CI/CT with GitHub Actions
dl cict manages CI/CT channel A: GitHub Actions jobs that ask for a
DreamLake label (runs-on: dreamlake-<pool>) run on DreamLake machines
through single-use runners, and their declared evidence is written to a
Source. How the flow works, the workflow contract and the evidence files are
described in the product guide,
CI/CT with GitHub Actions.
These commands need a server with the GitHub runner integration enabled;
otherwise every command returns HTTP 404 github_runners_disabled. The same
operations are available in TypeScript as dl.cict(namespace) and in Python
as client.cict.
Where a namespace stands
status lists every setup step (GitHub, DreamLake GPU, and the six steps for
your own Slurm cluster), marks which are done and prints the next one with the
page, command and API call that does it. Once ready it prints the runs-on
line to use. It is the starting point for an agent setting up CI/CT: repeat it
and do the next step. Only installing the GitHub App needs a person in a
browser; enrolling a login node needs SSH, and binding its service account
needs a namespace owner. TypeScript: dl.cict(ns).onboarding().
Connect GitHub
connect prints a GitHub App install link. Open it within 15 minutes in a
browser signed in to DreamLake as the same user, install the App on an
organization or account and select repositories. GitHub returns to the
DreamLake setup page, which binds the installation to the namespace after
checking that your GitHub user can access it. Starting the binding requires
the namespace's execution-admin permission.
An installation is bound to one namespace. Uninstalling or suspending the App on GitHub deletes or suspends the binding.
Runner pools
A pool maps runs-on labels to a Slurm provider and, optionally, to a Source
for evidence. Jobs whose labels match no pool are recorded as unmatched and
get no runner.
--nameis lowercase letters, digits and hyphens. Without--label, the pool serves the single labeldreamlake-<name>. Each--labelmust bedreamlakeor start withdreamlake-;self-hosted,linuxandx64are implicit. When several pools match a job, the most specific wins.--provideris a Slurm provider in the namespace on which you have your own association. Runners start as you (the pool creator) until service-account grants replace this.--evidence-sourcereceives each run's evidence under<prefix or ci>/<owner>/<repo>/<YYYY-MM-DD>/<HHMMSS>_<job>_<sha7>_<jobId>.--env NAME=valueadds a machine fact every job sees, such as a cache path on the cluster. Values are visible to anything running on the pool; never pass secrets. Names starting withDREAMLAKE_GITHUB,DREAMLAKE_RUN,DREAMLAKE_RESULTS,DREAMLAKE_METRICS,DREAMLAKE_JOB,DREAMLAKE_CACHE,GITHUB_,RUNNER_orSLURM_are refused. At most 32.
Deleting is refused with runner_pool_busy while the pool still serves
running jobs or collects their evidence; delete it after they finish.
update changes a pool for the next job; running jobs keep the provider,
evidence location and environment they started with. --env replaces every
variable. The name cannot change.
targets lists the namespace's Slurm providers and whether a pool can start
runners on each: usable, or the reason it cannot (no_association: no host
is bound to it; no_launch_layer: no GPU, memory and time settings;
legacy_provider).
Runs
A run is one GitHub Actions job that asked for a DreamLake label, joined with
the runner that served it. Its ID is <installationId>-<jobId>.
Each line shows the run ID, status or conclusion, repository, job, commit,
routing (pool name, unmatched, unbound or rejected:<reason>) and the
evidence state (pending, written, failed or none).
show adds the GitHub job link, the runner and its Slurm job, a runner error
code when start failed, and the evidence location. --result also reads the
run's result.json from the evidence Source; it is null when the run wrote
none or evidence is not written yet.
compare lists the cases that changed against the baseline: the latest
earlier run of the same repository, workflow and job on the default branch
whose evidence was written. Each case is regressed, fixed,
still_failing, unchanged, added or removed, with the change in its
score (such as normalized return) when both runs report one.
DreamLake GPU
DreamLake GPU runs jobs on machines DreamLake provides, chosen by runs-on
label; no pool is needed. The DreamLake team enables it for a namespace.
show prints the machine catalog with each machine type's price per GPU-hour,
whether the namespace may use it (not enabled, enabled but not started, in
use, or paused), its concurrency limit, running and queued jobs, where results
go and this month's GPU time. With prepaid billing on it adds what this month
was charged; otherwise what the GPU time is worth (not charged).
dl cict runs show adds a cost line for a DreamLake GPU job: the credit it
holds while it runs (its time limit at its price), then what it was charged
(the GPU time Slurm measured), or why it was not charged. Billing runs on
sandboxes for now: no real money moves.
start creates the cict-evidence Source for results, or uses --source,
and records you as the person who authorizes those writes. Creating a Source
needs the namespace admin role.
usage reports a month of DreamLake GPU time, from each job's start to its
end, by the GitHub user who triggered each job and by repository and
workflow. --jobs lists the jobs newest first instead; --sender LOGIN
keeps one person's and --limit caps how many (1–500). The month line and
each job show what was charged (or charge pending until Slurm reports), or
what the GPU time is worth when DreamLake GPU is not billed.
Administer DreamLake GPU
These commands are for admins of the platform namespace.
allow lets a namespace use DreamLake GPU, or resumes a paused one. A
namespace whose admin already started keeps its results Source.
providers lists the Slurm providers in the platform namespace and what one
job gets on each, read from its launch layer. A provider without a bound host,
a launch layer or a GPU in it cannot back a machine type.
machine creates or changes a machine type. A machine type is what one job
gets, not one machine: jobs land on any free node of the provider's partition.
The GPUs, CPUs, memory and time limit come from the provider's launch layer,
so they always match what Slurm allocates; --limit caps jobs at once across
all namespaces. Runners start with the provider's association. --price
sets the price in credits per GPU-hour (one credit is $0.005, so 240 is
$1.20); --price 0 makes it free and --price none clears it. A job keeps the price it started at, so a
change applies to later jobs only. A machine type without a price cannot run
jobs while billing is on.
Troubleshooting
| Symptom | Meaning |
|---|---|
routing unbound | The repository's App installation is not bound to a namespace; run dl cict github connect. |
routing unmatched | No pool or DreamLake GPU label covers every label in runs-on; check dl cict pools list and dl cict platform show. |
queued #N | A DreamLake GPU job waiting for a slot in the namespace or on the machine type. |
held: insufficient_credit | The namespace's credit cannot cover what the job holds while it runs; add credit in Settings → Billing and it starts within a minute. |
held: billing_budget_exceeded | The namespace's spending limit is reached (a new account's limit is zero); a namespace owner raises it in Settings → Billing. |
held: billing_setup_required | The namespace has no billing account yet; a namespace owner sets it up in Settings → Billing. |
held: machine_unpriced | The machine type has no price; a platform admin sets one with --price. |
held: evidence_authorizer_invalid | The person who authorized result writes can no longer write; run dl cict platform source. |
routing refused:platform_not_activated | DreamLake GPU is enabled but not started; run dl cict platform start. |
routing refused:platform_not_allowed | DreamLake GPU is not enabled for the namespace. |
routing rejected:public_repository | Public repositories are refused unless the server allows them. |
runner failed with target_not_supported | The pool targets a host; only Slurm pools start runners today. |
evidence failed | The run declared no readable results.json, or a declared path was invalid. See the product guide. |