Problem Statement: Centralized BFF Authorization & PDP-Agnostic API

Status: Draft for second-opinion review
Date: 2026-06-05
Authors: Sri Aradhyula (problem framing); AI-assisted synthesis from PR review + team discussion
Audience: Platform engineers, security architects, agents reviewing implementation options

Related work:

Link	Role
PR #1751	Workflow RBAC visibility, run access, unsaved UX, agent-access modal
PR #1751 review r3362688683	@subbaksh: workflows are UI/BFF concept; DA should not own workflow RBAC
Issue #1742	Epic: RBAC/Auth review remediation (PR #1444 follow-ups)
Issue #1755	Skills RBAC UI parity / shared-module hygiene (#1727 follow-up)
Issue #1455	Runtime BFF ReBAC endpoints for Slack/Webex bots
Issue #1454	Refactor DA runtime auth surface
Issue #1731	Duplicate OpenFGA `can_use` on DA chat path
Issue #1734–#1738	RAG thin PEP / Mongo in hot path

1. Executive summary

CAIPE has reusable RBAC modules on the BFF (Next.js UI server) and CI guards for new resource types, but authorization is still implemented inconsistently across surfaces:

Dynamic Agents (DA) sometimes reads MongoDB for UI-domain data (e.g. workflow_configs, teams) and duplicates OpenFGA checks the BFF already performs.
RAG still uses a monolithic server/rbac.py that mixes token validation, legacy roles, OpenFGA, and MongoDB — a different pattern than agents/KB/MCP on the BFF.
Public APIs and errors expose OpenFGA-specific concepts (agent#use, pdp_denied, tuple strings), coupling integrators and UI copy to the current PDP.
HTTP authorization is fragmented: admin-only rebac/check, domain-specific access-check routes, and no generic runtime evaluate API for service callers (bots, DA).

After testing 0.5.9 and reviewing PR #1751, the team wants to:

Stop creating anti-patterns (DA → Mongo for workflows; one-off RBAC per resource).
Enforce shared module usage going forward (linters, hygiene).
Make the authorization API PDP-agnostic so OpenFGA is an implementation detail, not a user-facing contract.
Centralize policy decisions on the BFF where platform domain data lives (workflows, teams, visibility).

This document captures the problem statement, current state, stakeholder input, and implementation options for a follow-up spec.

2. Problem statement

2.1 Core problem

Authorization logic and platform domain data are scattered across BFF, Dynamic Agents, RAG, and Slack bot, with multiple parallel patterns for the same decision (user X may use resource Y). This causes:

Architectural drift — reviewers repeatedly flag DA/RAG reading Mongo or re-implementing visibility rules that belong on the BFF.
Duplicate PDP hops — BFF checks OpenFGA, then DA checks OpenFGA again (latency, availability, confusion about source of truth).
Inconsistent developer experience — agents use shareable-resource.ts + resource-authz.ts; RAG uses rbac.py; workflows added DA-side Mongo reads in PR #1751.
Leaky abstraction — API consumers see OpenFGA relation names and PDP jargon instead of stable product-level denial reasons.
Weak enforcement — CI guarantees new types are registered and default-deny tested, but does not force routes/services to use shared modules vs ad-hoc checks.

2.2 Triggering incidents (this conversation)

A. PR #1751 — workflow RBAC + DA delegation

PR #1751 adds:

BFF: workflow visibility, run access gates, team slug normalization, unsaved UX, agent-access modal on save.
DA: workflow_execution_authz.py — reads workflow_configs and teams from Mongo; allows agent#can_use bypass when workflow_config_id is passed on chat requests.

Reviewer feedback (@subbaksh, r3362688683):

Workflows are purely a UI server concept. The UI invokes agents one step at a time. DA usually has no clue whether it's a workflow or a user chatting. For invocation — the DA server just has a tool (with workflow_id param) that it calls to invoke a workflow. The RBAC for invocation should be in UI server no?

Assessment: The reviewer is aligned with epic #1742. Workflow lifecycle auth already exists on the BFF (/api/workflow-runs, workflow-config-rebac.ts). DA Mongo reads are an anti-pattern; the bypass exists only because the workflow engine forwards the user JWT to DA and DA enforces agent#can_use independently.

B. Reusable modules — one path did not use them

Shared modules were introduced for unified shareable resource RBAC (spec 2026-06-03-unified-shareable-resource-rbac):

Write path: ui/src/lib/rbac/shareable-resource.ts → openfga-owned-resources.ts (reconcileShareableResource)
Read path: ui/src/lib/rbac/resource-authz.ts (requireResourcePermission, filterResourcesByPermission)

Skills team sharing was fixed in #1729 to use reconcileShareableResource. Follow-up #1755 tracks optional UI DRY (useShareableTeams hook) and documents explicit non-goals (skills as owner-team resources).

Gap: Not every feature path uses the shared stack. PR #1751’s DA workflow auth is the latest example of bypassing BFF modules entirely.

C. Team discussion (Kevin Kantesaria, 2026-06-05)

Paraphrased Slack thread:

Speaker	Point
Sri	Created #1755 to track cleanup after reusable modules; one instance did not use them.
Kevin	Push cleanup soon so we don’t keep creating anti-patterns; unsure how to enforce going forward.
Sri	Added basic RBAC enforcement for new resources; need linters and hygiene.
Kevin	Does that enforcement use the modules? Original problem: RAG and agents used different RBAC setups — want to avoid that.
Kevin / Sri	Make OpenFGA less obvious to end users; API should be agnostic if we switch PDP.

2.3 Goals (desired end state)

Single policy ownership — BFF owns platform domain data (workflows, teams, visibility) and composes FGA + product rules.
Single module stack — application code uses shareable-resource / resource-authz (or a thin facade), not raw openfga.ts or service-local Mongo policy.
PDP-agnostic public contract — HTTP and user-facing errors use { resource, action, reason }, not can_use / agent#use / pdp_denied.
Thin PEPs downstream — DA, RAG, bots either trust BFF decisions via API or verify signed claims; they do not read workflow_configs from Mongo for authz.
Enforceable in CI — import boundaries, create-path linter extensions, no new anti-patterns without explicit spec exception.

2.4 Non-goals (for initial spec)

Replacing OpenFGA as the relationship store (unless a separate strategic decision).
Removing all OpenFGA checks from DA in v1 (tracked as #1731; depends on #1458 dual-gate documentation).
Migrating skills to owner-team model (#1708) — separate from this problem.

3. Current state (as of 2026-06-05, branch `prebuild/feat/workflow-rbac-visibility-unsaved`)

3.1 BFF authorization layers

┌─────────────────────────────────────────────────────────────────┐
│  BFF route handler (ui/src/app/api/**/route.ts)                   │
└────────────────────────────┬────────────────────────────────────┘
                             │
         ┌───────────────────┼───────────────────┐
         ▼                   ▼                   ▼
  requireRbacPermission   requireResourcePermission   Domain helpers
  (api-middleware.ts)     (resource-authz.ts)         (workflow-config-rebac,
   org/capability gates                             slack-channel-rebac, …)
         │                   │                   │
         └───────────────────┼───────────────────┘
                             ▼
              checkOpenFgaTuple / checkUniversalRebacRelationship
                             (openfga.ts)
                             ▼
                        OpenFGA HTTP API

Key modules

Module	Path	Responsibility
OpenFGA transport	`ui/src/lib/rbac/openfga.ts`	`checkOpenFgaTuple`, tuple read/write/reconcile
Resource ReBAC	`ui/src/lib/rbac/resource-authz.ts`	`requireResourcePermission`, `filterResourcesByPermission`, `requireAgentPermission`, `requireSkillPermission`
Agent use (chat proxy)	`ui/src/lib/rbac/openfga-agent-authz.ts`	`requireAgentUsePermission` — used by `/api/v1/chat/*`
Agent use (audit path)	`ui/src/lib/rbac/pdp-shared.ts`	`evaluateAgentAccess` — direct + team-union with reason codes
Shareable write	`ui/src/lib/rbac/shareable-resource.ts`	`handleShareableResourceWrite`, `resolveShareableOwnershipWrite`
FGA reconcile	`ui/src/lib/rbac/openfga-owned-resources.ts`	`reconcileShareableResource`
Workflow policy	`ui/src/lib/rbac/workflow-config-rebac.ts`	Visibility, `workflowDelegatesAgentUse`, `reconcileWorkflowConfigAccess` (uses `reconcileShareableResource`)
Access explain (admin)	`ui/src/lib/rbac/access-explainer.ts`	`explainAccess` + Mongo provenance
Type registry	`ui/src/types/rbac-universal.ts`	`UniversalRebacResourceType`, `UniversalRebacRelationship`
Enforcement manifest	`ui/src/lib/rbac/fga-enforcement-manifest.ts`	Per-type `rebac_enforced` / surfaces for auditors

HTTP APIs that expose authorization today

Endpoint	Auth gate	Purpose	Runtime-safe?
`POST /api/admin/rebac/check`	`admin_ui#view`	Explain access (`explainAccess`)	No — admin only
`POST /api/integrations/slack/.../access-check`	Session + `slack_channel#read`	`checkSlackChannelAccess`	Partial — bot path evolving
`POST /api/integrations/webex/.../access-check`	Same pattern	Webex space access	Partial
`POST /api/admin/slack/.../access-check`	Admin	Diagnostics	No
`POST /api/admin/openfga/relationship`	Admin	Tuple grant/revoke	No (write)
`POST /api/workflow-runs`	`assertCanExecuteWorkflowRunsForConfigId`	Start run	Yes — domain API
`POST /api/workflow-configs/check-agent-access`	`withAuth`	Pre-save agent gap check (PR #1751)	Yes — BFF-only
DA `WorkflowApiClient` → `/api/workflow-runs`	OAuth2 client credentials	Workflow tools from agents	Yes — correct pattern

There is no generic POST /api/runtime/authorize today.

Spec draft exists: docs/docs/specs/2026-05-11-identity-group-rebac/contracts/rebac-policy-api.md (POST /api/admin/rebac/check with UniversalRebacRelationship).

3.2 CI enforcement (FGA coverage guarantee, spec 2026-06-04)

Layer	Artifact	What it enforces
1	`fga-type-coverage.test.ts`	OpenFGA model ≡ type registry ≡ chart
2	`fga-enforcement-manifest.test.ts`	Every type classified; enforced surfaces exist on disk
3	`scripts/validate-fga-create-paths.py`	Ownable types wire `reconcile*Relationships` from production routes
4	`default-deny-coverage.test.ts`	New types auto-covered for deny-by-default

Gap: Layers 1–4 do not require routes to call shareable-resource / resource-authz vs ad-hoc checkOpenFgaTuple. They do not block DA/RAG Mongo reads for authz.

Location	Mongo / policy	Issue
`dynamic_agents/auth/workflow_execution_authz.py`	Reads `workflow_configs`, `teams`	Duplicates BFF visibility; reviewer rejected
`dynamic_agents/auth/openfga_authz.py`	OpenFGA + workflow bypass	Third auth layer on chat path
`dynamic_agents/services/agent_runtime.py`	Reads `workflow_configs` for builtin tool validation	Should use BFF resolve API
`dynamic_agents/services/builtin_tools.py`	`WorkflowApiClient` → BFF	Correct pattern for invocation

3.4 RAG vs agents — divergent stacks

Concern	Agents / BFF resources	RAG server
Read/use gate	`resource-authz.ts` → OpenFGA	`server/rbac.py` (mixed)
Share/write	`shareable-resource.ts`	Partial (KB sharing route on BFF); ingestor paths differ
Mongo in authz hot path	Mostly moved to BFF	Still present (#1736)
Target architecture	Unified shareable resource RBAC	Thin PEP plan

Kevin’s concern: “RAG and agents used different RBAC setups” — still true at the Python RAG layer.

3.5 PR #1751 commit inventory (relevant)

Commit / area	BFF-aligned?	Notes
Workflow run access (`workflow-run-access.ts`)	Yes
`workflow-config-rebac.ts` + `reconcileShareableResource`	Yes
Unsaved UX, team slugs, `TeamMultiPicker`	Yes
`check-agent-access` + `WorkflowAgentAccessModal`	Yes	Pre-grant agent tuples on save
`workflow_execution_authz.py` on DA	No	Remove or defer to epic
`workflow_config_id` on DA chat body	No	Remove with above

3.6 User-visible OpenFGA leakage (examples)

From openfga-agent-authz.ts and resource-authz.ts:

{
  "success": false,
  "error": "Permission denied",
  "code": "agent#use",
  "reason": "pdp_denied",
  "action": "contact_admin"
}

Admin explain responses include FGA tuple strings. Audit events use pdp: "openfga". These are appropriate for operators, not for generic API consumers or end-user copy.

4. Architectural principles (agreed direction)

Workflows are a BFF concern — config, visibility, run orchestration, delegation checks live in UI server + workflow-engine.ts.
DA invokes agents; DA workflow tools invoke workflows via BFF — WorkflowApiClient is the model.
OpenFGA is the current PDP for relationships — but callers should not depend on FGA tuple shape.
Defense-in-depth is optional and explicit — duplicate DA OpenFGA (#1731) must be a documented choice (#1458), not accidental duplication.
Epic #1742 phases — runtime BFF endpoints (#1455) before bots/DA drop Mongo.

5. Options (detailed)

Option 1 — In-process Authz facade (BFF library)

Summary: Introduce ui/src/lib/authz/ as the only import surface for application authorization. OpenFGA moves behind a PolicyEngine interface.

Proposed structure

ui/src/lib/authz/
  index.ts                 # authorize(), filterAccessible(), reconcileShares()
  types.ts                 # Subject, Resource, Action, Decision (PDP-agnostic)
  errors.ts                # PERMISSION_DENIED, AUTHZ_UNAVAILABLE
  engines/
    openfga-adapter.ts     # implements PolicyEngine today
  domains/
    workflow.ts            # workflowDelegatesAgentUse, visibility (Mongo product policy)
    slack-channel.ts
    webex-space.ts

Public API (sketch)

interface AuthorizeRequest {
  subject: { type: "user" | "service_account"; id: string };
  resource: { type: UniversalRebacResourceType; id: string };
  action: ResourcePermissionAction;
  context?: Record<string, unknown>; // workflow_id, channel_id, etc.
}

interface AuthorizeDecision {
  allowed: boolean;
  reason: "OK" | "NO_CAPABILITY" | "AUTHZ_UNAVAILABLE" | "INVALID_REQUEST";
  action?: "sign_in" | "contact_admin" | "retry";
  // Admin-only extension:
  debug?: { engine: string; path?: string };
}

async function authorize(req: AuthorizeRequest): Promise<AuthorizeDecision>;
async function authorizeOrThrow(req: AuthorizeRequest): Promise<void>;

Migration

Wrap existing resource-authz.ts, pdp-shared.ts, domain modules.
Deprecate direct checkOpenFgaTuple imports outside lib/authz/**.
Map existing ApiError codes to PDP-agnostic envelope.

Pros

No new network hop; fits existing Next.js deployment.
Clear import boundary for ESLint.
Enables future non-OpenFGA engine without route changes.

Cons

Does not alone fix DA/RAG — they don’t run inside BFF.
Requires disciplined migration of ~dozens of route call sites.

Enforcement

// eslint no-restricted-imports
"@/lib/rbac/openfga" → only from "@/lib/authz/**"

Option 2 — Runtime HTTP authorize API (BFF routes)

Summary: Expose authorization decisions as HTTP for service callers (Slack bot, Webex bot, Dynamic Agents). Implements #1455 and generalizes access-check routes.

Proposed endpoint

POST /api/runtime/authorize
Authorization: Bearer <user-jwt | service-account-token>
Content-Type: application/json

{
  "subject": { "type": "user", "id": "<oidc-sub>" },
  "resource": { "type": "agent", "id": "platform-engineer" },
  "action": "use",
  "context": {
    "workflow_config_id": "wf-optional-hint"
  }
}

Response (allow):

{
  "allowed": true,
  "reason": "OK"
}

Response (deny):

{
  "allowed": false,
  "reason": "NO_CAPABILITY",
  "action": "contact_admin"
}

Response (PDP down):

{
  "allowed": false,
  "reason": "AUTHZ_UNAVAILABLE",
  "action": "retry"
}

Domain-specific endpoints (alternative/complement)

Instead of one generic endpoint, extend the access-check pattern:

Endpoint	Caller	Checks
`POST /api/runtime/authorize`	Generic	FGA + optional context handlers
`POST /api/runtime/workflows/authorize-run`	DA, internal	Workflow visibility + step agents
`POST /api/runtime/workflows/resolve-configs`	DA	Replace Mongo `workflow_configs` read
Existing Slack/Webex `access-check`	Bots	Channel/space scoped

Authentication model

User bearer — subject must match token sub (or OBO chain).
Service account — service_account:<client_id> in FGA; may evaluate on behalf of user when context.delegated_subject is present and bot OBO is valid (needs threat model).

Pros

DA/RAG/bots stop reading Mongo for policy.
Single HTTP contract for integrators.
Aligns with epic #1742 prerequisite #1455.

Cons

Latency and availability — every DA chat could add BFF round-trip unless cached or combined with Option 3.
Duplicates #1731 debate if DA still checks OpenFGA after BFF says allow.

Implementation notes

Implement handlers by delegating to Option 1 facade internally.
Admin POST /api/admin/rebac/check remains explain-only with richer debug payload.

Option 3 — Signed authorization claims (PEP-only downstream)

Summary: BFF is the sole PDP. After authorize(), BFF mints a short-lived signed claim (JWT or HMAC macaroon) attached to downstream requests. DA/RAG verify signature + expiry only.

Claim sketch

{
  "iss": "caipe-bff",
  "sub": "user:<oidc-sub>",
  "resource": "agent:platform-engineer",
  "action": "use",
  "iat": 1710000000,
  "exp": 1710000060,
  "jti": "unique-id"
}

DA require_agent_use_permission becomes require_authorization_claim — no OpenFGA client in DA.

Pros

Eliminates duplicate OpenFGA in DA (#1731).
True PDP agnostic at service boundary.
Workflow delegation encoded in claim issuance on BFF (no workflow_config_id on DA).

Cons

Largest engineering cost: key management, revocation, clock skew, claim replay.
Requires #1458 dual-gate retirement plan.
Workflow engine currently forwards user JWT — would need claim issuance at run start and per step.

When to choose

Phase 3+ of #1742, after runtime BFF API proves correct.

Option 4 — Enforcement-first (linters + manifest, minimal API change)

Summary: Keep current modules; block anti-patterns in CI without redesigning HTTP APIs yet.

Proposed guards

Guard	Mechanism	Blocks
Import boundary	ESLint `no-restricted-imports`	`checkOpenFgaTuple` outside `lib/rbac/` or `lib/authz/`
Service boundary	`scripts/validate-authz-boundaries.py`	`workflow_configs` / `teams` reads in `dynamic_agents/**` for authz
Route coverage	Extend `validate-fga-create-paths.py` or new script	New routes must import from `resource-authz` or `authz` facade
Manifest update	`fga-enforcement-manifest.ts`	Add surfaces for `task` (workflows), runtime authorize routes
PR template	Checkbox	“Uses shareable-resource / resource-authz / domain wrapper”

Pros

Fastest win; addresses Kevin’s “don’t create more anti-patterns.”
Low risk; parallel to other options.

Cons

Does not fix RAG rbac.py or user-visible OpenFGA strings.
Does not give DA a supported API without Option 2.

Option 5 — Universal relationship contract (spec-only standardization)

Summary: Standardize all checks on UniversalRebacRelationship from rbac-universal.ts — already in rebac-policy-api.md.

Request shape

{
  "subject": { "type": "user", "id": "user-123", "relation": "member" },
  "action": "use",
  "resource": { "type": "agent", "id": "platform-engineer" },
  "context": { "slack_channel": "C123" }
}

Mapping

Internal: checkUniversalRebacRelationship(relationship) in openfga.ts.
action: "use" maps to FGA relation can_use inside adapter — not exposed to clients.

Pros

One vocabulary for docs, admin UI, runtime API, tests.
Natural fit for Option 1 + Option 2.

Cons

action names still mirror current FGA model; true PDP swap requires adapter mapping table per engine.

6. PDP-agnostic public surface (cross-cutting)

Regardless of option, define a stable external contract:

6.1 Error envelope migration

Current	Target (public)	Admin/debug only
`code: "agent#use"`	`code: "PERMISSION_DENIED"`	`detail: "agent#use"`
`reason: "pdp_denied"`	`reason: "NO_CAPABILITY"`
`reason: "pdp_unavailable"`	`reason: "AUTHZ_UNAVAILABLE"`
`pdp: "openfga"` in JSON body	Omit	Keep in audit logs
FGA tuple in error message	Omit	`POST /api/admin/rebac/check`

6.2 Stable reason codes

Code	Meaning	User `action` hint
`OK`	Allowed	—
`NO_CAPABILITY`	Deny — no relationship	`contact_admin`
`NOT_SIGNED_IN`	Missing/invalid auth	`sign_in`
`AUTHZ_UNAVAILABLE`	PDP timeout/error	`retry`
`INVALID_REQUEST`	Bad resource id / malformed	`fix_request`

6.3 Two-tier API

Tier	Audience	OpenFGA visible?
Runtime	UI, bots, DA, RAG PEP	No
Admin explain	Operators, support	Yes (paths, tuples, provenance)

7. Workflow-specific resolution (PR #1751)

7.1 Correct split of responsibility

Operation	Owner	Mechanism
Create/update workflow config	BFF	`workflow-configs` routes + `reconcileWorkflowConfigAccess`
Start/list/cancel/resume run	BFF	`/api/workflow-runs` + `workflow-run-access.ts`
Pre-save agent access gaps	BFF	`check-agent-access` + grant via `/api/admin/openfga/relationship`
Orchestrate steps	BFF `workflow-engine.ts`	Calls DA with user auth headers
Agent `#can_use` at DA	DA OpenFGA only	No workflow Mongo — team access via pre-granted tuples or BFF-only delegation
DA workflow tools	DA → BFF	`WorkflowApiClient` /api/workflow-runs

7.2 `workflowDelegatesAgentUse` (exists, unused)

Defined in workflow-config-rebac.ts:

Agent listed in workflow steps and
workflowRunAllowedByVisibility for caller

Action: Wire at POST /api/workflow-runs and optionally before each consumeAgentStream in workflow-engine.ts. Remove DA workflow_execution_authz.py.

7.3 Global workflow gap

check-agent-access detects (all users) gaps but grant loop skips them. Spec should decide: org-wide agent grants vs admin-only warning.

8. Recommended phased approach

Phase	Scope	Options	Depends on
0 — Immediate	PR #1751 cleanup	Remove DA workflow Mongo; keep BFF-only RBAC	—
1 — Hygiene	CI + docs	Option 4 + #1755 UI DRY	—
2 — Facade	BFF routes	Option 1 + Option 5 + error envelope §6	Phase 1
3 — Runtime API	Bots, DA read paths	Option 2 (#1455)	Phase 2
4 — RAG convergence	`rbac.py` thin PEP	Option 1 client in Python or Option 2 HTTP	#1734
5 — Optional	DA duplicate FGA removal	Option 3	#1458, #1731

9. Open questions for second opinion

DA execution gate: Keep OpenFGA in DA for agent#use (#1731) or move to BFF-only + signed claims (Option 3)?
Workflow delegation without agent tuples: Is pre-grant on save (modal) sufficient for team workflows, or must runtime delegation work without prior grants?
Generic vs domain HTTP APIs: One POST /api/runtime/authorize vs family of access-check endpoints?
RAG: HTTP client to BFF vs shared Python library duplicating facade logic?
Service account delegation: How should bots evaluate access for a user without admin UI perms?
Claim revocation: Required for Option 3, or acceptable 60s TTL?
Breaking API changes: Version /api/v2/ for error envelope, or migrate in place?
Workflow task type in FGA: Is Mongo visibility a permanent product layer on top of FGA, or should all run access be FGA-only eventually?

10. Acceptance criteria (draft for full spec)

No production authz path in dynamic_agents/** reads workflow_configs or teams from Mongo.
All new BFF routes use authorize() facade or documented domain wrapper — CI enforced.
Public API errors use §6 reason codes; no pdp_denied / agent#use in user-facing responses.
Runtime callers (Slack, Webex, DA workflow tools) use BFF HTTP or documented service contract — not admin rebac routes.
workflowDelegatesAgentUse wired at run start (and documented for step execution).
RAG plan references same contract as BFF (no third parallel RBAC vocabulary).
fga-enforcement-manifest.ts updated with new surfaces and rebac_enforced entries.
ADR or #1458 documents dual-gate retirement if DA OpenFGA is removed.

11. References (code & docs)

Code

ui/src/lib/rbac/shareable-resource.ts
ui/src/lib/rbac/resource-authz.ts
ui/src/lib/rbac/openfga.ts
ui/src/lib/rbac/workflow-config-rebac.ts
ui/src/lib/rbac/fga-enforcement-manifest.ts
ui/src/app/api/admin/rebac/check/route.ts
ui/src/app/api/integrations/slack/.../access-check/route.ts
ai_platform_engineering/dynamic_agents/src/dynamic_agents/auth/workflow_execution_authz.py
ai_platform_engineering/dynamic_agents/src/dynamic_agents/services/builtin_tools.py (WorkflowApiClient)
ui/src/lib/server/workflow-engine.ts
scripts/validate-fga-create-paths.py

Docs

12. Instructions for reviewing agent

Please evaluate:

Whether Option 1 + 2 + 4 is the right default combo vs Option 3 early.
Whether workflow delegation should be FGA tuples (grant on save) vs runtime BFF check only.
Gaps in threat model for runtime authorize API (service account, OBO, user impersonation).
Migration risk for error code changes on existing UI/clients.
Whether this should be one spec or split: (a) facade + enforcement, (b) runtime HTTP API, (c) RAG/DA convergence.

Return: recommended option mix, ordered task list, and any contradictions with #1742 phasing.

1. Executive summary​

2. Problem statement​

2.1 Core problem​

2.2 Triggering incidents (this conversation)​

A. PR #1751 — workflow RBAC + DA delegation​

B. Reusable modules — one path did not use them​

C. Team discussion (Kevin Kantesaria, 2026-06-05)​

2.3 Goals (desired end state)​

2.4 Non-goals (for initial spec)​

3. Current state (as of 2026-06-05, branch prebuild/feat/workflow-rbac-visibility-unsaved)​

3.1 BFF authorization layers​

Key modules​

HTTP APIs that expose authorization today​

3.2 CI enforcement (FGA coverage guarantee, spec 2026-06-04)​

3.3 Dynamic Agents — workflow-related anti-patterns​

3.4 RAG vs agents — divergent stacks​

3.5 PR #1751 commit inventory (relevant)​

3.6 User-visible OpenFGA leakage (examples)​

4. Architectural principles (agreed direction)​

5. Options (detailed)​

Option 1 — In-process Authz facade (BFF library)​

Proposed structure​

Public API (sketch)​

Migration​

Pros​

Cons​

Enforcement​

Option 2 — Runtime HTTP authorize API (BFF routes)​

Proposed endpoint​

Domain-specific endpoints (alternative/complement)​

Authentication model​

Pros​

Cons​

Implementation notes​

Option 3 — Signed authorization claims (PEP-only downstream)​

Claim sketch​

Pros​

Cons​

When to choose​

Option 4 — Enforcement-first (linters + manifest, minimal API change)​

Proposed guards​

Pros​

Cons​

Option 5 — Universal relationship contract (spec-only standardization)​

Request shape​

Mapping​

Pros​

Cons​

6. PDP-agnostic public surface (cross-cutting)​

6.1 Error envelope migration​

6.2 Stable reason codes​

6.3 Two-tier API​

7. Workflow-specific resolution (PR #1751)​

7.1 Correct split of responsibility​

7.2 workflowDelegatesAgentUse (exists, unused)​

7.3 Global workflow gap​

8. Recommended phased approach​

9. Open questions for second opinion​

10. Acceptance criteria (draft for full spec)​

11. References (code & docs)​

Code​

Docs​

12. Instructions for reviewing agent​

1. Executive summary

2. Problem statement

2.1 Core problem

2.2 Triggering incidents (this conversation)

A. PR #1751 — workflow RBAC + DA delegation

B. Reusable modules — one path did not use them

C. Team discussion (Kevin Kantesaria, 2026-06-05)

2.3 Goals (desired end state)

2.4 Non-goals (for initial spec)

3. Current state (as of 2026-06-05, branch `prebuild/feat/workflow-rbac-visibility-unsaved`)

3.1 BFF authorization layers

Key modules

HTTP APIs that expose authorization today

3.2 CI enforcement (FGA coverage guarantee, spec 2026-06-04)

3.3 Dynamic Agents — workflow-related anti-patterns

3.4 RAG vs agents — divergent stacks

3.5 PR #1751 commit inventory (relevant)

3.6 User-visible OpenFGA leakage (examples)

4. Architectural principles (agreed direction)

5. Options (detailed)

Option 1 — In-process Authz facade (BFF library)

Proposed structure

Public API (sketch)

Migration

Pros

Cons

Enforcement

Option 2 — Runtime HTTP authorize API (BFF routes)

Proposed endpoint

Domain-specific endpoints (alternative/complement)

Authentication model

Pros

Cons

Implementation notes

Option 3 — Signed authorization claims (PEP-only downstream)

Claim sketch

Pros

Cons

When to choose

Option 4 — Enforcement-first (linters + manifest, minimal API change)

Proposed guards

Pros

Cons

Option 5 — Universal relationship contract (spec-only standardization)

Request shape

Mapping

Pros

Cons

6. PDP-agnostic public surface (cross-cutting)

6.1 Error envelope migration

6.2 Stable reason codes

6.3 Two-tier API

7. Workflow-specific resolution (PR #1751)

7.1 Correct split of responsibility

7.2 `workflowDelegatesAgentUse` (exists, unused)

7.3 Global workflow gap

8. Recommended phased approach

9. Open questions for second opinion

10. Acceptance criteria (draft for full spec)

11. References (code & docs)

Code

Docs

12. Instructions for reviewing agent