apx-aws-lambda
AWS Lambda for Action Platform: the aws/lambda deploy target (SAM), its overlay, read-only tools and commands, and an IAM-only connect stack for your AWS account so that no AWS key ever lives on the platform, on a machine or in a repository — and nothing of ours runs in the account. apx- is the prefix every Action Platform extension carries.
action-platform plugin install aws-lambda
action-platform cloud set aws/lambda # overlay files come from this plugin
action-platform deploy --dry-run # preflight: credentials, sam validate
action-platform deploy # sam build + sam deploy --config-env <stage>
action-platform diagnose
Contents
- What the plugin does
- Credentials without keys
- Where the deploy runs
- Tools and commands
- Requirements
- Development
What the plugin does
| Overlay | template.yaml, samconfig.toml, a one-line Makefile (ap-build package), requirements/ (IAM examples), .github/workflows/deploy.yml (optional). One template for every language: the Lambda Web Adapter layer runs the project as an HTTP server on $PORT — run.sh from ap-build (Python python3.12, Node nodejs22.x, Java/Kotlin java21, Ruby ruby3.3) or the static bootstrap on provided.al2023 (Go). Every web/* template of the official repository deploys; the health route is /health (API Gateway keeps /ping for itself on execute-api). Template parameters the target fills: Stage (the scope's name), and for a connected account PermissionsBoundaryArn and RolePath |
| Readiness | before a deploy, without building or changing anything: the tooling is there, the overlay is applied, the stack name resolves, the credentials work, the stack is not mid-operation or failed (a failed first creation — ROLLBACK_COMPLETE or ROLLBACK_FAILED — is a warning: the deploy deletes it), the overlay takes a connected account's boundary and role path, the deploy role may do what the template needs (simulated with iam:SimulatePrincipalPolicy against the stack, the function, its log group, the execution role and every public layer the template references — the Lambda Web Adapter's, for one) and sam validate --lint passes. The platform runs it after every release for each of the app's scopes and refuses a blocked deploy unless forced; action-platform readiness runs it on a machine |
| Deploy | sam build + sam deploy --config-env <stage> (scope prod → prod, any other → default). A stack left by a failed first creation (ROLLBACK_COMPLETE, ROLLBACK_FAILED) is deleted first; when that delete fails on a resource that was never created, the stack is deleted again with it retained. On a connected account, an overlay without PermissionsBoundaryArn and RolePath stops the deploy before the build. Every line sam prints reaches the platform's run log as it appears — the deploy form, the run dialog, action-platform logs -f |
| Rollback | CloudFormation rollback-stack — previous stack state |
| Diagnose | stack status and the HTTP API url |
| Destroy | sam delete of the stage's stack, the execution role SAM created with it |
| Tools | aws_lambda_stacks, aws_lambda_functions |
| Commands | action-platform aws-lambda connect|stacks|functions |
[deploy] in platform.toml:
[deploy]
target = "aws/lambda"
region = "us-east-1" # optional; samconfig.toml / AWS_REGION otherwise
health = "/health" # optional; checked through the HTTP API after the deploy, "" skips it
Deploying itself goes through the core's deploy / rollback / diagnose, which drive the target.
Credentials without keys
The target looks for credentials in this order:
Under [deploy] |
How | Who decides what the app may do |
|---|---|---|
role_arn |
sts assume-role-with-web-identity with a platform token |
the role's trust and permission policies |
neither, and the organization's AWS account ID (or deploy role ARN) set on the platform (Plugins → AWS Lambda), which every deploy job carries as AP_AWS_LAMBDA_ROLE_ARN with AP_APP |
sts assume-role-with-web-identity on the connect stack's role with the platform's token for the app |
the token's session tag and the role's policy, in your account |
| neither | the AWS CLI's own chain: SSO, profile, instance role | whatever that identity may do |
A connected account (recommended)
One CloudFormation stack, IAM only — no function, no table, nothing to upgrade. AWS itself trusts the platform's OIDC tokens; the token's session tag keeps each deploy inside its app.
1. Create the stack — once per AWS account, with an IAM user or role that may create IAM resources (not the root account). On the platform: Plugins → AWS Lambda → Configure → Connect AWS. The button copies a command and opens AWS CloudShell, behind the AWS sign-in; paste the command and press Enter. It downloads this release's connect/template.yaml from GitHub, deploys it as the stack action-platform-connect in us-east-1 with the platform's url and the organization's slug, and prints the account ID.
Without the button, the same stack from a machine logged in to the account:
action-platform aws-lambda connect https://platform.example.com acme # --region, --stack-name, --oidc-provider-arn
or in the console: CloudFormation → Create stack → Upload a template file → apx_aws_lambda/connect/template.yaml, parameters IssuerUrl (the platform's public url) and Organization (its slug), acknowledge IAM.
2. Tell the platform — Plugins → AWS Lambda → Configure → AWS account ID = the 12 digits the command printed (the stack's DeployRoleArn output works too). Every app of the organization whose platform.toml says target = "aws/lambda" deploys from then on; nothing to register per app.
What the stack creates:
| Resource | Purpose |
|---|---|
| IAM OIDC provider | trusts IssuerUrl for audience sts.amazonaws.com; pass OidcProviderArn when the account already has one for the platform |
ActionPlatformAppBoundary (/action-platform/) |
the cap on every execution role an app's stack creates — logs /aws/lambda/<prefix>*; S3 buckets, DynamoDB tables, SQS queues, SNS topics, Lambda functions <prefix>-*; Secrets Manager secrets and SSM parameters <prefix>/*; X-Ray — <prefix> being the role's action-platform:prefix tag. An app's own template.yaml still declares what its function needs; the boundary caps it |
ActionPlatformDeploy role (/action-platform/) |
trust: AssumeRoleWithWebIdentity for sub org:<org>:*, and TagSession only for the one tag action-platform:prefix with a value ap-<org>-*. Policy: every resource under ${aws:PrincipalTag/action-platform:prefix} — CloudFormation stacks, Lambda, logs; API Gateway; the public Lambda Web Adapter layers; SAM's bucket; IAM on roles under the path /action-platform/<prefix>/ only — iam:CreateRole with the boundary attached and the app's own prefix tag, iam:PassRole to Lambda only; iam:SimulatePrincipalPolicy on itself |
How a deploy runs: the platform signs a token for the app with the claim https://aws.amazon.com/tags → principal_tags.action-platform:prefix = ap-<org>-<project>-<app>; the target assumes the deploy role with it, names the stack ap-<org>-<project>-<app>-<scope>, passes PermissionsBoundaryArn and RolePath=/action-platform/<prefix>/ to the overlay's template and tags the stack with the prefix, so the execution role SAM creates carries the boundary and the tag and lives under the app's path. The path, not the name, tells an app's roles apart: CloudFormation cuts a generated role name to 64 characters, which can drop the prefix from it. sam delete removes that role with the stack. Who may deploy which app is the platform's decision — it only signs a token for an app the job or the caller is for.
Readiness simulates the deploy role with the session tag, the boundary and iam:PassedToService as context entries, and flags an overlay without the PermissionsBoundaryArn or RolePath parameter. An app whose overlay predates them applies it again: Configuration → Deploy target, or action-platform cloud set aws/lambda. The stack's policy is the one place to widen or narrow what deploys may do.
Updating the connect stack
The command pins the template to the plugin's release. After the platform moves to a plugin whose connect template changed, run Connect AWS again (or action-platform aws-lambda connect …): CloudFormation updates action-platform-connect in place, and the account ID stays the same.
A role of your own (OIDC)
Instead of the connect stack, register the platform as an identity provider once per account and write the trust policy yourself:
aws iam create-open-id-connect-provider --url https://platform.example.com --client-id-list sts.amazonaws.com
{
"Version": "2012-10-17",
"Statement": [{
"Effect": "Allow",
"Principal": { "Federated": "arn:aws:iam::123456789012:oidc-provider/platform.example.com" },
"Action": "sts:AssumeRoleWithWebIdentity",
"Condition": {
"StringEquals": { "platform.example.com:aud": "sts.amazonaws.com" },
"StringLike": { "platform.example.com:sub": "org:acme:project:shop:app:orders" }
}
}]
}
requirements/policy.json from the overlay is a starting permission policy. Then:
[deploy]
target = "aws/lambda"
role_arn = "arn:aws:iam::123456789012:role/shop-deploy"
sub is org:<org>:project:<project>:app:<app> for a deploy the platform runs; org:<org> plus an actor claim for action-platform deploy from a logged-in machine (StringLike with org:acme:* covers both).
The AWS CLI's own credentials
With neither role_arn nor a connected account, aws and sam use what the machine has: aws sso login, a profile, an instance role. Still no key in a file when you use SSO — but nothing ties the identity to one app.
Where the deploy runs
Independent of how credentials are obtained:
| Where | Token for the role comes from |
|---|---|
| the platform's worker (a deploy job) | signed by the platform for that app: sub = org:…:project:…:app:…, the action-platform:prefix session tag |
a logged-in machine (action-platform deploy) |
POST /api/v1/identity/token with the user's login: sub = org:<org>, actor, scopes |
the overlay's .github/workflows/deploy.yml |
GitHub's own OIDC — optional, unrelated to the platform; delete the workflow when the platform deploys |
Tools and commands
aws_lambda_stacks / action-platform aws-lambda stacks [prefix] |
CloudFormation stacks |
aws_lambda_functions / action-platform aws-lambda functions [prefix] |
Lambda functions |
action-platform aws-lambda connect <platform url> <organization> |
the connect stack, in the account the CLI is logged in to |
Requirements
awsandsam— on PATH when present; otherwise theawscliandaws-sam-clipackages the plugin depends on run aspython -m, which is how the hosted platform deploys without the CLIs in its image.AWS_REGIONorregionunder[deploy].- Network:
*.amazonaws.com. - Connecting the account: credentials for it that may create IAM resources, once — AWS CloudShell through the Connect AWS button, the console, or
awsforaction-platform aws-lambda connect. action-platform0.32.0 or newer: the Connect AWS button is an option action of the core.
Development
pip install -e ".[dev]"
pytest
The tests fake aws and sam; nothing reaches AWS.
| Module | Responsibility (ABC in abc.py) |
|---|---|
lambda_.py |
LambdaTarget: the core's DeployTarget contract, composing the parts below through parts(spec, access) |
spec.py |
Spec: samconfig, stack, region and — for a connected account — prefix, deploy role, boundary, resolved once per call without calling AWS |
credentials.py |
Credentials: OwnRole ([deploy] role_arn), ConnectedAccount, CliChain — which identity aws and sam act as |
stack.py |
Stack: the CloudFormation stack — status, outputs, clearing a failed first creation, rollback |
sam.py |
Sam: build, deploy, delete, validate |
health.py |
Health: the function answers its health route after the deploy |
checks.py |
Readiness: one class per check |
shell.py |
running aws and sam, streaming sam's output to the job log |
tools.py, cli.py |
read-only views for MCP and the terminal; connect |
Metadata
Release files for apx-aws-lambda 1.3.0
For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.
Source distribution (sdist)
| File | Size | Uploaded | |
|---|---|---|---|
| apx_aws_lambda-1.3.0.tar.gz | 42.1 kB | Details |
Built distribution (wheel)
| File | Interpreter | ABI | Platform | Reset |
|---|---|---|---|---|
| apx_aws_lambda-1.3.0-py3-none-any.whl | Python 3 | none | any | Details |
Total release size: 82.2 kB
Release files / apx_aws_lambda-1.3.0.tar.gz
| Download URL | apx_aws_lambda-1.3.0.tar.gz |
|---|---|
| Size | 42.1 kB |
| Tags | Source |
|
SHA-256 checksum How to use checksums |
fd736cd83490fddd350bf2843aeac9208eecfdce0e40b7a1c7d25804fdf34cb3
|
|
BLAKE2b-256 checksum How to use checksums |
2083ec4a10fcffd0cfe3c375bb8a47ffb86e6192027ea7f3882c8c9c005499d9
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
Yes |
| Uploaded via |
twine/7.0.0 CPython/3.13.14
|
Provenance
Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.
PyPI Publish Attestation
PyPI verified that this artifact, at this checksum, originated from the publisher listed below.
Signed by GitHub Actions, verified by PyPI on Sep 27, 2026.
Transparency logRelease files / apx_aws_lambda-1.3.0-py3-none-any.whl
| Download URL | apx_aws_lambda-1.3.0-py3-none-any.whl |
|---|---|
| Size | 40.2 kB |
| Tags | Python 3 |
|
SHA-256 checksum How to use checksums |
18ac8524b980df29e996819fafeef07effdd9b87574eba2932dfd3dba992d93d
|
|
BLAKE2b-256 checksum How to use checksums |
bb05b736ab5d45b3c3a8d6c3d956f75c41dccfe217c02d09a7bacc9d1bd017f7
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
Yes |
| Uploaded via |
twine/7.0.0 CPython/3.13.14
|
Provenance
Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.
PyPI Publish Attestation
PyPI verified that this artifact, at this checksum, originated from the publisher listed below.
Signed by GitHub Actions, verified by PyPI on Sep 27, 2026.
Transparency log