What the agent does
The agent reads the stack before adding anything. In a project like the typescript/api-cors-lambda-crud-dynamodb example in aws-samples/aws-cdk-examples, index.ts defines ApiLambdaCrudDynamoDBStack: a DynamoDB Table named items with itemId as partition key and RemovalPolicy.DESTROY, shared NodejsFunctionProps that bundle with @aws-sdk/* external, read lambdas/package-lock.json, set PRIMARY_KEY and TABLE_NAME and run on NODEJS_24_X, and five NodejsFunction constructs, getOneItemFunction, getAllItemsFunction, createItemFunction, updateItemFunction and deleteItemFunction, each from a handler under lambdas/. Each gets grantReadWriteData, a LambdaIntegration, and a method on the itemsApi RestApi: GET and POST on items, GET, PATCH and DELETE on items/{id}, with OPTIONS from addCorsOptions. package.json pins aws-cdk-lib 2.262.1 and the aws-cdk CLI 2.1133.0, with Jest 30 and @swc/jest; jest.config.js matches **/*.test.ts and loads aws-cdk-lib/testhelpers/jest-autoclean, but no test file exists. cdk.json runs npx tsc && npx tsx index.ts and already sets @aws-cdk/aws-codedeploy:removeAlarmsFromDeploymentGroup. Every deploy replaces all five functions' code at once, with no way to shift traffic gradually or roll back on errors.
For each function it adds a lambda.Alias named live on fn.currentVersion, and changes the five LambdaIntegration constructs to take the alias rather than the function, so API Gateway invokes the alias ARN and CodeDeploy can move traffic between versions behind it. A cloudwatch.Alarm per alias watches alias.metricErrors over one-minute periods with a threshold of 1 and one evaluation period, and a codedeploy.LambdaDeploymentGroup per alias uses LambdaDeploymentConfig.CANARY_10PERCENT_5MINUTES with that alarm in alarms, so a deploy sends a tenth of requests to the new version for five minutes and rolls back if the alarm fires. The helper that builds the alias, alarm and group is one function in index.ts called five times, so the five functions stay symmetric. grantReadWriteData stays on the functions, since the alias shares the function's role.
currentVersion only publishes a new version when the code or configuration changes, which is what makes a canary meaningful: an unchanged function keeps its version and CodeDeploy has nothing to shift. The agent sets removalPolicy: RemovalPolicy.RETAIN on the current version options so an older version stays available for a rollback after the deploy completes, and writes in README.md what a deploy now looks like in the CodeDeploy console and how to watch the alarm.
The proof is a test the example never had. index.test.ts uses Template.fromStack from aws-cdk-lib/assertions to assert five AWS::Lambda::Alias resources named live, five AWS::CodeDeploy::DeploymentGroup resources with DeploymentConfigName set to CodeDeployDefault.LambdaCanary10Percent5Minutes and an alarm in AlarmConfiguration, five alarms on the Errors metric, and that every API Gateway method's integration URI references an alias rather than a bare function ARN. npm run build type-checks, npx jest runs the test, and cdk synth runs the way scripts/build-typescript.sh and scripts/synth.sh do in the repository's pull request workflow, with the fake context file so no account is needed.
Start it with one API call
POST https://api.phoenix.so/o/{org}/sessions with this body:
{
"agent": "claude",
"input": "Add canary deployments to every Lambda function in this CDK stack. Read index.ts, cdk.json, package.json, jest.config.js, tsconfig.json, README.md, the handlers under lambdas/ and the repository's scripts/build-typescript.sh and scripts/synth.sh first. For each function create a lambda.Alias named live on currentVersion, point its LambdaIntegration at the alias, add a CloudWatch alarm on the alias's error metric with one-minute periods and a threshold of one, and a codedeploy.LambdaDeploymentGroup using LambdaDeploymentConfig.CANARY_10PERCENT_5MINUTES with that alarm; do it through one helper called once per function. Keep the DynamoDB grants on the functions, retain old versions, and explain in README.md how a deploy now shifts traffic and rolls back. Write index.test.ts with Template.fromStack assertions for the five aliases, five deployment groups with the canary config and alarm, five error alarms, and API Gateway integrations that reference aliases. Run npm run build, npx jest and cdk synth with the repository's fake context, and open a pull request.",
"resources": [
{
"url": "https://github.com/your-org/your-app"
}
]
}- Get an API key. In the console, open Settings, then API keys, and create a key for your org. Send it in the Authorization header as a Bearer token.
- Send the request. POST the body in request.json to https://api.phoenix.so/o/{org}/sessions, with your repository as the resources URL. The run starts in its own cloud sandbox.
- Read what came back. Follow the run on its live stream or wait for the run.completed webhook, then read the result and its trajectory.
What comes back
What comes back is a pull request that edits index.ts and README.md and adds index.test.ts. The description names the alias, alarm and deployment group each function gains, the canary configuration, and the one behaviour change a reviewer should expect: a deploy that changes a handler now takes at least five minutes to complete while traffic shifts.
The run's record keeps the tsc output, the Jest run with each assertion named, and the synthesized template from cdk synth, so a reviewer can read the AWS::CodeDeploy::DeploymentGroup resources and the integration URIs directly rather than trusting the description.
Before merging, a reviewer deploys to a non-production account, changes one handler, deploys again, and watches the deployment group in the CodeDeploy console shift ten percent of traffic to the new version. Forcing an error in that version during the window should trip the alarm and roll the alias back, which is the behaviour the alarm wiring exists for.
Run it on every Linear issue labelled deploy-safety
In the console, an automation with a linear trigger on linear.issue.labeled and the label deploy-safety starts one run per labelled issue in the Project's repository. The trigger adds the issue's text after the saved prompt, so an issue asking for a linear ramp instead of a canary, or for a latency alarm beside the error alarm, carries that itself and the prompt stays fixed. Changes made by an app integration never fire it, and the automation's max_runs_per_hour bounds a burst of labelling.
From your own server, the same job is one POST /o/{org}/sessions per issue, with the repository in resources and the issue's https://linear.app/<workspace>/issue/<ID> URL beside it so Rush reads the issue through your Linear connection. A webhook_url on the API key receives run.completed or run.failed for every run that key starts, with the session_url to read the pull request link from.
The use case this job follows, with its automation and trigger: Turn a Linear label into a PR.
Why run it on Rush?
The pieces are documented one at a time, and the mistakes are in how they connect: an integration still pointing at the function so the canary shifts nothing, an alarm on the function's metric rather than the alias's so a bad version never trips it, a version removal policy that deletes the only rollback target. The agent synthesizes the real template in its own sandbox and writes assertions against the resources, which a reading of the construct code does not do.
Every command and its output are kept with the run, so a reviewer can open the synthesized template from the record and check the integration URIs and deployment group properties themselves, before any deploy touches an account.
Frequently asked questions
Why point API Gateway at the alias instead of the function?
CodeDeploy shifts traffic by changing the weights behind an alias. An integration that names the function directly always invokes $LATEST, so the deployment group would move traffic the API never sends. The test asserts every integration URI references an alias for that reason.
What triggers a canary, and what does not?
A change to a handler under lambdas/, to the bundling options or to the environment publishes a new version through currentVersion, and the deployment group shifts traffic to it. A deploy that changes only the table or the API resources publishes nothing, and the alias stays where it is.
Can the alarm be something other than errors?
Yes. alarms on the deployment group takes any CloudWatch alarm, and a duration or throttle alarm on the alias's metrics fits the same helper. Errors is the one that catches a handler that throws, which is the most common bad deploy for CRUD functions like these.
Does this change the DynamoDB table or its removal policy?
No. The table, its DESTROY policy and the grantReadWriteData calls are untouched; the grants apply to the function's role, which the alias shares. Only the function wiring, the new deployment resources and the README change.