Changelog

All notable changes to this project will be documented in this file.

The format is based on Keep a Changelog.

[Unreleased]

[0.31.0] - 2026-10-17

  • The load test fills the books with customers enough for the top rung (500 by default), not 10,000, and a climb whose first rung completed no journey fails the test with why instead of recording capacity 0 (#60); The Workflows pages are gobank-workflow v0.7.0's own: instances as a filterable work queue, those in hand first; definitions on their own pages; an instance's revision and its stages in their latest state, with the drill, scan, load and perf sections under it. Instances move to /workflows/instances/; the old links redirect (#62)

Changed

  • The load test fills the books with customers enough for the top rung (500 by default), not 10,000, and a climb whose first rung completed no journey fails the test with why instead of recording capacity 0 (#60) (#60)
  • The Workflows pages are gobank-workflow v0.7.0's own: instances as a filterable work queue, those in hand first; definitions on their own pages; an instance's revision and its stages in their latest state, with the drill, scan, load and perf sections under it. Instances move to /workflows/instances/; the old links redirect (#62) (#62)

Roadmap

  • #23 prefixWriter.Write returns 0 after a full write, breaking the io.Writer contract

[0.30.0] - 2026-10-17

  • Perf runs make a payment a customer a day (up to 10,000), not gobank's default 10 a day for the whole bank; prefixWriter.Write reports the full count on success, so io.Copy and other count-checking callers no longer fail with a short write; Perf requests the staff pages in rotation during the days stage — dashboard, P&L, balance sheet, BBSI, explorer, customers, payments and treasury — so their counts and p95s at scale land in the record; Spans on /perf show only the precision they need (1h46m51s, not 1h46m51.360895289s), via go-span

Added

  • Perf requests the staff pages in rotation during the days stage — dashboard, P&L, balance sheet, BBSI, explorer, customers, payments and treasury — so their counts and p95s at scale land in the record (#54)

Changed

  • Perf runs make a payment a customer a day (up to 10,000), not gobank's default 10 a day for the whole bank (#53)
  • Spans on /perf show only the precision they need (1h46m51s, not 1h46m51.360895289s), via go-span (#56)

Fixed

  • prefixWriter.Write reports the full count on success, so io.Copy and other count-checking callers no longer fail with a short write (#23)

[0.29.0] - 2026-10-17

  • Dependencies: go-postgres v0.10.0, lofigui v0.18.0, aws-sdk-go-v2 and golang.org/x refreshed

Changed

  • Dependencies: go-postgres v0.10.0, lofigui v0.18.0, aws-sdk-go-v2 and golang.org/x refreshed (#50)

[0.28.0] - 2026-10-17

  • Short perf run from the dashboard: the Performance run form takes customers, days and a ceiling, blank for the full run

Added

  • Short perf run from the dashboard: the Performance run form takes customers, days and a ceiling, blank for the full run (#46)

[0.27.0] - 2026-10-17

  • Forget a finished environment: a row with no server, no running job and not in -envs can be dropped from the dashboard (#47)

[0.26.0] - 2026-10-16

  • New servers get a public IPv4 only, no IPv6: each box now costs one Hetzner primary IP, not two (#44)

[0.25.0] - 2026-10-16

  • Perf run records the demo database's table sizes in its last reading; the run's page shows them biggest first with rows, size and share of the database (#42; needs gobank with about.json tables, gobank #164)

[0.24.0] - 2026-10-16

  • Short perf runs (-customers, -days, -ceiling); days stage retries Run then aborts; no network create race (#38, #39, #40)

Added

  • A short perf run (#38): gobank-deploy perf takes -customers (default 10,000), -days (the days stage runs that many days instead of seven years) and -ceiling (default 6h; the servers' expiry follows it), so the same shape — every scale at once, create, add, days, routes, remove — gives figures in tens of minutes.

Fixed

  • The perf days stage no longer polls a stopped demo to the ceiling (#39): a demo found stopped short of the target is pressed Run again, at most twice in a row, then the stage fails and the box goes. Seen on a v0.55.0 run left at day 0 with the box billing.
  • Ups at once no longer race to create the private network (#40): the provider refusing a second create as a name already used is the network existing. perf -scales small,large lost its large run to it.

Roadmap

  • #26 Customer journey probe: the staff pages walked in sequence as a member of staff would, timed per page and end to end, defined as data so the gates can run it

[0.23.0] - 2026-10-16

  • Load test: the customer journey from a loader box over a private network, capacity at p95 2s a screen; firewalls admit home only (#26)

Added

  • The load test (#26): the customer journey — login, the accounts screen, one account's screen — defined as data and walked from a loader box beside the target on a private network, up a ladder of concurrent users (1 … 500), 10 journeys each, the customers in sequence through the books so nothing comes from a cache, until a screen's p95 is over 2 s or a journey fails. The load workflow (gobank-deploy load <env>, and a form on the page) creates the target, fills its books, makes <env>-loader in the target's location, climbs and removes both; the result is kept (load schema v1) and shown on Load tests (/perf/load): capacity against gobank's version, a line per scale. gobank-deploy journey -base URL runs the climb alone from anywhere.
  • The private network gobank (10.128.0.0/16, subnet 10.128.1.0/24, eu-central): every server is created on it, and an existing one is attached on its next up.

Changed

  • Firewalls are declared and set on every up, not created once: ssh and the demo from -admit (GOBANK_DEPLOY_ADMIT) only, ping from anywhere. Empty -admit keeps the old open rules. Networking and the firewall rules are documented as a d2 view and a decision table.

[0.22.0] - 2026-10-16

  • Perf runs keep per-route counts and p95; routes over versions chart by scale (#35)

Added

  • A performance run keeps gobank's per-route figures (#35): a new routes stage, the last before remove, reads every route the demo served over the run off about.json (gobank v0.51.0 onwards), its request count and p95, and stores them with the run (perf schema v2, a routes column). The run's page lists them; Routes under Performance (/perf/routes) charts a chosen route's p95 against the gobank version, a line per scale, with the same figures as a table.

[0.21.0] - 2026-10-16

  • Scans readable: a page per scan, the latest against the previous, ten worst, critical and the trend (#33)

Added

  • The scans readable (#33): Scans under Performance (/perf/scans) lists every scan and reads the latest against the one before — the critical routes (failed, newly not scanned, or p95 at the top rung over a 500 ms budget), the ten worst by p95 with the move since the previous scan, and every route across the scans, a column per scan — and each scan has a page (/perf/scans/<run>) with every route at each rung, worst route first, the routes not scanned with why. The run page and the journal link to it; before, a scan's figures were only the journal page's markdown in a <pre>.

Changed

  • gobank-workflow v0.5.0: runs record the step list they ran (schema v4, expand only, applied on start). Nothing here uses the new Purpose or Revision yet (#31).

[0.20.0] - 2026-10-16

  • Stories 18, 21, 28 and the roadmap epic for 26/27: the scan workflow and concurrency ladder, the journal figures adding up, and the Hetzner list as last listed

Fixed

  • Hetzner's rate limit (#28): the page asked for the server list and then every environment by name on each poll, five requests every three seconds while a job ran, past the token's 3,600 an hour. The cloud is now the project as last listed: one list answers the page and the flows for 30 seconds, a server by name is read off it, and a create, delete or relabel by this program forgets the list so it shows at once.
  • The journal's figures add up (#18): the add step records what it added, the books' total, and customers/s over its own wall clock, where it showed the books' total under "customers" beside the demo's in-batch rate, so 3,000 customers could sit beside a rate and a duration that made 2,400. The years step names its two clocks: bank days from and to, bank days/h and account days/h over the step's wall clock, in place of the demo's "account days / 12h". The steps table's column is "Wall clock" and the README has the vocabulary.

Added

  • The concurrency ladder (#21): the scenario's probe loads each route at 1, 4 and 16 in flight (-ladder) and records the throughput it achieved beside p50 and p95; the journal's probes table gains the method and the per-second column.
  • The scan workflow (#21): gobank-deploy scan <env> is a full performance scan of an environment as it stands — every route the demo's API document lists that the scan can drive, as the operator or as a customer, at each rung of the ladder, the rest listed as not scanned with why — kept as a run of the journal and written as the measured document api/measured.json into gobank's checkout for its API pages to show alongside each route.

[0.19.1] - 2026-10-16

  • A failed perf or scenario run removes its box instead of leaving it billing (#29)

Fixed

  • A failed perf or scenario run no longer leaves its box billing (#29): create is undone by remove when a later stage fails (gobank-workflow's Stage.Undo), so the box is deleted, the run's error says so, and the next attempt the same day starts from create instead of resuming against a server that is gone. With -keep a scenario's box is left, failed or not. The perf box also carries an expiry of twice the ceiling and an hour, so a run that hangs or an engine that dies is reaped by the page's tick like a temporary environment.

[0.19.0] - 2026-10-14

  • Performance run to targets: 10,000 customers, then seven years or six hours (#24)

Changed

  • The performance run (#24) runs to targets instead of ten minutes a stage: add until the books hold 10,000 customers, days until the bank's day is seven years on, each stage bounded by a six-hour ceiling, so a run is a soak of up to six hours a scale, all scales at once. The record's add and days spans are now the time each stage took; the log says when a stage stopped at the ceiling short of its target.

Fixed

  • perf run at every scale at once found no records when the runs crossed midnight: the keys are fixed when the runs start.

[0.18.0] - 2026-10-14

Changed

  • The scenario probe posts the customer search's name in the form, as gobank's search (gobank #154) takes it, never in a URL; the default name is Jones, a surname the demo's customers have.

[0.17.0] - 2026-10-14

  • An API token per environment: the console client drives the demo as the operator (gobank story 9.4); a login redirect is a refusal, which is why the recent drills failed

Added

  • The scenario workflow and the journal (#18, gobank milestone 1 #151): gobank-deploy scenario seven-years creates an environment at a scale, adds the customers flat out, runs seven years flat out, sets the day length to wall clock, probes the dashboard, the customer search by name and a customer's page from outside at concurrency 1 and 4, and removes the server; every step is recorded with its figures and the run is written as a page of journal/, the notebook of runs with the operator's notes, listed in journal/index.md and on the docs site. The ui's Journal page under About shows the runs this host made with the page to copy.
  • An API token per environment (gobank story 9.4): up generates one as it does the passwords, keeps it on the server's labels, writes it to the deploy env as GOBANK_API_TOKEN, and status and the ui show it. The console client sends it as a bearer token on every post (the drill's day length, the performance run's settings, batch add, Run and Stop), so the demo admits this program as the operator now that its console sits behind the staff login (gobank 1.7.1).

Fixed

  • A console post the demo answered by sending us to its login page was read as success: the drill's day length was never set, the run went on flat out between the observation and the stop, and the upgrade hop failed on "the run went on before the stop" (the v0.35.0 and v0.36.0 drills on preprod). A redirect to the login is now ErrLoginWanted, so a drill on an environment without a token fails at prepare and says why.

Roadmap

  • Story 1o — API token per environment; the console client drives the demo as the operator

[0.16.0] - 2026-10-09

[0.15.0] - 2026-10-09

[0.14.0] - 2026-10-09

  • Admin password per environment for the staff login (gobank 1.7.1)

Added

  • An admin password per environment (gobank story 1.7.1, the staff login): up generates one for a new server, and for a server without one on its next redeploy, keeps it on the server's labels beside the app password, writes it as GOBANK_ADMIN_PASSWORD in the deploy env, and status and the ui show it, so a tester can log the staff web in as admin.

Roadmap

  • Story 1n — Admin password per environment for the staff login

[0.13.0] - 2026-10-08

  • Release prep

[0.12.0] - 2026-10-06

  • A Workflows tab: each workflow from its definition, its instance counts and instances, steps laid over the definition

Added

  • A Workflows tab: the engine's view of this program. Each workflow it runs (demo, drill, perf) is shown from its definition, with its diagram, its steps and the code behind each, the count of its instances in every state and the instances themselves, newest first. An instance page lays the recorded steps over the definition, so the steps not reached are visible as such, and offers the instance as d2 with the steps coloured by state (/workflows/<id>/diagram.d2) and, where the explorer is configured, its step rows in the DB Explorer.

Changed

  • Every page shares one navbar, so the About page now has the Performance and DB Explorer tabs it lacked.

Roadmap

  • Story 1l — A Workflows tab: the engine's view of its workflows, instances and steps

[0.11.0] - 2026-10-06

  • perf writes its rows into gobank benchmark.md; large is ccx33 and xl ccx43; status by address when the hostname does not point at the server; create tries other locations then down the ladder

  • Status gives the address when the hostname does not point at the server; create tries the next location when one has no capacity

Added

  • perf writes each run's row into gobank's benchmark.md (the -src checkout; -doc to say where, "" to only print), filling the empty placeholder row for the scale or appending, so the performance doc is updated by the run and published by the next gobank release.

Changed

  • large is now ccx33 (8 dedicated vCPU, 32 GB) and xl ccx43 (16 dedicated vCPU, 64 GB): a measurement on dedicated cores says whether the demo uses them, which shared vCPUs cannot. The capacity ladder is ccx43, ccx33, cx53, cx43, cx33, cx23, a first cut to revise with the performance data.

Fixed

  • The first perf run failed twice over. Status reported the hostname URL although the DNS change had been refused (the AWS key tp secrets supplies, the LAN DNS user, may not change the gobank zone), so the workflow could not reach a demo that was serving by address; status now gives the hostname only when it resolves to the server's address, else the address, which also keeps the page's link and the drill honest when a name is stale. And Hetzner had no capacity for a cx53 in fsn1 ("error during placement, resource_unavailable"); the deployer now tries its locations in order, fsn1, nbg1, hel1, then the next size down the ladder at every location again, and fails only when the smallest is full everywhere. A smaller box beats none, and the server's actual type is what a perf run records.

[0.10.0] - 2026-10-06

  • Story 1k: the performance run as a workflow (perf), a Performance page and perf CLI; console reads gobank v0.12 rates

  • The performance run as a workflow: a fresh environment at a scale, measured for twenty minutes, removed; a Performance page with the row for gobank's benchmark.md

Added

  • The perf workflow (story 1k): create a fresh environment at a scale (an existing server is kept, so a failed run resumes on the same box), add customers flat out in batches of 1,000 for ten minutes with the rate taken from the demo batch by batch, days for ten minutes reading account days per 12h and the last day it is made of, remove. Started by Measure on the environments page (a fresh name and a scale) or perf <env> on the command line, which measures every scale at once (-scales small,large by default) on a server each, named <env>-<scale>, and prints the rows at the end; kept in the -db database with where it ran (scale, server type, RAM, gobank version), its schema versioned in perf_schema_migrations. The Performance page is the history, each run with its row for gobank's benchmark.md; the run page shows a perf run's figures.
  • The console reads gobank v0.12's rates off /about.json (customers_per_sec, account_days_per_12h, adding_customers, last_day_duration, last_day_accounts) and drives the demo's forms: the settings together (SetSettings), AddCustomers, Start, Stop.

Roadmap

  • Story 1k — The performance run as a workflow, with a Performance page

[0.9.0] - 2026-10-05

  • Drill accepts the day a stop finishes and the day a restart begins; sets the day length again after each hop

  • Drill accepts the day a stop finishes and the day a restart begins; sets the day length again after each hop

Fixed

  • The drill failed its first hop on prod with "landed on a day boundary" although the upgrade was sound: the demo's day length set by prepare lasted only as long as the process, so v0.10.2 came back flat out and ran two days before the reading. The gate pinned a property the demo does not have: a clean stop finishes the day in progress and a restart begins a new day, so the position moves on every hop. The hop now gates on the restart row's day count being at most one past the last observation (the stop) and the position at most one past that (the start); further is the run going on at full speed. And the drill sets the day length again after every hop, since a release before gobank v0.10.3 forgets it on restart (gobank v0.10.3 records it with the run).

[0.8.0] - 2026-10-05

  • Serving check waits while the service is alive, not for a fixed five minutes

Changed

  • The environment's address on the main page opens the deployed site in a new tab, so the dashboard stays put

Fixed

  • The check after a start waits while the service is alive (active and not restarted by systemd), up to half an hour, instead of a fixed five minutes: the demo's resume grows with its history (1m32s at 4,010 days, 2m21s at 5,811), so a clock would keep failing drills whose upgrade had landed. The job log says so each minute; a service that dies or restarts fails the check at once

[0.7.0] - 2026-10-04

  • Serving check outlasts the demo's resume; drills page shows each run's state and steps

Fixed

  • The check after a start waits five minutes, not one, for the demo to answer: it rebuilds the bank from its database before it listens (a minute and a half at four thousand days, growing with the history), so the first drill on prod failed its upgrade hop with "service not answering" while the upgrade landed anyway

Changed

  • The Drills page shows each drill's run: its state and error, and the steps with how long each took, so a failed drill reads in full there rather than only on the environments page; its link to the run page reads "details" ("run" looked like a button), and each drill is headed with the time it started, not just the day

[0.6.0] - 2026-10-04

  • Story 1j: the upgrade drill as a workflow, with a pglike database and the DB explorer

Added

  • The upgrade drill (story 1j): Drill to vX on an environment's row runs gobank's upgrade-drill.md as a gobank-workflow pipeline keyed <env> N→N+1 <date>: prepare (a day of at least 30m, set to 2h otherwise, then wait for a day with 10m left), before, upgrade, rollback, forward. Each hop fetches the release, redeploys, reads the demo's /about.json and is gated on the version serving, a restart row following the expected release with a known downtime and an intact handover, and the same simulated day as before; a release without about.json has its version checked only. A failed hop resumes on the next press the same day. Offered only with a release store to roll back from and a newer release than the one serving
  • A database: ui -db FILE (default <-build>/gobank-deploy.db, so /perm/gobank-deploy/gobank-deploy.db on hydrogen) is a pglike file holding gobank-workflow's run records and this program's drills and drill_observations, each component's schema versioned in its own migrations table through gobank-db ApplyTo. Runs and drills survive a restart; closes issue #1
  • Pages: Drills, the history with each drill's observations and the line for the story's record; /workflows/<id>, a run's steps and, for a drill, its observations (the runs table links there); DB Explorer at /internal/explorer (go-dbexplorer) over the live database. The About page draws the drill workflow
  • internal/remote.Console: the demo's /about.json and settings form as the drill uses them

Changed

  • cmd/flowd2 writes one <type>-workflow.d2 per definition instead of printing one to stdout

Roadmap

  • Story 1j — The upgrade drill as a workflow, with a database and a Drills page

[0.5.0] - 2026-10-04

  • Story 1i: rollback from the store (fetch a named tag) and a stop timeout that outlasts a day

Added

  • POST /fetch?tag=vX (story 1i): makes a named release the next deploy, fetching it into the store if it is not there, so a rollback from hydrogen is a fetch of the previous tag and a Redeploy. Store.Use marks a held version the latest.
  • The demo's systemd unit has TimeoutStopSec=900, from cloud-init on a fresh server and from the install script on an existing one: the demo finishes the simulated day in progress on SIGTERM, and the default 90s would have killed it mid-day.

Changed

  • ui on htmx (story 1h). The page no longer reloads itself: the environment table, job logs and clock are a fragment at /fragment, polled at the same 3s / 15s rate and paused while a form has focus, so a Create click or a half-chosen scale is no longer thrown away by a refresh. Buttons post over htmx and swap the fragment back; a refused action shows its message above the table instead of a bare error page. Without JavaScript the page works as before. htmx 2.0.4 is embedded in the binary

Roadmap

  • Story 1h — ui on htmx (lofigui tier 3): the page stops reloading itself
  • Story 1i — Rollback from the store (POST /fetch?tag=vX) and a stop that waits for the day

[0.4.0] - 2026-10-02

  • Gen update

Added

  • The ui's navbar shows a small analogue clock set at render time, so the second hand jumps round on each poll (every 3s while a job runs, 15s idle) and the page is visibly live.

[0.3.0] - 2026-10-02

  • App password per environment: up labels it, status and ui show it

Added

  • App password (story 1g). up gives each environment a 16-character password, kept on the server's app-password label so redeploys reuse it and a server from before this release gets one on its first redeploy. The install writes it as GOBANK_APP_PASSWORD in /etc/gobank/deploy.env, so the demo's customer BFF (gobank ADR-0002 stage 1, /v1/ on the demo's port) accepts logins; status and the ui show it. Cloud gains SetLabels.

Roadmap

  • Story 1g — App password per environment

[0.2.0] - 2026-10-01

  • POST /fetch: gobank's release puts its binaries into the store straight away

Added

  • POST /fetch on the page: fetches gobank's newest release into the store now, for gobank's tp release to call as its post_release step, so the page shows the new version as deployable without a deploy first. Answers with the tag, or 502 with why it could not fetch.

Fixed

  • Two puts of the same release into the store at once (a fetch after a release and one before a deploy) no longer share a temporary file.

[0.1.6] - 2026-10-01

  • Fetch gobank releases from the Forgejo into the store: release gobank, press Redeploy

  • The appliance fetches gobank's newest release itself: release gobank, press Redeploy

Added

  • Fetch as the stage before a deploy from a store: when gobank's newest tag is not in the store, its release's demo-linux-amd64 and demo-linux-arm64 (attached by gobank's tp release, now a goreleaser build) are fetched into the store and become the latest. A new gobank version reaches an environment without releasing gobank-deploy or updating hydrogen. A tag without binaries or a Forgejo out of reach is a notice; the deploy carries what the store has. A release the store already has is not fetched again, so task push of an older one is still a rollback.

Changed

  • With a store and a repo, an empty store no longer hides Create and Redeploy: the first deploy fills it. The page and status say the next deploy fetches a lagging release, rather than asking for a push.

Roadmap

  • Story 1f — The gobank release is the build stage: goreleaser attaches the demo binaries, up from a store fetches them

[0.1.5] - 2026-09-29

[0.1.4] - 2026-09-29

  • More version info

Changed

  • A refused Route 53 change no longer fails up or down: it is reported and the deploy carries on by address.

Added

  • The page and status report gobank's newest tag on the Forgejo (public tags API, no token) against what can be deployed from this host, and say when the store or checkout is behind the repo. -repo URL.

[0.1.3] - 2026-09-28

[0.1.2] - 2026-09-28

  • Docs site, About page with component diagrams, expiry label fix

Added

  • Docs site: index.md with the component structure, the up sequence and the demo workflow as d2 diagrams, plus decision tables; task docs:build renders it, tp pages previews it, deployed to gobank-deploy.docs.bytestone.uk.
  • About page on the ui with the same three diagrams, embedded in the binary so the appliance shows them.

[0.1.1] - 2026-09-27

  • Route 53 hostnames, version comparison, demo workflow with automatic removal

Added

  • Demo workflow, the first on gobank-workflow's pipeline runner: "remove after" on the Create forms makes a temporary environment that is created with its expiry as a server label, served until then and removed. The page reconciles every minute, resuming the workflow for any temporary server without a job, so a restart does not leave one billing. Recent workflow runs are listed on the page; cmd/flowd2 prints the d2.
  • Versions: status and the page report the version each environment is running (from the demo's page footer) against what the next up would deploy (the store's latest, or the checkout's git describe), and flag an environment that is behind.
  • Route 53: each environment is <env>.gobank.drummonds.net, an A record set on up (redeploy repairs it) and removed on down; status and the page link by hostname. -dns DOMAIN / -dns ""; skipped with a notice when there are no AWS credentials.

Roadmap

  • Story 1d — Versions on the page and in status
  • Story 1e — First workflow on gobank-workflow: the temporary demo environment

[0.1.0] - 2026-09-26

  • ui lists the project's gobank servers and adds environments by name; demo back in the default

Added

  • ui lists every gobank-* server in the Hetzner project alongside the -envs names, and has a New environment form (name + scale) so the list is no longer fixed at startup. demo is back in the default -envs (prod,preprod,demo).
  • build subcommand: cmd/demo for linux amd64 and arm64 into a release store (build/releases/<version>/demo-linux-<goarch> plus latest), and task push to copy the latest into hydrogen's /perm/gobank-deploy/releases.
  • -store DIR: deploy from a release store instead of building, for a host with no toolchain. Create / Redeploy become available on the page as soon as the store has a release (checked per request, no restart).
  • GOBANK_DEPLOY_SSH_KEY: base64 of a PEM private key to ssh with, for a host with no agent and no ~/.ssh (the appliance).
  • ui on a host that cannot build cmd/demo (no Go toolchain, or no gobank checkout at -src) shows status and offers Down only, says why, and refuses create / redeploy requests with 403. This is the mode for the hydrogen gokrazy appliance.
  • up writes /etc/gobank/deploy.env with GOBANK_MEMORY_LIMIT sized to half the box's RAM (the rest is PostgreSQL's), and makes the unit read it; works on boxes created before the unit had the EnvironmentFile line.
  • ui: lofigui page listing each environment's state with Create (scale preset, marked as starting billing), Redeploy, Down (confirmation tick required) and Cancel; one job per environment, log on the page, polling while a job runs.
  • up, down, status for a named environment on Hetzner Cloud, ported from gobank's deploy/hetzner scripts: explicit -create gate, scale presets, cloud-init Postgres setup, binary install over ssh, HTTP check.
  • Orchestration in internal/deploy behind Cloud, Dialer, Builder and Prober interfaces, tested with fakes.
  • Project-local host-key pinning with accept-new semantics and forget-on- recreate, tested. Negotiates the key types already pinned for a host (as OpenSSH does) and understands hashed entries, so a manual ssh against the same file does not turn into a false "host key changed".
  • ssh credentials: the agent named by IdentityAgent in ~/.ssh/config, the default agent and the key files are merged into one publickey method (x/crypto/ssh tries each method name once).
  • TestDialReal, gated by GOBANK_DEPLOY_SSH_IP, exercises the dialer against a live server.

Roadmap

  • Story 1 — up, down, status for a named environment
  • Story 1a — ui: environment states and controls on a polling page
  • Story 1b — ui as a status-and-down console on hydrogen (gokrazy)
  • Story 1c — up from a release store; task push fills hydrogen's, an older push is a rollback
  • Story 6 — Run on gokrazy: pure Go, a key on the appliance instead of an ssh agent