A fresh SnmpEngine() was created on every poll_device() call and never
closed. pysnmp opens a UDP transport socket per engine and doesn't release
it on GC, so each scheduler tick (default 60s, per device) leaked a file
descriptor. Over hours of polling the process hit its fd ceiling and the
listening socket could no longer accept connections — OSError: [Errno 24]
Too many open files on socket.accept(), locking up the app.
Wrap the engine in try/finally and release its transport socket via a new
_close_engine() helper that probes both pysnmp API shapes (6.2.x lextudio
camelCase, canonical 7.x snake_case); all close paths are best-effort so a
failed close never breaks the poll loop. Regression tests cover both shapes
and the never-raises contract.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
SNMP devices map onto a Steward host's detail page via _devices_for_host,
which previously matched the device's `host` field (the SNMP poll target)
against the Host's address/name — a coincidence of strings that breaks when
the poll target differs from how the Host is recorded.
Add two optional per-device config fields that decouple the binding from the
poll target:
• host_id — exact Steward Host UUID match (the explicit link)
• steward_host — friendly bind by Host name or address (case-insensitive)
An explicit binding is exclusive: a device bound to host A never implicitly
matches host B by a coincidental address. No explicit binding → existing
implicit `host`-string match (fully backward compatible).
host_panel passes host.id and badges explicitly-bound devices. plugin.yaml
documents the new fields. Tests cover explicit id/name, exclusivity, wrong
uuid, and legacy fallback.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_016Jg27rgypiW2efULXJDtMC
The container list and dashboard widget listed each swarm task as its own cryptic
row (svc.1.<taskid>) and could only show tasks on nodes that run a Steward agent.
Make both views service-centric and cluster-complete.
- swarm_view.build_swarm_services(): model-free builder that merges real container
rows (collected local-per-node, so host_id = the node a task runs on) with the
managers' placement (DockerSwarmService.placement_json). Where placement counts
exceed the local rows on a node — i.e. nodes with no agent — it synthesizes
"ghost" replicas so the replica list matches the cluster. Non-swarm containers
pass through grouped by host.
- /rows + rows.html: a Swarm-services panel collapses each service to one block,
every replica a host chip (status + cpu); ghosts render dashed "N · no agent".
Non-swarm/compose containers keep the per-host table below. New Services stat.
- /widget + widget.html: same service collapse with host chips; standalone
containers stay grouped by host.
- Unit tests for the builder (real+ghost merge, full coverage, partial agent-less,
synthesized service, down state).
No migration — node_id, placement_json and swarm-node hostnames are already stored.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_016Jg27rgypiW2efULXJDtMC
The widget-dedup test imported plugins.docker.routes, which re-imports
plugins.docker.models after the app already loaded the docker plugin —
"Table 'docker_containers' is already defined". Same plugin-loader gotcha the
host_agent query helpers avoid.
Extract dedup_by_container_id into plugins/docker/dedup.py (imports no ORM
models, so it's safe to import in either environment); routes.py imports it
from there and the test targets the model-free module.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_016Jg27rgypiW2efULXJDtMC
The running count was inflated: swarm-aware agents on every manager list
cluster-wide tasks, so the same container (identical container_id) was reported
once per manager and the dashboard widgets summed them. The earlier swarm dedup
only covered the Swarm topology page — the widgets read raw rows.
- _dedup_by_container_id(): collapse rows sharing a non-empty container_id
(globally unique, so only true duplicates merge); first occurrence wins under
the running-first ordering. Applied in both the containers and resources
widgets before counting/limiting. Rows without a container_id (older agents)
are kept as-is.
- Replaced the static "stopped" tally — not actionable — with "failed (24h)":
distinct containers with a die/oom event in the last 24h (DISTINCT so the
managers' duplicate events don't double it), rendered red when non-zero.
- widget.html empty-state now keys off total_count.
- Regression test: same container_id from two managers counts once; id-less rows
all kept.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_016Jg27rgypiW2efULXJDtMC
The container persister upserted current state keyed (host_id, name) but — unlike
the image/swarm/disk persisters — never deleted containers that vanished from the
host's latest listing. Every removed one-shot container (CI job runners, buildkit
builders, codex jobs) left a permanent "stopped" row, so the dashboard counts
ballooned (e.g. 856 stopped) and read as "not dedup'd". It wasn't dedup — it was
a missing reaper.
- ingest.py: after the upsert loop, delete this host's containers whose name is
notin the newest snapshot (the listing is authoritative — event derivation
already treats absence as removal). Mirrors images/swarm/disk. The existing
`if not snapshots: return` keeps swarm/disk-only samples from touching
containers; an empty container snapshot legitimately means none exist.
- widget_resources.html: this widget shows the busiest running containers by CPU.
Replaced the two unlabeled hairline bars with labeled CPU/MEM bars (label · bar ·
% value, warn/crit colors) so it's self-explanatory.
- Regression test: a container present then absent across snapshots is reaped.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_016Jg27rgypiW2efULXJDtMC
SNMP devices are config-defined by IP/hostname, not Steward Host records, so
they had no presence on a host's page. Map them by address and embed a fragment
(mirrors the Docker per-host fragment).
- _devices_for_host(devices_cfg, address, name): case-insensitive match of a
device's configured host to the Steward host's address or name (tolerates
non-dict / host-less entries).
- Route /plugins/snmp/host/<id>: renders the matched device(s) + latest readings,
or nothing when none map (so hosts without SNMP carry no empty card).
- snmp/host_panel.html: per-device card (name · address · reachability) with a
readings grid (K/M scaling) and a History link to the full device page.
- hosts/detail.html embeds it after the Docker fragment, gated on snmp enabled.
- Unit test for the matching helper (by address, by name, no-match, blank).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_016Jg27rgypiW2efULXJDtMC
plugin_metrics grows by (sources × resources × ~30s cadence); keeping 90d of raw
is a large table. Add a raw→hourly rollup (mirroring the Docker plugin) so only a
short raw window is kept at full resolution, with hourly averages archived longer.
- PluginMetricHourly model + core migration 0024 (plugin_metrics_hourly: avg/max/
count per source/resource/metric/hour, unique bucket constraint + lookup index).
- steward/core/metrics_retention.rollup_plugin_metrics: date_trunc('hour') agg of
raw older than the hour-aligned raw window, idempotent pg upsert into hourly,
delete the rolled raw, prune hourly beyond the rollup window.
- cleanup.py: plugin_metrics is no longer blanket-deleted at data.retention_days;
_run_metrics_retention drives the rollup with windows read live from settings.
- Settings: metrics.retention.raw_days (7) + rollup_days (90), tunable on the
Thresholds & Retention page (new "Host metrics retention" card).
- Chart read: _history_for_host merges the hourly rollup (older part of the range)
with raw date_bin (recent part, capped ≤1h), so 30d charts keep working —
recent at full resolution, older at hourly. Route passes raw_days from settings.
- Tests: unit (cutoff helpers) + integration (rollup aggregates/prunes; history
merges hourly + raw) against Postgres.
Speed was already handled by the indexes + SQL aggregation; this is the storage
lever (raw window ~10x smaller).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_016Jg27rgypiW2efULXJDtMC
The full-metrics page rendered once server-side with the history query inline,
so the charts blocked first paint and nothing refreshed without a reload (it
reads the latest snapshot the server holds — the agent pushes ~every 30s).
Split it into a shell + two HTMX fragments:
- Shell (host_detail.html): header + shared time-range toggle + two containers;
paints instantly.
- Current state (/<id>/metrics → _host_metrics.html): identity + gauges +
per-core + filesystems + interfaces/disks + temps, from the DISTINCT ON latest
query. hx-trigger "load, every 15s" → live numbers at ~the agent cadence.
- History charts (/<id>/charts → _host_charts.html): the 3 charts + Chart.js,
from the date_bin history query. hx-trigger "load, every 60s, rangeChange" so
they lazy-load (never block paint), refresh slowly, and follow the range
selector. Leak-safe: previous Chart instances are destroyed before re-render.
- Range toggle switched from full-reload links to the shared _time_range.html
(setTimeRange + rangeChange), so only the fragments refetch.
- routes: host_detail (shell) + host_detail_metrics + host_detail_charts, with a
_split_host_metrics helper.
- tests/test_templates_parse.py now also parses plugin templates (these
fragments aren't rendered in the unit lane).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_016Jg27rgypiW2efULXJDtMC
The previous commit's integration test failed at collection-time import:
importing plugins.host_agent.routes (which imports the host_agent ORM models at
top level) double-registers host_agent_registrations against the app-loaded
plugin's metadata ("Table already defined").
Move the two pure read helpers (_latest_metrics_for_host, _history_for_host)
plus SOURCE_MODULE / HISTORY_METRICS into plugins/host_agent/metrics_query.py,
which imports only the core PluginMetric model — no plugin models. routes.py
imports them back. The integration test now imports from metrics_query and no
longer trips the loader's registration guard. No behavior change.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_016Jg27rgypiW2efULXJDtMC
Follow-on to the plugin_metrics indexes. plugin_metrics is already retention-
bounded (core cleanup prunes > data.retention_days, default 90d) and charts top
out at 30d, so the cost wasn't growth — it was the read path shipping raw rows
to Python.
- _history_for_host: bucket + average in SQL via date_bin (epoch-aligned, ~120
buckets) instead of fetching every raw sample (a 30d range was hundreds of
thousands of rows) and downsampling in Python. Uses the new
(source_module, resource_name, recorded_at) index.
- _latest_metrics_for_host: DISTINCT ON (resource_name, metric_name) ORDER BY
recorded_at DESC — newest row per group in one index-ordered pass, replacing
the GROUP-BY-max subquery self-joined back to the whole history.
- Integration test validates both against Postgres.
Deliberately not a materialized raw→hourly rollup: these query-side changes
deliver the speed; a rollup would additionally cut storage and remains a future
option if scale demands it.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_016Jg27rgypiW2efULXJDtMC
docker_metrics.mem_usage_bytes and docker_containers.mem_usage_bytes /
mem_limit_bytes were int4 (max 2,147,483,647). A container using >2.1 GB of RAM
(e.g. 8.18 GB) overflowed the column, so asyncpg raised "value out of int32
range" and the entire docker ingest batch failed — no metrics stored for any
host with a large container.
The I/O counters and the milestone-77 rollup/disk tables already used
BigInteger; this trio (docker_001-era) was missed.
- models.py: DockerMetric.mem_usage_bytes, DockerContainer.mem_usage_bytes,
DockerContainer.mem_limit_bytes → BigInteger.
- migration docker_008_bigint_mem: ALTER COLUMN ... TYPE BIGINT (safe in-place
int4→int8 promotion).
- integration regression test: persist an ~8 GB container, assert the
current-state and time-series rows round-trip.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_016Jg27rgypiW2efULXJDtMC
Plugin UIs had no nav home (only dashboard widgets / typed URLs). Add an
optional get_nav() plugin export and surface it in the sidebar.
- plugin_manager: _PLUGIN_NAV registry + get_plugin_nav() getter + a tolerant
_collect_plugin_nav() (missing hook = fine; raising/malformed = logged &
skipped; idempotent per plugin so hot-reload re-runs cleanly). Collected in
both the startup load path and the hot-reload path.
- app.py: inject plugin_nav into the template context, filtered to enabled
plugins so a hot-disabled plugin's link can't linger before restart.
- base.html: render plugin links under the Infrastructure group, with the same
request.path active-state treatment as core links.
- docker/snmp/unifi/traefik: each exports get_nav() → its /plugins/<name>/ view.
- Tests: collector behavior (collect, missing hook, reload-replace, malformed
item, raising hook, sorted output).
With docker enabled, "Docker" now appears under Infrastructure → its fleet view.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_016Jg27rgypiW2efULXJDtMC
The flat top bar was a set of ungrouped peers and gave plugin data no home.
Move to a persistent left sidebar with grouped sections, per the navigation
redesign (operator-chosen shell).
- base.html: top <nav> → left <aside class="sidebar"> + content column.
Groups: Overview (Dashboard, Status), Infrastructure (Hosts; plugin links
arrive in slice 2), Monitoring (Monitors, Alerts), Automation (Ansible),
Admin (Settings, Audit; admin-only). Brand on top, user/logout at the bottom.
- Active link highlighted by request.path prefix (aria-current).
- Responsive: sidebar slides off-canvas under 900px via a ☰ toggle + scrim
(inline class toggle, no new JS deps). Candle-glow preserved.
- Logged-out pages (login/setup) render without the sidebar (gated on
session.user_id), content area full-width and centered as before.
- Add tests/test_templates_parse.py: syntax-parses every steward template so a
broken tag fails the unit lane (only the login page renders there today).
Plugin nav links (Docker/SNMP/UniFi/Traefik) come next in slice 2 via a get_nav
hook. UI-only; no behavior/route changes.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_016Jg27rgypiW2efULXJDtMC
Follow-up to the incident where an unwritable /data made Steward silently mint a
fresh ephemeral secret key on every boot, orphaning all encrypted secrets — only
discovered when an Ansible run failed. Make both failure modes loud and visible.
- config._resolve_secret_key: if a new key must be generated but can't be
persisted (no STEWARD_SECRET_KEY, unwritable /data), raise RuntimeError and
refuse to start, with an actionable fix (set STEWARD_SECRET_KEY, or make /data
writable by uid 1000). Also raises a clear error if an existing key file can't
be read. First-run on a writable volume still generates + persists normally.
- core.settings: detect secrets stored as ciphertext that won't decrypt with the
current key (scan_undecryptable_secrets at startup; _is_undecryptable helper).
Cached in memory; set_setting discards a key on a fresh write so the banner
clears without a restart.
- app.py: run the scan after migrate_plaintext_secrets, log a warning, and inject
undecryptable_secrets into the template context.
- base.html: admin banner naming which secrets to re-enter, each linked to its
settings tab.
- compose.deploy.yml: document the uid-1000 /data write requirement and the
STEWARD_SECRET_KEY option for multi-node Swarm.
- Tests: secret-key resolver (reuse existing file, generate when writable, raise
when unpersistable) and the undecryptable-secret detection helper.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_016Jg27rgypiW2efULXJDtMC
The SNMP plugin ships in the image but logged "pysnmp not installed — SNMP
polling disabled" on every poll, so polling never worked. Two coupled defects:
1. The Dockerfile installed only `.[ansible]`, so the `snmp` extra (pysnmp)
was never bundled even though the plugin is first-party and shipped.
2. poller.py used the synchronous pysnmp HLAPI (`next(getCmd(...))`), which
pysnmp-lextudio 6.x removed — it's asyncio-only now — so even with the dep
present, polling would have thrown and silently returned nothing. The 5.x
line that still has the sync API isn't safe on the image's Python 3.13.
Fix:
- Dockerfile: install `.[ansible,snmp]`.
- poller.py: `poll_device_sync` → `async def poll_device` on the asyncio HLAPI,
with a dual-version import (pysnmp 7.x `pysnmp.hlapi.v3arch.asyncio`/`get_cmd`
+ async `UdpTransportTarget.create`; pysnmp-lextudio 6.2.x
`pysnmp.hlapi.asyncio`/`getCmd` + direct `UdpTransportTarget`) so a dependency
bump can't silently re-break it.
- scheduler.py: await poll_device directly; drop the run_in_executor wrapper
and the now-unused asyncio import.
- Add tests/plugins/snmp/test_poller.py covering the version→mpModel mapping,
that the poller is a coroutine, and the graceful no-pysnmp path.
Note: CI confirms import/load and the no-pysnmp path, but has no SNMP target —
live polling against real devices is verified after deploy.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_016Jg27rgypiW2efULXJDtMC
CI never renders docker templates through the app (the unit-lane app uses
testing=True, which skips plugin loading), so a Jinja syntax error could ship
green. Add unit tests that parse every docker template, smoke-import the routes
module + confirm the #942 view functions exist, and cover the _human_bytes /
_human_uptime presentation helpers. Closes the render-regression gap the new UI
pages introduced.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_016Jg27rgypiW2efULXJDtMC
