fix(tests): replace remaining sk-ant-api03- fixtures with non-matching tokens

The secret-scan workflow flags sk-ant-[A-Za-z0-9_-]{40,} patterns. Two sk-ant-api03-* fixture tokens (47 and 62 chars) were present in test_sanitize_agent_error_reason_scrubs_all_secret_formats. They were not replaced by PR #1430 (which only fixed the sk-ant-DEADBEEF* tokens). Replace with tokens that still exercise the same scrubber paths: - BARE sk-* case (≥24 chars after "sk-"): use sk-FAKEPLACEHOLDER... (53 chars total; starts with "sk-" so the bare-pattern scrubber catches it, but lacks "sk-ant-" so the secret-scan pattern does not fire). - JSON-quoted apiKey value (≥24 chars): use anon_fakefakefake... (45 chars; satisfies the JSON-quoted redaction path; does not match any secret-scan credential pattern). Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Merge pull request 'fix(tests)+build: unblock secret scan and Runtime PR-Built on #1420 ' (#1430 ) from runtime/fix-test-fixture-v3 into fix/issue212-actionable-agent-error-reason
2026-05-17 16:34:31 +00:00 · 2026-05-17 16:18:01 +00:00 · 2026-05-17 15:48:31 +00:00 · 2026-05-17 07:56:16 -07:00 · 2026-05-17 07:20:14 -07:00 · 2026-05-17 13:46:55 +00:00
34 changed files with 2205 additions and 72 deletions
@@ -16,7 +16,40 @@ interface TokensTabProps {
  workspaceId: string;
 }

+// The settings panel passes the literal sentinel "global" when no canvas
+// node is selected. Workspace tokens are inherently per-workspace — there
+// is no /workspaces/global/tokens endpoint (querying the uuid column with
+// "global" 500s on Postgres). The org-wide equivalent lives in the
+// separate "Org API Keys" tab. Mirrors the sentinel-awareness that
+// api/secrets.ts already has (workspaceId === 'global' → /settings/secrets).
+const GLOBAL_WORKSPACE_ID = 'global';
+
 export function TokensTab({ workspaceId }: TokensTabProps) {
+  if (workspaceId === GLOBAL_WORKSPACE_ID) {
+    return (
+      <div className="p-4 space-y-4">
+        <div>
+          <h3 className="text-sm font-semibold text-ink">API Tokens</h3>
+          <p className="text-[10px] text-ink-mid mt-0.5">
+            Bearer tokens for authenticating API calls to this workspace.
+          </p>
+        </div>
+        <div className="text-center py-6">
+          <p className="text-xs text-ink-mid">Select a workspace node first</p>
+          <p className="text-[10px] text-ink-mid mt-1">
+            Workspace tokens are scoped to a single workspace. Select a node
+            on the canvas to manage its tokens, or use the{' '}
+            <span className="text-accent font-medium">Org API Keys</span> tab
+            for org-wide API keys.
+          </p>
+        </div>
+      </div>
+    );
+  }
+  return <WorkspaceTokensTab workspaceId={workspaceId} />;
+}
+
+function WorkspaceTokensTab({ workspaceId }: TokensTabProps) {
  const [tokens, setTokens] = useState<Token[]>([]);
  const [loading, setLoading] = useState(true);
  const [creating, setCreating] = useState(false);
@@ -302,3 +302,35 @@ describe("TokensTab — error", () => {
    expect(document.querySelector('[role="status"]')).toBeNull();
  });
 });
+
+// ─── "global" sentinel (no node selected) ────────────────────────────────────
+//
+// Regression: SettingsPanel passes the literal "global" when no canvas
+// node is selected. workspace tokens are per-workspace and there is no
+// /workspaces/global/tokens endpoint — calling it 500'd
+// ("invalid input syntax for type uuid: global"). The tab must NOT call
+// the API in that state and must point the user at the Org API Keys tab.
+describe("TokensTab — global sentinel (no node selected)", () => {
+  beforeEach(() => {
+    mockApiGet.mockReset();
+    mockApiPost.mockReset();
+    mockApiGet.mockRejectedValue(new Error("should not be called"));
+  });
+
+  it("does not call the API and shows a pointer to Org API Keys", async () => {
+    render(<TokensTab workspaceId="global" />);
+    await flush();
+    expect(mockApiGet).not.toHaveBeenCalled();
+    expect(mockApiPost).not.toHaveBeenCalled();
+    expect(document.body.textContent).toContain("Select a workspace node");
+    expect(document.body.textContent).toContain("Org API Keys");
+    // No error banner, no scary 500 surfacing.
+    expect(document.querySelector(".text-bad")).toBeNull();
+  });
+
+  it("has no create button in the global state", async () => {
+    render(<TokensTab workspaceId="global" />);
+    await flush();
+    expect(document.body.textContent).not.toContain("New Token");
+  });
+});
@@ -176,7 +176,7 @@ export function deriveProvidersFromModels(models: ModelSpec[]): string[] {
 // exactly the point of the platform adaptor. The deep `~/.hermes/
 // config.yaml` on the container is a separate runtime-internal file,
 // not this one.
-const RUNTIMES_WITH_OWN_CONFIG = new Set<string>(["external", "kimi", "kimi-cli"]);
+const RUNTIMES_WITH_OWN_CONFIG = new Set<string>(["external", "kimi", "kimi-cli", "openclaw"]);

 const FALLBACK_RUNTIME_OPTIONS: RuntimeOption[] = [
  { value: "", label: "LangGraph (default)", models: [], providers: [] },
@@ -45,11 +45,54 @@ export function FilesTab({ workspaceId, data }: Props) {
  if (data && isExternalLikeRuntime(data.runtime)) {
    return <NotAvailablePanel runtime={data.runtime} />;
  }
-  return <PlatformOwnedFilesTab workspaceId={workspaceId} />;
+  return <PlatformOwnedFilesTab workspaceId={workspaceId} runtime={data?.runtime} />;
 }

-function PlatformOwnedFilesTab({ workspaceId }: { workspaceId: string }) {
-  const [root, setRoot] = useState("/configs");
+/** Picks the initial root for the FilesTab dropdown based on the
+ *  workspace's runtime. Decision: per-runtime default (Hongming
+ *  2026-05-15, internal#425 Decisions §2).
+ *
+ *  - openclaw → `/agent-home` (the agent's identity/state — the
+ *    user-facing interesting files for that runtime live in
+ *    `~/.openclaw/` inside the container, which `/agent-home` maps to
+ *    via the Phase 2b docker-exec backend).
+ *  - everything else (claude-code, hermes, external-like, undefined)
+ *    → `/configs` (the legacy default — managed config that flows
+ *    through the per-runtime indirection in
+ *    workspace-server/internal/handlers/template_files_eic.go).
+ *
+ *  When the runtime is undefined (legacy callers that don't thread
+ *  `data` through, or a workspace whose runtime field hasn't loaded
+ *  yet) the default is `/configs` — matches today's behaviour, no
+ *  surprise.
+ *
+ *  Note on `/agent-home` pre-Phase-2b: the backend short-circuits
+ *  with HTTP 501 and the canonical "implementation pending" body.
+ *  The tab renders empty + the error banner explains. This is by
+ *  design — lets us land the canvas UX before the backend ships,
+ *  per the RFC's phased rollout. The 501 is graceful: it doesn't
+ *  poison error toasts or generate "workspace not found" noise.
+ *
+ *  Adding a new runtime that should default to `/agent-home`: add it
+ *  to the agentHomeDefaultRuntimes set below. Adding a runtime that
+ *  should default to a different root: extend this function. */
+const agentHomeDefaultRuntimes = new Set(["openclaw"]);
+
+function defaultRootForRuntime(runtime: string | undefined): string {
+  if (runtime && agentHomeDefaultRuntimes.has(runtime)) {
+    return "/agent-home";
+  }
+  return "/configs";
+}
+
+function PlatformOwnedFilesTab({
+  workspaceId,
+  runtime,
+}: {
+  workspaceId: string;
+  runtime?: string;
+}) {
+  const [root, setRoot] = useState(() => defaultRootForRuntime(runtime));
  const [selectedFile, setSelectedFile] = useState<string | null>(null);
  const [fileContent, setFileContent] = useState("");
  const [editContent, setEditContent] = useState("");
@@ -3,6 +3,22 @@
 import { useRef } from "react";
 import { getIcon } from "./tree";

+// secretShapeMarker is the canonical body the workspace-server Files
+// API returns when a file's path OR content matched a credential
+// regex (internal#425 RFC, Phase 2b — backed by
+// workspace-server/internal/secrets.ScanBytes). The marker is a
+// fixed prefix so the canvas can detect it without parsing JSON and
+// without round-tripping the matched bytes through the editor (which
+// would defeat the purpose — clipboard, browser history, log
+// surfaces would all see them).
+//
+// Today (Phase 1 / before 2b ships) the backend returns 501 for the
+// only root that uses this path, so the marker is dead code until
+// 2b lands. Wiring it in now keeps the canvas + backend contracts
+// aligned in one PR rather than a follow-up. The constant is
+// importable so a future test can pin the exact string.
+export const SECRET_SHAPE_DENIED_MARKER = "<denied: secret-shape>";
+
 interface Props {
  selectedFile: string | null;
  fileContent: string;
@@ -31,6 +47,22 @@ export function FileEditor({
  const editorRef = useRef<HTMLTextAreaElement>(null);
  const isDirty = editContent !== fileContent;

+  // internal#425 Phase 3: detect the secret-shape denial marker and
+  // render a placeholder instead of the editor. The marker comes
+  // from workspace-server Phase 2b (secrets.ScanBytes) which refuses
+  // to surface the file's bytes. We deliberately don't expose
+  // the matched pattern's Name here — the canvas just shows the
+  // generic denial. The Files API log surface has the Pattern.Name
+  // for operators who need to debug a false positive.
+  const isSecretShapeDenied = fileContent === SECRET_SHAPE_DENIED_MARKER;
+
+  // /agent-home is read-only from the canvas (Phase 2b ships read +
+  // delete; Phase-2b-followup may add write). Edits to /configs are
+  // unchanged. Until 2b ships, /agent-home returns 501 so this
+  // read-only gate is also dead code, but wiring it in now keeps
+  // the UI honest the moment 2b lands without a follow-up canvas PR.
+  const isReadOnlyRoot = root !== "/configs";
+
  if (!selectedFile) {
    return (
      <div className="flex-1 flex items-center justify-center">
@@ -75,11 +107,42 @@ export function FileEditor({
      {/* Editor area */}
      {loadingFile ? (
        <div className="p-4 text-xs text-ink-mid">Loading...</div>
+      ) : isSecretShapeDenied ? (
+        // Files API refused to surface this file's bytes because its
+        // path or content matched a credential regex
+        // (workspace-server/internal/secrets, internal#425 Phase 2b).
+        // We render a placeholder INSTEAD OF the textarea so the
+        // matched bytes never enter the DOM. Clipboard / view-source
+        // / element-inspector all see the placeholder, not the
+        // credential.
+        <div
+          role="region"
+          aria-label="File content denied"
+          className="flex-1 flex items-center justify-center p-6 bg-surface"
+        >
+          <div className="max-w-md text-center space-y-2">
+            <div className="text-2xl opacity-40">🛡️</div>
+            <p className="text-[11px] font-mono text-warm">
+              {SECRET_SHAPE_DENIED_MARKER}
+            </p>
+            <p className="text-[10px] text-ink-mid leading-relaxed">
+              The platform refused to surface this file because its
+              path or content matched a credential-shape pattern.
+              The bytes never left the workspace container.
+            </p>
+            <p className="text-[10px] text-ink-mid leading-relaxed">
+              If this is a false positive (test fixture, docs example,
+              or content that happens to share a credential's shape),
+              rename the file or adjust the content via the workspace
+              terminal so the regex no longer matches, then refresh.
+            </p>
+          </div>
+        </div>
      ) : (
        <textarea
          ref={editorRef}
          value={editContent}
-          readOnly={root !== "/configs"}
+          readOnly={isReadOnlyRoot}
          onChange={(e) => setEditContent(e.target.value)}
          onKeyDown={(e) => {
            if ((e.metaKey || e.ctrlKey) && e.key === "s") {
@@ -38,6 +38,15 @@ export function FilesToolbar({
          <option value="/home">/home</option>
          <option value="/workspace">/workspace</option>
          <option value="/plugins">/plugins</option>
+          {/* internal#425 Phase 1+3: container-internal $HOME root.
+              Backend lands the docker-exec dispatch in Phase 2b. Until
+              then the stub returns 501 with a canonical
+              "implementation pending" message — the dropdown renders
+              the option so the canvas affordance is design-frozen
+              even before the backend ships.
+              Runtime-default selection logic in FilesTab.tsx picks
+              this as the initial value for openclaw workspaces. */}
+          <option value="/agent-home">/agent-home</option>
        </select>
        <span className="text-[10px] text-ink-mid">{fileCount} files</span>
      </div>
@@ -0,0 +1,181 @@
+// @vitest-environment jsdom
+/**
+ * Tests for the /agent-home root selector + per-runtime default-root
+ * + secret-shape denial placeholder (internal#425 Phase 3).
+ *
+ * Separate file so the diff is reviewable as a unit and the existing
+ * FilesToolbar / FileEditor / FilesTab tests don't have to grow
+ * agent-home-specific cases. Once Phase 2b lands, the read-only +
+ * 501-stub assertions here can be tightened (or moved into the main
+ * test file as the agent-home root becomes a first-class affordance).
+ */
+import React from "react";
+import { render, screen, cleanup } from "@testing-library/react";
+import { afterEach, describe, expect, it, vi } from "vitest";
+
+import { FilesToolbar } from "../FilesToolbar";
+import {
+  FileEditor,
+  SECRET_SHAPE_DENIED_MARKER,
+} from "../FileEditor";
+
+afterEach(cleanup);
+
+describe("internal#425 Phase 3 — /agent-home root selector", () => {
+  it("dropdown includes /agent-home as an option", () => {
+    // Pins the affordance is in the DOM even pre-Phase-2b — the
+    // canvas design freezes today, the backend lands the dispatch
+    // later. Without this, a future refactor that drops the option
+    // would silently regress the RFC's Phase 1 contract (canvas
+    // visibility) without breaking any other test.
+    render(
+      <FilesToolbar
+        root="/configs"
+        setRoot={vi.fn()}
+        fileCount={0}
+        onNewFile={vi.fn()}
+        onUpload={vi.fn()}
+        onDownloadAll={vi.fn()}
+        onClearAll={vi.fn()}
+        onRefresh={vi.fn()}
+      />,
+    );
+    const select = screen.getByRole("combobox", {
+      name: /file root directory/i,
+    }) as HTMLSelectElement;
+    const values = Array.from(select.options).map((o) => o.value);
+    expect(values).toContain("/agent-home");
+  });
+
+  it("dropdown shows /agent-home as the SELECTED root when prop is /agent-home", () => {
+    render(
+      <FilesToolbar
+        root="/agent-home"
+        setRoot={vi.fn()}
+        fileCount={0}
+        onNewFile={vi.fn()}
+        onUpload={vi.fn()}
+        onDownloadAll={vi.fn()}
+        onClearAll={vi.fn()}
+        onRefresh={vi.fn()}
+      />,
+    );
+    const select = screen.getByRole("combobox", {
+      name: /file root directory/i,
+    }) as HTMLSelectElement;
+    expect(select.value).toBe("/agent-home");
+  });
+});
+
+describe("internal#425 Phase 3 — secret-shape denial placeholder", () => {
+  // Files API Phase 2b returns SECRET_SHAPE_DENIED_MARKER as the file
+  // body when the file's path or content matched a credential regex.
+  // The editor MUST render the marker as a placeholder, not pump it
+  // through the textarea — that would put the marker (and any future
+  // matched bytes if the backend contract changes) into the DOM
+  // value, clipboard, and inspector.
+
+  it("renders the denial placeholder INSTEAD of the textarea when fileContent is the marker", () => {
+    render(
+      <FileEditor
+        selectedFile="agent/.openclaw/secrets.env"
+        fileContent={SECRET_SHAPE_DENIED_MARKER}
+        editContent={SECRET_SHAPE_DENIED_MARKER}
+        setEditContent={vi.fn()}
+        loadingFile={false}
+        saving={false}
+        success={null}
+        root="/agent-home"
+        onSave={vi.fn()}
+        onDownload={vi.fn()}
+      />,
+    );
+    // Placeholder region present
+    expect(
+      screen.getByRole("region", { name: /file content denied/i }),
+    ).toBeTruthy();
+    // Marker text visible (so a debugging operator sees the canonical
+    // contract string without having to dig into the source).
+    expect(screen.getByText(SECRET_SHAPE_DENIED_MARKER)).toBeTruthy();
+    // Critically: NO textarea — the bytes never reach a controlled
+    // input. A regression that re-introduces the textarea path would
+    // make the matched marker (and any future content) selectable +
+    // copyable.
+    expect(screen.queryByRole("textbox")).toBeNull();
+  });
+
+  it("renders the textarea normally when fileContent is regular content", () => {
+    render(
+      <FileEditor
+        selectedFile="config.yaml"
+        fileContent="name: openclaw\n"
+        editContent="name: openclaw\n"
+        setEditContent={vi.fn()}
+        loadingFile={false}
+        saving={false}
+        success={null}
+        root="/configs"
+        onSave={vi.fn()}
+        onDownload={vi.fn()}
+      />,
+    );
+    expect(screen.getByRole("textbox")).toBeTruthy();
+    expect(screen.queryByRole("region", { name: /file content denied/i }))
+      .toBeNull();
+  });
+
+  it("/agent-home renders textarea READ-ONLY for non-denied content", () => {
+    // Phase 2b ships read + delete on /agent-home; write semantics
+    // are decided later. Until then, the canvas presents the editor
+    // as read-only so a user can't type into a buffer that the
+    // backend will refuse to PUT. Without this gate, the user would
+    // edit, hit Save, get a 501, and lose their context for why.
+    render(
+      <FileEditor
+        selectedFile=".openclaw/agent-card.json"
+        fileContent='{"name":"openclaw"}'
+        editContent='{"name":"openclaw"}'
+        setEditContent={vi.fn()}
+        loadingFile={false}
+        saving={false}
+        success={null}
+        root="/agent-home"
+        onSave={vi.fn()}
+        onDownload={vi.fn()}
+      />,
+    );
+    const textarea = screen.getByRole("textbox") as HTMLTextAreaElement;
+    expect(textarea.readOnly).toBe(true);
+  });
+
+  it("/configs renders textarea WRITABLE (regression guard for the read-only gate)", () => {
+    render(
+      <FileEditor
+        selectedFile="config.yaml"
+        fileContent="name: x\n"
+        editContent="name: x\n"
+        setEditContent={vi.fn()}
+        loadingFile={false}
+        saving={false}
+        success={null}
+        root="/configs"
+        onSave={vi.fn()}
+        onDownload={vi.fn()}
+      />,
+    );
+    const textarea = screen.getByRole("textbox") as HTMLTextAreaElement;
+    expect(textarea.readOnly).toBe(false);
+  });
+});
+
+describe("internal#425 Phase 3 — marker constant is the canonical string", () => {
+  // The marker string is part of the canvas <-> workspace-server
+  // contract. The workspace-server emits this exact body; the canvas
+  // detects it by exact-equality. A typo on either side would
+  // silently break detection — the canvas would render the literal
+  // string in the textarea instead of the placeholder. Pin the
+  // contract value here.
+  it("matches the contract value '<denied: secret-shape>'", () => {
+    expect(SECRET_SHAPE_DENIED_MARKER).toBe("<denied: secret-shape>");
+  });
+});
@@ -67,9 +67,21 @@ export function useChatSocket(
            const own = (targetId || msg.workspace_id) === workspaceId;
            if (own) {
              callbacksRef.current.onSendComplete?.();
-              callbacksRef.current.onSendError?.(
-                "Agent error (Exception) — see workspace logs for details.",
-              );
+              // internal#211/#212: surface the runtime's curated,
+              // user-actionable reason (provider HTTP status + error
+              // code + the provider's own guidance, e.g. a 403 "org
+              // disabled · use an API key / ask your admin"). The
+              // server now includes error_detail in the ACTIVITY_LOGGED
+              // broadcast; fall back to summary, and only as a last
+              // resort to a generic line. The old hardcoded
+              // "Agent error (Exception) — see workspace logs for
+              // details." string pointed at a logs UI that does not
+              // exist and discarded the actionable reason entirely.
+              const detail =
+                (p.error_detail as string) ||
+                (p.summary as string) ||
+                "The agent turn failed but the runtime reported no detail. Retry once; if it repeats the workspace runtime may need a restart.";
+              callbacksRef.current.onSendError?.(detail);
            }
          }
        } else if (type === "a2a_send") {
@@ -8,14 +8,18 @@ import { getTenantSlug } from "./tenant";
 export const PLATFORM_URL =
  process.env.NEXT_PUBLIC_PLATFORM_URL ?? "http://localhost:8080";

-// 15s is long enough for slow CP queries but short enough that a
-// hung backend doesn't leave the UI spinning forever. The abort
-// propagates through AbortController so React components can observe
-// the error and render a retry affordance. Callers that know the
-// endpoint is intentionally slow (org import walks a tree of
-// workspaces with server-side pacing) can pass `timeoutMs` to
-// override.
-const DEFAULT_TIMEOUT_MS = 15_000;
+// 35s is long enough for the slowest server-side path (EIC SSH
+// tunnel for tenant EC2 file operations, bounded server-side by
+// `eicFileOpTimeout = 30 * time.Second` in
+// workspace-server/internal/handlers/template_files_eic.go) so the
+// canvas surfaces the server's real error instead of aborting first
+// with a generic timeout. Shorter values caused "Save & Restart" to
+// time out at the client before the backend returned its 5xx. The
+// abort still propagates through AbortController so React components
+// can render a retry affordance. Callers that know an endpoint is
+// intentionally slow (org import walks a tree of workspaces with
+// server-side pacing) can pass `timeoutMs` to override.
+const DEFAULT_TIMEOUT_MS = 35_000;

 export interface RequestOptions {
  timeoutMs?: number;
@@ -62,6 +62,7 @@ TOP_LEVEL_MODULES = {
    "a2a_tools_memory",
    "a2a_tools_messaging",
    "a2a_tools_rbac",
+    "a2a_tools_identity",
    "adapter_base",
    "agent",
    "agents_md",
@@ -194,7 +194,12 @@ func (h *WorkspaceHandler) maybeMarkContainerDead(ctx context.Context, workspace
 	}
 	db.ClearWorkspaceKeys(ctx, workspaceID)
 	h.broadcaster.RecordAndBroadcast(ctx, string(events.EventWorkspaceOffline), workspaceID, map[string]interface{}{})
-	go h.RestartByID(workspaceID)
+	// Tracked via goAsync (not bare `go`) so the asyncWG can be drained
+	// before a test swaps the global db.DB. runRestartCycle reads db.DB
+	// before its provisioner gate, so an untracked detached goroutine
+	// races setupTestDB's t.Cleanup db.DB restore. Matches the already-
+	// correct site at a2a_proxy.go:648.
+	h.goAsync(func() { h.RestartByID(workspaceID) })
 	return true
 }

@@ -241,7 +246,10 @@ func (h *WorkspaceHandler) preflightContainerHealth(ctx context.Context, workspa
 	}
 	db.ClearWorkspaceKeys(ctx, workspaceID)
 	h.broadcaster.RecordAndBroadcast(ctx, string(events.EventWorkspaceOffline), workspaceID, map[string]interface{}{})
-	go h.RestartByID(workspaceID)
+	// Tracked via goAsync (see maybeMarkContainerDead): preflight's
+	// detached restart must be drainable so it doesn't race the global
+	// db.DB swap in test cleanup.
+	h.goAsync(func() { h.RestartByID(workspaceID) })
 	return &proxyA2AError{
 		Status: http.StatusServiceUnavailable,
 		Response: gin.H{
@@ -262,7 +270,8 @@ func (h *WorkspaceHandler) logA2AFailure(ctx context.Context, workspaceID, calle
 		errWsName = workspaceID
 	}
 	summary := "A2A request to " + errWsName + " failed: " + errMsg
-	go func(parent context.Context) {
+	parent := ctx
+	h.goAsync(func() {
 		logCtx, cancel := context.WithTimeout(context.WithoutCancel(parent), 30*time.Second)
 		defer cancel()
 		LogActivity(logCtx, h.broadcaster, ActivityParams{
@@ -277,7 +286,7 @@ func (h *WorkspaceHandler) logA2AFailure(ctx context.Context, workspaceID, calle
 			Status:       "error",
 			ErrorDetail:  &errMsg,
 		})
-	}(ctx)
+	})
 }

 // logA2ASuccess records a successful A2A round-trip and (for canvas-initiated
@@ -298,18 +307,19 @@ func (h *WorkspaceHandler) logA2ASuccess(ctx context.Context, workspaceID, calle
 	// silent workspaces. Only update when callerID is a real workspace (not
 	// canvas, not a system caller) and the target returned 2xx/3xx.
 	if callerID != "" && !isSystemCaller(callerID) && statusCode < 400 {
-		go func() {
+		h.goAsync(func() {
 			bgCtx, cancel := context.WithTimeout(context.Background(), 5*time.Second)
 			defer cancel()
 			if _, err := db.DB.ExecContext(bgCtx,
 				`UPDATE workspaces SET last_outbound_at = NOW() WHERE id = $1`, callerID); err != nil {
 				log.Printf("last_outbound_at update failed for %s: %v", callerID, err)
 			}
-		}()
+		})
 	}
 	summary := a2aMethod + " → " + wsNameForLog
 	toolTrace := extractToolTrace(respBody)
-	go func(parent context.Context) {
+	parent := ctx
+	h.goAsync(func() {
 		logCtx, cancel := context.WithTimeout(context.WithoutCancel(parent), 30*time.Second)
 		defer cancel()
 		LogActivity(logCtx, h.broadcaster, ActivityParams{
@@ -325,7 +335,7 @@ func (h *WorkspaceHandler) logA2ASuccess(ctx context.Context, workspaceID, calle
 			DurationMs:   &durationMs,
 			Status:       logStatus,
 		})
-	}(ctx)
+	})

 	if callerID == "" && statusCode < 400 {
 		h.broadcaster.BroadcastOnly(workspaceID, string(events.EventA2AResponse), map[string]interface{}{
@@ -510,7 +520,8 @@ func (h *WorkspaceHandler) logA2AReceiveQueued(ctx context.Context, workspaceID,
 		wsName = workspaceID
 	}
 	summary := a2aMethod + " → " + wsName + " (queued for poll)"
-	go func(parent context.Context) {
+	parent := ctx
+	h.goAsync(func() {
 		logCtx, cancel := context.WithTimeout(context.WithoutCancel(parent), 30*time.Second)
 		defer cancel()
 		LogActivity(logCtx, h.broadcaster, ActivityParams{
@@ -523,7 +534,7 @@ func (h *WorkspaceHandler) logA2AReceiveQueued(ctx context.Context, workspaceID,
 			RequestBody:  json.RawMessage(body),
 			Status:       "ok",
 		})
-	}(ctx)
+	})
 }

 // readUsageMap extracts input_tokens / output_tokens from the "usage" key of m.
@@ -691,6 +691,19 @@ func logActivityExec(ctx context.Context, exec activityExecutor, broadcaster eve
 		if respStr != nil {
 			payload["response_body"] = json.RawMessage(respJSON)
 		}
+		// internal#211/#212: error_detail carries the runtime's curated,
+		// user-actionable, secret-safe failure reason (provider HTTP
+		// status + error code + the provider's own guidance, e.g. a 403
+		// "org disabled · use an API key / ask your admin"). It is
+		// already persisted to the DB column above and capped by the
+		// runtime's report_activity helper (4096 chars). Previously it
+		// was dropped from the LIVE broadcast, so the canvas had nothing
+		// to render and fell back to a hardcoded opaque
+		// "Agent error (Exception) — see workspace logs" string. Include
+		// it so the chat bubble shows the real reason in real time.
+		if params.ErrorDetail != nil && *params.ErrorDetail != "" {
+			payload["error_detail"] = *params.ErrorDetail
+		}
 	}

 	return func() {
@@ -8,6 +8,7 @@ import (
 	"fmt"
 	"net/http"
 	"net/http/httptest"
+	"sync"
 	"testing"
 	"time"

@@ -22,8 +23,39 @@ import (
 	"github.com/redis/go-redis/v9"
 )

+// liveTestHandlers tracks every WorkspaceHandler built during the test
+// binary's lifetime so setupTestDB can drain their in-flight goAsync
+// goroutines (notably the detached RestartByID restart cycle, which
+// reads the global db.DB) BEFORE restoring db.DB. Without this drain a
+// fire-and-forget restart goroutine spawned by one test outlives that
+// test and races the db.DB swap in a later test's t.Cleanup — the
+// 0x...d548 data race on platform/internal/db.DB.
+var (
+	liveTestHandlersMu sync.Mutex
+	liveTestHandlers   []*WorkspaceHandler
+)
+
 func init() {
 	gin.SetMode(gin.TestMode)
+	newHandlerHook = func(h *WorkspaceHandler) {
+		liveTestHandlersMu.Lock()
+		liveTestHandlers = append(liveTestHandlers, h)
+		liveTestHandlersMu.Unlock()
+	}
+}
+
+// drainTestAsync waits for every tracked handler's goAsync goroutines to
+// finish. Called from setupTestDB's cleanup before db.DB is restored so
+// no detached restart/provision goroutine is mid-read of db.DB when the
+// pointer is swapped.
+func drainTestAsync() {
+	liveTestHandlersMu.Lock()
+	handlers := make([]*WorkspaceHandler, len(liveTestHandlers))
+	copy(handlers, liveTestHandlers)
+	liveTestHandlersMu.Unlock()
+	for _, h := range handlers {
+		h.waitAsyncForTest()
+	}
 }

 // setupTestDB creates a sqlmock DB and assigns it to the global db.DB.
@@ -37,7 +69,16 @@ func setupTestDB(t *testing.T) sqlmock.Sqlmock {
 	}
 	prevDB := db.DB
 	db.DB = mockDB
-	t.Cleanup(func() { db.DB = prevDB; mockDB.Close() })
+	t.Cleanup(func() {
+		// Drain detached async goroutines (e.g. goAsync(RestartByID),
+		// which reads db.DB in runRestartCycle before its provisioner
+		// gate) BEFORE swapping db.DB back. Doing the restore first
+		// would let an in-flight restart goroutine read db.DB while
+		// this line writes it — the data race this guards against.
+		drainTestAsync()
+		db.DB = prevDB
+		mockDB.Close()
+	})

 	// Disable SSRF checks for the duration of this test only. Restore
 	// the previous state via t.Cleanup so that TestIsSafeURL_* tests
@@ -56,9 +56,11 @@ const (
 // (an externally routable address) is used directly.
 func (h *WorkspaceHandler) gracefulPreRestart(ctx context.Context, workspaceID string) {
 	// Non-blocking send — don't stall the restart cycle.
-	// Run in a detached goroutine so the caller (runRestartCycle) can
-	// proceed to stopForRestart without waiting.
-	go func() {
+	// Run in a tracked async goroutine (goAsync, not bare `go`) so the
+	// caller (runRestartCycle) can proceed to stopForRestart without
+	// waiting, while the test harness can still drain it before swapping
+	// the global db.DB (resolveAgentURLForRestartSignal reads db.DB).
+	h.goAsync(func() {
 		signalCtx, cancel := context.WithTimeout(context.Background(), restartSignalTimeout)
 		defer cancel()

@@ -109,7 +111,7 @@ func (h *WorkspaceHandler) gracefulPreRestart(ctx context.Context, workspaceID s
 		} else {
 			log.Printf("A2AGracefulRestart: %s returned status %d — proceeding with stop", workspaceID, resp.StatusCode)
 		}
-	}()
+	})
 }

 // resolveAgentURLForRestartSignal returns the routable URL for the workspace
@@ -0,0 +1,117 @@
+package handlers
+
+// template_files_agent_home_stub_test.go — pins the Phase-1 stub
+// contract for the /agent-home root added by internal#425 RFC.
+//
+// Today (pre-Phase-2b), every Files API verb against `?root=/agent-home`
+// must return HTTP 501 with the canonical pending-message body. The
+// stub MUST NOT:
+//   1. Hit the DB (the workspace might not even exist yet from the
+//      canvas's POV — the root selector is testable without one).
+//   2. Touch the EIC tunnel / Docker / template-dir paths — those
+//      would 500/404/[] depending on the env and confuse the canvas.
+//   3. Accept writes/deletes that the future docker-exec backend
+//      would reject — fail closed.
+//
+// When Phase 2b lands, this file gets replaced by a real
+// docker-exec dispatch test; the stub-message constant in
+// templates.go disappears.
+
+import (
+	"net/http"
+	"net/http/httptest"
+	"strings"
+	"testing"
+
+	"github.com/gin-gonic/gin"
+)
+
+// TestAgentHomeAllowedRoot pins that /agent-home is in the allowedRoots
+// set. Without this, a future refactor that drops the key would
+// silently degrade the canvas root selector to a 400 instead of the
+// stub 501.
+func TestAgentHomeAllowedRoot(t *testing.T) {
+	if !allowedRoots["/agent-home"] {
+		t.Fatal("/agent-home must be in allowedRoots — RFC #425 contract")
+	}
+}
+
+// TestAgentHomeStub_AllVerbs_Return501 pins the canonical stub
+// response across all four verbs. Each must:
+//
+//   - status 501
+//   - body contains the canonical "/agent-home not implemented" prefix
+//   - NOT contain "workspace not found" (proves we short-circuit before
+//     the DB lookup)
+//
+// Driven as a table to keep symmetry — adding a fifth verb in the
+// future means adding one row here.
+func TestAgentHomeStub_AllVerbs_Return501(t *testing.T) {
+	cases := []struct {
+		name   string
+		method string
+		invoke func(c *gin.Context)
+	}{
+		{
+			name:   "ListFiles",
+			method: "GET",
+			invoke: func(c *gin.Context) { (&TemplatesHandler{}).ListFiles(c) },
+		},
+		{
+			name:   "ReadFile",
+			method: "GET",
+			invoke: func(c *gin.Context) { (&TemplatesHandler{}).ReadFile(c) },
+		},
+		{
+			name:   "WriteFile",
+			method: "PUT",
+			invoke: func(c *gin.Context) { (&TemplatesHandler{}).WriteFile(c) },
+		},
+		{
+			name:   "DeleteFile",
+			method: "DELETE",
+			invoke: func(c *gin.Context) { (&TemplatesHandler{}).DeleteFile(c) },
+		},
+	}
+
+	for _, tc := range cases {
+		t.Run(tc.name, func(t *testing.T) {
+			w := httptest.NewRecorder()
+			c, _ := gin.CreateTestContext(w)
+			c.Params = gin.Params{
+				{Key: "id", Value: "ws-stub"},
+				// Path param without leading slash so DeleteFile's
+				// filepath.IsAbs guard doesn't 400 before the root
+				// dispatch runs. The List/Read/Write paths strip the
+				// leading slash themselves and accept either form.
+				{Key: "path", Value: "notes.md"},
+			}
+			// WriteFile binds JSON; provide a minimal valid body so the
+			// short-circuit isn't masked by the bind-error path.
+			var body string
+			if tc.method == "PUT" {
+				body = `{"content":"x"}`
+			}
+			c.Request = httptest.NewRequest(
+				tc.method,
+				"/workspaces/ws-stub/files/notes.md?root=/agent-home",
+				strings.NewReader(body),
+			)
+			if body != "" {
+				c.Request.Header.Set("Content-Type", "application/json")
+			}
+
+			tc.invoke(c)
+
+			if w.Code != http.StatusNotImplemented {
+				t.Fatalf("expected 501, got %d: %s", w.Code, w.Body.String())
+			}
+			if !strings.Contains(w.Body.String(), "/agent-home not implemented") {
+				t.Errorf("body should contain canonical stub message; got %s", w.Body.String())
+			}
+			if strings.Contains(w.Body.String(), "workspace not found") {
+				t.Errorf("stub leaked through to DB lookup; body=%s", w.Body.String())
+			}
+		})
+	}
+}
@@ -18,11 +18,35 @@ import (
 )

 // allowedRoots are the container paths that the Files API can browse.
+//
+// `/agent-home` (added 2026-05-15, internal#425 RFC) is the container's
+// own $HOME — `/root` for openclaw, `/home/agent` for claude-code/hermes
+// — browsed via `docker exec` rather than host-side `find`. The
+// dispatch is stubbed today (returns 501); full implementation lands in
+// Phase 2b of the RFC. The allowedRoots key is added now so the canvas
+// can design its root-selector UI against the final shape and the
+// stub-vs-full transition is server-side only.
 var allowedRoots = map[string]bool{
-	"/configs":   true,
-	"/workspace": true,
-	"/home":      true,
-	"/plugins":   true,
+	"/configs":    true,
+	"/workspace":  true,
+	"/home":       true,
+	"/plugins":    true,
+	"/agent-home": true,
+}
+
+// agentHomeStubMessage is the body returned by every Files API verb
+// when `?root=/agent-home` is requested before Phase 2b lands. Keep the
+// status code 501 (Not Implemented) — the route exists, the verb is
+// understood, but the handler is unimplemented. Distinguishes from
+// 400/404 so a canvas behind a less-current server can render a clean
+// "feature pending" state instead of a generic error.
+const agentHomeStubMessage = "/agent-home not implemented yet (internal#425 RFC Phase 2b — docker-exec backend pending)"
+
+// isAgentHomeStubRequest returns true when the request targets the
+// stubbed /agent-home root. Centralised so every verb in this file
+// short-circuits with the same response shape.
+func isAgentHomeStubRequest(rootPath string) bool {
+	return rootPath == "/agent-home"
 }

 // maxUploadFiles limits the number of files in a single import/replace.
@@ -219,7 +243,14 @@ func (h *TemplatesHandler) ListFiles(c *gin.Context) {
 	//   ?depth= — max depth to recurse (default: 1, max: 5)
 	rootPath := c.DefaultQuery("root", "/configs")
 	if !allowedRoots[rootPath] {
-		c.JSON(http.StatusBadRequest, gin.H{"error": "root must be one of: /configs, /workspace, /home, /plugins"})
+		c.JSON(http.StatusBadRequest, gin.H{"error": "root must be one of: /configs, /workspace, /home, /plugins, /agent-home"})
+		return
+	}
+	// /agent-home dispatch is stubbed pre-Phase-2b. Short-circuit before
+	// the DB lookup + EIC dance so a canvas exercising the new root key
+	// gets a clean 501 instead of a half-effort response.
+	if isAgentHomeStubRequest(rootPath) {
+		c.JSON(http.StatusNotImplemented, gin.H{"error": agentHomeStubMessage})
 		return
 	}
 	subPath := c.DefaultQuery("path", "")
@@ -383,7 +414,11 @@ func (h *TemplatesHandler) ReadFile(c *gin.Context) {
 	ctx := c.Request.Context()
 	rootPath := c.DefaultQuery("root", "/configs")
 	if !allowedRoots[rootPath] {
-		c.JSON(http.StatusBadRequest, gin.H{"error": "root must be one of: /configs, /workspace, /home, /plugins"})
+		c.JSON(http.StatusBadRequest, gin.H{"error": "root must be one of: /configs, /workspace, /home, /plugins, /agent-home"})
+		return
+	}
+	if isAgentHomeStubRequest(rootPath) {
+		c.JSON(http.StatusNotImplemented, gin.H{"error": agentHomeStubMessage})
 		return
 	}

@@ -496,7 +531,11 @@ func (h *TemplatesHandler) WriteFile(c *gin.Context) {
 	ctx := c.Request.Context()
 	rootPath := c.DefaultQuery("root", "/configs")
 	if !allowedRoots[rootPath] {
-		c.JSON(http.StatusBadRequest, gin.H{"error": "root must be one of: /configs, /workspace, /home, /plugins"})
+		c.JSON(http.StatusBadRequest, gin.H{"error": "root must be one of: /configs, /workspace, /home, /plugins, /agent-home"})
+		return
+	}
+	if isAgentHomeStubRequest(rootPath) {
+		c.JSON(http.StatusNotImplemented, gin.H{"error": agentHomeStubMessage})
 		return
 	}
 	var wsName, instanceID, runtime string
@@ -573,7 +612,11 @@ func (h *TemplatesHandler) DeleteFile(c *gin.Context) {
 	ctx := c.Request.Context()
 	rootPath := c.DefaultQuery("root", "/configs")
 	if !allowedRoots[rootPath] {
-		c.JSON(http.StatusBadRequest, gin.H{"error": "root must be one of: /configs, /workspace, /home, /plugins"})
+		c.JSON(http.StatusBadRequest, gin.H{"error": "root must be one of: /configs, /workspace, /home, /plugins, /agent-home"})
+		return
+	}
+	if isAgentHomeStubRequest(rootPath) {
+		c.JSON(http.StatusNotImplemented, gin.H{"error": agentHomeStubMessage})
 		return
 	}
 	var wsName, instanceID, runtime string
@@ -10,8 +10,20 @@ import (
 	"github.com/Molecule-AI/molecule-monorepo/platform/internal/db"
 	"github.com/Molecule-AI/molecule-monorepo/platform/internal/wsauth"
 	"github.com/gin-gonic/gin"
+	"github.com/google/uuid"
 )

+// validWorkspaceID returns true when id is a syntactically valid UUID.
+// workspace_id is a `uuid` column; passing a non-UUID (e.g. the canvas
+// "global" sentinel sent when no node is selected) makes Postgres raise
+// `invalid input syntax for type uuid`, which previously leaked as an
+// opaque 500. Reject up front with a clean 400 instead. Mirrors the
+// uuid.Parse guard already used in handlers/activity.go.
+func validWorkspaceID(id string) bool {
+	_, err := uuid.Parse(id)
+	return err == nil
+}
+
 // TokenHandler exposes user-facing token management for workspaces.
 // Routes: GET/POST/DELETE /workspaces/:id/tokens (behind WorkspaceAuth).
 type TokenHandler struct{}
@@ -31,6 +43,10 @@ type tokenListItem struct {
 // never the plaintext or hash).
 func (h *TokenHandler) List(c *gin.Context) {
 	workspaceID := c.Param("id")
+	if !validWorkspaceID(workspaceID) {
+		c.JSON(http.StatusBadRequest, gin.H{"error": "invalid workspace id"})
+		return
+	}

 	limit := 50
 	if v := c.Query("limit"); v != "" {
@@ -53,6 +69,7 @@ func (h *TokenHandler) List(c *gin.Context) {
 		LIMIT $2 OFFSET $3
 	`, workspaceID, limit, offset)
 	if err != nil {
+		log.Printf("tokens: list query failed for workspace %s: %v", workspaceID, err)
 		c.JSON(http.StatusInternalServerError, gin.H{"error": "failed to list tokens"})
 		return
 	}
@@ -85,6 +102,10 @@ const maxTokensPerWorkspace = 50
 // exactly once in the response — it cannot be recovered afterwards.
 func (h *TokenHandler) Create(c *gin.Context) {
 	workspaceID := c.Param("id")
+	if !validWorkspaceID(workspaceID) {
+		c.JSON(http.StatusBadRequest, gin.H{"error": "invalid workspace id"})
+		return
+	}

 	// Rate limit: max active tokens per workspace
 	var count int
@@ -117,6 +138,10 @@ func (h *TokenHandler) Create(c *gin.Context) {
 func (h *TokenHandler) Revoke(c *gin.Context) {
 	workspaceID := c.Param("id")
 	tokenID := c.Param("tokenId")
+	if !validWorkspaceID(workspaceID) {
+		c.JSON(http.StatusBadRequest, gin.H{"error": "invalid workspace id"})
+		return
+	}

 	result, err := db.DB.ExecContext(c.Request.Context(), `
 		UPDATE workspace_auth_tokens
@@ -41,6 +41,15 @@ import (

 func init() { gin.SetMode(gin.TestMode) }

+// Workspace IDs are validated as UUIDs up front (tokens.go validWorkspaceID),
+// so handler tests must pass syntactically valid UUIDs. Fixed values keep
+// sqlmock WithArgs assertions deterministic.
+const (
+	wsUUID1 = "11111111-1111-1111-1111-111111111111"
+	wsUUID2 = "22222222-2222-2222-2222-222222222222"
+	wsUUID3 = "33333333-3333-3333-3333-333333333333"
+)
+
 // withMockDB swaps `db.DB` for a sqlmock and returns the mock plus a
 // restore func. Tests use this in place of setupTokenTestDB which
 // skips on a missing real DB.
@@ -81,13 +90,13 @@ func TestTokenHandler_List_HappyPath(t *testing.T) {
 	created := time.Date(2026, 4, 1, 12, 0, 0, 0, time.UTC)
 	last := created.Add(time.Hour)
 	mock.ExpectQuery(`SELECT id, prefix, created_at, last_used_at\s+FROM workspace_auth_tokens`).
-		WithArgs("ws-1", 50, 0).
+		WithArgs(wsUUID1, 50, 0).
 		WillReturnRows(sqlmock.NewRows([]string{"id", "prefix", "created_at", "last_used_at"}).
 			AddRow("tok-1", "abc12345", created, last).
 			AddRow("tok-2", "def67890", created, nil))

 	w := makeReq(t, NewTokenHandler().List, "GET",
-		"/workspaces/ws-1/tokens", gin.Params{{Key: "id", Value: "ws-1"}})
+		"/workspaces/ws-1/tokens", gin.Params{{Key: "id", Value: wsUUID1}})

 	if w.Code != http.StatusOK {
 		t.Fatalf("expected 200, got %d: %s", w.Code, w.Body.String())
@@ -121,7 +130,7 @@ func TestTokenHandler_List_EmptyResult(t *testing.T) {
 		WillReturnRows(sqlmock.NewRows([]string{"id", "prefix", "created_at", "last_used_at"}))

 	w := makeReq(t, NewTokenHandler().List, "GET",
-		"/workspaces/ws-2/tokens", gin.Params{{Key: "id", Value: "ws-2"}})
+		"/workspaces/ws-2/tokens", gin.Params{{Key: "id", Value: wsUUID2}})

 	if w.Code != http.StatusOK {
 		t.Fatalf("expected 200 on empty list, got %d", w.Code)
@@ -146,7 +155,7 @@ func TestTokenHandler_List_QueryError(t *testing.T) {
 		WillReturnError(errors.New("connection refused"))

 	w := makeReq(t, NewTokenHandler().List, "GET",
-		"/workspaces/ws-3/tokens", gin.Params{{Key: "id", Value: "ws-3"}})
+		"/workspaces/ws-3/tokens", gin.Params{{Key: "id", Value: wsUUID3}})

 	if w.Code != http.StatusInternalServerError {
 		t.Errorf("query error must surface as 500, got %d", w.Code)
@@ -158,13 +167,13 @@ func TestTokenHandler_List_RespectsLimit(t *testing.T) {
 	defer cleanup()

 	mock.ExpectQuery(`SELECT id, prefix, created_at, last_used_at`).
-		WithArgs("ws-1", 10, 5).
+		WithArgs(wsUUID1, 10, 5).
 		WillReturnRows(sqlmock.NewRows([]string{"id", "prefix", "created_at", "last_used_at"}))

 	w := httptest.NewRecorder()
 	c, _ := gin.CreateTestContext(w)
 	c.Request = httptest.NewRequest("GET", "/workspaces/ws-1/tokens?limit=10&offset=5", nil)
-	c.Params = gin.Params{{Key: "id", Value: "ws-1"}}
+	c.Params = gin.Params{{Key: "id", Value: wsUUID1}}
 	NewTokenHandler().List(c)

 	if w.Code != http.StatusOK {
@@ -186,7 +195,7 @@ func TestTokenHandler_List_ScanError(t *testing.T) {
 			AddRow("tok-1", "abc", "not-a-timestamp", nil))

 	w := makeReq(t, NewTokenHandler().List, "GET",
-		"/workspaces/ws-1/tokens", gin.Params{{Key: "id", Value: "ws-1"}})
+		"/workspaces/ws-1/tokens", gin.Params{{Key: "id", Value: wsUUID1}})

 	if w.Code != http.StatusInternalServerError {
 		t.Errorf("scan error must surface as 500, got %d: %s", w.Code, w.Body.String())
@@ -201,11 +210,11 @@ func TestTokenHandler_Create_RateLimited(t *testing.T) {

 	// Count query returns 50 (== max) → 429.
 	mock.ExpectQuery(`SELECT COUNT\(\*\) FROM workspace_auth_tokens`).
-		WithArgs("ws-1").
+		WithArgs(wsUUID1).
 		WillReturnRows(sqlmock.NewRows([]string{"count"}).AddRow(50))

 	w := makeReq(t, NewTokenHandler().Create, "POST",
-		"/workspaces/ws-1/tokens", gin.Params{{Key: "id", Value: "ws-1"}})
+		"/workspaces/ws-1/tokens", gin.Params{{Key: "id", Value: wsUUID1}})

 	if w.Code != http.StatusTooManyRequests {
 		t.Errorf("max active tokens should 429, got %d", w.Code)
@@ -225,7 +234,7 @@ func TestTokenHandler_Create_IssueFails(t *testing.T) {
 		WillReturnError(errors.New("disk full"))

 	w := makeReq(t, NewTokenHandler().Create, "POST",
-		"/workspaces/ws-1/tokens", gin.Params{{Key: "id", Value: "ws-1"}})
+		"/workspaces/ws-1/tokens", gin.Params{{Key: "id", Value: wsUUID1}})

 	if w.Code != http.StatusInternalServerError {
 		t.Errorf("IssueToken DB error must 500, got %d", w.Code)
@@ -242,7 +251,7 @@ func TestTokenHandler_Create_HappyPath(t *testing.T) {
 		WillReturnResult(sqlmock.NewResult(1, 1))

 	w := makeReq(t, NewTokenHandler().Create, "POST",
-		"/workspaces/ws-1/tokens", gin.Params{{Key: "id", Value: "ws-1"}})
+		"/workspaces/ws-1/tokens", gin.Params{{Key: "id", Value: wsUUID1}})

 	if w.Code != http.StatusCreated {
 		t.Fatalf("expected 201, got %d: %s", w.Code, w.Body.String())
@@ -257,7 +266,7 @@ func TestTokenHandler_Create_HappyPath(t *testing.T) {
 	if body.AuthToken == "" {
 		t.Errorf("auth_token must be present and non-empty in response")
 	}
-	if body.WorkspaceID != "ws-1" {
+	if body.WorkspaceID != wsUUID1 {
 		t.Errorf("workspace_id mismatch: %q", body.WorkspaceID)
 	}
 }
@@ -269,12 +278,12 @@ func TestTokenHandler_Revoke_HappyPath(t *testing.T) {
 	defer cleanup()

 	mock.ExpectExec(`UPDATE workspace_auth_tokens\s+SET revoked_at = now\(\)`).
-		WithArgs("tok-1", "ws-1").
+		WithArgs("tok-1", wsUUID1).
 		WillReturnResult(sqlmock.NewResult(0, 1))

 	w := makeReq(t, NewTokenHandler().Revoke, "DELETE",
 		"/workspaces/ws-1/tokens/tok-1", gin.Params{
-			{Key: "id", Value: "ws-1"},
+			{Key: "id", Value: wsUUID1},
 			{Key: "tokenId", Value: "tok-1"},
 		})

@@ -289,12 +298,12 @@ func TestTokenHandler_Revoke_NotFound(t *testing.T) {

 	// 0 rows affected → token not found OR already revoked.
 	mock.ExpectExec(`UPDATE workspace_auth_tokens`).
-		WithArgs("tok-ghost", "ws-1").
+		WithArgs("tok-ghost", wsUUID1).
 		WillReturnResult(sqlmock.NewResult(0, 0))

 	w := makeReq(t, NewTokenHandler().Revoke, "DELETE",
 		"/workspaces/ws-1/tokens/tok-ghost", gin.Params{
-			{Key: "id", Value: "ws-1"},
+			{Key: "id", Value: wsUUID1},
 			{Key: "tokenId", Value: "tok-ghost"},
 		})

@@ -312,7 +321,7 @@ func TestTokenHandler_Revoke_DBError(t *testing.T) {

 	w := makeReq(t, NewTokenHandler().Revoke, "DELETE",
 		"/workspaces/ws-1/tokens/tok-1", gin.Params{
-			{Key: "id", Value: "ws-1"},
+			{Key: "id", Value: wsUUID1},
 			{Key: "tokenId", Value: "tok-1"},
 		})

@@ -321,6 +330,59 @@ func TestTokenHandler_Revoke_DBError(t *testing.T) {
 	}
 }

+// ---- UUID validation (regression: "global" sentinel 500) ------------
+
+// The canvas Settings → Workspace Tokens tab sent the literal sentinel
+// "global" as the workspace id when no node was selected. workspace_id
+// is a `uuid` column, so the query raised
+// `invalid input syntax for type uuid: "global"` which leaked as an
+// opaque 500. List/Create/Revoke now reject any non-UUID id with a
+// clean 400 before touching the DB. No DB expectation is set on the
+// mock — a DB hit would fail ExpectationsWereMet, proving short-circuit.
+func TestTokenHandler_RejectsNonUUIDWorkspaceID(t *testing.T) {
+	h := NewTokenHandler()
+	cases := []struct {
+		name   string
+		run    func(c *gin.Context)
+		method string
+		params gin.Params
+	}{
+		{"List", h.List, "GET", gin.Params{{Key: "id", Value: "global"}}},
+		{"Create", h.Create, "POST", gin.Params{{Key: "id", Value: "global"}}},
+		{"Revoke", h.Revoke, "DELETE", gin.Params{
+			{Key: "id", Value: "global"},
+			{Key: "tokenId", Value: "tok-1"},
+		}},
+	}
+	for _, tc := range cases {
+		t.Run(tc.name, func(t *testing.T) {
+			mock, cleanup := withMockDB(t)
+			defer cleanup()
+
+			w := makeReq(t, tc.run, tc.method,
+				"/workspaces/global/tokens", tc.params)
+
+			if w.Code != http.StatusBadRequest {
+				t.Fatalf("%s with non-UUID id must 400, got %d: %s",
+					tc.name, w.Code, w.Body.String())
+			}
+			var body struct {
+				Error string `json:"error"`
+			}
+			_ = json.Unmarshal(w.Body.Bytes(), &body)
+			if body.Error != "invalid workspace id" {
+				t.Errorf("%s: want error=%q, got %q",
+					tc.name, "invalid workspace id", body.Error)
+			}
+			// No query/exec was expected → if the handler hit the DB
+			// this fails, proving the guard short-circuits before SQL.
+			if err := mock.ExpectationsWereMet(); err != nil {
+				t.Errorf("%s leaked a DB call past the uuid guard: %v", tc.name, err)
+			}
+		})
+	}
+}
+
 // Compile-time noise removal: the imports list pulls in the sql /
 // driver packages and the silenced ctx so a future scenario that
 // needs them doesn't have to re-add the import. Documented here so
@@ -11,6 +11,7 @@ import (
 	"github.com/Molecule-AI/molecule-monorepo/platform/internal/db"
 	"github.com/Molecule-AI/molecule-monorepo/platform/internal/wsauth"
 	"github.com/gin-gonic/gin"
+	"github.com/google/uuid"
 )

 func init() { gin.SetMode(gin.TestMode) }
@@ -167,11 +168,14 @@ func TestTokenHandler_RevokeWrongWorkspace(t *testing.T) {

 	h := NewTokenHandler()

-	// Try to revoke with a different workspace ID — should 404
+	// Try to revoke with a different (valid-UUID) workspace ID that does
+	// not own the token — should 404. A valid UUID is required so this
+	// exercises the ownership branch, not the up-front uuid-shape 400.
+	otherWS := uuid.NewString()
 	w := httptest.NewRecorder()
 	c, _ := gin.CreateTestContext(w)
-	c.Params = gin.Params{{Key: "id", Value: "wrong-workspace-id"}, {Key: "tokenId", Value: tokenID}}
-	c.Request = httptest.NewRequest("DELETE", "/workspaces/wrong/tokens/"+tokenID, nil)
+	c.Params = gin.Params{{Key: "id", Value: otherWS}, {Key: "tokenId", Value: tokenID}}
+	c.Request = httptest.NewRequest("DELETE", "/workspaces/"+otherWS+"/tokens/"+tokenID, nil)
 	h.Revoke(c)

 	if w.Code != http.StatusNotFound {
@@ -80,6 +80,15 @@ type WorkspaceHandler struct {
 	asyncWG sync.WaitGroup
 }

+// newHandlerHook, when non-nil, is invoked for every WorkspaceHandler
+// created via NewWorkspaceHandler. It is nil in production (zero cost);
+// the test harness sets it so setupTestDB can drain every handler's
+// in-flight async goroutines before swapping the global db.DB. Without
+// this, a detached restart goroutine (maybeMarkContainerDead ->
+// goAsync(RestartByID) -> runRestartCycle reads db.DB) races the
+// db.DB restore in another test's t.Cleanup.
+var newHandlerHook func(*WorkspaceHandler)
+
 func (h *WorkspaceHandler) goAsync(fn func()) {
 	h.asyncWG.Add(1)
 	go func() {
@@ -108,6 +117,9 @@ func NewWorkspaceHandler(b events.EventEmitter, p *provisioner.Provisioner, plat
 	if p != nil {
 		h.provisioner = p
 	}
+	if newHandlerHook != nil {
+		newHandlerHook(h)
+	}
 	return h
 }

@@ -237,10 +237,10 @@ func (h *WorkspaceHandler) Restart(c *gin.Context) {
 	// the silent-drop bugs PRs #2811/#2824 closed). RestartWorkspaceAuto
 	// enforces CP-FIRST ordering matching the other dispatchers — see
 	// docs/architecture/backends.md.
-	go func() {
+	h.goAsync(func() {
 		h.RestartWorkspaceAutoOpts(context.Background(), id, templatePath, configFiles, payload, resetClaudeSession)
-	}()
-	go h.sendRestartContext(id, restartData)
+	})
+	h.goAsync(func() { h.sendRestartContext(id, restartData) })

 	c.JSON(http.StatusOK, gin.H{"status": "provisioning", "config_dir": configLabel, "reset_session": resetClaudeSession})
 }
@@ -610,7 +610,9 @@ func (h *WorkspaceHandler) runRestartCycle(workspaceID string) {
 	h.provisionWorkspaceAutoSync(workspaceID, "", nil, payload)
 	// sendRestartContext is a one-way notification to the new container; safe
 	// to fire async — the next restart cycle won't depend on it completing.
-	go h.sendRestartContext(workspaceID, restartData)
+	// Tracked via goAsync so the test harness can drain it before the
+	// global db.DB swap (sendRestartContext reads db.DB).
+	h.goAsync(func() { h.sendRestartContext(workspaceID, restartData) })
 }

 // Pause handles POST /workspaces/:id/pause
@@ -178,12 +178,21 @@ func (p *CPProvisioner) Start(ctx context.Context, cfg WorkspaceConfig) (string,
 	// /admin/liveness and other admin-gated platform endpoints (core#831).
 	// p.adminToken is read from os.Getenv("ADMIN_TOKEN") at provisioner creation;
 	// it is also used for CP→platform HTTP auth but those are separate concerns.
-	env := cfg.EnvVars
-	if p.adminToken != "" {
-		env = make(map[string]string, len(cfg.EnvVars)+1)
-		for k, v := range cfg.EnvVars {
-			env[k] = v
+	//
+	// Forensic #145 hardening: tenant workspaces run on EC2 via this path, so
+	// the SCM-write-token denylist (see buildContainerEnv) is enforced here
+	// too. Always build a filtered copy — never pass cfg.EnvVars through
+	// verbatim — so a latent persona-merged GITEA_TOKEN can't reach the
+	// tenant container regardless of whether ADMIN_TOKEN is set.
+	env := make(map[string]string, len(cfg.EnvVars)+1)
+	for k, v := range cfg.EnvVars {
+		if isSCMWriteTokenKey(k) {
+			log.Printf("CPProvisioner.Start: dropped SCM-write credential %q from tenant workspace env (forensic #145 guard)", k)
+			continue
 		}
+		env[k] = v
+	}
+	if p.adminToken != "" {
 		env["ADMIN_TOKEN"] = p.adminToken
 	}
 	// Collect template files and generated configs, with OFFSEC-010 guards:
@@ -343,6 +352,7 @@ func collectCPConfigFiles(cfg WorkspaceConfig) (map[string]string, error) {
 	}
 	return files, nil
 }
+
 // Stop terminates the workspace's EC2 instance via the control plane.
 //
 // Looks up the actual EC2 instance_id from the workspaces table before
@@ -497,7 +507,9 @@ func (p *CPProvisioner) IsRunning(ctx context.Context, workspaceID string) (bool
 		// Don't leak the body — upstream errors may echo headers.
 		return true, fmt.Errorf("cp provisioner: status: unexpected %d", resp.StatusCode)
 	}
-	var result struct{ State string `json:"state"` }
+	var result struct {
+		State string `json:"state"`
+	}
 	// Cap body read at 64 KiB for parity with Start — a misconfigured
 	// or compromised CP streaming a huge body could otherwise exhaust
 	// memory in this hot path (called reactively per-request from
@@ -591,6 +591,28 @@ func ValidateWorkspaceAccess(access, workspacePath string) error {
 	}
 }

+// scmWriteTokenKeys is the explicit denylist of environment variable names
+// that carry a Git SCM *write* credential (push / merge / approve). These
+// must never reach a tenant workspace container — see the forensic #145
+// rationale in buildContainerEnv. Kept as an exact-match set rather than a
+// substring/prefix heuristic so the guard is auditable and can't silently
+// over-strip a legitimately-named var.
+var scmWriteTokenKeys = map[string]struct{}{
+	"GITEA_TOKEN":     {},
+	"GITHUB_TOKEN":    {},
+	"GH_TOKEN":        {}, // gh CLI honours GH_TOKEN as a GITHUB_TOKEN alias
+	"GITLAB_TOKEN":    {},
+	"GL_TOKEN":        {}, // glab CLI alias
+	"BITBUCKET_TOKEN": {},
+}
+
+// isSCMWriteTokenKey reports whether an env var name is a known Git SCM
+// write credential that must be stripped from tenant workspace env.
+func isSCMWriteTokenKey(key string) bool {
+	_, ok := scmWriteTokenKeys[key]
+	return ok
+}
+
 // buildContainerEnv assembles the initial environment variables injected
 // into every workspace container.
 //
@@ -627,6 +649,21 @@ func buildContainerEnv(cfg WorkspaceConfig) []string {
 		env = append(env, fmt.Sprintf("AWARENESS_URL=%s", cfg.AwarenessURL))
 	}
 	for k, v := range cfg.EnvVars {
+		// Forensic #145 hardening: tenant workspace containers run
+		// agent-controlled code and must NEVER receive a Git SCM *write*
+		// credential. Without merge/approve creds in-container the
+		// two-eyes review gate is structurally self-bypass-proof — an
+		// agent that forges an approval has no token to act on it. A
+		// latent path exists (loadPersonaEnvFile merges a per-role
+		// persona `GITEA_TOKEN` into cfg.EnvVars when MOLECULE_PERSONA_ROOT
+		// is set on a tenant host); it is inert today (persona dirs are
+		// operator-host-only) but unguarded. Strip SCM-write tokens here
+		// by construction so the invariant holds regardless of whether
+		// that path ever becomes reachable.
+		if isSCMWriteTokenKey(k) {
+			log.Printf("buildContainerEnv: dropped SCM-write credential %q from workspace env (forensic #145 guard)", k)
+			continue
+		}
 		env = append(env, fmt.Sprintf("%s=%s", k, v))
 	}
 	// Inject ADMIN_TOKEN from the platform server's environment so workspace
@@ -636,10 +636,15 @@ func TestBuildContainerEnv_AwarenessOnlyWhenBothSet(t *testing.T) {
 }

 func TestBuildContainerEnv_CustomEnvVarsAppended(t *testing.T) {
+	// NOTE: this test previously asserted GITHUB_TOKEN passed through
+	// verbatim. That assertion encoded the forensic #145 latent leak as
+	// expected behavior. Post-guard, ordinary custom env still flows but
+	// SCM-write credentials are stripped — see
+	// TestBuildContainerEnv_StripsSCMWriteTokens for the negative assertion.
 	cfg := WorkspaceConfig{
 		WorkspaceID: "ws-x",
 		PlatformURL: "http://localhost:8080",
-		EnvVars:     map[string]string{"CUSTOM": "value", "GITHUB_TOKEN": "fake-token-for-test"},
+		EnvVars:     map[string]string{"CUSTOM": "value", "ANTHROPIC_API_KEY": "sk-not-an-scm-token"},
 	}
 	env := buildContainerEnv(cfg)
 	seen := map[string]string{}
@@ -652,8 +657,8 @@ func TestBuildContainerEnv_CustomEnvVarsAppended(t *testing.T) {
 	if seen["CUSTOM"] != "value" {
 		t.Errorf("CUSTOM env missing, got env=%v", env)
 	}
-	if seen["GITHUB_TOKEN"] != "fake-token-for-test" {
-		t.Errorf("GITHUB_TOKEN env missing, got env=%v", env)
+	if seen["ANTHROPIC_API_KEY"] != "sk-not-an-scm-token" {
+		t.Errorf("non-SCM custom env must still pass through, got env=%v", env)
 	}
 	// Built-in defaults still present
 	if seen["MOLECULE_URL"] == "" {
@@ -661,6 +666,129 @@ func TestBuildContainerEnv_CustomEnvVarsAppended(t *testing.T) {
 	}
 }

+// ---------- forensic #145: SCM-write-token denylist guard ----------
+
+// TestBuildContainerEnv_StripsSCMWriteTokens is the core negative
+// assertion: a tenant workspace env constructed via buildContainerEnv MUST
+// NOT contain any Git SCM *write* credential, regardless of how it got into
+// cfg.EnvVars. This proves the two-eyes review gate stays structurally
+// self-bypass-proof — an agent in-container has no merge/approve token to
+// act on a forged approval. See forensic #145.
+//
+// This test FAILS on the pre-guard code (where buildContainerEnv passed
+// cfg.EnvVars through verbatim) and PASSES once the denylist filter is in
+// place — i.e. the guard is proven by construction, not by environment
+// accident.
+func TestBuildContainerEnv_StripsSCMWriteTokens(t *testing.T) {
+	scmTokens := []string{
+		"GITEA_TOKEN", "GITHUB_TOKEN", "GH_TOKEN",
+		"GITLAB_TOKEN", "GL_TOKEN", "BITBUCKET_TOKEN",
+	}
+
+	t.Run("normal path — SCM tokens explicitly set in EnvVars", func(t *testing.T) {
+		envVars := map[string]string{"CUSTOM": "ok", "ANTHROPIC_API_KEY": "sk-keep"}
+		for _, k := range scmTokens {
+			envVars[k] = "leaked-write-credential-" + k
+		}
+		cfg := WorkspaceConfig{
+			WorkspaceID: "ws-tenant",
+			PlatformURL: "http://localhost:8080",
+			Tier:        2,
+			EnvVars:     envVars,
+		}
+		assertNoSCMWriteToken(t, buildContainerEnv(cfg), scmTokens)
+
+		// Sanity: non-SCM custom env is NOT collateral-damaged by the filter.
+		if !envContains(buildContainerEnv(cfg), "CUSTOM=ok") {
+			t.Errorf("filter must not strip non-SCM custom env")
+		}
+		if !envContains(buildContainerEnv(cfg), "ANTHROPIC_API_KEY=sk-keep") {
+			t.Errorf("filter must not strip non-SCM API keys")
+		}
+	})
+
+	t.Run("persona-file path — simulates loadPersonaEnvFile merge", func(t *testing.T) {
+		// The latent path: handlers.loadPersonaEnvFile() merges a per-role
+		// persona env file (carrying GITEA_USER, GITEA_TOKEN, …) into the
+		// workspace env map when MOLECULE_PERSONA_ROOT is set on a tenant
+		// host. We can't invoke that cross-package helper here, but its
+		// observable effect is exactly "a GITEA_TOKEN appears in
+		// cfg.EnvVars". Constructing that condition directly proves the
+		// guard holds even if the latent path becomes reachable.
+		cfg := WorkspaceConfig{
+			WorkspaceID: "ws-tenant",
+			PlatformURL: "http://localhost:8080",
+			Tier:        2,
+			EnvVars: map[string]string{
+				// Persona identity fields that are SAFE to keep (read-only
+				// identity, not a write credential):
+				"GITEA_USER":       "backend-engineer",
+				"GITEA_USER_EMAIL": "backend-engineer@agents.moleculesai.app",
+				// The credential that must be stripped:
+				"GITEA_TOKEN":        "persona-merged-write-pat",
+				"GITEA_TOKEN_SCOPES": "write:repository",
+			},
+		}
+		got := buildContainerEnv(cfg)
+		assertNoSCMWriteToken(t, got, scmTokens)
+		// Non-credential persona identity may still flow through — only the
+		// write token is the denied surface.
+		if !envContains(got, "GITEA_USER=backend-engineer") {
+			t.Errorf("non-credential persona identity (GITEA_USER) should not be stripped")
+		}
+	})
+}
+
+// TestCPProvisionerEnv_StripsSCMWriteTokens covers the tenant-EC2 path:
+// CPProvisioner.Start builds the env map the control plane forwards to the
+// EC2 workspace container. The same forensic #145 denylist must hold there.
+func TestCPProvisionerEnv_StripsSCMWriteTokens(t *testing.T) {
+	// isSCMWriteTokenKey is the single source of truth shared by both
+	// buildContainerEnv (local Docker) and CPProvisioner.Start (tenant EC2).
+	// Assert it classifies every known SCM-write var as denied and leaves
+	// ordinary / read-only-identity vars alone.
+	for _, k := range []string{
+		"GITEA_TOKEN", "GITHUB_TOKEN", "GH_TOKEN",
+		"GITLAB_TOKEN", "GL_TOKEN", "BITBUCKET_TOKEN",
+	} {
+		if !isSCMWriteTokenKey(k) {
+			t.Errorf("isSCMWriteTokenKey(%q) = false, want true (SCM-write credential must be denied)", k)
+		}
+	}
+	for _, k := range []string{
+		"GITEA_USER", "GITEA_USER_EMAIL", "ANTHROPIC_API_KEY",
+		"CUSTOM", "PLATFORM_URL", "ADMIN_TOKEN", "",
+	} {
+		if isSCMWriteTokenKey(k) {
+			t.Errorf("isSCMWriteTokenKey(%q) = true, want false (must not over-strip non-SCM env)", k)
+		}
+	}
+}
+
+func assertNoSCMWriteToken(t *testing.T, env []string, scmTokens []string) {
+	t.Helper()
+	for _, e := range env {
+		key := e
+		if i := strings.IndexByte(e, '='); i >= 0 {
+			key = e[:i]
+		}
+		for _, banned := range scmTokens {
+			if key == banned {
+				t.Errorf("SCM-write credential %q leaked into workspace env (forensic #145 invariant violated): %q", banned, e)
+			}
+		}
+	}
+}
+
+func envContains(env []string, want string) bool {
+	for _, e := range env {
+		if e == want {
+			return true
+		}
+	}
+	return false
+}
+
 // ---------- buildWorkspaceMount — #65 workspace_access ----------

 func TestBuildWorkspaceMount_SelectionMatrix(t *testing.T) {
@@ -0,0 +1,226 @@
+// Package secrets provides the canonical SSOT for credential-shaped
+// regex patterns used by:
+//
+//   - the CI `Secret scan` workflow (.gitea/workflows/secret-scan.yml)
+//   - the runtime's bundled pre-commit hook
+//     (molecule-ai-workspace-runtime/molecule_runtime/scripts/pre-commit-checks.sh)
+//   - the upcoming Phase 2b docker-exec Files API backend, which has
+//     to refuse to surface files whose path OR content matches a
+//     credential shape (RFC internal#425, Hongming 2026-05-15)
+//
+// Before this package, the same regex set lived as duplicate bash
+// arrays in two unrelated repos; adding a pattern required editing
+// both, and pattern drift was caught only via secret-scan workflow
+// failures on PRs that had unrelated changes (#2090-class incident
+// vector). Centralising in Go makes the Files API the SSOT, with the
+// YAML + bash arrays generated/asserted from this package so drift
+// is detected at CI time, not at exfiltration time.
+//
+// This file is Phase 2a of the internal#425 RFC. Phase 2b will import
+// `Patterns` from `template_files_docker_exec.go` to gate
+// `listFilesViaDockerExec` / `readFileViaDockerExec` against
+// secret-shaped paths AND content. Until 2b lands, the package has
+// one consumer: this package's own unit tests, which pin the regex
+// strings so a refactor that drops or weakens one is caught here.
+package secrets
+
+import (
+	"fmt"
+	"regexp"
+	"sync"
+)
+
+// Pattern is one named credential shape — a human label plus the
+// compiled regex. The label appears in CI error output ("matched:
+// github-pat") so an operator can identify the family without seeing
+// the actual matched bytes (echoing the bytes widens the blast radius
+// per the secret-scan workflow's recovery prose).
+type Pattern struct {
+	// Name is a short kebab-case identifier (e.g. "github-pat",
+	// "anthropic-api-key"). Stable across versions — consumers may
+	// switch on it.
+	Name string
+	// Description is a one-line human-readable explanation of what
+	// the pattern matches. Used in CI error messages and the Files
+	// API "<denied: secret-shape>" placeholder rationale.
+	Description string
+	// regexSource is the regex literal in Go-RE2 syntax. Stored as a
+	// string so the slice declaration below stays readable; compiled
+	// once via sync.Once into a *regexp.Regexp.
+	regexSource string
+}
+
+// Patterns is the canonical credential-shape regex set.
+//
+// Adding a pattern here:
+//
+//  1. Add a new Pattern{} entry below with a kebab-case Name, a
+//     one-line Description, and the regex literal. Anchor on a
+//     low-false-positive prefix.
+//  2. Add a positive + negative test case in patterns_test.go.
+//  3. Mirror the regex string into:
+//     a. .gitea/workflows/secret-scan.yml SECRET_PATTERNS array
+//     b. molecule-ai-workspace-runtime/molecule_runtime/scripts/pre-commit-checks.sh
+//     (or wait for the codegen target that consumes this slice — TBD
+//     follow-up; tracked in the Phase 2a PR description.)
+//
+// The order is: alphabetical within each provider family, families
+// grouped by ecosystem (GitHub family, AI-provider family, chat
+// family, cloud family). Keep this stable so diffs are reviewable.
+var Patterns = []Pattern{
+	// --- GitHub token family ---
+	{
+		Name:        "github-pat-classic",
+		Description: "GitHub personal access token (classic)",
+		regexSource: `ghp_[A-Za-z0-9]{36,}`,
+	},
+	{
+		Name:        "github-app-installation-token",
+		Description: "GitHub App installation token (#2090 vector)",
+		regexSource: `ghs_[A-Za-z0-9]{36,}`,
+	},
+	{
+		Name:        "github-oauth-user-to-server",
+		Description: "GitHub OAuth user-to-server token",
+		regexSource: `gho_[A-Za-z0-9]{36,}`,
+	},
+	{
+		Name:        "github-oauth-user",
+		Description: "GitHub OAuth user token",
+		regexSource: `ghu_[A-Za-z0-9]{36,}`,
+	},
+	{
+		Name:        "github-oauth-refresh",
+		Description: "GitHub OAuth refresh token",
+		regexSource: `ghr_[A-Za-z0-9]{36,}`,
+	},
+	{
+		Name:        "github-pat-fine-grained",
+		Description: "GitHub fine-grained personal access token",
+		regexSource: `github_pat_[A-Za-z0-9_]{82,}`,
+	},
+
+	// --- AI-provider API key family ---
+	{
+		Name:        "anthropic-api-key",
+		Description: "Anthropic API key",
+		regexSource: `sk-ant-[A-Za-z0-9_-]{40,}`,
+	},
+	{
+		Name:        "openai-project-key",
+		Description: "OpenAI project API key",
+		regexSource: `sk-proj-[A-Za-z0-9_-]{40,}`,
+	},
+	{
+		Name:        "openai-service-account-key",
+		Description: "OpenAI service-account API key",
+		regexSource: `sk-svcacct-[A-Za-z0-9_-]{40,}`,
+	},
+	{
+		Name:        "minimax-api-key",
+		Description: "MiniMax API key (F1088 vector)",
+		regexSource: `sk-cp-[A-Za-z0-9_-]{60,}`,
+	},
+
+	// --- Chat-platform token family ---
+	{
+		Name:        "slack-token",
+		Description: "Slack token (xoxb/xoxa/xoxp/xoxr/xoxs)",
+		regexSource: `xox[baprs]-[A-Za-z0-9-]{20,}`,
+	},
+
+	// --- Cloud-provider credential family ---
+	{
+		Name:        "aws-access-key-id",
+		Description: "AWS access key ID",
+		regexSource: `AKIA[0-9A-Z]{16}`,
+	},
+	{
+		Name:        "aws-sts-temp-access-key-id",
+		Description: "AWS STS temporary access key ID",
+		regexSource: `ASIA[0-9A-Z]{16}`,
+	},
+}
+
+// compiledOnce protects the lazy build of compiledPatterns. We compile
+// lazily so package init is cheap; callers pay only on first match
+// (typically once per workspace-server boot).
+var (
+	compiledOnce     sync.Once
+	compiledPatterns []*compiledPattern
+	compileErr       error
+)
+
+type compiledPattern struct {
+	Name        string
+	Description string
+	Re          *regexp.Regexp
+}
+
+// compileAll compiles every Pattern.regexSource into a *regexp.Regexp.
+// Called once via compiledOnce. Any compile failure here is a build
+// bug (the unit tests assert each regex compiles) — surfacing via
+// returned error so callers don't panic in request handling.
+func compileAll() {
+	out := make([]*compiledPattern, 0, len(Patterns))
+	for _, p := range Patterns {
+		re, err := regexp.Compile(p.regexSource)
+		if err != nil {
+			compileErr = fmt.Errorf("secrets: pattern %q failed to compile: %w", p.Name, err)
+			return
+		}
+		out = append(out, &compiledPattern{Name: p.Name, Description: p.Description, Re: re})
+	}
+	compiledPatterns = out
+}
+
+// ScanBytes returns a non-nil Match if any pattern matches anywhere
+// inside b. Returns (nil, nil) on no match. Returns (nil, err) only
+// if a regex in the package fails to compile — that's a build bug,
+// not a runtime data issue.
+//
+// Match contains the pattern Name + Description so the caller can
+// emit a path-or-content-denial rationale WITHOUT round-tripping the
+// matched bytes (which would defeat the purpose). The matched bytes
+// stay inside this function.
+//
+// The Files API Phase 2b backend will call ScanBytes on:
+//
+//   - the absolute path string (catches a file literally named
+//     `ghs_abc.txt`)
+//   - the file content (catches a credential pasted into a workspace
+//     file by an agent or user — the Files API refuses to surface it
+//     and the canvas renders "<denied: secret-shape>")
+//
+// Ordering: patterns are tried in declaration order. First match
+// wins. This means narrower patterns (e.g. `sk-svcacct-…`) should
+// appear in `Patterns` before broader ones (`sk-…`) — today there's
+// no overlap, so order is descriptive only.
+func ScanBytes(b []byte) (*Match, error) {
+	compiledOnce.Do(compileAll)
+	if compileErr != nil {
+		return nil, compileErr
+	}
+	for _, cp := range compiledPatterns {
+		if cp.Re.Match(b) {
+			return &Match{Name: cp.Name, Description: cp.Description}, nil
+		}
+	}
+	return nil, nil
+}
+
+// ScanString is the string-input convenience wrapper around ScanBytes.
+// Identical semantics — the body never copies, []byte(s) is a
+// zero-copy reinterpret for the regex matcher.
+func ScanString(s string) (*Match, error) {
+	return ScanBytes([]byte(s))
+}
+
+// Match describes which pattern caught a value. Deliberately does
+// NOT include the matched substring — callers must not echo it.
+type Match struct {
+	// Name is the pattern's kebab-case identifier (e.g. "github-pat-classic").
+	Name string
+	// Description is the human-readable line for UI / log surfaces.
+	Description string
+}
@@ -0,0 +1,189 @@
+package secrets
+
+import (
+	"strings"
+	"testing"
+)
+
+// TestEveryPatternCompiles pins that every Pattern.regexSource is a
+// valid Go-RE2 expression. Without this, a bad regex would silently
+// disable ScanBytes for everything after it (the lazy compile would
+// set compileErr and ScanBytes would return that error every call).
+func TestEveryPatternCompiles(t *testing.T) {
+	for _, p := range Patterns {
+		if p.Name == "" {
+			t.Errorf("pattern with empty Name: regex=%q", p.regexSource)
+		}
+		if p.Description == "" {
+			t.Errorf("pattern %q has empty Description", p.Name)
+		}
+	}
+	// Force compile + check error.
+	if _, err := ScanBytes([]byte("placeholder")); err != nil {
+		t.Fatalf("ScanBytes init failed: %v", err)
+	}
+}
+
+// TestNoDuplicateNames — a duplicate pattern Name would make the
+// "first match wins" semantics surprising to readers and any caller
+// switching on Match.Name (none today but adding the guard is cheap).
+func TestNoDuplicateNames(t *testing.T) {
+	seen := map[string]bool{}
+	for _, p := range Patterns {
+		if seen[p.Name] {
+			t.Errorf("duplicate pattern Name: %q", p.Name)
+		}
+		seen[p.Name] = true
+	}
+}
+
+// TestKnownPatternsAllPresent — pins which specific Name values are
+// expected. A future refactor that renames or removes one without
+// updating consumers (CI workflow, runtime pre-commit hook, Files
+// API Phase 2b backend) would silently widen the leak surface.
+// Failing here forces the rename to be intentional.
+func TestKnownPatternsAllPresent(t *testing.T) {
+	expected := []string{
+		"github-pat-classic",
+		"github-app-installation-token",
+		"github-oauth-user-to-server",
+		"github-oauth-user",
+		"github-oauth-refresh",
+		"github-pat-fine-grained",
+		"anthropic-api-key",
+		"openai-project-key",
+		"openai-service-account-key",
+		"minimax-api-key",
+		"slack-token",
+		"aws-access-key-id",
+		"aws-sts-temp-access-key-id",
+	}
+	got := map[string]bool{}
+	for _, p := range Patterns {
+		got[p.Name] = true
+	}
+	for _, want := range expected {
+		if !got[want] {
+			t.Errorf("expected pattern %q missing from Patterns slice", want)
+		}
+	}
+}
+
+// TestPositiveMatches — for each pattern, supply a representative
+// shape and assert ScanBytes returns a Match with the right Name.
+// These are TEST FIXTURES, not real credentials — each is the
+// pattern's prefix + a long-enough trailing run of placeholder chars.
+// `EXAMPLE` is sprinkled in to make grep-finds in CI logs obviously
+// fake to a human reader (matches saved memory
+// feedback_assert_exact_not_substring: tighten by Name not body).
+func TestPositiveMatches(t *testing.T) {
+	cases := []struct {
+		fixture      string
+		expectedName string
+	}{
+		{"ghp_EXAMPLE111122223333444455556666777788889999", "github-pat-classic"},
+		{"ghs_EXAMPLE111122223333444455556666777788889999", "github-app-installation-token"},
+		{"gho_EXAMPLE111122223333444455556666777788889999", "github-oauth-user-to-server"},
+		{"ghu_EXAMPLE111122223333444455556666777788889999", "github-oauth-user"},
+		{"ghr_EXAMPLE111122223333444455556666777788889999", "github-oauth-refresh"},
+		{"github_pat_EXAMPLE" + strings.Repeat("1", 80), "github-pat-fine-grained"},
+		{"sk-ant-EXAMPLE" + strings.Repeat("1", 40), "anthropic-api-key"},
+		{"sk-proj-EXAMPLE" + strings.Repeat("1", 40), "openai-project-key"},
+		{"sk-svcacct-EXAMPLE" + strings.Repeat("1", 40), "openai-service-account-key"},
+		{"sk-cp-EXAMPLE" + strings.Repeat("1", 60), "minimax-api-key"},
+		{"xoxb-" + strings.Repeat("a", 25), "slack-token"},
+		{"xoxa-" + strings.Repeat("a", 25), "slack-token"},
+		// AWS regex requires [0-9A-Z]{16} — uppercase + digits only.
+		{"AKIA1234567890ABCDEF", "aws-access-key-id"},
+		{"ASIA1234567890ABCDEF", "aws-sts-temp-access-key-id"},
+	}
+	for _, tc := range cases {
+		t.Run(tc.expectedName, func(t *testing.T) {
+			m, err := ScanBytes([]byte(tc.fixture))
+			if err != nil {
+				t.Fatalf("ScanBytes(%q) errored: %v", tc.fixture, err)
+			}
+			if m == nil {
+				t.Fatalf("ScanBytes(%q) returned no match — expected %q", tc.fixture, tc.expectedName)
+			}
+			if m.Name != tc.expectedName {
+				t.Errorf("ScanBytes(%q) matched %q; expected %q", tc.fixture, m.Name, tc.expectedName)
+			}
+		})
+	}
+}
+
+// TestNegativeShapes — strings that look credential-adjacent but
+// shouldn't match (too short, wrong prefix, missing trailing bytes).
+// Failing here means a pattern is too loose, which would generate
+// false-positive denial in Files API and false-positive workflow
+// failures in CI.
+func TestNegativeShapes(t *testing.T) {
+	cases := []string{
+		// Too-short variants — anchored on the length suffix.
+		"ghp_tooshort",
+		"ghs_alsoshort1234",
+		"github_pat_short",
+		"sk-ant-short",
+		"sk-cp-not-enough-bytes-here",
+		// Looks like one of the prefixes but isn't (different letter).
+		"gha_EXAMPLE_thirty_six_or_more_chars_here_xxx",
+		// Slack family — wrong letter after xox.
+		"xoxz-aaaaaaaaaaaaaaaaaaaaaaaaa",
+		// AWS-shaped but wrong length suffix.
+		"AKIATOOSHORT",
+		// Empty / whitespace.
+		"",
+		"   ",
+		// Plain prose mentioning the prefix as part of a longer word.
+		"see also `ghp_HOWTO.md` in the repo",
+	}
+	for _, c := range cases {
+		t.Run(c, func(t *testing.T) {
+			m, err := ScanBytes([]byte(c))
+			if err != nil {
+				t.Fatalf("ScanBytes(%q) errored: %v", c, err)
+			}
+			if m != nil {
+				t.Errorf("ScanBytes(%q) unexpectedly matched %q", c, m.Name)
+			}
+		})
+	}
+}
+
+// TestScanString_NoOp — sanity-check ScanString is the zero-copy
+// wrapper around ScanBytes. Without this, a future refactor that
+// makes ScanString do its own thing (e.g. accidentally normalise
+// case) would diverge silently.
+func TestScanString_NoOp(t *testing.T) {
+	in := "ghp_EXAMPLE111122223333444455556666777788889999"
+	m1, err1 := ScanBytes([]byte(in))
+	if err1 != nil {
+		t.Fatalf("ScanBytes errored: %v", err1)
+	}
+	m2, err2 := ScanString(in)
+	if err2 != nil {
+		t.Fatalf("ScanString errored: %v", err2)
+	}
+	if m1 == nil || m2 == nil {
+		t.Fatalf("expected matches; got bytes=%+v string=%+v", m1, m2)
+	}
+	if m1.Name != m2.Name {
+		t.Errorf("ScanString and ScanBytes returned different Names: %q vs %q", m1.Name, m2.Name)
+	}
+}
+
+// TestMatch_NoRoundtrip — assert the Match struct does NOT include
+// the matched substring as a field. Adding such a field would
+// regress the "matched bytes never leave ScanBytes" invariant that
+// makes this package safe to call from log/UI surfaces. This is a
+// reflection-light contract test — checks the field names statically.
+func TestMatch_NoRoundtrip(t *testing.T) {
+	var m Match
+	// If someone adds a `Matched string` (or similar) field, this
+	// test reads as the canonical place to update + reconsider.
+	_ = m.Name
+	_ = m.Description
+	// The two-field shape is part of the public contract; new fields
+	// require deliberation about whether they leak the secret value.
+}
@@ -35,12 +35,14 @@ from a2a_tools import (
    tool_commit_memory,
    tool_delegate_task,
    tool_delegate_task_async,
+    tool_get_runtime_identity,
    tool_get_workspace_info,
    tool_inbox_peek,
    tool_inbox_pop,
    tool_list_peers,
    tool_recall_memory,
    tool_send_message_to_user,
+    tool_update_agent_card,
    tool_wait_for_message,
 )
 from platform_tools.registry import TOOLS as _PLATFORM_TOOL_SPECS
@@ -130,6 +132,10 @@ async def handle_tool_call(name: str, arguments: dict) -> str:
        return await tool_get_workspace_info(
            source_workspace_id=arguments.get("source_workspace_id") or None,
        )
+    elif name == "get_runtime_identity":
+        return await tool_get_runtime_identity()
+    elif name == "update_agent_card":
+        return await tool_update_agent_card(arguments.get("card"))
    elif name == "commit_memory":
        return await tool_commit_memory(
            arguments.get("content", ""),
@@ -167,3 +167,15 @@ from a2a_tools_inbox import (  # noqa: E402  (import after the top-of-module imp
    tool_inbox_pop,
    tool_wait_for_message,
 )
+
+
+# Identity tool handlers — extracted to a2a_tools_identity. Ports the
+# two T4-tier MCP tools (``tool_get_runtime_identity`` +
+# ``tool_update_agent_card``) from molecule-ai-workspace-runtime PR#17.
+# That repo is mirror-only (reference_runtime_repo_is_mirror_only);
+# this is the canonical edit point, and the wheel mirror is
+# regenerated by publish-runtime.yml on merge.
+from a2a_tools_identity import (  # noqa: E402  (import after the top-of-module imports)
+    tool_get_runtime_identity,
+    tool_update_agent_card,
+)
@@ -0,0 +1,187 @@
+"""Identity tool handlers — single-concern slice of the a2a_tools surface.
+
+Owns the two MCP tools that close the T4-tier workspace owner-permission
+gaps reported via the canvas:
+
+  * ``tool_get_runtime_identity`` — env-only; returns model, model_provider,
+    molecule_model, anthropic_base_url, tier, workspace_id, runtime
+    (ADAPTER_MODULE). No HTTP call. Always permitted by RBAC — even
+    read-only agents may know what model they are.
+
+  * ``tool_update_agent_card`` — POSTs the card to ``/registry/update-card``
+    with the workspace's own bearer (same auth path as ``tool_commit_memory``
+    via ``a2a_tools_rbac.auth_headers_for_heartbeat``). The platform
+    replaces the stored card and broadcasts an ``agent_card_updated``
+    event so the canvas reflects the new card live. Gated on
+    ``memory.write`` capability via the existing RBAC permission map so
+    read-only roles can't silently rewrite the platform card.
+
+Both originated as a port of molecule-ai-workspace-runtime PR#17
+(``feat(mcp): add update_agent_card + get_runtime_identity tools``).
+The mirror-only PR#17 was closed without merge per
+``reference_runtime_repo_is_mirror_only``; the canonical edit point is
+this monorepo at ``workspace/`` and the wheel mirror is regenerated
+automatically by the publish-runtime workflow.
+
+Imports the auth-header primitive from ``a2a_tools_rbac`` (iter 4a) —
+NOT from ``a2a_tools`` — to avoid a circular import with the
+kitchen-sink re-export module.
+"""
+from __future__ import annotations
+
+import json
+import os
+from typing import Any
+
+import httpx
+
+from a2a_client import PLATFORM_URL
+from a2a_tools_rbac import (
+    auth_headers_for_heartbeat as _auth_headers_for_heartbeat,
+    check_memory_write_permission as _check_memory_write_permission,
+)
+
+
+def _runtime_identity_payload() -> dict[str, Any]:
+    """Build the identity dict — env-only, no I/O.
+
+    Factored out from ``tool_get_runtime_identity`` so tests can assert
+    against the exact key set without re-parsing JSON. The MCP tool
+    handler ``tool_get_runtime_identity`` is the only public caller in
+    production; tests call this helper directly.
+    """
+    return {
+        "model": os.environ.get("MODEL", ""),
+        "model_provider": os.environ.get("MODEL_PROVIDER", ""),
+        "molecule_model": os.environ.get("MOLECULE_MODEL", ""),
+        "anthropic_base_url": os.environ.get("ANTHROPIC_BASE_URL", ""),
+        "tier": os.environ.get("TIER", ""),
+        "workspace_id": os.environ.get("WORKSPACE_ID", ""),
+        # Adapter module is the closest thing the runtime has to a
+        # "template slug" — e.g. "adapter" for claude-code-default,
+        # "hermes" for hermes-template, etc. Picked from
+        # $ADAPTER_MODULE env baked by each template's Dockerfile.
+        "runtime": os.environ.get("ADAPTER_MODULE", ""),
+    }
+
+
+async def tool_get_runtime_identity() -> str:
+    """Return this runtime's identity — model, provider, tier, IDs.
+
+    Env-only; no HTTP call. Useful so the agent can answer "what model
+    am I?" correctly instead of guessing from a stale system prompt
+    that the operator may have changed between boots.
+
+    Returns the identity as a JSON-encoded string (the dispatch contract
+    every MCP tool in this module follows). Tests that want to assert
+    individual fields can call ``_runtime_identity_payload()`` directly,
+    or ``json.loads`` the return value.
+
+    Always permitted by RBAC — there is no sensitive information here
+    that isn't already available to the process via ``os.environ``.
+    The point of the tool is to surface those env values to the agent
+    layer in a stable, documented shape rather than expecting every
+    agent runtime to know to ``echo $MODEL``.
+    """
+    return json.dumps(_runtime_identity_payload(), indent=2)
+
+
+async def tool_update_agent_card(card: Any) -> str:
+    """Update this workspace's agent_card on the platform.
+
+    POSTs the provided card to ``/registry/update-card`` with the
+    workspace's own bearer token (same auth path as ``tool_commit_memory``
+    and ``tool_get_workspace_info``). The platform validates required
+    fields server-side, replaces the stored card, and broadcasts an
+    ``agent_card_updated`` event so the canvas updates live.
+
+    Args:
+        card: A JSON-serialisable object (typically a dict) holding the
+            new card. The platform validates required fields server-side.
+
+    Returns:
+        JSON-encoded string. Body:
+          - ``{"success": true, "status": "updated"}`` on success;
+          - ``{"success": false, "error": "<msg>", "status_code": <int>}``
+            on platform error;
+          - ``{"success": false, "error": "<reason>"}`` on local validation
+            (non-dict card, missing WORKSPACE_ID, network error).
+
+    Permission gate: this tool requires the ``memory.write`` RBAC
+    capability — same gate as ``tool_commit_memory``. The check runs
+    inline rather than at the dispatcher layer to keep ``a2a_mcp_server``
+    permission-agnostic (the gate sits with the implementation, not the
+    transport). Read-only roles get a clear error string back instead
+    of a 403 from the platform.
+
+    We re-check ``isinstance(card, dict)`` here defensively rather than
+    trust the MCP schema validator alone — the schema only constrains
+    the transport, not the in-process call surface used by tests and
+    sibling modules.
+    """
+    payload = await _update_agent_card_impl(card)
+    return json.dumps(payload, indent=2)
+
+
+async def _update_agent_card_impl(card: Any) -> dict[str, Any]:
+    """Dict-returning core of ``tool_update_agent_card``.
+
+    Split out so tests can assert against the raw dict shape (status
+    codes, error messages) without re-parsing JSON on every assertion.
+    The string-returning ``tool_update_agent_card`` is a thin wrapper
+    invoked by the MCP dispatcher.
+    """
+    # RBAC: require memory.write permission. Same gate as
+    # tool_commit_memory (the agent already needs this capability to
+    # persist anything outbound). Read-only roles can still call
+    # get_runtime_identity / get_workspace_info to introspect — those
+    # are env-only / read-only and have no inline gate.
+    if not _check_memory_write_permission():
+        return {
+            "success": False,
+            "error": (
+                "RBAC — this workspace does not have the 'memory.write' "
+                "permission required to update the agent_card."
+            ),
+        }
+    if not isinstance(card, dict):
+        return {
+            "success": False,
+            "error": "card must be a JSON object (dict)",
+        }
+    ws_id = os.environ.get("WORKSPACE_ID", "")
+    if not ws_id:
+        return {
+            "success": False,
+            "error": "WORKSPACE_ID env not set; cannot identify caller",
+        }
+    try:
+        async with httpx.AsyncClient(timeout=10.0) as client:
+            resp = await client.post(
+                f"{PLATFORM_URL}/registry/update-card",
+                json={"workspace_id": ws_id, "agent_card": card},
+                headers=_auth_headers_for_heartbeat(),
+            )
+            if resp.status_code == 200:
+                body: dict[str, Any] = {}
+                try:
+                    body = resp.json()
+                except Exception:
+                    pass
+                return {
+                    "success": True,
+                    "status": body.get("status", "updated"),
+                }
+            # Non-200 — surface what the platform returned.
+            error_msg = ""
+            try:
+                error_msg = resp.json().get("error", "") or resp.text
+            except Exception:
+                error_msg = resp.text
+            return {
+                "success": False,
+                "status_code": resp.status_code,
+                "error": error_msg,
+            }
+    except Exception as e:
+        return {"success": False, "error": f"network error: {e}"}
@@ -340,6 +340,16 @@ _CLI_A2A_COMMAND_KEYWORDS: dict[str, str | None] = {
    "delegate_task_async": "delegate --async",
    "check_task_status": "status",
    "get_workspace_info": "info",
+    # `get_runtime_identity` + `update_agent_card` are MCP-first
+    # capabilities — the CLI subprocess interface doesn't expose them
+    # today. `get_runtime_identity` is env-only and an agent on a
+    # CLI-only runtime can already `echo $MODEL` etc, so there's no
+    # functional gap. `update_agent_card` requires a JSON object
+    # argument that wouldn't survive a positional-arg shell invocation
+    # cleanly. Mapped to None — flip to a keyword if a2a_cli grows
+    # `identity` / `card` subcommands in the future.
+    "get_runtime_identity": None,
+    "update_agent_card": None,
    # `broadcast_message` is not exposed via the CLI subprocess interface
    # today — it's an MCP-first capability. If a2a_cli grows a `broadcast`
    # subcommand, map it here and the alignment test will gate the change.
@@ -589,6 +599,28 @@ def _sanitize_for_external(msg: str) -> str:
    import re as _re

    msg = _re.sub(r"(?i)(?:bearer|token|api[_-]?key|sk-)[ :=]+[A-Za-z0-9_/.-]{20,}", "[REDACTED]", msg)
+    # Bare provider key with NO separator after the prefix — a real
+    # `sk-ant-api03-…` / `sk-…` key uses `-` (not `[ :=]`) so the rule
+    # above misses it. Require ≥24 key-ish chars after the `sk-`/`sk-ant-`
+    # prefix so curated examples like `sk-ant-EXAMPLE-SHORT` (13 chars
+    # after `sk-ant-`) still pass through un-redacted.
+    msg = _re.sub(r"(?i)\bsk-(?:ant-)?[A-Za-z0-9_-]{24,}", "[REDACTED]", msg)
+    # JSON-quoted credential values: {"token": "…"} / {"apiKey": "…"} /
+    # {"secret": "…"} / {"password": "…"}. Redact only the value, and only
+    # when it is ≥24 chars so a short curated sample like
+    # `"api_key": "sk-ant-EXAMPLE-SHORT"` (20-char value) still passes.
+    msg = _re.sub(
+        r'(?i)("(?:token|api[_-]?key|secret|password)"\s*:\s*")[^"]{24,}(")',
+        r"\1[REDACTED]\2",
+        msg,
+    )
+    # AWS secret access key in `aws_secret_access_key=…` form (env dumps,
+    # boto tracebacks). The base64-ish value runs until whitespace/quote.
+    msg = _re.sub(
+        r"(?i)(aws_secret_access_key\s*[:=]\s*)\S+",
+        r"\1[REDACTED]",
+        msg,
+    )
    # Absolute paths: /etc/shadow, /home/user/.aws/credentials, etc.
    msg = _re.sub(r"(?:/[^/\s]+){2,}", lambda m: m.group(0) if len(m.group(0)) < 60 else "[REDACTED_PATH]", msg)
    return msg
@@ -598,6 +630,7 @@ def sanitize_agent_error(
    exc: BaseException | None = None,
    category: str | None = None,
    stderr: str | None = None,
+    reason: str | None = None,
 ) -> str:
    """Render an agent-side failure into a user-safe error message.

@@ -605,6 +638,18 @@ def sanitize_agent_error(
    category string (e.g. from `classify_subprocess_error`). If both are
    given, `category` wins. If neither, the tag defaults to "unknown".

+    When ``reason`` is provided (internal#211/#212), it is a *pre-curated,
+    user-actionable, secret-safe* explanation built by the caller from a
+    provider-side failure — e.g. a 403 "Your organization has disabled
+    Claude subscription access · Use an Anthropic API key instead, or ask
+    your admin to enable access" with error code ``oauth_org_not_allowed``.
+    This text is exactly what the user needs to self-serve, so it is
+    surfaced VERBATIM as the message instead of being collapsed to the
+    opaque exception class name. It still passes through the
+    key/token/bearer/path scrubber as a belt-and-braces second pass so a
+    buggy caller can't leak a credential that snuck into the reason.
+    ``reason`` wins over ``stderr``; both lose to neither being set.
+
    When ``stderr`` is provided (e.g. the first ~1 KB of a subprocess stderr
    or HTTP error body), it is sanitized and appended to the output so the
    A2A caller gets actionable context without needing to dig through workspace
@@ -619,6 +664,13 @@ def sanitize_agent_error(
    else:
        tag = "unknown"

+    if reason:
+        # Curated, user-actionable reason — surface it as the message.
+        # Still scrub: a 403/auth/quota message is safe, but the scrubber
+        # is cheap insurance against a caller that didn't curate cleanly.
+        clean = _sanitize_for_external(reason[:_MAX_STDERR_PREVIEW])
+        return f"Agent error ({tag}): {clean}"
+
    if stderr:
        # Truncate and sanitize before including — prevents DoS via
        # a malicious or buggy peer injecting a huge error body, and
@@ -57,12 +57,14 @@ from a2a_tools import (
    tool_commit_memory,
    tool_delegate_task,
    tool_delegate_task_async,
+    tool_get_runtime_identity,
    tool_get_workspace_info,
    tool_inbox_peek,
    tool_inbox_pop,
    tool_list_peers,
    tool_recall_memory,
    tool_send_message_to_user,
+    tool_update_agent_card,
    tool_wait_for_message,
 )

@@ -289,6 +291,61 @@ _GET_WORKSPACE_INFO = ToolSpec(
    section=A2A_SECTION,
 )

+_GET_RUNTIME_IDENTITY = ToolSpec(
+    name="get_runtime_identity",
+    short=(
+        "Return this runtime's identity — model, model_provider, tier, "
+        "workspace_id, runtime template. Reads from process env; no HTTP call."
+    ),
+    when_to_use=(
+        "Use this to answer 'what model am I?' truthfully instead of "
+        "guessing from a stale system prompt — the operator may have "
+        "routed you to a different model via persona env between boots. "
+        "Always permitted by RBAC: even read-only agents may know what "
+        "model they are. Distinct from get_workspace_info — that one "
+        "calls the platform for ID/role/tier/parent (workspace metadata); "
+        "this one returns the live process env (MODEL, MODEL_PROVIDER, "
+        "MOLECULE_MODEL, ANTHROPIC_BASE_URL, TIER, WORKSPACE_ID, "
+        "ADAPTER_MODULE)."
+    ),
+    input_schema={"type": "object", "properties": {}},
+    impl=tool_get_runtime_identity,
+    section=A2A_SECTION,
+)
+
+_UPDATE_AGENT_CARD = ToolSpec(
+    name="update_agent_card",
+    short=(
+        "Replace this workspace's agent_card on the platform. The "
+        "platform validates required fields and broadcasts an "
+        "agent_card_updated event so the canvas reflects the change live."
+    ),
+    when_to_use=(
+        "Use when the workspace's capabilities, skills, description, or "
+        "name change and the canvas display needs to follow. The "
+        "platform stores the new card and pushes an "
+        "``agent_card_updated`` event to subscribers. Gated behind the "
+        "``memory.write`` RBAC capability — read-only roles cannot "
+        "rewrite the card. Tier-1+ owners always have this capability."
+    ),
+    input_schema={
+        "type": "object",
+        "properties": {
+            "card": {
+                "type": "object",
+                "description": (
+                    "The new agent_card object (name, version, "
+                    "description, skills, etc). Server-side validation "
+                    "rejects payloads missing required fields."
+                ),
+            },
+        },
+        "required": ["card"],
+    },
+    impl=tool_update_agent_card,
+    section=A2A_SECTION,
+)
+
 _BROADCAST_MESSAGE = ToolSpec(
    name="broadcast_message",
    short=(
@@ -642,6 +699,8 @@ TOOLS: list[ToolSpec] = [
    _CHECK_TASK_STATUS,
    _LIST_PEERS,
    _GET_WORKSPACE_INFO,
+    _GET_RUNTIME_IDENTITY,
+    _UPDATE_AGENT_CARD,
    _BROADCAST_MESSAGE,
    _SEND_MESSAGE_TO_USER,
    # Inbox (standalone-only; in-container returns informational error)
@@ -5,6 +5,8 @@
 - **check_task_status**: Poll the status of a task started with delegate_task_async; returns result when done.
 - **list_peers**: List the workspaces this agent can communicate with — name, ID, status, role for each.
 - **get_workspace_info**: Get this workspace's own info — ID, name, role, tier, parent, status.
+- **get_runtime_identity**: Return this runtime's identity — model, model_provider, tier, workspace_id, runtime template. Reads from process env; no HTTP call.
+- **update_agent_card**: Replace this workspace's agent_card on the platform. The platform validates required fields and broadcasts an agent_card_updated event so the canvas reflects the change live.
 - **broadcast_message**: Send a message to ALL agent workspaces in the org simultaneously. Requires broadcast_enabled=true on this workspace (set by user/admin).
 - **send_message_to_user**: Send a message directly to the user's canvas chat — pushed instantly via WebSocket. Use this to: (1) acknowledge a task immediately ('Got it, I'll start working on this'), (2) send interim progress updates while doing long work, (3) deliver follow-up results after delegation completes, (4) attach files (zip, pdf, csv, image) for the user to download via the `attachments` field (NEVER paste file URLs in `message`). The message appears in the user's chat as if you're proactively reaching out.
 - **wait_for_message**: Block until the next inbound message (canvas user OR peer agent) arrives, or until ``timeout_secs`` elapses.
@@ -27,6 +29,12 @@ Call this first when you need to delegate but don't know the target's ID. Access
 ### get_workspace_info
 Use to introspect your own identity (e.g. before reporting back to the user, or to determine whether you're a tier-0 root that can write GLOBAL memory).

+### get_runtime_identity
+Use this to answer 'what model am I?' truthfully instead of guessing from a stale system prompt — the operator may have routed you to a different model via persona env between boots. Always permitted by RBAC: even read-only agents may know what model they are. Distinct from get_workspace_info — that one calls the platform for ID/role/tier/parent (workspace metadata); this one returns the live process env (MODEL, MODEL_PROVIDER, MOLECULE_MODEL, ANTHROPIC_BASE_URL, TIER, WORKSPACE_ID, ADAPTER_MODULE).
+
+### update_agent_card
+Use when the workspace's capabilities, skills, description, or name change and the canvas display needs to follow. The platform stores the new card and pushes an ``agent_card_updated`` event to subscribers. Gated behind the ``memory.write`` RBAC capability — read-only roles cannot rewrite the card. Tier-1+ owners always have this capability.
+
 ### broadcast_message
 Use for urgent, org-wide signals: critical status changes, emergency stop instructions, coordinated task announcements. Every non-removed workspace receives the message in its activity log (poll-mode agents see it on their next poll; push-mode canvases get a real-time banner). This tool returns an error if broadcast_enabled is false — a user or admin must enable it via the workspace abilities settings first.

@@ -0,0 +1,390 @@
+"""Tests for ``tool_get_runtime_identity`` and ``tool_update_agent_card``.
+
+These two MCP tools close the T4-tier workspace owner-permission gaps
+reported via the canvas:
+
+  - the agent could not update its own ``agent_card`` (no MCP tool
+    wrapped the existing ``POST /registry/update-card`` endpoint);
+  - the agent could not identify which model it was running (the
+    ``MODEL`` env var is injected by ``provisioner.workspace_provision``
+    but nothing surfaced it back to the agent).
+
+Ported from molecule-ai-workspace-runtime PR#17 (mirror-only repo;
+canonical edit point per ``reference_runtime_repo_is_mirror_only``).
+Adapted to core's conventions:
+
+  * tool functions return ``str`` (JSON-encoded), matching every other
+    tool in ``a2a_tools_*`` modules. Tests ``json.loads`` to inspect.
+  * permission check ``memory.write`` runs inline in
+    ``tool_update_agent_card`` (same pattern as
+    ``a2a_tools_memory.tool_commit_memory``).
+  * ``WORKSPACE_ID`` is read directly from ``os.environ`` — core does
+    not have the runtime's validated-cache layer (``molecule_runtime.
+    builtin_tools.validation``).
+"""
+from __future__ import annotations
+
+import json
+
+import pytest
+
+
+# --- Drift gate: re-export aliases on a2a_tools ------------------------------
+
+class TestBackCompatAliases:
+    """Pin that ``a2a_tools.tool_*`` resolves to the same callable as
+    ``a2a_tools_identity.tool_*``. Refactor wrapping (e.g. a doc-string
+    wrapper that loses the function identity) silently breaks call
+    sites that ``patch("a2a_tools.tool_update_agent_card", ...)`` —
+    this gate makes that drift fail fast."""
+
+    def test_tool_get_runtime_identity_alias(self):
+        import a2a_tools
+        import a2a_tools_identity
+        assert a2a_tools.tool_get_runtime_identity is a2a_tools_identity.tool_get_runtime_identity
+
+    def test_tool_update_agent_card_alias(self):
+        import a2a_tools
+        import a2a_tools_identity
+        assert a2a_tools.tool_update_agent_card is a2a_tools_identity.tool_update_agent_card
+
+
+# --- tool_get_runtime_identity ----------------------------------------------
+
+class TestGetRuntimeIdentity:
+    """The tool returns env-derived runtime identity. No HTTP call."""
+
+    @pytest.mark.asyncio
+    async def test_returns_all_known_env_fields(self, monkeypatch):
+        from a2a_tools_identity import tool_get_runtime_identity
+
+        monkeypatch.setenv("MODEL", "claude-opus-4-7")
+        monkeypatch.setenv("MODEL_PROVIDER", "anthropic")
+        monkeypatch.setenv("TIER", "T4")
+        monkeypatch.setenv("WORKSPACE_ID", "ws-abc")
+        monkeypatch.setenv("ADAPTER_MODULE", "adapter")
+        monkeypatch.setenv("MOLECULE_MODEL", "claude-opus-4-7")
+        monkeypatch.setenv("ANTHROPIC_BASE_URL", "https://api.anthropic.com")
+
+        out = await tool_get_runtime_identity()
+        # MCP tools return JSON-encoded strings (matches the contract
+        # every other tool_* in a2a_tools_* uses).
+        assert isinstance(out, str)
+        parsed = json.loads(out)
+
+        assert parsed["model"] == "claude-opus-4-7"
+        assert parsed["model_provider"] == "anthropic"
+        assert parsed["tier"] == "T4"
+        assert parsed["workspace_id"] == "ws-abc"
+        assert parsed["runtime"] == "adapter"
+        assert parsed["molecule_model"] == "claude-opus-4-7"
+        assert parsed["anthropic_base_url"] == "https://api.anthropic.com"
+
+    @pytest.mark.asyncio
+    async def test_missing_env_returns_empty_strings(self, monkeypatch):
+        """Tool MUST NOT raise when env vars are absent — every key is
+        present but the value is the empty string. The agent then knows
+        the slot exists but is unset."""
+        from a2a_tools_identity import tool_get_runtime_identity
+
+        for var in (
+            "MODEL", "MODEL_PROVIDER", "TIER", "WORKSPACE_ID",
+            "ADAPTER_MODULE", "MOLECULE_MODEL", "ANTHROPIC_BASE_URL",
+        ):
+            monkeypatch.delenv(var, raising=False)
+
+        parsed = json.loads(await tool_get_runtime_identity())
+        assert parsed["model"] == ""
+        assert parsed["model_provider"] == ""
+        assert parsed["tier"] == ""
+        assert parsed["workspace_id"] == ""
+        assert parsed["runtime"] == ""
+        assert parsed["molecule_model"] == ""
+        assert parsed["anthropic_base_url"] == ""
+
+    @pytest.mark.asyncio
+    async def test_no_http_call_made(self, monkeypatch):
+        """``get_runtime_identity`` is env-only — must not open
+        httpx.AsyncClient even if the call would otherwise succeed.
+        Tripwire any client construction."""
+        import httpx
+
+        from a2a_tools_identity import tool_get_runtime_identity
+
+        class _Tripwire:
+            def __init__(self, *_a, **_kw):
+                raise AssertionError(
+                    "tool_get_runtime_identity must not open httpx.AsyncClient"
+                )
+
+        monkeypatch.setattr(httpx, "AsyncClient", _Tripwire)
+        # Must not raise.
+        await tool_get_runtime_identity()
+
+    @pytest.mark.asyncio
+    async def test_helper_dict_matches_string_payload(self, monkeypatch):
+        """``_runtime_identity_payload`` is the dict-returning helper
+        used by both the public tool and tests. Verify the public tool
+        json.dumps the same dict — no field is dropped or renamed by
+        the encoding step."""
+        from a2a_tools_identity import (
+            _runtime_identity_payload,
+            tool_get_runtime_identity,
+        )
+
+        monkeypatch.setenv("MODEL", "claude-opus-4-7")
+        monkeypatch.setenv("TIER", "T4")
+        monkeypatch.setenv("WORKSPACE_ID", "ws-helper-check")
+
+        helper = _runtime_identity_payload()
+        tool_str = await tool_get_runtime_identity()
+        assert json.loads(tool_str) == helper
+
+
+# --- tool_update_agent_card -------------------------------------------------
+
+
+class _MockResponse:
+    def __init__(self, status_code: int, payload: dict):
+        self.status_code = status_code
+        self._payload = payload
+        self.text = json.dumps(payload)
+
+    def json(self):
+        return self._payload
+
+
+class _MockClient:
+    """Drop-in for httpx.AsyncClient context manager.
+
+    Records the URL + json body + headers the tool POSTed so the test
+    can assert against them. Returns the canned _MockResponse passed
+    in at construction time.
+    """
+
+    def __init__(self, *, response: _MockResponse, captured: dict):
+        self._response = response
+        self._captured = captured
+
+    async def __aenter__(self):
+        return self
+
+    async def __aexit__(self, *_args):
+        return False
+
+    async def post(self, url, *, json=None, headers=None, **_kw):  # noqa: A002
+        self._captured["url"] = url
+        self._captured["json"] = json
+        self._captured["headers"] = headers
+        return self._response
+
+
+@pytest.fixture
+def _grant_memory_write(monkeypatch):
+    """Force the inline RBAC gate inside ``tool_update_agent_card`` to
+    succeed. The gate calls
+    ``a2a_tools_rbac.check_memory_write_permission`` which inspects
+    ``$MOLECULE_ROLES`` / the role table; the patch sidesteps that
+    machinery so tests can focus on the platform-call shape.
+    """
+    import a2a_tools_identity
+    monkeypatch.setattr(
+        a2a_tools_identity, "_check_memory_write_permission", lambda: True
+    )
+
+
+class TestUpdateAgentCard:
+    @pytest.mark.asyncio
+    async def test_posts_to_registry_update_card(
+        self, monkeypatch, _grant_memory_write,
+    ):
+        """Hits POST {PLATFORM_URL}/registry/update-card with the
+        workspace bearer and the {workspace_id, agent_card} body shape
+        the platform handler expects (workspace-server
+        ``internal/handlers/registry.go``)."""
+        import a2a_tools_identity
+
+        monkeypatch.setenv("WORKSPACE_ID", "ws-42")
+        # Ensure PLATFORM_URL re-import sees a deterministic value —
+        # a2a_client imports it at module load so we patch the symbol
+        # on a2a_tools_identity directly (the module's own reference).
+        monkeypatch.setattr(a2a_tools_identity, "PLATFORM_URL", "http://test.invalid")
+
+        captured: dict = {}
+        response = _MockResponse(200, {"status": "updated"})
+
+        def _client_factory(*_a, **_kw):
+            return _MockClient(response=response, captured=captured)
+
+        monkeypatch.setattr(a2a_tools_identity.httpx, "AsyncClient", _client_factory)
+        monkeypatch.setattr(
+            a2a_tools_identity, "_auth_headers_for_heartbeat",
+            lambda: {"Authorization": "Bearer ws-token-xyz"},
+        )
+
+        card = {"name": "agent-foo", "version": "0.1.0", "description": "demo"}
+        result_str = await a2a_tools_identity.tool_update_agent_card(card)
+        result = json.loads(result_str)
+
+        # URL: PLATFORM_URL + /registry/update-card
+        assert captured["url"] == "http://test.invalid/registry/update-card"
+
+        # The platform handler expects {workspace_id, agent_card}; the
+        # agent_card is the raw object the agent submitted.
+        body = captured["json"]
+        assert body["workspace_id"] == "ws-42"
+        assert body["agent_card"] == card
+
+        # Auth header from auth_headers_for_heartbeat is forwarded
+        # verbatim — same path commit_memory uses.
+        assert captured["headers"]["Authorization"] == "Bearer ws-token-xyz"
+
+        assert result["success"] is True
+        assert result["status"] == "updated"
+
+    @pytest.mark.asyncio
+    async def test_propagates_server_error(
+        self, monkeypatch, _grant_memory_write,
+    ):
+        """Non-200 from platform surfaces as a structured error to the
+        agent. The agent sees {success:false, status_code, error} and
+        can decide whether to retry, fall back, or escalate."""
+        import a2a_tools_identity
+
+        monkeypatch.setenv("WORKSPACE_ID", "ws-42")
+        monkeypatch.setattr(a2a_tools_identity, "PLATFORM_URL", "http://test.invalid")
+
+        captured: dict = {}
+        response = _MockResponse(400, {"error": "invalid card"})
+
+        monkeypatch.setattr(
+            a2a_tools_identity.httpx, "AsyncClient",
+            lambda *a, **kw: _MockClient(response=response, captured=captured),
+        )
+        monkeypatch.setattr(
+            a2a_tools_identity, "_auth_headers_for_heartbeat", lambda: {},
+        )
+
+        result = json.loads(
+            await a2a_tools_identity.tool_update_agent_card({"name": "x"})
+        )
+        assert result["success"] is False
+        assert result["status_code"] == 400
+        assert "invalid card" in str(result["error"]).lower()
+
+    @pytest.mark.asyncio
+    async def test_rejects_non_dict_card(self, _grant_memory_write):
+        """The MCP schema constrains transport callers to pass a dict;
+        in-process callers (tests, sibling modules) can still pass any
+        type. Reject non-dict defensively so the platform isn't asked
+        to validate JSON-encoded strings or lists."""
+        from a2a_tools_identity import tool_update_agent_card
+
+        result = json.loads(await tool_update_agent_card("not-a-dict"))
+        assert result["success"] is False
+        assert "dict" in str(result["error"]).lower()
+
+    @pytest.mark.asyncio
+    async def test_workspace_id_missing_returns_error(
+        self, monkeypatch, _grant_memory_write,
+    ):
+        """If WORKSPACE_ID is not set the tool refuses to issue the
+        request — it would otherwise POST with an empty workspace_id
+        and let the platform return a confusing 400."""
+        from a2a_tools_identity import tool_update_agent_card
+
+        monkeypatch.delenv("WORKSPACE_ID", raising=False)
+
+        result = json.loads(await tool_update_agent_card({"name": "x"}))
+        assert result["success"] is False
+        assert "workspace_id" in str(result["error"]).lower()
+
+    @pytest.mark.asyncio
+    async def test_denies_when_memory_write_permission_missing(self, monkeypatch):
+        """The agent's RBAC role must grant ``memory.write`` to update
+        the card. Read-only roles get an RBAC error string back
+        immediately, never touching the platform."""
+        import a2a_tools_identity
+
+        monkeypatch.setenv("WORKSPACE_ID", "ws-42")
+        monkeypatch.setattr(
+            a2a_tools_identity, "_check_memory_write_permission", lambda: False,
+        )
+
+        # Tripwire httpx — must not be called when RBAC denies.
+        import httpx
+
+        class _Tripwire:
+            def __init__(self, *_a, **_kw):
+                raise AssertionError("RBAC denial must short-circuit before httpx call")
+
+        monkeypatch.setattr(httpx, "AsyncClient", _Tripwire)
+
+        result = json.loads(
+            await a2a_tools_identity.tool_update_agent_card({"name": "x"}),
+        )
+        assert result["success"] is False
+        assert "memory.write" in str(result["error"]).lower()
+
+    @pytest.mark.asyncio
+    async def test_network_exception_returns_structured_error(
+        self, monkeypatch, _grant_memory_write,
+    ):
+        """A network exception (DNS failure, connect timeout, etc) is
+        wrapped into a structured error dict instead of bubbling up
+        to the MCP transport layer."""
+        import a2a_tools_identity
+
+        monkeypatch.setenv("WORKSPACE_ID", "ws-42")
+        monkeypatch.setattr(a2a_tools_identity, "PLATFORM_URL", "http://test.invalid")
+
+        class _ExplodingClient:
+            async def __aenter__(self):
+                return self
+
+            async def __aexit__(self, *_a):
+                return False
+
+            async def post(self, *_a, **_kw):
+                raise RuntimeError("simulated DNS failure")
+
+        monkeypatch.setattr(
+            a2a_tools_identity.httpx, "AsyncClient",
+            lambda *a, **kw: _ExplodingClient(),
+        )
+
+        result = json.loads(
+            await a2a_tools_identity.tool_update_agent_card({"name": "x"})
+        )
+        assert result["success"] is False
+        assert "network" in str(result["error"]).lower()
+
+
+# --- Registry contract ------------------------------------------------------
+
+
+class TestRegistryContract:
+    """Pin the new tools' registration in platform_tools.registry. The
+    structural tests in ``test_platform_tools.py`` already check
+    registry↔MCP alignment; these are tighter assertions specific to
+    the two new tools so a future contributor deleting one entry sees
+    a focused failure."""
+
+    def test_get_runtime_identity_in_registry(self):
+        from platform_tools.registry import by_name
+        spec = by_name("get_runtime_identity")
+        assert spec.section == "a2a"
+        # No input parameters — env-only call.
+        assert spec.input_schema == {"type": "object", "properties": {}}
+        # impl points at the actual tool function, not a shim.
+        from a2a_tools_identity import tool_get_runtime_identity
+        assert spec.impl is tool_get_runtime_identity
+
+    def test_update_agent_card_in_registry(self):
+        from platform_tools.registry import by_name
+        spec = by_name("update_agent_card")
+        assert spec.section == "a2a"
+        assert "card" in spec.input_schema["properties"]
+        assert spec.input_schema["required"] == ["card"]
+        from a2a_tools_identity import tool_update_agent_card
+        assert spec.impl is tool_update_agent_card
@@ -788,6 +788,123 @@ def test_sanitize_agent_error_stderr_combined_with_existing_tests():
    assert "workspace logs" in out


+# ─── reason passthrough (internal#211/#212: surface actionable provider error) ───
+
+
+def test_sanitize_agent_error_reason_surfaced_verbatim():
+    """A curated provider reason is shown to the user, not collapsed to the
+    exception class name. This is the internal#211 regression: a 403
+    org-disabled message must reach the canvas."""
+    reason = (
+        "provider HTTP 403 — oauth_org_not_allowed — Your organization has "
+        "disabled Claude subscription access for Claude Code · Use an "
+        "Anthropic API key instead, or ask your admin to enable access"
+    )
+
+    class _ResultErr(Exception):
+        pass
+
+    out = sanitize_agent_error(exc=_ResultErr("opaque"), reason=reason)
+    # The actionable provider guidance and status code must be visible.
+    assert "403" in out
+    assert "oauth_org_not_allowed" in out
+    assert "disabled Claude subscription access" in out
+    assert "ask your admin to enable access" in out
+    # NOT the old opaque form.
+    assert "see workspace logs" not in out
+
+
+def test_sanitize_agent_error_reason_still_scrubs_secrets():
+    """Even on the reason path the key/token scrubber runs — a buggy caller
+    that lets a bearer token into the reason still gets it redacted."""
+    leaky = (
+        "provider HTTP 401 — auth failed — Authorization: Bearer "
+        "PLACEHOLDER_LONG_TOKEN_0123456789abcdefghijklm please re-auth"
+    )
+    out = sanitize_agent_error(reason=leaky)
+    assert "[REDACTED]" in out
+    assert "PLACEHOLDER_LONG_TOKEN_0123456789abcdefghijklm" not in out
+    # The non-secret guidance still survives the scrub.
+    assert "401" in out
+    assert "please re-auth" in out
+
+
+def test_sanitize_agent_error_reason_scrubs_all_secret_formats():
+    """The scrubber must redact every realistic credential shape — not just
+    the `Bearer <tok>` form the original test happened to exercise
+    (internal#212 review finding: bare `sk-ant-api03-…` keys, JSON-quoted
+    "token"/"apiKey" values, and `aws_secret_access_key=` all leaked).
+    All curated/actionable guidance must still survive the scrub.
+    """
+    # 1. Bare sk-ant-api03 key — no `[ :=]` separator after the prefix
+    #    (a real Anthropic key uses `-`), so the legacy regex missed it.
+    bare = (
+        "provider HTTP 401 — auth failed — invalid key "
+        "sk-FAKEPLACEHOLDERabcdefghijklmnopqrstuvwxy0123456789 "
+        "please re-auth"
+    )
+    out = sanitize_agent_error(reason=bare)
+    assert "sk-FAKEPLACEHOLDERabcdefghijklmnopqrstuvwxy0123456789" not in out
+    assert "[REDACTED]" in out
+    assert "401" in out  # actionable status survives
+    assert "please re-auth" in out  # actionable guidance survives
+
+    # 2. JSON-quoted "token" / "apiKey" values.
+    jblob = (
+        'provider error — config dump {"token": '
+        '"abcDEF0123456789ghIJKL0123456789mnopQRST", "apiKey": '
+        '"anon_fakefakefakefakefakefakefakefakefakefake"} — '
+        "use an API key instead"
+    )
+    out = sanitize_agent_error(reason=jblob)
+    assert "abcDEF0123456789ghIJKL0123456789mnopQRST" not in out
+    assert "anon_fakefakefakefakefakefakefakefakefakefake" not in out
+    assert "[REDACTED]" in out
+    assert "use an API key instead" in out  # actionable guidance survives
+
+    # 3. aws_secret_access_key=… form.
+    awsblob = (
+        "provider HTTP 403 — boto credential error "
+        "aws_secret_access_key=wJalrXUtnFEMI/K7MDENG/bPxRfiCYEXAMPLEKEY — "
+        "ask your admin to enable access"
+    )
+    out = sanitize_agent_error(reason=awsblob)
+    assert "wJalrXUtnFEMI/K7MDENG/bPxRfiCYEXAMPLEKEY" not in out
+    assert "[REDACTED]" in out
+    assert "403" in out  # actionable status survives
+    assert "ask your admin to enable access" in out  # guidance survives
+
+    # 4. Regression: the original Bearer form still redacts.
+    # Uses PLACEHOLDER_LONG_TOKEN (>=40 chars, no sk-ant- prefix) to avoid
+    # triggering the secret-scan workflow pattern
+    # `sk-ant-[A-Za-z0-9_-]{40,}`.
+    bearer = (
+        "provider HTTP 401 — Authorization: Bearer "
+        "PLACEHOLDER_LONG_TOKEN_9876543210abcdefghij re-auth"
+    )
+    out = sanitize_agent_error(reason=bearer)
+    assert "PLACEHOLDER_LONG_TOKEN_9876543210abcdefghij" not in out
+    assert "[REDACTED]" in out
+    assert "re-auth" in out
+
+
+def test_sanitize_agent_error_reason_wins_over_stderr():
+    """When both reason and stderr are passed, the curated reason wins."""
+    out = sanitize_agent_error(
+        reason="provider HTTP 403 — use an API key",
+        stderr="raw subprocess noise that should not be shown",
+    )
+    assert "use an API key" in out
+    assert "raw subprocess noise" not in out
+
+
+def test_sanitize_agent_error_no_reason_unchanged():
+    """Omitting reason preserves the original generic behavior."""
+    out = sanitize_agent_error(exc=ValueError("boom"))
+    assert "ValueError" in out
+    assert "workspace logs" in out
+
+

 # ======================================================================
 # classify_subprocess_error