Ring:Declaring Plugin Permissions

Declaring Plugin Permissions

Drive a Lerian plugin team through authoring permissions.yaml — the client-side, SEMANTIC declaration the plugin PUBLISHES at boot under the access-manager inversion. This is an interactive workflow: follow the steps in order and use AskUserQuestion to gather every choice. Do not free-hand a manifest.

Overview

Schema authority: lib-auth/v3 auth/declaration/manifest.go (>= v3.4.0-beta.1), mirrored server-side by plugin-access-manager identity/pkg/model/declaration.go. The reconciler validates this exact shape at boot and refuses a bad manifest.

THE CRITICAL INVARIANT: every (resource, action) pair in the manifest MUST exactly match a real AuthClient.Authorize(service, resource, action) guard in the plugin (lib-auth auth/middleware/middleware.go). If they diverge, authz silently breaks — the guard demands a permission the manifest never declared. Adopting the semantic standard therefore means the route guards AND the manifest move together.

• CANONICAL model (semantic, complies): br-sisbajud (internal/auth/declaration/permissions.yaml).
• ANTI-EXAMPLE (HTTP verbs, do NOT copy): midaz-fees — its guards pass Authorize("plugin-fees","estimates","post"). Legacy. Never emit verb actions.

When to use

• A plugin must publish its own permissions at boot (inversion) — authoring a new permissions.yaml.
• Fixing or migrating a plugin whose guards/manifest use HTTP verbs to the semantic standard.

Skip when

• The plugin has no Authorize(...) guards / no RBAC surface.
• Only the wiring is needed and the manifest already exists — point to authdecl.WireFromEnv and stop.

The naming standard (non-negotiable)

The schema doc comment says verbatim: "The action is SEMANTIC (create/read/update/delete), never an HTTP verb."

Domain verbs are first-class and encouraged where CRUD does not fit: rotate, trigger, justify, generate, reprocess, read_pii, request_read, request_write, receive. Keep them; do not force them into CRUD.

This is REQUIRED — and nothing downstream will hold it for you. The HTTP-verb reject was removed from BOTH lib-auth and identity (lib-auth#145 / plugin-access-manager#310), so a manifest with action: get validates, boots and publishes without complaint — several in production do exactly that. What holds the standard is review plus the check-manifest-actions Makefile guard you add in Step 9. Only delete among HTTP methods is allowed — it is also a valid semantic action. Never emit post/get/put/patch as an action.

Manifest schema (author against THIS)

Top-level YAML: service (str, REQUIRED), version (int, REQUIRED), permissions (list), roles (list), m2m (object). All bare-name rules below: the server composes the prefix — never pre-prefix.

Why deny is refused and not merely discouraged. The manifest used to accept it and the reconciler wrote it to Casdoor as a real permission, but no decision point ever applied it: every evaluator reads an effect other than allow as "did not match" and carries on, authorizing on the first allow that does match. There is no deny-wins pass. So a deny was a refusal you could read in the manifest, in review, and in the stored permission — while the runtime granted. Validation now refuses it at boot (fail-closed, lib-auth#183) and the declaration PUT answers 422 (plugin-access-manager#448).

Only the effect field is constrained. deny is still a fine ACTION name: br-sfn/services/spb declares { resource: str-emission-approvals, action: deny, effect: allow }, because approving or denying an STR emission is that domain's verb.

Composed names the server builds: permission {service}/{resource}:{action}, role {service}/{name}, group {owner}/{group}.

House style: write permissions: in FLOW style

The schema is indifferent to YAML style, the org is not. Write each permission as a single flow-mapping line and align the columns:

yamlCopy code
permissions:
  - { resource: account-types, action: read,   effect: allow, roles: [editor, contributor, viewer] }
  - { resource: account-types, action: create, effect: allow, roles: [editor, contributor] }
  - { resource: account-types, action: delete, effect: allow, roles: [editor] }

not the five-line block form. This is what the existing manifests do — midaz, tracer, plugin-access-manager, streaming-hub, reporter, lender, br-sisbajud, billing-worker and every br-sfn/services/* — and a real manifest is 20-100 pairs, where flow gives one readable line and one clean diff hunk per permission instead of five. Keep roles: and m2m: in block style, as those same manifests do.

See template.permissions.yaml in this folder for a compact, valid, commented example.

Interactive workflow — follow in order

Step 1 — Determine service

Grep the plugin for its product/slug constant before asking:

bashCopy code
grep -rn -iE "ProductName|ApplicationName|ModuleName|feesApplicationName|Slug\s*=" \
  --include="*.go" <plugin-root> | head

Propose the found constant as the default. Confirm with AskUserQuestion, and warn: it MUST equal the M2M app DisplayName and be the first arg of every Authorize(...) call (BOLA). Keep any plugin- prefix.

Step 2 — Discover the REAL authorization surface

Enumerate what the code actually enforces today:

bashCopy code
grep -rn "\.Authorize(" --include="*.go" <plugin-root>

Extract each (service, resource, action) triple from the guard chains (and route tables). Present the full list. This list is ground truth — the manifest must cover exactly these pairs (Step 8 re-checks).

Step 3 — Normalize actions to the SEMANTIC standard

For EACH discovered action:

• If it is an HTTP verb (post/get/put/patch/delete-method) → propose the semantic equivalent from the table AND flag that the route guard must change too (guard + manifest move together — this is a code change, not just YAML).
• If it is already semantic → keep it.
• If CRUD does not fit → offer the relevant domain verb.

Use AskUserQuestion per action/resource to let the engineer confirm or rename, offering sensible options (e.g. for delete: delete vs remove). Record the final semantic (resource, action) set. If any guard currently passes a verb, list those guards explicitly as follow-up code edits the team owns — do not silently emit verb actions to make the mismatch "go away".

Step 4 — Roles and group grants

Ask (via AskUserQuestion):

• Which roles exist? Default viewer + editor; offer domain roles (operator/investigator/compliance-officer/…) as in br-sisbajud.
• Which BARE group(s) grant each role? (Groups are group-only, bare.)
• Grant mapping — offer this DEFAULT, let them override: read/query actions → viewer + editor; mutating actions → editor-only.

Every permission MUST list >= 1 role, and every listed role MUST be declared.

Step 5 — M2M contract

Ask:

• exposed: is this plugin an M2M target (callable by other plugins)?
• needs: which services does it CALL via M2M (e.g. midaz)? Omit the block if neither applies.

Step 6 — Emit permissions.yaml

Write the manifest to the plugin's declaration dir (mirror br-sisbajud: internal/auth/declaration/permissions.yaml). Bare resources/groups/roles, semantic actions, valid effects, >= 1 role per permission, permissions: entries in FLOW style (see "House style" above). Then show the wiring the plugin adds (manifest authoring is the focus; this is the glue):

goCopy code
//go:embed permissions.yaml
var Manifest []byte

// ...at startup (authdecl = github.com/LerianStudio/lib-auth/v3/auth/declaration, >= v3.4.0-beta.1):
stop, err := authdecl.WireFromEnv(ctx, authdecl.WireInput{
    Slug:     <service>,   // MUST equal manifest.service (BOLA)
    Manifest: Manifest,
    Logger:   logger,
})

Deployment sets the FIXED env contract (default OFF, fail-open). A NEW adopter creates these vars with the canonical IDP_ names from the start. The four RI/D7-declaration vars carry the product-wide IDP_ prefix (identity provider, lib-auth #4232 — shared across every plugin, NOT a per-plugin prefix): IDP_DECLARATION_ENABLED, IDP_HOST, IDP_M2M_CLIENT_ID, IDP_M2M_CLIENT_SECRET, plus the token-minter vars PLUGIN_AUTH_ENABLED, PLUGIN_AUTH_HOST (out of scope for #4232, unchanged).

The IDP_ names require lib-auth ≥ v3.4.0-beta.6 (the release that carries #4232). This is a LATER threshold than the >= v3.4.0-beta.1 manifest-schema pin above. For ONE release after #4232 the old names (DECLARATION_ENABLED, PLUGIN_IDENTITY_HOST, M2M_CLIENT_ID, M2M_CLIENT_SECRET) still work as deprecated aliases (canonical IDP_ wins; WireFromEnv WARNs when only the alias is set), so a plugin pinned to an older lib-auth keeps booting — migrate to the IDP_ names before the following release drops the aliases.

Step 7 — Validate

Run structural checks against every rule below, and if a Go toolchain + lib-auth (>= v3.4.0-beta.1) are available, verify against the REAL validator (parse+Validate, zero network) with a tiny throwaway program:

goCopy code
// authdecl "github.com/LerianStudio/lib-auth/v3/auth/declaration"
// _, err := authdecl.New(authdecl.Config{Slug: "<service>", Manifest: raw, /* IdentityAddr, ClientID/Secret dummy */})
// New parses + Validate()s the manifest eagerly and enforces slug==service (BOLA); a
// non-manifest error means the manifest itself is structurally valid.

Or a YAML lint + this checklist. Validation rules (all aggregated at boot):

• service non-empty and not ./...
• version >= 1.
• each action is non-empty. The SEMANTIC standard above is a CONVENTION, not a boot check: the HTTP-verb reject was removed from both lib-auth and identity (lib-auth#145 / plugin-access-manager#310), so a manifest with action: get publishes without complaint. Hold the standard in review and with the Makefile guard below — nothing downstream will hold it for you.
• each permission: non-empty resource and action; effect is allow;

  = 1 role; every role reference is a DECLARED role.

• no duplicate composed permission {service}/{resource}:{action}.
• no duplicate composed role {service}/{name}.
• no Casdoor-safe-name collision: chars / ? : # & % = + ; and whitespace collapse to - (lossy), so two different names can collide — and a name that derives to empty is rejected.

Step 8 — Alignment gate (BLOCKING)

Re-confirm every declared (resource, action) maps to a real Authorize(...) call (Step 2 list) and vice-versa. Enumerate ANY mismatch as blocking:

• guard exists, manifest missing → add the permission.
• manifest declares a pair no guard uses → remove it or add the guard.
• guard still passes an HTTP verb → the team MUST update the guard to the semantic action so both sides use it (do not "fix" it by declaring the verb).

Do not consider the manifest done while any mismatch remains.

Step 9 — Scaffold the durable CI guard (Makefile)

The alignment gate above is a one-time check. Lock the semantic standard in so a future edit that reintroduces an HTTP verb FAILS the build. For the MANIFEST's action names there is exactly ONE automated layer, and it is the one you add here. Guard alignment is NOT automated at all: the target below reads $(MANIFEST) and nothing else, so a guard still passing get against a manifest declaring read is caught by Steps 3 and 8 and by review — nowhere else.

• Boot-time (lib-auth): none. The validator used to reject post/get/put/patch, and that reject was removed from lib-auth and identity (lib-auth#145 / plugin-access-manager#310). A manifest with a verb action starts and publishes fine — nothing catches the regression at runtime.
• CI (Makefile): the check-manifest-actions guard below. Cheap, lives in the plugin's own repo, and is the only thing that fails a build on a verb reintroduced in the manifest.

Check the plugin's Makefile for the existing check-* convention (most Lerian plugins wire check-tests, check-migrations, … into a ci:/check aggregate — grep ^check- and ^ci:). Add a check-manifest-actions target matching that idiom and wire it into the aggregate:

makefileCopy code
MANIFEST ?= internal/auth/declaration/permissions.yaml

.PHONY: check-manifest-actions
# Fail if the manifest is missing/unreadable, or if it uses HTTP-verb actions.
# 'delete' is allowed (also a valid semantic action). This is the ONLY automated
# check for the semantic standard — lib-auth no longer rejects verbs at boot. It
# reads the manifest only: a mismatched Authorize() guard is Step 8's job.
# The pattern matches BOTH styles: flow `- { ..., action: post, ... }` (the house
# style) and block `action: post` on its own line, with or without quotes and with
# or without a trailing `# comment`. It deliberately does NOT match a `#` comment
# line that merely mentions an action, and does not fire on a resource or a longer
# action that starts with a verb (`repost`, `getaway`, `read_pii`).
check-manifest-actions:
	@test -r "$(MANIFEST)" || { echo "ERROR: manifest not found or unreadable: $(MANIFEST)"; exit 1; }
	@echo "Checking manifest actions are semantic (not HTTP verbs)..."
	@if grep -inE '^([[:space:]]*-[[:space:]]*\{(.*[{,])?)?[[:space:]]*action:[[:space:]]*["'\'']?(post|get|put|patch)["'\'']?[[:space:]]*([,}#]|$$)' "$(MANIFEST)"; then \
		echo "ERROR: HTTP-verb action in $(MANIFEST) — use a SEMANTIC action (create/read/update/delete or a domain verb). 'delete' is allowed."; \
		exit 1; \
	fi
	@echo "OK: manifest actions are semantic."

Add check-manifest-actions to the ci:/check: prerequisite list (and .PHONY). If the plugin has no check-*/ci idiom, still add the target and call it where tests run. Confirm it FAILS on a seeded action: post — seeded in the SAME style the manifest uses — and PASSES on the real manifest before finishing.

Step 10 — Bump the shared-workflows CI pin

The org shared CI (LerianStudio/github-actions-shared-workflows, reusable go-pr-validation.yml) now carries a NON-BLOCKING permission-manifest-nudge that reminds any lib-auth repo still missing a permissions.yaml. You are already touching this repo — bump its pin so the pipeline is current.

• Find the consumer pins: grep -rn 'LerianStudio/github-actions-shared-workflows' .github/workflows. Expect exact-tag uses: …@vX.Y.Z on go-pr-validation.yml / go-release.yml / routine.yml. Leave any …@v1 major-float pins as-is.
• Resolve the latest release tag: gh release view --repo LerianStudio/github-actions-shared-workflows --json tagName -q .tagName (or gh api repos/LerianStudio/github-actions-shared-workflows/releases/latest -q .tag_name). It must be >= the release that introduced permission-manifest-nudge.
• Bump EVERY exact-tag shared-workflows pin in .github/workflows/*.yml to that tag, keeping all of them on the SAME version. Do not touch unrelated uses: lines.

This is hygiene, not a gate: for THIS repo — which now declares a manifest — the nudge reports "compliant" and posts nothing. The bump only keeps the shared pipeline current. Confirm the target tag exists before writing, and preserve the pin format.

Red Flags — STOP

• An action is post/get/put/patch/delete-the-method → it is an HTTP verb.
• A resource, role, or group carries a {service}/ or {owner}/ prefix → it will double-prefix; write it BARE.
• A granted_to entry has a user: key → there is no user grantee; groups only.
• A permission lists zero roles, or a role not in roles: → validation fails.
• You are emitting a verb action to sidestep a guard mismatch → fix the guard instead.
• service differs from the M2M app DisplayName / the Authorize 1st arg → BOLA break.

All of these mean: stop and correct before writing/finishing the manifest.

Anti-Rationalization