Skip to main content

Project

Project​

Defines a logical project.

A project is the boundary for resource isolation, IAM permissions and network segmentation.

PropertyTypeDescription
apiVersionstringAPI schema version. Must be lowops.manifests.v1.
kindstringResource kind. Must be Project.
metadatamapReserved keys such as name, environment, organizational-unit and organization resolve the project identity and its parent Environment.
specProjectDefinitionDesired project configuration, as a shared ProjectDefinition.

AccessibleRegistries​

Registry access grouped by compute target.

The registries that workloads of this project pull from.

PropertyTypeDescription
cloudrunlist of ComputedAccessibleRegistryRegistries for the Cloud Run applications of the project.
computelist of ComputedAccessibleRegistryRegistries for the Compute Engine applications of the project.

AgentApplicationsEntry​

PropertyTypeDescription
keystring
valuestring

CloudrunApplicationsEntry​

PropertyTypeDescription
keystring
valueComputedApplication

CloudrunJobApplicationsEntry​

PropertyTypeDescription
keystring
valueComputedApplication

ComputeApplicationsEntry​

PropertyTypeDescription
keystring
valueComputedApplication

IdentityProviderAuthorizedDomainsEntry​

PropertyTypeDescription
keystring
valuestring

McpApplicationsEntry​

PropertyTypeDescription
keystring
valuestring

VpcSpokesEntry​

PropertyTypeDescription
keystring
valuestring

ComputedAccessibleRegistry​

Computed Artifact Registry repository.

Holds the name and location that resolve image paths.

PropertyTypeDescription
namestringName of the Artifact Registry repository.
locationstringGCP region or multi-region of the registry. It forms part of the image path.
trustedRepositorieslist of stringTrusted repository children of an external registry. Binary Authorization allows each one separately, not the whole registry. Empty for internal repositories, which use attestation.

ComputedApplication​

Computed deployment configuration of an application.

Holds the mesh strategy, the migration requirements, and the container image layout.

PropertyTypeDescription
namestringResolved application name.
meshStrategystringService mesh strategy, for example SIDECAR, PROXYLESS, DISABLED, or EXCLUDED. It controls traffic routes and service URLs.
containerComputedContainerDefinitionResolved container image of the application.

ProjectDefinition​

Definition of a Project that holds workloads.

It maps to a GCP project and sets the region, IAM bindings, maintenance windows and default routing rules.

PropertyTypeDescription
displayNamestringDisplay name of the GCP project. Same rules as OrganizationalUnitDefinition.display_name.
descriptionstringFree-text description.
permissionsAccessPermissionsPermissions for the Project. The engine binds the IAM of the Project and of its folder from this block only. It does not merge this block with the permissions of the parent Environment or OrganizationalUnit. In a self-hosted installation, Google inherits the IAM that the engine binds on those parent folders.
networkProjectNetworkNOT YET IMPLEMENTED. The engine ignores this field. See ProjectNetwork.
hibernationHibernationConfigHibernation schedule for the Project. It overrides schedules from parent manifests.
regionstringPrimary GCP region for the Project and its resources. Defaults to the region of the parent Environment.
maintenanceMaintenanceMaintenance windows and exclusions. Required. See Maintenance.
defaultUrlRedirectstringURL to redirect to when a request matches no other routing rule in the Project. Required.
allowedEgresslist of stringExternal hostnames or IP ranges that applications in the Project can connect to. The engine enforces this policy on outbound traffic.
identityProviderConfigProjectIdpConfigIdentity provider settings for all tenants in the Project. See ProjectIdpConfig.
rsiRsiConfigPolicy for recursive self-improvement (RSI) in the Project. See RsiConfig. RSI is opt-in and it is off by default. It is on only if the block has enabled: true, and only in a non-production environment. The engine does not act on this field. The planned readers are Beck and the MCP server.
mcpGcpAccessMcpGcpAccessOpt-in for read-only GCP access through the Infrastream MCP. See McpGcpAccess. NOT a write path: the engine makes read-only reader service accounts. They hold the managed roles in managed hosting. In a self-hosted installation they hold two generated roles that the installer supplies. If the block is absent, or enabled is false, the feature is off.

RsiConfig​

Policy for recursive self-improvement (RSI) in a Project.

RSI lets an agent open pull requests without a human confirmation. RSI is opt-in and it is off by default. The policy is on only if enabled is true. The engine does not act on this policy. It loads and validates the policy only. The planned readers are Beck and the MCP server.

RSI is off in these cases: no spec.rsi, rsi:, rsi: null, rsi: {}, enabled: false, and entries with no enabled. An entry never turns RSI on. A block with entries and no enabled is valid. The engine writes a warning that names the Project, because the author may think that RSI is on. A reader must apply the same rules.

The engine clears a block that is off and has no entry when it runs the defaulters. A YAML loader reads rsi: and rsi: {} as a block with a value.

RSI can be on only in a non-production environment. The environment is the value of metadata.environment. The engine rejects enabled: true unless that value is one of these names: development, dev, integration, int, qa, test, sandbox, preview and ephemeral. The engine compares the name exactly, with the case as written. It rejects every other name, so it rejects production and staging. Staging is not a non-production environment, and an unknown name is not safe.

The path of the manifest file must agree with metadata.environment. The path must have one directory named environment, and the name of the environment must follow it, for example environment/dev/project/web.yaml. The engine rejects enabled: true if the path has no such pair, or more than one, or if the name differs from metadata.environment. A reader must apply these two rules too, because the engine checks only the manifests that it scans.

A change to this block must need a human confirmation. This is a requirement on the clients that write manifests, and the engine does not enforce it.

A bad value in this block fails the scan in the same way as a bad value in any other manifest. The engine makes no exception for this block. If any value breaks a rule, a reader must treat the whole rsi block as invalid, and RSI is off for the Project. A reader must never drop the bad entry and use the rest, because a shorter list of protected paths is unsafe. A reader must read each entry as text. An entry that is not a string, such as ~, null, true or 1, breaks the rules.

The engine accepts only the YAML values true and false for enabled. A reader must treat any other value, such as yes, on, 1 or the text "true", as a break of the rules. If enabled appears more than once, the engine takes the last value. A reader must treat RSI as off unless every occurrence is true.

The engine joins repeated keys in a YAML manifest. If spec.rsi appears twice, or a list key appears twice, the engine takes the union of all the lists. A reader must take the union of every occurrence too, and never only the last one. A JSON manifest is different: the engine rejects a repeated key.

The policy is for one Project. The engine does not inherit it from the Environment or the OrganizationalUnit, and no child manifest receives it. The engine accepts the block in every environment, but it rejects enabled: true outside a non-production environment.

PropertyTypeDescription
enabledboolSwitch that turns RSI on for the Project. RSI is on only if the value is the boolean true. If the key is absent, or the value is false, RSI is off. The entries of this block never turn RSI on. RSI can be on only in a non-production environment. The engine rejects enabled: true in production, in staging and in any environment that it does not know. See RsiConfig for the allowed names and the rule for the path.
repositorieslist of stringGitHub repositories where an agent can open a pull request without a human confirmation. Each entry has the form owner/repo, for example acme/billing-api. If empty, no code repository is in scope of RSI. A reader must compare entries without regard to case, because GitHub treats Acme/API and acme/api as one repository. If the key repositories appears more than once, the engine joins the lists, and a reader must take the union of all of them.
protectedPathslist of stringPatterns of repository paths that always need a human confirmation, for example go.mod or scripts/ci/**. They add to the permanent list. A reader must always treat .github/** and every file named CODEOWNERS as protected. This field cannot remove them. A reader must read the keys protectedPaths and protected_paths, and take the union of every list. A pattern is a path from the repository root, with / as separator. It matches the whole path, case-sensitively, with the code points as stored and no Unicode normalisation. * matches any characters inside one segment, also a name that starts with a dot. ** matches any number of whole segments, also none, and /** at the end matches the paths below a directory. Every other character is literal.

McpGcpAccess​

Opt-in for read-only access to the GCP resources of a Project through the Infrastream MCP.

NOT a write path. The engine makes read-only reader service accounts. The MCP server uses them to read the resources of the Project for a person who has a tier in spec.permissions. No role of these service accounts can write.

If enabled is true, the engine makes the service account mcp-gcp-reader in the Project. The MCP server uses it for the tier viewers. The engine makes the service account mcp-gcp-editor-reader too, if the conditions below are true. The MCP server uses it for the tiers contributors and administrators.

A reader service account never has a permission that the own tier role of the person lacks. The roles depend on the hosting mode.

In managed hosting, mcp-gcp-reader holds the role managed_viewer. The engine makes mcp-gcp-editor-reader only if the installer supplies the role managed_editor_read. It holds that role.

In a self-hosted installation, the tier roles of the people are the basic roles roles/viewer, roles/editor and roles/owner. The managed roles have permissions that these roles lack. So mcp-gcp-reader holds the role managed_viewer_basic, and mcp-gcp-editor-reader holds the role managed_editor_read_basic. The installer makes both roles. Each one is never more than roles/viewer or roles/editor.

The engine makes both service accounts only if the installer supplies both roles. Both must be custom roles of the organization of the installation, and they must have these exact IDs.

The installer can supply the service account of the MCP server. If it does, the engine gives that one account, and no other account, the right to create tokens for the two reader service accounts. If the installer does not supply it, the engine gives no such right.

The engine makes none of these service accounts, and writes a warning, if a setting is missing or wrong. In managed hosting, the settings are the three tier roles managed_viewer, managed_editor and managed_admin, the role managed_editor_read and the service account of the MCP server. Without a tier role, the engine binds nobody for that tier, and a reader service account would give reads that the person does not have. In a self-hosted installation, the settings are the two basic roles and the service account of the MCP server. A basic role with another ID, or in another organization, is wrong.

A skipped Project is not an error. The engine builds the rest of the Project, and the rest of the organization, as for any other Project.

If the block is absent, or enabled is false, the engine makes none of these service accounts. It deletes the service accounts that an earlier run made.

The MCP server can serve a person only if it can find the tier of the person in the manifests. This works only if the workforce federation of the Organization is infrastreamManaged, because only then the manifests list the users. For Google Workspace and Microsoft Entra ID the tier is unknown. The MCP server then gives that person no access.

The policy is for one Project. The engine does not inherit it from the Environment or the OrganizationalUnit. In a self-hosted installation, a person can have more rights from the folders of the Environment and of the OrganizationalUnit. The MCP server does not count them. It uses the tier of the Project only, so it can give less than the person has, and never more.

PropertyTypeDescription
enabledboolSwitch that turns the reader service accounts on for the Project. They exist only if the value is the boolean true. If the key is absent, or the value is false, the engine makes none and deletes the ones that exist. The switch does not give a person any right. The tier of the person in spec.permissions decides what the MCP server reads for that person.

ComputedContainerDefinition​

Computed container image source of a deployment.

Gives the image URL for the deployment API.

PropertyTypeDescription
sourceRegistryComputedAccessibleRegistryArtifact Registry that stores the image.
imagestringAbsolute image reference, with registry host and repository path.

HibernationConfig​

Hibernation schedule configuration.

The engine combines windows and exclusions from the OU, Environment and Project into the final state.

PropertyTypeDescription
hibernateboolIf true, the resource hibernates immediately and overrides all windows and exclusions. Defaults to false.
windowslist of WindowsEntryTime windows that repeat, in which the resource hibernates. The map key is a unique name.
exclusionslist of ExclusionsEntryOne-time windows that suspend hibernation, even if a window is active. Use them for planned maintenance or high-traffic periods. The map key is a unique name.

ProjectNetwork​

VPC flow log settings for a Project.

NOT YET IMPLEMENTED. The engine ignores this message.

PropertyTypeDescription
logsNetworkLogsNOT YET IMPLEMENTED. The engine ignores this field. See NetworkLogs.

AccessPermissions​

Administrator, contributor, and viewer privileges for an asset.

The engine aggregates the users and groups into IAM policies.

PropertyTypeDescription
administratorsDetailedAccessPermissionsUsers and groups with administrative privileges, usually full control.
contributorsDetailedAccessPermissionsUsers and groups with contributor privileges, usually read and write access.
viewersDetailedAccessPermissionsUsers and groups with viewer privileges, usually read-only access.

ProjectIdpConfig​

Project-wide Identity Provider (IdP) configuration.

It applies identity and MFA settings to all Identity Platform tenants in the project.

PropertyTypeDescription
mfaMfaConfigNOT YET IMPLEMENTED. The engine ignores this field. See MfaConfig.
gdprComplianceboolEnables GDPR compliance for all tenants in the project. If true, users can delete their own accounts. The login app shows a terms of service checkbox and a "Delete my account" button. All child IdentityProvider manifests must then set terms_of_service and privacy_policy URLs. The engine enforces this at compute time.

Maintenance​

Weekly maintenance window and exclusions for resource upgrades.

It controls when Google Cloud upgrades resources, such as GKE clusters and Cloud SQL instances.

PropertyTypeDescription
startstringRFC3339 timestamp. Its time of day and day of week set the start of the weekly window. RFC3339
endstringRFC3339 timestamp for the end of the weekly window. RFC3339
exclusionslist of MaintenanceExclusionOne-time windows in which maintenance must not run, even inside the weekly window. Use them for business-critical periods.

DetailedAccessPermissions​

User and group access definitions.

The engine resolves the listed manifests to Google Workspace identity emails for IAM bindings.

PropertyTypeDescription
memberslist of stringNames of OrganizationUser manifests.
groupslist of stringNames of OrganizationUserGroup manifests.

WindowsEntry​

PropertyTypeDescription
keystring
valueHibernationWindow

NetworkLogs​

VPC Flow Logs export settings.

NOT YET IMPLEMENTED. The engine ignores this message. Intended use: set the aggregation interval and sampling rate of flow logs.

PropertyTypeDescription
intervalstringAggregation interval. One of the INTERVAL_* values, from INTERVAL_5_SEC to INTERVAL_15_MIN.
samplingdoubleFraction of connections to capture, from 0.0 (no logs) to 1.0 (all logs).

MfaConfig​

Multi-factor authentication (MFA) requirements for users.

NOT YET IMPLEMENTED as a project-wide setting. Configure MFA for each tenant on the IdentityProvider manifest.

PropertyTypeDescription
modestringMFA enforcement mode: DISABLED, ENABLED (optional) or MANDATORY (required for all users).
testPhoneNumberslist of TestPhoneNumbersEntryMap of a phone number to its expected 6-digit OTP code. It tests MFA flows without real SMS.
allowedRegionslist of stringTwo-letter Unicode CLDR region codes of the regions where users can use MFA. See https://cldr.unicode.org/.

MaintenanceExclusion​

One-time window in which platform maintenance must not run.

It overrides the weekly maintenance window.

PropertyTypeDescription
namestringUnique label for the reason, for example black-friday-freeze.
startstringWindow start, in RFC3339 format. RFC3339
endstringWindow end, in RFC3339 format. RFC3339

ExclusionsEntry​

PropertyTypeDescription
keystring
valueHibernationExclusion

TestPhoneNumbersEntry​

PropertyTypeDescription
keystring
valuestring

HibernationExclusion​

Suspension of the hibernation schedule.

Workloads stay active in the given time frame.

PropertyTypeDescription
startstringWindow start, in RFC3339 format. RFC3339
endstringWindow end, in RFC3339 format. RFC3339

HibernationWindow​

Period when an asset scales down.

Cron strings tell the control plane when to stop virtual machines or scale Cloud Run instances to zero.

PropertyTypeDescription
startstringCron expression for the start of the window.
endstringCron expression for the end of the window.