Skip to content

Operate

Changelog

Versioned against CONTRACT.md's own revisions, not against a release calendar. This documentation set doesn't print a date it can't verify from the repository itself. Where "in progress" stands in for a date below, the repository holds no date to verify, and the gap is deliberate. Current version: 0.1.0 (.env.example → VERSION; also Chart.yaml's appVersion).

Log analytics that measure (unreleased)

Added

  • Log analytics measure as well as count: the unique count of a field, the average, minimum, maximum, sum, p50, p90, p95 or p99 of a number a line carries (an attribute or a JSON message's own key, such as duration_ms), or the error ratio, by any field, ranked highest first, in its unit, with a row for all lines in the table. See logs.md.
  • @status:503 finds a line's HTTP status wherever it writes it: an attribute, a JSON key or an access line. The status facet lists the same, 4xx and 5xx in colour. A key with no dot also reads a JSON message's own key (@route:...).
  • shape:"<pattern>" finds exactly a pattern's lines. A pattern's plus, minus and lines, the error strip's top error and the dashboards' log panels use it.

Changed

  • A pattern keeps a JSON line's keys only when the request failed, so the 2xx routes share one row and the failures stay on the first screen, and FATAL, ERROR and WARN patterns come first.

Fixed

  • A pattern's minus no longer hides the failed requests that share its message, and its plus no longer counts them.
  • Lines stored before the engine read levels from text get theirs read when you query them, so the histogram, the facets, the patterns and the error strip of a window use one level source across that moment.
  • A measured top list or pie shows each value in its unit (111 ms) and a share only for a count or a sum. A measured time series rolls up to a third of the points, puts its unit on the axis and in the legend, and leaves the end values to the legend. A measured table's trends leave an empty bucket as a gap, and its extra measures come from the same read, so the Other and All lines rows fill in. A nested split breaks a tie on its measure by lines and names a line without a level "No level". @status:ok finds an app's own status again.
  • Error and warning patterns stay on the list when busier INFO patterns fill its top 50, and a pattern row reads its dates' formats and its numbers' ranges from the newest lines before you open it.

Log patterns that say why (unreleased)

Added

  • A log pattern keeps a JSON line's HTTP status, method, route or path template and error class beside its message, and an access line's HTTP status, so a 503 is a row apart from a 200: http request status=503 route=GET /api/.... The error strip's top error reads the same way, and its click and a pattern's plus find exactly those lines. See logs.md.
  • A line in the list shows its status and error keys right after its message, a 4xx or 5xx status in colour, with its keys dimmed so the values read first. An open line chips its request: method, route, status and duration.
  • The facet rail has "Search facets" across facet names and values with "Showing N of M", and counts from 10,000 up read short. The live tail says what share of the lines it shows when more arrive than a read returns.
  • The first screen shows more lines: the saved views sit on the query row, the how-to line moved into the help, the histogram is shorter with the error strip on its legend row, and the view tabs head the results.

Fixed

  • An error burst starts at its first busy bucket: a few stray errors before it are no part of it.
  • The live rate averages the last 4 whole buckets. An error row keeps its level bar without a pink tint. service:a, reads as service:a, as the engine reads it.

Log facets as Datadog's, patterns by level and service (unreleased)

Added

  • The log facets count each one under the query without its own terms, so a checked value leaves the others listed. A click joins a value to the checked ones (service:cart,checkout, any of them), "Only" keeps one, the minus excludes one, and a facet with more than 10 values has a search box. New facets: workload, container, node, image tag, version and HTTP status code. The query language takes key:a,b for any of the values. See logs.md.
  • Patterns are a row per shape, level and service, as Datadog splits them; a row's plus adds its level and service, and an open row lists what each * held in its newest lines.
  • The error strip names a burst that already ended and since when the errors are back to their rate.
  • A level read from Valkey's and Redis's marker, and from failed:.

Fixed

  • The ingestion rules act on the level a line carries, never on one read from its text, so a rule that drops a level keeps the lines it kept before.
  • The error strip no longer calls a burst that ended a step, and its reads run beside the page's instead of after it.
  • The dashboards' count by level stacks errors on top, as the explorer's histogram. The histogram's axis counts whole lines. A pattern's share carries one decimal.

Log levels read from the line (unreleased)

Added

  • A log line that arrives with no level gets the one its text names: klog prefixes, HTTP access lines by status, level= and JSON levels (pino's numbers too), [WARN], ERROR: (Postgres's UTC [pid] ERROR: among them), .NET's fail:, Kafka's ] INFO [. An indented line or Postgres's STATEMENT: takes the level of the line before it from the same container. The engine stores that level, so the histogram, the facets, severity:, the error strip, the patterns, the analytics and the ingestion rules count it, and the list marks it with a ring. Lines stored before keep none. See logs.md.
  • A pattern names its worst level and the service that wrote most of its lines over the whole window, and opens to its newest 20 lines with their time and level.

Fixed

  • The error strip finds a step after an early burst: the earlier rate is the interquartile mean of the earlier buckets, and the step is the earliest bucket from which the average holds at 3 times that.
  • A saved log view keeps the analytics' field and top. The histogram stacks errors on top. A new query no longer shows the last one's services and top error under the error strip. Shares carry one decimal, a pie offers top 5 or 8, and on a phone the analytics table drops its trend column before its counts.

Log analytics and errors at a glance (unreleased)

Added

  • The log explorer counts every line the query matches by one field, as a time series, a top list, a table with each value's trend, or a pie, beside Lines and Patterns. It is one GET /api/v1/logs/aggregate read, the one log dashboard panels make. See logs.md.
  • Under the histogram, the window's error lines: how many, since when they stepped up, the 3 services that wrote most of them and their most frequent shape, each a click to those lines.
  • A log row carries its level's colour down its left edge, and an error row a red tint. A pattern names the service that wrote most of its lines. An open line walks to the next with the arrow keys and gives its time in full with how long ago. Live shows lines a second.

Workloads map (unreleased)

Added

  • The workloads list opens as a map (Map beside Table, or ?view=map): one square per pod, grouped by namespace, node or workload, colored by status, by CPU or memory against its request or limit, or by restarts. The group with the most failing pods comes first and so does its worst pod. The facets and the search filter the map as they filter the table, the legend counts each color and keeps one at a click, and ?group= and ?fill= keep the choice in the link. Pointing at a pod reads its status, node, restarts with the last exit reason, and usage against request and limit; a click keeps it there with links to the pod, its logs and its workload's panel.
  • A workload's YAML tab, on its page and in the workloads drawer, prints the controller as the API server holds it, less managedFields and the last-applied copy, with every value a credential would hold masked as a diff masks it. It needs config.read and reads GET /api/v1/clusters/{id}/workloads/{ns}/{kind}/{name}/manifest only when the tab opens; each read is audited as workload.manifest.view.

https behind a CDN (unreleased)

Fixed

  • An ingress with no TLS behind a CDN overwrites X-Forwarded-Proto with http, so web set its cookies without Secure or __Host- and printed http:// in the MCP snippet and the SSO redirect URIs. publicUrl now reaches web, and an https address there outranks the header. Empty, nothing changes. See configuration.md.

Upgrade notes

  • An install with an https publicUrl whose ingress says http signs everyone in once more at the upgrade, and again at a rollback: web does not read a session under orbit_session on https.

What a service's page counts (unreleased)

Fixed

  • Workers and crons reporting an API's service.name no longer read as the API. Time by dependency counts only what callees took inside its requests, instead of showing nearly all of it on Redis. Calls and the service map's panel and labels lead with the calls made inside a request and name the rest as background. Recent traces, recent errors and error types list the traces a request started. GET /api/v1/apm/servicemap adds backgroundRate to an edge, and three series join the service graph (apm.md).
  • The service page's pods count reads Pods sending spans: it counts the pods whose spans carry the name, not the replicas of the workload above it.

What reads as a secret (unreleased)

Fixed

  • The Configuration tab showed a value in clear when its name said secret in words the rule did not know: a passphrase, DB_PASS, API_SIGNATURE. More names now mask, there and in Nocturne's tools, a diff, a terminal's command line and a database column: passphrase, passcode, *_PASS, a name ending in SIGNATURE or _SIG, signing and encryption keys, HMACKEY, MASTERKEY, PRIVKEY, KEY_PEM and its like, PASS_1, SSHPASS, keytool's -storepass, Redis's --requirepass, *_CREDS, BEARER, BASIC_AUTH, *_SALT, *_PEPPER, SENHA, SEGREDO, CHAVE. Each keeps its audited Reveal.
  • A password inside a value masks too: redis://:x@host, a Go MySQL DSN, an Oracle JDBC URL, Azure's AccountKey=, a signed URL's signature, dbPassword: x, --db-password=x, an upper-case *_KEY=x in a text.
  • An initialization vector, KEY_ID and SIGNATURE_VERSION stay in clear: none is a secret. PUBLIC_KEY and TOKEN_TTL stay masked, as before: one click to read.
  • A patch Nocturne checks no longer shows a secret in its diff when the patch drops or adds a whole list or object, such as an env list removed with its container or a ConfigMap's data, or edits args that hold --password x. A host or an image named like a credential (auth-token:8080, registry.example/team/app-token:1.2.3) and a scheduling term's matchLabelKeys stay readable.

A lighter Deck in the browser (unreleased)

Changed

  • Through web, a JSON answer of 1 KiB or more crosses gzipped when the request accepts it. A trace of 2,000 spans goes from 845 to 80 KB.
  • A first load fetches four font files, not six: the italic and the mono's weight 500 load on the first screen that draws them.
  • Log pages load their regex engine with the first /regex/ term. Until it arrives, about one round trip, that term is not marked in the lines.
  • A dashboard loads its editors, its settings and its log panel code when it needs them, and the savings list its scale form when a drawer opens.

Trace views (unreleased)

Added

  • The flame graph zooms with the wheel, pans with a drag and walks by keyboard. Spans that ran at the same time stack instead of hiding each other, and it draws only what is on screen, so a trace thousands of spans deep no longer outgrows the canvas.
  • Operations, a third view of a trace: calls, self time and critical path per service and operation, and a comparison against another trace or a typical one (?compare=<id>). See apm.md.
  • Self time colours, Root errors, a fold down to the hops between services, and the explorer's terms in a trace's find box.
  • A span's attributes group by namespace, and each copies its value, finds its spans in the trace and searches the explorer with or without it.

Fixed

  • Walking spans with the drawer open kept moving the keyboard into the drawer, so only the first arrow key walked. It also sent the drawer's server reads on every step; they now wait until the cursor rests for 250 ms.
  • A service in a trace's legend found every span whose attributes named it. It now finds that service.

Application log files (unreleased)

Added

  • telemetry.logs.files: orbit-node reads log files that applications write on a hostPath, such as Laravel's storage/logs/laravel.log and its daily laravel-YYYY-MM-DD.log, and joins each entry with its stack trace. Each source's directory is mounted read-only; / and system directories are refused. See logs.md, "Application log files".

Fixed

  • orbit-collector no longer gives a log resource that names a file on the node (log.file.path) the pod of the connection that sent it.

Alert notifications (unreleased)

Added

  • Alerts → Notifications: an admin posts the Deck's alerts to Slack, Microsoft Teams (Workflows) or Google Chat through incoming webhooks, with routing by severity, cluster and namespace, a wait before the first message, repeats, resolved messages, emoji per severity, a title template, a Slack mention and a link back to the Deck. Acknowledged and silenced alerts page nobody. See alert-notifications.md.
  • alerting.notifications.enabled, alerting.interval, alerting.allowedHosts and publicUrl in the chart; orbit_api_alert_notifications_total{kind,result} on orbit-api's metrics.

CloudWatch metrics (unreleased)

Added

  • cloudwatch.enabled: orbit-agent reads AWS CloudWatch metrics (eight RDS metrics by default) into orbit-stack, named like prometheus/cloudwatch_exporter, so its Grafana dashboards work unchanged. It reads every cloudwatch.interval (3m by default) at 1-minute resolution; only recently active series are requested. See cloudwatch.md for the cost per interval.
  • The read-only role gains cloudwatch:GetMetricData and cloudwatch:ListMetrics, and deploy/aws/setup-pod-identity.sh also associates <release>-orbit-agent with it.

FinOps: allocation, savings, cost per request, budgets (unreleased)

Added

  • FinOps has four tabs: Overview, Savings, Cost per request and Budgets. See finops.md.
  • The allocation explorer splits the cost by namespace, workload, team, node group, architecture or capacity type over 24h, 7d or 30d, with idle cost, efficiency and a 14-day trend per row.
  • The daily cost of the last 30 days, month to date, and the end-of-month forecast with its band.
  • The savings center ranks every way to spend less: rightsize, downscale, autoscale, idle, Spot, Graviton, consolidation and orphaned volumes. Each item has its evidence and its patch; resource and replica changes apply through the existing write routes.
  • Cost per request: each APM service's cost per thousand requests, with its p95 and error rate.
  • Budgets per cluster, namespace or team, with ok, warn and over statuses. Writes need settings.manage and are audited.
  • Cost anomalies, with severity and the likeliest cause.
  • Eight routes under /api/v1/clusters/{id}/finops/, gated by the FinOps feature switch.

Addendum v0.20: the backend is Rust (unreleased)

Changed

  • Every backend image is built from services/rust/<name>: orbit-api, nocturne, orbit-stack and the orbit-* engines, on gcr.io/distroless/cc-debian12:nonroot. Same names, ports, variables, routes and data directories as before. The Go originals live at the git tag go-oracle, as the differential oracle only.
  • The chart's engine defaults are orbit: telemetry.agent.engine, telemetry.nodeExporter.engine, telemetry.kubeStateMetrics.engine and apm.collector.engine. An upgrade switches those workloads in one rolling update; set prometheus / upstream to keep the upstream images. With orbit-agent there is no kubelet job; nothing queried it.
  • orbit-stack has no garbage collector: GOGC, GOMEMLIMIT and ORBIT_STACK_MEMORY_LIMIT are ignored. See orbit-stack.md → "Memory".
  • The benchmark and differential harnesses moved to services/rust/orbit-stack/bench/ and measure the Rust engine by default; ENGINE=go runs the Go one from its tag.

Addendum v0.5: identity, audit, dashboards and APM (unreleased, 2026-09-24)

Added

  • Users, sessions and roles in orbit-api: local accounts with PBKDF2 passwords, signed and revocable 12-hour sessions carried by the HttpOnly cookie orbit_session, three built-in roles and custom ones over eleven permissions, checked on every route. The first admin comes from ORBIT_ADMIN_USER and ORBIT_ADMIN_PASSWORD_VERIFIER, from three demo accounts on demo data, or from a one-time setup code in the log. New pages: /login, /deck/account, /deck/admin/users, /deck/admin/roles. See identity.md.
  • The audit trail: every sign-in, change and refusal as an event with actor, target, outcome, client address and request id, queryable at /deck/admin/audit and GET /api/v1/audit, exported as CSV, kept ORBIT_AUDIT_RETENTION. See audit.md.
  • Dashboards: five built-in, read-only dashboards, an editor, variables, a JSON model, import from Grafana and versioned saves, stored by orbit-api. See dashboards.md.
  • Traces and APM: orbit-stack takes OTLP/HTTP at POST /v1/traces, keeps spans for ORBIT_STACK_TRACE_RETENTION and derives span metrics and service graph series under Tempo's names; the Deck gained services, service map, trace search and trace pages. The demo shop, orbit-shop, runs in compose and behind demo.shop.enabled. See apm.md.
  • Alert acknowledgements and silences, kept by orbit-api, gated by workloads.write and audited; a runbook section per derived rule. See alerts.md.
  • Contact leads (Addendum v0.7): stored before the form is answered, listed at /deck/admin/leads and GET /api/v1/leads, optionally POSTed to ORBIT_CONTACT_WEBHOOK_URL.
  • New variables: ORBIT_DATA_DIR, ORBIT_SESSION_SECRET, ORBIT_ADMIN_USER, ORBIT_ADMIN_PASSWORD_VERIFIER, ORBIT_AUDIT_RETENTION, ORBIT_CONTACT_WEBHOOK_URL, WEB_TRUST_PROXY, ORBIT_STACK_TRACE_RETENTION, ORBIT_STACK_TRACE_HEAD_BYTES and the shop's ORBIT_SHOP_*. See configuration.md.
  • A twelfth error code, forbidden (403): a valid session whose role lacks the route's permission.

Changed / breaking

  • Every orbit-api route but the probes, the sign-in routes and contact needs a session, on top of the bearer. A script that called GET /api/v1/version or GET /api/v1/clusters with the bearer alone now gets 401: sign in first (identity.md → "Signing in from a shell"). GET /api/v1/auth/state answers with the bearer alone and is the new way to check it.
  • orbit-api runs one replica on every install path, over a single-writer store under ORBIT_DATA_DIR. The chart forces replicas: 1, refuses autoscaling for it, drops its PodDisruptionBudget, and with components.orbit-api.persistence.enabled mounts a claim with a Recreate rollout. values-prod.yaml and values-eks.yaml turn persistence on.
  • ORBIT_TRUSTED_PROXIES no longer names web: web is believed through its bearer, and passes the client address only as the ingress wrote it (WEB_TRUST_PROXY). Set it back to empty; a pod CIDR there would skip the ingress's own address.
  • Ad-hoc PromQL, and so every dashboard panel, runs on the engine on demo data too, when ORBIT_STACK_URL is set.
  • Trace segments are format version 2. An engine refuses a version 1 segment, written by the first trace release, at start by name, skips it and leaves it on disk; the write-ahead log replays either way. One trace is capped at ORBIT_STACK_TRACE_MAX_BYTES (5MiB), and the service graph's waiting spans at ORBIT_STACK_TRACE_GRAPH_PENDING (100000); the service map answers partial and unpairedSpans when spans went unpaired.
  • The first trace comparisons with Grafana Tempo were withdrawn and replaced by round 2 as an independent critic re-ran it; benchmark.md has the figures with the loss table.

Migrating an existing install: give orbit-api a volume (components.orbit-api.persistence), decide how the first admin is created (a Secret pair or the setup code), set ORBIT_TRUSTED_PROXIES back to empty, and replace any bearer-only call to orbit-api with a signed-in one.

Addendum v0.2 (unreleased, in progress)

Landed since the addendum was written

  • The engine's head, WAL and blocks were rebuilt (2026-09-23, not yet measured by a report under bench/runs/): the WAL writes each series' labels once per file and logs samples as columns of deltas and XORs under zstd, and drops what it has synced from the page cache; the head's closed chunks are packed into shared spill segments with one pread per chunk; blocks are compacted in the background into Prometheus-style levels; a block that fails verification at open is served around with warnings, a degraded /readyz and orbit_stack_skipped_blocks; a seal records its logs in meta.json, so a crash no longer seals a window twice; regex matchers are anchored as ^(?s:re)$. The PromQL is Prometheus 3's for float samples, and the API gained metadata, rules, query_exemplars, format_query, status/buildinfo and a per-IP read limiter (ORBIT_STACK_QUERY_RPM, ORBIT_STACK_QUERY_BURST). See docs/orbit-stack.md.

  • The 20k query figures were corrected (2026-09-23): the loop-1 timings sorted 429 refusals in with the answers; timed on answers only, p50 / p90 is 24 / 33 ms, 4.0x / 3.3x faster than Prometheus and 1.5x / 1.7x slower than VictoriaMetrics.

  • Orbit Stack exists (Addendum v0.3): services/orbit-stack, one static binary that takes Prometheus remote-write, seals zstd-compressed columnar blocks and answers a PromQL subset. Fourth container in compose, Helm and the plain manifests.

  • The benchmark ran, against Prometheus 3.1.0 and VictoriaMetrics v1.152.0, and the site's numbers changed to what it measured: 2.9x less disk per sample on sealed blocks than Prometheus (1.2x less than VictoriaMetrics), RSS 2.4x and 2.5x lower at 20k series, ingest 1.19x Prometheus and parity with VictoriaMetrics, query 4.1x faster than Prometheus and 1.5x slower than VictoriaMetrics at 20k; at 200k series every row beats Prometheus and RSS is 64% above VictoriaMetrics. "10x", "6.8x RAM at 10M series" and "40M active series per node" are retired. docs/benchmark.md is now a results page.

  • Sealing no longer materialises the block in RAM: columns stream through zstd series by series and the process hands the seal's peak back to the OS. A 200k-series run showed the old seal doubling RSS; docs/orbit-stack.md now documents GOMEMLIMIT for containers.

  • ORBIT_DEMO_DATA=false starts: internal/source/live reads a Kubernetes cluster (kubeconfig or in-cluster with the chart's read-only ClusterRole), derives alerts and a list-price FinOps estimate, and answers query_range from the engine. The demo source serves CLAIMS.md's estate with the frontend's shapes.

  • The Deck goes live with NEXT_PUBLIC_DEMO_MODE=false, which make k3d-up builds: the console shows the k3d cluster, fed by a Prometheus agent into the engine.

  • Ad-hoc PromQL in the Deck: the Observability page takes an expression, runs it through query_range, and saves it as a panel in the browser.

  • The web front door exists: apps/web/src/app/api/[...path]/route.ts forwards /api/* to ORBIT_API_URL with ORBIT_AUTH_TOKEN attached server-side, drops any client-supplied Authorization, streams SSE unbuffered and cancels upstream when the browser leaves. pnpm check:proxy guards it.

  • The Deck's Nocturne AI page streams from the service through that door, frame by frame, and its routing panel reads GET /api/v1/nocturne/models, which orbit-api now proxies alongside chat and analyze.

  • POST /api/v1/contact accepts the site form's own fields (clusters, series, locale, source); the message is optional. The form posts to its own origin.

  • ORBIT_TRUSTED_PROXIES / NOCTURNE_TRUSTED_PROXIES: the per-IP limiter keys on the forwarded client only behind a peer the operator named. Without it every visitor behind web shared one bucket.

  • CI (.github/workflows/ci.yml): the full gate, image build and scan with no suppression, and a generated digests.yaml that values-prod.yaml is rendered against and that ships with the release.

  • A rollout no longer cuts a proxied Nocturne AI answer without a terminal frame: orbit-api drains for 180s (was 60s; terminationGracePeriodSeconds 200 everywhere), and any proxied stream still open 5s before the budget ends is closed with event: done + finishReason: "shutdown". The Deck prints that as "Cut short by a service rollout. Ask again."

  • make test no longer performs discovery against a non-local kubectl context. Manifest validation against a cluster runs only on k3d-*, kind-*, minikube, docker-desktop or rancher-desktop (or with ORBIT_ALLOW_REMOTE=1).

Round 1 shipped both Go services with no authentication of any kind: an unauthenticated caller could read the whole estate and choose which model tier to bill. This addendum closes that and two related gaps. It is mandatory: a build that starts without it is a defect.

Added

  • Mandatory auth middleware ahead of the router on both orbit-api and nocturne: static bearer (constant-time compare, refuses to start under 32 bytes), oidc (JWT verified against issuer JWKS), or disabled (local-demo only, refused whenever ORBIT_DEMO_DATA=false). See security.md.
  • web now proxies /api/* to orbit-api from the Next.js server, attaching ORBIT_AUTH_TOKEN itself; the browser never holds it.
  • /metrics moved off the API listener onto its own bound address (ORBIT_METRICS_ADDR / NOCTURNE_METRICS_ADDR, default :9090) on both Go services. No ingress routes it, and the auth exemption list never covers it, because it isn't on the listener that list applies to.
  • New required metrics shapes: *_provider_requests_total{tier,outcome} counter, *_stream_duration_seconds histogram (buckets to 300s), an in-flight gauge. These metrics now count an event: error inside a 200 SSE response as a failure.
  • nocturne now carries the same per-IP rate limiter orbit-api already had, on every route, not only on contact.
  • /readyz on both services must now be able to return 503: orbit-api checks its source and nocturne; nocturne checks provider reachability (cached).
  • Explicit shutdown-grace requirements: both services take <PREFIX>_SHUTDOWN_GRACE, defaulting above the request timeout (60s for orbit-api, 180s for nocturne, which streams). A drain overrun logs shutdown incomplete and exits 0; it is never reported as a startup failure.

Changed

  • nocturne is now reached exclusively through orbit-api's internal reverse proxy (POST /api/v1/nocturne/chat and /analyze on orbit-api); it has no ingress rule and no CORS surface of its own.

Removed / breaking

  • NEXT_PUBLIC_API_URL is gone. It was build-time inlined into the browser bundle and pointed the browser at a service it must no longer reach directly. If you were setting it, stop: web now reads ORBIT_API_URL (server-side only, runtime) instead. See configuration.md.
  • tier is no longer a field in the chat or analyze request body. An unauthenticated or careless client choosing its own model tier is the billing-control gap this addendum closes. tier.Select now classifies every request server-side; the tier that ran is still reported back in the done SSE frame and in a response header, but it can no longer be requested. See api.md.

Migrating an existing install: generate ORBIT_AUTH_TOKEN and NOCTURNE_AUTH_TOKEN (configuration.md → "Generating a static token") before upgrading. A static-mode service with no token refuses to start; that is the correct failure mode for an install nobody finished configuring, so don't route around it. If you were calling orbit-api or nocturne directly from a browser against their own origins, stop; call web's /api/* path instead.

0.1.0: initial contract

The first agreed contract: three components (web, orbit-api, nocturne), the shared /healthz / /readyz / /metrics / /api/v1/version surface on both Go services, the orbit-api REST surface (clusters, nodes, workloads, events, alerts, FinOps, recommendations, query_range, contact), the nocturne chat/analyze/models surface with two-tier model routing, container and Kubernetes hardening rules (distroless, non-root, dropped capabilities, read-only root filesystem, requests and limits, liveness/readiness on the contract's own paths, PDBs and topology spread), and the request/body/rate limits enumerated in api.md. Shipped with no authentication; Addendum v0.2 above explains why that's no longer true.

Where to go next

  • What's implemented on disk right now, component by component, as opposed to what the contract specifies: status.md.
  • The full, current environment-variable surface, addendum included: configuration.md.