fix(api): restore image resolution so e2e-cloud-test can boot boxes

[G4614]

Branches off chore/e2e-required-merge-gate and folds in #735's image-resolution rebuild so the e2e-cloud test suite can boot boxes again — box-start.action.ts currently fails every create with Box image resolution is unavailable: the image/template subsystem was removed because the rebuild PR hasn't landed.

Test plan

	Before	After
POST /v1/{prefix}/boxes on the Tokyo stack	400 Bad Request: "Box failed to start: Box image resolution is unavailable: the image/template subsystem was removed"	Box accepted, transitions UNKNOWN → CREATING → STARTED
apps/api/src/box/entities/box.entity.ts	no image column	image: string first-class column (#735's 02d599f5)
apps/api/src/box/runner-adapter/runnerAdapter.v2.ts	createBox removed in 7ec370b7	restored via CREATE_BOX job carrying the OCI ref (#735's 02d599f5..be82c1f0)
e2e-cloud-test.yml quota SQL	shell-escape-stripped, PG saw lowercase organizationid	base64-encoded round-trip, PG sees the literal quoted identifier (this branch)
e2e-cloud-test.yml snapshot pre-registration	hits non-existent /api/snapshots, 404	block removed (this branch)

Summary by CodeRabbit

• New Features

  • Curated OCI image selector in the dashboard; new Tokyo cloud E2E pipeline and reusable deploy/test workflows.

• Changes

  • Box and Workspace now include a required image field; public Box IDs migrated to short public-id format; onboarding/examples use pinned image refs; dashboard now targets cloud-backed box APIs.

• Tests

  • E2E and unit tests updated to use pinned/default images and new bootstrap behavior.

• Documentation

  • CI docs updated to describe the cloud regression pipeline, triggers, and required settings.

[coderabbitai]

Note

Reviews paused

It looks like this branch is under active development. To avoid overwhelming you with review comments due to an influx of new commits, CodeRabbit has automatically paused this review. You can configure this behavior by changing the reviews.auto_review.auto_pause_after_reviewed_commits setting.

Use the following commands to manage reviews:

• @coderabbitai resume to resume automatic reviews.
• @coderabbitai review to trigger a single review.

Use the checkboxes below for quick actions:

• ▶️ Resume reviews
• 🔍 Trigger review

📝 Walkthrough

Walkthrough

Adds reusable GitHub workflows for cloud deploys and E2E runs; makes Box image required and switches public Box identity to the 12-char id; routes box creation through runner CREATE_BOX jobs; migrates the dashboard to a CloudBox client and removes the TypeScript SDK; updates e2e defaults and fixture setup.

Changes

Cloud box platform rollout

Layer / File(s)	Summary
Cloud deployment and test workflows .github/workflows/*	New and updated GitHub Actions workflows add deploy, build, and E2E orchestration for the Tokyo cloud stack, including reusable deploy workflows, runner/runner-binary build paths, and e2e-cloud test wiring.
Box image & identity contracts openapi/*, apps/api-client-go/*, apps/api/src/box/*, apps/runner/pkg/*, apps/api/src/migrations/*	API/OpenAPI, generated clients, DTOs, TypeORM entities and migrations add an image field and convert box public identity to a 12-character id; generated client models and migrations updated accordingly.
Box creation, lookup & runner flow apps/api/src/box/**/*, apps/api/src/admin/services/*, apps/api/src/boxlite-rest/mappers/*, apps/runner/pkg/*	Backend validates curated images, persists resolved image on create, enqueues runner CREATE_BOX jobs (RunnerAdapterV2.createBox), updates cache-key helpers/invalidation to use id, and adapts observability/mapper logic to the new id/image shape.
Dashboard CloudBox migration apps/dashboard/**/*, apps/dashboard/src/lib/cloudBox.ts, apps/libs/api-client/*	Dashboard replaces direct SDK usage with a local CloudBox abstraction, threads image through create/load flows, limits code-snippets to Python, updates imports/aliases, and sets image defaults to pinned digests.
E2E image defaults & fixture setup scripts/test/e2e/**/*, apps/infra-local/scripts/stack-up.sh	E2E fixtures and drivers derive DEFAULT_IMAGE from credentials/API-provided env, remove snapshot registration, and add API-provided system image env keys used for local tests; tests updated to use DEFAULT_IMAGE.
Legacy SDK removal & build cleanup apps/libs/sdk-typescript/**/*, apps/dashboard/vite.config.mts, apps/eslint.config.mjs	The TypeScript SDK package and related build/docs/tests/config were removed; dashboard Vite alias and ESLint overrides adjusted accordingly.

Sequence Diagram(s)

sequenceDiagram
  participant E2ECloud as e2e-cloud.yml
  participant DeployRunner as deploy-runner.yml
  participant DeployAPI as deploy-api.yml
  participant DeployServices as deploy-app-services.yml
  participant E2ETest as e2e-cloud-test.yml
  E2ECloud->>DeployRunner: invoke when runner_chain changes
  E2ECloud->>DeployAPI: invoke when api changes
  E2ECloud->>DeployServices: invoke when supporting services change
  E2ECloud->>E2ETest: run shared Tokyo E2E suite

Loading

Estimated code review effort

🎯 5 (Critical) | ⏱️ ~90+ minutes

Possibly related PRs

• boxlite-ai/boxlite#743: Overlaps removal of the publishable apps/libs/sdk-typescript SDK and migration of dashboard usages to apps/dashboard/src/lib/cloudBox.
• boxlite-ai/boxlite#729: Related changes touching CreateBoxSheet.tsx and box creation UI wiring.
• boxlite-ai/boxlite#678: Touches the e2e stack/fixture setup and related test harness paths.

Suggested reviewers

• law-chain-hot
• DorianZheng

Poem

I hop through YAMLs, tests, and code,
I plant an image, watch boxes load.
Jobs enqueue with gentle taps,
Deploys hum soft like trailing maps.
🐇🎉

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch fix/api-image-resolution-from-735

[coderabbitai]

Actionable comments posted: 13

🧹 Nitpick comments (24)

apps/dashboard/src/lib/cloudBox.ts (1)

404-423: 💤 Low value

TypeScript command suppresses diagnostic 5107 (module resolution errors).

Line 414 includes --ignore-diagnostics 5107 when running ts-node. Diagnostic 5107 is TypeScript's "Cannot find module" error. Suppressing this diagnostic might hide legitimate import errors in user code. If this is intentional for the playground/sandbox environment (e.g., to allow dynamic module resolution), consider adding a comment explaining why this diagnostic is suppressed.

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@apps/dashboard/src/lib/cloudBox.ts` around lines 404 - 423, The ts-node
invocation in getRunCommand for CodeLanguage.TYPESCRIPT currently includes the
--ignore-diagnostics 5107 flag which hides "Cannot find module" errors; update
the CodeLanguage.TYPESCRIPT branch inside getRunCommand to either remove the
--ignore-diagnostics 5107 flag so module resolution errors surface, or if
suppression is intentional for the sandbox/playground, add a clear inline
comment next to the ts-node command explaining why diagnostic 5107 is suppressed
and the safety/constraints that justify it (reference getRunCommand,
CodeLanguage.TYPESCRIPT, and the --ignore-diagnostics 5107 token to locate the
change).

apps/dashboard/src/components/Playground/VNC/Interaction/Screenshot.tsx (1)

127-129: ⚡ Quick win

Update the outdated comment.

The comment on Line 127 references ComputerUse but the code now casts to CloudBoxComputerUse.

📝 Suggested comment update

-  // Disable logic ensures that this method is called when ComputerUseClient exists -> we use as ComputerUse to silence TS compiler
+  // Disable logic ensures that this method is called when ComputerUseClient exists -> we use as CloudBoxComputerUse to silence TS compiler
   const screenshotActionAPICall: PlaygroundActionInvokeApi = async (screenshotActionFormData) => {
     const ScreenshotActionsClient = (ComputerUseClient as CloudBoxComputerUse).screenshot

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@apps/dashboard/src/components/Playground/VNC/Interaction/Screenshot.tsx`
around lines 127 - 129, The comment above the screenshotActionAPICall is
outdated—update it to reflect the current cast to CloudBoxComputerUse instead of
ComputerUse; edit the comment that references ComputerUse/disable logic so it
accurately states that the method is called when ComputerUseClient exists and
that we cast to CloudBoxComputerUse to satisfy TypeScript for the
ScreenshotActionsClient assignment (referencing ComputerUseClient,
CloudBoxComputerUse, and screenshotActionAPICall).

apps/dashboard/src/components/Playground/VNC/Interaction/Display.tsx (1)

38-42: ⚡ Quick win

Update the outdated comment.

The comment on Line 38 references ComputerUse but the code now casts to CloudBoxComputerUse.

📝 Suggested comment update

-  // Disable logic ensures that this method is called when ComputerUseClient exists -> we use as ComputerUse to silence TS compiler
+  // Disable logic ensures that this method is called when ComputerUseClient exists -> we use as CloudBoxComputerUse to silence TS compiler
   const displayActionAPICall: PlaygroundActionInvokeApi = async (displayActionFormData) => {
     const displayActionResponse = await (ComputerUseClient as CloudBoxComputerUse).display[

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@apps/dashboard/src/components/Playground/VNC/Interaction/Display.tsx` around
lines 38 - 42, Update the outdated comment above displayActionAPICall to reflect
that we now cast ComputerUseClient to CloudBoxComputerUse (not ComputerUse) to
satisfy the TypeScript compiler; mention the real reason for the cast (ensuring
ComputerUseClient exists and typing as CloudBoxComputerUse) and reference the
symbols displayActionAPICall, ComputerUseClient, CloudBoxComputerUse, and
DisplayActions so future readers understand why the cast is used.

apps/dashboard/src/components/Playground/VNC/Interaction/Keyboard.tsx (1)

81-83: ⚡ Quick win

Update the outdated comment.

The comment on Line 81 references ComputerUse but the code now casts to CloudBoxComputerUse.

📝 Suggested comment update

-  // Disable logic ensures that this method is called when ComputerUseClient exists -> we use as ComputerUse to silence TS compiler
+  // Disable logic ensures that this method is called when ComputerUseClient exists -> we use as CloudBoxComputerUse to silence TS compiler
   const keyboardActionAPICall: PlaygroundActionInvokeApi = async (keyboardActionFormData) => {
     const KeyboardActionsClient = (ComputerUseClient as CloudBoxComputerUse).keyboard

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@apps/dashboard/src/components/Playground/VNC/Interaction/Keyboard.tsx` around
lines 81 - 83, The comment above keyboardActionAPICall is outdated: it mentions
casting to ComputerUse but the code actually casts ComputerUseClient to
CloudBoxComputerUse; update the comment to accurately state that the disable
logic ensures this method is called only when ComputerUseClient exists and that
we cast to CloudBoxComputerUse (not ComputerUse) to satisfy the TypeScript
compiler before accessing the keyboard client (KeyboardActionsClient).

apps/dashboard/src/components/Playground/VNC/Interaction/Mouse.tsx (1)

165-167: ⚡ Quick win

Update the outdated comment.

The comment on Line 165 references ComputerUse but the code now casts to CloudBoxComputerUse.

📝 Suggested comment update

-  // Disable logic ensures that this method is called when ComputerUseClient exists -> we use as ComputerUse to silence TS compiler
+  // Disable logic ensures that this method is called when ComputerUseClient exists -> we use as CloudBoxComputerUse to silence TS compiler
   const mouseActionAPICall: PlaygroundActionInvokeApi = async (mouseActionFormData) => {
     const MouseActionsClient = (ComputerUseClient as CloudBoxComputerUse).mouse

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@apps/dashboard/src/components/Playground/VNC/Interaction/Mouse.tsx` around
lines 165 - 167, The comment above mouseActionAPICall is outdated (mentions
ComputerUse) — update it to accurately reflect the current cast to
CloudBoxComputerUse; modify the comment that precedes the line "const
MouseActionsClient = (ComputerUseClient as CloudBoxComputerUse).mouse" so it
explains that the disable logic ensures ComputerUseClient exists and that we
cast to CloudBoxComputerUse to satisfy TypeScript when accessing .mouse.

apps/dashboard/src/components/Playground/Box/CodeSnippets/python.ts (1)

17-20: ⚡ Quick win

Remove redundant type alias.

When createBoxFromTemplate is false, line 19 generates the import string 'CreateBoxFromImageParams as CreateBoxFromImageParams', which aliases a type to itself. This is redundant and can be simplified:

♻️ Proposed fix

       p.config.useBoxCreateParams
         ? p.config.createBoxFromTemplate
           ? 'CreateBoxFromTemplateParams as CreateBoxFromImageParams'
-          : 'CreateBoxFromImageParams as CreateBoxFromImageParams'
+          : 'CreateBoxFromImageParams'
         : '',

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@apps/dashboard/src/components/Playground/Box/CodeSnippets/python.ts` around
lines 17 - 20, The import string generation currently produces a redundant
self-alias when p.config.createBoxFromTemplate is false; update the ternary that
builds the import entry (the expression using p.config.createBoxFromTemplate) so
the false branch returns 'CreateBoxFromImageParams' (no "as
CreateBoxFromImageParams") instead of 'CreateBoxFromImageParams as
CreateBoxFromImageParams', keeping the true branch ('CreateBoxFromTemplateParams
as CreateBoxFromImageParams') unchanged.

apps/dashboard/src/components/Playground/Box/CodeSnippets/index.ts (1)

10-12: 💤 Low value

Consider a more extensible type for future language additions.

The type Record<CodeLanguage.PYTHON, CodeSnippetGenerator> creates a record that can only have a single key (the literal value "python"). If additional languages are added in the future, this type will need to be updated. Consider using a more flexible type:

type SupportedLanguages = CodeLanguage.PYTHON;
export const codeSnippetGenerators: Record<SupportedLanguages, CodeSnippetGenerator> = {
  [CodeLanguage.PYTHON]: PythonSnippetGenerator,
}

This makes it easier to add languages by just updating the union type.

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@apps/dashboard/src/components/Playground/Box/CodeSnippets/index.ts` around
lines 10 - 12, The current declaration for codeSnippetGenerators uses
Record<CodeLanguage.PYTHON, CodeSnippetGenerator>, which hardcodes a single key
and is not extensible; create a union type alias (e.g., SupportedLanguages) that
includes CodeLanguage.PYTHON (and future languages), then change the type
annotation on codeSnippetGenerators to Record<SupportedLanguages,
CodeSnippetGenerator> and keep the existing initializer mapping
[CodeLanguage.PYTHON]: PythonSnippetGenerator so new languages can be added by
extending SupportedLanguages and adding entries to the object.

apps/api/src/box/runner-adapter/runnerAdapter.v2.createBox.spec.ts (1)

22-84: ⚡ Quick win

Expand test coverage for GPU quota and userId mapping.

The test suite validates image passthrough and resource quotas but doesn't cover two key concerns:

1. GPU quota omission: The test doesn't assert whether payload.gpuQuota should be present or absent. Add a test case that sets box.gpu = 1 and validates the payload behavior.

2. userId vs organizationId: The test doesn't validate the payload.userId mapping. Add an assertion to confirm the intended source (either box.userId or box.organizationId).

🧪 Proposed test expansion

+  it('omits GPU quota from payload (CREATE_BOX does not support GPU)', async () => {
+    const createJob = jest.fn().mockResolvedValue(undefined)
+    const adapter = createAdapter(createJob)
+    const box = buildBox()
+    box.gpu = 1
+
+    await adapter.createBox(box)
+
+    const payload = createJob.mock.calls[0][5] as Record<string, unknown>
+    expect('gpuQuota' in payload).toBe(false)
+  })
+
+  it('maps organizationId to userId (verify correct source)', async () => {
+    const createJob = jest.fn().mockResolvedValue(undefined)
+    const adapter = createAdapter(createJob)
+    const box = buildBox()
+
+    await adapter.createBox(box)
+
+    const payload = createJob.mock.calls[0][5] as Record<string, unknown>
+    expect(payload.userId).toBe(box.organizationId)
+    expect(payload.organizationId).toBe(box.organizationId)
+  })

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@apps/api/src/box/runner-adapter/runnerAdapter.v2.createBox.spec.ts` around
lines 22 - 84, Add tests in RunnerAdapterV2.createBox spec to cover GPU quota
and userId mapping: create a new it(...) that calls adapter.createBox with a Box
built via buildBox() but set box.gpu = 1 and box.userId = 'some-user-id', then
read the payload from createJob.mock.calls[0][5] and assert payload.gpuQuota ===
1 (and/or that 'gpuQuota' exists) and assert payload.userId equals the expected
source (if your implementation maps box.userId use that value, otherwise assert
it equals box.organizationId); reference the createBox call and the createJob
mock used in the existing tests to locate where to add these assertions.

.github/workflows/deploy-api.yml (5)

97-97: ⚡ Quick win

Pin checkout action to commit SHA and disable credential persistence.

For supply-chain security, pin actions/checkout to a specific commit SHA and set persist-credentials: false.

🔒 Recommended fix

-      - uses: actions/checkout@v5
+      - uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683  # v5.2.2
+        with:
+          persist-credentials: false

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In @.github/workflows/deploy-api.yml at line 97, Update the workflow step that
uses actions/checkout@v5 to pin to a specific commit SHA and add
persist-credentials: false; specifically replace the "uses: actions/checkout@v5"
entry with "uses: actions/checkout@<COMMIT_SHA>" (use the stable commit SHA from
the actions/checkout repository) and add the input "with: persist-credentials:
false" under that step so the checkout step stops persisting GitHub credentials.

Source: Linters/SAST tools

---

168-172: ⚡ Quick win

Pin AWS credentials action to commit SHA.

For supply-chain security consistency, pin aws-actions/configure-aws-credentials to a specific commit SHA.

🔒 Recommended fix

-        uses: aws-actions/configure-aws-credentials@v4
+        uses: aws-actions/configure-aws-credentials@e3dd6a429d7300a6a4c196c26e071d42e0343502  # v4.0.2

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In @.github/workflows/deploy-api.yml around lines 168 - 172, Replace the
floating tag for the GitHub Action in the workflow step that currently
references aws-actions/configure-aws-credentials@v4 with a pinned commit SHA;
update the step using symbol "uses: aws-actions/configure-aws-credentials@v4" to
use a specific commit hash (e.g.
aws-actions/configure-aws-credentials@<COMMIT_SHA>) so the "uses" entry is
pinned to an immutable release, keeping the rest of the inputs (role-to-assume,
aws-region, role-session-name) unchanged.

Source: Linters/SAST tools

---

100-108: ⚡ Quick win

Pin paths-filter action to commit SHA.

For supply-chain security consistency, pin dorny/paths-filter to a specific commit SHA.

🔒 Recommended fix

-        uses: dorny/paths-filter@v3
+        uses: dorny/paths-filter@de90cc6fb38fc0963ad72b210f1f284cd68cea36  # v3.0.2

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In @.github/workflows/deploy-api.yml around lines 100 - 108, Replace the
floating tag for the GitHub Action with a specific commit SHA to pin the
dependency: locate the step that currently declares "uses:
dorny/paths-filter@v3" (the step that provides the api_source filters) and
change it to "uses: dorny/paths-filter@<commit-sha>" using a specific commit SHA
from the dorny/paths-filter repository; ensure you pick a stable commit SHA (not
a branch or tag) and update the workflow to use that SHA so the paths-filter
behavior is pinned for supply-chain security.

Source: Linters/SAST tools

---

225-225: ⚡ Quick win

Pin ECR login action to commit SHA.

For supply-chain security consistency, pin aws-actions/amazon-ecr-login to a specific commit SHA.

🔒 Recommended fix

-        uses: aws-actions/amazon-ecr-login@v2
+        uses: aws-actions/amazon-ecr-login@062b18b96a7aff071d4dc91bc00c4c1a7945b076  # v2.0.1

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In @.github/workflows/deploy-api.yml at line 225, The workflow currently
references the ECR login action by tag ("uses:
aws-actions/amazon-ecr-login@v2"); replace that tag with the action's immutable
commit SHA (e.g. "uses: aws-actions/amazon-ecr-login@<commit-sha>") to pin the
dependency for supply-chain security, locating the usage string
"aws-actions/amazon-ecr-login@v2" in the deploy workflow and updating it to the
specific commit SHA from the action's GitHub repository, then commit the change
and verify the workflow runs against the pinned SHA.

Source: Linters/SAST tools

---

165-165: ⚡ Quick win

Pin checkout action to commit SHA and disable credential persistence.

For supply-chain security, pin actions/checkout to a specific commit SHA and set persist-credentials: false.

🔒 Recommended fix

-      - uses: actions/checkout@v5
+      - uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683  # v5.2.2
+        with:
+          persist-credentials: false

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In @.github/workflows/deploy-api.yml at line 165, Replace the floating version
"uses: actions/checkout@v5" with a pinned commit SHA for the actions/checkout
action (e.g., "uses: actions/checkout@<commit-sha>") and add a with block
setting persist-credentials: false to disable credential persistence; update the
workflow line that references actions/checkout and add the with:
persist-credentials: false setting so the checkout action is both pinned and
does not persist credentials.

Source: Linters/SAST tools

.github/workflows/build-runner-binary.yml (2)

107-110: ⚡ Quick win

Pin artifact download action to commit SHA.

For supply-chain security consistency, pin actions/download-artifact to a specific commit SHA.

🔒 Recommended fix

-        uses: actions/download-artifact@v4
+        uses: actions/download-artifact@fa0a91b85d4f404e444e00e005971372dc801d16  # v4.1.8

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In @.github/workflows/build-runner-binary.yml around lines 107 - 110, Replace
the floating version for the GitHub Action usage "uses:
actions/download-artifact@v4" with a pinned commit SHA (e.g. "uses:
actions/download-artifact@<commit-sha>") to ensure supply-chain security; update
the workflow entry that currently includes the keys "name: c-sdk-linux-x64-gnu"
and "path: /tmp/c-sdk/" to reference the pinned commit instead of the tag,
commit the change, and verify the workflow still downloads the artifact as
before.

Source: Linters/SAST tools

---

112-125: ⚡ Quick win

Validate archive uniqueness before extraction.

Lines 118 and 121 use head -1 to select from potentially multiple matches. If the artifact contains multiple archives or the extraction produces duplicate libboxlite.a files, the script silently picks the first without warning. Consider adding a uniqueness assertion.

🛡️ Optional improvement

          set -euo pipefail
-         ARCHIVE=$(ls /tmp/c-sdk/boxlite-c-v*-linux-x64-gnu.tar.gz | head -1)
+         ARCHIVES=($(ls /tmp/c-sdk/boxlite-c-v*-linux-x64-gnu.tar.gz))
+         [ "${`#ARCHIVES`[@]}" -eq 1 ] || { echo "::error::Expected 1 archive, found ${`#ARCHIVES`[@]}"; exit 1; }
+         ARCHIVE="${ARCHIVES[0]}"
          [ -n "$ARCHIVE" ] || { echo "::error::No c-sdk archive in artifact"; exit 1; }
          tar xzf "$ARCHIVE" -C /tmp/c-sdk/
-         A=$(find /tmp/c-sdk -name libboxlite.a | head -1)
+         LIBS=($(find /tmp/c-sdk -name libboxlite.a))
+         [ "${`#LIBS`[@]}" -eq 1 ] || { echo "::error::Expected 1 libboxlite.a, found ${`#LIBS`[@]}"; exit 1; }
+         A="${LIBS[0]}"
          [ -n "$A" ] || { echo "::error::libboxlite.a not found in archive"; exit 1; }

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In @.github/workflows/build-runner-binary.yml around lines 112 - 125, The script
currently uses HEAD selection for ARCHIVE and A which can silently pick the
first match; update the block that sets ARCHIVE (variable ARCHIVE=$(ls
/tmp/c-sdk/boxlite-c-v*-linux-x64-gnu.tar.gz | head -1)) and the find for A
(A=$(find /tmp/c-sdk -name libboxlite.a | head -1)) to assert uniqueness: list
matches into an array or count results and if count == 0 emit an error, if count
> 1 emit an error asking to disambiguate; only proceed to tar xzf and cp when
exactly one archive and exactly one libboxlite.a are found, and include
descriptive error messages referencing ARCHIVE and A when failing.

.github/workflows/_deploy-single-service.yml (4)

60-64: ⚡ Quick win

Pin AWS credentials action to commit SHA.

For supply-chain security consistency, pin aws-actions/configure-aws-credentials to a specific commit SHA.

🔒 Recommended fix

-        uses: aws-actions/configure-aws-credentials@v4
+        uses: aws-actions/configure-aws-credentials@e3dd6a429d7300a6a4c196c26e071d42e0343502  # v4.0.2

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In @.github/workflows/_deploy-single-service.yml around lines 60 - 64, Replace
the floating tag "aws-actions/configure-aws-credentials@v4" with a specific
commit SHA to pin the action for supply-chain security; in the workflow step
that uses aws-actions/configure-aws-credentials (the step with inputs
role-to-assume, aws-region, role-session-name), update the action reference to
the exact commit SHA (e.g., aws-actions/configure-aws-credentials@<COMMIT_SHA>)
and commit the change so the workflow uses the pinned revision.

Source: Linters/SAST tools

---

79-81: ⚡ Quick win

Pin ECR login action to commit SHA.

For supply-chain security consistency, pin aws-actions/amazon-ecr-login to a specific commit SHA.

🔒 Recommended fix

-        uses: aws-actions/amazon-ecr-login@v2
+        uses: aws-actions/amazon-ecr-login@062b18b96a7aff071d4dc91bc00c4c1a7945b076  # v2.0.1

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In @.github/workflows/_deploy-single-service.yml around lines 79 - 81, The
GitHub Actions step "ECR login" (id: ecr_login) currently uses the floating
reference "uses: aws-actions/amazon-ecr-login@v2"; replace this with a pinned
commit SHA (e.g. "uses: aws-actions/amazon-ecr-login@<commit-sha>") to lock the
action to a specific immutable revision—update the uses line for the
aws-actions/amazon-ecr-login action to reference the chosen commit SHA.

Source: Linters/SAST tools

---

98-125: ⚖️ Poor tradeoff

Temporary IAM role borrowing bypasses least-privilege.

The workflow borrows Api service's task and execution roles because per-service roles were deleted from IAM. While documented as temporary, this violates least-privilege (borrowed roles grant S3 + CloudWatch permissions the borrower may not need). Consider prioritizing the sst deploy --stage e2e-ci recreation to restore proper per-service isolation.

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In @.github/workflows/_deploy-single-service.yml around lines 98 - 125, The
current job block (resolving API role ARNs via API_TD, TASK_ROLE and EXEC_ROLE)
temporarily borrows Api roles which violates least-privilege; change the step to
avoid automatic borrowing by first attempting to recreate per-service roles (run
`sst deploy --stage e2e-ci` or call an administrative deploy helper) and fail
fast if that fails, or require an explicit workflow input/approval flag to opt
into borrowing; implement this by replacing the logic that sets API_TD /
TASK_ROLE / EXEC_ROLE with a pre-check that runs the deploy (or checks an
explicit input like inputs.allow_role_borrowing) and only falls back to
borrowing if that input is set, and ensure the step emits a clear error and
exits non-zero when neither recreation nor explicit approval is present.

---

57-57: ⚡ Quick win

Pin action to commit SHA and disable credential persistence.

For stronger supply-chain security, pin actions/checkout to a specific commit SHA instead of a mutable tag. Additionally, set persist-credentials: false to prevent the GITHUB_TOKEN from being accessible to subsequent steps via .git/config.

🔒 Recommended fix

-      - uses: actions/checkout@v5
+      - uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683  # v5.2.2
+        with:
+          persist-credentials: false

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In @.github/workflows/_deploy-single-service.yml at line 57, Update the GitHub
Actions checkout step that currently uses "actions/checkout@v5": pin the "uses"
to a specific commit SHA for actions/checkout and add the input
"persist-credentials: false" to the same step to prevent GITHUB_TOKEN from being
stored in .git/config; locate the checkout step in the workflow where "uses:
actions/checkout@v5" appears and replace the mutable tag with the desired commit
SHA and include the persist-credentials setting.

Source: Linters/SAST tools

.github/workflows/deploy-runner.yml (1)

311-326: ⚡ Quick win

External IP resolution dependency.

The workflow uses https://api.ipify.org to resolve the GitHub Actions runner's egress IP for the temporary security group rule. If this third-party service is unavailable or rate-limits the request, the deployment will fail.

Consider adding a fallback or using AWS's own service:

MY_IP=$(curl -fsSL --max-time 10 https://checkip.amazonaws.com)

AWS's checkip.amazonaws.com provides the same functionality and may be more reliable for AWS-centric workflows.

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In @.github/workflows/deploy-runner.yml around lines 311 - 326, The workflow
step that obtains the runner egress IP (the sgopen step that sets MY_IP and then
calls aws ec2 authorize-security-group-ingress) relies on api.ipify.org; replace
or add a fallback to use AWS's checkip service to avoid third‑party outages by
attempting MY_IP from https://checkip.amazonaws.com first and falling back to
https://api.ipify.org if empty, then proceed to validate MY_IP and continue
using the existing authorize-security-group-ingress call and RULE_ID checks.

.github/workflows/deploy-app-services.yml (1)

145-179: ⚡ Quick win

Review secrets inheritance for reusable workflows.

All three deploy jobs use secrets: inherit, which passes all repository and organization secrets to the called workflow. Static analysis flags this as a potential over-sharing risk.

If _deploy-single-service.yml only requires specific secrets (e.g., AWS credentials, GitHub tokens), consider explicit secret passing:

secrets:
  AWS_ROLE_ARN: ${{ secrets.AWS_ROLE_ARN }}
  # ... other required secrets

This follows the principle of least privilege and makes the secret dependencies explicit.

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In @.github/workflows/deploy-app-services.yml around lines 145 - 179, Replace
the broad "secrets: inherit" usage in the deploy jobs and instead pass only the
specific secrets required by the reusable workflow (_deploy-single-service.yml);
locate the deploy job blocks that set service_name: Proxy, service_name:
OtelCollector, and service_name: SshGateway and change each secrets: inherit to
an explicit mapping (e.g., map AWS_ROLE_ARN, AWS_WEB_IDENTITY_TOKEN_FILE,
AWS_ROLE_SESSION_NAME, and the minimum GitHub token or other required secrets)
so the reusable workflow receives only those named secrets.

.github/workflows/e2e-cloud.yml (2)

98-98: 💤 Low value

Consider extracting the long boolean expression for readability.

The any output computes a 7-way OR of change flags in a single inline expression. While functionally correct, it's difficult to scan visually.

Alternative: compute in a shell step

       app_services_list: ${{ steps.appsvc.outputs.list }}
-      any: ${{ steps.filter.outputs.api == 'true' || steps.filter.outputs.runner_chain == 'true' || steps.filter.outputs.sdk_py == 'true' || steps.filter.outputs.tests_or_workflow == 'true' || steps.filter.outputs.proxy == 'true' || steps.filter.outputs.otel_collector == 'true' || steps.filter.outputs.ssh_gateway == 'true' }}
+      any: ${{ steps.compute_any.outputs.result }}
     steps:
       - uses: actions/checkout@v5
       - uses: dorny/paths-filter@v3
         id: filter
         ...
+      - id: compute_any
+        run: |
+          if [ '${{ steps.filter.outputs.api }}' = 'true' ] || \
+             [ '${{ steps.filter.outputs.runner_chain }}' = 'true' ] || \
+             [ '${{ steps.filter.outputs.sdk_py }}' = 'true' ] || \
+             [ '${{ steps.filter.outputs.tests_or_workflow }}' = 'true' ] || \
+             [ '${{ steps.filter.outputs.proxy }}' = 'true' ] || \
+             [ '${{ steps.filter.outputs.otel_collector }}' = 'true' ] || \
+             [ '${{ steps.filter.outputs.ssh_gateway }}' = 'true' ]; then
+            echo "result=true" >> "$GITHUB_OUTPUT"
+          else
+            echo "result=false" >> "$GITHUB_OUTPUT"
+          fi

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In @.github/workflows/e2e-cloud.yml at line 98, Extract the long inline OR into
a dedicated step that computes the combined flag and exposes it as an output
(e.g., add a step with id compute_any that runs a short shell snippet to OR the
individual flags from steps.filter.outputs and then uses echo "::set-output
name=any::${result}" or the newer GITHUB_OUTPUT mechanism); replace the current
inline expression for any with a reference to steps.compute_any.outputs.any so
the workflow reads clearly and the seven-way OR is maintained but moved out of
the inline YAML expression.

---

156-158: ⚖️ Poor tradeoff

Template expansion in shell conditions is safe here, but consider refactoring to satisfy linters.

Static analysis flagged these lines for potential code injection via template expansion. In this case, the values come from dorny/paths-filter outputs which are strictly controlled 'true' or 'false' strings, and they're quoted in the shell conditions, so no actual injection risk exists. However, the pattern triggers security linters.

Alternative pattern to satisfy linters

       - id: appsvc
         run: |
           set -euo pipefail
           LIST=()
-          [ '${{ steps.filter.outputs.proxy }}' = 'true' ]          && LIST+=(Proxy)
-          [ '${{ steps.filter.outputs.otel_collector }}' = 'true' ] && LIST+=(OtelCollector)
-          [ '${{ steps.filter.outputs.ssh_gateway }}' = 'true' ]    && LIST+=(SshGateway)
+        env:
+          PROXY: ${{ steps.filter.outputs.proxy }}
+          OTEL: ${{ steps.filter.outputs.otel_collector }}
+          SSHGW: ${{ steps.filter.outputs.ssh_gateway }}
+        run: |
+          set -euo pipefail
+          LIST=()
+          [ "${PROXY}" = 'true' ]  && LIST+=(Proxy)
+          [ "${OTEL}" = 'true' ]   && LIST+=(OtelCollector)
+          [ "${SSHGW}" = 'true' ]  && LIST+=(SshGateway)
           IFS=','
           echo "list=${LIST[*]:-}" >> "$GITHUB_OUTPUT"

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In @.github/workflows/e2e-cloud.yml around lines 156 - 158, Refactor the
conditional checks to first assign the filter outputs to safe shell variables
and then test those variables, e.g. capture
steps.filter.outputs.proxy/otel_collector/ssh_gateway into local variables
(PROXY, OTEL_COLLECTOR, SSH_GATEWAY) and replace the template-expanded tests
with quoted variable comparisons like [ "$PROXY" = "true" ] && LIST+=(Proxy), [
"$OTEL_COLLECTOR" = "true" ] && LIST+=(OtelCollector), [ "$SSH_GATEWAY" = "true"
] && LIST+=(SshGateway); this preserves behavior but avoids direct template
expansion inside the test expressions and satisfies linters.

Source: Linters/SAST tools

apps/api/Dockerfile.source (1)

81-86: 💤 Low value

Consider using ARG for rebuild triggers instead of timestamp comments.

The timestamp comments serve as cache-busting triggers, but Docker's ARG with --build-arg provides a cleaner mechanism that's visible in build output and doesn't clutter the source file.

Alternative pattern

 ENTRYPOINT ["yarn", "ts-node", "--project", "api/tsconfig.app.json", "--transpile-only", "-r", "tsconfig-paths/register", "api/src/main.ts"]

-# 2026-06-11T15:34:47Z — trigger full Api rebuild + deploy after IAM resources recreated
-# 2026-06-11T15:48:35Z — full rebuild after main merge
-
-# 2026-06-11T16:25:09Z — force fresh api build (post-#732, VolumeManager uses .get not .getOrThrow)
-
-# 2026-06-11T16:50:17Z — post-DB-reset rebuild (baseline migration will run on first boot)
+ARG REBUILD_TRIGGER=none
+LABEL rebuild.trigger="${REBUILD_TRIGGER}"

Then in CI: docker build --build-arg REBUILD_TRIGGER="$(date -u +%Y%m%dT%H%M%S)" ...

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@apps/api/Dockerfile.source` around lines 81 - 86, Replace the manual
timestamp comment cache-busters with a Docker build ARG: add an ARG named
REBUILD_TRIGGER (with an optional default) near the top of the Dockerfile and
reference it in a no-op layer (e.g., an ENV or a harmless RUN that consumes
REBUILD_TRIGGER) so CI can pass --build-arg REBUILD_TRIGGER="$(date -u
+%Y%m%dT%H%M%S)" to force rebuilds; remove the timestamp comment blocks from the
file and ensure the new ARG/ENV usage is clearly placed so build output shows
the trigger value.

🤖 Prompt for all review comments with AI agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Inline comments:
In @.github/workflows/deploy-app-services.yml:
- Around line 73-86: The workflow pins third-party Actions using floating tags
(actions/checkout@v5 and dorny/paths-filter@v3); replace those uses with the
corresponding full commit SHAs to reduce supply-chain risk by updating the uses
entries for actions/checkout and dorny/paths-filter in the deploy-app-services
workflow (the steps identified by uses: actions/checkout@v5 and id: filter /
uses: dorny/paths-filter@v3) to their full commit SHA references, keeping the
rest of the step configuration (id: filter, with: filters, etc.) unchanged.

In @.github/workflows/e2e-cloud-test.yml:
- Around line 26-35: Remove the temporary GitHub Actions push trigger block
before merging to main: delete the entire "push:" section (including the
branches entry referencing "chore/e2e-required-merge-gate" and the paths list
that includes ".github/workflows/e2e-cloud-test.yml") so only the
workflow_dispatch or intended triggers remain; ensure the “drop before landing”
comment is preserved or noted in the PR to confirm the removal.

In @.github/workflows/e2e-cloud.yml:
- Line 100: Update the checkout step using actions/checkout@v5 to explicitly set
persist-credentials: false so the GitHub token is not written to .git/config;
locate the checkout invocation (uses: actions/checkout@v5) in the workflow and
add the persist-credentials: false input under that step to prevent credential
leakage to subsequent steps.

In `@apps/api-client-go/api/openapi.yaml`:
- Around line 10004-10007: The CreateBox OpenAPI schema is missing the image
property that the Box schema requires; update the CreateBox schema in
openapi.yaml to include the same image property (description: "The OCI image ref
the box boots from", type: string, include the example value shown) and add
"image" to CreateBox's required array if CreateBox should enforce it, ensuring
the CreateBox schema matches the Box contract so generated Go clients can send
the image field; reference the CreateBox schema and the Box schema definitions
around the existing image declaration to copy the exact metadata.

In `@apps/api/src/box/dto/create-box.dto.ts`:
- Around line 22-24: The image property in CreateBox DTO is missing the Swagger
decorator so it won't appear in OpenAPI; add the `@ApiPropertyOptional`()
decorator above the image field (alongside existing `@IsOptional`() and
`@IsString`()) to match other optional fields and ensure the image property is
included in generated API docs.

In `@apps/api/src/box/services/box.service.ts`:
- Around line 133-134: The warm-pool creation currently forces the fallback
image via assertSupportedImage(undefined); change createForWarmPool (and any
code path that sets box.image for warm pools) to use the WarmPool row's
savedImage instead of calling assertSupportedImage with undefined (i.e., set
box.image = warmPoolRow.savedImage or pass savedImage into assertSupportedImage
if validation is needed). Update the corresponding test in
box.service.warm-pool.spec.ts to create a WarmPool with a non-default savedImage
and assert the created Box.image equals that savedImage (not the curated
fallback digest) to verify the real saved image is persisted.
- Around line 608-630: The fallback by name in getRunnerId and getRegionId can
return a box from another organization because the findOne where clause uses {
name: boxIdOrName } without scoping; update both methods (getRunnerId and
getRegionId) to either (A) accept an organizationId parameter and include {
organizationId, name: boxIdOrName } in the findOne where conditions when the
input is not a UUID, or (B) disallow name lookups and require a public box id
only—implement the chosen approach consistently in both methods and in any
callers so the boxRepository.findOne queries are always scoped by organizationId
when matching by name.

In `@apps/dashboard/src/components/Playground/Box/CodeSnippets/python.ts`:
- Around line 64-66: Extract the repeated OCI image string into a shared
exported constant named DEFAULT_BOX_IMAGE (e.g., export const DEFAULT_BOX_IMAGE
=
'ghcr.io/boxlite-ai/boxlite-agent-base@sha256:834dcb65465985fc2f648451d76c81d166bc7672391c9064a0a115ce6306c85f')
in a common module (suggested module name: cloudBox.ts), then import and use
DEFAULT_BOX_IMAGE wherever the literal appears instead of the hardcoded string;
update the usage in the conditional that builds the image line (the expression
using p.config.createBoxFromImage in python.ts) and in
onboarding-code-examples.ts for all language examples (Python, Go, Rust) to
reference DEFAULT_BOX_IMAGE so the single constant replaces all five
occurrences.

In `@apps/dashboard/src/hooks/useBoxSession.ts`:
- Around line 86-99: The code currently applies DEFAULT_CREATE_TIMEOUT_SECONDS
twice (once to boxApi.createBox and again to waitUntilStarted) causing a
possible double timeout; change to a single end-to-end timeout budget by
computing a deadline before the HTTP call (e.g. const deadline = Date.now() +
DEFAULT_CREATE_TIMEOUT_SECONDS * 1000), pass the remaining milliseconds to
boxApi.createBox as its timeout, then compute remaining seconds and pass that to
waitUntilStarted (or fail early if no time remains). Update references in
createMutation (boxApi.createBox and waitUntilStarted) so they use the shared
deadline/remaining-time rather than each using DEFAULT_CREATE_TIMEOUT_SECONDS
independently.

In `@apps/dashboard/src/lib/cloudBox.ts`:
- Around line 161-162: In toCreateBoxRequest, the default language
(CodeLanguage.PYTHON) is only applied when params is null/undefined due to the
nullish coalescing; when callers pass an empty object ({}) the default is
skipped. Fix by treating params as an optional partial and merging defaults with
provided fields (e.g., base default { language: CodeLanguage.PYTHON } merged
with params/CreateBoxParams) so the language default is applied unless
explicitly overridden; update any usage of resolvedParams in toCreateBoxRequest
accordingly.
- Around line 213-247: waitUntilStarted can exceed the provided timeout because
it only checks elapsed time after awaiting api.boxApi.getBox; update
waitUntilStarted to compute remainingTime before each poll and pass a
request-level timeout (e.g., via AbortController or axios timeout) to
api.boxApi.getBox (or bail immediately if remainingTime <= 0) so in-flight
requests cannot push you past timeoutSeconds; also ensure semantics for
timeoutSeconds === 0 remain documented/unchanged (0 = infinite) and use symbols
currentBox, startTime, timeoutSeconds, POLL_INTERVAL_MS, and api.boxApi.getBox
to locate and implement the change.

In `@apps/dashboard/src/providers/PlaygroundProvider.tsx`:
- Around line 300-303: The same hardcoded image digest is duplicated between
createBoxFromImageParams in PlaygroundProvider and SUPPORTED_IMAGES in
CreateBoxSheet; extract the default/supported image list into a single exported
constant (e.g., SUPPORTED_IMAGES or DEFAULT_BOX_IMAGES) in a new module
SupportedImages.ts and import that constant in both PlaygroundProvider.tsx and
CreateBoxSheet.tsx, replacing the literal strings so both
createBoxFromImageParams and the SUPPORTED_IMAGES reference the shared constant;
ensure the new module exports a typed array (matching CreateBoxFromImageParams
image shape) and update imports in the two components accordingly.

In `@scripts/test/e2e/cases/conftest.py`:
- Around line 33-47: _wrap the module-level call in _default_image so it never
raises: catch exceptions around reading CRED_PATH.read_text() and
tomllib.loads() inside the _default_image function (e.g., FileNotFoundError,
OSError, tomllib.TOMLDecodeError, and a generic Exception fallback), log or
silently ignore the error, and return the hardcoded fallback image; ensure
DEFAULT_IMAGE is still assigned from _default_image() so importing the module
cannot fail due to malformed/missing credentials or TOML parsing errors._

---

Nitpick comments:
In @.github/workflows/_deploy-single-service.yml:
- Around line 60-64: Replace the floating tag
"aws-actions/configure-aws-credentials@v4" with a specific commit SHA to pin the
action for supply-chain security; in the workflow step that uses
aws-actions/configure-aws-credentials (the step with inputs role-to-assume,
aws-region, role-session-name), update the action reference to the exact commit
SHA (e.g., aws-actions/configure-aws-credentials@<COMMIT_SHA>) and commit the
change so the workflow uses the pinned revision.
- Around line 79-81: The GitHub Actions step "ECR login" (id: ecr_login)
currently uses the floating reference "uses: aws-actions/amazon-ecr-login@v2";
replace this with a pinned commit SHA (e.g. "uses:
aws-actions/amazon-ecr-login@<commit-sha>") to lock the action to a specific
immutable revision—update the uses line for the aws-actions/amazon-ecr-login
action to reference the chosen commit SHA.
- Around line 98-125: The current job block (resolving API role ARNs via API_TD,
TASK_ROLE and EXEC_ROLE) temporarily borrows Api roles which violates
least-privilege; change the step to avoid automatic borrowing by first
attempting to recreate per-service roles (run `sst deploy --stage e2e-ci` or
call an administrative deploy helper) and fail fast if that fails, or require an
explicit workflow input/approval flag to opt into borrowing; implement this by
replacing the logic that sets API_TD / TASK_ROLE / EXEC_ROLE with a pre-check
that runs the deploy (or checks an explicit input like
inputs.allow_role_borrowing) and only falls back to borrowing if that input is
set, and ensure the step emits a clear error and exits non-zero when neither
recreation nor explicit approval is present.
- Line 57: Update the GitHub Actions checkout step that currently uses
"actions/checkout@v5": pin the "uses" to a specific commit SHA for
actions/checkout and add the input "persist-credentials: false" to the same step
to prevent GITHUB_TOKEN from being stored in .git/config; locate the checkout
step in the workflow where "uses: actions/checkout@v5" appears and replace the
mutable tag with the desired commit SHA and include the persist-credentials
setting.

In @.github/workflows/build-runner-binary.yml:
- Around line 107-110: Replace the floating version for the GitHub Action usage
"uses: actions/download-artifact@v4" with a pinned commit SHA (e.g. "uses:
actions/download-artifact@<commit-sha>") to ensure supply-chain security; update
the workflow entry that currently includes the keys "name: c-sdk-linux-x64-gnu"
and "path: /tmp/c-sdk/" to reference the pinned commit instead of the tag,
commit the change, and verify the workflow still downloads the artifact as
before.
- Around line 112-125: The script currently uses HEAD selection for ARCHIVE and
A which can silently pick the first match; update the block that sets ARCHIVE
(variable ARCHIVE=$(ls /tmp/c-sdk/boxlite-c-v*-linux-x64-gnu.tar.gz | head -1))
and the find for A (A=$(find /tmp/c-sdk -name libboxlite.a | head -1)) to assert
uniqueness: list matches into an array or count results and if count == 0 emit
an error, if count > 1 emit an error asking to disambiguate; only proceed to tar
xzf and cp when exactly one archive and exactly one libboxlite.a are found, and
include descriptive error messages referencing ARCHIVE and A when failing.

In @.github/workflows/deploy-api.yml:
- Line 97: Update the workflow step that uses actions/checkout@v5 to pin to a
specific commit SHA and add persist-credentials: false; specifically replace the
"uses: actions/checkout@v5" entry with "uses: actions/checkout@<COMMIT_SHA>"
(use the stable commit SHA from the actions/checkout repository) and add the
input "with: persist-credentials: false" under that step so the checkout step
stops persisting GitHub credentials.
- Around line 168-172: Replace the floating tag for the GitHub Action in the
workflow step that currently references aws-actions/configure-aws-credentials@v4
with a pinned commit SHA; update the step using symbol "uses:
aws-actions/configure-aws-credentials@v4" to use a specific commit hash (e.g.
aws-actions/configure-aws-credentials@<COMMIT_SHA>) so the "uses" entry is
pinned to an immutable release, keeping the rest of the inputs (role-to-assume,
aws-region, role-session-name) unchanged.
- Around line 100-108: Replace the floating tag for the GitHub Action with a
specific commit SHA to pin the dependency: locate the step that currently
declares "uses: dorny/paths-filter@v3" (the step that provides the api_source
filters) and change it to "uses: dorny/paths-filter@<commit-sha>" using a
specific commit SHA from the dorny/paths-filter repository; ensure you pick a
stable commit SHA (not a branch or tag) and update the workflow to use that SHA
so the paths-filter behavior is pinned for supply-chain security.
- Line 225: The workflow currently references the ECR login action by tag
("uses: aws-actions/amazon-ecr-login@v2"); replace that tag with the action's
immutable commit SHA (e.g. "uses: aws-actions/amazon-ecr-login@<commit-sha>") to
pin the dependency for supply-chain security, locating the usage string
"aws-actions/amazon-ecr-login@v2" in the deploy workflow and updating it to the
specific commit SHA from the action's GitHub repository, then commit the change
and verify the workflow runs against the pinned SHA.
- Line 165: Replace the floating version "uses: actions/checkout@v5" with a
pinned commit SHA for the actions/checkout action (e.g., "uses:
actions/checkout@<commit-sha>") and add a with block setting
persist-credentials: false to disable credential persistence; update the
workflow line that references actions/checkout and add the with:
persist-credentials: false setting so the checkout action is both pinned and
does not persist credentials.

In @.github/workflows/deploy-app-services.yml:
- Around line 145-179: Replace the broad "secrets: inherit" usage in the deploy
jobs and instead pass only the specific secrets required by the reusable
workflow (_deploy-single-service.yml); locate the deploy job blocks that set
service_name: Proxy, service_name: OtelCollector, and service_name: SshGateway
and change each secrets: inherit to an explicit mapping (e.g., map AWS_ROLE_ARN,
AWS_WEB_IDENTITY_TOKEN_FILE, AWS_ROLE_SESSION_NAME, and the minimum GitHub token
or other required secrets) so the reusable workflow receives only those named
secrets.

In @.github/workflows/deploy-runner.yml:
- Around line 311-326: The workflow step that obtains the runner egress IP (the
sgopen step that sets MY_IP and then calls aws ec2
authorize-security-group-ingress) relies on api.ipify.org; replace or add a
fallback to use AWS's checkip service to avoid third‑party outages by attempting
MY_IP from https://checkip.amazonaws.com first and falling back to
https://api.ipify.org if empty, then proceed to validate MY_IP and continue
using the existing authorize-security-group-ingress call and RULE_ID checks.

In @.github/workflows/e2e-cloud.yml:
- Line 98: Extract the long inline OR into a dedicated step that computes the
combined flag and exposes it as an output (e.g., add a step with id compute_any
that runs a short shell snippet to OR the individual flags from
steps.filter.outputs and then uses echo "::set-output name=any::${result}" or
the newer GITHUB_OUTPUT mechanism); replace the current inline expression for
any with a reference to steps.compute_any.outputs.any so the workflow reads
clearly and the seven-way OR is maintained but moved out of the inline YAML
expression.
- Around line 156-158: Refactor the conditional checks to first assign the
filter outputs to safe shell variables and then test those variables, e.g.
capture steps.filter.outputs.proxy/otel_collector/ssh_gateway into local
variables (PROXY, OTEL_COLLECTOR, SSH_GATEWAY) and replace the template-expanded
tests with quoted variable comparisons like [ "$PROXY" = "true" ] &&
LIST+=(Proxy), [ "$OTEL_COLLECTOR" = "true" ] && LIST+=(OtelCollector), [
"$SSH_GATEWAY" = "true" ] && LIST+=(SshGateway); this preserves behavior but
avoids direct template expansion inside the test expressions and satisfies
linters.

In `@apps/api/Dockerfile.source`:
- Around line 81-86: Replace the manual timestamp comment cache-busters with a
Docker build ARG: add an ARG named REBUILD_TRIGGER (with an optional default)
near the top of the Dockerfile and reference it in a no-op layer (e.g., an ENV
or a harmless RUN that consumes REBUILD_TRIGGER) so CI can pass --build-arg
REBUILD_TRIGGER="$(date -u +%Y%m%dT%H%M%S)" to force rebuilds; remove the
timestamp comment blocks from the file and ensure the new ARG/ENV usage is
clearly placed so build output shows the trigger value.

In `@apps/api/src/box/runner-adapter/runnerAdapter.v2.createBox.spec.ts`:
- Around line 22-84: Add tests in RunnerAdapterV2.createBox spec to cover GPU
quota and userId mapping: create a new it(...) that calls adapter.createBox with
a Box built via buildBox() but set box.gpu = 1 and box.userId = 'some-user-id',
then read the payload from createJob.mock.calls[0][5] and assert
payload.gpuQuota === 1 (and/or that 'gpuQuota' exists) and assert payload.userId
equals the expected source (if your implementation maps box.userId use that
value, otherwise assert it equals box.organizationId); reference the createBox
call and the createJob mock used in the existing tests to locate where to add
these assertions.

In `@apps/dashboard/src/components/Playground/Box/CodeSnippets/index.ts`:
- Around line 10-12: The current declaration for codeSnippetGenerators uses
Record<CodeLanguage.PYTHON, CodeSnippetGenerator>, which hardcodes a single key
and is not extensible; create a union type alias (e.g., SupportedLanguages) that
includes CodeLanguage.PYTHON (and future languages), then change the type
annotation on codeSnippetGenerators to Record<SupportedLanguages,
CodeSnippetGenerator> and keep the existing initializer mapping
[CodeLanguage.PYTHON]: PythonSnippetGenerator so new languages can be added by
extending SupportedLanguages and adding entries to the object.

In `@apps/dashboard/src/components/Playground/Box/CodeSnippets/python.ts`:
- Around line 17-20: The import string generation currently produces a redundant
self-alias when p.config.createBoxFromTemplate is false; update the ternary that
builds the import entry (the expression using p.config.createBoxFromTemplate) so
the false branch returns 'CreateBoxFromImageParams' (no "as
CreateBoxFromImageParams") instead of 'CreateBoxFromImageParams as
CreateBoxFromImageParams', keeping the true branch ('CreateBoxFromTemplateParams
as CreateBoxFromImageParams') unchanged.

In `@apps/dashboard/src/components/Playground/VNC/Interaction/Display.tsx`:
- Around line 38-42: Update the outdated comment above displayActionAPICall to
reflect that we now cast ComputerUseClient to CloudBoxComputerUse (not
ComputerUse) to satisfy the TypeScript compiler; mention the real reason for the
cast (ensuring ComputerUseClient exists and typing as CloudBoxComputerUse) and
reference the symbols displayActionAPICall, ComputerUseClient,
CloudBoxComputerUse, and DisplayActions so future readers understand why the
cast is used.

In `@apps/dashboard/src/components/Playground/VNC/Interaction/Keyboard.tsx`:
- Around line 81-83: The comment above keyboardActionAPICall is outdated: it
mentions casting to ComputerUse but the code actually casts ComputerUseClient to
CloudBoxComputerUse; update the comment to accurately state that the disable
logic ensures this method is called only when ComputerUseClient exists and that
we cast to CloudBoxComputerUse (not ComputerUse) to satisfy the TypeScript
compiler before accessing the keyboard client (KeyboardActionsClient).

In `@apps/dashboard/src/components/Playground/VNC/Interaction/Mouse.tsx`:
- Around line 165-167: The comment above mouseActionAPICall is outdated
(mentions ComputerUse) — update it to accurately reflect the current cast to
CloudBoxComputerUse; modify the comment that precedes the line "const
MouseActionsClient = (ComputerUseClient as CloudBoxComputerUse).mouse" so it
explains that the disable logic ensures ComputerUseClient exists and that we
cast to CloudBoxComputerUse to satisfy TypeScript when accessing .mouse.

In `@apps/dashboard/src/components/Playground/VNC/Interaction/Screenshot.tsx`:
- Around line 127-129: The comment above the screenshotActionAPICall is
outdated—update it to reflect the current cast to CloudBoxComputerUse instead of
ComputerUse; edit the comment that references ComputerUse/disable logic so it
accurately states that the method is called when ComputerUseClient exists and
that we cast to CloudBoxComputerUse to satisfy TypeScript for the
ScreenshotActionsClient assignment (referencing ComputerUseClient,
CloudBoxComputerUse, and screenshotActionAPICall).

In `@apps/dashboard/src/lib/cloudBox.ts`:
- Around line 404-423: The ts-node invocation in getRunCommand for
CodeLanguage.TYPESCRIPT currently includes the --ignore-diagnostics 5107 flag
which hides "Cannot find module" errors; update the CodeLanguage.TYPESCRIPT
branch inside getRunCommand to either remove the --ignore-diagnostics 5107 flag
so module resolution errors surface, or if suppression is intentional for the
sandbox/playground, add a clear inline comment next to the ts-node command
explaining why diagnostic 5107 is suppressed and the safety/constraints that
justify it (reference getRunCommand, CodeLanguage.TYPESCRIPT, and the
--ignore-diagnostics 5107 token to locate the change).

🪄 Autofix (Beta)

Fix all unresolved CodeRabbit comments on this PR:

• Push a commit to this branch (recommended)
• Create a new PR with the fixes

[coderabbitai]

⚠️ Potential issue | 🟠 Major

🧩 Analysis chain

🏁 Script executed:

#!/bin/bash
# Search for security or action pinning policy documents
rg -l -i 'action.*pin|supply.*chain|security.*policy' --type=md

Repository: boxlite-ai/boxlite

Length of output: 1954

---

Pin GitHub Actions to commit SHAs in deploy-app-services

The workflow uses floating tags (actions/checkout@v5, dorny/paths-filter@v3), which can change over time and increases supply-chain risk; pin these actions to full commit SHAs.

- uses: actions/checkout@<full-sha>
- uses: dorny/paths-filter@<full-sha>

🧰 Tools

🪛 zizmor (1.25.2)

[warning] 73-73: credential persistence through GitHub Actions artifacts (artipacked): does not set persist-credentials: false

(artipacked)

---

[error] 73-73: unpinned action reference (unpinned-uses): action is not pinned to a hash (required by blanket policy)

(unpinned-uses)

---

[error] 76-76: unpinned action reference (unpinned-uses): action is not pinned to a hash (required by blanket policy)

(unpinned-uses)

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In @.github/workflows/deploy-app-services.yml around lines 73 - 86, The workflow
pins third-party Actions using floating tags (actions/checkout@v5 and
dorny/paths-filter@v3); replace those uses with the corresponding full commit
SHAs to reduce supply-chain risk by updating the uses entries for
actions/checkout and dorny/paths-filter in the deploy-app-services workflow (the
steps identified by uses: actions/checkout@v5 and id: filter / uses:
dorny/paths-filter@v3) to their full commit SHA references, keeping the rest of
the step configuration (id: filter, with: filters, etc.) unchanged.

[coderabbitai]

⚠️ Potential issue | 🟡 Minor

🧩 Analysis chain

🏁 Script executed:

#!/bin/bash
# Check if this workflow exists on main without the push trigger
git show main:.github/workflows/e2e-cloud-test.yml 2>/dev/null | grep -A5 '^on:' | grep -q 'push:' && echo "WARNING: push trigger still present on main" || echo "OK: no push trigger on main"

Repository: boxlite-ai/boxlite

Length of output: 1954

---

🌐 Web query:

Find the contents of https://raw.githubusercontent.com/boxlite-ai/boxlite/main/.github/workflows/e2e-cloud-test.yml and whether it contains a push:trigger underon:``

💡 Result:

Based on my search of the boxlite-ai/boxlite repository as of June 2026, the file .github/workflows/e2e-cloud-test.yml does not exist in the main branch [1]. The repository contains a standard set of GitHub Actions workflows for continuous integration and delivery (CI/CD) [2], but exhaustive searches of the current repository structure confirm that e2e-cloud-test.yml is not a valid path within that project [1][3]. Therefore, it does not contain a push: trigger or any other configuration.

Citations:

• 1: https://github.com/boxlite-ai/boxlite/tree/refs/heads/main
• 2: https://github.com/boxlite-ai/boxlite/actions
• 3: https://github.com/boxlite-ai/boxlite/blob/main/.pre-commit-config.yaml

---

🏁 Script executed:

#!/bin/bash
# Retry: check whether main has a `push:` trigger in this workflow (no output beyond the result)
set -euo pipefail
git show main:.github/workflows/e2e-cloud-test.yml 2>/dev/null | awk '
  /^on:/ {in_on=1}
  in_on && $0 ~ /^[[:space:]]*push:/ {found=1}
  END { if (found) print "WARNING: push trigger still present on main"; else print "OK: no push trigger on main" }
'

Repository: boxlite-ai/boxlite

Length of output: 1924

---

Ensure the temporary push trigger is removed when landing this workflow to main.

This workflow file doesn’t exist on main yet, so the temporary push trigger isn’t currently present there; keep the “drop before landing” requirement and remove the push: block as part of the merge.

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In @.github/workflows/e2e-cloud-test.yml around lines 26 - 35, Remove the
temporary GitHub Actions push trigger block before merging to main: delete the
entire "push:" section (including the branches entry referencing
"chore/e2e-required-merge-gate" and the paths list that includes
".github/workflows/e2e-cloud-test.yml") so only the workflow_dispatch or
intended triggers remain; ensure the “drop before landing” comment is preserved
or noted in the PR to confirm the removal.

[coderabbitai]

⚠️ Potential issue | 🟠 Major | ⚡ Quick win

Set persist-credentials: false to prevent credential leakage.

The checkout action defaults to persisting GitHub token credentials in the workspace .git/config, which remain accessible to subsequent workflow steps. If an attacker compromises a later step or a dependency, they could exfiltrate the token.

🛡️ Recommended fix

       - uses: actions/checkout@v5
+        with:
+          persist-credentials: false

📝 Committable suggestion

‼️ IMPORTANT
Carefully review the code before committing. Ensure that it accurately replaces the highlighted code, contains no missing lines, and has no issues with indentation. Thoroughly test & benchmark the code to ensure it meets the requirements.

Suggested change

	- uses: actions/checkout@v5
	- uses: actions/checkout@v5
	with:
	persist-credentials: false

🧰 Tools

🪛 zizmor (1.25.2)

[warning] 100-100: credential persistence through GitHub Actions artifacts (artipacked): does not set persist-credentials: false

(artipacked)

---

[error] 100-100: unpinned action reference (unpinned-uses): action is not pinned to a hash (required by blanket policy)

(unpinned-uses)

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In @.github/workflows/e2e-cloud.yml at line 100, Update the checkout step using
actions/checkout@v5 to explicitly set persist-credentials: false so the GitHub
token is not written to .git/config; locate the checkout invocation (uses:
actions/checkout@v5) in the workflow and add the persist-credentials: false
input under that step to prevent credential leakage to subsequent steps.

Source: Linters/SAST tools

[coderabbitai]

⚠️ Potential issue | 🟠 Major | 🏗️ Heavy lift

Add image to CreateBox too.

Lines 10004-10116 make image a required part of every Box, but CreateBox in this same spec still has no image field. That leaves generated Go clients unable to send the image selector the restored create flow now depends on, so the request/response contract is inconsistent.

Also applies to: 10116-10116

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@apps/api-client-go/api/openapi.yaml` around lines 10004 - 10007, The
CreateBox OpenAPI schema is missing the image property that the Box schema
requires; update the CreateBox schema in openapi.yaml to include the same image
property (description: "The OCI image ref the box boots from", type: string,
include the example value shown) and add "image" to CreateBox's required array
if CreateBox should enforce it, ensuring the CreateBox schema matches the Box
contract so generated Go clients can send the image field; reference the
CreateBox schema and the Box schema definitions around the existing image
declaration to copy the exact metadata.

[coderabbitai]

⚠️ Potential issue | 🟡 Minor | ⚡ Quick win

Add missing @ApiPropertyOptional() decorator for API documentation.

The image field is missing the @ApiPropertyOptional() decorator. Every other optional field in this file includes both class-validator decorators (@IsOptional(), @IsString()) and the Swagger decorator (@ApiPropertyOptional()). Without it, the image field won't appear in the generated OpenAPI specification, making the API contract incomplete for client developers.

📝 Proposed fix

+ `@ApiPropertyOptional`({
+   description: 'The OCI image ref the box boots from',
+   example: 'ghcr.io/boxlite-ai/boxlite-agent-python@sha256:80d562a57f4bc12def4e54dbdb9e7d26d3268fe0767a2955ab5ad718041145d6',
+ })
  `@IsOptional`()
  `@IsString`()
  image?: string

📝 Committable suggestion

‼️ IMPORTANT
Carefully review the code before committing. Ensure that it accurately replaces the highlighted code, contains no missing lines, and has no issues with indentation. Thoroughly test & benchmark the code to ensure it meets the requirements.

Suggested change

	@IsOptional()
	@IsString()
	image?: string
	`@ApiPropertyOptional`({
	description: 'The OCI image ref the box boots from',
	example: 'ghcr.io/boxlite-ai/boxlite-agent-python@sha256:80d562a57f4bc12def4e54dbdb9e7d26d3268fe0767a2955ab5ad718041145d6',
	})
	`@IsOptional`()
	`@IsString`()
	image?: string

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@apps/api/src/box/dto/create-box.dto.ts` around lines 22 - 24, The image
property in CreateBox DTO is missing the Swagger decorator so it won't appear in
OpenAPI; add the `@ApiPropertyOptional`() decorator above the image field
(alongside existing `@IsOptional`() and `@IsString`()) to match other optional
fields and ensure the image property is included in generated API docs.

[coderabbitai]

⚠️ Potential issue | 🟡 Minor | ⚡ Quick win

Cumulative timeout may reach 120 seconds (HTTP + polling), exceeding the 60-second constant name.

The createMutation sets the HTTP request timeout to DEFAULT_CREATE_TIMEOUT_SECONDS * 1000 (60s, line 91), then calls waitUntilStarted with another DEFAULT_CREATE_TIMEOUT_SECONDS (60s, line 97). If the createBox HTTP call takes close to 60s and then polling takes another 60s, the total elapsed time approaches 120s. The constant name DEFAULT_CREATE_TIMEOUT_SECONDS suggests a single unified timeout, but in practice it's applied twice. Consider whether you want a single end-to-end timeout budget or if the current behavior (separate HTTP + polling timeouts) is intentional.

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@apps/dashboard/src/hooks/useBoxSession.ts` around lines 86 - 99, The code
currently applies DEFAULT_CREATE_TIMEOUT_SECONDS twice (once to boxApi.createBox
and again to waitUntilStarted) causing a possible double timeout; change to a
single end-to-end timeout budget by computing a deadline before the HTTP call
(e.g. const deadline = Date.now() + DEFAULT_CREATE_TIMEOUT_SECONDS * 1000), pass
the remaining milliseconds to boxApi.createBox as its timeout, then compute
remaining seconds and pass that to waitUntilStarted (or fail early if no time
remains). Update references in createMutation (boxApi.createBox and
waitUntilStarted) so they use the shared deadline/remaining-time rather than
each using DEFAULT_CREATE_TIMEOUT_SECONDS independently.

[coderabbitai]

⚠️ Potential issue | 🟡 Minor | ⚡ Quick win

Default language only applies when params is undefined, not when empty.

When params is explicitly {}, the default language: CodeLanguage.PYTHON won't be applied (line 162 only triggers when params ?? ... resolves the nullish coalescing). Consider whether this is the intended behavior or if you want to apply the default language even when an empty object is passed.

📝 Proposed fix to apply default consistently

 export function toCreateBoxRequest(params?: CreateBoxParams, target?: string): CreateBoxRequest {
-  const resolvedParams = params ?? { language: CodeLanguage.PYTHON }
+  const resolvedParams = params ?? {}
   const labels = { ...(resolvedParams.labels ?? {}) }
 
-  if (resolvedParams.language) {
-    labels[CODE_LANGUAGE_LABEL] = String(resolvedParams.language)
-  }
+  const language = resolvedParams.language ?? CodeLanguage.PYTHON
+  labels[CODE_LANGUAGE_LABEL] = String(language)

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@apps/dashboard/src/lib/cloudBox.ts` around lines 161 - 162, In
toCreateBoxRequest, the default language (CodeLanguage.PYTHON) is only applied
when params is null/undefined due to the nullish coalescing; when callers pass
an empty object ({}) the default is skipped. Fix by treating params as an
optional partial and merging defaults with provided fields (e.g., base default {
language: CodeLanguage.PYTHON } merged with params/CreateBoxParams) so the
language default is applied unless explicitly overridden; update any usage of
resolvedParams in toCreateBoxRequest accordingly.

[coderabbitai]

⚠️ Potential issue | 🟡 Minor

🧩 Analysis chain

🏁 Script executed:

#!/bin/bash
# Description: Find all call sites of waitUntilStarted and check for timeout documentation.

# Find call sites
rg -nP '\bwaitUntilStarted\s*\(' --type=ts --type=tsx

# Check for JSDoc or comments mentioning timeout behavior
rg -nP 'waitUntilStarted.*timeout|timeout.*waitUntilStarted' --type=ts --type=tsx -C3

Repository: boxlite-ai/boxlite

Length of output: 1954

---

Potential timeout overshoot in waitUntilStarted

• waitUntilStarted checks the timeout only after each awaited api.boxApi.getBox(...), so wall-clock time can exceed timeoutSeconds by the duration of in-flight requests.
• Recompute remaining time before each poll (or use a request-level timeout like AbortController) and document the intended semantics for timeoutSeconds === 0.

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@apps/dashboard/src/lib/cloudBox.ts` around lines 213 - 247, waitUntilStarted
can exceed the provided timeout because it only checks elapsed time after
awaiting api.boxApi.getBox; update waitUntilStarted to compute remainingTime
before each poll and pass a request-level timeout (e.g., via AbortController or
axios timeout) to api.boxApi.getBox (or bail immediately if remainingTime <= 0)
so in-flight requests cannot push you past timeoutSeconds; also ensure semantics
for timeoutSeconds === 0 remain documented/unchanged (0 = infinite) and use
symbols currentBox, startTime, timeoutSeconds, POLL_INTERVAL_MS, and
api.boxApi.getBox to locate and implement the change.

[coderabbitai]

🛠️ Refactor suggestion | 🟠 Major | ⚡ Quick win

Duplicated image digest in PlaygroundProvider.tsx and CreateBoxSheet.tsx.

The digest ghcr.io/boxlite-ai/boxlite-agent-base@sha256:834dcb65... is hardcoded in PlaygroundProvider.tsx (line 301-302) and duplicated in CreateBoxSheet.tsx (line 43 as SUPPORTED_IMAGES[0]). Additionally, CreateBoxSheet.tsx defines two more supported images (lines 44-45). If these digests are rotated, all occurrences must be updated together. The root cause is the absence of a shared constant for the default and supported images. Extract these to a single shared constant file (e.g., apps/dashboard/src/constants/SupportedImages.ts) and import it in both PlaygroundProvider.tsx and CreateBoxSheet.tsx to eliminate the duplication and ensure consistency.

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@apps/dashboard/src/providers/PlaygroundProvider.tsx` around lines 300 - 303,
The same hardcoded image digest is duplicated between createBoxFromImageParams
in PlaygroundProvider and SUPPORTED_IMAGES in CreateBoxSheet; extract the
default/supported image list into a single exported constant (e.g.,
SUPPORTED_IMAGES or DEFAULT_BOX_IMAGES) in a new module SupportedImages.ts and
import that constant in both PlaygroundProvider.tsx and CreateBoxSheet.tsx,
replacing the literal strings so both createBoxFromImageParams and the
SUPPORTED_IMAGES reference the shared constant; ensure the new module exports a
typed array (matching CreateBoxFromImageParams image shape) and update imports
in the two components accordingly.

[coderabbitai]

⚠️ Potential issue | 🟠 Major | ⚡ Quick win

Add error handling to prevent module import failure.

_default_image() is called at module level (line 47). If CRED_PATH.read_text() or tomllib.loads() throws an exception (e.g., encoding error, malformed TOML), the entire module will fail to import, breaking all tests in this directory.

🛡️ Proposed fix to add error handling

 def _default_image() -> str:
     """The image must be in the API's supported allowlist, which differs per
     stack (bootstrap.sh points the local stack at public refs; prod pins ghcr
     digests). fixture_setup.py records the stack's base image in the profile;
     the ghcr fallback only covers runs against prod-like stacks without it."""
     if env_image := os.environ.get("BOXLITE_E2E_IMAGE"):
         return env_image
     if CRED_PATH.exists():
-        profile = tomllib.loads(CRED_PATH.read_text()).get("profiles", {}).get(DEFAULT_PROFILE, {})
-        if profile.get("default_image"):
-            return profile["default_image"]
+        try:
+            profile = tomllib.loads(CRED_PATH.read_text()).get("profiles", {}).get(DEFAULT_PROFILE, {})
+            if profile.get("default_image"):
+                return profile["default_image"]
+        except Exception:
+            pass  # Fall through to ghcr.io fallback
     return "ghcr.io/boxlite-ai/boxlite-agent-base@sha256:834dcb65465985fc2f648451d76c81d166bc7672391c9064a0a115ce6306c85f"

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@scripts/test/e2e/cases/conftest.py` around lines 33 - 47, _wrap the
module-level call in _default_image so it never raises: catch exceptions around
reading CRED_PATH.read_text() and tomllib.loads() inside the _default_image
function (e.g., FileNotFoundError, OSError, tomllib.TOMLDecodeError, and a
generic Exception fallback), log or silently ignore the error, and return the
hardcoded fallback image; ensure DEFAULT_IMAGE is still assigned from
_default_image() so importing the module cannot fail due to malformed/missing
credentials or TOML parsing errors._