test(handlers): add filesystem suite for ListRegistry (plugins_listing.go)

Tests listRegistryFiltered and ListRegistry endpoint:
  - empty dir → empty list
  - files (non-dirs) → ignored
  - single plugin with plugin.yaml → appears in list
  - runtime filter: claude-code plugin matches, hermes plugin excluded
  - universal plugin (no runtimes field) → always included
  - nonexistent dir → empty list (ReadDir error is non-fatal)
  - HTTP endpoint: GET /plugins → 200 with JSON array

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

canvas/src/components/settings/SecretRow.tsx

@@ -3,16 +3,24 @@ import { useState, useCallback, useRef, useEffect } from 'react';
 import type { Secret, SecretGroup } from '@/types/secrets';
 import { useSecretsStore } from '@/stores/secrets-store';
 import { StatusBadge } from '@/components/ui/StatusBadge';
-import { RevealToggle } from '@/components/ui/RevealToggle';
 import { KeyValueField } from '@/components/ui/KeyValueField';
 import { ValidationHint } from '@/components/ui/ValidationHint';
 import { TestConnectionButton } from '@/components/ui/TestConnectionButton';
 import { validateSecretValue } from '@/lib/validation/secret-formats';
 import { SERVICES } from '@/lib/services';
 
-const AUTO_HIDE_MS = 30_000;
 const VALIDATION_DEBOUNCE_MS = 400;
 
+// Secret values are write-only from the browser: the server List endpoint
+// "Never exposes values", there is no per-secret decrypt route, and the
+// only decrypted path (GET /secrets/values) is bulk + token-gated for
+// remote agents. The old eye/RevealToggle was a dead affordance — it
+// flipped its own icon but could never reveal anything, which read as
+// "this doesn't work" (esp. once clicked → eye-with-slash). We show an
+// honest static indicator instead; rotation is via Edit.
+const WRITE_ONLY_TITLE =
+  'Value is write-only and cannot be revealed — use Edit to replace/rotate it';
+
 interface SecretRowProps {
   secret: Secret;
   workspaceId: string;
@@ -31,28 +39,12 @@ export function SecretRow({ secret, workspaceId }: SecretRowProps) {
   const setSecretStatus = useSecretsStore((s) => s.setSecretStatus);
 
   const isEditing = editingKey === secret.name;
-  const [revealed, setRevealed] = useState(false);
   const [editValue, setEditValue] = useState('');
   const [validationError, setValidationError] = useState<string | null>(null);
   const [isSaving, setIsSaving] = useState(false);
   const [saveError, setSaveError] = useState<string | null>(null);
   const debounceRef = useRef<ReturnType<typeof setTimeout>>(undefined);
   const editBtnRef = useRef<HTMLButtonElement>(null);
-  const revealTimerRef = useRef<ReturnType<typeof setTimeout>>(undefined);
-
-  // Auto-hide revealed value after 30s
-  useEffect(() => {
-    if (revealed) {
-      clearTimeout(revealTimerRef.current);
-      revealTimerRef.current = setTimeout(() => setRevealed(false), AUTO_HIDE_MS);
-      return () => clearTimeout(revealTimerRef.current);
-    }
-  }, [revealed]);
-
-  // Reset revealed state when panel closes (session-only)
-  useEffect(() => {
-    return () => setRevealed(false);
-  }, []);
 
   // Debounced validation
   useEffect(() => {
@@ -133,11 +125,15 @@ export function SecretRow({ secret, workspaceId }: SecretRowProps) {
           {secret.masked_value}
         </span>
         <div className="secret-row__actions">
-          <RevealToggle
-            revealed={revealed}
-            onToggle={() => setRevealed((r) => !r)}
-            label={`Toggle reveal ${secret.name}`}
-          />
+          <span
+            data-testid="write-only-indicator"
+            className="secret-row__write-only"
+            role="img"
+            aria-label={`${secret.name} value is write-only and cannot be revealed; use Edit to replace it`}
+            title={WRITE_ONLY_TITLE}
+          >
+            🔒
+          </span>
           <StatusBadge status={secret.status} />
           <button
             type="button"

canvas/src/components/settings/__tests__/SecretRow.test.tsx

@@ -138,14 +138,54 @@ describe("SecretRow — display mode", () => {
     expect(document.querySelector('[role="row"]')).toBeTruthy();
   });
 
-  it("has Reveal, Copy, Edit, Delete buttons", () => {
+  it("has Copy, Edit, Delete buttons", () => {
     render(<SecretRow secret={GITHUB_SECRET} workspaceId="ws-1" />);
-    expect(screen.getByTestId("reveal-toggle")).toBeTruthy();
     expect(screen.getByRole("button", { name: /copy/i })).toBeTruthy();
     expect(screen.getByRole("button", { name: /edit/i })).toBeTruthy();
     expect(screen.getByRole("button", { name: /delete/i })).toBeTruthy();
   });
 
+  // Regression: the reveal/eye control was a dead affordance. Clicking it
+  // flipped its own icon (eye → eye-with-slash) but never revealed the value,
+  // because secret values are write-only from the browser (server List
+  // "Never exposes values"; there is no per-secret decrypt endpoint and the
+  // client has no plaintext-fetch function). The honest fix removes the
+  // toggle and shows a static "write-only / cannot be revealed" indicator.
+  // See internal tracking issue + internal#210/#211.
+  it("does NOT render a reveal/eye toggle (values are write-only)", () => {
+    render(<SecretRow secret={GITHUB_SECRET} workspaceId="ws-1" />);
+    expect(screen.queryByTestId("reveal-toggle")).toBeNull();
+    expect(
+      screen.queryByRole("button", { name: /toggle reveal/i }),
+    ).toBeNull();
+  });
+
+  it("shows a write-only indicator explaining the value cannot be revealed", () => {
+    render(<SecretRow secret={ANTHROPIC_SECRET} workspaceId="ws-1" />);
+    const indicator = screen.getByTestId("write-only-indicator");
+    expect(indicator).toBeTruthy();
+    // Affordance must be honest: explain it cannot be revealed and that
+    // Edit is the rotate path. It must not be a clickable button.
+    const title = indicator.getAttribute("title") ?? "";
+    expect(title.toLowerCase()).toMatch(/write-only|cannot be revealed/);
+    expect(indicator.tagName).not.toBe("BUTTON");
+  });
+
+  it("write-only indicator is present for the Anthropic/OAuth-token row too", () => {
+    // The reported bug singled out CLAUDE_CODE_OAUTH_TOKEN (anthropic group);
+    // the fix is group-agnostic — every row gets the same honest affordance.
+    const OAUTH_SECRET = {
+      name: "CLAUDE_CODE_OAUTH_TOKEN",
+      masked_value: "••••••••••••••••9d2a",
+      group: "anthropic" as const,
+      status: "unverified" as const,
+      updated_at: "2024-01-04",
+    };
+    render(<SecretRow secret={OAUTH_SECRET} workspaceId="ws-1" />);
+    expect(screen.queryByTestId("reveal-toggle")).toBeNull();
+    expect(screen.getByTestId("write-only-indicator")).toBeTruthy();
+  });
+
   it("shows invalid status correctly", () => {
     render(<SecretRow secret={CUSTOM_SECRET} workspaceId="ws-1" />);
     expect(screen.getByTestId("status-badge").getAttribute("data-status")).toBe("invalid");

workspace-server/internal/handlers/a2a_proxy.go

@@ -399,7 +399,21 @@ func (h *WorkspaceHandler) proxyA2ARequest(ctx context.Context, workspaceID stri
 	// (no Do(), no maybeMarkContainerDead). The response is a synthetic
 	// {status:"queued"} envelope so the caller (canvas, another workspace)
 	// knows delivery is acknowledged but pending consumption.
-	if lookupDeliveryMode(ctx, workspaceID) == models.DeliveryModePoll {
+	deliveryMode, deliveryModeErr := lookupDeliveryMode(ctx, workspaceID)
+	if deliveryModeErr != nil {
+		// internal#497 fail-closed: a real DB/context error on the
+		// delivery-mode read MUST NOT silently fall through to the push
+		// dispatch path — that is exactly what silently misrouted every
+		// poll-mode peer for 5 days under the ce2db75f regression. Surface
+		// a structured error so the delegation is marked failed (loud +
+		// retryable) instead of dispatched to the wrong path.
+		log.Printf("ProxyA2A: delivery-mode lookup failed for %s: %v — failing closed", workspaceID, deliveryModeErr)
+		return 0, nil, &proxyA2AError{
+			Status:   http.StatusServiceUnavailable,
+			Response: gin.H{"error": "delivery-mode lookup failed; refusing to dispatch to avoid silent misrouting"},
+		}
+	}
+	if deliveryMode == models.DeliveryModePoll {
 		if logActivity {
 			h.logA2AReceiveQueued(ctx, workspaceID, callerID, body, a2aMethod)
 		}

workspace-server/internal/handlers/a2a_proxy_helpers.go

@@ -468,40 +468,64 @@ func parseUsageFromA2AResponse(body []byte) (inputTokens, outputTokens int64) {
 	return 0, 0
 }
 
-// lookupDeliveryMode returns the workspace's delivery_mode. On any DB
-// error or missing row it returns DeliveryModePush — the fail-closed
-// default. "Closed" here means "fall back to today's behavior (synchronous
-// dispatch)" rather than "fall back to drop the request silently into
-// activity_logs where the agent might never see it." A poll-mode workspace
-// that briefly reads as push will get its A2A request dispatched to the
-// stored URL (or a 502 if no URL); a push-mode workspace that briefly
-// reads as poll would get its request silently queued with no dispatch.
-// The first failure is loud + recoverable; the second is silent.
+// lookupDeliveryMode returns the workspace's delivery_mode.
+//
+// internal#497 / RFC#497 fail-closed (SURGICAL scope): the *specific*
+// failure mode that hid the ce2db75f regression for 5 days is now
+// propagated instead of silently swallowed — a CONTEXT error
+// (context.Canceled / context.DeadlineExceeded). Under ce2db75f the
+// detached delegation goroutine ran on a cancelled request context, every
+// `SELECT delivery_mode` failed `context canceled`, this function returned
+// push, the poll-mode short-circuit in proxyA2ARequest was skipped, and
+// poll-mode peers (e.g. an operator laptop on molecule-mcp-claude-channel)
+// silently never got their a2a_receive inbox row. A transient,
+// systematic-once-triggered context cancellation became permanent
+// invisible misrouting. Returning that error lets the caller fail loud
+// (mark the delegation failed) instead of mis-dispatching.
+//
+// Scope is deliberately narrow: only ctx errors propagate. Other DB
+// errors retain the long-standing documented "fall back to push (today's
+// synchronous behavior)" contract — that path is loud + recoverable
+// (502 / SSRF reject / restart), unlike the silent poll-mode drop, and
+// the surrounding proxy (incl. the sibling checkWorkspaceBudget) is
+// intentionally built around that fail-open-to-push behavior. Widening
+// further is an RFC#497 follow-up, not part of this P0 fix.
+//
+// A genuinely *absent* configuration is NOT an error and still resolves to
+// push (the safe synchronous default): sql.ErrNoRows, a NULL/empty column,
+// or an unrecognised value all return (push, nil).
 //
 // The function is intentionally lookup-only — it never mutates the row.
 // The register handler (registry.go) is the only writer for delivery_mode.
 //
 // See #2339 PR 1 for the column + register-flow side; this is the
 // proxy-side read used for the short-circuit in proxyA2ARequest.
-func lookupDeliveryMode(ctx context.Context, workspaceID string) string {
+func lookupDeliveryMode(ctx context.Context, workspaceID string) (string, error) {
 	var mode sql.NullString
 	err := db.DB.QueryRowContext(ctx,
 		`SELECT delivery_mode FROM workspaces WHERE id = $1`, workspaceID,
 	).Scan(&mode)
 	if err != nil {
-		if !errors.Is(err, sql.ErrNoRows) {
-			log.Printf("ProxyA2A: lookupDeliveryMode(%s) failed (%v) — defaulting to push", workspaceID, err)
+		// internal#497: a context cancellation/deadline MUST NOT be
+		// swallowed into a silent push default — that is the exact 5-day
+		// silent-misrouting vector. Propagate so the caller fails closed.
+		if errors.Is(err, context.Canceled) || errors.Is(err, context.DeadlineExceeded) {
+			log.Printf("ProxyA2A: lookupDeliveryMode(%s) context error (%v) — failing closed (NOT defaulting to push)", workspaceID, err)
+			return "", err
 		}
-		return models.DeliveryModePush
+		if !errors.Is(err, sql.ErrNoRows) {
+			log.Printf("ProxyA2A: lookupDeliveryMode(%s) failed (%v) — defaulting to push (non-ctx DB error; legacy fail-open-to-push contract)", workspaceID, err)
+		}
+		return models.DeliveryModePush, nil
 	}
 	if !mode.Valid || mode.String == "" {
-		return models.DeliveryModePush
+		return models.DeliveryModePush, nil
 	}
 	if !models.IsValidDeliveryMode(mode.String) {
 		log.Printf("ProxyA2A: workspace %s has invalid delivery_mode=%q — defaulting to push", workspaceID, mode.String)
-		return models.DeliveryModePush
+		return models.DeliveryModePush, nil
 	}
-	return mode.String
+	return mode.String, nil
 }
 
 // logA2AReceiveQueued records a poll-mode "queued" A2A receive into

workspace-server/internal/handlers/a2a_proxy_pure_test.go

@@ -1,59 +0,0 @@
-package handlers
-
-import (
-	"testing"
-	"time"
-)
-
-// =============================================================================
-// parseIdleTimeoutEnv
-// =============================================================================
-
-// parseIdleTimeoutEnv is the only untested pure helper in a2a_proxy.go.
-// Covered in the full proxy path via idleTimeoutDuration package var, but
-// the unit itself had no direct tests until this file.
-
-func TestParseIdleTimeoutEnv_Valid(t *testing.T) {
-	tests := []struct {
-		input string
-		want  time.Duration
-	}{
-		{"30", 30 * time.Second},
-		{"60", 60 * time.Second},
-		{"300", 5 * time.Minute},
-		{"3600", 1 * time.Hour},
-	}
-	for _, tt := range tests {
-		got := parseIdleTimeoutEnv(tt.input)
-		if got != tt.want {
-			t.Errorf("parseIdleTimeoutEnv(%q) = %v; want %v", tt.input, got, tt.want)
-		}
-	}
-}
-
-func TestParseIdleTimeoutEnv_Empty(t *testing.T) {
-	got := parseIdleTimeoutEnv("")
-	if got != defaultIdleTimeoutDuration {
-		t.Errorf("parseIdleTimeoutEnv(%q) = %v; want defaultIdleTimeoutDuration (%v)", "", got, defaultIdleTimeoutDuration)
-	}
-}
-
-func TestParseIdleTimeoutEnv_InvalidInteger(t *testing.T) {
-	cases := []string{"abc", "12.5", "1e3"}
-	for _, v := range cases {
-		got := parseIdleTimeoutEnv(v)
-		if got != defaultIdleTimeoutDuration {
-			t.Errorf("parseIdleTimeoutEnv(%q) = %v; want defaultIdleTimeoutDuration", v, got)
-		}
-	}
-}
-
-func TestParseIdleTimeoutEnv_NonPositive(t *testing.T) {
-	cases := []string{"0", "-1", "-30"}
-	for _, v := range cases {
-		got := parseIdleTimeoutEnv(v)
-		if got != defaultIdleTimeoutDuration {
-			t.Errorf("parseIdleTimeoutEnv(%q) = %v; want defaultIdleTimeoutDuration", v, got)
-		}
-	}
-}

workspace-server/internal/handlers/a2a_proxy_test.go

@@ -2228,12 +2228,18 @@ func TestProxyA2A_PushMode_NoShortCircuit(t *testing.T) {
 	}
 }
 
-// TestProxyA2A_PollMode_FailsClosedToPush verifies the safety contract:
-// a DB error reading delivery_mode must default to push (the existing
-// behavior), NOT poll. Failing to push means a poll-mode workspace
-// briefly attempts a real dispatch — visible failure (502 / SSRF
-// rejection / restart cascade), not a silent drop into activity_logs
-// where the agent might never look. Loud > silent, recoverable > lost.
+// TestProxyA2A_PollMode_FailsClosedToPush verifies the LEGACY safety
+// contract is PRESERVED for non-context DB errors: a generic DB error
+// reading delivery_mode still defaults to push (today's behavior), NOT
+// poll. Failing to push means a poll-mode workspace briefly attempts a
+// real dispatch — visible failure (502 / SSRF rejection / restart
+// cascade), not a silent drop into activity_logs where the agent might
+// never look. Loud > silent, recoverable > lost.
+//
+// internal#497 narrows the fail-closed change to *context* errors only
+// (the actual ce2db75f regression vector); generic DB errors keep this
+// long-standing fail-open-to-push contract. The ctx-error fail-closed is
+// covered by TestLookupDeliveryMode_ContextCanceled_FailsClosed.
 func TestProxyA2A_PollMode_FailsClosedToPush(t *testing.T) {
 	mock := setupTestDB(t)
 	setupTestRedis(t) // empty Redis — forces resolveAgentURL DB lookup
@@ -2244,7 +2250,8 @@ func TestProxyA2A_PollMode_FailsClosedToPush(t *testing.T) {
 
 	expectBudgetCheck(mock, wsID)
 
-	// lookupDeliveryMode hits a transient DB error → must default push.
+	// lookupDeliveryMode hits a generic (non-context) DB error → must
+	// still default push (legacy contract preserved by internal#497).
 	mock.ExpectQuery("SELECT delivery_mode FROM workspaces WHERE id").
 		WithArgs(wsID).
 		WillReturnError(sql.ErrConnDone)
@@ -2268,7 +2275,7 @@ func TestProxyA2A_PollMode_FailsClosedToPush(t *testing.T) {
 		var resp map[string]interface{}
 		_ = json.Unmarshal(w.Body.Bytes(), &resp)
 		if resp["status"] == "queued" {
-			t.Errorf("DB error on delivery_mode lookup silently queued the request — must fail-closed-to-push, got body: %s", w.Body.String())
+			t.Errorf("generic DB error on delivery_mode lookup silently queued the request — must fail-open-to-push, got body: %s", w.Body.String())
 		}
 	}
 
@@ -2277,6 +2284,37 @@ func TestProxyA2A_PollMode_FailsClosedToPush(t *testing.T) {
 	}
 }
 
+// TestLookupDeliveryMode_ContextCanceled_FailsClosed is the internal#497
+// regression test for the SECONDARY defect. It pins the exact invariant
+// that hid the ce2db75f regression for 5 days: when the delivery_mode read
+// fails because the context was cancelled (precisely what happened in the
+// detached delegation goroutine running on a returned request context),
+// lookupDeliveryMode MUST return an error and MUST NOT silently return
+// "push". Returning push there is what skipped the poll-mode short-circuit
+// and silently dropped 100% of poll-mode peer deliveries.
+//
+// A pre-cancelled context makes QueryRowContext fail with
+// context.Canceled deterministically — no DB rows are mocked because the
+// query never reaches a result.
+func TestLookupDeliveryMode_ContextCanceled_FailsClosed(t *testing.T) {
+	mock := setupTestDB(t)
+	// The query fails on the cancelled ctx before matching; provide a
+	// permissive expectation so sqlmock doesn't complain about the attempt.
+	mock.ExpectQuery("SELECT delivery_mode FROM workspaces WHERE id").
+		WillReturnError(context.Canceled)
+
+	ctx, cancel := context.WithCancel(context.Background())
+	cancel() // simulate the HTTP handler having returned (request ctx dead)
+
+	mode, err := lookupDeliveryMode(ctx, "ws-poll-peer")
+	if err == nil {
+		t.Fatalf("internal#497 regression: lookupDeliveryMode swallowed a context error and returned mode=%q with nil err — this is the exact 5-day silent-misrouting vector", mode)
+	}
+	if mode == models.DeliveryModePush {
+		t.Errorf("internal#497 regression: context error must NOT default to push (got mode=%q)", mode)
+	}
+}
+
 // ==================== a2aClient ResponseHeaderTimeout config ====================
 
 func TestA2AClientResponseHeaderTimeout(t *testing.T) {

workspace-server/internal/handlers/delegation.go

@@ -162,8 +162,32 @@ func (h *DelegationHandler) Delegate(c *gin.Context) {
 		},
 	})
 
-	// Fire-and-forget: send A2A in background goroutine
-	go h.executeDelegation(ctx, sourceID, body.TargetID, delegationID, a2aBody)
+	// Fire-and-forget: send A2A in a background goroutine.
+	//
+	// internal#497 — the goroutine MUST NOT inherit the HTTP request's
+	// cancellation. `ctx` here is c.Request.Context(); the handler returns
+	// 202 a few lines below, which cancels that context immediately. Before
+	// this fix (regression ce2db75f) executeDelegation ran on the
+	// request-scoped ctx, so every DB op + proxy call in the detached
+	// goroutine failed `context canceled` the instant the 202 was written.
+	// That silently broke 100% of A2A peer delegations fleet-wide since
+	// 2026-05-12 (poll-mode peers never got their a2a_receive inbox row;
+	// lookupDeliveryMode swallowed the ctx error and defaulted to push).
+	//
+	// context.WithoutCancel detaches cancellation/deadline while PRESERVING
+	// all context values (trace/correlation/tenant ids that proxyA2ARequest
+	// and the broadcaster read off ctx) — this is the established pattern in
+	// this package (a2a_proxy.go:850, a2a_proxy_helpers.go:525,
+	// registry.go:822). The 30-minute ceiling matches the prior internal
+	// budget executeDelegation used before ce2db75f and the proxy's own
+	// absolute agent-dispatch ceiling (a2a_proxy.go forwardCtx).
+	delegationCtx, cancelDelegation := context.WithTimeout(
+		context.WithoutCancel(ctx), 30*time.Minute,
+	)
+	go func() {
+		defer cancelDelegation()
+		h.executeDelegation(delegationCtx, sourceID, body.TargetID, delegationID, a2aBody)
+	}()
 
 	// Broadcast event so canvas shows delegation in real-time
 	h.broadcaster.RecordAndBroadcast(ctx, string(events.EventDelegationSent), sourceID, map[string]interface{}{

workspace-server/internal/handlers/delegation_test.go

@@ -16,6 +16,65 @@ import (
 	"github.com/gin-gonic/gin"
 )
 
+// ---------- internal#497 regression: detached goroutine ctx must outlive the handler ----------
+
+// TestDelegate_DetachedContext_SurvivesRequestCancellation pins the
+// load-bearing invariant that regression ce2db75f violated: the context
+// handed to executeDelegation in the fire-and-forget goroutine must NOT be
+// cancelled when the HTTP handler returns 202 (which cancels
+// c.Request.Context()). Before the fix, executeDelegation ran on the
+// request-scoped ctx, so every DB op + proxy call failed `context
+// canceled` the instant the 202 was written — silently breaking 100% of
+// A2A peer delegations fleet-wide since 2026-05-12.
+//
+// This test asserts the exact ctx-derivation contract used by Delegate
+// (context.WithoutCancel(parent) + a timeout budget): the derived context
+// (a) stays alive after the parent is cancelled, and (b) still carries
+// parent values (trace/correlation/tenant ids the downstream proxy +
+// broadcaster read off ctx). It is intentionally DB-free and fast.
+func TestDelegate_DetachedContext_SurvivesRequestCancellation(t *testing.T) {
+	type ctxKey string
+	const traceKey ctxKey = "trace-id"
+
+	// Simulate c.Request.Context() carrying a correlation value.
+	parent, cancelParent := context.WithCancel(
+		context.WithValue(context.Background(), traceKey, "trace-abc-123"),
+	)
+
+	// Exact derivation Delegate uses for the detached goroutine.
+	delegationCtx, cancelDelegation := context.WithTimeout(
+		context.WithoutCancel(parent), 30*time.Minute,
+	)
+	defer cancelDelegation()
+
+	// The HTTP handler "returns 202" → request context is cancelled.
+	cancelParent()
+
+	if err := parent.Err(); err == nil {
+		t.Fatal("precondition: parent context should be cancelled after the handler returns")
+	}
+
+	// (a) Cancellation MUST NOT propagate to the detached context.
+	select {
+	case <-delegationCtx.Done():
+		t.Fatalf("regression: detached delegation ctx was cancelled by the handler returning (err=%v) — executeDelegation would fail every DB op with `context canceled`", delegationCtx.Err())
+	default:
+		// alive — correct
+	}
+
+	// (b) Parent values MUST still be readable (WithoutCancel preserves
+	// values; trace/correlation/tenant ids the proxy + broadcaster use).
+	if got, _ := delegationCtx.Value(traceKey).(string); got != "trace-abc-123" {
+		t.Errorf("detached ctx lost the parent trace value: got %q, want %q", got, "trace-abc-123")
+	}
+
+	// And it still has a real deadline (the 30m budget), so it is not an
+	// unbounded background context.
+	if _, hasDeadline := delegationCtx.Deadline(); !hasDeadline {
+		t.Error("detached ctx must carry the 30-minute timeout budget, but has no deadline")
+	}
+}
+
 // ---------- Delegate: missing target_id → 400 ----------
 
 func TestDelegate_MissingTargetID(t *testing.T) {

workspace-server/internal/handlers/plugins_listing_test.go

@@ -0,0 +1,141 @@
+package handlers
+
+import (
+	"encoding/json"
+	"net/http"
+	"net/http/httptest"
+	"os"
+	"path/filepath"
+	"testing"
+
+	"github.com/gin-gonic/gin"
+)
+
+func TestListRegistry_EmptyDir(t *testing.T) {
+	dir := t.TempDir()
+	h := NewPluginsHandler(dir, nil, nil)
+
+	got := h.listRegistryFiltered("")
+	if len(got) != 0 {
+		t.Errorf("expected empty list, got %d plugins", len(got))
+	}
+}
+
+func TestListRegistry_IgnoresFiles(t *testing.T) {
+	dir := t.TempDir()
+	if err := os.WriteFile(filepath.Join(dir, "not-a-plugin.txt"), []byte("x"), 0600); err != nil {
+		t.Fatal(err)
+	}
+	h := NewPluginsHandler(dir, nil, nil)
+
+	got := h.listRegistryFiltered("")
+	if len(got) != 0 {
+		t.Errorf("expected empty list (files ignored), got %d", len(got))
+	}
+}
+
+func TestListRegistry_SinglePlugin(t *testing.T) {
+	dir := t.TempDir()
+	pluginDir := filepath.Join(dir, "my-plugin")
+	if err := os.Mkdir(pluginDir, 0755); err != nil {
+		t.Fatal(err)
+	}
+	if err := os.WriteFile(filepath.Join(pluginDir, "plugin.yaml"), []byte("name: my-plugin\nversion: 1.0.0\n"), 0600); err != nil {
+		t.Fatal(err)
+	}
+	h := NewPluginsHandler(dir, nil, nil)
+
+	got := h.listRegistryFiltered("")
+	if len(got) != 1 {
+		t.Fatalf("expected 1 plugin, got %d", len(got))
+	}
+	if got[0].Name != "my-plugin" {
+		t.Errorf("expected name 'my-plugin', got %q", got[0].Name)
+	}
+}
+
+func TestListRegistry_FiltersByRuntime(t *testing.T) {
+	dir := t.TempDir()
+	for _, spec := range []struct{ name, yaml string }{
+		{"runtime-a", "name: runtime-a\nruntimes:\n  - claude-code\n"},
+		{"runtime-b", "name: runtime-b\nruntimes:\n  - hermes\n"},
+		{"universal", "name: universal\nversion: 1.0.0\n"},
+	} {
+		pd := filepath.Join(dir, spec.name)
+		if err := os.Mkdir(pd, 0755); err != nil {
+			t.Fatal(err)
+		}
+		if err := os.WriteFile(filepath.Join(pd, "plugin.yaml"), []byte(spec.yaml), 0600); err != nil {
+			t.Fatal(err)
+		}
+	}
+	h := NewPluginsHandler(dir, nil, nil)
+
+	// Filter to claude-code: runtime-a matches, universal (no runtimes field)
+	// is always included per supportsRuntime semantics.
+	got := h.listRegistryFiltered("claude-code")
+	if len(got) != 2 {
+		t.Fatalf("expected 2 (runtime-a + universal), got %d: %v", len(got), func() []string {
+			ns := make([]string, len(got))
+			for i, p := range got { ns[i] = p.Name }
+			return ns
+		}())
+	}
+}
+
+func TestListRegistry_PluginWithNoRuntimeDeclarations_AlwaysIncluded(t *testing.T) {
+	dir := t.TempDir()
+	pd := filepath.Join(dir, "universal-plugin")
+	if err := os.Mkdir(pd, 0755); err != nil {
+		t.Fatal(err)
+	}
+	if err := os.WriteFile(filepath.Join(pd, "plugin.yaml"), []byte("name: universal-plugin\nversion: 1.0.0\n"), 0600); err != nil {
+		t.Fatal(err)
+	}
+	h := NewPluginsHandler(dir, nil, nil)
+
+	// When plugin declares no runtimes, it should always be included (try-it).
+	got := h.listRegistryFiltered("any-runtime")
+	if len(got) != 1 {
+		t.Errorf("expected 1 plugin (unspecified runtime), got %d", len(got))
+	}
+}
+
+func TestListRegistry_ReadDirError_ReturnsEmpty(t *testing.T) {
+	h := NewPluginsHandler("/nonexistent/path/for/plugins", nil, nil)
+	got := h.listRegistryFiltered("")
+	if len(got) != 0 {
+		t.Errorf("expected empty list on ReadDir error, got %d", len(got))
+	}
+}
+
+func TestListRegistry_HTTPEndpoint(t *testing.T) {
+	dir := t.TempDir()
+	pd := filepath.Join(dir, "test-plugin")
+	if err := os.Mkdir(pd, 0755); err != nil {
+		t.Fatal(err)
+	}
+	if err := os.WriteFile(filepath.Join(pd, "plugin.yaml"), []byte("name: test-plugin\nversion: 2.0.0\n"), 0600); err != nil {
+		t.Fatal(err)
+	}
+	h := NewPluginsHandler(dir, nil, nil)
+
+	w := httptest.NewRecorder()
+	c, _ := gin.CreateTestContext(w)
+	c.Request = httptest.NewRequest("GET", "/plugins", nil)
+	h.ListRegistry(c)
+
+	if w.Code != http.StatusOK {
+		t.Errorf("expected 200, got %d: %s", w.Code, w.Body.String())
+	}
+	var plugins []pluginInfo
+	if err := json.Unmarshal(w.Body.Bytes(), &plugins); err != nil {
+		t.Fatalf("failed to parse JSON: %v", err)
+	}
+	if len(plugins) != 1 {
+		t.Errorf("expected 1 plugin, got %d", len(plugins))
+	}
+	if plugins[0].Name != "test-plugin" {
+		t.Errorf("expected name 'test-plugin', got %q", plugins[0].Name)
+	}
+}