v1.226.1

fix: grant workflows:write so release-major-tag can move the v1 tag Erik Osterman (Cloud Posse) (@osterman) (#2962)

What

Adds workflows: write to the permissions: block in .github/workflows/release-major-tag.yml.

Why

The release-major-tag job force-pushes the moving v1 tag (so external consumers can reference cloudposse/atmos/actions/cache@v1) using only contents: write. GitHub rejects any push — including a tag force-push — whose resulting tree differs from the target ref in .github/workflows/**, unless the token also has workflows: write.

Confirmed via gh run view --log on both a passing run (v1.225.0) and failing runs (v1.223.0, v1.226.0):

refs/tags/v1:refs/tags/v1 [remote rejected] (refusing to allow a GitHub App
to create or update workflow `.github/workflows/codeql.yml` without
`workflows` permission)

This made the job fail intermittently — specifically whenever a .github/workflows/* file changed since v1 was last successfully moved — forcing consumers to pin exact release tags (e.g. @v1.226.0) instead of the moving @v1 tag. The tagger action authenticates purely via github.token, so widening this workflow's own permissions: block is sufficient; no PAT or other change is required.

References

• .github/workflows/release-major-tag.yml
• Failing runs: v1.223.0 (29426305603), v1.226.0 (32307634756)

Summary by CodeRabbit

• Chores
  • Improved release automation to support more reliable creation and maintenance of major-version tags.
  • Updated the release process to use dedicated authorization for tag updates.
  • No user-facing product features, functionality, or interface changes are included in this update.

🚀 Enhancements

fix: name the offending component/stack in backend_type mismatch errors Erik Osterman (Cloud Posse) (@osterman) (#2965)

what

• checkTerraformBackendTypeMatch / checkRemoteStateBackendTypeMatch now name the offending component and stack directly in the error's hint text, e.g.:

  component "bootstrap" in stack "sandbox": remote_state_backend_type is "local" but remote_state_backend: only configures s3...

• Added a Go unit test (TestProcessTerraformRemoteStateBackend_InheritedLocalTypeMismatch) reproducing the exact real-world shape that surfaced this, plus hint-text assertions on the existing mismatch tests.
• Added a CLI-level fixture/test-case (tests/fixtures/scenarios/remote-state-backend-type-mismatch/, tests/test-cases/remote-state-backend-type-mismatch.yaml) proving the fix end-to-end through the real binary.

why

• These two checks (added in #2953) run during whole-repo stack processing, so a mismatch in one component blocks describe stacks/terraform plan for every other stack too, including ones totally unrelated to the misconfigured component.
• The component/stack causing the failure were previously attached only via WithContext, which the CLI's default (non --verbose) error renderer drops entirely. The printed error gave zero indication of which of potentially hundreds of components across a repo was actually at fault.
• Found live in cloudposse/infra-live: atmos describe stacks --stack core-gbl-marketplace failed with only "remote_state_backend_type is local but remote_state_backend: only configures s3" — no component, no stack. It took manually instrumenting the binary with debug prints to discover the real cause was an unrelated plat/sandbox component (vpc-no-provider) that overrides backend_type: local without a matching remote_state_backend_type. With this fix, that same failure now reads:

  component "vpc-no-provider" in stack "orgs/cplive/plat/sandbox/us-east-2": remote_state_backend_type is "local" but remote_state_backend: only configures s3...
  which is immediately actionable.

references

None — found and reported internally, no existing issue.

Summary by CodeRabbit

• Bug Fixes
  • Improved Terraform backend and remote-state mismatch errors with clearer component and stack context.
  • Enhanced error hints to identify conflicting backend types, including inherited configuration issues.
• Tests
  • Added regression coverage for backend mismatches across regular and remote-state configurations.
  • Added scenario coverage for mismatches involving unrelated stacks.
  • Improved parallel HTTP redirect test isolation to prevent connection cleanup races.

fix(terraform): support Terraform 1.15+ module source interpolation Erik Osterman (Cloud Posse) (@osterman) (#2915)

what

• Atmos no longer fails to parse a Terraform component whose module block uses variable interpolation in source (e.g. source = "./mods/${var.org}") when the variable is declared const = true — valid syntax under Terraform 1.15+, not just OpenTofu 1.8+.
• Generalized the existing "Variables not allowed" diagnostic skip in internal/exec/utils.go / internal/exec/terraform_detection.go so it no longer depends on detecting OpenTofu — renamed isKnownOpenTofuFeature → isKnownModuleSourceInterpolationDiagnostic, and the component_info flag validation_skipped_opentofu → validation_skipped_module_source_interpolation.
• Hardened that skip so it can never silently swallow a genuine, unrelated HCL error that happens to co-occur in the same module: diagnostics are now inspected individually and grouped by source position (allDiagnosticsAreModuleSourceInterpolation), instead of pattern-matching the collapsed Diagnostics.Error() string, which only renders the first diagnostic's text.
• Added a new regression fixture/test reproducing the exact issue on plain terraform (no command: override), plus a fixture/test proving a real unrelated error is still surfaced when it co-occurs with the known-safe diagnostic.
• Investigated whether Atmos's SBOM generation is affected by dynamic module sources; confirmed it isn't (it reads already-resolved sources from terraform modules -json, never the static parser), and added a permanent guard test (pkg/sbom/terraform_test.go) for that invariant.
• Bumped the nanoid pnpm override in website/package.json to resolve the transitive website/pnpm-lock.yaml dependency to nanoid@3.3.18, fixing an open Dependabot alert (infinite loop on zero-size input). The two open image-size alerts have no upstream patch yet and are not auto-fixable.
• Fixed pre-existing EditorConfig violations (tabs instead of the required 2-space indent) in docs/prd/opentofu-module-source-interpolation.md, surfaced once that file entered the branch's diff.

why

• Atmos pre-parses every Terraform component with terraform-config-inspect before running any Terraform/OpenTofu command. That library decodes a module's source attribute with a nil hcl.EvalContext, so any variable reference there always produces the "Variables not allowed" diagnostic — regardless of whether the configured tool/version actually supports it.
• Atmos already tolerated this diagnostic for OpenTofu 1.8+ (PR #1756), but Terraform 1.15 (April 2026) added the equivalent capability via const = true variables, so plain-Terraform users hit the same diagnostic as a hard failure even though their syntax is valid.
• The diagnostic text can't distinguish "valid under a modern tool" from "genuinely invalid" — Atmos already accepted that ambiguity unconditionally for OpenTofu, so extending the same leniency to Terraform is consistent, provided a real unrelated error can never be silently discarded alongside it (the second commit's fix).

references

• closes #2913
• docs/prd/opentofu-module-source-interpolation.md (updated with a 2026-08-10 addendum)
• docs/fixes/2026-08-10-terraform-module-source-interpolation.md

Summary by CodeRabbit

• Bug Fixes

  • Added support for Terraform 1.15+ interpolation in module source paths.
  • Prevented known parser diagnostics from suppressing unrelated configuration errors.
  • Preserved resolved dynamic module sources in component and SBOM metadata.

• Documentation

  • Documented supported Terraform behavior, validation handling, and expected component results.

• Tests

  • Added regression coverage for valid interpolation, mixed diagnostics, component metadata, and SBOM output.

fix(container): build.load works without bake; portable bake vars Erik Osterman (Cloud Posse) (@osterman) (#2963)

what

• Adds a standalone load: true field to the plain (non-bake) engine: buildx container build step, so docker buildx build --load works without adopting bake:.
• Wires the new field through schema, build-config, arg-building, and validation (load: true now requires engine: buildx, mirroring the existing driver/cache check).
• Switches bake.vars from the --var CLI flag to environment-variable injection (NAME=value), since docker buildx bake has always resolved HCL variable {} blocks from the environment, while --var is a newer flag missing from older buildx builds (e.g. Debian Trixie's 0.13.1).
• Documents both changes in website/docs/workflows/workflows/workflow/steps/type/container.mdx and records the fixes in docs/fixes/.

why

• With a non-default Buildx driver (e.g. docker-container), a build's output lands in BuildKit's own cache rather than the local Docker image store, so a following push step can't find the image unless --load is passed. Previously load only existed under bake:, forcing anyone who needed --load to also adopt an external docker-bake.hcl file just to flip one boolean.
• docker buildx bake --var isn't implemented on every buildx release (e.g. Debian Trixie ships 0.13.1, which predates docker/buildx#3610), so builds using bake.vars failed with unknown flag: --var on those hosts. Environment-variable injection is the pre-existing, universally-supported mechanism docker buildx bake uses to resolve HCL variable "NAME" {} blocks, so it fixes portability with no change in what a bake file can express.

references

• docs/fixes/2026-08-20-container-build-load-without-bake.md
• docs/fixes/2026-08-19-container-bake-vars-env-injection.md

Summary by CodeRabbit

• New Features

  • Added load support for plain Buildx builds, allowing images to be added to the local Docker image store.
  • Improved Bake variable handling through environment-based resolution and broader Buildx compatibility.

• Bug Fixes

  • Added validation for incompatible image-loading configurations.
  • Improved reliability for slower environments during container session startup.

• Documentation

  • Documented image-loading requirements and Bake variable behavior.
  • Added fix notes covering container build and Bake variable updates.

fix(ci): recover per-run assertion detail in test summary fallback Erik Osterman (Cloud Posse) (@osterman) (#2959)

what

• The CI job-summary fallback for terraform test output (used when per-run run "name"... pass/fail status lines weren't captured) now recovers the failing assertion's file, line, and message from terraform's Error: diagnostic block when it survived in the captured output.
• Previously the fallback always synthesized a bare aggregate row like test summary (per-run detail unavailable): N passed, M failed with no location or message, even when that detail was still present in the raw text.
• Added errorLocationRe to parse the on <file> line <N>: locator out of a terraform error block, and reuse the existing ExtractErrorBlocks helper to populate the synthesized row's Error/File/Line fields. No template changes were needed — templates/test.md already renders those fields for fail/error rows.
• Added a fix-log record at docs/fixes/2026-08-19-ci-test-summary-fallback-recovers-error-detail.md.

why

• A prior fix (docs/fixes/2026-08-14-ci-summary-test-table-fallback-dropped.md) stopped the results table from disappearing entirely on this fallback path, but the synthesized row still carried no per-test detail — CI test summaries showed only aggregate pass counts and a reproduction command, with no individual test-run/assertion detail, even when that detail was actually recoverable from the captured output.
• A reproduction test (TestTestTemplate_SummaryFallback_LosesRunDetail) confirmed the gap before this fix; it's retained to document the remaining, genuinely irreducible case where no Error: block survives at all.

references

• docs/fixes/2026-08-14-ci-summary-test-table-fallback-dropped.md — the prior fix this builds on.
• docs/fixes/2026-08-19-ci-test-summary-fallback-recovers-error-detail.md — this fix's record.

Summary by CodeRabbit

• Bug Fixes

  • Improved Terraform test failure summaries by recovering file and line details when a single error is available.
  • Preserved complete error messages without corrupting report tables or duplicating content.
  • Avoided assigning potentially incorrect locations when multiple errors are present.
  • Increased the Terraform registry cache CI timeout from 20 to 30 minutes.

• Documentation

  • Added fix documentation covering CI summary recovery and registry cache timeout handling.

fix: type: store step misrouted to container decoder in custom commands and hooks Erik Osterman (Cloud Posse) (@osterman) (#2961)

what

• Fix decodeStepWith (pkg/schema/workflow.go) so a step is only routed to the container with: decoder when type: container, instead of whenever action: is non-empty.
• Fix the kind: step hook bridge (pkg/hooks/step_engine.go) to backfill WorkflowStep.With from the hook's with: payload when the normal decode leaves it nil.
• Add regression tests covering both the workflow-file and custom-command/Viper decode paths for type: store, and both the static and runtime hook decode paths.

why

• A documented custom-command/workflow step shaped like:

  - type: store
    action: write
    with:
      store: image-metadata
      key: image-dev
      value: "..."
      stack: dev
      component: app

  failed before ever reaching execution with container action: writedoes not accept awith: block. decodeStepWith treated any step with a non-empty action: as a container step regardless of type:, so type: store (and any other non-container type that sets action:) was misrouted into the container decoder.
• Investigating the same class of bug surfaced a second, independent issue: the documented kind: step / type: store component-hook pattern (see /workflows/steps/type/store) also silently dropped its store/key/value config, because the hook bridge round-trips the hook's with: payload directly into WorkflowStep's top-level fields — which works for step types with flat fields (archive, say) but not for step types like store/tflint whose config lives only in the generic With map. StoreHandler.Validate then failed with a generic "store is required" error that never showed the store the user actually configured.

references

• N/A

Summary by CodeRabbit

• Bug Fixes

  • Preserved with values for store hooks when decoding workflow steps.
  • Kept step parameters consistent across workflow, runtime, YAML, and map-based decoding.
  • Limited container-specific processing to container steps.
  • Prevented valid parameters on non-container steps from being lost or misinterpreted.
  • Kept store step parameters available without populating container-only fields.
  • Increased the Terraform registry cache timeout for Windows jobs to improve reliability.

• Documentation

  • Documented the step-parameter decoding and Windows timeout fixes.

fix(emulator): join Atmos's container to the shared network when reuse fails Erik Osterman (Cloud Posse) (@osterman) (#2960)

what

• Atmos's own container now joins the dedicated per-stack Docker/Podman network when it can't reuse its existing one, via a new NetworkConnector runtime capability (docker/podman network connect).
• Hardens the last-resort emulator endpoint guess to prefer host.docker.internal (only when it actually resolves) before falling back to the default-gateway IP guess.
• Adds a real, unmocked regression test that runs entirely inside a nested container (no --network, host socket mounted) to prove the self-detection/join mechanism works against a real daemon, not just a mocked runtime.

why

• atmos terraform test --ci run inside a CI job container (talking to Docker only through a mounted socket) started the AWS emulator successfully but reported an endpoint (http://172.17.0.1:<port>) unreachable from that same job container -- connection refused against the AWS provider's GetCallerIdentity call.
• Root cause: a job container started with a plain docker run (no --network) sits on Docker's default bridge, which CurrentContainerNetwork correctly excludes from reuse (no embedded DNS/aliases). Reuse failing meant the endpoint fell back to a guessed default-gateway IP, which isn't where Docker Desktop's port-forwarding actually listens for sibling containers.
• Instead of only checking whether the existing network happens to be reusable, AttachSharedNetwork now actively makes it reusable by connecting Atmos's own container to the dedicated network too -- so the existing DNS-alias endpoint logic just works, for every built-in emulator driver, not just AWS.
• Verified live against the reported reproduction (disposable copy of the affected application repo, docker:cli job container, no --network, mounted host socket): the emulator now reports a DNS alias instead of an IP, and atmos terraform test app -s fixtures --ci completes fully (Success! 1 passed, 0 failed, 0 skipped.) where it previously failed with connection refused.
• New pkg/container/sibling_network_test.go + sibling_network_docker_test.go (opt-in via ATMOS_TEST_SIBLING_CONTAINER=1) reproduces the bug end-to-end inside a real nested container -- confirmed it fails with no such host when the join logic is reverted, and passes with it in place.

references

• Closes the job-container endpoint-reachability gap left open by #2942 ("Shared per-stack networking for containers, emulators & run steps").

Summary by CodeRabbit

• New Features
  • Running containers can join dedicated Docker or Podman networks with optional DNS aliases.
  • Stack containers automatically connect to shared networks when supported.
  • Improved access to published services through host.docker.internal.
• Bug Fixes
  • Network connections safely handle already-connected containers.
  • Network attachment failures no longer block container creation.
  • Host gateway detection avoids hanging during unavailable DNS resolution.
• Tests
  • Added coverage for aliases, network connectivity, runtime behavior, and container communication.
• Documentation
  • Documented emulator endpoint and container networking improvements.