13 November 2026

Compass 0.9.2: A first install off AWS reaches a map and a verified import

A first install on a laptop or a server outside AWS now goes from an empty folder to a map and a verified import pull request without a workaround: connect creates the runtime user before anything that names it and waits for AWS to know it, an import is verified with the read role when the machine has no AWS identity of its own, the map shows your account's tables rather than Compass's own, and two drifts that were not drift are gone.

connect works on a first run off AWS

On a machine with no AWS role of its own, connect makes a runtime user and names it in the alert queue's policy and the roles' trust. It named the user before creating it, and AWS refused. Now the user comes first, and connect waits up to a minute for AWS to know a new identity, and 90 seconds for its first AssumeRole, instead of failing.

compass connect account=<account id> region=us-east-1

Verification no longer shows a denied read

connect checks the roles it made through the read role itself. iam:GetPolicy and iam:GetPolicyVersion were granted on the two roles, and a policy is not a role, so the table reported both as denied on every install, however complete it was. They are now a statement of their own on the account's and AWS's managed policies.

compass connect verify

A captain with many reads gets a session

A captain's session is narrowed to the reads it declares, passed inline, at most 2,048 characters. A captain whose reads did not fit was refused a session. connect now creates a managed policy for it, svc-compass-<captain>-<install id> under /service/, unattached, and records it under aws.managed_policies; the session names that instead.

svc-compass-<captain>-<install id>

Imports are verified off AWS

An import pull request is verified by a Terraform plan. On a machine with no AWS credentials in its environment, init against the S3 backend found none and the pull request opened unverified. The plan now runs as the read role, reading the state and the imported resources with -refresh=false, so nothing else the root module manages is touched.

terraform plan -refresh=false

Two drifts that were not drift

Coverage compared a declared value with the live one as text, so a number AWS returns as a string, an SQS queue's visibility timeout of "30", drifted from the 30 in Terraform. And a load balancer's internal = true was held against the live scheme, "internal". Both now compare as what they mean, and the drift list keeps the drift that is real.

The map shows your account's tables

A local install keeps its own tables in DynamoDB Local and points every DynamoDB client at it. The mapper's client inherited that, so Compass's own tables appeared on the map as unmanaged resources of your account, and the account's tables went unseen. The mapper now reads the account's DynamoDB, whatever the install itself uses.

Compose defaults name this release

docker-compose.yml and .env.example kept 0.9.0 as the default after 0.9.1 was out, so a fresh clone with no COMPASS_VERSION in its .env ran a release behind and the dashboard offered the update at once. They name 0.9.2.

COMPASS_VERSION=0.9.2

After the update

Compass is open source (Apache-2.0).

Code, changelog and issues: github.com/beamreach-ai/compass. Already running it? The update card on Settings → Updates shows these notes before you update.