Thank you for using CodeRabbit. We offer it for free to the OSS community and would appreciate your support in helping us grow. If you find it useful, would you consider giving us a shout-out on your favorite social media?

Thank you for starting this@genu ! It's something wanted by many.

I think the basic implementation should be straightforward. The call path forfindXXX proxy is:

PolicyProxyHandler.findXXX -> PolicyProxyHandler.doFind -> PolicyUtils.injectForRead

PolicyUtils.injectForRead is the unified function for injecting "read" related policies. Here' you can see currently only "read" policies are injected:

I think what needs to be done is to further inject "list" policies.

Also, the "list" policy should only govern "findFirst" and "findMany", not "findUnique". We should probably introduce a boolean flag to distinguish these two cases.

I guess the proxy code is not entirely easy to read and understand 😅. Let me know if you need more clarifications.

Another complication is, whether "list" should governgroupBy call, sincegroupBy is partly "listing" data as well. What do you think?

Another complication is, whether "list" should governgroupBy call, sincegroupBy is partly "listing" data as well. What do you think?

I'm not exactly sure, but I think I would expectlist to govern anything that "lists" things.

@ymc9 Do we need to distinguish between betweenfindFirst andfindMany here?

or shouldinjectForRead be modified here to accept afindMany parameter to indicate that thelist permission should apply?

Also, why wouldlist apply to afindFirst?

@ymc9 Do we need to distinguish between betweenfindFirst andfindMany here?

zenstack/packages/runtime/src/enhancements/node/policy/handler.ts

Lines 133 to 151 inb41fd93

	privateasyncdoFind(args:any,actionName:FindOperations,handleRejection:()=>any){
	constorigArgs=args;
	const_args=this.policyUtils.safeClone(args);
	if(!this.policyUtils.injectForRead(this.prisma,this.model,_args)){
	if(this.shouldLogQuery){
	this.logger.info(`[policy] \`${actionName}\`${this.model}: unconditionally denied`);
	}
	returnhandleRejection();
	}
	this.policyUtils.injectReadCheckSelect(this.model,_args);
	if(this.shouldLogQuery){
	this.logger.info(`[policy] \`${actionName}\`${this.model}:\n${formatObject(_args)}`);
	}
	constresult=awaitthis.modelClient[actionName](_args);
	returnthis.policyUtils.postProcessForRead(result,this.model,origArgs);
	}

or shouldinjectForRead be modified here to accept afindMany parameter to indicate that thelist permission should apply?

Also, why wouldlist apply to afindFirst?

I think 'list' should governfindFirst because otherwise you'll be able to iteratively callfindFirst to simulatefindMany:

letr=awaitdb.foo.findFirst();constknownIds:string[]=[];while(r){knownIds.push(r.id);r=awaitdb.foo.findFirst({where:{not:{id:{in:knownIds}}}});}

We can probably introduce a boolean parameter todoFind to distinguish the two cases where "list" should be enforced or not.

@ymc9 Here are some ideas from thinking through this:

findMany

• Controlled by list permission since you're viewing multiple resources
• Would attribute permissions be required to determine what fields are visible in alist context? (In the case thatlist is allowed andread is denied). Unlesslist impliesread but that may not necessarily the case.

findFirst

• Still governed by list permission even though returning single item
• Because it involves scanning/viewing multiple items to find the match

Nested to-many relations during read

• List permission required for the related collection (should a permission allow list, if parent allows aread? Similar to howcheck works currently)

Aggregate (This is kind of tricky)

• List permission required as you're viewing/organizing multiple items
• May need attribute permissions for the grouping fields, plus permissions for the fields being aggregated within groups

Count

• Requires list permission as it reveals information about collection size
• Maybe we have to treat this as a separated permission as count info is sensitive?

Maybe for an initial implementation we should stick to the obvious/straight forward approach and polish it later

thoughts?

Hi@genu , sorry I almost missed this comment. I think my thoughts generally aligns with your description. I'm sharing the differences below.

@ymc9 Here are some ideas from thinking through this:

findMany

• Controlled by list permission since you're viewing multiple resources
• Would attribute permissions be required to determine what fields are visible in alist context? (In the case thatlist is allowed andread is denied). Unlesslist impliesread but that may not necessarily the case.

I think field-level and model-level permissions should work independently (as they do today). Model-level "list" determines if you can list entities, and field-level rules determine what fields you can see. You can't define "list" rules at the field level because it's meaningless.

findFirst

• Still governed by list permission even though returning single item
• Because it involves scanning/viewing multiple items to find the match

Nested to-many relations during read

• List permission required for the related collection (should a permission allow list, if parent allows aread? Similar to howcheck works currently)

The permission checking of fetching relations doesn't inherit from its parent, instead, for "read", the relation's model's "read" rules are evaluated to filter out items that shouldn't be seen, regardless if the parent is readable. I think "list" can behave the same way.

Aggregate (This is kind of tricky)

• List permission required as you're viewing/organizing multiple items
• May need attribute permissions for the grouping fields, plus permissions for the fields being aggregated within groups

Count

• Requires list permission as it reveals information about collection size
• Maybe we have to treat this as a separated permission as count info is sensitive?

I agree "count" should require "list" permission. I don't see a separate permission kind is needed for it for now. We can iterate in the future as we gather feedback.

Maybe for an initial implementation we should stick to the obvious/straight forward approach and polish it later

I totally agree. There're many details to sort out. Just having a singlefindMany to work will be a great start!

thoughts?

Another major problem is how we add the "list" permission without introducing breaking changes, plus allowing incremental adoption. Here are my thoughts:

• If there're no "list" rules defined for a model, "list" is allowed if "read" is. This will allow people to continue usingfindMany etc. with only "read" rules defined.
• If "list" rules are defined for a model, it then implies "read", and operations likefindMany,aggregate, etc. will require "list" permissions to work. Basically, by starting using "list" permission you opt in the new behavior for that model.

What do you think?

@ymc9 Do you mind taking over this PR, I'm not exactly sure how to proceed.

I'll keep an eye out for an updates and review/test as needed.

@ymc9 your comments above make sense on how you expect it to work. I do have a suggestion though.

If there're no "list" rules defined for a model, "list" is allowed if "read" is. This will allow people to continue using findMany etc. with only "read" rules defined.
If "list" rules are defined for a model, it then implies "read", and operations like findMany, aggregate, etc. will require "list" permissions to work. Basically, by starting using "list" permission you opt in the new behavior for that model.

What I'm gathering from this is that this will just transparently work and start filtering out entries if it's not specifically defined right but the user has read access? If I'm correct in this understanding then I recommend below.

The permission checking of fetching relations doesn't inherit from its parent, instead, for "read", the relation's model's "read" rules are evaluated to filter out items that shouldn't be seen, regardless if the parent is readable. I think "list" can behave the same way.

Since this is a breaking change for some users, I would recommend a flag to enable the filtering behaviour and have it off by default. This would allow people to do updates without items unexpectedly disappearing from their lists. E.g. it could be a feature flag in the zmodel as behaviour on or off globally (I don't mean for each model).

@ymc9 Do you mind taking over this PR, I'm not exactly sure how to proceed.

I'll keep an eye out for an updates and review/test as needed.

Sure@genu , and thank you for initiating this effort! I'm currently in the middle of experimenting some larger refactor, so will need to put this PR on hold for a bit. I'll keep you updated when I resume the work.

@ymc9 your comments above make sense on how you expect it to work. I do have a suggestion though.

If there're no "list" rules defined for a model, "list" is allowed if "read" is. This will allow people to continue using findMany etc. with only "read" rules defined.
If "list" rules are defined for a model, it then implies "read", and operations like findMany, aggregate, etc. will require "list" permissions to work. Basically, by starting using "list" permission you opt in the new behavior for that model.

What I'm gathering from this is that this will just transparently work and start filtering out entries if it's not specifically defined right but the user has read access? If I'm correct in this understanding then I recommend below.

The permission checking of fetching relations doesn't inherit from its parent, instead, for "read", the relation's model's "read" rules are evaluated to filter out items that shouldn't be seen, regardless if the parent is readable. I think "list" can behave the same way.

Since this is a breaking change for some users, I would recommend a flag to enable the filtering behaviour and have it off by default. This would allow people to do updates without items unexpectedly disappearing from their lists. E.g. it could be a feature flag in the zmodel as behaviour on or off globally (I don't mean for each model).

Hi@hongkongkiwi , sorry I didn't explain clearly in the previous reply. The intention is to avoid any breaking changes. If you don't write any "list" rules, the "read" rules will behave the same as they do today. The behavior will only change if you start to have "list" rules in a model.

I like the idea of using a runtime flag to control the behavior as well. It's probably cleaner. I'll keep thinking about it. Thanks!

What happened to this list permission? Is work stopped?

What happened to this list permission? Is work stopped?

Correct me if I'm wrong@ymc9 but I believe this feature is being pushed tov3 which is in the works here:https://github.com/zenstackhq/zenstack-v3

What happened to this list permission? Is work stopped?

Correct me if I'm wrong@ymc9 but I believe this feature is being pushed tov3 which is in the works here:https://github.com/zenstackhq/zenstack-v3

Yes, that's what I'm thinking. It'll need to be rewritten if implemented in v2.