Release notes
Seylo is released in two parts:
- Seylo itself, the portal at
portal.seylo.in. Every organisation gets a new release when it ships. There is nothing for you to install. - The agent and the
seylo-observabilitychart, which you install in your clusters and upgrade when you choose.
0.5.0 — next release
Section titled “0.5.0 — next release”-
Retention in one click. On Integrations → Observability stores → Retention, each store has one button per choice: 7, 15, 30 or 90 days for logs and traces, and also 180 or 400 for metrics. Click one to see what it does to your disk and your data, measured on your cluster, then Keep N days; Seylo restarts the store and reads its own setting back. A shorter period says it deletes data, a metrics period that does not fit the volume is refused with the numbers, and a store managed by Argo CD, Flux or an operator asks for an explicit tick. Only an admin can change it. Needs
seylo-observability0.2.23. See choose how long to keep data. -
One switch for the archive: Keep a copy in my cloud account. It copies every installed signal, logs, metrics and traces, every hour to a bucket in your own cloud account. On Azure it is one downloaded template and one command, read back before Turn on; Archive now proves the identity the same day. Per signal status says what is copied and up to when. See keep a copy in your cloud account.
-
Choose how often to copy and how long to keep the copy. Logs can be copied every hour, 15 minutes or 5 minutes; metrics and traces every hour or 15 minutes. Each choice shows the most that would be lost if the cluster were deleted and what it costs in storage operations from your own bill, or unpriced with the reason. Six keep-for choices set a storage rule priced the same way. See how often to copy.
-
Read the archive when the cluster is gone. The Logs, Metrics and Traces pages search the copy when the agent is offline or the cluster was removed, labelled From the archive, up to a time. Removed clusters appear in the picker as Removed - archive only. See read the archive when a cluster is gone.
-
A replacement cluster continues the history. An admin confirms in one click that a new cluster replaces a removed one; ranges that start before the new cluster existed read the old archive, with the gaps named. See link a replacement cluster.
-
Unit cost. FinOps → Unit cost shows what one test, transaction, billable hour or API call costs for each tenant (a client, a branch, a store) and for the organisation, from your cost and a count from your own data. The shared part is shown apart, with the rule that split it. A day with no count is “no count reported”, never 0. Counts come from a read-only query that runs in your cluster, your application, a push or a CSV, and are only counts. See unit cost and send counts. The query job needs
seylo-observability0.2.24. -
Firebase, read in place. Connect a Firebase project on Integrations → Firebase and see crashes, crash-free users and performance on Observability → Mobile apps, read from your own BigQuery with no copy to Seylo and no key. Nightshift can cite an app’s crashes as a lead. See connect Firebase and Mobile apps.
0.4.10 — 2026-11-12
Section titled “0.4.10 — 2026-11-12”- Optimise is now FinOps. The section, its sidebar entry and its address
changed:
/finopsreplaces/optimise, and old links, bookmarks and emails open the new pages. In these docs the pages moved fromoptimise/tofinops/, and the old addresses redirect. - Savings shows one combined plan. Right-size first, then fewer or smaller nodes, each step with its share of one figure; spot is shown beside the figure and joins it only when you tick it. A pool can keep some nodes of its size beside a smaller one. See the plan.
- Fewer things pin a node. The cloud’s own add-ons (for example AKS’s managed Prometheus), finished pods, and read-only mounts of the node’s logs or CA certificates no longer keep a node. See what keeps a node, and what does not. The read-only mounts need an agent newer than 0.2.22.
- StatefulSets are listed as not counted in the right-sizing step, with what they could free, instead of being left out silently.
- Nightshift checks a namespace’s quota before it raises anything. Every fix that raises requests, limits, replicas or storage must fit every quota in the namespace, including during its rollout. A fix a quota refuses says so at once, as The change could not roll out, instead of “did not work”. See every fix is checked against the quota first.
- A namespace over its quota is a critical alert: new pods are blocked, even when the namespace uses a small part of what it declares. Nightshift’s first remedy is a rebalance: room made by lowering workloads that declare far more than they use, one fix a person approves. See a namespace at its quota.
- A budget below what a namespace already counts needs an acknowledgement. It would refuse every new pod there. Admin → Clusters → Coding agents lists namespaces already over their budget, with Raise the budget. See give the namespace a budget.
0.4.9 — 2026-11-12
Section titled “0.4.9 — 2026-11-12”-
One page for everything Seylo does on its own: Operate → Automation. It shows, per cluster, the Nightshift fixes that may run without asking and the grants behind them, the managed savings grant, one Pause all automation for every cluster, and one log of every change Seylo made without a person, and every refusal, with why and under what. Changes waiting for approval stay on Approvals, counted at the top. It replaces FinOps → Cost & savings → Autoscale and the controls on Nightshift → Autonomy, which keeps a one-line summary; old links open the new page. Every standing permission is now called a grant. See Automation.
-
Allow this kind, from Approvals. On a resource or replica change waiting for approval, an
admincan open the grant form pre-filled for that workload, with the waiting change’s own size as the limit, so the same kind of change runs on its own next time. An approver still signs it. See allow this kind of change. -
The separate autoscale engine is retired. It worked out its own sizing every 15 minutes and filed it for approval again every few days. Sizing for savings is now one judgement, in Savings’ Rightsizing. Requests it already filed stay on Approvals. A grant naming CPU request, memory request or replicas now covers only Nightshift’s and remediation’s changes. See what changed.
-
Rightsizing raises too. Besides lowering a request above use, it raises a CPU or memory limit a container presses on, never both for one resource. A raise is a risk taken away, not a saving, so it has no dollar figure. Apply selected sends the raises first. See Rightsizing.
-
Managed savings. An
admincan let Seylo apply a cluster’s savings on its own, lever by lever, inside a grant signed for 90 days: rightsizing, fewer nodes, spot through your own node autoscaler (AKS node auto-provisioning, Karpenter on EKS, EKS Auto Mode), better machine sizes, and off-hours schedules for a non-production cluster. Every change is checked and put back if it does not hold. Stop managing, Restore original and Renew are one click each, and admins are emailed 7 days before it ends. It costs 5% of the saving Seylo measures, per lever: no saving, no fee, and nothing after you stop. Spot and sizes need agent 0.2.21 with a permission that is off by default. See managed savings. -
Every $0 lever says why. A lever with no saving names its cause from the same packing: pods that cannot move, a pool that was not packed again, every node already needed, or a spot pool that would cost more than it saves. A new Cannot move list names each pod that keeps its node, what to do about it, and what clearing them would add. See when a lever reads $0.
-
Better sizes and spot prices on AWS and Google Cloud. Savings now tries other EC2 and Compute Engine machine sizes, priced from each cloud’s list, within the pool’s CPU architecture. AWS spot prices are read through your connected account, per zone, and Google Cloud’s from its billing catalogue. See Savings.
-
Only sizes your account can get. Before a size is offered, Seylo reads your connected cloud account for whether it is offered in your region and zones and fits your vCPU quota, and leaves out the ones that do not, saying why. A size restricted only in zones your pool does not use is kept. AWS stacks gain a
seylo-quota-readpolicy (two Service Quotas reads); a stack saved before 10 November 2026 says how to update it. See sizes your account can get. -
The usage card counts each node, container and volume once, even when two collectors export the same one. A cluster with a second kube-state-metrics could show twice its real capacity.
-
More Nightshift fixes may run on their own, only when an admin allows them. New grant actions: roll back a crash or stuck rollout to the last release Seylo saw run healthy, within 30 minutes of the bad release; raise a CPU request to measured use plus 20%, at most 4 times, under a CPU ceiling you set; and add an HPA where load splits across pods, from today’s pods up to Most replicas, which the grant must set. On Automation, Probe paths lets a probe fix run on its own after 3 were approved here, only with a path copied from a passing probe of the same image, and the memory card can allow a second step the same day once the first held. Nothing changes until an admin turns one on. See what Nightshift may do on its own.
-
A volume alert reads who owns the claim. Before, Nightshift could not tell, and said so with an owner it could not name. Now an owner it cannot read keeps the growth for a person, and says why.
-
Agent 0.2.21. Reads which node autoscaler your cluster runs, and adds
rbac.write.autoscaler.enabled, off by default, for managed savings. See agent 0.2.21. -
Twelve more things Nightshift detects and fixes, all waiting for a person: a CronJob that misses its schedules, a CronJob whose runs pile up, a Job or CronJob run that keeps failing, a namespace at its ResourceQuota (always a person), a missing ConfigMap or Secret, a PodDisruptionBudget that blocks a drain, metrics-server down, CoreDNS unhealthy, an init container that keeps failing and a node out of disk, plus a registry that refuses the login and a certificate that is not ready, which are diagnosed only. See what Nightshift handles and the fixes.
-
A change outside an app’s namespace, or in
kube-system, needs anadminwho also has release access there, and never runs on its own. See changes outside the app’s namespace. -
Automation: HPAs at their maximum has a Highest maxReplicas field; grant cards name each action in plain words; Most replicas is required for a grant that adds an HPA. See HPAs at their maximum.
-
Incidents: an alert beyond an incident’s 100-alert limit belongs to it and starts no investigation of its own. Alerts that are unrelated are no longer pulled into an incident when the calls between their services succeed. See incidents.
-
A feature-flag revert restarts what holds the flag. After the flag service, it restarts the services that cache the flag, chosen from evidence, at most 3 workloads in all. When a revert does not clear the alert, the notice no longer says the setting was ruled out. See a setting or feature flag.
-
Cost and Savings are faster. They share recent figures and say how old they are: Figures from N min ago · Refresh. Tabs stay loaded, and long ranges on Metrics load faster. See how fresh the figures are.
-
Agents: a cluster whose agent keeps presenting a rejected or revoked credential shows a note on its page saying to re-enroll or uninstall it, and pages answer at once for a cluster that has been offline for a while. See troubleshooting.
-
Agent 0.2.22. Reads PodDisruptionBudgets, Jobs, CronJobs, ResourceQuotas and cert-manager’s certificates, read-only. See agent 0.2.22.
0.4.8 — 2026-11-11
Section titled “0.4.8 — 2026-11-11”- Savings shows how much you could save, and where. FinOps → Cost & savings → Savings opens on one figure - available savings per month, per year and as a share of what the nodes cost - made of four levers you tick: Fewer, better nodes, Rightsizing, Spot for safe workloads (priced from Azure’s spot list) and Schedule (switch a non-production cluster off outside working hours). Below them, the machines you run today against the machines after, and how much of what you pay for is requested and used. Savings are worked out by packing your pods onto the right machines, so a system pod without a PodDisruptionBudget is listed under Clear first instead of hiding the saving. On Azure, Fewer, better nodes also tries cheaper VM sizes. A lever with no price says why, and the total then reads “at least”. Apply selected sends the ticked levers for approval. See Savings.
- When a Nightshift fix does not work, you are told why. If a fix does not clear the problem, Nightshift puts back what it changed, as before. It now also writes a line on the alert saying what it tried, why it did not count and what to do next, and emails admins and whoever approved the fix. When Nightshift reverts a setting or feature flag and the alert keeps firing, it puts the setting back and says the setting was not the cause. The message names the setting but never its value. Each person can turn the email (Nightshift fixes put back) down to the daily digest or off in their notification settings. A fix whose object no longer exists ends did not verify, with “nothing to undo”, instead of “undo failed”.
- What Nightshift handles, in one page. The problems it fixes on its own once you allow it, the ones it fixes after you approve, and the ones where it hands you the fix. See what Nightshift handles.
- Firewalls checked against where your cluster leaves from. For a
database firewall or App Service access restrictions, Nightshift now names
the address your pods leave from - the load balancer, the NAT gateway, or
the Azure Firewall your route table sends all traffic through - and gives
the command that admits it. An address it cannot read, such as a
third-party appliance’s, is “not known”, with the reason. A database’s
timeout expirednow counts as a symptom, a stopped Function App is named, and a cloud host first seen in the pod’s logs is read too. Still Azure only, still read-only: AWS and Google Cloud resources are not read. See Nightshift and your cloud. - Changes made outside Seylo. “What changed” now also lists Jobs, CronJobs, Secrets, ConfigMaps, workloads and Services created in the namespace in the 2 hours before the alert, and which Secrets and ConfigMaps the failing workload reads. Secrets are listed by name and time, never read. A new credential-like object just before a rejected login is flagged as a likely rotation on one side. See what changed.
- Nightshift’s rules feed the investigation. When its rules already recognised the failure, that finding is evidence the run must confirm or refute.
- Feature flags put back, properly. A flag or setting that changed just before the errors goes back to the value Seylo recorded, the flag service restarts so it loads it, and the revert counts only after the alert stays quiet for 5 minutes. If it does not hold, it is undone, under the same owner acknowledgement you approved it with. See setting or feature flag.
- New fixes you approve. Pause a Chaos Mesh experiment (the agent needs
updateon it; the role is on the page), scale a workload back from zero to the count it last ran, add the one permission a 403 names to a workload’s Role, and set a ConfigMap key to the value your own skill names. A code fault on an alert starts the coding-agent task in one click. See Nightshift’s fixes. - Sizing reads only this workload’s history. CPU and memory sizing read usage only since the workload was created, so one deleted and created again does not inherit the old load. A pending pod’s refusal names the resource that is really full, and why each node was passed over.
- Prices from your Google Cloud bill. A connected Google Cloud project prices machine types and Persistent Disks from your Cloud Billing export in BigQuery, read once a day through the same keyless connection. The export is yours to turn on; Test finds it or says how. See prices from your cloud bill.
- Agent 0.2.20. Reads which Key Vault secret version each pod mounted, never the value, and refuses a manifest with no namespace. See agent 0.2.20.
0.4.7 — 2026-11-09
Section titled “0.4.7 — 2026-11-09”- More fixes Nightshift may apply on its own, only when an admin grants them. Memory limits sized from evidence may now reach four times today’s limit (a step with no evidence still stops at twice). On Nightshift → Autonomy, two new cards: Nearly-full volumes grows a volume on its own only inside the largest step and largest size you set, and only when the added cost is priced from a source - an unpriced growth waits for a person; Claims that never bound recreates a claim on its own only when it is proven never to have bound. New standing-permission actions: roll back when the new image cannot start, restart pods that hold an old Key Vault secret, and eleven one-fix grants, each held to the bound its label states. Nothing changes until an admin turns one on. See what Nightshift may do on its own.
- An autoscaler or a bigger request, decided from measurements. When a workload uses more CPU than it requests, Nightshift asks whether adding a pod would lower each pod’s CPU, from 24 hours of per-pod CPU and the workload’s traffic, and proposes an HPA only when the answer is yes. See autoscaler or a bigger request.
- Nightshift reads your cloud dependencies, metadata only. When the
evidence names a Key Vault, a storage account, a PostgreSQL, MySQL or Azure
SQL server, a VM or the network path to it, an App Service or Function App,
or a Front Door, Nightshift reads its settings through your Azure cloud
account and gives the root cause, its evidence, and the exact
azcommand and Terraform or Bicep for a person to apply. Application Insights and Front Door logs are queried for failures only. See Nightshift and your cloud. - One Seylo Reader role on Azure. A new Azure connection grants a custom role, Reader minus the 31 reads Microsoft documents as returning a key, token or app setting. Creating it needs Owner or User Access Administrator. On AWS, every stack now denies the reads that return instance user data and Lambda environment variables. AWS and Google Cloud connections get a Let Nightshift investigate cloud dependencies tick box, ticked by default. Test now lists what Nightshift can read, and says when a grant is wider than metadata-only. An existing connection keeps working, and its card shows the commands that move it. See cloud accounts and what Seylo reads in your cloud.
- Prices from your own bill: Azure disks and AWS. Managed-disk tiers are
now priced from your Azure bill, beside machines. A connected AWS account
prices EC2 instances and EBS volumes from Cost Explorer, with one more
permission in the stack (
ce:GetCostAndUsage); AWS charges $0.01 per request, and Seylo reads once a day per region. See prices from your cloud bill. - Across clusters. Seylo links your clusters when one sends traffic to another, such as a front proxy kept in front of a migrated app. Nightshift says where a 5xx is born - in the app, at the new ingress or at the front proxy - gives the exact annotation to change on the front, and lists what the migration left behind. Both clusters connected give the full answer. See across clusters.
0.4.6 — 2026-11-09
Section titled “0.4.6 — 2026-11-09”- The portal shows in search engines as its sign-in page only. A search
engine may list
portal.seylo.inas the Seylo sign-in page, titled Sign in · Seylo. Every other portal address tells search engines not to index it. Once you sign in, the browser tab reads Seylo, as before.
0.4.5 — 2026-11-07
Section titled “0.4.5 — 2026-11-07”- Nightshift on or off per cluster. An
admincan switch Nightshift off for a cluster under Nightshift → Settings → Clusters, with an optional reason. On a cluster that is off, Nightshift investigates nothing, proposes nothing and applies nothing, on its own or when someone presses Investigate & fix. Alerts are still raised and shown there, and fixes already waiting for approval stay with the approvers. Every cluster starts on, and each change is in your audit log. See Nightshift on or off per cluster. - Prices from your own Azure bill. A connected Azure subscription (Admin → Integrations → Cloud accounts, the Reader role, no key) now prices your nodes from the subscription’s own bill: what each machine type actually cost, with reservations and savings plans spread over their hours, in US dollars as Azure states them. Your bill wins over an uploaded rate and the public list price. The Price book card then reads Prices: from your Azure bill. On Make the figures match your invoice, the first row is Your Azure bill; the Azure price sheet and cost export rows appear only on a cluster that already runs them. See your Azure bill.
- Whole-cloud cost dashboards from your Azure bill. The whole-cloud panels on Cost & savings → Dashboards read your connected subscriptions’ bill, in its own currency, with up to 190 days of history. Choose Whole cloud bill (no cluster) to draw them without a cluster. See whole-cloud spend.
- Datadog and OpenAI spend with one script. Tick the tools, choose your
Datadog site, and copy one bash script. It names the cluster
kubectlpoints at, asks before writing, reads each key hidden in your own terminal, writes each Secret, then runs onehelm upgrade. A key is never typed into the page and never sent to Seylo. Anthropic, Grafana Cloud, Splunk, Dynatrace and New Relic are listed as not available, with the reason. See Datadog and OpenAI spend. - Autoscale limits are asked once. The policy form has one Max increase per change % and one Max decrease per change % for every ticked dimension; a different limit, a floor or a ceiling for one dimension is behind a checkbox. The form says that the 50% cap on one unattended change is fixed in Seylo, and a disabled Request approval says why. See dimensions and limits.
- Timezones are chosen from a list in the autoscale policy form and in Schedule deploy: your browser’s zone first, then common zones, then every zone, each with its offset from UTC.
- Expiry limits at the date picker. A silence (90 days), a snooze (7 days), a deploy allowance (90 days), an accepted scan finding (30 days) and Nightshift’s proven fixes (their limit) no longer accept a date past the limit only to refuse it after you submit. The picker greys out later days, and a later date typed in shows the limit and the latest date allowed under the field, with the button off until it is changed.
- Azure commands work from Git Bash on Windows. Commands Seylo shows that
pass an Azure resource ID now start with
MSYS_NO_PATHCONV=1, so Git Bash no longer turns/subscriptions/...into a file path: the cloud account’s Reader grant and its removal, and the Key Vault role in copy a namespace. Other shells ignore it, and the portal’s PowerShell forms leave it out. - A tidier Savings tab. One Generate proposal button, on the Optimisation proposal card, and the rules that were not judged are folded under one line.
0.4.4 — 2026-11-05
Section titled “0.4.4 — 2026-11-05”- A coding session on a branch that does not exist is refused before a
sandbox starts. Seylo now asks GitHub whether the branch exists first. A
repository with no commits has no branch at all, even though GitHub names
mainas its default. A missing branch is refused withrepo-branch-missing, which says what to do: add a first commit to the repository, or leave Branch empty to start from the default branch. No sandbox is started. If Seylo cannot ask GitHub, the session starts as before. - When a session cannot fetch its repository, the error says why. It names the cause git reported: a branch that does not exist, a Seylo GitHub App installation that does not cover the repository, or a sandbox that could not reach the Git host. Before, it named two possible causes every time.
0.4.3 — 2026-11-05
Section titled “0.4.3 — 2026-11-05”- More room for the conversation in the coding agent. In
Build → Coding agent:
- This month’s use of Seylo’s model is in the status bar, beside the sandbox
minutes, as a short figure such as
$0.05/$5.00. The full sentence is in its tooltip. - The reminder never to paste a password or token is an (i) beside the model menu.
- Resume, Hibernate and Stop are icon buttons, with what each does in its tooltip.
- How a pull request is made is now the tooltip of Open pull request. When the button is disabled, the tooltip gives the reason first, then how a pull request is made.
- A button at the top of the session hides the details panel, so the conversation takes the full width. Your browser remembers the choice. A status bar item that opens Checks brings the panel back.
- This month’s use of Seylo’s model is in the status bar, beside the sandbox
minutes, as a short figure such as
- The coding agent knows about Deploy. Asked to deploy, it now prepares
what Deploy needs: Kubernetes manifests in a
k8s/folder, and a Dockerfile if the repository has none. Then it tells you to press Deploy and choose the cluster and namespace. Before, it said deploying was impossible. It still holds no cluster credential. See deploy from the coding agent.
0.4.2 — 2026-11-04
Section titled “0.4.2 — 2026-11-04”- Security hardening. Sign-in limits and audit records now use your real address rather than a proxy’s. Very heavy bursts of sign-ins get a short “try again in a few seconds” instead of waiting. The portal can no longer be shown inside another website. A saved connection key is only ever sent to the address it was saved for: change the address and you enter the key again.
- The enrollment token is never put in a web address. Seylo refuses a token sent that way. Download the manifest from the cluster’s install page instead. See install the agent.
- After an update, an open page reloads by itself once when it asks for a file the new version no longer has, including a stylesheet. Before, a page such as Billing could show an error until you reloaded it.
0.4.1 — 2026-11-04
Section titled “0.4.1 — 2026-11-04”- Billing changes. See plans and billing.
0.4.0 — 2026-11-03
Section titled “0.4.0 — 2026-11-03”0.4.0 is the launch release. It comes after 0.3.15, a security release, and keeps everything 0.3.15 changed.
- Heads-up for Google Workspace sign-in: save the client secret. Google redeems a sign-in only with the client’s secret, so the secret is now required for Google, and Seylo completes Google sign-ins on its server with the secret you saved. It is stored encrypted, is write-only and is never shown again. If your Google provider was saved without one, paste it in and select Save provider. Any other OpenID Connect provider saved with a client secret signs in the same way; Microsoft Entra ID, and providers saved without a secret, are unchanged. See Google Workspace.
- Paid from the first day. The free preview ends with this release and there is no standard trial; Seylo may grant an organisation a 30-day free trial on request, covering the platform fee and Nightshift but not coding agent seats. Measured usage is shown on Admin → Billing from the first day. On the 1st of each month Seylo issues the invoice for the previous month’s platform fee and verified fixes and emails it with a payment link; you pay through Razorpay, on the Billing page or from the email. See plans and billing.
- Pricing: people are free and unlimited; Nightshift is $6 per verified fix. The platform is billed only per managed vCPU ($5 a month), and a first or last partial month is charged for the days it covers. There is no charge for people and no limit on how many you add, on any plan. Nightshift’s first 5 verified fixes are free, once per organisation, then $6 per verified fix, which includes the tokens its investigations use on Seylo’s model. Seylo’s MCP tools are covered by the platform fee.
- Coding agent seats are paid before use. $20 per seat per month: an organisation admin buys seats for named people, and a seat works once the payment is confirmed. A seat covers the calendar month (UTC) it is bought for and is not prorated; renewal opens on the 25th, and a seat not renewed stops at 00:00 UTC on the 1st until it is paid. Each seat payment has its own invoice, already marked paid.
- The coding agent runs on Seylo’s model by default. Each seat includes $5 of Seylo’s model (GLM-5.3) use per month, so there is no model to set up first. The model menu in the composer shows each model’s cost as a multiple (×); when the included use is spent, switch to your organisation’s own key from the same menu (top-ups are not available yet). On Seylo’s model one call is limited to 200,000 tokens of context and 32,768 of output. Tokens on your own key are billed by your provider, never by Seylo. See choose the model.
- Growing a volume always asks a person. Nightshift still proposes growing a nearly full volume by one bounded step, within the limits an admin set, but it never does so unattended any more, under a grant or earned autonomy: a larger disk costs money from then on and never shrinks back. See letting proven fixes run.
- Editing a cluster is agent-only, like adding one. A cluster saved with an older method (a cloud account or a kubeconfig) no longer shows credential fields: its Edit dialog says why and offers Switch to the agent. See connection methods.
- Namespace grants decide what a person sees, not only what they change. With scoping on, a person who holds grants sees only the namespaces those grants cover: in namespace lists, environment variables and ConfigMaps, the resource browser and its YAML, schedules, deployments and insights. Another namespace is refused with the reason. Approvals lists only the requests in their namespaces, and opening, approving or rejecting any other is refused, naming the grant they hold. A request for a whole cluster, with no single namespace, needs a grant over every namespace of that cluster. People with no grants see what they saw before. See where a role applies.
- A coding agent’s deploy refuses a credential in
env. A value in thedeployspec’senvthat looks like a password, key or token is refused, and the value is never repeated back. Have the agent callrequest_secret, so a person creates the Secret, then deploy withpropose_deployand manifests that read it throughsecretKeyRef. See the deploy spec. - The memory grant changes memory only. Under the memory grant, a model-sized change is applied without a person only when everything it sets is one container’s memory request or limit. A change that sets anything else as well waits for a person, and says why.
- Every diagnosis ends in a remedy. When Nightshift knows the cause, its Fix is never “nothing for Seylo to do”. It says Where the fix goes - this cluster, the owner’s Git repository, the application’s code, or outside Kubernetes - the exact Change, How it is applied (by Seylo once a person approves, as a pull request, by the coding agent, or by a person following the change) and what it is Verified by, with a confidence. A secret value is never written into a change. See the fix is a remedy.
- Nightshift’s own changes fit the workload. A generated probe change keeps the probe’s existing handler and corrects its port, instead of swapping an HTTP probe for another kind. A container is named in a change, never counted by position, and a field restated at its current value is left out instead of refusing the whole change. StatefulSets get generated changes as Deployments do. See generated fixes.
- A pending pod’s smaller request must schedule. When a pod is Pending because no node has room for its CPU or memory request, Nightshift proposes a smaller request only if a node the pod may run on has room for it - allocatable minus the requests already placed there - and never below the container’s measured use plus 20%. Otherwise it refuses and says the cluster needs capacity. See a request no node has room for.
- A failing probe is investigated with the network in view. The investigation opens with a read of the cluster’s network: which engine, if any, enforces NetworkPolicy, the namespace’s policies, and packets dropped by policy where those are measured. A policy that nothing enforces is ruled out, with the read that clears it, instead of being blamed. See the read menu.
- Nightshift has its own place. A Nightshift group in the sidebar,
right after Observability: Incidents, Fixes, Skills, Memory
(
admin), Autonomy and Settings. Admin → AI settings is gone: its parts are those pages, and the coding agent’s model is on Build → Coding agent → Model settings. Old bookmarks open the matching page. See where Nightshift lives. - Incidents, not alerts. Alerts that the service map connects, within minutes of each other, become one incident. Seylo walks it to the service where the failure started, lists who else is hit, and Nightshift investigates it once, across its namespaces. See incidents.
- The service map. Observability → Service map shows what each workload calls and what calls it, from your cluster’s objects, traces, a service mesh and your Backstage or Seylo architecture files. Every edge says how it is known. See service map.
- Follow one request. From a log line or a trace, see the path one request took, where it dropped and why. See follow a request.
- Health signals. Consumer lag, queue backlogs, dead letters, pools and database connections at their limit, slow queries, retries and circuit breakers, read from the exporters you already run, become the health of the map’s nodes, so a database that never alerts can be named as an incident’s root. See health signals.
- The fix on the root, through the right door. An incident proposes one
fix on its root: in the cluster; as a pull request to the owner’s
repository when Argo CD or Flux deploys it from GitHub (off until an
adminturns on Let Seylo open pull requests); or as a task for the coding agent when the fault is in the code. See the fix on the root. - Nightshift fixes what it diagnoses. When no fixed class of fix applies, its model may propose one change of its own - a probe, resources, a PodDisruptionBudget, an autoscaler and a few more fields - accepted only by a five-check guard: allowed fields, evidence for every field, no worse posture, a dry run, and a verification with an exact undo. It can create what is missing from evidence, and Admin → Posture has a production-readiness campaign that proposes one such change per workload, filed one at a time. See generated fixes.
- Memory limits sized from evidence. After an
OOMKilled, the new limit is sized from what the container declares (a heap flag,--vm-bytes), its crash log, or the same image running healthy elsewhere. A leak is refused: more memory only delays the crash. See memory limits sized from evidence. - Nightshift remembers. One short lesson per alert and incident it worked on, recalled and cited the next time. It belongs to your organisation, only admins see it, and it can be switched off, exported and deleted. A fix that keeps working can be drafted into a skill. See memory.
- Seylo’s core runbooks. 45 runbooks, one per kind of failure, the same for every organisation, each in the order mitigate, diagnose, fix, prevent. Your own skills come first, and a skill can now be scoped to a service on the map. Nightshift’s answer is laid out as Mitigation, Fix and Prevent, and names the runbooks and skills it used. See core runbooks.
- Mitigations. A capacity buffer, a rate limit or maintenance page on an ingress-nginx Ingress, or scaling down another workload, composed and checked by Seylo and always approved by a person. See mitigations.
- Rollback to the Deployment’s own previous revision. When a release made
with
kubectlor a pipeline breaks a workload, Nightshift proposes going back to the revision the cluster kept, askubectl rollout undodoes. It always waits for a person and never runs unattended. See rollback. - Recreate a claim whose StorageClass does not exist. A pod Pending on
such a claim is fixed by recreating the claim under the same name with your
default StorageClass - only a claim that never bound, only with an
approvertyping its name, and only with the agent’s destructive tier (rbac.write.destructive.enabled=true), which is off by default. There is no automatic undo, and the fix says why. See recreate a claim. - APM coverage and autonomous tracing. Traces & APM → Coverage shows
which workloads serve traffic and send no traces, and proposes one
OpenTelemetry Operator annotation for each, verified by spans arriving and
removed if they do not. Unattended only under a grant that names the
namespace and languages. Coding agents get the
apm_coveragetool. See APM coverage. - Alerts close on their own once the workload is fixed. An event last seen more than 10 minutes ago no longer holds an alert open, and the alerts of a deleted namespace are resolved with the reason, never deleted. One crash is one alert, even when Kubernetes reports it twice. See how an alert ends on its own.
- A workload created again starts clean. A workload deleted and created again under the same name no longer inherits its predecessor’s fixes and refusals.
- Limits on automatic investigations, and how to raise them. By default 40 a day, 10 an hour and 4 per cluster an hour per organisation. Seylo raises them for your organisation on request through Help → Contact support, and Nightshift → Settings shows your numbers and when and why they were changed. See ask for higher Nightshift limits.
- Faster assistants that do not time out. Ask AI and Build with AI run in the background: the page shows how long the model has been asked, with Cancel, and the answer says which model answered and how long it took. This fixes the 0.3.14 known issue.
- Investigations get more time. A run may take up to 12 minutes and each model call up to 3, so a reasoning model answers instead of the run falling back to the rules.
- Fixed: releasing a new image tag of a workload Seylo had not deployed
before failed with
Could not find an image: line; it now works, and sets the tag on the container you named. An alert that fires again after being resolved is judged from its new start, so a rollback after a new release is proposed again. A crash-looping pod caught running between restarts no longer makes a correct rollback look unneeded. - Cost dashboards. FinOps → Cost & savings → Dashboards holds Seylo’s FinOps templates, read-only, and your organisation’s own dashboards, copied from a template, started from scratch or written with Build with AI. The cluster templates show compute cost by namespace and workload, idle capacity, requested against used, and each week against the week before or an 8-week average, priced by Seylo from OpenCost. A period with no data is never drawn as zero: each panel says where its history starts and why anything is not priced. Whole-cloud spend, from your cloud bill, is coming; until it is connected, those panels say so and why. See cost dashboards.
- The service map in lanes. Left to right as a request flows: Entry (Ingresses, HTTPRoutes, and a CronJob or Job that calls something), Stateless (Deployments, in call order), Stateful (StatefulSets, and databases, caches and queues in the cluster) and Storage and outside (volume claims, outside hosts, SaaS and managed services). DaemonSets, and anything with no known edge, sit in a Background band below. The volume claims a workload mounts are nodes, joined to it by a mounts edge. Hover over a node, or move to it with the keyboard, to light its whole request path; chips show or hide each kind of node; Hide unconnected leaves out what has no edge; and All namespaces you can see draws one band per namespace. See how the map is laid out.
- Posture leaves Seylo’s own node-exporter out of the score, visibly. The node-exporter of the seylo-observability chart (0.2.19, below) needs the host network and read-only host mounts, which two of Seylo’s own policy rules flag. Those findings are listed as exempt: Seylo’s own with the reason, counted on their own row, left out of the score, and offered no fix, but only while the DaemonSet is as the chart installs it. See on the Posture page.
- Approval requests expire after 72 hours, and approving one that is out of date is refused: its namespace is gone, or the object a Nightshift fix was worked out from has changed since. Nothing is changed, and the request says why. See requests expire after 72 hours and what is checked again.
- Choose your own notification email. Under your name → Notifications, each kind of email is Immediate, Daily digest or Off, for you alone. Turning routed alerts off is shown in the channel’s delivery log. See your notification email.
- Report an AI response. A coding agent’s replies, Nightshift’s runs and the answers on the Metrics and Logs pages have a Report button that files a support request with the response, redacted. See report an AI response.
- The assistants stay on your cluster. Ask AI and Build with AI answer questions about your clusters’ metrics, logs, traces, costs and workloads. A request that is plainly something else - an essay, a translation, homework, an unrelated coding exercise - is refused before it reaches the model, with what the assistant is for. Each person may ask each assistant a set number of times a day: 100 for Ask AI on each page, 40 for Build with AI. See what the assistants answer.
- Nightshift looks up metric names before it queries. While it investigates, it can list the metric names in the alert’s namespace that contain a word, or one metric’s labels and values, instead of guessing a name that returns nothing. Each lookup counts as one of its own queries.
- A new symptom in an incident is investigated. An alert that joins an incident after Nightshift’s run started, with a kind of failure none of its alerts had then, starts one follow-up run of the incident. The incident is walked again first, and Nightshift is asked whether it is the same root or a separate failure. See Nightshift investigates the incident once.
- A problem that comes back after a verified fix is investigated again, even within the 6 hours that otherwise separate two automatic runs of the same alert: once for each verified fix. See how often it starts on its own.
- Fixed: follow-up runs of an incident, added during this release, failed to start without saying so; they now start. A fix to a Service’s selector is checked against that Service’s own endpoints; before, the check could wait on another Service and undo a correct fix. A model call that fails with a server error or a rate limit (HTTP 5xx or 429) is retried once, so one provider error no longer ends a run on the rules alone. Nightshift’s answer is no longer thrown away when the model puts an unescaped quote in it, or spends its whole reply reasoning: the call is made again at a lower reasoning effort. A proposal that only waits for a person’s change, or does nothing on its own, is no longer offered. APM coverage no longer says the OpenTelemetry Operator is missing when it is installed, and a pull request it cannot open says whether the Argo CD Application was not found, not readable, or Argo CD is not on the cluster, and keeps the other ways to instrument.
- Works best with agent and chart 0.2.19, below. Every older agent keeps working.
0.3.15 — 2026-11-01
Section titled “0.3.15 — 2026-11-01”A security release. Three settings now refuse values they accepted before; each refusal says what to use instead. If none of them applies to you, nothing changes for you.
- Clusters connect through the agent only. Adding a cluster, changing
one or testing its connection with any method other than
In-cluster agent - a cloud provider’s credentials, a pasted kubeconfig
or a kubeconfig context - is refused with
Connection method "<method>" is not available on this server. A cluster saved earlier with one of those methods is refused when Seylo would reach it, too: edit it, choose In-cluster agent, and install the agent. See connection methods. - Image registry settings. A registry’s credential variable names are no
longer saved from the portal or the API. The ACR type is used only for an
Azure Container Registry host (such as
*.azurecr.io) and the GAR type only for a Google host (*-docker.pkg.devorgcr.io); for any other registry, choose OCI. Seylo reaches a registry only at a public HTTPS address, does not follow its redirects, and quotes only the start of an error the registry returns. See the image registry. - Identity provider issuers. A SAML entity ID or OpenID Connect issuer
that starts with
local://orbreak-glass://is refused when you save the provider: Seylo writes those itself for its own sign-ins. So is an issuer that Seylo’s own staff sign in through. Use the entity ID or issuer your identity provider publishes in its metadata. See what Seylo enforces.
0.3.14 — 2026-10-22
Section titled “0.3.14 — 2026-10-22”- Heads-up: your existing uptime checks now raise alerts. A check that is
down raises one alert, through your notification channels and routes, and
resolves it when it passes again. Every check you already had alerts at
criticalafter 2 failed checks in a row; change either per check, or turn the Uptime checks detector off under Observability → Alerts → Detectors. See uptime checks. - Agents connect to
agent.seylo.in. New install commands point the agent atwss://agent.seylo.in/agent/connect, a host for agent connections only, which keeps long-lived connections steadier. If your egress is restricted, allowagent.seylo.in:443. Agents you already installed keep working onportal.seylo.in, with nothing to do; to move one, see move an agent. - Import from Uptime Kuma: a 1.x backup file, or the text of its
/metrics. HTTP and keyword monitors become uptime checks; every other monitor and every notification is listed with the reason it was not imported, and credentials are never imported. See import from Uptime Kuma. - Nightshift writes its own queries. While it investigates, it can write metrics, logs and trace queries and read-only Kubernetes reads, always limited to the alert’s namespace by the stores, and every answer is numbered evidence its findings cite. A run may now make 18 reads. See queries Nightshift writes itself.
- Teach Nightshift your runbooks. Organisation skills:
SKILL.mdfiles in the open Agent Skills format, written under Admin → AI settings → Nightshift → Skills (from 0.4.0, Nightshift → Skills) or synced read-only from one Git repository. Scoped by cluster, namespace, labels and alert, cited as evidence, and never run. See skills. - On
portal.seylo.in, Seylo’s model is GLM-5.3, on Azure AI Foundry in Azure’s US Data Zone, for Nightshift and the Ask AI and Build with AI assistants. Your own key stays the alternative, under Admin → AI settings → Nightshift (from 0.4.0, Nightshift → Settings). If your organisation has no key of its own saved and did not choose Your own key, Seylo’s model now answers, so Nightshift investigates new alerts on its own unless anadminturns that off. See Nightshift’s model. - Known issue: with Seylo’s model, Ask AI and Build with AI can keep waiting without an answer, because the model takes longer than the page’s request may stay open. Nightshift’s investigations are not affected - they run in the background. The fix, in the next release, runs the assistants in the background too.
- Calmer pages. Posture opens on the score and the findings, with the rule list one click away. Billing and Admin → Integrations → GitHub show the facts, with each explanation behind an (i).
0.3.13 — 2026-10-21
Section titled “0.3.13 — 2026-10-21”- Follow your support requests. Help → My support requests shows each request’s status and the whole conversation, and you can reply or close it there. See Get help.
- A calmer Admin → AI settings (from 0.4.0, the pages of the Nightshift group). Each section is a title with an (i) that explains it. Fixes are listed with their risk, how often each was proven here, and one choice: Ask first or Auto once proven. Standing grants are managed under FinOps → Autoscale. See Fixes, class by class.
- Billing says plainly that there is nothing to pay during the free preview, and when it ends.
0.3.12 — 2026-10-21
Section titled “0.3.12 — 2026-10-21”- Help, in the portal. A Help menu on every page links to this site and opens Contact support: a short form with up to five screenshots that shows you what context it adds before you send, refuses anything shaped like a credential, and gives you a reference. See Get help.
- Nightshift’s checks run on portal.seylo.in, including the lost-node alert.
0.3.11 — 2026-10-20
Section titled “0.3.11 — 2026-10-20”- Cloud accounts, in early access: connect an Azure subscription, an AWS account or a Google Cloud project on Admin → Integrations → Cloud accounts, and Seylo shows what runs in it. The connection is read-only and keyless: you grant a read-only role that trusts Seylo’s identity, and you revoke it by deleting that role. Seylo refuses a grant that could change anything or read your data. Connecting a cluster through the agent still needs no cloud access at all.
0.3.10 — 2026-10-20
Section titled “0.3.10 — 2026-10-20”- A dropped agent connection costs less. A read that was lost when the connection dropped waits up to 15 seconds for the agent to reconnect and is asked once more. If that fails too, the error says it was already retried. A change lost the same way is not sent again, and its error says it may already have been applied: read the resource’s current state before you try again.
- Clusters shows how steady each agent’s connection is: the Agent: Connected chip’s tooltip says how many times it reconnected in the last 24 hours, how long the last connection lasted, and which side noticed the drop.
- Nightshift on a lost node. A node that has been NotReady for one minute raises an alert. Nightshift explains what your platform is doing about it: AKS node auto-repair, Karpenter or EKS Auto Mode, or the cluster autoscaler. For EKS managed node groups and GKE, it says that the repair cannot be seen from inside the cluster. It does not cordon the node or force-delete its pods while the platform repairs it, and the alert closes only after Nightshift has verified that the node and its workloads are healthy again.
- Nightshift finds a workload’s Services by listing them, so a Service that is not named after the workload is found too.
- Logs: with Source: Rehydrated archive, the facet counts come from the archive. Before, they counted the live store.
- Coding agents: the Codex setup command on the token screen exports
SEYLO_MCP_TOKENinstead ofSEYLO_TOKEN. See connect your coding agent. - Works best with agent 0.2.17, below. Every older agent keeps working.
Agent and chart
Section titled “Agent and chart”The agent and the seylo-observability chart share version numbers and are
released together, but you upgrade them separately. See
upgrade the agent and
upgrade the chart.
You are never required to upgrade in step with Seylo’s server: an older agent keeps doing everything it could do before, and a feature it lacks is reported as needing a newer version. Releases marked fixes a defect fixed something broken in older agents. A cluster whose agent is older than the oldest release not known to be broken — today 0.1.8 — is told to upgrade by email as well as in the portal.
The latest agent is 0.2.22. Its signed image index, covering linux/amd64
and linux/arm64:
ghcr.io/seylo-inc/seylo-agent:0.2.22@sha256:ed9222f8b6e0beb0fb12873388439ce270304bc5b54dc41a80a96707c552661fVerify it before you install it.
0.2.23 — next release
Section titled “0.2.23 — next release”Chart only: the agent itself is unchanged. seylo-observability 0.2.23 is
published with it.
- Stores read the Retention page’s choice. Each store’s retention comes
from the chart’s defaults first and then from the
seylo-retentionConfigMap that the Retention page writes, which Helm never renders, so a laterhelm upgradedoes not put the old period back. Upgrading restarts each store once. See choose how long to keep data. - The archive copies every signal, in slices. One file per signal per
finished hour by default, with
archive.cadence.logs,archive.cadence.metricsandarchive.cadence.traces(1h,15m, and5mfor logs),archive.lagMinutes(15) andarchive.metrics.enabled, which is off by default so an upgrade never starts copying metrics unasked. Traces are copied as spans. See how often to copy. Known issue on AKS: node-exporter and port 19100. AKS’s node image runs its own node exporter on port 19100 of every node, so chart 0.2.19 to 0.2.23 starts node-exporter pods that crash-loop withlisten tcp <node-ip>:19100: bind: address already in use, andhelm upgrade --waitfails. The stores themselves are not affected. If you upgraded to any of them on AKS, set the port yourself:--set nodeExporter.port=19101. From chart 0.2.24 the default is
- See node-exporter on AKS.
0.2.22 — 2026-11-11
Section titled “0.2.22 — 2026-11-11”Image index sha256:ed9222f8b6e0beb0fb12873388439ce270304bc5b54dc41a80a96707c552661f
(linux/amd64 and linux/arm64).
- Read-only access for what Nightshift fixes next.
getandliston PodDisruptionBudgets (policy), Jobs and CronJobs (batch), ResourceQuotas, and cert-manager’scertificates,certificaterequests,issuers,clusterissuers(ClusterRole only), and ACMEordersandchallenges. In the ClusterRole and in each watched namespace’s Role, every posture included, no Secret and no write. See what 0.2.22 adds. - On an older agent, Nightshift cannot read budgets or certificates and says the read was refused. Nothing changes in the protocol, so an older agent keeps working.
seylo-observability0.2.22 is the same as 0.2.21.
0.2.21 — 2026-11-10
Section titled “0.2.21 — 2026-11-10”Image index sha256:3c9fb7f45335318587b77ed081cd50920765bacd26332ccfb6e170f333f93b59
(linux/amd64 and linux/arm64).
- Your node autoscaler, read. The agent reads Karpenter and AKS node
auto-provisioning NodePools and NodeClaims, the AKS, EKS and EKS Auto Mode
node classes, and GKE ComputeClasses, so the Savings tab can say which
autoscaler manages your nodes and what a cheaper or spot pool would change.
New read-only RBAC in the ClusterRole:
get,listandwatchon those kinds. A node class’s bootstrap script is never returned. rbac.write.autoscaler.enabled, off by default. Lets Seylo create and change those NodePools and node classes when you switch on managed savings with spot or better sizes. It never deletes one - deleting a NodePool drains every node it owns, so that stays your step - and the agent refuses such a delete itself. No posture turns it on. A write it is not allowed names the setting.- Chart:
seylo-observability0.2.21 is the same as 0.2.20.
See what 0.2.21 adds.
0.2.20 — 2026-11-10
Section titled “0.2.20 — 2026-11-10”Image index sha256:397c12c00e59d9fc34f769f889865b5ff97dbd3047918cf01e542f337b4196f1
(linux/amd64 and linux/arm64).
- Key Vault secret versions, never values. The agent reads which version
of a Key Vault secret each pod mounted through the Secrets Store CSI
driver, and ServiceAccounts, so Nightshift can tell a pod still holding an
old version from a wrong value, and check a workload identity. New
read-only RBAC:
getandlistonsecretproviderclasspodstatuses,getonserviceaccounts. See what 0.2.20 adds. - A manifest with no namespace is refused instead of landing in the agent’s own namespace, and a cluster-wide object cannot be written by naming a namespace on it. Seylo always names the namespace, so nothing you deploy through it changes.
- A namespace-scoped agent says what it refused: a list across all namespaces, or a cluster-wide kind, and which setting limits it.
- Chart: the network-costs collector is removed.
opencost.networkCosts.enabled=truenow fails with the reason: its image is not open source. See network traffic.
0.2.19 — 2026-11-02
Section titled “0.2.19 — 2026-11-02”Image index sha256:d75a1ad150c54397dbf1d6b4a03ac3fa07af979625f4f9aabfa5ea4f639e1861
(linux/amd64 and linux/arm64).
-
Rolling back a change made outside Seylo removes what that change added. Seylo writes with server-side apply, which only removes fields Seylo set, so a
command,args, label or annotation another tool (such askubectl) added survived the rollback and the pod failed exactly as before. The agent now removes those fields from the pod template after the rollback is applied, whichever way it was approved. An older agent cannot, so Seylo refuses such a rollback before writing anything and says which version to upgrade to, or how to roll back by hand. -
Heads-up: if you forward traces through your own OpenTelemetry Collector, or Beyla or OBI, to port 4317 or 4318, move it to the forwarder port - 4328 (HTTP) or 4327 (gRPC) - and list that collector’s namespace in
traces.trustedForwarders.namespaceswhen you upgrade theseylo-observabilitychart. Until you do, its spans are attributed to the collector’s own namespace (or to none, if it runs on the host network), so people limited to some namespaces stop seeing those services. See which namespace a span belongs to. -
Every served version of an API group is read, not only the version the group prefers. An operator whose resources are served only at another version is found: the OpenTelemetry Operator’s
Instrumentation, atv1alpha1, read as not installed before. With an older agent, Seylo calls such a kind unknown rather than absent. -
An out-of-memory kill from days ago no longer reopens its alert every time the agent reconnects. A kill is judged by when it ended, so each new kill still alerts at once.
-
A rejected credential is retried calmly. An agent whose credential Seylo refuses - expired, already used or revoked - waits up to a minute between attempts instead of retrying within seconds. Only a new credential fixes it: see
auth-invalid-credentialandauth-agent-revokedin troubleshooting the agent. -
The YAML library the agent uses is updated to a release with a security fix.
-
Whole-cloud cost reads (OpenCost’s
/cloudCost, whole-cluster agents only) take exactly the parameters OpenCost reads, and accept cloud tag keys as your bill writes them, such ascost-center. -
The agent’s ClusterRole can
getandlistthe OpenTelemetry Operator’sinstrumentations, so APM coverage can see which Instrumentation the operator would inject from. A namespace-scoped install does not get this. -
Chart: host metrics, on by default. A node-exporter DaemonSet on every Linux node feeds the Hosts dashboard: port 19100 on the node’s own address, five collectors, user 65534, on the host network, not privileged, no host PID namespace, and
/proc,/sysand/mounted read-only. Turn it off withnodeExporter.enabled=false. See host and ingress metrics. -
Chart: ingress-nginx metrics without annotations. Controller pods with the standard ingress-nginx labels are scraped on port 10254 once their metrics are on. See ingress metrics.
-
Chart: a span’s namespace is the pod it came from, worked out from the connection, not from what the span claims, so no service can put its spans in another namespace’s view. Your own collector or eBPF agent sends to the forwarder ports, which exist only for the namespaces you list.
0.2.18 — 2026-10-25
Section titled “0.2.18 — 2026-10-25”- Nightshift sees pods’ and containers’ real times again - when a container started and finished, when a pod or workload was created. 0.2.16 and 0.2.17 sent them empty, so a pod killed during start-up, or stuck on an old revision, could go undiagnosed. Until you upgrade, Seylo fills what it can from your metrics store and never guesses a time.
- One crash-looping container raises one alert, and alerts close once the workload is fixed instead of being held open by old events.
- Read-only access to PriorityClasses (
getandlist, in the ClusterRole only), so Nightshift can check for a capacity buffer before proposing one. See what 0.2.18 adds. - When Nightshift undoes a fix that created an object, the API server checks the object’s UID before deleting it.
- Upgrade with
--set image.tag=0.2.18as well as the digest: reused values keep an earlier tag. See upgrade the agent. seylo-observability0.2.18 is released at the same version with no change to its templates or defaults. You do not need to upgrade the chart for this release.
0.2.17 — 2026-10-20
Section titled “0.2.17 — 2026-10-20”- The agent pings its connection to Seylo every 15 seconds. When nothing comes back within 10 seconds, it closes the connection and dials again. A connection that a proxy or load balancer drops without closing it is noticed in about 25 seconds, instead of at the agent’s next heartbeat. It works with every Seylo server.
- Alerts, metrics and command results produced while the agent was disconnected are kept and sent once it has reconnected. Older agents dropped them. The agent keeps a bounded amount and drops the oldest first.
- Two new chart values:
connection.pingSeconds(default15;0turns the ping off, for a network path that drops WebSocket control frames) andconnection.pongTimeoutSeconds(default10, kept belowpingSeconds). See what 0.2.17 adds. - Seylo can tell the agent not to send usage that the cluster’s metrics store already holds. Seylo does not ask for that today, so nothing changes yet.
seylo-observability0.2.17 is released at the same version with no change to its templates or defaults. You do not need to upgrade the chart for this release.
0.2.16 — 2026-10-20
Section titled “0.2.16 — 2026-10-20”- Nightshift reads a Service’s endpoints and a StorageClass: it finds a Service with no endpoints without the observability chart, and can propose growing a nearly full volume whose StorageClass allows expansion. Seylo also sees each object’s own status, for example a pod’s last termination reason. Secrets are never read.
- New read-only RBAC,
getandlistonly:endpointsanddiscovery.k8s.ioendpointslices(the ClusterRole, and the Role in each watched namespace), andstorage.k8s.iostorageclasses(the ClusterRole only). See what 0.2.16 adds. seylo-observability0.2.16: a retention period per store, a nightly archive of finished log days to your own bucket with workload identity, and rehydrate. Archive and rehydrate are off by default; with them off, the chart renders the same stores as 0.2.15.- Search archived log days on the Logs page: load them with the Retention page’s rehydrate, then choose Source: Rehydrated archive. Each person still sees only their namespaces.
0.2.15 — 2026-10-15
Section titled “0.2.15 — 2026-10-15”- Connect an in-cluster Grafana, Prometheus or Alertmanager
to Seylo by its Service name and import from it. The key stays in a Secret
in your cluster (
seylo-connections) that only the agent reads; Seylo never stores it. - The agent reads only fixed, read-only paths of those tools, never follows redirects, and reduces Alertmanager’s configuration to routing before anything leaves the cluster.
0.2.14 — 2026-10-15
Section titled “0.2.14 — 2026-10-15”- Resources no longer shows a 403 for CustomResourceDefinitions: the agent may read CRDs (they are schemas, never data), so custom kinds are known rather than guessed. See RBAC posture.
seylo-observability0.2.14 stops the log store being OOMKilled under query load: 2 GiB with bounded caches, at most 4 queries at once, and a Go soft memory limit on every store. The log collector gets 512 MiB.- Structured log lines whose message is under
@m,@mt,MessageTemplate,event,errorand similar keys no longer read “missing _msg field”. - Replaces 0.2.13, whose observability chart could not be upgraded with plain
--reuse-values. Upgrade straight to 0.2.14 or later.
0.2.12 — 2026-10-14
Section titled “0.2.12 — 2026-10-14”seylo-observability0.2.12 fixes doubled “of allocatable” figures on dashboards when OpenCost is on, and gives StatefulSet pods their controller in OpenCost’s allocation. The agent itself is unchanged from 0.2.11.
0.2.11 — 2026-10-14
Section titled “0.2.11 — 2026-10-14”- Chargeback by team, label or namespace from OpenCost in your cluster, with efficiency and CSV export. Rows are filtered to each person’s namespaces in the agent, before anything leaves the cluster.
- Needs
seylo-observability0.2.11, which lets the agent reach OpenCost’s API. seylo-observability0.2.11 also adds optional billing integrations, all off by default: the Azure price sheet, the Azure cost export, and Datadog and OpenAI spend. Keys stay in Secrets you create, or an identity you grant.
0.2.10 — 2026-10-13
Section titled “0.2.10 — 2026-10-13”- Optional Datadog connector: read your Datadog in place and compare it with Seylo’s own collection. Keys stay in a Secret in your cluster.
- Import Alertmanager routes and Datadog monitors, dashboards and synthetic tests into Seylo; credentials never leave your cluster.
- Optional: let the Seylo Agent run a fixed list of read-only
commands in containers while it
troubleshoots (
cat,ls, environment variable names,df,ps, listening sockets, in-cluster DNS and HTTP status). Off unless you setsre.exec.enabled=trueand turn it on for the cluster in Seylo; no shell, output redacted in the cluster. - Turning on OpenCost with
--reuse-valueson an olderseylo-observabilityinstall works; 0.2.9 refused it and changed nothing. seylo-observability0.2.10 also adds OpenCost’s network-costs collector as an opt-in (opencost.networkCosts.enabled), off by default: it is privileged and uses the host network, and its image is not open source.
0.2.9 — 2026-10-09
Section titled “0.2.9 — 2026-10-09”- Upgrading the
seylo-observabilitychart with--reuse-valuesfrom 0.2.7 or earlier works again; 0.2.8 refused to render that upgrade and changed nothing. - The agent itself is unchanged from 0.2.7.
0.2.8 — 2026-10-09
Section titled “0.2.8 — 2026-10-09”- The
seylo-observabilitychart writes far less to disk: it stores only the kubelet and cAdvisor metrics Seylo reads, and saves every 30s instead of every 5s. - Optional OpenCost in the
seylo-observabilitychart (opencost.enabled) for cost per namespace and workload, computed in your cluster. Read-only, and no node-exporter. - The agent itself is unchanged from 0.2.7.
0.2.7 — 2026-10-09 (fixes a defect)
Section titled “0.2.7 — 2026-10-09 (fixes a defect)”- Upgrading an existing install with
--reuse-valuesno longer fails: a carried-overcustomResources: truegrants no custom-resource writes instead of being refused.
0.2.6 — 2026-10-06
Section titled “0.2.6 — 2026-10-06”- The default install no longer grants write on every resource type; custom-resource writes are now an explicit, opt-in list. See upgrading from before 0.2.6.
- A namespace copy of a Key Vault workload needs no extra Helm values: SecretProviderClasses and PVCs are readable, and creatable without delete, by default.
- The agent reports only namespaces in its watch list that actually exist.
- Key Vault (CSI) secrets can be converted to Kubernetes Secrets in a copy, with values carried inside the cluster.
0.2.5 — 2026-10-04
Section titled “0.2.5 — 2026-10-04”- With the observability chart installed, usage is read from its store instead of scraping every kubelet a second time.
- An agent installed for some namespaces now reports CPU and memory usage, read from the store.
- The observability chart keeps pod network I/O and node CPU and memory, so those dashboard panels have data.
0.2.4 — 2026-10-03
Section titled “0.2.4 — 2026-10-03”- Reports the namespace list its own install fixed, so a copy or a new namespace it would refuse is caught before you submit it.
- When it refuses a namespace, it says whether its install or Seylo set the limit, and what to run or change.
0.2.3 — 2026-10-03
Section titled “0.2.3 — 2026-10-03”- Copying a namespace within one cluster carries Secret values over inside the cluster: only key names reach Seylo.
- Reports each Deployment’s container port, so a new namespace can start from existing components with ports filled in.
0.2.2 — 2026-10-03
Section titled “0.2.2 — 2026-10-03”- Logs are collected by vlagent, VictoriaLogs’ own collector, instead of Fluent Bit: one vendor from collection to storage.
- Namespace filtering reads the namespace field vlagent writes, and still matches logs collected before the switch.
0.2.1 — 2026-10-02
Section titled “0.2.1 — 2026-10-02”- Finds the
seylo-observabilitystores by itself — no URLs to enter. - Traces and APM are filtered by namespace, so an agent installed for some namespaces can show them.
0.2.0 — 2026-10-01
Section titled “0.2.0 — 2026-10-01”- Queries the in-cluster VictoriaMetrics, VictoriaLogs and VictoriaTraces stores for the Observability pages.
0.1.8 — 2026-09-20 (fixes a defect)
Section titled “0.1.8 — 2026-09-20 (fixes a defect)”- Removing a ConfigMap key or an environment variable works on objects created by kubectl, Helm or Argo CD. On 0.1.7 and earlier, such a removal was accepted and the key was still there afterwards.