feat(cli): build a remote sandbox base image from a Dockerfile and save it by digest by ItamarZand88 · Pull Request #750 · alienplatform/alien

ItamarZand88 · 2026-09-26T18:15:11Z

Summary

alien projects capabilities enable remote-sandbox --src ./sandbox builds the sandbox base image from a Dockerfile, pushes it to the project's private repository and saves it by digest, so a private image no longer needs a hand-run docker login/tag/push. Re-running with an unchanged directory, on any machine, reuses the image and changes nothing.

When a developer runs the command with --src:

The CLI hashes the source directory (relative paths and bytes, the Dockerfile, build args, target) and looks up remote-sandbox-src-<hash> in the project's repository.
On a miss it builds with Docker, pushes, tags the manifest with that hash and reads back the digest; on a hit it skips both.
It saves <registry>/<repository>@sha256:… as the capability's base image.

This PR changes a private sandbox image from four manual Docker commands to one CLI command.

What I did

--image (was --base-image, kept as a hidden alias) for a prebuilt image; --src with optional --dockerfile to build; --rebuild to build again even when the directory is unchanged, re-pulling FROM.
Pulled the manager proxy push settings and the cached stack push out of release.rs so release and this command share them.
ECR CreateRepository is now sent once. It was retried on the 400 "already exists" answer for ~7s before being treated as success, which made a repeat request time out. Repository lookups try the routable name first.
The configure request carries the project's saved Azure sandbox settings. The endpoint replaces the whole sandbox configuration, and the command used to send the AWS half only, which turned a configured Azure sandbox off.
The Docker build cache key now includes the Dockerfile's contents, and a build whose source changed mid-build is not cached. Side effect: an existing Docker worker rebuilds once on its next alien release.

How I tested

Unit: alien-build 150, alien-aws-clients 76 (one asserts a 400 "already exists" makes exactly one HTTP request), alien-bindings 346, alien-cli 333.
End to end against a real API, manager and ECR, with a Dockerfile FROM alpine:3.20:
- repository already exists: no timeout;
- unchanged rerun: Unchanged …:remote-sandbox-src-<hash> in 3s, same digest, nothing pushed;
- edited Dockerfile: new build and digest;
- a copy of the directory at another path with no local cache: reused, same digest;
- --rebuild: new build with --pull;
- --base-image alias with a public image still configures;
- a project with a saved Azure sandbox keeps it after both --src and --image.
Registry tests against a local registry: re-tagging a manifest and a multi-arch index keeps the digest; a missing tag reads as absent; a 403 is an error, not a miss.
What it checked on the auth path:
- The push authenticates to the manager proxy with the same workspace and token as alien release; no new credential is created or stored.
- The saved reference is renamed under the host and repository the API returned, and the CLI refuses a pushed reference outside that repository (test configure_refuses_an_image_outside_the_pushed_repository).
- Registry errors are classified by status and never carry the raw registry response, which can contain signed URLs.
- Nothing turned up.

…I client

… push

`alien projects capabilities enable remote-sandbox --src <dir>` builds the Dockerfile in <dir> as the sandbox base image, pushes it to the project's private repository through the manager, and configures the capability with the pushed reference. A rerun with an unchanged build context reuses the earlier push through the release push cache, so the reference does not change.

…rror Adds SandboxImagePushFailed so the registry cause is the headline, resolves the manager before creating the project repository, marks the two pushed-reference invariants as GenericError, and drops Platform clones.

…sh cache atomically A missing --src directory is a validation error, not a retryable file failure. The push cache is written through a temp file and renamed, so an interrupted write cannot empty it and force a re-push under a new tag.

A retried create answers "already exists" to its own first attempt, and every HTTP error is retryable in the shared client, so an existing repository cost three backoff retries before it was reported as a conflict. CreateRepository now goes out once through sign_send_json_once and callers decide whether to retry.

create_repository always creates the prefixed name, so describing the logical name first cost a failed request on every lookup. The logical name stays as the fallback.

…re-pulls bases BuildSettings.rebuild builds every source resource again: the local artifact cache is not consulted and Docker builds pass --pull, so a change that only exists upstream in a FROM image is picked up. Off for every existing caller.

docker_source_input_hash covers the files under src by relative path, the Dockerfile, build args, stage and target platforms, and not the absolute path, so two checkouts of the same tree agree on it. registry::manifest_digest reads a tag's digest and reports a missing manifest as absent; registry::tag_manifest points a second tag at a pushed manifest by copying its bytes, so an image index keeps its digest.

…y digest `alien projects capabilities enable remote-sandbox --src` now hashes the build inputs and looks for remote-sandbox-src-<hash> in the project repository first. A hit skips the build and push, from any machine; a miss builds, pushes, and tags the pushed manifest with that name. The capability is configured with the manifest digest. --rebuild skips that lookup and the local artifact cache and re-pulls FROM images. --image replaces --base-image, which still parses as a hidden alias. A missing Dockerfile is refused before any request. The repository is now ensured before the build, since the lookup needs it.

…misses A pull role scoped to the repository prefix answers 403 for the logical name, which hid the in-scope 404. Also state why CreateRepository is sent once.

…or out A refused lookup or tag is IMAGE_LOOKUP_REJECTED / IMAGE_PUSH_REJECTED (not retryable); a transport error, 5xx or 429 stays retryable. The OCI error is never attached, as registry errors can carry signed URLs.

The source tag is shared by every machine with the same tree, so the input hash is recomputed after the build and a mismatch fails the run. The sandbox push skips the release push cache, so --rebuild always pushes and a cached reference the registry dropped cannot break the tag step. A failed push-cache save is logged as a warning.

…e local cache The artifact is cached under the pre-edit tree's key, so reverting the edit would have reused it and tagged it as that tree.

…sable A token endpoint failure reads as a failed token request, not a refused credential.

…tifact cache The cache key is taken before the build reads the tree. It is taken again before the cache entry is written, and on a mismatch the artifact is left uncached, so an interrupted run or a reverted edit cannot reuse it. This covers alien release as well; the sandbox CLI keeps only its own check before tagging.

…ache key A Dockerfile outside the source directory, or under a directory the source walk skips, was keyed by path only, so an edit to it reused the old artifact. A missing Dockerfile is left to the build to report. A mid-build re-key that fails counts as a change: the artifact stays uncached rather than failing a finished build.

…a failed read The wrapper covers lookup, push and tag, so it is SANDBOX_IMAGE_PUBLISH_FAILED. IMAGE_LOOKUP_REJECTED no longer claims a refusal for an unusable manifest. An empty ECR repository id looks up the shared repository, as create does.

A re-key that fails is logged with its error rather than as a source change.

…WS one The remote sandbox configure endpoint rebuilds the whole configuration from the request, so an omitted `azure` key turns Azure off. `alien projects capabilities enable remote-sandbox` always omitted it, which deleted a project's saved Azure sandbox whenever the AWS base image was set with --image or --src. Read the project's saved Azure settings right before the configure call and send them with the new base image. A project with no Azure sandbox still omits the key.

ItamarZand88 · 2026-09-26T18:15:27Z

@greptile review

greptile-apps · 2026-09-26T18:20:43Z

Confidence Score: 5/5

[High risk] Adds container registry operations and changes build caching logic.

The PR appears safe to merge based on the reviewed changes.

Summary

The PR adds a CLI flow to build and publish a remote sandbox base image by source hash, then configure it by digest. Since the previous review, it separates forced source rebuilding from base-image pulling and carries saved Azure settings into the configure request.

_{Reviews (3) · Last reviewed commit: "fix(cli): pull base images only on --reb..."}

…image The source input hash reused the artifact cache's file walk, which skips node_modules, target, .git and symlinks. Docker sends all of them unless .dockerignore says otherwise, so an edit there left the hash unchanged and reused a stale image. The hash now walks the context with Docker's .dockerignore rules (the Dockerfile-specific ignore file first), and covers file contents, the executable bit, symlink targets and directories. It skips the build's own output directory when that sits inside the source, since the build writes there between the hashes taken before and after it.

…ith its hash On a registry hit the command configured the tagged digest using the hash taken before the lookup, so a tree edited in between got the older image. Both paths now rehash before they use the tag and fail the same way when the source moved.

Two releases in one directory shared push-cache.json.tmp, so one could rename the other's file away and fail its write. Each write now goes through a uniquely named temp file in the cache directory.

ItamarZand88 · 2026-09-26T19:25:36Z

@greptile review

… across generated types

ItamarZand88 · 2026-09-26T19:54:03Z

@greptile review

ItamarZand88 added 20 commits September 26, 2026 21:05

chore(sdk): generate the sandbox base image repository call in the AP…

17c0b2f

…I client

refactor(cli): share manager proxy push settings and the cached stack…

26fefe3

… push

fix(bindings): look up an ECR repository by its routable name first

fd24c1c

create_repository always creates the prefixed name, so describing the logical name first cost a failed request on every lookup. The logical name stays as the fallback.

fix(bindings): report the routable name's miss when every ECR lookup …

a73dc64

…misses A pull role scoped to the repository prefix answers 403 for the logical name, which hid the in-scope 404. Also state why CreateRepository is sent once.

fix(cli): drop a sandbox build whose source changed mid-build from th…

335fb59

…e local cache The artifact is cached under the pre-edit tree's key, so reverting the edit would have reused it and tagged it as that tree.

fix(build): name the source image when its manifest media type is unu…

999ed8d

…sable A token endpoint failure reads as a failed token request, not a refused credential.

fix(build): say why an artifact was left uncached after its build

3a4401a

A re-key that fails is logged with its error rather than as a source change.

greptile-apps Bot reviewed Sep 26, 2026

View reviewed changes

Comment thread crates/alien-build/src/source_input.rs Outdated

Comment thread crates/alien-cli/src/commands/platform/projects.rs

Comment thread crates/alien-cli/src/commands/release.rs Outdated

ItamarZand88 added 5 commits September 26, 2026 21:45

fix(cli): give each push cache write its own temporary file

6b88b60

Two releases in one directory shared push-cache.json.tmp, so one could rename the other's file away and fail its write. Each write now goes through a uniquely named temp file in the cache directory.

fix(build): match buildx's context walk when hashing a source image

052c8bf

fix(cli): always build the sandbox image on a registry miss

73b0cee

greptile-apps Bot reviewed Sep 26, 2026

View reviewed changes

Comment thread crates/alien-cli/src/commands/platform/projects.rs

fix(cli): pull base images only on --rebuild and carry Azure settings…

65b2bdc

… across generated types

ItamarZand88 merged commit f2909b7 into main Sep 26, 2026
29 checks passed

ItamarZand88 deleted the itamar/alien-1034-give-the-sandbox-capability-a-private-image-without-typing branch September 26, 2026 20:19

Navigation Menu

Repository navigation

feat(cli): build a remote sandbox base image from a Dockerfile and save it by digest - #750

feat(cli): build a remote sandbox base image from a Dockerfile and save it by digest#750
ItamarZand88 merged 26 commits into
mainfrom
itamar/alien-1034-give-the-sandbox-capability-a-private-image-without-typing

ItamarZand88 commented Sep 26, 2026

Uh oh!

ItamarZand88 commented Sep 26, 2026

Uh oh!

greptile-apps Bot commented Sep 26, 2026 •

edited

Loading

Uh oh!

Uh oh!

Uh oh!

Uh oh!

ItamarZand88 commented Sep 26, 2026

Uh oh!

Uh oh!

ItamarZand88 commented Sep 26, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

Conversation

ItamarZand88 commented Sep 26, 2026

Summary

What I did

How I tested

Uh oh!

ItamarZand88 commented Sep 26, 2026

Uh oh!

greptile-apps Bot commented Sep 26, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Confidence Score: 5/5

Summary

Uh oh!

Uh oh!

Uh oh!

Uh oh!

ItamarZand88 commented Sep 26, 2026

Uh oh!

Uh oh!

ItamarZand88 commented Sep 26, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

greptile-apps Bot commented Sep 26, 2026 •

edited

Loading