test: match docs request path exactly and assert proxy headers stay off upstream

Two CodeRabbit findings on the previous commit:

1. The docs-request selectors used strings.Contains, so a wrong API
   prefix or version (e.g. docs_ai/v2) could still match while the mock
   answers any path with a generic success. Select by the full exact
   path /open-apis/docs_ai/v1/documents/<doc>/fetch instead.

2. The upstream assertion proved token injection but would not notice
   the forward leaking sidecar wire-protocol headers. Assert every
   X-Lark-Proxy-* header and X-Lark-Body-Sha256 is absent from the
   forwarded docs request.

.github/workflows/ci.yml

@@ -47,6 +47,34 @@ jobs:
             exit 1
           fi
 
+  plugin-integration:
+    needs: fast-gate
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@93cb6efe18208431cddfb8368fd83d5badbf9bfd # v5
+        with:
+          persist-credentials: false
+      - uses: actions/setup-go@4a3601121dd01d1626a1e23e37211e3254c1c06c # v6
+        with:
+          go-version-file: go.mod
+      # No fetch_meta: the git-archive clean tree must embed only the
+      # committed meta_data stub (reproduces the bare-module customer state).
+      - name: Run plugin-integration L4 tests
+        run: go test -count=1 -timeout=15m ./tests/plugin_e2e/...
+
+  sidecar-integration:
+    needs: fast-gate
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@93cb6efe18208431cddfb8368fd83d5badbf9bfd # v5
+        with:
+          persist-credentials: false
+      - uses: actions/setup-go@4a3601121dd01d1626a1e23e37211e3254c1c06c # v6
+        with:
+          go-version-file: go.mod
+      - name: Run sidecar tag build + HMAC round-trip
+        run: make sidecar-test
+
   # ── Layer 2: Quality Gate ──────────────────────────────────────────
   unit-test:
     needs: fast-gate
@@ -263,13 +291,19 @@ jobs:
     runs-on: ubuntu-latest
     steps:
       - uses: actions/checkout@93cb6efe18208431cddfb8368fd83d5badbf9bfd # v5
+        with:
+          fetch-depth: 0
       - uses: actions/setup-go@4a3601121dd01d1626a1e23e37211e3254c1c06c # v6
         with:
           go-version-file: go.mod
       - uses: actions/setup-python@a309ff8b426b58ec0e2a45f0f869d46889d02405 # v6
         with:
           python-version: '3.x'
+      - name: Resolve CLI E2E domains
+        id: e2e_domains
+        run: node scripts/e2e_domains.js
       - name: Build lark-cli
+        if: ${{ steps.e2e_domains.outputs.mode != 'skip' }}
         run: make build
       - name: Run dry-run E2E tests
         env:
@@ -277,7 +311,28 @@ jobs:
           LARKSUITE_CLI_APP_ID: dry-run
           LARKSUITE_CLI_APP_SECRET: dry-run
           LARKSUITE_CLI_BRAND: feishu
-        run: go test -v -count=1 -timeout=5m ./tests/cli_e2e/... -run 'DryRun|Regression'
+          E2E_MODE: ${{ steps.e2e_domains.outputs.mode }}
+          E2E_REASON: ${{ steps.e2e_domains.outputs.reason }}
+          E2E_DRY_ROOT_PACKAGE: ${{ steps.e2e_domains.outputs.dry_root_package }}
+          E2E_DRY_PACKAGES: ${{ steps.e2e_domains.outputs.dry_packages }}
+        run: |
+          if [ "$E2E_MODE" = "skip" ]; then
+            echo "No dry-run CLI E2E needed: $E2E_REASON"
+            exit 0
+          fi
+          if [ -z "$E2E_DRY_ROOT_PACKAGE" ] && [ -z "$E2E_DRY_PACKAGES" ]; then
+            echo "::error::No dry-run CLI E2E packages resolved for mode $E2E_MODE"
+            exit 1
+          fi
+          echo "Dry-run CLI E2E domains: $E2E_MODE ($E2E_REASON)"
+          if [ -n "$E2E_DRY_ROOT_PACKAGE" ]; then
+            echo "Dry-run CLI E2E root package: $E2E_DRY_ROOT_PACKAGE"
+            go test -v -count=1 -timeout=5m "$E2E_DRY_ROOT_PACKAGE"
+          fi
+          if [ -n "$E2E_DRY_PACKAGES" ]; then
+            echo "Dry-run CLI E2E packages: $E2E_DRY_PACKAGES"
+            go test -v -count=1 -timeout=5m $E2E_DRY_PACKAGES -run 'DryRun|Regression'
+          fi
 
   e2e-live:
     needs: [unit-test, lint, script-test, deterministic-gate]
@@ -292,15 +347,22 @@ jobs:
       TEST_USER_ACCESS_TOKEN: ${{ secrets.TEST_USER_ACCESS_TOKEN }}
     steps:
       - uses: actions/checkout@93cb6efe18208431cddfb8368fd83d5badbf9bfd # v5
+        with:
+          fetch-depth: 0
       - uses: actions/setup-go@4a3601121dd01d1626a1e23e37211e3254c1c06c # v6
         with:
           go-version-file: go.mod
       - uses: actions/setup-python@a309ff8b426b58ec0e2a45f0f869d46889d02405 # v6
         with:
           python-version: '3.x'
+      - name: Resolve CLI E2E domains
+        id: e2e_domains
+        run: node scripts/e2e_domains.js
       - name: Build lark-cli
+        if: ${{ steps.e2e_domains.outputs.mode != 'skip' }}
         run: make build
       - name: Configure bot credentials
+        if: ${{ steps.e2e_domains.outputs.mode != 'skip' }}
         run: |
           if [ -z "$TEST_BOT1_APP_ID" ] || [ -z "$TEST_BOT1_APP_SECRET" ]; then
             echo "::error::Missing required secrets: TEST_BOT1_APP_ID / TEST_BOT1_APP_SECRET"
@@ -310,16 +372,24 @@ jobs:
       - name: Run CLI E2E tests
         env:
           LARK_CLI_BIN: ${{ github.workspace }}/lark-cli
+          E2E_MODE: ${{ steps.e2e_domains.outputs.mode }}
+          E2E_REASON: ${{ steps.e2e_domains.outputs.reason }}
+          E2E_LIVE_PACKAGES: ${{ steps.e2e_domains.outputs.live_packages }}
         run: |
-          packages=$(go list ./tests/cli_e2e/... | grep -v '^github.com/larksuite/cli/tests/cli_e2e$' | grep -v '/demo$')
+          if [ "$E2E_MODE" = "skip" ]; then
+            echo "No live CLI E2E needed: $E2E_REASON"
+            exit 0
+          fi
+          packages="$E2E_LIVE_PACKAGES"
           if [ -z "$packages" ]; then
-            echo "No CLI E2E packages to test after exclusions."
+            echo "::error::No live CLI E2E packages resolved for mode $E2E_MODE"
             exit 1
           fi
-          packages_arg=$(printf '%s\n' "$packages" | paste -sd' ' -)
-          go run gotest.tools/gotestsum@v1.12.3 --rerun-fails=2 --rerun-fails-max-failures=20 --packages="$packages_arg" --format testname --junitfile cli-e2e-report.xml -- -count=1 -v
+          echo "Live CLI E2E domains: $E2E_MODE ($E2E_REASON)"
+          echo "Live CLI E2E packages: $packages"
+          go run gotest.tools/gotestsum@v1.12.3 --rerun-fails=2 --rerun-fails-max-failures=20 --packages="$packages" --format testname --junitfile cli-e2e-report.xml -- -count=1 -v
       - name: Publish CLI E2E test report
-        if: ${{ !cancelled() }}
+        if: ${{ !cancelled() && steps.e2e_domains.outputs.mode != 'skip' }}
         uses: dorny/test-reporter@a43b3a5f7366b97d083190328d2c652e1a8b6aa2 # v3.0.0
         with:
           name: CLI E2E Tests
@@ -372,7 +442,7 @@ jobs:
   # ── Results Gate (single required check for branch protection) ─────
   results:
     if: ${{ always() }}
-    needs: [fast-gate, unit-test, lint, script-test, deterministic-gate, coverage, deadcode, e2e-dry-run, e2e-live, security, license-header]
+    needs: [fast-gate, unit-test, lint, script-test, deterministic-gate, coverage, deadcode, e2e-dry-run, e2e-live, security, license-header, plugin-integration, sidecar-integration]
     runs-on: ubuntu-latest
     steps:
       - name: Evaluate results
@@ -392,10 +462,18 @@ jobs:
           echo "| L3 | e2e-live | ${{ needs.e2e-live.result }} |" >> $GITHUB_STEP_SUMMARY
           echo "| L4 | security | ${{ needs.security.result }} |" >> $GITHUB_STEP_SUMMARY
           echo "| L4 | license-header | ${{ needs.license-header.result }} |" >> $GITHUB_STEP_SUMMARY
+          echo "| L4 | plugin-integration (observe-only) | ${{ needs.plugin-integration.result }} |" >> $GITHUB_STEP_SUMMARY
+          echo "| L4 | sidecar-integration (observe-only) | ${{ needs.sidecar-integration.result }} |" >> $GITHUB_STEP_SUMMARY
 
           # Any failure or cancellation in any job blocks the merge.
           # Legitimately skipped jobs (deadcode on push, e2e-live on fork,
           # license-header on push) are OK.
+          #
+          # plugin-integration and sidecar-integration are intentionally NOT
+          # in this loop yet: they run on every PR and their status is shown
+          # in the table above, but a failure is observe-only (non-blocking)
+          # during the initial soak. Add them back here to make them required
+          # once they have proven stable.
           FAILED=0
           for result in \
             "${{ needs.fast-gate.result }}" \

Makefile

@@ -23,7 +23,7 @@ PREFIX   ?= /usr/local
 TEST_GOARCH := $(or $(GOARCH),$(shell go env GOARCH))
 RACE_FLAG := $(if $(filter riscv64,$(TEST_GOARCH)),,-race)
 
-.PHONY: all build vet fmt-check script-test test unit-test integration-test examples-build quality-gate install uninstall clean fetch_meta gitleaks
+.PHONY: all build vet fmt-check script-test test unit-test integration-test examples-build quality-gate install uninstall clean fetch_meta gitleaks sidecar-test
 
 all: test
 
@@ -51,7 +51,7 @@ script-test:
 	bash scripts/resolve-changed-from.test.sh
 	bash scripts/ci-workflow.test.sh
 	bash scripts/semantic-review-workflow.test.sh
-	$(NODE) --test scripts/semantic-review-verify-artifact.test.js scripts/pr-quality-summary.test.js scripts/semantic-review-publish.test.js scripts/ci-quality-summary-publish.test.js
+	$(NODE) --test scripts/e2e_domains.test.js scripts/semantic-review-verify-artifact.test.js scripts/pr-quality-summary.test.js scripts/semantic-review-publish.test.js scripts/ci-quality-summary-publish.test.js
 
 # ./extension/... keeps the public plugin SDK in the default test matrix.
 unit-test: fetch_meta
@@ -105,6 +105,14 @@ uninstall:
 clean:
 	rm -f $(BINARY)
 
+# sidecar-test compiles and runs the authsidecar* build-tagged code that the
+# default CI matrix never sees (they carry //go:build tags).
+sidecar-test:
+	go build -tags authsidecar -o /dev/null .
+	go test $(RACE_FLAG) -count=1 -tags authsidecar ./extension/credential/sidecar/ ./extension/transport/sidecar/ ./internal/cmdutil/
+	go test $(RACE_FLAG) -count=1 -tags authsidecar_demo ./sidecar/server-demo/
+	go test $(RACE_FLAG) -count=1 -tags authsidecar ./tests/sidecar_e2e/
+
 # Run secret-leak checks locally before pushing.
 # Step 1: check-doc-tokens catches realistic-looking example tokens in reference
 #         docs and asks you to use _EXAMPLE_TOKEN placeholders instead.

cmd/root_integration_test.go

@@ -14,11 +14,13 @@ import (
 	"github.com/larksuite/cli/cmd/api"
 	"github.com/larksuite/cli/cmd/auth"
 	"github.com/larksuite/cli/cmd/service"
+	"github.com/larksuite/cli/internal/apicatalog"
 	"github.com/larksuite/cli/internal/build"
 	"github.com/larksuite/cli/internal/cmdutil"
 	"github.com/larksuite/cli/internal/core"
 	"github.com/larksuite/cli/internal/envvars"
 	"github.com/larksuite/cli/internal/httpmock"
+	"github.com/larksuite/cli/internal/meta"
 	"github.com/larksuite/cli/internal/output"
 	"github.com/larksuite/cli/internal/skillscheck"
 	"github.com/larksuite/cli/internal/update"
@@ -103,6 +105,11 @@ func parseTypedEnvelope(t *testing.T, stderr *bytes.Buffer) typedErrorEnvelope {
 }
 
 func buildStrictModeIntegrationRootCmd(t *testing.T, f *cmdutil.Factory) *cobra.Command {
+	t.Helper()
+	return buildStrictModeIntegrationRootCmdWithCatalog(t, f, nil)
+}
+
+func buildStrictModeIntegrationRootCmdWithCatalog(t *testing.T, f *cmdutil.Factory, catalog *apicatalog.Catalog) *cobra.Command {
 	t.Helper()
 	rootCmd := &cobra.Command{Use: "lark-cli"}
 	rootCmd.SilenceErrors = true
@@ -113,7 +120,11 @@ func buildStrictModeIntegrationRootCmd(t *testing.T, f *cmdutil.Factory) *cobra.
 	}
 	rootCmd.AddCommand(auth.NewCmdAuth(f))
 	rootCmd.AddCommand(api.NewCmdApi(f, nil))
-	service.RegisterServiceCommands(rootCmd, f)
+	if catalog != nil {
+		service.RegisterServiceCommandsFromCatalog(context.Background(), rootCmd, f, *catalog)
+	} else {
+		service.RegisterServiceCommands(rootCmd, f)
+	}
 	shortcuts.RegisterShortcuts(rootCmd, f)
 	if mode := f.ResolveStrictMode(context.Background()); mode.IsActive() {
 		pruneForStrictMode(rootCmd, mode)
@@ -121,6 +132,29 @@ func buildStrictModeIntegrationRootCmd(t *testing.T, f *cmdutil.Factory) *cobra.
 	return rootCmd
 }
 
+func strictModeFixtureCatalog() apicatalog.Catalog {
+	return apicatalog.New(apicatalog.SourceEmbedded, []meta.Service{
+		{
+			Name:        "fixture",
+			ServicePath: "/open-apis/fixture/v1",
+			Resources: map[string]meta.Resource{
+				"things": {
+					Methods: map[string]meta.Method{
+						"create": {
+							Path:         "things",
+							HTTPMethod:   "POST",
+							AccessTokens: []meta.Token{meta.TokenTenant},
+							RequestBody: map[string]meta.Field{
+								"name": {Type: "string"},
+							},
+						},
+					},
+				},
+			},
+		},
+	})
+}
+
 func newStrictModeDefaultFactory(t *testing.T, profile string, mode core.StrictMode) (*cmdutil.Factory, *bytes.Buffer, *bytes.Buffer) {
 	t.Helper()
 	t.Setenv(envvars.CliAppID, "")
@@ -355,10 +389,11 @@ func TestIntegration_StrictModeBot_ProfileOverride_ServiceExplicitUserReturnsEnv
 
 func TestIntegration_StrictModeUser_ProfileOverride_ServiceBotOnlyMethodReturnsEnvelope(t *testing.T) {
 	f, stdout, stderr := newStrictModeDefaultFactory(t, "target", core.StrictModeUser)
-	rootCmd := buildStrictModeIntegrationRootCmd(t, f)
+	catalog := strictModeFixtureCatalog()
+	rootCmd := buildStrictModeIntegrationRootCmdWithCatalog(t, f, &catalog)
 
 	code := executeRootIntegration(t, f, rootCmd, []string{
-		"im", "images", "create", "--data", `{"image_type":"message","image":"x"}`, "--dry-run",
+		"fixture", "things", "create", "--data", `{"name":"probe"}`, "--dry-run",
 	})
 
 	if code != output.ExitValidation {

scripts/ci-workflow.test.sh

@@ -215,6 +215,73 @@ if ! grep -Fq "if: \${{ $fork_safe_guard }}" <<<"$section"; then
   exit 1
 fi
 
+if ! grep -Fq "name: Resolve CLI E2E domains" <<<"$dry_run_section" ||
+   ! grep -Fq "id: e2e_domains" <<<"$dry_run_section" ||
+   ! grep -Fq "run: node scripts/e2e_domains.js" <<<"$dry_run_section"; then
+  echo "e2e-dry-run should resolve changed-file CLI E2E domains before running tests"
+  exit 1
+fi
+
+if ! grep -Fq "steps.e2e_domains.outputs.dry_packages" <<<"$dry_run_section"; then
+  echo "e2e-dry-run should use resolved dry_packages instead of always running the full suite"
+  exit 1
+fi
+
+if ! grep -Fq "E2E_REASON: \${{ steps.e2e_domains.outputs.reason }}" <<<"$dry_run_section" ||
+   ! grep -Fq 'echo "Dry-run CLI E2E domains: $E2E_MODE ($E2E_REASON)"' <<<"$dry_run_section"; then
+  echo "e2e-dry-run should pass dynamic domain output through env before shell use"
+  exit 1
+fi
+
+if ! grep -Fq "E2E_DRY_ROOT_PACKAGE: \${{ steps.e2e_domains.outputs.dry_root_package }}" <<<"$dry_run_section" ||
+   ! grep -Fq 'go test -v -count=1 -timeout=5m "$E2E_DRY_ROOT_PACKAGE"' <<<"$dry_run_section"; then
+  echo "e2e-dry-run should run the root CLI E2E harness package without the DryRun/Regression filter"
+  exit 1
+fi
+
+if ! grep -Fq "No dry-run CLI E2E needed" <<<"$dry_run_section"; then
+  echo "e2e-dry-run should explicitly skip when domain mode is skip"
+  exit 1
+fi
+
+if ! grep -Fq "name: Resolve CLI E2E domains" <<<"$section" ||
+   ! grep -Fq "id: e2e_domains" <<<"$section" ||
+   ! grep -Fq "run: node scripts/e2e_domains.js" <<<"$section"; then
+  echo "e2e-live should resolve changed-file CLI E2E domains before credentials and tests"
+  exit 1
+fi
+
+if ! grep -Fq "steps.e2e_domains.outputs.live_packages" <<<"$section"; then
+  echo "e2e-live should use resolved live_packages instead of always running the full suite"
+  exit 1
+fi
+
+if ! grep -Fq "E2E_REASON: \${{ steps.e2e_domains.outputs.reason }}" <<<"$section" ||
+   ! grep -Fq 'echo "Live CLI E2E domains: $E2E_MODE ($E2E_REASON)"' <<<"$section"; then
+  echo "e2e-live should pass dynamic domain output through env before shell use"
+  exit 1
+fi
+
+if ! awk '
+  /^      - name: Build lark-cli/ { in_step = 1 }
+  in_step && /if: \$\{\{ steps\.e2e_domains\.outputs\.mode != '\''skip'\'' \}\}/ { found = 1 }
+  in_step && /^      - name:/ && !/Build lark-cli/ { in_step = 0 }
+  END { exit found ? 0 : 1 }
+' <<<"$dry_run_section"; then
+  echo "e2e-dry-run should skip building lark-cli when domain mode is skip"
+  exit 1
+fi
+
+if ! awk '
+  /^      - name: Build lark-cli/ { in_step = 1 }
+  in_step && /if: \$\{\{ steps\.e2e_domains\.outputs\.mode != '\''skip'\'' \}\}/ { found = 1 }
+  in_step && /^      - name:/ && !/Build lark-cli/ { in_step = 0 }
+  END { exit found ? 0 : 1 }
+' <<<"$section"; then
+  echo "e2e-live should skip building lark-cli when domain mode is skip"
+  exit 1
+fi
+
 if ! grep -Fq "permissions:" <<<"$section" ||
    ! grep -Fq "contents: read" <<<"$section" ||
    ! grep -Fq "checks: write" <<<"$section"; then
@@ -237,13 +304,23 @@ if ! grep -Fq "::error::Missing required secrets: TEST_BOT1_APP_ID / TEST_BOT1_A
   exit 1
 fi
 
+if ! awk '
+  /^      - name: Configure bot credentials/ { in_step = 1 }
+  in_step && /if: \$\{\{ steps\.e2e_domains\.outputs\.mode != '\''skip'\'' \}\}/ { found = 1 }
+  in_step && /^      - name:/ && !/Configure bot credentials/ { in_step = 0 }
+  END { exit found ? 0 : 1 }
+' <<<"$section"; then
+  echo "e2e-live should only configure bot credentials when domain mode is not skip"
+  exit 1
+fi
+
 if grep -Fq "steps.live_e2e_credentials.outputs.configured" <<<"$section"; then
   echo "e2e-live build, configure, test, and report steps should not be gated by a skip-state output"
   exit 1
 fi
 
-if ! grep -Fq "if: \${{ !cancelled() }}" <<<"$section"; then
-  echo "e2e-live report step should run after attempted live tests unless the workflow is cancelled"
+if ! grep -Fq "if: \${{ !cancelled() && steps.e2e_domains.outputs.mode != 'skip' }}" <<<"$section"; then
+  echo "e2e-live report step should run after attempted live tests unless the workflow is cancelled or domain mode is skip"
   exit 1
 fi
 

scripts/domain-map.js

@@ -0,0 +1,54 @@
+// Copyright (c) 2026 Lark Technologies Pte. Ltd.
+// SPDX-License-Identifier: MIT
+
+const fs = require("node:fs");
+const path = require("node:path");
+
+const DOMAIN_MAP_PATH = path.join(__dirname, "domain-map.json");
+const domainMap = JSON.parse(fs.readFileSync(DOMAIN_MAP_PATH, "utf8"));
+
+function normalizeRepoPath(input) {
+  return String(input || "").trim().replace(/\\/g, "/").replace(/^\.\//, "").toLowerCase();
+}
+
+const pathMappingsBySpecificity = (domainMap.pathMappings || [])
+  .map((entry) => ({ ...entry, prefix: normalizeRepoPath(entry.prefix) }))
+  .sort((a, b) => b.prefix.length - a.prefix.length);
+
+function findPathMapping(filePath) {
+  const normalized = normalizeRepoPath(filePath);
+  return pathMappingsBySpecificity.find((entry) => normalized.startsWith(entry.prefix));
+}
+
+function labelDomainsForPath(filePath) {
+  const mapping = findPathMapping(filePath);
+  return mapping ? [...(mapping.labelDomains || [])] : [];
+}
+
+function e2eDomainsForPath(filePath) {
+  const mapping = findPathMapping(filePath);
+  return mapping ? [...(mapping.e2eDomains || [])] : [];
+}
+
+function matchesFullFallback(filePath) {
+  const normalized = normalizeRepoPath(filePath);
+  return (domainMap.fullFallbackPrefixes || []).some((prefix) => normalized.startsWith(prefix));
+}
+
+function isSkippablePath(filePath) {
+  const normalized = normalizeRepoPath(filePath);
+  const basename = path.posix.basename(normalized);
+  return (domainMap.skipPrefixes || []).some((prefix) => normalized.startsWith(prefix))
+    || (domainMap.skipSuffixes || []).some((suffix) => normalized.endsWith(suffix))
+    || (domainMap.skipFilenames || []).includes(basename);
+}
+
+module.exports = {
+  domainMap,
+  e2eDomainsForPath,
+  findPathMapping,
+  isSkippablePath,
+  labelDomainsForPath,
+  matchesFullFallback,
+  normalizeRepoPath,
+};

scripts/domain-map.json

@@ -0,0 +1,71 @@
+{
+  "pathMappings": [
+    { "prefix": "shortcuts/im/", "labelDomains": ["im"], "e2eDomains": ["im"] },
+    { "prefix": "shortcuts/vc/", "labelDomains": ["vc"], "e2eDomains": ["vc"] },
+    { "prefix": "shortcuts/calendar/", "labelDomains": ["calendar"], "e2eDomains": ["calendar"] },
+    { "prefix": "shortcuts/doc/", "labelDomains": ["ccm"], "e2eDomains": ["docs"] },
+    { "prefix": "shortcuts/sheets/", "labelDomains": ["ccm"], "e2eDomains": ["sheets"] },
+    { "prefix": "shortcuts/drive/", "labelDomains": ["ccm"], "e2eDomains": ["drive"] },
+    { "prefix": "shortcuts/wiki/", "labelDomains": ["ccm"], "e2eDomains": ["wiki"] },
+    { "prefix": "shortcuts/base/", "labelDomains": ["base"], "e2eDomains": ["base"] },
+    { "prefix": "shortcuts/mail/", "labelDomains": ["mail"], "e2eDomains": ["mail"] },
+    { "prefix": "shortcuts/task/", "labelDomains": ["task"], "e2eDomains": ["task"] },
+    { "prefix": "shortcuts/contact/", "labelDomains": ["contact"], "e2eDomains": ["contact"] },
+    { "prefix": "shortcuts/apps/", "labelDomains": [], "e2eDomains": ["apps"] },
+    { "prefix": "shortcuts/markdown/", "labelDomains": [], "e2eDomains": ["markdown"] },
+    { "prefix": "shortcuts/minutes/", "labelDomains": [], "e2eDomains": ["minutes"] },
+    { "prefix": "shortcuts/okr/", "labelDomains": [], "e2eDomains": ["okr"] },
+    { "prefix": "shortcuts/slides/", "labelDomains": [], "e2eDomains": ["slides"] },
+    { "prefix": "shortcuts/note/", "labelDomains": [], "e2eDomains": ["note"] },
+    { "prefix": "shortcuts/event/", "labelDomains": [], "e2eDomains": ["event"] },
+
+    { "prefix": "skills/lark-im/", "labelDomains": ["im"], "e2eDomains": ["im"] },
+    { "prefix": "skills/lark-vc/", "labelDomains": ["vc"], "e2eDomains": ["vc"] },
+    { "prefix": "skills/lark-doc/", "labelDomains": ["ccm"], "e2eDomains": ["docs"] },
+    { "prefix": "skills/lark-wiki/", "labelDomains": ["ccm"], "e2eDomains": ["wiki"] },
+    { "prefix": "skills/lark-drive/", "labelDomains": ["ccm"], "e2eDomains": ["drive"] },
+    { "prefix": "skills/lark-sheets/", "labelDomains": ["ccm"], "e2eDomains": ["sheets"] },
+    { "prefix": "skills/lark-base/", "labelDomains": ["base"], "e2eDomains": ["base"] },
+    { "prefix": "skills/lark-mail/", "labelDomains": ["mail"], "e2eDomains": ["mail"] },
+    { "prefix": "skills/lark-calendar/", "labelDomains": ["calendar"], "e2eDomains": ["calendar"] },
+    { "prefix": "skills/lark-task/", "labelDomains": ["task"], "e2eDomains": ["task"] },
+    { "prefix": "skills/lark-contact/", "labelDomains": ["contact"], "e2eDomains": ["contact"] },
+    { "prefix": "skills/lark-apps/", "labelDomains": [], "e2eDomains": ["apps"] },
+    { "prefix": "skills/lark-markdown/", "labelDomains": [], "e2eDomains": ["markdown"] },
+    { "prefix": "skills/lark-minutes/", "labelDomains": [], "e2eDomains": ["minutes"] },
+    { "prefix": "skills/lark-okr/", "labelDomains": [], "e2eDomains": ["okr"] },
+    { "prefix": "skills/lark-slides/", "labelDomains": [], "e2eDomains": ["slides"] },
+    { "prefix": "skills/lark-note/", "labelDomains": [], "e2eDomains": ["note"] },
+    { "prefix": "skills/lark-event/", "labelDomains": [], "e2eDomains": ["event"] }
+  ],
+  "fullFallbackPrefixes": [
+    "shortcuts/common/",
+    "cmd/",
+    "internal/",
+    "pkg/",
+    "extension/",
+    "registry/",
+    "go.mod",
+    "go.sum",
+    "Makefile",
+    ".github/workflows/",
+    "scripts/"
+  ],
+  "skipPrefixes": [
+    "docs/",
+    ".changeset/"
+  ],
+  "skipSuffixes": [
+    ".md",
+    ".mdx",
+    ".txt",
+    ".rst"
+  ],
+  "skipFilenames": [
+    "readme.md",
+    "readme.zh.md",
+    "changelog.md",
+    "license",
+    "cla.md"
+  ]
+}

scripts/e2e_domains.js

@@ -0,0 +1,224 @@
+#!/usr/bin/env node
+// Copyright (c) 2026 Lark Technologies Pte. Ltd.
+// SPDX-License-Identifier: MIT
+
+const fs = require("node:fs");
+const path = require("node:path");
+const { execFileSync } = require("node:child_process");
+const {
+  e2eDomainsForPath,
+  findPathMapping,
+  isSkippablePath,
+  matchesFullFallback,
+  normalizeRepoPath,
+} = require("./domain-map");
+
+const ROOT = process.env.E2E_DOMAINS_ROOT || path.join(__dirname, "..");
+process.chdir(ROOT);
+
+function execLines(command, args) {
+  return execFileSync(command, args, { encoding: "utf8" })
+    .split(/\r?\n/)
+    .map((line) => line.trim())
+    .filter(Boolean);
+}
+
+function modulePath() {
+  return execLines("go", ["list", "-m"])[0];
+}
+
+function rootPackage(moduleName) {
+  return `${moduleName}/tests/cli_e2e`;
+}
+
+function allLivePackages(moduleName) {
+  return execLines("go", ["list", "./tests/cli_e2e/..."])
+    .filter((pkg) => pkg !== rootPackage(moduleName))
+    .filter((pkg) => !pkg.endsWith("/demo"));
+}
+
+function allDryPackages(moduleName) {
+  return allLivePackages(moduleName);
+}
+
+const domainExistsCache = new Map();
+
+function domainExists(domain) {
+  if (domainExistsCache.has(domain)) {
+    return domainExistsCache.get(domain);
+  }
+  let exists = false;
+  try {
+    execFileSync("go", ["list", `./tests/cli_e2e/${domain}`], { stdio: "ignore" });
+    exists = true;
+  } catch {
+    exists = false;
+  }
+  domainExistsCache.set(domain, exists);
+  return exists;
+}
+
+function readChangedFiles() {
+  const changedFilesPath = process.env.E2E_DOMAIN_CHANGED_FILES;
+  if (changedFilesPath) {
+    return fs.readFileSync(changedFilesPath, "utf8")
+      .split(/\r?\n/)
+      .map(normalizeRepoPath)
+      .filter(Boolean);
+  }
+
+  if (process.env.GITHUB_EVENT_NAME !== "pull_request") {
+    return null;
+  }
+
+  const baseRef = process.env.GITHUB_BASE_REF || "main";
+  try {
+    execFileSync("git", ["rev-parse", "--verify", `origin/${baseRef}`], { stdio: "ignore" });
+    return execLines("git", ["diff", "--name-only", `origin/${baseRef}...HEAD`]).map(normalizeRepoPath);
+  } catch {
+    return null;
+  }
+}
+
+function addDomain(domains, domain) {
+  if (domain && domainExists(domain)) {
+    domains.add(domain);
+    return true;
+  }
+  return false;
+}
+
+function classifyPath(filePath, domains) {
+  const normalized = normalizeRepoPath(filePath);
+  if (!normalized) return { matched: false };
+
+  const e2eMatch = normalized.match(/^tests\/cli_e2e\/([^/]+)\//);
+  if (e2eMatch) {
+    const domain = e2eMatch[1];
+    if (domain === "demo") return { matched: false };
+    if (domainExists(domain)) {
+      addDomain(domains, domain);
+      return { matched: true };
+    }
+    if (isSkippablePath(normalized)) return { matched: false };
+    return { fullReason: `unknown CLI E2E domain path: ${normalized}` };
+  }
+
+  if (normalized.startsWith("tests/cli_e2e/")) {
+    return { fullReason: `shared CLI E2E harness changed: ${normalized}` };
+  }
+
+  if (matchesFullFallback(normalized)) {
+    return { fullReason: `shared/runtime path changed: ${normalized}` };
+  }
+
+  const mappedDomains = e2eDomainsForPath(normalized);
+  if (mappedDomains.length > 0) {
+    const missingDomains = [];
+    for (const domain of mappedDomains) {
+      if (!addDomain(domains, domain)) missingDomains.push(domain);
+    }
+    if (missingDomains.length > 0) {
+      return { fullReason: `mapped CLI E2E domain has no package: ${missingDomains.join(",")} (${normalized})` };
+    }
+    return { matched: true };
+  }
+
+  if (findPathMapping(normalized)) {
+    return { fullReason: `mapped path has no CLI E2E package: ${normalized}` };
+  }
+
+  if (normalized.match(/^shortcuts\/[^/]+\//) || normalized.match(/^skills\/lark-[^/]+\//)) {
+    return { fullReason: `unmapped CLI E2E domain path: ${normalized}` };
+  }
+
+  if (isSkippablePath(normalized)) return { matched: false };
+
+  return { fullReason: `unclassified path changed: ${normalized}` };
+}
+
+function resolveDomains(changedFiles) {
+  const moduleName = modulePath();
+  const rootDryPackage = rootPackage(moduleName);
+  if (changedFiles === null) {
+    return {
+      mode: "full",
+      reason: "non-pull_request run or unavailable diff",
+      domains: ["all"],
+      dryRootPackage: rootDryPackage,
+      dryPackages: allDryPackages(moduleName),
+      livePackages: allLivePackages(moduleName),
+    };
+  }
+
+  const domains = new Set();
+  let matchedRelevant = false;
+  let fullReason = "";
+
+  for (const file of changedFiles) {
+    const result = classifyPath(file, domains);
+    if (result.matched) matchedRelevant = true;
+    if (result.fullReason && !fullReason) fullReason = result.fullReason;
+  }
+
+  if (fullReason) {
+    return {
+      mode: "full",
+      reason: fullReason,
+      domains: ["all"],
+      dryRootPackage: rootDryPackage,
+      dryPackages: allDryPackages(moduleName),
+      livePackages: allLivePackages(moduleName),
+    };
+  }
+
+  if (matchedRelevant && domains.size > 0) {
+    const sortedDomains = [...domains].sort();
+    const packages = sortedDomains.map((domain) => `${moduleName}/tests/cli_e2e/${domain}`);
+    return {
+      mode: "subset",
+      reason: "business domain changes",
+      domains: sortedDomains,
+      dryRootPackage: rootDryPackage,
+      dryPackages: packages,
+      livePackages: packages,
+    };
+  }
+
+  return {
+    mode: "skip",
+    reason: "docs-only or no live CLI E2E impact",
+    domains: [],
+    dryRootPackage: "",
+    dryPackages: [],
+    livePackages: [],
+  };
+}
+
+function emit(resolved) {
+  const values = {
+    mode: resolved.mode,
+    reason: resolved.reason,
+    domains: resolved.domains.join(","),
+    dry_root_package: resolved.dryRootPackage,
+    dry_packages: resolved.dryPackages.join(" "),
+    live_packages: resolved.livePackages.join(" "),
+  };
+
+  const lines = Object.entries(values).map(([key, value]) => `${key}=${value}`);
+  console.log(lines.join("\n"));
+
+  if (process.env.GITHUB_OUTPUT) {
+    fs.appendFileSync(process.env.GITHUB_OUTPUT, `${lines.join("\n")}\n`);
+  }
+}
+
+if (require.main === module) {
+  emit(resolveDomains(readChangedFiles()));
+}
+
+module.exports = {
+  classifyPath,
+  readChangedFiles,
+  resolveDomains,
+};

scripts/e2e_domains.test.js

@@ -0,0 +1,94 @@
+// Copyright (c) 2026 Lark Technologies Pte. Ltd.
+// SPDX-License-Identifier: MIT
+
+const assert = require("node:assert/strict");
+const fs = require("node:fs");
+const os = require("node:os");
+const path = require("node:path");
+const { execFileSync } = require("node:child_process");
+const test = require("node:test");
+
+const scriptPath = path.join(__dirname, "e2e_domains.js");
+
+function parseOutput(raw) {
+  const result = {};
+  for (const line of raw.trim().split(/\r?\n/)) {
+    const idx = line.indexOf("=");
+    if (idx === -1) continue;
+    result[line.slice(0, idx)] = line.slice(idx + 1);
+  }
+  return result;
+}
+
+function runDomains(files) {
+  const dir = fs.mkdtempSync(path.join(os.tmpdir(), "e2e-domains-"));
+  const file = path.join(dir, "changed.txt");
+  fs.writeFileSync(file, `${files.join("\n")}\n`);
+  try {
+    return parseOutput(execFileSync(process.execPath, [scriptPath], {
+      cwd: path.join(__dirname, ".."),
+      encoding: "utf8",
+      env: { ...process.env, E2E_DOMAIN_CHANGED_FILES: file },
+    }));
+  } finally {
+    fs.rmSync(dir, { recursive: true, force: true });
+  }
+}
+
+test("maps shortcut changes to one business domain package", () => {
+  const output = runDomains(["shortcuts/im/messages/send.go"]);
+  assert.equal(output.mode, "subset");
+  assert.equal(output.domains, "im");
+  assert.match(output.dry_root_package, /github\.com\/larksuite\/cli\/tests\/cli_e2e$/);
+  assert.match(output.live_packages, /github\.com\/larksuite\/cli\/tests\/cli_e2e\/im/);
+  assert.doesNotMatch(output.live_packages, /github\.com\/larksuite\/cli\/tests\/cli_e2e\/drive/);
+});
+
+test("maps doc shortcuts to docs package", () => {
+  const output = runDomains(["shortcuts/doc/update.go"]);
+  assert.equal(output.mode, "subset");
+  assert.equal(output.domains, "docs");
+  assert.match(output.live_packages, /github\.com\/larksuite\/cli\/tests\/cli_e2e\/docs/);
+});
+
+test("maps direct e2e domain package changes", () => {
+  const output = runDomains(["tests/cli_e2e/drive/helpers.go"]);
+  assert.equal(output.mode, "subset");
+  assert.equal(output.domains, "drive");
+  assert.match(output.live_packages, /github\.com\/larksuite\/cli\/tests\/cli_e2e\/drive/);
+});
+
+test("falls back to full for shared e2e harness changes", () => {
+  const output = runDomains(["tests/cli_e2e/core.go"]);
+  assert.equal(output.mode, "full");
+  assert.equal(output.domains, "all");
+  assert.match(output.reason, /shared CLI E2E harness changed/);
+});
+
+test("falls back to full for runtime changes", () => {
+  const output = runDomains(["cmd/root.go"]);
+  assert.equal(output.mode, "full");
+  assert.equal(output.domains, "all");
+  assert.match(output.reason, /shared\/runtime path changed/);
+});
+
+test("skips docs-only changes", () => {
+  const output = runDomains(["docs/usage.md", "README.md"]);
+  assert.equal(output.mode, "skip");
+  assert.equal(output.domains, "");
+  assert.equal(output.dry_root_package, "");
+  assert.equal(output.live_packages, "");
+});
+
+test("uses shared map for skill domain changes", () => {
+  const output = runDomains(["skills/lark-sheets/SKILL.md"]);
+  assert.equal(output.mode, "subset");
+  assert.equal(output.domains, "sheets");
+  assert.match(output.live_packages, /github\.com\/larksuite\/cli\/tests\/cli_e2e\/sheets/);
+});
+
+test("falls back to full when a mapped path has no e2e package", () => {
+  const output = runDomains(["shortcuts/whiteboard/export.go"]);
+  assert.equal(output.mode, "full");
+  assert.match(output.reason, /unmapped CLI E2E domain path/);
+});

scripts/pr-labels/index.js

@@ -4,6 +4,7 @@
 
 const fs = require("node:fs/promises");
 const path = require("node:path");
+const { labelDomainsForPath } = require("../domain-map");
 
 // ============================================================================
 // Constants & Configuration
@@ -35,33 +36,6 @@ const CORE_PREFIXES = ["internal/auth/", "internal/engine/", "internal/config/",
 const HEAD_BUSINESS_DOMAINS = new Set(["im", "contact", "ccm", "base", "docx"]);
 const LOW_RISK_TYPES = new Set(["docs", "ci", "test", "chore"]);
 
-// CODEOWNERS-based path to domain label mapping
-// Maps shortcuts and skills paths to business domain labels
-const PATH_TO_DOMAIN_MAP = {
-  // shortcuts
-  "shortcuts/im/": "im",
-  "shortcuts/vc/": "vc",
-  "shortcuts/calendar/": "calendar",
-  "shortcuts/doc/": "ccm",
-  "shortcuts/sheets/": "ccm",
-  "shortcuts/drive/": "ccm",
-  "shortcuts/wiki/": "ccm",
-  "shortcuts/base/": "base",
-  "shortcuts/mail/": "mail",
-  "shortcuts/task/": "task",
-  "shortcuts/contact/": "contact",
-  // skills
-  "skills/lark-im/": "im",
-  "skills/lark-vc/": "vc",
-  "skills/lark-doc/": "ccm",
-  "skills/lark-wiki/": "ccm",
-  "skills/lark-base/": "base",
-  "skills/lark-mail/": "mail",
-  "skills/lark-calendar/": "calendar",
-  "skills/lark-task/": "task",
-  "skills/lark-contact/": "contact",
-};
-
 const SENSITIVE_PATTERN = /(^|\/)(auth|permission|permissions|security)(\/|_|\.|$)/;
 
 const CLASS_STANDARDS = {
@@ -285,13 +259,7 @@ function skillDomainForPath(filePath) {
 
 // Get business domain label based on CODEOWNERS path mapping
 function getBusinessDomain(filePath) {
-  const normalized = normalizePath(filePath);
-  for (const [prefix, domain] of Object.entries(PATH_TO_DOMAIN_MAP)) {
-    if (normalized.startsWith(prefix)) {
-      return domain;
-    }
-  }
-  return "";
+  return labelDomainsForPath(filePath)[0] || "";
 }
 
 async function detectNewShortcutDomain(files) {

scripts/resolve-changed-from.test.sh

@@ -8,7 +8,17 @@ repo_root="$(cd "$(dirname "${BASH_SOURCE[0]}")/.." && pwd)"
 script="$repo_root/scripts/resolve-changed-from.sh"
 
 tmp="${TMPDIR:-/tmp}/resolve-changed-from-test-$$"
-trap 'rm -rf "$tmp"' EXIT
+
+cleanup_tmp() {
+  local attempt
+  for attempt in 1 2 3; do
+    rm -rf "$tmp" && return 0
+    sleep 1
+  done
+  rm -rf "$tmp"
+}
+
+trap cleanup_tmp EXIT
 mkdir -p "$tmp"
 
 git_init() {

shortcuts/calendar/calendar_create.go

@@ -20,6 +20,7 @@ import (
 func buildEventData(runtime *common.RuntimeContext, startTs, endTs string) map[string]interface{} {
 	eventData := map[string]interface{}{
 		"summary":          runtime.Str("summary"),
+		"description":      runtime.Str("description"),
 		"start_time":       map[string]string{"timestamp": startTs},
 		"end_time":         map[string]string{"timestamp": endTs},
 		"attendee_ability": "can_modify_event",
@@ -29,15 +30,9 @@ func buildEventData(runtime *common.RuntimeContext, startTs, endTs string) map[s
 			{"minutes": 5},
 		},
 	}
-	if runtime.Cmd.Flags().Changed("description") {
-		eventData["description"] = runtime.Str("description")
-	}
 	if rrule := runtime.Str("rrule"); rrule != "" {
 		eventData["recurrence"] = rrule
 	}
-	if descriptionRich := runtime.Str("description-rich"); descriptionRich != "" {
-		eventData["description_rich"] = descriptionRich
-	}
 	return eventData
 }
 
@@ -104,8 +99,7 @@ var CalendarCreate = common.Shortcut{
 		{Name: "summary", Desc: "event title"},
 		{Name: "start", Desc: "start time (ISO 8601)", Required: true},
 		{Name: "end", Desc: "end time (ISO 8601)", Required: true},
-		{Name: "description", Desc: "event description (plain text)"},
-		{Name: "description-rich", Desc: "rich-text description as Markdown (@file or - for stdin); supports bold/italic/underline/strikethrough, links, ordered/unordered lists, and GFM tables. When set without --description, a plain-text preview is auto-derived so the client renders it immediately", Input: []string{common.File, common.Stdin}},
+		{Name: "description", Desc: "event description"},
 		{Name: "attendee-ids", Desc: "attendee IDs, comma-separated (supports user ou_, chat oc_, room omm_)"},
 		{Name: "calendar-id", Desc: "calendar ID (default: primary)"},
 		{Name: "rrule", Desc: "recurrence rule (rfc5545)"},

shortcuts/calendar/calendar_update.go

@@ -29,8 +29,7 @@ var CalendarUpdate = common.Shortcut{
 		{Name: "event-id", Desc: "event ID to update", Required: true},
 		{Name: "calendar-id", Desc: "calendar ID (default: primary)"},
 		{Name: "summary", Desc: "event title"},
-		{Name: "description", Desc: "event description (plain text)"},
-		{Name: "description-rich", Desc: "rich-text description as Markdown (@file or - for stdin); supports bold/italic/underline/strikethrough, links, ordered/unordered lists, and GFM tables. When set without --description, a plain-text preview is auto-derived so the client renders it immediately", Input: []string{common.File, common.Stdin}},
+		{Name: "description", Desc: "event description"},
 		{Name: "start", Desc: "new start time (ISO 8601); requires --end"},
 		{Name: "end", Desc: "new end time (ISO 8601); requires --start"},
 		{Name: "rrule", Desc: "recurrence rule (rfc5545)"},
@@ -71,7 +70,7 @@ func validateCalendarUpdate(runtime *common.RuntimeContext) error {
 		return err
 	}
 	if !hasCalendarUpdateOperation(runtime) {
-		return errs.NewValidationError(errs.SubtypeInvalidArgument, "nothing to update: specify at least one of --summary, --description, --description-rich, --start/--end, --rrule, --add-attendee-ids, or --remove-attendee-ids")
+		return errs.NewValidationError(errs.SubtypeInvalidArgument, "nothing to update: specify at least one of --summary, --description, --start/--end, --rrule, --add-attendee-ids, or --remove-attendee-ids")
 	}
 	return nil
 }
@@ -115,10 +114,6 @@ func buildCalendarUpdateEventData(runtime *common.RuntimeContext) (map[string]in
 			hasFields = true
 		}
 	}
-	if runtime.Cmd.Flags().Changed("description-rich") {
-		body["description_rich"] = runtime.Str("description-rich")
-		hasFields = true
-	}
 	if runtime.Cmd.Flags().Changed("rrule") {
 		rrule := strings.TrimSpace(runtime.Str("rrule"))
 		if rrule != "" {
@@ -358,9 +353,6 @@ func calendarUpdateResult(eventID string, event map[string]interface{}, addedCou
 	if description, _ := event["description"].(string); description != "" {
 		result["description"] = description
 	}
-	if descriptionRich, _ := event["description_rich"].(string); descriptionRich != "" {
-		result["description_rich"] = descriptionRich
-	}
 	if start := formatCalendarEventTime(event["start_time"]); start != "" {
 		result["start"] = start
 	}

shortcuts/mail/helpers.go

@@ -51,9 +51,8 @@ func hintSendDraft(runtime *common.RuntimeContext, mailboxID, draftID string) {
 // original message as read after a reply/reply-all/forward operation.
 func hintMarkAsRead(runtime *common.RuntimeContext, mailboxID, originalMessageID string) {
 	fmt.Fprintf(runtime.IO().ErrOut,
-		"tip: mark original as read? lark-cli mail user_mailbox.messages batch_modify_message"+
-			` --params '{"user_mailbox_id":"%s"}' --data '{"message_ids":["%s"],"remove_label_ids":["UNREAD"]}'`+"\n",
-		sanitizeForTerminal(mailboxID), sanitizeForTerminal(originalMessageID))
+		"tip: mark original as read? lark-cli mail +message-modify --mailbox '%s' --message-ids '%s' --remove-label-ids UNREAD\n",
+		shellQuoteForHint(mailboxID), shellQuoteForHint(originalMessageID))
 }
 
 // hintReadReceiptRequest prints a stderr tip when a message that the caller

shortcuts/mail/helpers_test.go

@@ -465,14 +465,19 @@ func TestPrintWatchOutputSchema(t *testing.T) {
 // TestHintMarkAsRead verifies hint mark as read.
 func TestHintMarkAsRead(t *testing.T) {
 	rt, _, stderr := newOutputRuntime(t)
-	// Inject ANSI escape + message ID to verify sanitization
-	hintMarkAsRead(rt, "me", "msg-\x1b[31m123")
+	hintMarkAsRead(rt, "mail box;$(whoami)", "msg-\x1b[31m123 'quoted'\nnext")
 	out := stderr.String()
 	if strings.Contains(out, "\x1b[") {
 		t.Errorf("hintMarkAsRead should sanitize ANSI escapes, got: %q", out)
 	}
-	if !strings.Contains(out, "msg-123") {
-		t.Errorf("hintMarkAsRead should contain sanitized message ID, got: %q", out)
+	if strings.Contains(out, "\nnext") {
+		t.Errorf("hintMarkAsRead should strip embedded newlines, got: %q", out)
+	}
+	if !strings.Contains(out, "--mailbox 'mail box;$(whoami)'") {
+		t.Errorf("hintMarkAsRead should quote mailbox for shell copy/paste, got: %q", out)
+	}
+	if !strings.Contains(out, "--message-ids 'msg-123 '\\''quoted'\\''next'") {
+		t.Errorf("hintMarkAsRead should quote message ID for shell copy/paste, got: %q", out)
 	}
 }
 

shortcuts/mail/mail_message_manage_test.go

@@ -0,0 +1,482 @@
+// Copyright (c) 2026 Lark Technologies Pte. Ltd.
+// SPDX-License-Identifier: MIT
+
+package mail
+
+import (
+	"encoding/json"
+	"errors"
+	"fmt"
+	"strings"
+	"testing"
+
+	"github.com/larksuite/cli/errs"
+	"github.com/larksuite/cli/internal/auth"
+	"github.com/larksuite/cli/internal/httpmock"
+	"github.com/larksuite/cli/internal/output"
+	"github.com/larksuite/cli/shortcuts/common"
+)
+
+func messageManageID(suffix string) string {
+	return "msg_abcdefghijklmnop_" + suffix
+}
+
+func stubMessageManagePost(reg *httpmock.Registry, endpoint string, body map[string]interface{}) *httpmock.Stub {
+	stub := &httpmock.Stub{
+		Method: "POST",
+		URL:    "/user_mailboxes/me/messages/" + endpoint,
+		Body:   body,
+	}
+	reg.Register(stub)
+	return stub
+}
+
+func decodeMessageManageSummary(t *testing.T, data map[string]interface{}) ([]interface{}, []interface{}) {
+	t.Helper()
+	success, ok := data["success_message_ids"].([]interface{})
+	if !ok {
+		t.Fatalf("success_message_ids = %#v, want array", data["success_message_ids"])
+	}
+	failed, ok := data["failed_message_ids"].([]interface{})
+	if !ok {
+		t.Fatalf("failed_message_ids = %#v, want array", data["failed_message_ids"])
+	}
+	return success, failed
+}
+
+func requireMessageManageValidationParam(t *testing.T, err error, param string) *errs.ValidationError {
+	t.Helper()
+	if err == nil {
+		t.Fatalf("expected validation error for %s, got nil", param)
+	}
+	var validationErr *errs.ValidationError
+	if !errors.As(err, &validationErr) {
+		t.Fatalf("expected *errs.ValidationError for %s, got %T", param, err)
+	}
+	problem, ok := errs.ProblemOf(err)
+	if !ok {
+		t.Fatalf("expected typed Problem for %s, got %T", param, err)
+	}
+	if problem.Category != errs.CategoryValidation || problem.Subtype != errs.SubtypeInvalidArgument {
+		t.Fatalf("problem = %s/%s, want validation/invalid_argument", problem.Category, problem.Subtype)
+	}
+	if validationErr.Param != param {
+		t.Fatalf("param = %q, want %q", validationErr.Param, param)
+	}
+	return validationErr
+}
+
+func requireMessageManageFailedPrecondition(t *testing.T, err error) {
+	t.Helper()
+	if err == nil {
+		t.Fatal("expected failed precondition error, got nil")
+	}
+	var validationErr *errs.ValidationError
+	if !errors.As(err, &validationErr) {
+		t.Fatalf("expected *errs.ValidationError, got %T", err)
+	}
+	problem, ok := errs.ProblemOf(err)
+	if !ok {
+		t.Fatalf("expected typed Problem, got %T", err)
+	}
+	if problem.Category != errs.CategoryValidation || problem.Subtype != errs.SubtypeFailedPrecondition {
+		t.Fatalf("problem = %s/%s, want validation/failed_precondition", problem.Category, problem.Subtype)
+	}
+}
+
+func TestMessageManage_NormalizeMessageIDs(t *testing.T) {
+	id1 := messageManageID("1")
+	id2 := messageManageID("2")
+	got, err := normalizeMessageManageIDs([]string{id1, id2, id1})
+	if err != nil {
+		t.Fatalf("normalizeMessageManageIDs returned error: %v", err)
+	}
+	if len(got) != 2 || got[0] != id1 || got[1] != id2 {
+		t.Fatalf("ids = %v, want [%s %s]", got, id1, id2)
+	}
+	got, err = normalizeMessageManageIDs([]string{id1 + "," + id2, id1})
+	if err != nil {
+		t.Fatalf("normalizeMessageManageIDs CSV/repeated returned error: %v", err)
+	}
+	if len(got) != 2 || got[0] != id1 || got[1] != id2 {
+		t.Fatalf("CSV/repeated ids = %v, want [%s %s]", got, id1, id2)
+	}
+
+	cases := [][]string{
+		{""},
+		{" id_with_leading_space_12345"},
+		{"msg_abcdefghijklmnop_1,msg_abcdefghijklmnop_2 "},
+		{"1234567890123456"},
+		{"short"},
+		{"msg_abcdefghijklmnop!"},
+		{"msg_abcdefghijklmnop\t"},
+		{"msg_abcdefghijklmnop_1\nmsg_abcdefghijklmnop_2"},
+		{"msg_abcdefghijklmnop_1", "msg_abcdefghijklmnop_2 "},
+	}
+	for _, tc := range cases {
+		_, err := normalizeMessageManageIDs(tc)
+		requireMessageManageValidationParam(t, err, "--message-ids")
+	}
+}
+
+func TestMessageModify_Metadata(t *testing.T) {
+	if MailMessageModify.Command != "+message-modify" {
+		t.Fatalf("Command = %q", MailMessageModify.Command)
+	}
+	if MailMessageModify.Risk != "write" {
+		t.Errorf("Risk = %q, want write", MailMessageModify.Risk)
+	}
+	if len(MailMessageModify.AuthTypes) != 1 || MailMessageModify.AuthTypes[0] != "user" {
+		t.Errorf("AuthTypes = %v, want [user]", MailMessageModify.AuthTypes)
+	}
+	requiredScopes := map[string]bool{
+		"mail:user_mailbox.message:modify": true,
+	}
+	for _, scope := range MailMessageModify.Scopes {
+		delete(requiredScopes, scope)
+	}
+	if len(requiredScopes) != 0 {
+		t.Errorf("Scopes missing %v", requiredScopes)
+	}
+	if len(MailMessageModify.ConditionalScopes) != 1 || MailMessageModify.ConditionalScopes[0] != "mail:user_mailbox.folder:read" {
+		t.Errorf("ConditionalScopes = %v, want [mail:user_mailbox.folder:read]", MailMessageModify.ConditionalScopes)
+	}
+	flags := map[string]common.Flag{}
+	for _, fl := range MailMessageModify.Flags {
+		flags[fl.Name] = fl
+	}
+	for _, name := range []string{"mailbox", "message-ids", "add-label-ids", "remove-label-ids", "add-folder"} {
+		if _, ok := flags[name]; !ok {
+			t.Fatalf("missing --%s flag", name)
+		}
+	}
+	if flags["message-ids"].Type != "string_array" || !flags["message-ids"].Required {
+		t.Errorf("--message-ids = %#v, want required string_array", flags["message-ids"])
+	}
+}
+
+func TestMessageTrash_Metadata(t *testing.T) {
+	if MailMessageTrash.Command != "+message-trash" {
+		t.Fatalf("Command = %q", MailMessageTrash.Command)
+	}
+	if MailMessageTrash.Risk != "high-risk-write" {
+		t.Errorf("Risk = %q, want high-risk-write", MailMessageTrash.Risk)
+	}
+	if len(MailMessageTrash.AuthTypes) != 1 || MailMessageTrash.AuthTypes[0] != "user" {
+		t.Errorf("AuthTypes = %v, want [user]", MailMessageTrash.AuthTypes)
+	}
+	if len(MailMessageTrash.Scopes) != 1 || MailMessageTrash.Scopes[0] != "mail:user_mailbox.message:modify" {
+		t.Errorf("Scopes = %v, want [mail:user_mailbox.message:modify]", MailMessageTrash.Scopes)
+	}
+}
+
+func TestMessageModify_LabelOnlyDoesNotRequireFolderReadScope(t *testing.T) {
+	f, stdout, _, reg := mailShortcutTestFactory(t)
+	token := auth.GetStoredToken("test-app", "ou_testuser")
+	if token == nil {
+		t.Fatal("expected test token")
+	}
+	token.Scope = strings.ReplaceAll(token.Scope, " mail:user_mailbox.folder:read", "")
+	if err := auth.SetStoredToken(token); err != nil {
+		t.Fatalf("SetStoredToken() error = %v", err)
+	}
+
+	id := messageManageID("1")
+	post := stubMessageManagePost(reg, "batch_modify", map[string]interface{}{"code": 0, "data": map[string]interface{}{}})
+
+	err := runMountedMailShortcut(t, MailMessageModify, []string{
+		"+message-modify",
+		"--message-ids", id,
+		"--remove-label-ids", "UNREAD",
+	}, f, stdout)
+	if err != nil {
+		t.Fatalf("unexpected err: %v", err)
+	}
+	var body map[string]interface{}
+	if err := json.Unmarshal(post.CapturedBody, &body); err != nil {
+		t.Fatalf("unmarshal captured body: %v", err)
+	}
+	removeLabels := body["remove_label_ids"].([]interface{})
+	if len(removeLabels) != 1 || removeLabels[0] != "UNREAD" {
+		t.Fatalf("remove_label_ids = %#v, want [UNREAD]", removeLabels)
+	}
+}
+
+func TestMessageModify_ReadReceiptRequestLabelIsSystemLabel(t *testing.T) {
+	f, stdout, _, reg := mailShortcutTestFactory(t)
+	id := messageManageID("1")
+	post := stubMessageManagePost(reg, "batch_modify", map[string]interface{}{"code": 0, "data": map[string]interface{}{}})
+
+	err := runMountedMailShortcut(t, MailMessageModify, []string{
+		"+message-modify",
+		"--message-ids", id,
+		"--remove-label-ids", "read_receipt_request",
+	}, f, stdout)
+	if err != nil {
+		t.Fatalf("unexpected err: %v", err)
+	}
+	var body map[string]interface{}
+	if err := json.Unmarshal(post.CapturedBody, &body); err != nil {
+		t.Fatalf("unmarshal captured body: %v", err)
+	}
+	removeLabels := body["remove_label_ids"].([]interface{})
+	if len(removeLabels) != 1 || removeLabels[0] != "READ_RECEIPT_REQUEST" {
+		t.Fatalf("remove_label_ids = %#v, want [READ_RECEIPT_REQUEST]", removeLabels)
+	}
+}
+
+func TestMessageModify_LabelFolderNormalizationAndValidationAPIs(t *testing.T) {
+	f, stdout, _, reg := mailShortcutTestFactory(t)
+	id := messageManageID("1")
+	reg.Register(&httpmock.Stub{Method: "GET", URL: "/user_mailboxes/me/labels/customA", Body: map[string]interface{}{"code": 0, "data": map[string]interface{}{"label_id": "customA"}}})
+	reg.Register(&httpmock.Stub{Method: "GET", URL: "/user_mailboxes/me/folders/folderA", Body: map[string]interface{}{"code": 0, "data": map[string]interface{}{"folder_id": "folderA"}}})
+	post := stubMessageManagePost(reg, "batch_modify", map[string]interface{}{"code": 0, "data": map[string]interface{}{}})
+
+	err := runMountedMailShortcut(t, MailMessageModify, []string{
+		"+message-modify",
+		"--message-ids", id,
+		"--add-label-ids", "unread,customA",
+		"--remove-label-ids", "FLAGGED",
+		"--add-folder", "folderA",
+	}, f, stdout)
+	if err != nil {
+		t.Fatalf("unexpected err: %v", err)
+	}
+	var body map[string]interface{}
+	if err := json.Unmarshal(post.CapturedBody, &body); err != nil {
+		t.Fatalf("unmarshal captured body: %v", err)
+	}
+	if got := body["add_folder"]; got != "folderA" {
+		t.Errorf("add_folder = %v, want folderA", got)
+	}
+	addLabels := body["add_label_ids"].([]interface{})
+	if addLabels[0] != "UNREAD" || addLabels[1] != "customA" {
+		t.Errorf("add_label_ids = %#v, want [UNREAD customA]", addLabels)
+	}
+	removeLabels := body["remove_label_ids"].([]interface{})
+	if removeLabels[0] != "FLAGGED" {
+		t.Errorf("remove_label_ids = %#v, want [FLAGGED]", removeLabels)
+	}
+	success, failed := decodeMessageManageSummary(t, decodeShortcutEnvelopeData(t, stdout))
+	if len(success) != 1 || success[0] != id || len(failed) != 0 {
+		t.Errorf("summary success=%v failed=%v", success, failed)
+	}
+}
+
+func TestMessageModify_RejectsLabelIntersectionAndTrashFolder(t *testing.T) {
+	f, stdout, _, _ := mailShortcutTestFactory(t)
+	id := messageManageID("1")
+	err := runMountedMailShortcut(t, MailMessageModify, []string{
+		"+message-modify",
+		"--message-ids", id,
+		"--add-label-ids", "unread",
+		"--remove-label-ids", "UNREAD",
+	}, f, stdout)
+	requireMessageManageValidationParam(t, err, "--add-label-ids")
+	if !strings.Contains(err.Error(), "label cannot be both added and removed") {
+		t.Fatalf("error = %v, want label intersection validation", err)
+	}
+
+	err = runMountedMailShortcut(t, MailMessageModify, []string{
+		"+message-modify",
+		"--message-ids", id,
+		"--add-folder", "trash",
+	}, f, stdout)
+	requireMessageManageValidationParam(t, err, "--add-folder")
+	if !strings.Contains(err.Error(), "use +message-trash") {
+		t.Fatalf("error = %v, want TRASH validation", err)
+	}
+}
+
+func TestMessageModify_EmptyOperationDoesNotCallPost(t *testing.T) {
+	f, stdout, _, _ := mailShortcutTestFactory(t)
+	id1 := messageManageID("1")
+	id2 := messageManageID("2")
+	err := runMountedMailShortcut(t, MailMessageModify, []string{
+		"+message-modify",
+		"--message-ids", id1 + "," + id2 + "," + id1,
+	}, f, stdout)
+	if err != nil {
+		t.Fatalf("unexpected err: %v", err)
+	}
+	success, failed := decodeMessageManageSummary(t, decodeShortcutEnvelopeData(t, stdout))
+	if len(success) != 2 || success[0] != id1 || success[1] != id2 || len(failed) != 0 {
+		t.Fatalf("summary success=%v failed=%v", success, failed)
+	}
+}
+
+func TestMessageModify_BatchesAndAggregatesPartialFailure(t *testing.T) {
+	f, stdout, _, reg := mailShortcutTestFactory(t)
+	ids := make([]string, 41)
+	for i := range ids {
+		ids[i] = messageManageID(fmt.Sprintf("%02d", i))
+	}
+	first := stubMessageManagePost(reg, "batch_modify", map[string]interface{}{"code": 0, "data": map[string]interface{}{}})
+	second := stubMessageManagePost(reg, "batch_modify", map[string]interface{}{"code": 1230001, "msg": "bad request"})
+	third := stubMessageManagePost(reg, "batch_modify", map[string]interface{}{"code": 0, "data": map[string]interface{}{}})
+
+	err := runMountedMailShortcut(t, MailMessageModify, []string{
+		"+message-modify",
+		"--message-ids", strings.Join(ids, ","),
+		"--add-folder", "archive",
+	}, f, stdout)
+	if err != nil {
+		t.Fatalf("unexpected err: %v", err)
+	}
+	for idx, stub := range []*httpmock.Stub{first, second, third} {
+		var body map[string]interface{}
+		if err := json.Unmarshal(stub.CapturedBody, &body); err != nil {
+			t.Fatalf("batch %d body unmarshal: %v", idx+1, err)
+		}
+		messageIDs := body["message_ids"].([]interface{})
+		want := []int{20, 20, 1}[idx]
+		if len(messageIDs) != want {
+			t.Fatalf("batch %d size = %d, want %d", idx+1, len(messageIDs), want)
+		}
+		if body["add_folder"] != "ARCHIVED" {
+			t.Fatalf("batch %d add_folder = %v, want ARCHIVED", idx+1, body["add_folder"])
+		}
+	}
+	success, failed := decodeMessageManageSummary(t, decodeShortcutEnvelopeData(t, stdout))
+	if len(success) != 21 || len(failed) != 20 {
+		t.Fatalf("success=%d failed=%d, want 21/20", len(success), len(failed))
+	}
+}
+
+func TestMessageModify_AllBatchesFailReturnsError(t *testing.T) {
+	f, stdout, _, reg := mailShortcutTestFactory(t)
+	id := messageManageID("1")
+	stubMessageManagePost(reg, "batch_modify", map[string]interface{}{"code": 1230001, "msg": "bad request"})
+
+	err := runMountedMailShortcut(t, MailMessageModify, []string{
+		"+message-modify",
+		"--message-ids", id,
+		"--add-folder", "archive",
+	}, f, stdout)
+	requireMessageManageFailedPrecondition(t, err)
+}
+
+func TestMessageModify_DryRunShowsPlanWithoutValidationGET(t *testing.T) {
+	f, stdout, _, _ := mailShortcutTestFactory(t)
+	id1 := messageManageID("1")
+	id2 := messageManageID("2")
+	err := runMountedMailShortcut(t, MailMessageModify, []string{
+		"+message-modify",
+		"--message-ids", id1 + "," + id2,
+		"--add-label-ids", "customA",
+		"--add-folder", "folderA",
+		"--dry-run",
+	}, f, stdout)
+	if err != nil {
+		t.Fatalf("dry-run failed: %v", err)
+	}
+	out := stdout.String()
+	for _, want := range []string{
+		`/user_mailboxes/me/messages/batch_modify`,
+		`validation_api_plan`,
+		`/user_mailboxes/me/labels/customA`,
+		`/user_mailboxes/me/folders/folderA`,
+		`will_validate`,
+		`batch_size`,
+	} {
+		if !strings.Contains(out, want) {
+			t.Fatalf("dry-run output missing %q; got %s", want, out)
+		}
+	}
+}
+
+func TestMessageTrash_RequiresYesAndBatches(t *testing.T) {
+	f, stdout, _, reg := mailShortcutTestFactory(t)
+	id1 := messageManageID("1")
+	id2 := messageManageID("2")
+	err := runMountedMailShortcut(t, MailMessageTrash, []string{
+		"+message-trash",
+		"--message-ids", id1 + "," + id2,
+	}, f, stdout)
+	if err == nil {
+		t.Fatal("expected confirmation error, got nil")
+	}
+	if code := output.ExitCodeOf(err); code != output.ExitConfirmationRequired {
+		t.Fatalf("exit code = %d, want %d", code, output.ExitConfirmationRequired)
+	}
+
+	post := stubMessageManagePost(reg, "batch_trash", map[string]interface{}{"code": 0, "data": map[string]interface{}{}})
+	err = runMountedMailShortcut(t, MailMessageTrash, []string{
+		"+message-trash",
+		"--message-ids", id1 + "," + id2,
+		"--yes",
+	}, f, stdout)
+	if err != nil {
+		t.Fatalf("unexpected err with --yes: %v", err)
+	}
+	var body map[string]interface{}
+	if err := json.Unmarshal(post.CapturedBody, &body); err != nil {
+		t.Fatalf("unmarshal captured body: %v", err)
+	}
+	if got := len(body["message_ids"].([]interface{})); got != 2 {
+		t.Fatalf("message_ids len = %d, want 2", got)
+	}
+	success, failed := decodeMessageManageSummary(t, decodeShortcutEnvelopeData(t, stdout))
+	if len(success) != 2 || len(failed) != 0 {
+		t.Fatalf("summary success=%v failed=%v", success, failed)
+	}
+}
+
+func TestMessageTrash_AllBatchesFailReturnsError(t *testing.T) {
+	f, stdout, _, reg := mailShortcutTestFactory(t)
+	id := messageManageID("1")
+	stubMessageManagePost(reg, "batch_trash", map[string]interface{}{"code": 1230001, "msg": "bad request"})
+
+	err := runMountedMailShortcut(t, MailMessageTrash, []string{
+		"+message-trash",
+		"--message-ids", id,
+		"--yes",
+	}, f, stdout)
+	requireMessageManageFailedPrecondition(t, err)
+}
+
+func TestMessageManage_RejectsWhitespaceBeforeAPI(t *testing.T) {
+	id1 := messageManageID("1")
+	id2 := messageManageID("2")
+	cases := []struct {
+		name     string
+		shortcut common.Shortcut
+		args     []string
+	}{
+		{
+			name:     "trash newline in repeated flag",
+			shortcut: MailMessageTrash,
+			args:     []string{"+message-trash", "--message-ids", id1 + "\n" + id2, "--yes"},
+		},
+		{
+			name:     "trash tab in csv flag",
+			shortcut: MailMessageTrash,
+			args:     []string{"+message-trash", "--message-ids", id1 + ",\t" + id2, "--yes"},
+		},
+		{
+			name:     "modify space in repeated flag",
+			shortcut: MailMessageModify,
+			args:     []string{"+message-modify", "--message-ids", id1, "--message-ids", id2 + " ", "--add-folder", "archive"},
+		},
+		{
+			name:     "modify space in csv flag",
+			shortcut: MailMessageModify,
+			args:     []string{"+message-modify", "--message-ids", id1 + ", " + id2, "--add-folder", "archive"},
+		},
+	}
+	for _, tc := range cases {
+		t.Run(tc.name, func(t *testing.T) {
+			f, stdout, _, _ := mailShortcutTestFactory(t)
+			err := runMountedMailShortcut(t, tc.shortcut, tc.args, f, stdout)
+			if err == nil {
+				t.Fatal("expected validation error, got nil")
+			}
+			if code := output.ExitCodeOf(err); code != output.ExitValidation {
+				t.Fatalf("exit code = %d, want %d; err=%v", code, output.ExitValidation, err)
+			}
+			if !strings.Contains(err.Error(), "must not contain whitespace or control characters") {
+				t.Fatalf("error = %v, want whitespace/control validation", err)
+			}
+		})
+	}
+}

shortcuts/mail/mail_message_modify.go

@@ -0,0 +1,141 @@
+// Copyright (c) 2026 Lark Technologies Pte. Ltd.
+// SPDX-License-Identifier: MIT
+
+package mail
+
+import (
+	"context"
+
+	"github.com/larksuite/cli/shortcuts/common"
+)
+
+type messageModifyInput struct {
+	MessageIDs         []string
+	AddLabelIDs        []string
+	RemoveLabelIDs     []string
+	AddFolder          string
+	CustomLabelIDs     []string
+	CustomFolderID     string
+	ValidationAPIPlans []validationAPIPlan
+}
+
+// MailMessageModify is the `+message-modify` shortcut: apply labels, unread
+// state labels, or a folder move to existing messages in batches of 20.
+var MailMessageModify = common.Shortcut{
+	Service:     "mail",
+	Command:     "+message-modify",
+	Description: "Modify existing mail messages by adding/removing label IDs or moving them to a folder. Batches message IDs in groups of 20 and keeps output compact.",
+	Risk:        "write",
+	Scopes:      []string{"mail:user_mailbox.message:modify"},
+	ConditionalScopes: []string{
+		"mail:user_mailbox.folder:read",
+	},
+	AuthTypes: []string{"user"},
+	HasFormat: true,
+	Flags: []common.Flag{
+		{Name: "mailbox", Desc: "Mailbox email address that owns the messages (default: me)."},
+		{Name: "message-ids", Type: "string_array", Required: true, Desc: "Message IDs to modify; comma-separated or repeat the flag."},
+		{Name: "add-label-ids", Type: "string_slice", Desc: "Label IDs to add. System labels unread/important/other/flagged are normalized to upper case."},
+		{Name: "remove-label-ids", Type: "string_slice", Desc: "Label IDs to remove. System labels unread/important/other/flagged are normalized to upper case."},
+		{Name: "add-folder", Desc: "Folder ID to move messages to. System folders inbox/sent/spam/archive/archived are normalized; TRASH is rejected, use +message-trash."},
+	},
+	Validate: validateMessageModify,
+	DryRun:   dryRunMessageModify,
+	Execute:  executeMessageModify,
+}
+
+func validateMessageModify(ctx context.Context, rt *common.RuntimeContext) error {
+	_, err := buildMessageModifyInput(rt)
+	return err
+}
+
+func dryRunMessageModify(ctx context.Context, rt *common.RuntimeContext) *common.DryRunAPI {
+	mailboxID := resolveMailboxID(rt)
+	input, _ := buildMessageModifyInput(rt)
+	api := common.NewDryRunAPI().
+		Desc("Modify messages sequentially in batches of 20; dry-run does not call label/folder validation APIs").
+		Set("batch_size", mailMessageManageBatchSize).
+		Set("batches", chunkMessageManageIDs(input.MessageIDs)).
+		Set("validation_api_plan", input.ValidationAPIPlans)
+	for _, batch := range chunkMessageManageIDs(input.MessageIDs) {
+		api = api.POST(mailboxPath(mailboxID, "messages", "batch_modify")).
+			Body(messageManageBody(batch, input.AddLabelIDs, input.RemoveLabelIDs, input.AddFolder))
+	}
+	return api
+}
+
+func executeMessageModify(ctx context.Context, rt *common.RuntimeContext) error {
+	mailboxID := resolveMailboxID(rt)
+	input, err := buildMessageModifyInput(rt)
+	if err != nil {
+		return err
+	}
+	if err := validateCustomMessageManageLabels(rt, mailboxID, input.CustomLabelIDs); err != nil {
+		return err
+	}
+	if err := validateCustomMessageManageFolder(rt, mailboxID, input.CustomFolderID); err != nil {
+		return err
+	}
+
+	if len(input.AddLabelIDs) == 0 && len(input.RemoveLabelIDs) == 0 && input.AddFolder == "" {
+		emitMessageManageSummary(rt, messageManageSummary{
+			SuccessMessageIDs: input.MessageIDs,
+			FailedMessageIDs:  []messageManageFailure{},
+		}, true)
+		return nil
+	}
+
+	summary := messageManageSummary{FailedMessageIDs: []messageManageFailure{}}
+	for _, batch := range chunkMessageManageIDs(input.MessageIDs) {
+		_, err := rt.CallAPITyped("POST", mailboxPath(mailboxID, "messages", "batch_modify"), nil,
+			messageManageBody(batch, input.AddLabelIDs, input.RemoveLabelIDs, input.AddFolder))
+		if err != nil {
+			for _, id := range batch {
+				summary.FailedMessageIDs = append(summary.FailedMessageIDs, messageManageFailure{MessageID: id, Reason: err.Error()})
+			}
+			continue
+		}
+		summary.SuccessMessageIDs = append(summary.SuccessMessageIDs, batch...)
+	}
+	emitMessageManageSummary(rt, summary, false)
+	if len(summary.SuccessMessageIDs) == 0 && len(summary.FailedMessageIDs) > 0 {
+		return mailFailedPreconditionError("all message modify batches failed")
+	}
+	return nil
+}
+
+func buildMessageModifyInput(rt *common.RuntimeContext) (messageModifyInput, error) {
+	messageIDs, err := normalizeMessageManageIDs(rt.StrArray("message-ids"))
+	if err != nil {
+		return messageModifyInput{}, err
+	}
+	addLabels, customAddLabels, err := normalizeMessageManageLabels(rt.StrSlice("add-label-ids"), "--add-label-ids")
+	if err != nil {
+		return messageModifyInput{}, err
+	}
+	removeLabels, customRemoveLabels, err := normalizeMessageManageLabels(rt.StrSlice("remove-label-ids"), "--remove-label-ids")
+	if err != nil {
+		return messageModifyInput{}, err
+	}
+	if err := validateLabelIntersection(addLabels, removeLabels); err != nil {
+		return messageModifyInput{}, err
+	}
+	folder, customFolder, err := normalizeMessageManageFolder(rt.Str("add-folder"))
+	if err != nil {
+		return messageModifyInput{}, err
+	}
+	customLabels := append(customAddLabels, customRemoveLabels...)
+	customFolderID := ""
+	if customFolder {
+		customFolderID = folder
+	}
+	return messageModifyInput{
+		MessageIDs:         messageIDs,
+		AddLabelIDs:        addLabels,
+		RemoveLabelIDs:     removeLabels,
+		AddFolder:          folder,
+		CustomLabelIDs:     customLabels,
+		CustomFolderID:     customFolderID,
+		ValidationAPIPlans: messageManageValidationPlan(resolveMailboxID(rt), customLabels, customFolderID),
+	}, nil
+}

shortcuts/mail/mail_message_trash.go

@@ -0,0 +1,75 @@
+// Copyright (c) 2026 Lark Technologies Pte. Ltd.
+// SPDX-License-Identifier: MIT
+
+package mail
+
+import (
+	"context"
+
+	"github.com/larksuite/cli/shortcuts/common"
+)
+
+// MailMessageTrash is the `+message-trash` shortcut: soft-delete existing
+// messages in batches of 20 via batch_trash. Risk is high-risk-write, so the
+// runner requires --yes before Execute.
+var MailMessageTrash = common.Shortcut{
+	Service:     "mail",
+	Command:     "+message-trash",
+	Description: "Soft-delete existing mail messages. Batches message IDs in groups of 20 and calls batch_trash sequentially. Requires --yes.",
+	Risk:        "high-risk-write",
+	Scopes:      []string{"mail:user_mailbox.message:modify"},
+	AuthTypes:   []string{"user"},
+	HasFormat:   true,
+	Flags: []common.Flag{
+		{Name: "mailbox", Desc: "Mailbox email address that owns the messages (default: me)."},
+		{Name: "message-ids", Type: "string_array", Required: true, Desc: "Message IDs to soft-delete; comma-separated or repeat the flag."},
+	},
+	Validate: validateMessageTrash,
+	DryRun:   dryRunMessageTrash,
+	Execute:  executeMessageTrash,
+}
+
+func validateMessageTrash(ctx context.Context, rt *common.RuntimeContext) error {
+	_, err := normalizeMessageManageIDs(rt.StrArray("message-ids"))
+	return err
+}
+
+func dryRunMessageTrash(ctx context.Context, rt *common.RuntimeContext) *common.DryRunAPI {
+	mailboxID := resolveMailboxID(rt)
+	messageIDs, _ := normalizeMessageManageIDs(rt.StrArray("message-ids"))
+	api := common.NewDryRunAPI().
+		Desc("Soft-delete messages sequentially in batches of 20").
+		Set("batch_size", mailMessageManageBatchSize).
+		Set("batches", chunkMessageManageIDs(messageIDs))
+	for _, batch := range chunkMessageManageIDs(messageIDs) {
+		api = api.POST(mailboxPath(mailboxID, "messages", "batch_trash")).
+			Body(map[string]interface{}{"message_ids": batch})
+	}
+	return api
+}
+
+func executeMessageTrash(ctx context.Context, rt *common.RuntimeContext) error {
+	mailboxID := resolveMailboxID(rt)
+	messageIDs, err := normalizeMessageManageIDs(rt.StrArray("message-ids"))
+	if err != nil {
+		return err
+	}
+
+	summary := messageManageSummary{FailedMessageIDs: []messageManageFailure{}}
+	for _, batch := range chunkMessageManageIDs(messageIDs) {
+		_, err := rt.CallAPITyped("POST", mailboxPath(mailboxID, "messages", "batch_trash"), nil,
+			map[string]interface{}{"message_ids": batch})
+		if err != nil {
+			for _, id := range batch {
+				summary.FailedMessageIDs = append(summary.FailedMessageIDs, messageManageFailure{MessageID: id, Reason: err.Error()})
+			}
+			continue
+		}
+		summary.SuccessMessageIDs = append(summary.SuccessMessageIDs, batch...)
+	}
+	emitMessageManageSummary(rt, summary, false)
+	if len(summary.SuccessMessageIDs) == 0 && len(summary.FailedMessageIDs) > 0 {
+		return mailFailedPreconditionError("all message trash batches failed")
+	}
+	return nil
+}

shortcuts/mail/mail_shortcut_test.go

@@ -44,7 +44,7 @@ func mailShortcutTestFactory(t *testing.T) (*cmdutil.Factory, *bytes.Buffer, *by
 		RefreshToken:     "test-refresh-token",
 		ExpiresAt:        time.Now().Add(1 * time.Hour).UnixMilli(),
 		RefreshExpiresAt: time.Now().Add(24 * time.Hour).UnixMilli(),
-		Scope:            "mail:user_mailbox.messages:write mail:user_mailbox.messages:read mail:user_mailbox.message:modify mail:user_mailbox.message:readonly mail:user_mailbox.message.address:read mail:user_mailbox.message.subject:read mail:user_mailbox.message.body:read mail:user_mailbox:readonly",
+		Scope:            "mail:user_mailbox.messages:write mail:user_mailbox.messages:read mail:user_mailbox.message:modify mail:user_mailbox.message:readonly mail:user_mailbox.message.address:read mail:user_mailbox.message.subject:read mail:user_mailbox.message.body:read mail:user_mailbox:readonly mail:user_mailbox.folder:read",
 		GrantedAt:        time.Now().Add(-1 * time.Hour).UnixMilli(),
 	}
 	if err := auth.SetStoredToken(token); err != nil {

shortcuts/mail/message_manage_helpers.go

@@ -0,0 +1,283 @@
+// Copyright (c) 2026 Lark Technologies Pte. Ltd.
+// SPDX-License-Identifier: MIT
+
+package mail
+
+import (
+	"fmt"
+	"io"
+	"strings"
+	"unicode"
+
+	"github.com/larksuite/cli/internal/output"
+	"github.com/larksuite/cli/shortcuts/common"
+)
+
+const mailMessageManageBatchSize = 20
+
+var messageManageSystemLabels = map[string]string{
+	"UNREAD":               "UNREAD",
+	"IMPORTANT":            "IMPORTANT",
+	"OTHER":                "OTHER",
+	"FLAGGED":              "FLAGGED",
+	"READ_RECEIPT_REQUEST": "READ_RECEIPT_REQUEST",
+}
+
+var messageManageSystemFolders = map[string]string{
+	"INBOX":    "INBOX",
+	"SENT":     "SENT",
+	"SPAM":     "SPAM",
+	"ARCHIVE":  "ARCHIVED",
+	"ARCHIVED": "ARCHIVED",
+}
+
+type messageManageSummary struct {
+	SuccessMessageIDs []string               `json:"success_message_ids"`
+	FailedMessageIDs  []messageManageFailure `json:"failed_message_ids"`
+}
+
+type messageManageFailure struct {
+	MessageID string `json:"message_id"`
+	Reason    string `json:"reason"`
+}
+
+type validationAPIPlan struct {
+	Method       string `json:"method"`
+	Path         string `json:"path"`
+	WillValidate bool   `json:"will_validate"`
+}
+
+func normalizeMessageManageIDs(raw []string) ([]string, error) {
+	if len(raw) == 0 {
+		return nil, mailValidationParamError("--message-ids", "--message-ids is required")
+	}
+	parts, err := splitMessageManageIDTokens(raw)
+	if err != nil {
+		return nil, err
+	}
+	ids := make([]string, 0, len(parts))
+	seen := make(map[string]struct{}, len(parts))
+	for i, part := range parts {
+		if part == "" {
+			return nil, mailValidationParamError("--message-ids", "--message-ids entry %d is empty; remove extra commas or provide valid message IDs", i+1)
+		}
+		id := strings.TrimSpace(part)
+		if id == "" {
+			return nil, mailValidationParamError("--message-ids", "--message-ids entry %d is empty; remove extra commas or provide valid message IDs", i+1)
+		}
+		if id != part {
+			return nil, mailValidationParamError("--message-ids", "--message-ids entry %d (%q): must not contain leading or trailing whitespace", i+1, part)
+		}
+		if err := validateMessageManageID(id, i); err != nil {
+			return nil, err
+		}
+		if _, ok := seen[id]; ok {
+			continue
+		}
+		seen[id] = struct{}{}
+		ids = append(ids, id)
+	}
+	if len(ids) == 0 {
+		return nil, mailValidationParamError("--message-ids", "--message-ids is required")
+	}
+	return ids, nil
+}
+
+func splitMessageManageIDTokens(raw []string) ([]string, error) {
+	parts := make([]string, 0, len(raw))
+	for i, token := range raw {
+		for _, r := range token {
+			if unicode.IsSpace(r) || unicode.IsControl(r) {
+				return nil, mailValidationParamError("--message-ids", "--message-ids entry %d (%q): must not contain whitespace or control characters", i+1, token)
+			}
+		}
+		parts = append(parts, strings.Split(token, ",")...)
+	}
+	return parts, nil
+}
+
+func validateMessageManageID(id string, index int) error {
+	if len(id) < 16 {
+		return mailValidationParamError("--message-ids", "--message-ids entry %d (%q): length must be at least 16 characters", index+1, id)
+	}
+	if strings.Trim(id, "0123456789") == "" {
+		return mailValidationParamError("--message-ids", "--message-ids entry %d (%q): numeric primary IDs are not supported; pass the Open API message_id from mail output", index+1, id)
+	}
+	for _, r := range id {
+		if unicode.IsSpace(r) || unicode.IsControl(r) {
+			return mailValidationParamError("--message-ids", "--message-ids entry %d (%q): must not contain whitespace or control characters", index+1, id)
+		}
+		if (r >= 'A' && r <= 'Z') || (r >= 'a' && r <= 'z') || (r >= '0' && r <= '9') {
+			continue
+		}
+		switch r {
+		case '+', '/', '=', '_', '-':
+			continue
+		default:
+			return mailValidationParamError("--message-ids", "--message-ids entry %d (%q): contains characters outside the Open API message_id character set", index+1, id)
+		}
+	}
+	return nil
+}
+
+func normalizeMessageManageLabels(raw []string, flagName string) ([]string, []string, error) {
+	labels := make([]string, 0, len(raw))
+	custom := make([]string, 0, len(raw))
+	seen := make(map[string]struct{}, len(raw))
+	for i, part := range raw {
+		id := strings.TrimSpace(part)
+		if id == "" {
+			return nil, nil, mailValidationParamError(flagName, "%s entry %d is empty; remove extra commas or provide valid label IDs", flagName, i+1)
+		}
+		if id != part {
+			return nil, nil, mailValidationParamError(flagName, "%s entry %d (%q): must not contain leading or trailing whitespace", flagName, i+1, part)
+		}
+		normalized := id
+		if system, ok := messageManageSystemLabels[strings.ToUpper(id)]; ok {
+			normalized = system
+		} else {
+			custom = append(custom, id)
+		}
+		if _, ok := seen[normalized]; ok {
+			continue
+		}
+		seen[normalized] = struct{}{}
+		labels = append(labels, normalized)
+	}
+	if len(labels) > 20 {
+		return nil, nil, mailValidationParamError(flagName, "%s accepts at most 20 label IDs (got %d)", flagName, len(labels))
+	}
+	return labels, custom, nil
+}
+
+func validateLabelIntersection(add, remove []string) error {
+	removeSet := make(map[string]struct{}, len(remove))
+	for _, id := range remove {
+		removeSet[id] = struct{}{}
+	}
+	for _, id := range add {
+		if _, ok := removeSet[id]; ok {
+			return mailValidationParamError("--add-label-ids", "label cannot be both added and removed: %s", id)
+		}
+	}
+	return nil
+}
+
+func normalizeMessageManageFolder(raw string) (string, bool, error) {
+	if raw == "" {
+		return "", false, nil
+	}
+	folder := strings.TrimSpace(raw)
+	if folder == "" {
+		return "", false, mailValidationParamError("--add-folder", "--add-folder must not be empty")
+	}
+	if folder != raw {
+		return "", false, mailValidationParamError("--add-folder", "--add-folder %q must not contain leading or trailing whitespace", raw)
+	}
+	if strings.EqualFold(folder, "TRASH") {
+		return "", false, mailValidationParamError("--add-folder", "TRASH is not supported by +message-modify; use +message-trash")
+	}
+	if system, ok := messageManageSystemFolders[strings.ToUpper(folder)]; ok {
+		return system, false, nil
+	}
+	return folder, true, nil
+}
+
+func chunkMessageManageIDs(ids []string) [][]string {
+	if len(ids) == 0 {
+		return nil
+	}
+	chunks := make([][]string, 0, (len(ids)+mailMessageManageBatchSize-1)/mailMessageManageBatchSize)
+	for start := 0; start < len(ids); start += mailMessageManageBatchSize {
+		end := start + mailMessageManageBatchSize
+		if end > len(ids) {
+			end = len(ids)
+		}
+		chunks = append(chunks, ids[start:end])
+	}
+	return chunks
+}
+
+func validateCustomMessageManageLabels(rt *common.RuntimeContext, mailboxID string, ids []string) error {
+	if len(ids) == 0 {
+		return nil
+	}
+	if err := validateLabelReadScope(rt); err != nil {
+		return err
+	}
+	seen := map[string]struct{}{}
+	for _, id := range ids {
+		if _, ok := seen[id]; ok {
+			continue
+		}
+		seen[id] = struct{}{}
+		if _, err := rt.CallAPITyped("GET", mailboxPath(mailboxID, "labels", id), nil, nil); err != nil {
+			return mailDecorateProblemMessage(err, "label not found: %s", id)
+		}
+	}
+	return nil
+}
+
+func validateCustomMessageManageFolder(rt *common.RuntimeContext, mailboxID, id string) error {
+	if id == "" {
+		return nil
+	}
+	if err := validateFolderReadScope(rt); err != nil {
+		return err
+	}
+	if _, err := rt.CallAPITyped("GET", mailboxPath(mailboxID, "folders", id), nil, nil); err != nil {
+		return mailDecorateProblemMessage(err, "folder not found: %s", id)
+	}
+	return nil
+}
+
+func messageManageBody(ids, addLabels, removeLabels []string, addFolder string) map[string]interface{} {
+	body := map[string]interface{}{"message_ids": ids}
+	if len(addLabels) > 0 {
+		body["add_label_ids"] = addLabels
+	}
+	if len(removeLabels) > 0 {
+		body["remove_label_ids"] = removeLabels
+	}
+	if addFolder != "" {
+		body["add_folder"] = addFolder
+	}
+	return body
+}
+
+func messageManageValidationPlan(mailboxID string, customLabels []string, customFolder string) []validationAPIPlan {
+	plans := make([]validationAPIPlan, 0, len(customLabels)+1)
+	seenLabels := map[string]struct{}{}
+	for _, id := range customLabels {
+		if _, ok := seenLabels[id]; ok {
+			continue
+		}
+		seenLabels[id] = struct{}{}
+		plans = append(plans, validationAPIPlan{
+			Method:       "GET",
+			Path:         mailboxPath(mailboxID, "labels", id),
+			WillValidate: true,
+		})
+	}
+	if customFolder != "" {
+		plans = append(plans, validationAPIPlan{
+			Method:       "GET",
+			Path:         mailboxPath(mailboxID, "folders", customFolder),
+			WillValidate: true,
+		})
+	}
+	return plans
+}
+
+func emitMessageManageSummary(rt *common.RuntimeContext, summary messageManageSummary, noAPICalls bool) {
+	rt.OutFormat(summary, &output.Meta{Count: len(summary.SuccessMessageIDs)}, func(w io.Writer) {
+		fmt.Fprintf(w, "success_message_ids: %d\n", len(summary.SuccessMessageIDs))
+		fmt.Fprintf(w, "failed_message_ids: %d\n", len(summary.FailedMessageIDs))
+		if noAPICalls {
+			fmt.Fprintln(w, "No changes requested; no API calls were made.")
+		}
+		for _, item := range summary.FailedMessageIDs {
+			fmt.Fprintf(w, "- %s: %s\n", item.MessageID, item.Reason)
+		}
+	})
+}

shortcuts/mail/shortcuts.go

@@ -10,6 +10,8 @@ func Shortcuts() []common.Shortcut {
 	return []common.Shortcut{
 		MailMessage,
 		MailMessages,
+		MailMessageModify,
+		MailMessageTrash,
 		MailThread,
 		MailTriage,
 		MailWatch,

sidecar/server-demo/handler_test.go

@@ -18,6 +18,7 @@ import (
 	"testing"
 
 	extcred "github.com/larksuite/cli/extension/credential"
+	"github.com/larksuite/cli/internal/core"
 	"github.com/larksuite/cli/internal/credential"
 	"github.com/larksuite/cli/internal/envvars"
 	"github.com/larksuite/cli/sidecar"
@@ -585,11 +586,15 @@ func TestProxyHandler_StripsClientSuppliedAuthHeaders(t *testing.T) {
 }
 
 func TestBuildAllowedHosts(t *testing.T) {
-	feishu := struct{ Open, Accounts, MCP string }{
-		"https://open.feishu.cn", "https://accounts.feishu.cn", "https://mcp.feishu.cn",
+	feishu := core.Endpoints{
+		Open:     "https://open.feishu.cn",
+		Accounts: "https://accounts.feishu.cn",
+		MCP:      "https://mcp.feishu.cn",
 	}
-	lark := struct{ Open, Accounts, MCP string }{
-		"https://open.larksuite.com", "https://accounts.larksuite.com", "https://mcp.larksuite.com",
+	lark := core.Endpoints{
+		Open:     "https://open.larksuite.com",
+		Accounts: "https://accounts.larksuite.com",
+		MCP:      "https://mcp.larksuite.com",
 	}
 	hosts := buildAllowedHosts(feishu, lark)
 	// feishu hosts

skill-template/domains/mail.md

@@ -65,7 +65,7 @@
 
 1. `+triage --from spam@x.com` → 列出 N 条结果
 2. 展示："将删除 N 封邮件（发件人 spam@x.com，主题：…），确认？"
-3. 用户确认后 → `*.batch_trash`
+3. 用户确认后 → `+message-trash --message-ids ... --yes`
 
 ## 身份选择：优先使用 user 身份
 
@@ -82,12 +82,13 @@
 1. **确认身份** — 首次操作邮箱前先调用 `lark-cli mail user_mailboxes profile --params '{"user_mailbox_id":"me"}'` 获取当前用户的真实邮箱地址（`primary_email_address`），不要通过系统用户名猜测。后续判断"发件人是否为用户本人"时以此地址为准。
 2. **浏览** — `+triage` 查看收件箱摘要，获取 `message_id` / `thread_id`
 3. **阅读** — `+message` 读单封邮件，`+thread` 读整个会话
-4. **回复** — `+reply` / `+reply-all`（默认存草稿，加 `--confirm-send` 则立即发送）
-5. **转发** — `+forward`（默认存草稿，加 `--confirm-send` 则立即发送）
-6. **新邮件** — `+send` 存草稿（默认），加 `--confirm-send` 发送
-7. **确认投递** — 立即发送后用 `send_status` 查询投递状态，定时发送后在预定时间后再查询；取消定时发送用 `cancel_scheduled_send`
-8. **编辑草稿** — `+draft-edit` 修改已有草稿。正文编辑通过 `--patch-file`：回复/转发草稿用 `set_reply_body` op 保留引用区，普通草稿用 `set_body` op
-9. **已读回执** —
+4. **整理** — 标签、已读/未读状态和移动文件夹优先用 `+message-modify`；软删除优先用 `+message-trash`
+5. **回复** — `+reply` / `+reply-all`（默认存草稿，加 `--confirm-send` 则立即发送）
+6. **转发** — `+forward`（默认存草稿，加 `--confirm-send` 则立即发送）
+7. **新邮件** — `+send` 存草稿（默认），加 `--confirm-send` 发送
+8. **确认投递** — 立即发送后用 `send_status` 查询投递状态，定时发送后在预定时间后再查询；取消定时发送用 `cancel_scheduled_send`
+9. **编辑草稿** — `+draft-edit` 修改已有草稿。正文编辑通过 `--patch-file`：回复/转发草稿用 `set_reply_body` op 保留引用区，普通草稿用 `set_body` op
+10. **已读回执** —
    - **请求回执（写信侧）**：`--request-receipt` 仅在**用户显式要求**时添加，**不要从 subject / body 内容推断意图**。
    - **响应回执（拉信侧）**：拉信看到 `label_ids` 含 `READ_RECEIPT_REQUEST`（或 `-607`）时，**必须先问用户**是否回执（不要自动回执，涉及隐私）。用户同意 → `+send-receipt` 响应；用户不同意但想消掉提示 → `+decline-receipt` 只清本地标签、不发邮件。
 
@@ -417,7 +418,7 @@ lark-cli mail +message --message-id <id>
 
 ## 原生 API 调用规则
 
-没有 Shortcut 覆盖的操作才使用原生 API。调用步骤以本节为准（API Resources 章节的 resource/method 列表可辅助查阅）。
+没有 Shortcut 覆盖的操作才使用原生 API。标签、已读状态、移动文件夹优先使用 `+message-modify`；软删除优先使用 `+message-trash`。调用步骤以本节为准（API Resources 章节的 resource/method 列表可辅助查阅）。
 
 ### Step 1 — 用 `-h` 确定要调用的 API（必须，不可跳过）
 

skills/lark-calendar/references/lark-calendar-create.md

@@ -32,8 +32,7 @@ lark-cli calendar +create --summary "..." --start "..." --end "..." \
 | `--summary <text>` | 否 | 日程标题。注意：标题中不应该出现时间、地点、人物信息 |
 | `--start <time>` | 是 | 开始时间（ISO 8601，如 `2026-03-12T14:00+08:00`） |
 | `--end <time>` | 是 | 结束时间（ISO 8601） |
-| `--description <text>` | 否 | 日程纯文本描述。提供会议议程、活动内容、注意事项或链接等。与 summary 配合使用，仅关注当前日程信息 |
-| `--description-rich <markdown>` | 否 | 日程富文本描述，使用 **Markdown** 格式。支持加粗、斜体、下划线（`<u>...</u>`）、删除线、链接 `[文本](url)`、有序/无序列表、以及 GFM 表格（`\| 列1 \| 列2 \|` + 分隔行 `\| --- \| --- \|`）。支持 `@文件路径` 或 `-`（stdin）读取。需要富文本格式（如带链接、表格）时使用此参数而非 `--description`。只传 `--description-rich`（不带 `--description`）时，CLI 会自动从富文本派生一段纯文本预览回填 `description`，确保端上创建后即时展示富文本，无需再手动编辑保存 |
+| `--description <text>` | 否 | 日程详细描述。提供会议议程、活动内容、注意事项或链接等。与 summary 配合使用，仅关注当前日程信息 |
 | `--attendee-ids <id_list>` | 否 | 参与人 ID 列表（逗号分隔）。支持用户（`ou_`）、群组（`oc_`）和会议室（`omm_`）。AI 提取时请务必保留对应前缀 |
 | `--calendar-id <id>` | 否 | 日历 ID（省略则使用主日历） |
 | `--rrule <rrule>` | 否 | 重复日程的重复性规则，规则设置方式参考rfc5545。示例值："FREQ=DAILY;INTERVAL=1;UNTIL=<具体日期>" |

skills/lark-calendar/references/lark-calendar-update.md

@@ -43,8 +43,7 @@ lark-cli calendar +update \
 | `--event-id <id>` | 是 | 要更新的日程 ID。重复性日程请根据操作范围选择 ID，详见 [重复性日程操作规范](lark-calendar-recurring.md) |
 | `--calendar-id <id>` | 否 | 日历 ID（省略则使用 `primary`） |
 | `--summary <text>` | 否 | 新日程标题。仅在显式传入 `--summary` 时更新；若传空字符串，会把标题清空 |
-| `--description <text>` | 否 | 新日程纯文本描述。仅在显式传入 `--description` 时更新；若传空字符串，会把描述清空。注意：只传 `--description`（不带 `--description-rich`）会覆盖并清空原有富文本描述 |
-| `--description-rich <markdown>` | 否 | 新日程富文本描述，使用 **Markdown** 格式（加粗、斜体、下划线 `<u>...</u>`、删除线、链接 `[文本](url)`、有序/无序列表、以及 GFM 表格 `\| 列1 \| 列2 \|` + 分隔行 `\| --- \| --- \|`）。支持 `@文件路径` 或 `-`（stdin）读取。仅在显式传入时更新。需保留富文本格式时使用此参数而非 `--description`。 |
+| `--description <text>` | 否 | 新日程描述。目前 API 方式不支持编辑富文本描述；如果日程描述通过客户端编辑为富文本内容，则使用 API 更新描述会导致富文本格式丢失。仅在显式传入 `--description` 时更新；若传空字符串，会把描述清空 |
 | `--start <time>` | 否 | 新开始时间（ISO 8601，如 `2026-03-12T14:00+08:00`）。更新日程时间时必须同时传 `--end` |
 | `--end <time>` | 否 | 新结束时间（ISO 8601）。更新日程时间时必须同时传 `--start` |
 | `--rrule <rrule>` | 否 | 新重复规则（RFC5545）。**不要使用 COUNT；如需限制次数，推算后转为 UNTIL** |
@@ -53,13 +52,12 @@ lark-cli calendar +update \
 | `--notify` | 否 | 是否发送更新通知，默认 `true`。可用 `--notify=false` 静默更新 |
 | `--dry-run` | 否 | 预览 API 调用，不执行 |
 
-至少需要提供一个动作：`--summary`、`--description`、`--description-rich`、`--start/--end`、`--rrule`、`--add-attendee-ids` 或 `--remove-attendee-ids`。
+至少需要提供一个动作：`--summary`、`--description`、`--start/--end`、`--rrule`、`--add-attendee-ids` 或 `--remove-attendee-ids`。
 
 ## 使用规则
 
 - `--add-attendee-ids` 是**增量添加**，不是替换最终参与人列表。不要用它表达“只保留这些人”。
-- 对 `--summary`、`--description`、`--description-rich`，CLI 以“是否显式传入该 flag”判断是否更新，而不是以“值是否为空”判断；如果显式传入空字符串，会把对应字段清空。
-- 富文本描述与纯文本描述互斥存储：需要保留/写入富文本时用 `--description-rich`；只传 `--description` 会清空原有富文本。
+- 对 `--summary`、`--description`，CLI 以“是否显式传入该 flag”判断是否更新，而不是以“值是否为空”判断；如果显式传入空字符串，会把对应字段清空。
 - 只想增删参会人或会议室时，不需要同时传 `--summary`、`--start`、`--end` 等日程字段。
 - 只想修改标题、描述、时间或重复规则时，不需要同时传 `--add-attendee-ids` 或 `--remove-attendee-ids`。
 - 如需替换某个参与人、群组或会议室，使用 `--remove-attendee-ids <旧ID>` + `--add-attendee-ids <新ID>`。

skills/lark-mail/SKILL.md

@@ -79,7 +79,7 @@ metadata:
 
 1. `+triage --from spam@x.com` → 列出 N 条结果
 2. 展示："将删除 N 封邮件（发件人 spam@x.com，主题：…），确认？"
-3. 用户确认后 → `*.batch_trash`
+3. 用户确认后 → `+message-trash --message-ids ... --yes`
 
 ## 身份选择：优先使用 user 身份
 
@@ -96,13 +96,14 @@ metadata:
 1. **确认身份** — 首次操作邮箱前先调用 `lark-cli mail user_mailboxes profile --params '{"user_mailbox_id":"me"}'` 获取当前用户的真实邮箱地址（`primary_email_address`），不要通过系统用户名猜测。后续判断"发件人是否为用户本人"时以此地址为准。
 2. **浏览** — `+triage` 查看收件箱摘要，获取 `message_id` / `thread_id`
 3. **阅读** — `+message` 只读单封邮件；已有多个 `message_id` 时用 `+messages` 批量读取，不要循环调用 `+message`；`+thread` 读整个会话
-4. **回复** — `+reply` / `+reply-all`（默认存草稿，加 `--confirm-send` 则立即发送）
-5. **转发** — `+forward`（默认存草稿，加 `--confirm-send` 则立即发送）
-6. **新邮件** — `+send` 存草稿（默认），加 `--confirm-send` 发送
-7. **HTML body 预检（可选）** — 复杂 HTML body 提交前可先跑 `+lint-html` 看 lint 会改 / 删什么；写信路径（`+send` / `+draft-create` / `+reply` / `+reply-all` / `+forward` / `+draft-edit` body op）已内置 autofix，普通正文不必先跑。详见 [references/lark-mail-html.md](references/lark-mail-html.md) 中的「写入路径内置 HTML lint」章节
-8. **确认投递** — 立即发送后用 `send_status` 查询投递状态，定时发送后在预定时间后再查询；取消定时发送用 `cancel_scheduled_send`
-9. **编辑草稿** — `+draft-edit` 修改已有草稿。正文编辑通过 `--patch-file`：回复/转发草稿用 `set_reply_body` op 保留引用区，普通草稿用 `set_body` op
-10. **已读回执** —
+4. **整理** — 标签、已读/未读状态和移动文件夹优先用 `+message-modify`；软删除优先用 `+message-trash`
+5. **回复** — `+reply` / `+reply-all`（默认存草稿，加 `--confirm-send` 则立即发送）
+6. **转发** — `+forward`（默认存草稿，加 `--confirm-send` 则立即发送）
+7. **新邮件** — `+send` 存草稿（默认），加 `--confirm-send` 发送
+8. **HTML body 预检（可选）** — 复杂 HTML body 提交前可先跑 `+lint-html` 看 lint 会改 / 删什么；写信路径（`+send` / `+draft-create` / `+reply` / `+reply-all` / `+forward` / `+draft-edit` body op）已内置 autofix，普通正文不必先跑。详见 [references/lark-mail-html.md](references/lark-mail-html.md) 中的「写入路径内置 HTML lint」章节
+9. **确认投递** — 立即发送后用 `send_status` 查询投递状态，定时发送后在预定时间后再查询；取消定时发送用 `cancel_scheduled_send`
+10. **编辑草稿** — `+draft-edit` 修改已有草稿。正文编辑通过 `--patch-file`：回复/转发草稿用 `set_reply_body` op 保留引用区，普通草稿用 `set_body` op
+11. **已读回执** —
    - **请求回执（写信侧）**：`--request-receipt` 仅在**用户显式要求**时添加，**不要从 subject / body 内容推断意图**。
    - **响应回执（拉信侧）**：拉信看到 `label_ids` 含 `READ_RECEIPT_REQUEST`（或 `-607`）时，**必须先问用户**是否回执（不要自动回执，涉及隐私）。用户同意 → `+send-receipt` 响应；用户不同意但想消掉提示 → `+decline-receipt` 只清本地标签、不发邮件。
 
@@ -119,6 +120,8 @@ metadata:
 - 查看发送邮件后的投递状态：发送成功后查看邮件投递状态；也覆盖发送拦截。ref: [lark-mail-send-status](references/lark-mail-send-status.md)
 - 使用邮件模板：区分个人模板和静态 HTML 模板，发信类 shortcut 用 `--template-id` 套用模板。ref: [lark-mail-template](references/lark-mail-template.md)
 - 撤回已发送邮件：撤回邮件并查询异步撤回状态。ref: [lark-mail-recall](references/lark-mail-recall.md)
+- 修改邮件标签/已读状态/文件夹：优先使用 `+message-modify`。ref: [`+message-modify`](references/lark-mail-message-modify.md)
+- 软删除邮件：优先使用 `+message-trash`。ref: [`+message-trash`](references/lark-mail-message-trash.md)
 - 收信规则：创建、验证、删除自动处理收到邮件的规则。ref: [lark-mail-rules](references/lark-mail-rules.md)
 - 分享邮件到 IM：分享邮件或会话到群聊、个人会话。ref: [lark-mail-share-to-chat](references/lark-mail-share-to-chat.md)
 - 发送日程邀请邮件：在邮件中嵌入 `text/calendar` 日程邀请。ref: [lark-mail-calendar-invite](references/lark-mail-calendar-invite.md)
@@ -192,7 +195,7 @@ lark-cli mail +messages --message-ids <id1>,<id2>,<id3> --html=false
 
 ## 原生 API 调用规则
 
-没有 Shortcut 覆盖的操作才使用原生 API。调用步骤以本节为准；资源和 method 用 `lark-cli mail -h` / `lark-cli mail <resource> -h` 发现，不在入口保留完整资源表。
+没有 Shortcut 覆盖的操作才使用原生 API。标签、已读状态、移动文件夹优先使用 `+message-modify`；软删除优先使用 `+message-trash`。调用步骤以本节为准；资源和 method 用 `lark-cli mail -h` / `lark-cli mail <resource> -h` 发现，不在入口保留完整资源表。
 
 ### Step 1 — 用 `-h` 确定要调用的 API（必须，不可跳过）
 

skills/lark-mail/references/lark-mail-forward.md

@@ -215,7 +215,7 @@ lark-cli mail user_mailbox.drafts cancel_scheduled_send --params '{"user_mailbox
 **2. 标记已读**（可选）— 询问用户是否需要将原邮件标记为已读。如果用户同意：
 
 ```bash
-lark-cli mail user_mailbox.messages batch_modify --params '{"user_mailbox_id":"me"}' --data '{"message_ids":["<原邮件ID>"],"remove_label_ids":["UNREAD"]}'
+lark-cli mail +message-modify --message-ids <原邮件ID> --remove-label-ids UNREAD
 ```
 
 ## 编辑转发草稿

skills/lark-mail/references/lark-mail-message-modify.md

@@ -0,0 +1,48 @@
+# mail +message-modify
+
+`mail +message-modify` is the preferred shortcut for changing labels, read-state labels, or folder placement on existing messages.
+
+Use it instead of raw `user_mailbox.messages batch_modify` when the operation targets concrete `message_id` values from `+triage`, `+message`, or `+messages`.
+
+## Common Commands
+
+```bash
+lark-cli mail +message-modify --message-ids <id1>,<id2> --add-label-ids unread
+lark-cli mail +message-modify --message-ids <id> --remove-label-ids FLAGGED
+lark-cli mail +message-modify --message-ids <id> --add-folder archive
+lark-cli mail +message-modify --mailbox shared@example.com --message-ids <id> --add-folder folder_xxx
+lark-cli mail +message-modify --message-ids <id> --add-label-ids custom_label_id --dry-run
+```
+
+## Flags
+
+| Flag | Required | Notes |
+| --- | --- | --- |
+| `--mailbox` | No | Mailbox that owns the messages. Defaults to `me`. |
+| `--message-ids` | Yes | `string_array`; supports comma-separated values and repeated flags. |
+| `--add-label-ids` | No | Adds labels. System labels `unread`, `important`, `other`, `flagged` normalize to upper case. |
+| `--remove-label-ids` | No | Removes labels. Cannot overlap with `--add-label-ids`. |
+| `--add-folder` | No | Moves to one folder. `inbox`, `sent`, `spam`, `archive`, `archived` normalize to system folder IDs. |
+
+`TRASH` is intentionally rejected by this shortcut. Use `mail +message-trash --message-ids <id> --yes` for soft deletion.
+
+## Behavior
+
+- Message IDs are locally validated, de-duplicated in first-seen order, and sent in batches of 20.
+- Custom label IDs are checked with `labels.get`; custom folder IDs are checked with `folders.get`.
+- If no label or folder operation is requested, the command succeeds locally, emits all message IDs as `success_message_ids`, and makes no POST request.
+- Single batch POST failures mark every message in that batch with the same failure reason; later batches still run.
+- JSON output is intentionally compact:
+
+```json
+{
+  "success_message_ids": ["id1"],
+  "failed_message_ids": [
+    {"message_id": "id2", "reason": "api error"}
+  ]
+}
+```
+
+## When Raw API Is Still Appropriate
+
+Use raw `mail user_mailbox.messages batch_modify` only when you need a request shape that the shortcut intentionally does not expose, or when reproducing backend/API behavior exactly for diagnostics.

skills/lark-mail/references/lark-mail-message-trash.md

@@ -0,0 +1,41 @@
+# mail +message-trash
+
+`mail +message-trash` is the preferred shortcut for soft-deleting existing messages.
+
+Use it after obtaining real `message_id` values from `+triage`, `+message`, or `+messages`, and after the user has confirmed the deletion preview.
+
+## Common Commands
+
+```bash
+lark-cli mail +message-trash --message-ids <id1>,<id2> --yes
+lark-cli mail +message-trash --mailbox shared@example.com --message-ids <id> --yes
+lark-cli mail +message-trash --message-ids <id1> --message-ids <id2> --dry-run
+```
+
+## Flags
+
+| Flag | Required | Notes |
+| --- | --- | --- |
+| `--mailbox` | No | Mailbox that owns the messages. Defaults to `me`. |
+| `--message-ids` | Yes | `string_array`; supports comma-separated values and repeated flags. |
+| `--yes` | Yes for execution | Required by the high-risk write confirmation framework. |
+
+## Behavior
+
+- Message IDs are locally validated, de-duplicated in first-seen order, and sent in batches of 20.
+- The shortcut calls `POST /open-apis/mail/v1/user_mailboxes/<mailbox>/messages/batch_trash` sequentially.
+- Single batch POST failures mark every message in that batch with the same failure reason; later batches still run.
+- JSON output is intentionally compact:
+
+```json
+{
+  "success_message_ids": ["id1"],
+  "failed_message_ids": [
+    {"message_id": "id2", "reason": "api error"}
+  ]
+}
+```
+
+## When Raw API Is Still Appropriate
+
+Use raw `mail user_mailbox.messages batch_trash` only when reproducing backend/API behavior exactly for diagnostics. For normal soft deletion, prefer this shortcut because it handles validation, batching, compact output, and `--yes` confirmation consistently.

skills/lark-mail/references/lark-mail-reply-all.md

@@ -203,7 +203,7 @@ lark-cli mail user_mailbox.drafts cancel_scheduled_send --params '{"user_mailbox
 **2. 标记已读**（可选）— 询问用户是否需要将原邮件标记为已读。如果用户同意：
 
 ```bash
-lark-cli mail user_mailbox.messages batch_modify --params '{"user_mailbox_id":"me"}' --data '{"message_ids":["<原邮件ID>"],"remove_label_ids":["UNREAD"]}'
+lark-cli mail +message-modify --message-ids <原邮件ID> --remove-label-ids UNREAD
 ```
 
 ## 相关命令

skills/lark-mail/references/lark-mail-reply.md

@@ -218,7 +218,7 @@ lark-cli mail user_mailbox.drafts cancel_scheduled_send --params '{"user_mailbox
 **2. 标记已读**（可选）— 询问用户是否需要将原邮件标记为已读。如果用户同意：
 
 ```bash
-lark-cli mail user_mailbox.messages batch_modify --params '{"user_mailbox_id":"me"}' --data '{"message_ids":["<原邮件ID>"],"remove_label_ids":["UNREAD"]}'
+lark-cli mail +message-modify --message-ids <原邮件ID> --remove-label-ids UNREAD
 ```
 
 ## 编辑回复草稿

tests/cli_e2e/calendar/calendar_description_rich_dryrun_test.go

@@ -1,117 +0,0 @@
-// Copyright (c) 2026 Lark Technologies Pte. Ltd.
-// SPDX-License-Identifier: MIT
-
-package calendar
-
-import (
-	"context"
-	"testing"
-	"time"
-
-	clie2e "github.com/larksuite/cli/tests/cli_e2e"
-	"github.com/stretchr/testify/require"
-	"github.com/tidwall/gjson"
-)
-
-// richTextMarkdown is a Markdown rich-text payload carrying a doc link and
-// styling. The CLI forwards it verbatim in description_rich; the OpenAPI service
-// converts Markdown <-> ClientVars.
-const richTextMarkdown = "见 [设计文档](https://bytedance.feishu.cn/docx/abc) 和 **重点**"
-
-// TestCalendar_CreateDescriptionRichDryRun verifies that +create forwards the
-// rich-text payload as-is under the description_rich body field.
-func TestCalendar_CreateDescriptionRichDryRun(t *testing.T) {
-	t.Setenv("LARKSUITE_CLI_CONFIG_DIR", t.TempDir())
-	t.Setenv("LARKSUITE_CLI_APP_ID", "app")
-	t.Setenv("LARKSUITE_CLI_APP_SECRET", "secret")
-	t.Setenv("LARKSUITE_CLI_BRAND", "feishu")
-
-	ctx, cancel := context.WithTimeout(context.Background(), 30*time.Second)
-	t.Cleanup(cancel)
-
-	result, err := clie2e.RunCmd(ctx, clie2e.Request{
-		Args: []string{
-			"calendar", "+create",
-			"--calendar-id", "cal_dry",
-			"--summary", "rich dry-run",
-			"--start", "2026-04-25T10:00:00+08:00",
-			"--end", "2026-04-25T11:00:00+08:00",
-			"--description", "plain fallback",
-			"--description-rich", richTextMarkdown,
-			"--dry-run",
-		},
-		DefaultAs: "bot",
-	})
-	require.NoError(t, err)
-	result.AssertExitCode(t, 0)
-
-	out := result.Stdout
-	require.Equal(t, "POST", gjson.Get(out, "api.0.method").String(), "stdout:\n%s", out)
-	require.Equal(t, "/open-apis/calendar/v4/calendars/cal_dry/events", gjson.Get(out, "api.0.url").String(), "stdout:\n%s", out)
-	require.Equal(t, "plain fallback", gjson.Get(out, "api.0.body.description").String(), "stdout:\n%s", out)
-	require.Equal(t, richTextMarkdown, gjson.Get(out, "api.0.body.description_rich").String(), "stdout:\n%s", out)
-}
-
-// TestCalendar_CreateDescriptionRichOnlyDryRun verifies that when only
-// --description-rich is provided (no --description), +create omits the
-// description body field entirely. Sending an empty description would suppress
-// the server's plain-preview backfill and break first-load rendering.
-func TestCalendar_CreateDescriptionRichOnlyDryRun(t *testing.T) {
-	t.Setenv("LARKSUITE_CLI_CONFIG_DIR", t.TempDir())
-	t.Setenv("LARKSUITE_CLI_APP_ID", "app")
-	t.Setenv("LARKSUITE_CLI_APP_SECRET", "secret")
-	t.Setenv("LARKSUITE_CLI_BRAND", "feishu")
-
-	ctx, cancel := context.WithTimeout(context.Background(), 30*time.Second)
-	t.Cleanup(cancel)
-
-	result, err := clie2e.RunCmd(ctx, clie2e.Request{
-		Args: []string{
-			"calendar", "+create",
-			"--calendar-id", "cal_dry",
-			"--summary", "rich only dry-run",
-			"--start", "2026-04-25T10:00:00+08:00",
-			"--end", "2026-04-25T11:00:00+08:00",
-			"--description-rich", richTextMarkdown,
-			"--dry-run",
-		},
-		DefaultAs: "bot",
-	})
-	require.NoError(t, err)
-	result.AssertExitCode(t, 0)
-
-	out := result.Stdout
-	require.Equal(t, richTextMarkdown, gjson.Get(out, "api.0.body.description_rich").String(), "stdout:\n%s", out)
-	require.False(t, gjson.Get(out, "api.0.body.description").Exists(), "description must be omitted when not set; stdout:\n%s", out)
-}
-
-// TestCalendar_UpdateDescriptionRichDryRun verifies that +update forwards the
-// rich-text payload as-is under the description_rich body field.
-func TestCalendar_UpdateDescriptionRichDryRun(t *testing.T) {
-	t.Setenv("LARKSUITE_CLI_CONFIG_DIR", t.TempDir())
-	t.Setenv("LARKSUITE_CLI_APP_ID", "app")
-	t.Setenv("LARKSUITE_CLI_APP_SECRET", "secret")
-	t.Setenv("LARKSUITE_CLI_BRAND", "feishu")
-
-	ctx, cancel := context.WithTimeout(context.Background(), 30*time.Second)
-	t.Cleanup(cancel)
-
-	result, err := clie2e.RunCmd(ctx, clie2e.Request{
-		Args: []string{
-			"calendar", "+update",
-			"--calendar-id", "cal_dry",
-			"--event-id", "evt_dry",
-			"--description-rich", richTextMarkdown,
-			"--notify=false",
-			"--dry-run",
-		},
-		DefaultAs: "bot",
-	})
-	require.NoError(t, err)
-	result.AssertExitCode(t, 0)
-
-	out := result.Stdout
-	require.Equal(t, "PATCH", gjson.Get(out, "api.0.method").String(), "stdout:\n%s", out)
-	require.Equal(t, "/open-apis/calendar/v4/calendars/cal_dry/events/evt_dry", gjson.Get(out, "api.0.url").String(), "stdout:\n%s", out)
-	require.Equal(t, richTextMarkdown, gjson.Get(out, "api.0.body.description_rich").String(), "stdout:\n%s", out)
-}

tests/cli_e2e/core.go

@@ -24,7 +24,17 @@ import (
 const EnvBinaryPath = "LARK_CLI_BIN"
 const projectRootMarkerDir = "tests"
 const cliBinaryName = "lark-cli"
-const CleanupTimeout = 30 * time.Second
+
+const (
+	// CleanupTimeout is the outer teardown budget. Keep it above any
+	// per-resource wait so cleanup command retries still have room to run.
+	CleanupTimeout = 60 * time.Second
+
+	defaultRetryAttempts        = 4
+	defaultRetryInitialDelay    = time.Second
+	defaultRetryMaxDelay        = 6 * time.Second
+	defaultRetryBackoffMultiple = 2
+)
 
 func SkipWithoutUserToken(t *testing.T) {
 	t.Helper()
@@ -102,6 +112,34 @@ type Result struct {
 	RunErr     error
 }
 
+type cleanupWarningError struct {
+	err error
+}
+
+func (e *cleanupWarningError) Error() string {
+	return e.err.Error()
+}
+
+func (e *cleanupWarningError) Unwrap() error {
+	return e.err
+}
+
+// CleanupWarning marks a cleanup verification issue as non-fatal after the
+// destructive cleanup command itself has already succeeded.
+func CleanupWarning(err error) error {
+	if err == nil {
+		return nil
+	}
+	return &cleanupWarningError{err: err}
+}
+
+// IsCleanupWarning reports whether err should be logged without failing the
+// parent test.
+func IsCleanupWarning(err error) bool {
+	var warning *cleanupWarningError
+	return errors.As(err, &warning)
+}
+
 // RetryOptions configures retry behavior for flaky external API calls.
 type RetryOptions struct {
 	Attempts        int
@@ -111,8 +149,25 @@ type RetryOptions struct {
 	ShouldRetry     func(*Result) bool
 }
 
+// WaitOptions configures a bounded poll loop for eventually consistent cleanup
+// or verification checks.
+type WaitOptions struct {
+	Timeout      time.Duration
+	Interval     time.Duration
+	TimeoutError func() error
+}
+
 // RunCmd executes lark-cli and captures stdout/stderr/exit code.
+// Service errors that return {"error":{"retryable":true}} are retried with
+// bounded exponential backoff so individual tests do not need to remember
+// RunCmdWithRetry for normal transient server contention.
 func RunCmd(ctx context.Context, req Request) (*Result, error) {
+	return RunCmdWithRetry(ctx, req, RetryOptions{
+		ShouldRetry: ResultHasRetryableError,
+	})
+}
+
+func runCmdOnce(ctx context.Context, req Request) (*Result, error) {
 	binaryPath, err := ResolveBinaryPath(req)
 	if err != nil {
 		return nil, err
@@ -186,16 +241,16 @@ func buildCommandEnv(req Request) []string {
 // RunCmdWithRetry reruns a command when the result matches the configured retry condition.
 func RunCmdWithRetry(ctx context.Context, req Request, opts RetryOptions) (*Result, error) {
 	if opts.Attempts <= 0 {
-		opts.Attempts = 4
+		opts.Attempts = defaultRetryAttempts
 	}
 	if opts.InitialDelay <= 0 {
-		opts.InitialDelay = 1 * time.Second
+		opts.InitialDelay = defaultRetryInitialDelay
 	}
 	if opts.MaxDelay <= 0 {
-		opts.MaxDelay = 6 * time.Second
+		opts.MaxDelay = defaultRetryMaxDelay
 	}
 	if opts.BackoffMultiple <= 1 {
-		opts.BackoffMultiple = 2
+		opts.BackoffMultiple = defaultRetryBackoffMultiple
 	}
 	if opts.ShouldRetry == nil {
 		opts.ShouldRetry = func(result *Result) bool {
@@ -206,7 +261,7 @@ func RunCmdWithRetry(ctx context.Context, req Request, opts RetryOptions) (*Resu
 	delay := opts.InitialDelay
 	var lastResult *Result
 	for attempt := 1; attempt <= opts.Attempts; attempt++ {
-		result, err := RunCmd(ctx, req)
+		result, err := runCmdOnce(ctx, req)
 		if err != nil {
 			return nil, err
 		}
@@ -234,6 +289,63 @@ func RunCmdWithRetry(ctx context.Context, req Request, opts RetryOptions) (*Resu
 	return lastResult, nil
 }
 
+// ResultHasRetryableError reports whether lark-cli returned a structured
+// service error with error.retryable=true in either output stream.
+func ResultHasRetryableError(result *Result) bool {
+	if result == nil {
+		return false
+	}
+	return rawHasRetryableError(result.Stdout) || rawHasRetryableError(result.Stderr)
+}
+
+func rawHasRetryableError(raw string) bool {
+	payload := extractJSONPayload(raw)
+	if payload == "" {
+		return false
+	}
+	return gjson.Get(payload, "error.retryable").Bool()
+}
+
+// WaitForCondition polls condition until it returns true, an error, the context
+// is canceled, or the configured timeout expires.
+func WaitForCondition(ctx context.Context, opts WaitOptions, condition func() (bool, error)) error {
+	if condition == nil {
+		return errors.New("wait condition is nil")
+	}
+	if opts.Timeout <= 0 {
+		opts.Timeout = CleanupTimeout
+	}
+	if opts.Interval <= 0 {
+		opts.Interval = time.Second
+	}
+
+	deadline := time.NewTimer(opts.Timeout)
+	defer deadline.Stop()
+	ticker := time.NewTicker(opts.Interval)
+	defer ticker.Stop()
+
+	for {
+		done, err := condition()
+		if err != nil {
+			return err
+		}
+		if done {
+			return nil
+		}
+
+		select {
+		case <-ctx.Done():
+			return ctx.Err()
+		case <-deadline.C:
+			if opts.TimeoutError != nil {
+				return opts.TimeoutError()
+			}
+			return fmt.Errorf("condition still false after %s", opts.Timeout)
+		case <-ticker.C:
+		}
+	}
+}
+
 // GenerateSuffix returns a high-entropy UTC timestamp suffix suitable for remote test resource names.
 func GenerateSuffix() string {
 	now := time.Now().UTC()
@@ -251,6 +363,10 @@ func ReportCleanupFailure(parentT *testing.T, prefix string, result *Result, err
 	parentT.Helper()
 
 	if err != nil {
+		if IsCleanupWarning(err) {
+			parentT.Logf("%s: %v", prefix, err)
+			return
+		}
 		parentT.Errorf("%s: %v", prefix, err)
 		return
 	}
@@ -271,26 +387,11 @@ func isCleanupSuppressedResult(result *Result) bool {
 		return false
 	}
 
-	raw := strings.TrimSpace(result.Stdout)
-	if raw == "" {
-		raw = strings.TrimSpace(result.Stderr)
+	payload := extractJSONPayload(result.Stdout)
+	if payload == "" {
+		payload = extractJSONPayload(result.Stderr)
 	}
-	if raw == "" {
-		return false
-	}
-
-	start := strings.LastIndex(raw, "\n{")
-	if start >= 0 {
-		start++
-	} else {
-		start = strings.Index(raw, "{")
-	}
-	if start < 0 {
-		return false
-	}
-
-	payload := raw[start:]
-	if !gjson.Valid(payload) {
+	if payload == "" {
 		return false
 	}
 
@@ -306,6 +407,32 @@ func isCleanupSuppressedResult(result *Result) bool {
 	return errType == "api_error" && (errCode == 800004135 || strings.Contains(errMessage, " limited"))
 }
 
+func extractJSONPayload(raw string) string {
+	raw = strings.TrimSpace(raw)
+	if raw == "" {
+		return ""
+	}
+	if gjson.Valid(raw) {
+		return raw
+	}
+
+	start := strings.LastIndex(raw, "\n{")
+	if start >= 0 {
+		start++
+	} else {
+		start = strings.Index(raw, "{")
+	}
+	if start < 0 {
+		return ""
+	}
+
+	payload := raw[start:]
+	if !gjson.Valid(payload) {
+		return ""
+	}
+	return payload
+}
+
 // ResolveBinaryPath finds the CLI binary path using request, env, then PATH.
 func ResolveBinaryPath(req Request) (string, error) {
 	if req.BinaryPath != "" {

tests/cli_e2e/core_test.go

@@ -5,10 +5,12 @@ package clie2e
 
 import (
 	"context"
+	"errors"
 	"os"
 	"path/filepath"
 	"strings"
 	"testing"
+	"time"
 
 	"github.com/stretchr/testify/assert"
 	"github.com/stretchr/testify/require"
@@ -223,6 +225,88 @@ func TestRunCmd(t *testing.T) {
 		assert.NotContains(t, env, "LARKSUITE_CLI_APP_ID=cli_app_test")
 		assert.NotContains(t, env, "LARKSUITE_CLI_USER_ACCESS_TOKEN=uat_test")
 	})
+
+	t.Run("retries structured retryable service errors by default", func(t *testing.T) {
+		fake := newFakeCLI(t)
+		statePath := filepath.Join(t.TempDir(), "retry-count")
+		result, err := RunCmd(context.Background(), Request{
+			BinaryPath: fake.BinaryPath,
+			Args:       []string{"fail-once-retryable", statePath},
+		})
+		require.NoError(t, err)
+		result.AssertExitCode(t, 0)
+		result.AssertStdoutStatus(t, true)
+
+		countBytes, err := os.ReadFile(statePath)
+		require.NoError(t, err)
+		assert.Equal(t, "2\n", string(countBytes))
+	})
+
+	t.Run("does not retry non-retryable service errors by default", func(t *testing.T) {
+		fake := newFakeCLI(t)
+		statePath := filepath.Join(t.TempDir(), "retry-count")
+		result, err := RunCmd(context.Background(), Request{
+			BinaryPath: fake.BinaryPath,
+			Args:       []string{"always-non-retryable", statePath},
+		})
+		require.NoError(t, err)
+		result.AssertExitCode(t, 1)
+
+		countBytes, err := os.ReadFile(statePath)
+		require.NoError(t, err)
+		assert.Equal(t, "1\n", string(countBytes))
+	})
+}
+
+func TestRunCmdWithRetry(t *testing.T) {
+	t.Run("does not include RunCmd default retry as a nested retry", func(t *testing.T) {
+		fake := newFakeCLI(t)
+		statePath := filepath.Join(t.TempDir(), "retry-count")
+		result, err := RunCmdWithRetry(context.Background(), Request{
+			BinaryPath: fake.BinaryPath,
+			Args:       []string{"fail-once-retryable", statePath},
+		}, RetryOptions{
+			Attempts:     1,
+			InitialDelay: time.Millisecond,
+			MaxDelay:     time.Millisecond,
+			ShouldRetry:  ResultHasRetryableError,
+		})
+		require.NoError(t, err)
+		result.AssertExitCode(t, 1)
+
+		countBytes, err := os.ReadFile(statePath)
+		require.NoError(t, err)
+		assert.Equal(t, "1\n", string(countBytes))
+	})
+}
+
+func TestWaitForCondition(t *testing.T) {
+	t.Run("polls until condition succeeds", func(t *testing.T) {
+		attempts := 0
+		err := WaitForCondition(context.Background(), WaitOptions{
+			Timeout:  50 * time.Millisecond,
+			Interval: time.Millisecond,
+		}, func() (bool, error) {
+			attempts++
+			return attempts == 2, nil
+		})
+
+		require.NoError(t, err)
+		assert.Equal(t, 2, attempts)
+	})
+
+	t.Run("returns custom timeout error", func(t *testing.T) {
+		wantErr := errors.New("still visible")
+		err := WaitForCondition(context.Background(), WaitOptions{
+			Timeout:      time.Millisecond,
+			Interval:     time.Millisecond,
+			TimeoutError: func() error { return wantErr },
+		}, func() (bool, error) {
+			return false, nil
+		})
+
+		assert.ErrorIs(t, err, wantErr)
+	})
 }
 
 type fakeCLI struct {
@@ -260,6 +344,35 @@ if [ "$1" = "emit-stdin" ]; then
   exit 0
 fi
 
+if [ "$1" = "fail-once-retryable" ]; then
+  state="$2"
+  count=0
+  if [ -f "$state" ]; then
+    count="$(cat "$state")"
+  fi
+  count=$((count + 1))
+  echo "$count" > "$state"
+  if [ "$count" -eq 1 ]; then
+    echo "Deleting folder fake..." >&2
+    echo '{"ok":false,"error":{"type":"api","code":1061045,"message":"resource contention occurred, please retry.","retryable":true}}' >&2
+    exit 1
+  fi
+  echo '{"ok":true}'
+  exit 0
+fi
+
+if [ "$1" = "always-non-retryable" ]; then
+  state="$2"
+  count=0
+  if [ -f "$state" ]; then
+    count="$(cat "$state")"
+  fi
+  count=$((count + 1))
+  echo "$count" > "$state"
+  echo '{"ok":false,"error":{"type":"api","code":123,"message":"validation failed","retryable":false}}' >&2
+  exit 1
+fi
+
 exit_code=0
 while [ "$#" -gt 0 ]; do
   case "$1" in

tests/cli_e2e/docs/coverage.md

@@ -21,7 +21,7 @@
 | Status | Cmd | Type | Testcase | Key parameter shapes | Notes / uncovered reason |
 | --- | --- | --- | --- | --- | --- |
 | ✓ | docs +create | shortcut | docs/helpers_test.go::createDocWithRetry; docs_create_fetch_test.go::TestDocs_CreateAndFetchWorkflowAsUser/create as user; docs_update_dryrun_test.go::TestDocs_DryRunDefaultsToV2OpenAPI/create; docs_update_dryrun_test.go::TestDocs_CreateTitleDryRunPrependsContent | `--parent-token`; `--doc-format markdown`; `--content`; `--title` | helper asserts returned doc id from `data.document.document_id`; dry-run asserts title is prepended into request body content |
-| ✓ | docs +fetch | shortcut | docs_create_fetch_test.go::TestDocs_CreateAndFetchWorkflow/fetch as bot; docs_update_test.go::TestDocs_UpdateWorkflow/verify as bot; docs_create_fetch_test.go::TestDocs_CreateAndFetchWorkflowAsUser/fetch as user; docs_update_dryrun_test.go::TestDocs_DryRunDefaultsToV2OpenAPI/fetch | `--doc <docToken>`; `--doc-format markdown`; default `extra_param.enable_user_cite_reference_map=true` | |
+| ✓ | docs +fetch | shortcut | docs_fetch_dryrun_test.go::TestDocsFetchDryRunIgnoresAPIVersionCompatFlag; docs_create_fetch_test.go::TestDocs_CreateAndFetchWorkflow/fetch as bot; docs_update_test.go::TestDocs_UpdateWorkflow/verify as bot; docs_create_fetch_test.go::TestDocs_CreateAndFetchWorkflowAsUser/fetch as user; docs_update_dryrun_test.go::TestDocs_DryRunDefaultsToV2OpenAPI/fetch | `--doc <docToken>`; `--doc-format markdown`; default `extra_param.enable_user_cite_reference_map=true`; `--api-version v1` compatibility flag still dry-runs the v2 fetch endpoint | |
 | ✓ | docs +history-list | shortcut | docs_update_dryrun_test.go::TestDocs_DryRunDefaultsToV2OpenAPI/history list; docs_history_workflow_test.go::TestDocs_HistoryWorkflow | `--doc`; `--page-size`; `--page-token` | live workflow gated by `LARK_DOC_HISTORY_E2E=1` |
 | ✓ | docs +history-revert | shortcut | docs_update_dryrun_test.go::TestDocs_DryRunDefaultsToV2OpenAPI/history revert; docs_history_workflow_test.go::TestDocs_HistoryWorkflow | `--doc`; `--history-version-id`; `--wait-timeout-ms` | live workflow gated by `LARK_DOC_HISTORY_E2E=1` |
 | ✓ | docs +history-revert-status | shortcut | docs_update_dryrun_test.go::TestDocs_DryRunDefaultsToV2OpenAPI/history revert status; docs_history_workflow_test.go::TestDocs_HistoryWorkflow | `--doc`; `--task-id` | live workflow polls only when revert returns `running` |

tests/cli_e2e/docs/docs_fetch_dryrun_test.go

@@ -1,7 +1,7 @@
 // Copyright (c) 2026 Lark Technologies Pte. Ltd.
 // SPDX-License-Identifier: MIT
 
-package doc
+package docs
 
 import (
 	"context"
@@ -23,7 +23,7 @@ func TestDocsFetchDryRunIgnoresAPIVersionCompatFlag(t *testing.T) {
 		Args: []string{
 			"docs", "+fetch",
 			"--doc", "doxcnDryRunCompat",
-			"--api-version", "legacy",
+			"--api-version", "v1",
 			"--dry-run",
 		},
 		DefaultAs: "bot",

tests/cli_e2e/drive/helpers.go

@@ -14,6 +14,18 @@ import (
 	"github.com/tidwall/gjson"
 )
 
+const (
+	driveDeleteVisibilityTimeout = 30 * time.Second
+	driveDeleteVisibilityPoll    = 3 * time.Second
+)
+
+var driveDeleteVisibilityWait = clie2e.WaitOptions{
+	// This wait only covers the post-delete visibility lag after Drive accepts
+	// deletion. The delete command itself is bounded by clie2e.CleanupContext.
+	Timeout:  driveDeleteVisibilityTimeout,
+	Interval: driveDeleteVisibilityPoll,
+}
+
 // CreateDriveFolder creates a Drive folder, optionally under a parent folder, and
 // deletes it during parent cleanup.
 func CreateDriveFolder(t *testing.T, parentT *testing.T, ctx context.Context, name string, defaultAs string, parentFolderToken string) string {
@@ -60,14 +72,18 @@ func CreateDriveFolder(t *testing.T, parentT *testing.T, ctx context.Context, na
 // returned a suppressed not_found or partial API error but the resource still
 // exists.
 func DeleteDriveResourceAndVerify(ctx context.Context, token, docType, defaultAs string) (*clie2e.Result, error) {
+	return deleteDriveResourceAndVerify(ctx, token, docType, defaultAs, driveDeleteVisibilityWait)
+}
+
+func deleteDriveResourceAndVerify(ctx context.Context, token, docType, defaultAs string, visibilityWait clie2e.WaitOptions) (*clie2e.Result, error) {
 	if defaultAs == "" {
 		defaultAs = "bot"
 	}
 
-	deleteResult, deleteErr := clie2e.RunCmdWithRetry(ctx, clie2e.Request{
+	deleteResult, deleteErr := clie2e.RunCmd(ctx, clie2e.Request{
 		Args:      []string{"drive", "+delete", "--file-token", token, "--type", docType, "--yes"},
 		DefaultAs: defaultAs,
-	}, clie2e.RetryOptions{})
+	})
 	if deleteErr != nil || deleteResult == nil {
 		return deleteResult, deleteErr
 	}
@@ -82,35 +98,21 @@ func DeleteDriveResourceAndVerify(ctx context.Context, token, docType, defaultAs
 		}
 		return deleteResult, fmt.Errorf("drive resource %s/%s still exists after delete failed: exit=%d stdout=%s stderr=%s", docType, token, deleteResult.ExitCode, deleteResult.Stdout, deleteResult.Stderr)
 	}
-	if err := WaitDriveResourceDeleted(ctx, token, docType, defaultAs); err != nil {
-		return deleteResult, err
+	if err := waitDriveResourceDeleted(ctx, token, docType, defaultAs, visibilityWait); err != nil {
+		return deleteResult, clie2e.CleanupWarning(
+			fmt.Errorf("drive resource %s/%s still visible after accepted delete: %w", docType, token, err),
+		)
 	}
 	return deleteResult, nil
 }
 
-func WaitDriveResourceDeleted(ctx context.Context, token, docType, defaultAs string) error {
-	deadline := time.NewTimer(20 * time.Second)
-	defer deadline.Stop()
-	ticker := time.NewTicker(time.Second)
-	defer ticker.Stop()
-
-	for {
-		deleted, err := IsDriveResourceDeleted(ctx, token, docType, defaultAs)
-		if err != nil {
-			return err
-		}
-		if deleted {
-			return nil
-		}
-
-		select {
-		case <-ctx.Done():
-			return ctx.Err()
-		case <-deadline.C:
-			return fmt.Errorf("drive resource %s/%s still exists after delete", docType, token)
-		case <-ticker.C:
-		}
+func waitDriveResourceDeleted(ctx context.Context, token, docType, defaultAs string, opts clie2e.WaitOptions) error {
+	opts.TimeoutError = func() error {
+		return fmt.Errorf("drive resource %s/%s still exists %s after delete", docType, token, opts.Timeout)
 	}
+	return clie2e.WaitForCondition(ctx, opts, func() (bool, error) {
+		return IsDriveResourceDeleted(ctx, token, docType, defaultAs)
+	})
 }
 
 func IsDriveResourceDeleted(ctx context.Context, token, docType, defaultAs string) (bool, error) {

tests/cli_e2e/drive/helpers_test.go

@@ -5,8 +5,13 @@ package drive
 
 import (
 	"context"
+	"os"
+	"path/filepath"
 	"testing"
+	"time"
 
+	clie2e "github.com/larksuite/cli/tests/cli_e2e"
+	"github.com/stretchr/testify/assert"
 	"github.com/stretchr/testify/require"
 )
 
@@ -16,3 +21,59 @@ func createDriveFolder(t *testing.T, parentT *testing.T, ctx context.Context, na
 	require.NotEmpty(t, folderToken)
 	return folderToken
 }
+
+func TestDeleteDriveResourceAndVerify(t *testing.T) {
+	t.Run("successful delete with stale meta returns cleanup warning", func(t *testing.T) {
+		fake := mustWriteDriveCleanupFakeCLI(t)
+		t.Setenv(clie2e.EnvBinaryPath, fake)
+
+		result, err := deleteDriveResourceAndVerify(context.Background(), "fld_stale", "folder", "bot", clie2e.WaitOptions{
+			Timeout:  10 * time.Millisecond,
+			Interval: time.Millisecond,
+		})
+		require.NotNil(t, result)
+		assert.Equal(t, 0, result.ExitCode)
+		require.Error(t, err)
+		assert.True(t, clie2e.IsCleanupWarning(err), "err: %v", err)
+	})
+
+	t.Run("failed delete with existing meta remains fatal", func(t *testing.T) {
+		fake := mustWriteDriveCleanupFakeCLI(t)
+		t.Setenv(clie2e.EnvBinaryPath, fake)
+		t.Setenv("FAKE_DRIVE_DELETE_EXIT", "1")
+
+		result, err := DeleteDriveResourceAndVerify(context.Background(), "fld_existing", "folder", "bot")
+		require.NotNil(t, result)
+		assert.Equal(t, 1, result.ExitCode)
+		require.Error(t, err)
+		assert.False(t, clie2e.IsCleanupWarning(err), "err: %v", err)
+		assert.Contains(t, err.Error(), "still exists after delete failed")
+	})
+}
+
+func mustWriteDriveCleanupFakeCLI(t *testing.T) string {
+	t.Helper()
+
+	script := `#!/bin/sh
+if [ "$1" = "drive" ] && [ "$2" = "+delete" ]; then
+  if [ "${FAKE_DRIVE_DELETE_EXIT:-0}" != "0" ]; then
+    echo '{"ok":false,"error":{"type":"api","message":"delete failed"}}' >&2
+    exit "$FAKE_DRIVE_DELETE_EXIT"
+  fi
+  echo '{"ok":true}'
+  exit 0
+fi
+
+if [ "$1" = "api" ] && [ "$2" = "post" ] && [ "$3" = "/open-apis/drive/v1/metas/batch_query" ]; then
+  echo '{"ok":true,"data":{"metas":[{"url":"https://example.com/still-visible"}]}}'
+  exit 0
+fi
+
+echo "unexpected fake CLI args: $*" >&2
+exit 2
+`
+
+	binaryPath := filepath.Join(t.TempDir(), "fake-lark-cli")
+	require.NoError(t, os.WriteFile(binaryPath, []byte(script), 0o755))
+	return binaryPath
+}

tests/cli_e2e/mail/coverage.md

@@ -1,9 +1,9 @@
 # Mail CLI E2E Coverage
 
 ## Metrics
-- Denominator: 63 leaf commands
-- Covered: 14
-- Coverage: 22.2%
+- Denominator: 65 leaf commands
+- Covered: 16
+- Coverage: 24.6%
 
 ## Summary
 - TestMail_DraftLifecycleWorkflowAsUser: proves a self-contained user draft workflow across `mail user_mailboxes profile`, `mail +draft-create`, `mail user_mailbox.drafts list`, `mail user_mailbox.drafts get`, `mail +draft-edit`, and `mail user_mailbox.drafts delete`; key `t.Run(...)` proof points are `get mailbox profile as user`, `create draft with shortcut as user`, `list draft as user`, `get created draft as user`, `inspect created draft as user`, `update draft subject with shortcut as user`, `inspect updated draft as user`, `delete draft as user`, and `verify draft removed from list as user`.
@@ -20,6 +20,8 @@
 | ✓ | mail +draft-send | shortcut | mail_draft_send_workflow_test.go::TestMail_DraftSendWorkflowAsUser/send draft with shortcut as user; mail_draft_send_dryrun_test.go::TestMail_DraftSendDryRun | `--draft-id`; `--mailbox me`; `--yes`; dry-run repeated/comma-separated `--draft-id` | sends a self-addressed draft through the batch shortcut and locks dry-run request shape |
 | ✓ | mail +forward | shortcut | mail_send_workflow_test.go::TestMail_SendWorkflowAsUser/forward received message with shortcut as user; mail_send_workflow_test.go::TestMail_SendWorkflowAsUser/inspect forward draft as user | `--message-id`; `--to`; `--body`; `--plain-text` | uses self-generated inbox message as source and inspects forwarded draft projection |
 | ✓ | mail +message | shortcut | mail_send_workflow_test.go::TestMail_SendWorkflowAsUser/get sent message as user; mail_send_workflow_test.go::TestMail_SendWorkflowAsUser/get received message as user | `--mailbox me`; `--message-id` | verifies both SENT and INBOX copies after self-send |
+| ✓ | mail +message-modify | shortcut | shortcuts/mail/mail_message_manage_test.go::TestMessageModify_DryRunShowsPlanWithoutValidationGET; shortcuts/mail/mail_message_manage_test.go::TestMessageModify_BatchesAndAggregatesPartialFailure | `--message-ids`; `--add-label-ids`; `--remove-label-ids`; `--add-folder`; `--dry-run` | unit/dry-run coverage locks validation, batching, request shape, and partial failure aggregation; live E2E needs controlled disposable messages/labels/folders |
+| ✓ | mail +message-trash | shortcut | shortcuts/mail/mail_message_manage_test.go::TestMessageTrash_RequiresYesAndBatches | `--message-ids`; `--yes`; `--dry-run` | unit coverage locks high-risk confirmation and batch_trash request shape; live E2E needs controlled disposable messages |
 | ✓ | mail +messages | shortcut | mail_send_workflow_test.go::TestMail_SendWorkflowAsUser/get both self sent messages as user | `--mailbox me`; `--message-ids` | batch reads both sent and received message copies |
 | ✓ | mail +reply | shortcut | mail_send_workflow_test.go::TestMail_SendWorkflowAsUser/reply to received message with shortcut as user; mail_send_workflow_test.go::TestMail_SendWorkflowAsUser/inspect reply draft as user | `--message-id`; `--body`; `--plain-text` | creates reply draft from self-generated inbox message and inspects quoted content |
 | ✕ | mail +reply-all | shortcut |  | none | self-send traffic leaves no stable non-self recipient set for deterministic reply-all assertions |

tests/cli_e2e/wiki/helpers_test.go

@@ -5,6 +5,7 @@ package wiki
 
 import (
 	"context"
+	"errors"
 	"fmt"
 	"strings"
 	"testing"
@@ -203,6 +204,11 @@ type wikiNodeInfo struct {
 	ObjType   string
 }
 
+const (
+	wikiDeleteVisibilityTimeout = 30 * time.Second
+	wikiDeleteVisibilityPoll    = 3 * time.Second
+)
+
 // deleteWikiNodeAndVerify removes a wiki node, then polls get_node until the
 // original node token is gone. Wiki cleanup cannot use drive +delete because
 // wiki origin nodes need the backing obj_token and parent nodes must delete
@@ -333,28 +339,34 @@ func listWikiNodeChildren(ctx context.Context, spaceID, parentNodeToken string)
 }
 
 func waitWikiNodeDeleted(ctx context.Context, nodeToken string) error {
-	deadline := time.NewTimer(20 * time.Second)
-	defer deadline.Stop()
-	ticker := time.NewTicker(time.Second)
-	defer ticker.Stop()
+	var lastTransientErr error
 
-	for {
+	opts := clie2e.WaitOptions{
+		Timeout:  wikiDeleteVisibilityTimeout,
+		Interval: wikiDeleteVisibilityPoll,
+		TimeoutError: func() error {
+			if lastTransientErr != nil {
+				return fmt.Errorf("wiki node %s delete verification kept hitting transient errors: %w", nodeToken, lastTransientErr)
+			}
+			return fmt.Errorf("wiki node %s still exists after delete", nodeToken)
+		},
+	}
+
+	return clie2e.WaitForCondition(ctx, opts, func() (bool, error) {
 		deleted, err := isWikiNodeDeleted(ctx, nodeToken)
 		if err != nil {
-			return err
+			if isWikiVerifyTransientError(err) {
+				lastTransientErr = err
+				return false, nil
+			} else {
+				return false, err
+			}
 		}
 		if deleted {
-			return nil
+			return true, nil
 		}
-
-		select {
-		case <-ctx.Done():
-			return ctx.Err()
-		case <-deadline.C:
-			return fmt.Errorf("wiki node %s still exists after delete", nodeToken)
-		case <-ticker.C:
-		}
-	}
+		return false, nil
+	})
 }
 
 func isWikiNodeDeleted(ctx context.Context, nodeToken string) (bool, error) {
@@ -375,9 +387,31 @@ func isWikiNodeDeleted(ctx context.Context, nodeToken string) (bool, error) {
 	if isWikiNodeDeletedResult(result) {
 		return true, nil
 	}
+	if isWikiVerifyTransientResult(result) {
+		return false, wikiVerifyTransientError{
+			err: fmt.Errorf("verify wiki node %s after delete hit transient response: exit=%d stdout=%s stderr=%s", nodeToken, result.ExitCode, result.Stdout, result.Stderr),
+		}
+	}
 	return false, fmt.Errorf("verify wiki node %s after delete: exit=%d stdout=%s stderr=%s", nodeToken, result.ExitCode, result.Stdout, result.Stderr)
 }
 
+type wikiVerifyTransientError struct {
+	err error
+}
+
+func (e wikiVerifyTransientError) Error() string {
+	return e.err.Error()
+}
+
+func (e wikiVerifyTransientError) Unwrap() error {
+	return e.err
+}
+
+func isWikiVerifyTransientError(err error) bool {
+	var transient wikiVerifyTransientError
+	return err != nil && errors.As(err, &transient)
+}
+
 func wikiAPISuccess(stdout string) bool {
 	if ok := gjson.Get(stdout, "ok"); ok.Exists() {
 		return ok.Bool()
@@ -404,6 +438,55 @@ func isWikiNodeDeletedResult(result *clie2e.Result) bool {
 		strings.Contains(combined, "not found")
 }
 
+func isWikiVerifyTransientResult(result *clie2e.Result) bool {
+	if result == nil {
+		return false
+	}
+	payload := result.Stdout
+	if strings.TrimSpace(payload) == "" {
+		payload = result.Stderr
+	}
+	if gjson.Get(payload, "error.type").String() != "internal" ||
+		gjson.Get(payload, "error.subtype").String() != "invalid_response" {
+		return false
+	}
+	message := strings.ToLower(gjson.Get(payload, "error.message").String())
+	return strings.Contains(message, "http 429") ||
+		strings.Contains(message, "http 500") ||
+		strings.Contains(message, "http 502") ||
+		strings.Contains(message, "http 503") ||
+		strings.Contains(message, "http 504")
+}
+
+func TestWikiVerifyTransientResult(t *testing.T) {
+	t.Run("matches invalid response from transient http status", func(t *testing.T) {
+		result := &clie2e.Result{
+			ExitCode: 5,
+			Stderr:   `{"ok":false,"error":{"type":"internal","subtype":"invalid_response","message":"SDK returned an invalid JSON response: failed to parse TAT response (HTTP 429): invalid character 'r' looking for beginning of value"}}`,
+		}
+
+		require.True(t, isWikiVerifyTransientResult(result))
+	})
+
+	t.Run("does not match unrelated invalid response", func(t *testing.T) {
+		result := &clie2e.Result{
+			ExitCode: 5,
+			Stderr:   `{"ok":false,"error":{"type":"internal","subtype":"invalid_response","message":"SDK returned an invalid JSON response: malformed body"}}`,
+		}
+
+		require.False(t, isWikiVerifyTransientResult(result))
+	})
+
+	t.Run("does not match api errors", func(t *testing.T) {
+		result := &clie2e.Result{
+			ExitCode: 1,
+			Stderr:   `{"ok":false,"error":{"type":"api","subtype":"conflict","message":"resource contention occurred, please retry","retryable":true}}`,
+		}
+
+		require.False(t, isWikiVerifyTransientResult(result))
+	})
+}
+
 func findWikiNodeByToken(t *testing.T, ctx context.Context, spaceID string, nodeToken string, parentNodeTokens ...string) gjson.Result {
 	t.Helper()
 

tests/plugin_e2e/degrade_subsystem_test.go

@@ -0,0 +1,265 @@
+// Copyright (c) 2026 Lark Technologies Pte. Ltd.
+// SPDX-License-Identifier: MIT
+
+package plugin_e2e
+
+import (
+	"fmt"
+	"os"
+	"path/filepath"
+	"strings"
+	"testing"
+	"time"
+
+	"github.com/tidwall/gjson"
+)
+
+// seededCatalogVersion is far newer than the embedded stub's 0.0.0, so the
+// runtime overlay in internal/registry unconditionally applies it.
+const seededCatalogVersion = "9.9.9"
+
+// seededCatalogJSON is a remote_meta.json (registry.MergedRegistry) carrying one
+// obviously-synthetic service. Seeding it into a bare-module fork's on-disk cache
+// gives the runtime catalog real data WITHOUT any network, so a test can prove
+// SchemaCatalog() consults that runtime catalog (issue #1764) rather than the
+// embedded-only (empty stub) catalog. Fields mirror internal/meta.Service.
+const seededCatalogJSON = `{
+  "version": "9.9.9",
+  "services": [
+    {
+      "name": "plugine2e",
+      "version": "v1",
+      "title": "plugin_e2e synthetic service",
+      "description": "synthetic fixture for the runtime-catalog test; not a real API",
+      "servicePath": "/open-apis/plugine2e/v1",
+      "resources": {
+        "widgets": {
+          "methods": {
+            "get": {
+              "id": "plugine2e.widgets.get",
+              "path": "/open-apis/plugine2e/v1/widgets/:id",
+              "httpMethod": "GET",
+              "description": "synthetic read method",
+              "risk": "read",
+              "accessTokens": ["tenant"],
+              "parameters": {
+                "id": {"type": "string", "location": "path", "required": true, "description": "synthetic id"}
+              }
+            }
+          }
+        }
+      }
+    }
+  ]
+}`
+
+// runWithSeededCatalog runs bin against a fresh LARKSUITE_CLI_CONFIG_DIR whose
+// cache already holds cacheJSON as remote_meta.json (plus a fresh, high-version
+// cache-meta so the overlay applies and the TTL never triggers a refetch). Remote
+// meta is left ON so the on-disk cache overlay is consulted, but a long
+// LARKSUITE_CLI_META_TTL keeps the run offline and deterministic. This models a
+// bare-module binary that has runtime metadata available from a warm cache.
+func runWithSeededCatalog(t *testing.T, bin, cacheJSON string, args ...string) result {
+	t.Helper()
+	cfg := t.TempDir()
+	cacheDir := filepath.Join(cfg, "cache")
+	if err := os.MkdirAll(cacheDir, 0o755); err != nil {
+		t.Fatalf("mkdir cache dir: %v", err)
+	}
+	writeFile(t, filepath.Join(cacheDir, "remote_meta.json"), cacheJSON)
+	writeFile(t, filepath.Join(cacheDir, "remote_meta.meta.json"),
+		fmt.Sprintf(`{"last_check_at":%d,"version":%q,"brand":""}`, time.Now().Unix(), seededCatalogVersion))
+	env := append(os.Environ(),
+		"LARKSUITE_CLI_NO_UPDATE_NOTIFIER=1",
+		"LARKSUITE_CLI_NO_SKILLS_NOTIFIER=1",
+		"LARKSUITE_CLI_CONFIG_DIR="+cfg,
+		"LARKSUITE_CLI_META_TTL=1000000",
+	)
+	return runWithEnv(t, bin, env, args...)
+}
+
+// plainPlugin registers a minimal observer-only plugin with NO Restrict rule
+// -- unlike readonly_test.go's plugins, it cannot deny "schema" as
+// out-of-domain, so any failure the command produces below is the command's
+// own behavior against the empty stub catalog, not a policy denial.
+const plainPlugin = `// Code generated by plugin_e2e; DO NOT EDIT.
+package plugin
+
+import (
+	"context"
+
+	"github.com/larksuite/cli/extension/platform"
+)
+
+func init() {
+	platform.Register(
+		platform.NewPlugin("plain", "0.1.0").
+			Observer(platform.After, "noop", platform.All(),
+				func(_ context.Context, _ platform.Invocation) {}).
+			FailOpen().
+			MustBuild())
+}
+`
+
+// TestDegradeStubMetadataSchema pins the #1764 stub-metadata degrade path.
+// The clean tree embeds only the empty meta_data_default.json stub
+// (internal/registry/catalog.go's SchemaCatalog falls through to
+// RuntimeCatalog when EmbeddedServicesTyped() is empty), and run()'s isolated
+// environment disables the remote overlay fetch and points the cache dir at
+// an empty tmp dir, so cmd/schema/schema.go's runSchema sees
+// catalog.Services() == 0 unconditionally -- the exact "offline with a cold
+// cache, remote meta off" branch documented at cmd/schema/schema.go:96-101.
+//
+// Observed real output for both `schema` and `schema im.messages.reply`
+// (identical -- runSchema checks catalog.Services()==0 before parsing args):
+//
+//	exit=2
+//	stdout=(empty)
+//	stderr={"ok":false,"error":{"type":"validation","subtype":"failed_precondition",
+//	  "message":"No API metadata available",
+//	  "hint":"this binary has no embedded API metadata; run any command with
+//	  network access to the open platform once so metadata can be fetched and
+//	  cached"}}
+//
+// This is the PINNED "graceful degrade" criterion: a structured JSON envelope
+// (gjson.Valid, no "panic:" substring) carrying a validation/failed_precondition
+// error with an actionable hint, NOT the raw Go panic crash that
+// install_test.go's TestInstallMustBuildInitPanicCrashesBinary pins for a
+// genuinely broken plugin, and NOT an "Unknown"-shaped internal error.
+// Note: exit==2 alone does not prove "not a crash" -- a genuine Go panic also
+// exits 2. The two real discriminators against a crash are the absence of a
+// "panic:" substring in stderr and stderr being valid JSON (gjson.Valid); both
+// are asserted below.
+func TestDegradeStubMetadataSchema(t *testing.T) {
+	bin := buildFork(t, "plain", plainPlugin)
+	cases := []struct {
+		name string
+		args []string
+	}{
+		{"schema root", []string{"schema"}},
+		{"schema with path", []string{"schema", "im.messages.reply"}},
+	}
+	for _, tc := range cases {
+		t.Run(tc.name, func(t *testing.T) {
+			res := run(t, bin, tc.args...)
+			t.Logf("exit=%d stdout=%s stderr=%s", res.exit, res.stdout, res.stderr)
+			if res.exit != 2 {
+				t.Fatalf("exit=%d want 2 (graceful validation exit); stdout=%s stderr=%s", res.exit, res.stdout, res.stderr)
+			}
+			if strings.Contains(res.stderr, "panic:") {
+				t.Fatalf("stderr contains a raw Go panic trace, not a graceful degrade; stderr=%s", res.stderr)
+			}
+			if !gjson.Valid(res.stderr) {
+				t.Fatalf("stderr not a structured JSON envelope: %s", res.stderr)
+			}
+			if got := gjson.Get(res.stderr, "error.type").String(); got != "validation" {
+				t.Errorf("error.type=%q want validation", got)
+			}
+			if got := gjson.Get(res.stderr, "error.subtype").String(); got != "failed_precondition" {
+				t.Errorf("error.subtype=%q want failed_precondition", got)
+			}
+			if msg := gjson.Get(res.stderr, "error.message").String(); msg != "No API metadata available" {
+				t.Errorf("error.message=%q want %q", msg, "No API metadata available")
+			}
+			if hint := gjson.Get(res.stderr, "error.hint").String(); !strings.Contains(hint, "no embedded API metadata") {
+				t.Errorf("error.hint=%q want to contain %q", hint, "no embedded API metadata")
+			}
+		})
+	}
+}
+
+// TestRuntimeCatalogResolvesSchema pins the PRIMARY #1764 fix: a bare-module fork
+// (embedded stub only) resolves `schema` against the RUNTIME catalog seeded from
+// the on-disk cache, not the embedded-only catalog. Before f0b6f35f the module
+// build read the embedded-only catalog and returned "Unknown service: <svc>" even
+// though the runtime registry had metadata; after it, registry.SchemaCatalog()
+// falls back to the merged runtime catalog and the lookup succeeds.
+//
+// This is the counterpart to TestDegradeStubMetadataSchema: that test pins the
+// cold-cache corner (no runtime data -> graceful "No API metadata available");
+// this one pins the warm-cache main path (runtime data present -> schema works),
+// so a regression that re-embeds the embedded-only lookup fails HERE with
+// "Unknown service" rather than silently passing.
+func TestRuntimeCatalogResolvesSchema(t *testing.T) {
+	bin := buildFork(t, "plain", plainPlugin)
+	res := runWithSeededCatalog(t, bin, seededCatalogJSON, "schema", "plugine2e")
+	t.Logf("exit=%d stdout=%s stderr=%s", res.exit, res.stdout, res.stderr)
+	out := res.stdout + res.stderr
+	if strings.Contains(out, "Unknown service") {
+		t.Fatalf("schema returned \"Unknown service\" -> runtime catalog NOT consulted (issue #1764 regression); out=%s", out)
+	}
+	if strings.Contains(out, "No API metadata available") {
+		t.Fatalf("schema saw no metadata -> the seeded runtime cache was not loaded; out=%s", out)
+	}
+	if res.exit != 0 {
+		t.Fatalf("exit=%d want 0 (schema resolved from runtime catalog); stdout=%s stderr=%s", res.exit, res.stdout, res.stderr)
+	}
+	if !strings.Contains(out, "plugine2e") {
+		t.Errorf("schema output does not mention the seeded service; out=%s", out)
+	}
+}
+
+// credentialBlockPlugin registers a credential.Provider whose ResolveAccount
+// (and ResolveToken) unconditionally return a *credential.BlockError.
+// internal/credential/credential_provider.go's doResolveAccount returns this
+// error straight from the provider loop -- before any defaultAcct fallback
+// and, transitively, before the LarkClient/HttpClient phases that would issue
+// a real network call ever run (see internal/cmdutil/factory_default.go's
+// Phase 2 -> Phase 4 ordering).
+const credentialBlockPlugin = `// Code generated by plugin_e2e; DO NOT EDIT.
+package plugin
+
+import (
+	"context"
+
+	"github.com/larksuite/cli/extension/credential"
+)
+
+type blockProvider struct{}
+
+func (blockProvider) Name() string { return "block-cred" }
+
+func (blockProvider) ResolveAccount(ctx context.Context) (*credential.Account, error) {
+	return nil, &credential.BlockError{Provider: "block-cred", Reason: "blocked for test"}
+}
+
+func (blockProvider) ResolveToken(ctx context.Context, req credential.TokenSpec) (*credential.Token, error) {
+	return nil, &credential.BlockError{Provider: "block-cred", Reason: "blocked for test"}
+}
+
+func init() {
+	credential.Register(blockProvider{})
+}
+`
+
+// TestSubsystemCredentialBlock pins the credential.BlockError offline effect.
+// Observed real output for `docs +fetch --doc nonexistent`, run twice across
+// separate `go test -count` invocations (byte-identical both times, unlike
+// the transport-abort case -- credential resolution happens once, before any
+// endpoint is chosen, so there is no varying destination URL to leak into the
+// message):
+//
+//	exit=5
+//	stdout=(empty)
+//	stderr={"ok":false,"identity":"bot","error":{"type":"internal","subtype":"unknown",
+//	  "message":"blocked by block-cred: blocked for test"}}
+func TestSubsystemCredentialBlock(t *testing.T) {
+	bin := buildFork(t, "credential-block", credentialBlockPlugin)
+	res := run(t, bin, "docs", "+fetch", "--doc", "nonexistent")
+	t.Logf("exit=%d stdout=%s stderr=%s", res.exit, res.stdout, res.stderr)
+	if res.exit != 5 {
+		t.Fatalf("exit=%d want 5; stdout=%s stderr=%s", res.exit, res.stdout, res.stderr)
+	}
+	if !gjson.Valid(res.stderr) {
+		t.Fatalf("stderr not JSON: %s", res.stderr)
+	}
+	if got := gjson.Get(res.stderr, "error.type").String(); got != "internal" {
+		t.Errorf("error.type=%q want internal", got)
+	}
+	if got := gjson.Get(res.stderr, "error.subtype").String(); got != "unknown" {
+		t.Errorf("error.subtype=%q want unknown", got)
+	}
+	if msg := gjson.Get(res.stderr, "error.message").String(); msg != "blocked by block-cred: blocked for test" {
+		t.Errorf("error.message=%q want %q", msg, "blocked by block-cred: blocked for test")
+	}
+}

tests/plugin_e2e/diagnostics_test.go

@@ -0,0 +1,39 @@
+// Copyright (c) 2026 Lark Technologies Pte. Ltd.
+// SPDX-License-Identifier: MIT
+
+package plugin_e2e
+
+import (
+	"testing"
+
+	"github.com/tidwall/gjson"
+)
+
+// TestDiagnostics asserts the VERIFIED stdout shapes of the two policy/plugin
+// diagnostic commands on a fork carrying the readonly Restrict rule:
+//   - `config policy show`: source_name == the plugin name that installed the
+//     active rule.
+//   - `config plugins show`: {"plugins":[{"name","version","capabilities",...,
+//     "hooks":{...}}],"total":N} with the readonly plugin present.
+func TestDiagnostics(t *testing.T) {
+	bin := buildFork(t, "readonly", readonlyPlugin)
+
+	pol := run(t, bin, "config", "policy", "show")
+	if pol.exit != 0 || !gjson.Valid(pol.stdout) {
+		t.Fatalf("policy show exit=%d stdout=%s stderr=%s", pol.exit, pol.stdout, pol.stderr)
+	}
+	if src := gjson.Get(pol.stdout, "source_name").String(); src != "readonly" {
+		t.Errorf("policy source_name=%q want readonly (stdout=%s)", src, pol.stdout)
+	}
+
+	plug := run(t, bin, "config", "plugins", "show")
+	if plug.exit != 0 || !gjson.Valid(plug.stdout) {
+		t.Fatalf("plugins show exit=%d stdout=%s", plug.exit, plug.stdout)
+	}
+	if total := gjson.Get(plug.stdout, "total").Int(); total < 1 {
+		t.Errorf("plugins total=%d want >=1 (stdout=%s)", total, plug.stdout)
+	}
+	if name := gjson.Get(plug.stdout, "plugins.0.name").String(); name != "readonly" {
+		t.Errorf("plugins.0.name=%q want readonly", name)
+	}
+}

tests/plugin_e2e/harness.go

@@ -0,0 +1,210 @@
+// Copyright (c) 2026 Lark Technologies Pte. Ltd.
+// SPDX-License-Identifier: MIT
+
+// Package plugin_e2e exercises the extension/platform plugin contract the way a
+// real customer does: it builds a fork of lark-cli with a plugin blank-imported,
+// then runs that fork as a subprocess and asserts the real stderr/stdout
+// envelopes and exit codes. This is L4 coverage — the in-process unit and
+// integration tests (extension/..., cmd/...) assert Go error values in the test
+// process and structurally cannot observe envelope serialization, exit codes, or
+// the blank-import -> init -> Register -> InstallAll assembly chain.
+//
+// Mechanism (the "customer build", mirrors xcaddy's build mode):
+//  1. `git archive HEAD` a clean tree containing only committed files (so the
+//     fork embeds the tracked meta_data stub, reproducing the bare-module state).
+//  2. Generate a customer module: go.mod (cli's requires + `replace` to the
+//     archived tree) + go.sum copy + main.go (blank-imports the plugin package)
+//     + plugin package (its init() calls platform.Register).
+//  3. `go build` the fork (offline-capable via the warm module cache), then run
+//     it as a subprocess and assert.
+package plugin_e2e
+
+import (
+	"context"
+	"errors"
+	"fmt"
+	"os"
+	"os/exec"
+	"path/filepath"
+	"strings"
+	"testing"
+	"time"
+)
+
+// cleanTree is the git-archived, committed-only source tree of the repo under
+// test, shared by every fork build. Populated by TestMain (smoke_test.go) —
+// TestMain must live in a _test.go file to be recognized by `go test`, so the
+// entry point sits there while the rest of the harness mechanism lives here.
+var cleanTree string
+
+// baseDir holds the archive tree plus every generated customer module.
+var baseDir string
+
+// repoRoot resolves the lark-cli module root from the test's working directory
+// (which `go test` sets to the package dir, tests/plugin_e2e).
+func repoRoot() (string, error) {
+	out, err := exec.Command("git", "rev-parse", "--show-toplevel").Output()
+	if err != nil {
+		return "", err
+	}
+	return strings.TrimSpace(string(out)), nil
+}
+
+// gitArchive extracts HEAD's committed tree into dst by streaming `git archive`
+// into `tar -x`. Only tracked files are included — gitignored build artifacts
+// (e.g. the fetched meta_data.json) are absent, exactly as a module consumer
+// would see them. It wires the two processes with an explicit pipe rather than a
+// shell, so dst never reaches a shell command line.
+func gitArchive(root, dst string) error {
+	archive := exec.Command("git", "archive", "HEAD")
+	archive.Dir = root
+	extract := exec.Command("tar", "-x", "-C", dst)
+	pipe, err := archive.StdoutPipe()
+	if err != nil {
+		return err
+	}
+	extract.Stdin = pipe
+	// Each process gets its own stderr buffer: os/exec spawns a copy goroutine
+	// per command, so a shared strings.Builder would be written concurrently by
+	// both (git archive and tar run in parallel) -- a data race, since
+	// strings.Builder is not concurrency-safe.
+	var archiveErr, extractErr strings.Builder
+	archive.Stderr = &archiveErr
+	extract.Stderr = &extractErr
+	if err := extract.Start(); err != nil {
+		return err
+	}
+	if err := archive.Run(); err != nil {
+		_ = extract.Wait()
+		return fmt.Errorf("git archive: %w: %s", err, archiveErr.String())
+	}
+	if err := extract.Wait(); err != nil {
+		return fmt.Errorf("tar extract: %w: %s", err, extractErr.String())
+	}
+	return nil
+}
+
+// builtForks caches fork binaries by name so identical forks are built once.
+var builtForks = map[string]string{}
+
+// buildFork generates a customer module whose plugin package body is pluginSrc,
+// builds the fork, and returns the binary path. Forks are cached by name.
+func buildFork(t *testing.T, name, pluginSrc string) string {
+	t.Helper()
+	if bin, ok := builtForks[name]; ok {
+		return bin
+	}
+	mod := filepath.Join(baseDir, "fork-"+name)
+	if err := os.MkdirAll(filepath.Join(mod, "plugin"), 0o755); err != nil {
+		t.Fatalf("mkdir customer module: %v", err)
+	}
+
+	// go.mod: reuse cli's require graph, rename the module, replace cli with the
+	// local archived tree. This avoids `go mod tidy` (no network at test time).
+	rawMod, err := os.ReadFile(filepath.Join(cleanTree, "go.mod"))
+	if err != nil {
+		t.Fatalf("read archived go.mod: %v", err)
+	}
+	gomod := strings.Replace(string(rawMod), "module github.com/larksuite/cli", "module larkcustomer", 1)
+	gomod += "\nrequire github.com/larksuite/cli v0.0.0\n\nreplace github.com/larksuite/cli => " + cleanTree + "\n"
+	writeFile(t, filepath.Join(mod, "go.mod"), gomod)
+
+	// go.sum: transitive dependency hashes are identical to cli's.
+	rawSum, err := os.ReadFile(filepath.Join(cleanTree, "go.sum"))
+	if err != nil {
+		t.Fatalf("read archived go.sum: %v", err)
+	}
+	writeFile(t, filepath.Join(mod, "go.sum"), string(rawSum))
+
+	writeFile(t, filepath.Join(mod, "main.go"), customerMain)
+	writeFile(t, filepath.Join(mod, "plugin", "plugin.go"), pluginSrc)
+
+	bin := filepath.Join(mod, "fork-bin")
+	build := exec.Command("go", "build", "-o", bin, ".")
+	build.Dir = mod
+	// -mod=mod fixes require annotations copied from cli's go.mod; the default
+	// GOPROXY resolves any dep missing from the cache (goproxy in CI/dev).
+	build.Env = append(os.Environ(), "GOFLAGS=-mod=mod")
+	if out, err := build.CombinedOutput(); err != nil {
+		t.Fatalf("build fork %q failed: %v\n%s", name, err, out)
+	}
+	builtForks[name] = bin
+	return bin
+}
+
+const customerMain = `// Code generated by plugin_e2e; DO NOT EDIT.
+package main
+
+import (
+	"os"
+
+	"github.com/larksuite/cli/cmd"
+	_ "larkcustomer/plugin" // blank import triggers plugin init() -> platform.Register
+)
+
+func main() { os.Exit(cmd.Execute()) }
+`
+
+// result is a subprocess run outcome.
+type result struct {
+	stdout string
+	stderr string
+	exit   int
+}
+
+// run executes the fork binary with args in an isolated, offline environment and
+// captures stdout/stderr/exit. Each call gets a fresh empty
+// LARKSUITE_CLI_CONFIG_DIR and LARKSUITE_CLI_REMOTE_META=off, so the fork never
+// inherits the host's ~/.lark-cli cache or makes a startup metadata fetch to the
+// open platform. That reproduces the bare-module customer state (no embedded
+// metadata, cold cache) deterministically on any machine, including CI: without
+// it, whether a command's assertion is reached depends on whether a live network
+// fetch happened to succeed. Tests that need runtime metadata seed it explicitly
+// via runWithSeededCatalog.
+func run(t *testing.T, bin string, args ...string) result {
+	t.Helper()
+	return runWithEnv(t, bin, isolatedEnv(t), args...)
+}
+
+// isolatedEnv is the bare-module, offline environment shared by run() and (as a
+// base) by runWithSeededCatalog.
+func isolatedEnv(t *testing.T) []string {
+	t.Helper()
+	return append(os.Environ(),
+		"LARKSUITE_CLI_NO_UPDATE_NOTIFIER=1",
+		"LARKSUITE_CLI_NO_SKILLS_NOTIFIER=1",
+		"LARKSUITE_CLI_CONFIG_DIR="+t.TempDir(),
+		"LARKSUITE_CLI_REMOTE_META=off",
+	)
+}
+
+// runWithEnv runs bin as a subprocess with the given full environment, capturing
+// stdout/stderr/exit.
+func runWithEnv(t *testing.T, bin string, env []string, args ...string) result {
+	t.Helper()
+	ctx, cancel := context.WithTimeout(context.Background(), 60*time.Second)
+	defer cancel()
+	c := exec.CommandContext(ctx, bin, args...)
+	c.Env = env
+	var stdout, stderr strings.Builder
+	c.Stdout = &stdout
+	c.Stderr = &stderr
+	err := c.Run()
+	exit := 0
+	if err != nil {
+		var ee *exec.ExitError
+		if errors.As(err, &ee) {
+			exit = ee.ExitCode()
+		} else {
+			t.Fatalf("run %v: %v", args, err)
+		}
+	}
+	return result{stdout: stdout.String(), stderr: stderr.String(), exit: exit}
+}
+
+func writeFile(t *testing.T, path, content string) {
+	t.Helper()
+	if err := os.WriteFile(path, []byte(content), 0o644); err != nil {
+		t.Fatalf("write %s: %v", path, err)
+	}
+}

tests/plugin_e2e/install_test.go

@@ -0,0 +1,339 @@
+// Copyright (c) 2026 Lark Technologies Pte. Ltd.
+// SPDX-License-Identifier: MIT
+
+package plugin_e2e
+
+import (
+	"strings"
+	"testing"
+
+	"github.com/tidwall/gjson"
+)
+
+// multipleRestrictPlugin registers TWO distinct plugins that each call
+// Restrict() with an independently valid Rule. cmdpolicy.Resolve rejects
+// more than one distinct Restrict-owner regardless of each rule's own
+// validity (internal/cmdpolicy/resolver.go's distinctOwners check runs
+// before ValidateRule).
+const multipleRestrictPlugin = `// Code generated by plugin_e2e; DO NOT EDIT.
+package plugin
+
+import "github.com/larksuite/cli/extension/platform"
+
+func init() {
+	platform.Register(
+		platform.NewPlugin("restrict-a", "0.1.0").
+			Restrict(&platform.Rule{
+				Name:    "a-rule",
+				Allow:   []string{"docs/**"},
+				MaxRisk: platform.RiskRead,
+			}).
+			MustBuild())
+	platform.Register(
+		platform.NewPlugin("restrict-b", "0.1.0").
+			Restrict(&platform.Rule{
+				Name:    "b-rule",
+				Allow:   []string{"im/**"},
+				MaxRisk: platform.RiskRead,
+			}).
+			MustBuild())
+}
+`
+
+// TestInstallMultipleRestrictPluginsPin pins reason_code=multiple_restrict_plugins.
+// Observed real output (any command, e.g. "schema" -- the fatal guard walks
+// every RunE in the tree so it fires regardless of which command runs):
+//
+//	exit=2
+//	stderr={"ok":false,"error":{"type":"validation","subtype":"failed_precondition",
+//	  "message":"multiple plugins called Restrict; only one plugin may own the
+//	  policy: [restrict-a restrict-b]",
+//	  "hint":"plugin policy configuration is broken (reason_code
+//	  multiple_restrict_plugins); fix the plugin's Restrict rule or remove the
+//	  conflicting plugin"}}
+func TestInstallMultipleRestrictPluginsPin(t *testing.T) {
+	bin := buildFork(t, "multiple-restrict", multipleRestrictPlugin)
+	res := run(t, bin, "schema")
+	t.Logf("exit=%d stdout=%s stderr=%s", res.exit, res.stdout, res.stderr)
+	assertReasonCodeEnvelope(t, res, "multiple_restrict_plugins")
+}
+
+// invalidRulePlugin registers a single plugin whose Restrict Rule carries a
+// syntactically-invalid MaxRisk value. Neither the Builder nor the staging
+// Registrar validate Rule *contents* (only nilness) -- semantic validation
+// happens later, in cmdpolicy.ValidateRule, called from
+// cmd/platform_bootstrap.go's applyUserPolicyPruning -> cmdpolicy.Resolve.
+const invalidRulePlugin = `// Code generated by plugin_e2e; DO NOT EDIT.
+package plugin
+
+import "github.com/larksuite/cli/extension/platform"
+
+func init() {
+	platform.Register(
+		platform.NewPlugin("invalid-rule", "0.1.0").
+			Restrict(&platform.Rule{
+				Name:    "bad-risk",
+				Allow:   []string{"docs/**"},
+				MaxRisk: platform.Risk("bogus"),
+			}).
+			MustBuild())
+}
+`
+
+// TestInstallInvalidRulePin pins reason_code=invalid_rule. Observed real output
+// (schema):
+//
+//	exit=2
+//	stderr={"ok":false,"error":{"type":"validation","subtype":"failed_precondition",
+//	  "message":"plugin \"invalid-rule\" rule invalid: invalid max_risk \"bogus\":
+//	  must be one of read|write|high-risk-write",
+//	  "hint":"plugin policy configuration is broken (reason_code invalid_rule);
+//	  fix the plugin's Restrict rule or remove the conflicting plugin"}}
+func TestInstallInvalidRulePin(t *testing.T) {
+	bin := buildFork(t, "invalid-rule", invalidRulePlugin)
+	res := run(t, bin, "schema")
+	t.Logf("exit=%d stdout=%s stderr=%s", res.exit, res.stdout, res.stderr)
+	assertReasonCodeEnvelope(t, res, "invalid_rule")
+}
+
+// installFailedPlugin is a hand-written bare platform.Plugin (not
+// Builder-based -- Install returning a plain error is not expressible
+// through the Builder's fluent API) whose Install always returns an error.
+// FailurePolicy=FailClosed makes the host abort rather than warn+skip.
+const installFailedPlugin = `// Code generated by plugin_e2e; DO NOT EDIT.
+package plugin
+
+import (
+	"errors"
+
+	"github.com/larksuite/cli/extension/platform"
+)
+
+type installFailed struct{}
+
+func (installFailed) Name() string    { return "install-failed" }
+func (installFailed) Version() string { return "0.1.0" }
+func (installFailed) Capabilities() platform.Capabilities {
+	return platform.Capabilities{FailurePolicy: platform.FailClosed}
+}
+func (installFailed) Install(r platform.Registrar) error {
+	return errors.New("deliberate install failure")
+}
+
+func init() { platform.Register(installFailed{}) }
+`
+
+// TestInstallFailedPin pins reason_code=install_failed. Observed real output
+// (schema):
+//
+//	exit=2
+//	stderr={"ok":false,"error":{"type":"validation","subtype":"failed_precondition",
+//	  "message":"plugin \"install-failed\" (install_failed): Install returned
+//	  error: deliberate install failure",
+//	  "hint":"plugin \"install-failed\" failed to install (reason_code
+//	  install_failed); fix or remove the plugin before running commands"}}
+func TestInstallFailedPin(t *testing.T) {
+	bin := buildFork(t, "install-failed", installFailedPlugin)
+	res := run(t, bin, "schema")
+	t.Logf("exit=%d stdout=%s stderr=%s", res.exit, res.stdout, res.stderr)
+	assertReasonCodeEnvelope(t, res, "install_failed")
+}
+
+// installPanicPlugin is a hand-written bare Plugin whose Install panics.
+// safeCallInstall (internal/platform/host.go) recovers and converts the
+// panic into a typed install_panic error rather than crashing the binary.
+const installPanicPlugin = `// Code generated by plugin_e2e; DO NOT EDIT.
+package plugin
+
+import "github.com/larksuite/cli/extension/platform"
+
+type installPanic struct{}
+
+func (installPanic) Name() string    { return "install-panic" }
+func (installPanic) Version() string { return "0.1.0" }
+func (installPanic) Capabilities() platform.Capabilities {
+	return platform.Capabilities{FailurePolicy: platform.FailClosed}
+}
+func (installPanic) Install(r platform.Registrar) error {
+	panic("deliberate install panic")
+}
+
+func init() { platform.Register(installPanic{}) }
+`
+
+// TestInstallPanicPin pins reason_code=install_panic. Observed real output
+// (schema):
+//
+//	exit=2
+//	stderr={"ok":false,"error":{"type":"validation","subtype":"failed_precondition",
+//	  "message":"plugin \"install-panic\" (install_panic): Install panicked:
+//	  deliberate install panic",
+//	  "hint":"plugin \"install-panic\" failed to install (reason_code
+//	  install_panic); fix or remove the plugin before running commands"}}
+func TestInstallPanicPin(t *testing.T) {
+	bin := buildFork(t, "install-panic", installPanicPlugin)
+	res := run(t, bin, "schema")
+	t.Logf("exit=%d stdout=%s stderr=%s", res.exit, res.stdout, res.stderr)
+	assertReasonCodeEnvelope(t, res, "install_panic")
+}
+
+// pluginNamePanicPlugin is a hand-written bare Plugin whose Name() panics.
+// InstallAll's outer loop calls safeCallName BEFORE it ever reads
+// Capabilities(), so this aborts unconditionally regardless of what
+// Capabilities() would have declared (host.go's isUntrustedConfigError
+// list) -- Capabilities() here is a throwaway zero value, never invoked.
+const pluginNamePanicPlugin = `// Code generated by plugin_e2e; DO NOT EDIT.
+package plugin
+
+import "github.com/larksuite/cli/extension/platform"
+
+type pluginNamePanic struct{}
+
+func (pluginNamePanic) Name() string    { panic("deliberate name panic") }
+func (pluginNamePanic) Version() string { return "0.1.0" }
+func (pluginNamePanic) Capabilities() platform.Capabilities {
+	return platform.Capabilities{}
+}
+func (pluginNamePanic) Install(r platform.Registrar) error { return nil }
+
+func init() { platform.Register(pluginNamePanic{}) }
+`
+
+// TestInstallPluginNamePanicPin pins reason_code=plugin_name_panic. Observed real
+// output (schema):
+//
+//	exit=2
+//	stderr={"ok":false,"error":{"type":"validation","subtype":"failed_precondition",
+//	  "message":"plugin \"<unknown>\" (plugin_name_panic): Plugin.Name()
+//	  panicked: deliberate name panic",
+//	  "hint":"plugin \"<unknown>\" failed to install (reason_code
+//	  plugin_name_panic); fix or remove the plugin before running commands"}}
+func TestInstallPluginNamePanicPin(t *testing.T) {
+	bin := buildFork(t, "plugin-name-panic", pluginNamePanicPlugin)
+	res := run(t, bin, "schema")
+	t.Logf("exit=%d stdout=%s stderr=%s", res.exit, res.stdout, res.stderr)
+	assertReasonCodeEnvelope(t, res, "plugin_name_panic")
+}
+
+// capabilitiesPanicPlugin is a hand-written bare Plugin whose Capabilities()
+// panics. readFailurePolicy (internal/platform/host.go) re-invokes
+// Capabilities() to decide FailOpen vs FailClosed, panics again, and its
+// recover leaves the pre-set FailClosed default in place -- so this aborts
+// unconditionally too, without the plugin ever declaring a real policy.
+const capabilitiesPanicPlugin = `// Code generated by plugin_e2e; DO NOT EDIT.
+package plugin
+
+import "github.com/larksuite/cli/extension/platform"
+
+type capabilitiesPanic struct{}
+
+func (capabilitiesPanic) Name() string    { return "capabilities-panic" }
+func (capabilitiesPanic) Version() string { return "0.1.0" }
+func (capabilitiesPanic) Capabilities() platform.Capabilities {
+	panic("deliberate capabilities panic")
+}
+func (capabilitiesPanic) Install(r platform.Registrar) error { return nil }
+
+func init() { platform.Register(capabilitiesPanic{}) }
+`
+
+// TestInstallCapabilitiesPanicPin pins reason_code=capabilities_panic. Observed
+// real output (schema):
+//
+//	exit=2
+//	stderr={"ok":false,"error":{"type":"validation","subtype":"failed_precondition",
+//	  "message":"plugin \"capabilities-panic\" (capabilities_panic):
+//	  Plugin.Capabilities() panicked: deliberate capabilities panic",
+//	  "hint":"plugin \"capabilities-panic\" failed to install (reason_code
+//	  capabilities_panic); fix or remove the plugin before running commands"}}
+func TestInstallCapabilitiesPanicPin(t *testing.T) {
+	bin := buildFork(t, "capabilities-panic", capabilitiesPanicPlugin)
+	res := run(t, bin, "schema")
+	t.Logf("exit=%d stdout=%s stderr=%s", res.exit, res.stdout, res.stderr)
+	assertReasonCodeEnvelope(t, res, "capabilities_panic")
+}
+
+// restrictsMismatchPlugin is a hand-written bare Plugin that declares
+// Capabilities.Restricts=true (paired with the required FailClosed) but
+// whose Install never calls r.Restrict. stagingRegistrar.validateSelf
+// (internal/platform/staging.go) checks this exact declared-vs-actual
+// consistency after Install returns.
+const restrictsMismatchPlugin = `// Code generated by plugin_e2e; DO NOT EDIT.
+package plugin
+
+import "github.com/larksuite/cli/extension/platform"
+
+type restrictsMismatch struct{}
+
+func (restrictsMismatch) Name() string    { return "restricts-mismatch" }
+func (restrictsMismatch) Version() string { return "0.1.0" }
+func (restrictsMismatch) Capabilities() platform.Capabilities {
+	return platform.Capabilities{Restricts: true, FailurePolicy: platform.FailClosed}
+}
+func (restrictsMismatch) Install(r platform.Registrar) error { return nil }
+
+func init() { platform.Register(restrictsMismatch{}) }
+`
+
+// TestInstallRestrictsMismatchPin pins reason_code=restricts_mismatch.
+// Observed real output (schema):
+//
+//	exit=2
+//	stderr={"ok":false,"error":{"type":"validation","subtype":"failed_precondition",
+//	  "message":"plugin \"restricts-mismatch\" (restricts_mismatch):
+//	  Capabilities.Restricts=true but Install did not call r.Restrict",
+//	  "hint":"plugin \"restricts-mismatch\" failed to install (reason_code
+//	  restricts_mismatch); fix or remove the plugin before running commands"}}
+func TestInstallRestrictsMismatchPin(t *testing.T) {
+	bin := buildFork(t, "restricts-mismatch", restrictsMismatchPlugin)
+	res := run(t, bin, "schema")
+	t.Logf("exit=%d stdout=%s stderr=%s", res.exit, res.stdout, res.stderr)
+	assertReasonCodeEnvelope(t, res, "restricts_mismatch")
+}
+
+// mustBuildPanicPlugin calls MustBuild() on a Builder with an invalid plugin
+// name ("BadName!!" fails ^[a-z0-9][a-z0-9-]*$). This panics from
+// plugin.init(), which runs from the blank-import BEFORE main() has a
+// chance to install any recover-and-envelope guard -- so, unlike every
+// other case here, this crashes the process outright: no JSON envelope,
+// non-zero exit, a raw Go panic trace on stderr.
+const mustBuildPanicPlugin = `// Code generated by plugin_e2e; DO NOT EDIT.
+package plugin
+
+import "github.com/larksuite/cli/extension/platform"
+
+func init() {
+	platform.Register(platform.NewPlugin("BadName!!", "0.1.0").MustBuild())
+}
+`
+
+// TestInstallMustBuildInitPanicCrashesBinary pins the MustBuild init-panic crash
+// shape. This is NOT the plugin_install envelope -- it is a bare Go panic
+// crash, because it happens in init(), before main()'s recover guard
+// exists. Observed real output (schema):
+//
+//	exit=2
+//	stderr=panic: plugin "BadName!!": invalid plugin name "BadName!!": must
+//	  match ^[a-z0-9][a-z0-9-]*$
+//
+//	goroutine 1 [running]:
+//	larkcustomer/plugin.init.0(...)
+//	        .../plugin/plugin.go:7
+//	...
+func TestInstallMustBuildInitPanicCrashesBinary(t *testing.T) {
+	bin := buildFork(t, "mustbuild-panic", mustBuildPanicPlugin)
+	res := run(t, bin, "schema")
+	t.Logf("exit=%d stdout=%s stderr=%s", res.exit, res.stdout, res.stderr)
+	if res.exit == 0 {
+		t.Fatalf("expected non-zero exit on init panic; exit=%d stdout=%s stderr=%s", res.exit, res.stdout, res.stderr)
+	}
+	if gjson.Valid(res.stderr) {
+		t.Fatalf("expected a raw panic trace, not a JSON envelope; stderr=%s", res.stderr)
+	}
+	if !strings.Contains(res.stderr, "panic:") {
+		t.Fatalf("stderr missing Go panic trace; stderr=%s", res.stderr)
+	}
+	if !strings.Contains(res.stderr, `invalid plugin name "BadName!!"`) {
+		t.Errorf("stderr missing the Builder's invalid-name message; stderr=%s", res.stderr)
+	}
+}

tests/plugin_e2e/observe_wrap_test.go

@@ -0,0 +1,298 @@
+// Copyright (c) 2026 Lark Technologies Pte. Ltd.
+// SPDX-License-Identifier: MIT
+
+package plugin_e2e
+
+import (
+	"strings"
+	"testing"
+
+	"github.com/tidwall/gjson"
+)
+
+// auditPlugin registers a single After observer matching every command that
+// logs "[audit] <path>" to stderr. Based on (a simplified form of) the
+// shipped extension/platform/examples/audit-observer example.
+const auditPlugin = `// Code generated by plugin_e2e; DO NOT EDIT.
+package plugin
+
+import (
+	"context"
+	"fmt"
+	"os"
+
+	"github.com/larksuite/cli/extension/platform"
+)
+
+func init() {
+	platform.Register(
+		platform.NewPlugin("audit", "0.1.0").
+			Observer(platform.After, "log", platform.All(),
+				func(_ context.Context, inv platform.Invocation) {
+					fmt.Fprintf(os.Stderr, "[audit] %s\n", inv.Cmd().Path())
+				}).
+			FailOpen().
+			MustBuild())
+}
+`
+
+// TestObservePin pins the audit observer's stderr line format. Observed
+// real output (docs +fetch --doc nonexistent, a real read-risk command that
+// fails downstream with an API error unrelated to the plugin):
+//
+//	exit=1
+//	stderr=[audit] docs/+fetch
+//	{"ok":false,"identity":"user","error":{"type":"api","subtype":"unknown",...}}
+//
+// The observer line always leads, on its own line, before whatever the
+// command itself writes to stderr.
+func TestObservePin(t *testing.T) {
+	bin := buildFork(t, "audit", auditPlugin)
+	res := run(t, bin, "docs", "+fetch", "--doc", "nonexistent")
+	if !strings.Contains(res.stderr, "[audit] docs/+fetch\n") {
+		t.Fatalf("stderr missing audit line; stderr=%s", res.stderr)
+	}
+}
+
+// auditRestrictPlugin combines an After observer with a Restrict rule in one
+// plugin, so a denied command's stderr carries both the observer's
+// side-effect and the denial envelope: the framework's contract is that
+// After observers fire even for denied commands (see
+// extension/platform/invocation.go's DeniedByPolicy doc).
+const auditRestrictPlugin = `// Code generated by plugin_e2e; DO NOT EDIT.
+package plugin
+
+import (
+	"context"
+	"fmt"
+	"os"
+
+	"github.com/larksuite/cli/extension/platform"
+)
+
+func init() {
+	platform.Register(
+		platform.NewPlugin("audit-restrict", "0.1.0").
+			Observer(platform.After, "log", platform.All(),
+				func(_ context.Context, inv platform.Invocation) {
+					fmt.Fprintf(os.Stderr, "[audit] %s\n", inv.Cmd().Path())
+				}).
+			Restrict(&platform.Rule{
+				Name:    "agent-readonly",
+				Allow:   []string{"docs/**", "im/**"},
+				MaxRisk: platform.RiskRead,
+			}).
+			MustBuild())
+}
+`
+
+// TestObserveOnDeniedPin pins the audit-contract case: a denied command's
+// stderr carries BOTH the observer's audit line AND the denial envelope,
+// concatenated in a single stream, audit line first. Observed real output
+// (docs +update --doc-token x --content y, denied write_not_allowed):
+//
+//	exit=2
+//	stderr=[audit] docs/+update
+//	{"ok":false,"error":{"type":"validation","subtype":"failed_precondition",...}}
+//
+// The leading "[audit] ..." line means gjson.Valid on the raw stderr is
+// false; the JSON envelope must be sliced out from the first '{' before
+// parsing it as JSON.
+func TestObserveOnDeniedPin(t *testing.T) {
+	bin := buildFork(t, "audit-restrict", auditRestrictPlugin)
+	res := run(t, bin, "docs", "+update", "--doc-token", "x", "--content", "y")
+	if res.exit != 2 {
+		t.Fatalf("exit=%d stdout=%s stderr=%s", res.exit, res.stdout, res.stderr)
+	}
+	if !strings.Contains(res.stderr, "[audit] docs/+update\n") {
+		t.Fatalf("stderr missing audit line on a denied command; stderr=%s", res.stderr)
+	}
+	i := strings.Index(res.stderr, "{")
+	if i < 0 {
+		t.Fatalf("stderr has no JSON envelope after the audit line; stderr=%s", res.stderr)
+	}
+	envelope := res.stderr[i:]
+	if !gjson.Valid(envelope) {
+		t.Fatalf("sliced envelope not JSON: %s", envelope)
+	}
+	if got := gjson.Get(envelope, "error.type").String(); got != "validation" {
+		t.Errorf("error.type=%q want validation", got)
+	}
+	if got := gjson.Get(envelope, "error.subtype").String(); got != "failed_precondition" {
+		t.Errorf("error.subtype=%q want failed_precondition", got)
+	}
+	if hint := gjson.Get(envelope, "error.hint").String(); !strings.Contains(hint, "reason_code write_not_allowed") {
+		t.Errorf("hint=%q want to contain reason_code write_not_allowed", hint)
+	}
+}
+
+// observerPanicPlugin's After observer panics unconditionally. runObserverSafe
+// (internal/hook/install.go) must isolate the panic so command dispatch
+// still completes normally.
+const observerPanicPlugin = `// Code generated by plugin_e2e; DO NOT EDIT.
+package plugin
+
+import (
+	"context"
+
+	"github.com/larksuite/cli/extension/platform"
+)
+
+func init() {
+	platform.Register(
+		platform.NewPlugin("observer-panic", "0.1.0").
+			Observer(platform.After, "log", platform.All(),
+				func(_ context.Context, _ platform.Invocation) {
+					panic("boom")
+				}).
+			FailOpen().
+			MustBuild())
+}
+`
+
+// TestObserverPanicIsolationPin pins panic isolation: an After observer that
+// always panics must not affect the command's own outcome. The assertion is
+// baseline-relative -- the panicking-observer fork's exit code must equal the
+// noop-observer baseline fork's for the same `schema` command (a local,
+// network-free, read-risk command), whatever that shared exit code is.
+// Observed real output at pin time:
+//
+//	panicking: exit=0 stderr=warning: hook "observer-panic.log" panicked: boom
+//	baseline:  exit=0 stderr=(empty)
+//
+// The panic is fully swallowed by runObserverSafe (internal/hook/install.go),
+// surfacing only as a stderr warning line, never as a non-zero exit or crash.
+func TestObserverPanicIsolationPin(t *testing.T) {
+	bin := buildFork(t, "observer-panic", observerPanicPlugin)
+	res := run(t, bin, "schema")
+
+	baselineBin := buildFork(t, "smoke", noopPlugin)
+	baseline := run(t, baselineBin, "schema")
+
+	if res.exit != baseline.exit {
+		t.Fatalf("panicking-observer exit=%d differs from baseline exit=%d; stderr=%s", res.exit, baseline.exit, res.stderr)
+	}
+	if !strings.Contains(res.stderr, `warning: hook "observer-panic.log" panicked: boom`) {
+		t.Errorf("stderr missing panic-isolation warning; stderr=%s", res.stderr)
+	}
+}
+
+// wrapAbortPlugin's Wrapper short-circuits every command with an AbortError
+// instead of calling next.
+const wrapAbortPlugin = `// Code generated by plugin_e2e; DO NOT EDIT.
+package plugin
+
+import (
+	"context"
+
+	"github.com/larksuite/cli/extension/platform"
+)
+
+func init() {
+	platform.Register(
+		platform.NewPlugin("wrap-abort", "0.1.0").
+			Wrap("guard", platform.All(), func(next platform.Handler) platform.Handler {
+				return func(ctx context.Context, inv platform.Invocation) error {
+					return &platform.AbortError{
+						HookName: "guard",
+						Reason:   "blocked for test",
+					}
+				}
+			}).
+			FailOpen().
+			MustBuild())
+}
+`
+
+// TestWrapAbortPin pins the wrap-abort envelope shape. An *AbortError
+// returned by a Wrapper is converted by wrapAbortError
+// (internal/hook/install.go) into the SAME envelope shape as a Restrict
+// denial -- error.type=="validation", error.subtype=="failed_precondition"
+// -- NOT a distinct "hook" error type. Observed real output (docs +fetch
+// --doc nonexistent, wrapper aborts unconditionally before calling next):
+//
+//	exit=2
+//	stderr={"ok":false,"error":{"type":"validation","subtype":"failed_precondition",
+//	  "message":"hook \"wrap-abort.guard\" aborted: blocked for test",
+//	  "hint":"plugin hook \"wrap-abort.guard\" aborted this command; adjust the
+//	  request to satisfy the hook's policy, or remove the plugin"}}
+//
+// HookName is namespaced to "<plugin-name>.<hookName>" ("wrap-abort.guard")
+// regardless of the HookName the plugin set on the AbortError itself
+// (namespacedWrap overwrites it) -- see internal/hook/install.go.
+func TestWrapAbortPin(t *testing.T) {
+	bin := buildFork(t, "wrap-abort", wrapAbortPlugin)
+	res := run(t, bin, "docs", "+fetch", "--doc", "nonexistent")
+	if res.exit != 2 {
+		t.Fatalf("exit=%d stdout=%s stderr=%s", res.exit, res.stdout, res.stderr)
+	}
+	if !gjson.Valid(res.stderr) {
+		t.Fatalf("stderr not JSON: %s", res.stderr)
+	}
+	if got := gjson.Get(res.stderr, "error.type").String(); got != "validation" {
+		t.Errorf("error.type=%q want validation", got)
+	}
+	if got := gjson.Get(res.stderr, "error.subtype").String(); got != "failed_precondition" {
+		t.Errorf("error.subtype=%q want failed_precondition", got)
+	}
+	if msg := gjson.Get(res.stderr, "error.message").String(); !strings.Contains(msg, `hook "wrap-abort.guard" aborted: blocked for test`) {
+		t.Errorf("error.message=%q want to contain the namespaced hook name and Reason", msg)
+	}
+	if hint := gjson.Get(res.stderr, "error.hint").String(); !strings.Contains(hint, `plugin hook "wrap-abort.guard" aborted this command`) {
+		t.Errorf("error.hint=%q want to contain the abort hint", hint)
+	}
+}
+
+// wrapPanicPlugin's Wrapper factory panics on every invocation (the factory
+// closure itself, not the returned Handler).
+const wrapPanicPlugin = `// Code generated by plugin_e2e; DO NOT EDIT.
+package plugin
+
+import (
+	"github.com/larksuite/cli/extension/platform"
+)
+
+func init() {
+	platform.Register(
+		platform.NewPlugin("wrap-panic", "0.1.0").
+			Wrap("guard", platform.All(), func(next platform.Handler) platform.Handler {
+				panic("wrap boom")
+			}).
+			FailOpen().
+			MustBuild())
+}
+`
+
+// TestWrapPanicPin pins the wrap-panic envelope shape: a panicking Wrapper
+// factory does not crash the process. recoverWrap (internal/hook/install.go)
+// converts the panic into the same validation/failed_precondition shape as
+// wrap-abort, with a distinct message/hint pair. Observed real output (docs
+// +fetch --doc nonexistent, wrapper factory panics unconditionally):
+//
+//	exit=2
+//	stderr={"ok":false,"error":{"type":"validation","subtype":"failed_precondition",
+//	  "message":"hook \"wrap-panic.guard\" panicked: wrap boom",
+//	  "hint":"plugin hook \"wrap-panic.guard\" crashed while handling this
+//	  command; report the panic to the plugin author or remove the plugin"}}
+func TestWrapPanicPin(t *testing.T) {
+	bin := buildFork(t, "wrap-panic", wrapPanicPlugin)
+	res := run(t, bin, "docs", "+fetch", "--doc", "nonexistent")
+	if res.exit != 2 {
+		t.Fatalf("exit=%d stdout=%s stderr=%s", res.exit, res.stdout, res.stderr)
+	}
+	if !gjson.Valid(res.stderr) {
+		t.Fatalf("stderr not JSON: %s", res.stderr)
+	}
+	if got := gjson.Get(res.stderr, "error.type").String(); got != "validation" {
+		t.Errorf("error.type=%q want validation", got)
+	}
+	if got := gjson.Get(res.stderr, "error.subtype").String(); got != "failed_precondition" {
+		t.Errorf("error.subtype=%q want failed_precondition", got)
+	}
+	if msg := gjson.Get(res.stderr, "error.message").String(); !strings.Contains(msg, `hook "wrap-panic.guard" panicked: wrap boom`) {
+		t.Errorf("error.message=%q want to contain the namespaced hook name and panic value", msg)
+	}
+	if hint := gjson.Get(res.stderr, "error.hint").String(); !strings.Contains(hint, `plugin hook "wrap-panic.guard" crashed while handling this command`) {
+		t.Errorf("error.hint=%q want to contain the panic hint", hint)
+	}
+}

tests/plugin_e2e/restrict_test.go

@@ -0,0 +1,205 @@
+// Copyright (c) 2026 Lark Technologies Pte. Ltd.
+// SPDX-License-Identifier: MIT
+
+package plugin_e2e
+
+import (
+	"strings"
+	"testing"
+
+	"github.com/tidwall/gjson"
+)
+
+// readonlyPlugin registers a Restrict rule that only allows read-risk
+// commands under the docs/** and im/** domains. It mirrors the official
+// example readonly-policy configuration.
+const readonlyPlugin = `// Code generated by plugin_e2e; DO NOT EDIT.
+package plugin
+
+import "github.com/larksuite/cli/extension/platform"
+
+func init() {
+	platform.Register(
+		platform.NewPlugin("readonly", "0.1.0").
+			Restrict(&platform.Rule{
+				Name:    "agent-readonly",
+				Allow:   []string{"docs/**", "im/**"},
+				MaxRisk: platform.RiskRead,
+			}).
+			MustBuild())
+}
+`
+
+// TestReadonlyDenial asserts the VERIFIED denial envelope shape: stderr is
+// valid JSON, error.type=="validation", error.subtype=="failed_precondition",
+// error.hint contains the literal "reason_code <X>" substring, and the
+// process exits 2. reason_code lives only in the hint string, not a
+// structured field.
+func TestReadonlyDenial(t *testing.T) {
+	bin := buildFork(t, "readonly", readonlyPlugin)
+	// Note: reason_code mixed_children_policy is intentionally NOT covered here.
+	// It requires a parent command whose *enumerated children* have mixed
+	// allow/deny outcomes, which needs the full command tree from API metadata.
+	// This L4 harness builds a bare-module fork (embedded stub only), so offline
+	// a parent like "sheets" has no known children and collapses to
+	// domain_not_allowed -- identical to the "leaf out of allow list" case and
+	// not a distinct reason_code. Covered instead by the in-process cmdpolicy
+	// unit tests, which construct a mixed-children tree directly.
+	cases := []struct {
+		name       string
+		args       []string
+		reasonCode string
+	}{
+		{"write in allowed domain", []string{"docs", "+update", "--doc-token", "x", "--content", "y"}, "write_not_allowed"},
+		{"leaf out of allow list", []string{"schema"}, "domain_not_allowed"},
+	}
+	for _, tc := range cases {
+		t.Run(tc.name, func(t *testing.T) {
+			assertReasonCodeEnvelope(t, run(t, bin, tc.args...), tc.reasonCode)
+		})
+	}
+}
+
+// TestReadonlyAllows asserts the allow-path: a read command inside an
+// allowed domain must NOT be denied by the policy gate. It may still fail
+// downstream (e.g. api/auth error), but that failure must not carry the
+// denial envelope shape and must not exit 2.
+func TestReadonlyAllows(t *testing.T) {
+	bin := buildFork(t, "readonly", readonlyPlugin)
+	res := run(t, bin, "docs", "+fetch", "--doc", "nonexistent")
+	if res.exit == 2 {
+		t.Fatalf("read command was denied (exit=2); stderr=%s", res.stderr)
+	}
+	if gjson.Valid(res.stderr) && gjson.Get(res.stderr, "error.subtype").String() == "failed_precondition" {
+		t.Errorf("read command produced a denial envelope; stderr=%s", res.stderr)
+	}
+}
+
+// identityPlugin registers a Restrict rule scoped to bot identities only.
+// im +messages-search declares AuthTypes:["user"] (see
+// shortcuts/im/im_messages_search.go), so it has no intersection with the
+// rule's bot-only whitelist regardless of which --as value the caller
+// passes: platform.Rule.Identities is checked against the command's own
+// static supported-identities annotation, not the runtime --as flag.
+const identityPlugin = `// Code generated by plugin_e2e; DO NOT EDIT.
+package plugin
+
+import "github.com/larksuite/cli/extension/platform"
+
+func init() {
+	platform.Register(
+		platform.NewPlugin("identity-restrict", "0.1.0").
+			Restrict(&platform.Rule{
+				Name:       "bot-only",
+				Allow:      []string{"im/**"},
+				MaxRisk:    platform.RiskRead,
+				Identities: []platform.Identity{platform.IdentityBot},
+			}).
+			MustBuild())
+}
+`
+
+// denylistPlugin registers a Restrict rule that allows the docs/** domain
+// but explicitly denies docs/+search (a real read-risk leaf, see
+// shortcuts/doc/docs_search.go). Deny has priority over Allow, so the
+// command is rejected before MaxRisk is even consulted.
+const denylistPlugin = `// Code generated by plugin_e2e; DO NOT EDIT.
+package plugin
+
+import "github.com/larksuite/cli/extension/platform"
+
+func init() {
+	platform.Register(
+		platform.NewPlugin("denylist-restrict", "0.1.0").
+			Restrict(&platform.Rule{
+				Name:    "deny-search",
+				Allow:   []string{"docs/**"},
+				Deny:    []string{"docs/+search"},
+				MaxRisk: platform.RiskRead,
+			}).
+			MustBuild())
+}
+`
+
+// multiRulePlugin registers two scope-exclusive Restrict rules (im-only,
+// docs-only). A command outside both domains (e.g. the top-level "schema"
+// command, itself read-risk and already proven to hit domain_not_allowed
+// under a single Allow:["docs/**","im/**"] rule in TestReadonlyDenial) is
+// rejected by both rules, so cmdpolicy's OR-engine collapses the two
+// per-rule denials into the aggregate reason_code "no_matching_rule".
+const multiRulePlugin = `// Code generated by plugin_e2e; DO NOT EDIT.
+package plugin
+
+import "github.com/larksuite/cli/extension/platform"
+
+func init() {
+	platform.Register(
+		platform.NewPlugin("multi-rule-restrict", "0.1.0").
+			Restrict(&platform.Rule{
+				Name:    "im-only",
+				Allow:   []string{"im/**"},
+				MaxRisk: platform.RiskRead,
+			}).
+			Restrict(&platform.Rule{
+				Name:    "docs-only",
+				Allow:   []string{"docs/**"},
+				MaxRisk: platform.RiskRead,
+			}).
+			MustBuild())
+}
+`
+
+// assertReasonCodeEnvelope asserts the VERIFIED envelope shape shared by every
+// reason_code across this package -- both policy denials (this file) and
+// install-time failures (install_test.go): exit 2, valid JSON on stderr,
+// error.type=="validation", error.subtype=="failed_precondition", and
+// error.hint containing "reason_code <wantReasonCode>". Both paths render
+// through the SAME cmd/platform_guards.go WithHint(...) family, embedding
+// reason_code in the hint STRING, not a structured error.detail.reason_code
+// field (contradicting internal/platform/error.go:34's comment).
+func assertReasonCodeEnvelope(t *testing.T, res result, wantReasonCode string) {
+	t.Helper()
+	if res.exit != 2 {
+		t.Fatalf("exit=%d stdout=%s stderr=%s", res.exit, res.stdout, res.stderr)
+	}
+	if !gjson.Valid(res.stderr) {
+		t.Fatalf("stderr not JSON: %s", res.stderr)
+	}
+	if got := gjson.Get(res.stderr, "error.type").String(); got != "validation" {
+		t.Errorf("error.type=%q want validation", got)
+	}
+	if got := gjson.Get(res.stderr, "error.subtype").String(); got != "failed_precondition" {
+		t.Errorf("error.subtype=%q want failed_precondition", got)
+	}
+	if hint := gjson.Get(res.stderr, "error.hint").String(); !strings.Contains(hint, "reason_code "+wantReasonCode) {
+		t.Errorf("hint=%q want to contain reason_code %s", hint, wantReasonCode)
+	}
+}
+
+// TestIdentityMismatchDenial pins reason_code=identity_mismatch: a bot-only
+// rule rejects a command whose declared AuthTypes don't include "bot".
+func TestIdentityMismatchDenial(t *testing.T) {
+	bin := buildFork(t, "identity", identityPlugin)
+	res := run(t, bin, "im", "+messages-search", "--as", "user")
+	t.Logf("exit=%d stdout=%s stderr=%s", res.exit, res.stdout, res.stderr)
+	assertReasonCodeEnvelope(t, res, "identity_mismatch")
+}
+
+// TestDenylistDenial pins reason_code=command_denylisted: a Deny glob hit
+// rejects the command even though it also matches Allow.
+func TestDenylistDenial(t *testing.T) {
+	bin := buildFork(t, "denylist", denylistPlugin)
+	res := run(t, bin, "docs", "+search")
+	t.Logf("exit=%d stdout=%s stderr=%s", res.exit, res.stdout, res.stderr)
+	assertReasonCodeEnvelope(t, res, "command_denylisted")
+}
+
+// TestMultiRuleDenial pins reason_code=no_matching_rule: a command rejected
+// by every rule in a multi-Restrict() plugin gets the aggregate reason_code,
+// not either rule's own per-rule reason_code.
+func TestMultiRuleDenial(t *testing.T) {
+	bin := buildFork(t, "multirule", multiRulePlugin)
+	res := run(t, bin, "schema")
+	t.Logf("exit=%d stdout=%s stderr=%s", res.exit, res.stdout, res.stderr)
+	assertReasonCodeEnvelope(t, res, "no_matching_rule")
+}

tests/plugin_e2e/smoke_test.go

@@ -0,0 +1,69 @@
+// Copyright (c) 2026 Lark Technologies Pte. Ltd.
+// SPDX-License-Identifier: MIT
+
+package plugin_e2e
+
+import (
+	"os"
+	"path/filepath"
+	"testing"
+)
+
+// TestMain archives HEAD's committed tree once for the whole package before
+// any fork build runs. It lives here (not in harness.go) because `go test`
+// only discovers TestMain in a _test.go file — a TestMain defined in a plain
+// .go file is silently never invoked.
+// NOTE: exactly one TestMain is allowed per package — do not add another in other _test.go files here.
+func TestMain(m *testing.M) {
+	root, err := repoRoot()
+	if err != nil {
+		panic("locate repo root: " + err.Error())
+	}
+	baseDir, err = os.MkdirTemp("", "plugin-e2e-")
+	if err != nil {
+		panic("mkdtemp: " + err.Error())
+	}
+	cleanTree = filepath.Join(baseDir, "larkcli-clean")
+	if err := os.MkdirAll(cleanTree, 0o755); err != nil {
+		panic("mkdir clean tree: " + err.Error())
+	}
+	if err := gitArchive(root, cleanTree); err != nil {
+		panic("git archive: " + err.Error())
+	}
+	code := m.Run()
+	_ = os.RemoveAll(baseDir)
+	os.Exit(code)
+}
+
+// noopPlugin registers a plugin that installs nothing observable, proving
+// the blank-import -> init -> Register -> InstallAll assembly chain links
+// and the fork boots.
+const noopPlugin = `// Code generated by plugin_e2e; DO NOT EDIT.
+package plugin
+
+import (
+	"context"
+
+	"github.com/larksuite/cli/extension/platform"
+)
+
+func init() {
+	platform.Register(
+		platform.NewPlugin("smoke", "0.0.1").
+			Observer(platform.After, "noop", platform.All(),
+				func(_ context.Context, _ platform.Invocation) {}).
+			FailOpen().
+			MustBuild())
+}
+`
+
+func TestSmokeForkBoots(t *testing.T) {
+	bin := buildFork(t, "smoke", noopPlugin)
+	res := run(t, bin, "--help")
+	if res.exit != 0 {
+		t.Fatalf("--help exit=%d stderr=%s", res.exit, res.stderr)
+	}
+	if res.stdout == "" {
+		t.Fatalf("--help produced empty stdout")
+	}
+}

tests/sidecar_e2e/roundtrip_test.go

@@ -0,0 +1,648 @@
+// Copyright (c) 2026 Lark Technologies Pte. Ltd.
+// SPDX-License-Identifier: MIT
+
+//go:build authsidecar
+
+// Package sidecar_e2e proves the sidecar auth-proxy wire protocol end-to-end,
+// offline and secret-free: a real fork binary (built with -tags authsidecar,
+// exercising the REAL extension/transport/sidecar interceptor) signs a
+// request with HMAC-SHA256 and routes it to an in-test sidecar, which
+// verifies the signature using the REAL sidecar.Verify / sidecar.CanonicalRequest
+// from github.com/larksuite/cli/sidecar, injects a synthetic token, and
+// forwards to an in-test mock upstream.
+//
+// DEVIATION FROM THE ORIGINAL PLAN: the plan called for driving the real
+// sidecar/server-demo binary (built with -tags authsidecar_demo) as the
+// middle process. That is infeasible for an OFFLINE test, for three
+// independent reasons, all verified in source:
+//
+//  1. sidecar/server-demo/handler.go:171 resolves a REAL token via
+//     h.cred.ResolveToken(...), which errors out unless the machine has run
+//     `lark-cli auth login` — there is no way to make it return a token
+//     without live credentials.
+//  2. sidecar/server-demo/main.go builds handler.allowedHosts from
+//     core.ResolveEndpoints(BrandFeishu/BrandLark) only — real feishu/lark
+//     hosts. An in-test mock (127.0.0.1:<port>) is never in that allowlist
+//     and would be rejected with 403 (handler.go step 4).
+//  3. sidecar/server-demo/handler.go:184 pins the forward scheme to
+//     "https://" + targetHost, ignoring the client-supplied scheme. It can
+//     never be redirected to an http:// mock.
+//
+// server-demo's verify+inject logic is ALREADY covered by
+// `go test -tags authsidecar_demo ./sidecar/server-demo/` (see the
+// sidecar-test Makefile target, item 3) — that is unit-level coverage of the
+// same code paths this file would otherwise exercise via a real subprocess.
+//
+// So instead, this test builds its OWN in-test sidecar (an httptest.Server)
+// that mirrors server-demo/handler.go's verify+inject steps 0-8 exactly,
+// using the real protocol package (sidecar.Verify, sidecar.CanonicalRequest,
+// sidecar.BodySHA256, the Header* / Sentinel* / Identity* constants) — the
+// same symbols server-demo itself uses. This is the standard shape for this
+// kind of test: one real external process (the fork binary, compiled with
+// the production interceptor code) plus two in-process httptest.Server
+// stand-ins (sidecar, upstream). It proves the real wire protocol end-to-end
+// without requiring live credentials, real feishu/lark hosts, or TLS.
+//
+// Every key/token/app-id here is an obviously-synthetic placeholder; nothing
+// in this file can authenticate against anything real.
+package sidecar_e2e
+
+import (
+	"bytes"
+	"context"
+	"encoding/json"
+	"errors"
+	"fmt"
+	"io"
+	"net/http"
+	"net/http/httptest"
+	"net/url"
+	"os"
+	"os/exec"
+	"path/filepath"
+	"strings"
+	"sync"
+	"testing"
+	"time"
+
+	"github.com/larksuite/cli/sidecar"
+)
+
+// Synthetic, obviously-fake fixtures. None of these are real secrets.
+const (
+	testProxyKey  = "test-proxy-key-not-a-real-secret-000000000000"
+	testAppID     = "cli_test_app_not_real"
+	injectedToken = "fake-injected-token-not-real"
+
+	// testDocToken is the --doc argument runFork passes; the docs +fetch call
+	// becomes POST /open-apis/docs_ai/v1/documents/<testDocToken>/fetch. Sharing
+	// it keeps the request marker below in sync with the command invocation.
+	testDocToken = "nonexistent"
+
+	// docsReqPath is the exact path of the TARGET docs +fetch request among
+	// every request the fork routes through the proxy. `docs +fetch --as user`
+	// resolves a sentinel UAT, and the credential layer then verifies it with
+	// a mandatory /open-apis/authen/v1/user_info probe (see
+	// internal/credential/credential_provider.go enrichUserInfo) — so a second
+	// request also flows through the sidecar. Asserting on whichever arrived
+	// last would let that identity probe masquerade as the docs request; we
+	// select the docs call by its full path (exact match, not a substring — a
+	// wrong API prefix or version must not slip through).
+	docsReqPath = "/open-apis/docs_ai/v1/documents/" + testDocToken + "/fetch"
+
+	// wantProxyTargetHost is the real Feishu open-platform host the interceptor
+	// must name as the proxy target for BRAND=feishu. The request is never
+	// actually forwarded there (the in-test sidecar redirects to the mock); the
+	// header only records where the fork BELIEVED it was going, and it is HMAC
+	// signing input, so it must be exactly the real host.
+	wantProxyTargetHost = "open.feishu.cn"
+)
+
+// TestSidecarHMACRoundTrip drives the whole wire protocol as three named
+// steps so the flow is readable at a glance; each step's mechanics live in a
+// dedicated helper below.
+func TestSidecarHMACRoundTrip(t *testing.T) {
+	// Two in-process stand-ins: the mock upstream (for open.feishu.cn) and the
+	// in-test sidecar (server-demo's verify+inject, via the real protocol pkg).
+	upstream := startMockUpstream(t)
+	sc := startInTestSidecar(t, []byte(testProxyKey), upstream.URL)
+
+	// One real external process: lark-cli built with -tags authsidecar, run
+	// fully offline against the in-test sidecar.
+	bin := buildAuthsidecarFork(t)
+	res := runFork(t, bin, sc.URL)
+
+	// Diagnostic dump (shown only on failure or -v): the full request set, so a
+	// failure makes plain which requests flowed and which one the assertions
+	// targeted, instead of guessing about the last-arriving request.
+	for _, s := range sc.seenAll() {
+		t.Logf("sidecar saw: %s %s target=%q identity=%q verifyRan=%v verifyErr=%v",
+			s.req.method, s.req.path, s.req.headers.Get(sidecar.HeaderProxyTarget),
+			s.req.headers.Get(sidecar.HeaderProxyIdentity), s.verifyRan, s.verifyErr)
+	}
+	for _, r := range upstream.sink.all() {
+		t.Logf("upstream saw: %s %s auth=%q", r.method, r.path, r.headers.Get("Authorization"))
+	}
+	t.Logf("fork exit=%d\nstdout=%s\nstderr=%s", res.exit, res.stdout, res.stderr)
+
+	// Assert the fork's command itself succeeded end-to-end, not just that some
+	// bytes reached the sidecar.
+	assertForkSucceeded(t, res)
+
+	// Assert the three properties of a correct round trip, scoped to the DOCS
+	// request (not an auxiliary identity probe).
+	assertInterceptorSigned(t, sc)                  // (a)+(c) fork -> sidecar
+	assertInjectedTokenReachedUpstream(t, upstream) // (b)     sidecar -> upstream
+}
+
+// --- request capture -------------------------------------------------------
+
+// capturedRequest snapshots the parts of an *http.Request that matter for
+// assertions, taken before the request (and its body reader) is consumed or
+// goes out of scope.
+type capturedRequest struct {
+	method  string
+	path    string
+	headers http.Header
+	body    []byte
+}
+
+// requestSink stores EVERY request a stub server saw, in arrival order,
+// guarded so the httptest handler goroutine and the test goroutine can hand
+// them over safely. Capturing all requests (not just the last) is what closes
+// the false-green window: the fork may route more than one request through the
+// proxy, and the target docs request is not guaranteed to be the last.
+type requestSink struct {
+	mu   sync.Mutex
+	reqs []*capturedRequest
+}
+
+func (s *requestSink) capture(r *http.Request, body []byte) *capturedRequest {
+	snap := &capturedRequest{
+		method:  r.Method,
+		path:    r.URL.RequestURI(),
+		headers: r.Header.Clone(),
+		body:    body,
+	}
+	s.mu.Lock()
+	s.reqs = append(s.reqs, snap)
+	s.mu.Unlock()
+	return snap
+}
+
+func (s *requestSink) all() []*capturedRequest {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	return append([]*capturedRequest(nil), s.reqs...)
+}
+
+// find returns the first captured request whose path equals path, or nil.
+func (s *requestSink) find(path string) *capturedRequest {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	for _, r := range s.reqs {
+		if r.path == path {
+			return r
+		}
+	}
+	return nil
+}
+
+// --- mock upstream (stands in for open.feishu.cn) --------------------------
+
+type mockUpstream struct {
+	*httptest.Server
+	sink requestSink
+}
+
+func startMockUpstream(t *testing.T) *mockUpstream {
+	t.Helper()
+	m := &mockUpstream{}
+	m.Server = httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		body, _ := io.ReadAll(r.Body)
+		m.sink.capture(r, body)
+		w.Header().Set("Content-Type", "application/json")
+		w.WriteHeader(http.StatusOK)
+		// Respond per path so each forwarded request parses as success: the
+		// identity probe needs authen/v1/user_info's {data:{open_id,name}} to
+		// resolve cleanly; the docs +fetch is satisfied by the generic code:0
+		// envelope. A single canned body would make the identity probe error.
+		_, _ = w.Write(mockResponseFor(r.URL.Path))
+	}))
+	t.Cleanup(m.Close)
+	return m
+}
+
+// mockResponseFor returns a minimal success body matching the API the given
+// path addresses. Unknown paths get a generic code:0 envelope.
+func mockResponseFor(path string) []byte {
+	if strings.Contains(path, "/authen/v1/user_info") {
+		return []byte(`{"code":0,"msg":"success","data":{"open_id":"ou_mock","name":"mock user"}}`)
+	}
+	return []byte(`{"code":0,"msg":"success","data":{}}`)
+}
+
+// --- in-test sidecar (mirrors server-demo/handler.go verify+inject) --------
+
+// sidecarSeen is one request the in-test sidecar received, together with the
+// per-request verification outcome. Tracking this per request (not as a single
+// last-write-wins field) lets assertions check the verification that belongs
+// to the DOCS request specifically.
+type sidecarSeen struct {
+	req       *capturedRequest
+	verifyRan bool
+	verifyErr error
+}
+
+type inTestSidecar struct {
+	*httptest.Server
+	key         []byte
+	upstreamURL string
+
+	mu   sync.Mutex // guards seen
+	seen []sidecarSeen
+}
+
+func startInTestSidecar(t *testing.T, key []byte, upstreamURL string) *inTestSidecar {
+	t.Helper()
+	s := &inTestSidecar{key: key, upstreamURL: upstreamURL}
+	s.Server = httptest.NewServer(http.HandlerFunc(s.handle))
+	t.Cleanup(s.Close)
+	return s
+}
+
+// handle is the request flow: capture -> verify (steps 0-4) -> inject+forward.
+func (s *inTestSidecar) handle(w http.ResponseWriter, r *http.Request) {
+	body, _ := io.ReadAll(r.Body)
+	snap := &capturedRequest{
+		method:  r.Method,
+		path:    r.URL.RequestURI(),
+		headers: r.Header.Clone(),
+		body:    body,
+	}
+
+	authHeader, verifyRan, verifyErr, ok := s.verifyProxyRequest(w, r, body)
+	s.mu.Lock()
+	s.seen = append(s.seen, sidecarSeen{req: snap, verifyRan: verifyRan, verifyErr: verifyErr})
+	s.mu.Unlock()
+	if !ok {
+		return
+	}
+	s.forwardWithInjectedToken(w, r, body, authHeader)
+}
+
+// verifyProxyRequest mirrors server-demo/handler.go steps 0-4: protocol
+// version, body SHA256, target validation, and HMAC signature verification.
+// It returns the auth header the client committed to, whether verification
+// (step 4) actually ran, and its result. On any pre-step-4 failure it writes
+// the HTTP error and returns ok=false with verifyRan=false.
+func (s *inTestSidecar) verifyProxyRequest(w http.ResponseWriter, r *http.Request, body []byte) (authHeader string, verifyRan bool, verifyErr error, ok bool) {
+	// Step 0: protocol version.
+	version := r.Header.Get(sidecar.HeaderProxyVersion)
+	if version != sidecar.ProtocolV1 {
+		http.Error(w, "unsupported "+sidecar.HeaderProxyVersion+": "+version, http.StatusBadRequest)
+		return "", false, nil, false
+	}
+
+	// Step 1-2: timestamp + body SHA256.
+	ts := r.Header.Get(sidecar.HeaderProxyTimestamp)
+	claimedSHA := r.Header.Get(sidecar.HeaderBodySHA256)
+	if claimedSHA == "" || claimedSHA != sidecar.BodySHA256(body) {
+		http.Error(w, "body SHA256 mismatch", http.StatusBadRequest)
+		return "", false, nil, false
+	}
+
+	// Step 3: target host, identity, auth-header (all covered by the sig).
+	targetHost, perr := parseTargetHost(r.Header.Get(sidecar.HeaderProxyTarget))
+	if perr != nil {
+		http.Error(w, "invalid "+sidecar.HeaderProxyTarget+": "+perr.Error(), http.StatusForbidden)
+		// verifyRan=false: step 4 never ran; surface the parse error so the
+		// diagnostic dump shows why this request was rejected early.
+		return "", false, perr, false
+	}
+	identity := r.Header.Get(sidecar.HeaderProxyIdentity)
+	authHeader = r.Header.Get(sidecar.HeaderProxyAuthHeader)
+
+	// Step 4: verify HMAC signature over the canonical request.
+	err := sidecar.Verify(s.key, sidecar.CanonicalRequest{
+		Version:      version,
+		Method:       r.Method,
+		Host:         targetHost,
+		PathAndQuery: r.URL.RequestURI(),
+		BodySHA256:   claimedSHA,
+		Timestamp:    ts,
+		Identity:     identity,
+		AuthHeader:   authHeader,
+	}, r.Header.Get(sidecar.HeaderProxySignature))
+	if err != nil {
+		http.Error(w, "HMAC verification failed: "+err.Error(), http.StatusUnauthorized)
+		return "", true, err, false
+	}
+	return authHeader, true, nil, true
+}
+
+// forwardWithInjectedToken mirrors server-demo's inject+forward. Unlike
+// server-demo (which forwards to "https://"+targetHost), this test forwards to
+// the in-test MOCK's URL — proving the sidecar's inject step without needing a
+// real upstream or a route to targetHost. It strips any client-supplied auth
+// headers first (the sidecar is the sole source of auth material), injects the
+// synthetic token into the committed header, and relays the response back.
+func (s *inTestSidecar) forwardWithInjectedToken(w http.ResponseWriter, r *http.Request, body []byte, authHeader string) {
+	freq, err := http.NewRequest(r.Method, s.upstreamURL+r.URL.RequestURI(), bytes.NewReader(body))
+	if err != nil {
+		http.Error(w, "failed to build forward request", http.StatusInternalServerError)
+		return
+	}
+	for k, vs := range r.Header {
+		if isProxyHeader(k) {
+			continue
+		}
+		for _, v := range vs {
+			freq.Header.Add(k, v)
+		}
+	}
+	freq.Header.Del("Authorization")
+	freq.Header.Del(sidecar.HeaderMCPUAT)
+	freq.Header.Del(sidecar.HeaderMCPTAT)
+
+	if authHeader == "Authorization" {
+		freq.Header.Set("Authorization", "Bearer "+injectedToken)
+	} else {
+		freq.Header.Set(authHeader, injectedToken)
+	}
+
+	resp, err := http.DefaultClient.Do(freq)
+	if err != nil {
+		http.Error(w, "forward failed: "+err.Error(), http.StatusBadGateway)
+		return
+	}
+	defer resp.Body.Close()
+	respBody, _ := io.ReadAll(resp.Body)
+	for k, vs := range resp.Header {
+		for _, v := range vs {
+			w.Header().Add(k, v)
+		}
+	}
+	w.WriteHeader(resp.StatusCode)
+	_, _ = w.Write(respBody)
+}
+
+// seenAll returns a copy of every request the sidecar received, in order.
+func (s *inTestSidecar) seenAll() []sidecarSeen {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	return append([]sidecarSeen(nil), s.seen...)
+}
+
+// findSeen returns the first received request whose path equals path, or nil.
+func (s *inTestSidecar) findSeen(path string) *sidecarSeen {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	for i := range s.seen {
+		if s.seen[i].req.path == path {
+			return &s.seen[i]
+		}
+	}
+	return nil
+}
+
+// isProxyHeader reports whether name is one of the sidecar wire-protocol
+// headers that must not be copied through to the forwarded (mock upstream)
+// request. Mirrors sidecar/server-demo/handler.go's isProxyHeader.
+func isProxyHeader(name string) bool {
+	switch http.CanonicalHeaderKey(name) {
+	case http.CanonicalHeaderKey(sidecar.HeaderProxyVersion),
+		http.CanonicalHeaderKey(sidecar.HeaderProxyTarget),
+		http.CanonicalHeaderKey(sidecar.HeaderProxyIdentity),
+		http.CanonicalHeaderKey(sidecar.HeaderProxySignature),
+		http.CanonicalHeaderKey(sidecar.HeaderProxyTimestamp),
+		http.CanonicalHeaderKey(sidecar.HeaderBodySHA256),
+		http.CanonicalHeaderKey(sidecar.HeaderProxyAuthHeader):
+		return true
+	}
+	return false
+}
+
+// parseTargetHost validates X-Lark-Proxy-Target and returns its host.
+// Mirrors sidecar/server-demo/handler.go's parseTarget: the header must be
+// "https://<host>" with no path, query, fragment, or userinfo. Only the host
+// is used, both as HMAC signing input and to record what the fork believed
+// its real destination was — the actual forward in this test always goes to
+// the in-test mock, never to this host.
+func parseTargetHost(target string) (string, error) {
+	u, err := url.Parse(target)
+	if err != nil {
+		return "", fmt.Errorf("parse: %w", err)
+	}
+	if u.Scheme != "https" {
+		return "", fmt.Errorf("scheme must be https, got %q", u.Scheme)
+	}
+	if u.Host == "" {
+		return "", fmt.Errorf("missing host")
+	}
+	if u.User != nil {
+		return "", fmt.Errorf("userinfo not allowed")
+	}
+	if u.Path != "" && u.Path != "/" {
+		return "", fmt.Errorf("path not allowed (got %q)", u.Path)
+	}
+	if u.RawQuery != "" {
+		return "", fmt.Errorf("query not allowed")
+	}
+	if u.Fragment != "" {
+		return "", fmt.Errorf("fragment not allowed")
+	}
+	return u.Host, nil
+}
+
+// --- fork build + run ------------------------------------------------------
+
+// buildAuthsidecarFork builds the REAL lark-cli with -tags authsidecar (the
+// production interceptor) and returns the binary path.
+func buildAuthsidecarFork(t *testing.T) string {
+	t.Helper()
+	bin := filepath.Join(t.TempDir(), "forkbin")
+	build := exec.Command("go", "build", "-tags", "authsidecar", "-o", bin, ".")
+	build.Dir = repoRoot(t)
+	if out, err := build.CombinedOutput(); err != nil {
+		t.Fatalf("build fork binary: %v\n%s", err, out)
+	}
+	return bin
+}
+
+// forkResult is the fork subprocess outcome.
+type forkResult struct {
+	exit   int
+	stdout string
+	stderr string
+}
+
+// runFork runs the fork against the in-test sidecar, fully offline, and returns
+// its exit code and captured output. LARKSUITE_CLI_REMOTE_META=off is essential:
+// without it the fork's startup metadata refresh hits the real
+// open.feishu.cn/api/tools/open/api_definition (internal/registry/remote.go),
+// which both breaks the "offline, secret-free" contract and makes the run
+// depend on live network. With it set, the command still completes and the
+// docs request still flows through the sidecar, but nothing leaves the machine.
+func runFork(t *testing.T, binPath, sidecarURL string) forkResult {
+	t.Helper()
+	scURL, err := url.Parse(sidecarURL)
+	if err != nil {
+		t.Fatalf("parse sidecar URL: %v", err)
+	}
+
+	ctx, cancel := context.WithTimeout(context.Background(), 60*time.Second)
+	defer cancel()
+	cmd := exec.CommandContext(ctx, binPath, "docs", "+fetch", "--doc", testDocToken, "--as", "user")
+	cmd.Env = append(os.Environ(),
+		"LARKSUITE_CLI_AUTH_PROXY=http://"+scURL.Host,
+		"LARKSUITE_CLI_PROXY_KEY="+testProxyKey,
+		"LARKSUITE_CLI_APP_ID="+testAppID,
+		"LARKSUITE_CLI_BRAND=feishu",
+		"LARKSUITE_CLI_CONFIG_DIR="+t.TempDir(),
+		"LARKSUITE_CLI_REMOTE_META=off",
+		"LARKSUITE_CLI_NO_UPDATE_NOTIFIER=1",
+		"LARKSUITE_CLI_NO_SKILLS_NOTIFIER=1",
+	)
+	var stdout, stderr bytes.Buffer
+	cmd.Stdout = &stdout
+	cmd.Stderr = &stderr
+	runErr := cmd.Run()
+	exit := 0
+	if runErr != nil {
+		var ee *exec.ExitError
+		if errors.As(runErr, &ee) {
+			exit = ee.ExitCode()
+		} else {
+			t.Fatalf("run fork: %v", runErr)
+		}
+	}
+	return forkResult{exit: exit, stdout: stdout.String(), stderr: stderr.String()}
+}
+
+// repoRoot resolves the lark-cli module root from the test's working
+// directory (which `go test` sets to the package dir, tests/sidecar_e2e).
+func repoRoot(t *testing.T) string {
+	t.Helper()
+	out, err := exec.Command("git", "rev-parse", "--show-toplevel").Output()
+	if err != nil {
+		t.Fatalf("resolve repo root: %v", err)
+	}
+	return strings.TrimSpace(string(out))
+}
+
+// --- assertions ------------------------------------------------------------
+
+// assertForkSucceeded checks the fork command itself completed the round trip:
+// exit 0 and an ok:true JSON envelope on stdout. This is what makes the docs
+// request a genuine success path, not merely bytes that happened to flow.
+func assertForkSucceeded(t *testing.T, res forkResult) {
+	t.Helper()
+	if res.exit != 0 {
+		t.Fatalf("fork exit=%d want 0; stdout=%s stderr=%s", res.exit, res.stdout, res.stderr)
+	}
+	// Parse rather than substring-match: the CLI pretty-prints stdout, so the
+	// envelope reads "ok": true (with a space), and the field's truth — not its
+	// serialized spelling — is what proves the round trip succeeded.
+	var env struct {
+		OK bool `json:"ok"`
+	}
+	if err := json.Unmarshal([]byte(res.stdout), &env); err != nil {
+		t.Fatalf("fork stdout is not a JSON envelope: %v; stdout=%s stderr=%s", err, res.stdout, res.stderr)
+	}
+	if !env.OK {
+		t.Fatalf("fork stdout ok != true (round trip did not succeed); stdout=%s stderr=%s", res.stdout, res.stderr)
+	}
+}
+
+// assertInterceptorSigned checks the fork -> sidecar hop (assertions a + c) for
+// the DOCS request specifically: the real interceptor ran (all proxy headers
+// present, identity=user, method+path+target as expected), stripped every
+// real/sentinel auth header before signing, and produced a signature that
+// verified against the shared key.
+func assertInterceptorSigned(t *testing.T, sc *inTestSidecar) {
+	t.Helper()
+	seen := sc.findSeen(docsReqPath)
+	if seen == nil {
+		t.Fatalf("sidecar never received the docs request (path %q) — interceptor did not route it to AUTH_PROXY; saw %v",
+			docsReqPath, sidecarPaths(sc.seenAll()))
+	}
+	got := seen.req
+	if !seen.verifyRan {
+		t.Fatal("sidecar received the docs request but never reached HMAC verification (rejected earlier — see handler headers)")
+	}
+	if seen.verifyErr != nil {
+		t.Fatalf("HMAC verification failed on the fork's own signed docs request: %v", seen.verifyErr)
+	}
+	t.Logf("fork->sidecar docs headers: %v", got.headers)
+
+	// Target/method/path: prove we asserted on the real docs call to the real
+	// Feishu open platform, not an auxiliary request.
+	if got.method != http.MethodPost {
+		t.Errorf("docs request method = %q, want POST", got.method)
+	}
+	if targetHost, err := parseTargetHost(got.headers.Get(sidecar.HeaderProxyTarget)); err != nil {
+		t.Errorf("docs request %s invalid: %v", sidecar.HeaderProxyTarget, err)
+	} else if targetHost != wantProxyTargetHost {
+		t.Errorf("docs request proxy target host = %q, want %q", targetHost, wantProxyTargetHost)
+	}
+
+	// No real/sentinel auth ever left the fork: the interceptor strips the
+	// sentinel before signing, so this hop must carry no auth header at all.
+	if auth := got.headers.Get("Authorization"); auth != "" {
+		t.Fatalf("fork->sidecar hop leaked an Authorization header (want none, interceptor should have stripped it): %q", auth)
+	}
+	if v := got.headers.Get(sidecar.HeaderMCPUAT); v != "" {
+		t.Fatalf("fork->sidecar hop leaked %s (want none): %q", sidecar.HeaderMCPUAT, v)
+	}
+	if v := got.headers.Get(sidecar.HeaderMCPTAT); v != "" {
+		t.Fatalf("fork->sidecar hop leaked %s (want none): %q", sidecar.HeaderMCPTAT, v)
+	}
+
+	// Proxy headers must be present (proves the interceptor actually ran).
+	for _, h := range []string{
+		sidecar.HeaderProxyVersion, sidecar.HeaderProxyTarget, sidecar.HeaderProxyIdentity,
+		sidecar.HeaderProxySignature, sidecar.HeaderProxyTimestamp, sidecar.HeaderBodySHA256,
+		sidecar.HeaderProxyAuthHeader,
+	} {
+		if got.headers.Get(h) == "" {
+			t.Fatalf("fork->sidecar hop missing required proxy header %s", h)
+		}
+	}
+	if id := got.headers.Get(sidecar.HeaderProxyIdentity); id != sidecar.IdentityUser {
+		t.Fatalf("fork->sidecar identity = %q, want %q", id, sidecar.IdentityUser)
+	}
+}
+
+// assertInjectedTokenReachedUpstream checks the sidecar -> upstream hop
+// (assertion b) for the DOCS request: the mock saw exactly the sidecar-injected
+// synthetic token, never a sentinel or a real one — proving injection happened.
+func assertInjectedTokenReachedUpstream(t *testing.T, up *mockUpstream) {
+	t.Helper()
+	got := up.sink.find(docsReqPath)
+	if got == nil {
+		t.Fatalf("mock upstream never received the forwarded docs request (path %q) — sidecar did not forward it after verification; saw %v",
+			docsReqPath, requestPaths(up.sink.all()))
+	}
+	t.Logf("sidecar->mock docs headers: %v", got.headers)
+
+	wantAuth := "Bearer " + injectedToken
+	gotAuth := got.headers.Get("Authorization")
+	if gotAuth != wantAuth {
+		t.Fatalf("mock upstream Authorization = %q, want %q", gotAuth, wantAuth)
+	}
+	// Belt-and-suspenders: the value the mock saw must not be either sentinel,
+	// proving the only token that ever reached "upstream" was the injected one.
+	if gotAuth == "Bearer "+sidecar.SentinelUAT || gotAuth == "Bearer "+sidecar.SentinelTAT {
+		t.Fatalf("mock upstream received a sentinel token instead of the injected one: %q", gotAuth)
+	}
+	// The sidecar wire-protocol headers are between fork and sidecar only —
+	// the forward must strip every one of them. Token injection alone passing
+	// would still be a leak if signatures/timestamps/digests reached upstream.
+	for _, h := range []string{
+		sidecar.HeaderProxyVersion, sidecar.HeaderProxyTarget, sidecar.HeaderProxyIdentity,
+		sidecar.HeaderProxySignature, sidecar.HeaderProxyTimestamp, sidecar.HeaderBodySHA256,
+		sidecar.HeaderProxyAuthHeader,
+	} {
+		if v := got.headers.Get(h); v != "" {
+			t.Errorf("proxy protocol header %s leaked to upstream (want stripped): %q", h, v)
+		}
+	}
+}
+
+// sidecarPaths / requestPaths render captured paths for failure messages.
+func sidecarPaths(seen []sidecarSeen) []string {
+	paths := make([]string, len(seen))
+	for i, s := range seen {
+		paths[i] = s.req.method + " " + s.req.path
+	}
+	return paths
+}
+
+func requestPaths(reqs []*capturedRequest) []string {
+	paths := make([]string, len(reqs))
+	for i, r := range reqs {
+		paths[i] = r.method + " " + r.path
+	}
+	return paths
+}