diff --git a/plugins/glean/.claude-plugin/plugin.json b/plugins/glean/.claude-plugin/plugin.json
index 301b08a..256dc47 100644
--- a/plugins/glean/.claude-plugin/plugin.json
+++ b/plugins/glean/.claude-plugin/plugin.json
@@ -1,6 +1,6 @@
 {
   "name": "glean-vnext",
-  "version": "0.2.36",
+  "version": "0.2.37",
   "description": "Glean plugin for discovering skills and running tools.",
   "author": {
     "name": "Glean"
diff --git a/plugins/glean/.codex-plugin/plugin.json b/plugins/glean/.codex-plugin/plugin.json
index 7b31deb..68fd1d3 100644
--- a/plugins/glean/.codex-plugin/plugin.json
+++ b/plugins/glean/.codex-plugin/plugin.json
@@ -1,6 +1,6 @@
 {
   "name": "glean-vnext",
-  "version": "0.2.36",
+  "version": "0.2.37",
   "description": "Glean Codex plugin for discovering skills and running tools.",
   "author": {
     "name": "Glean"
diff --git a/plugins/glean/.cursor-plugin/plugin.json b/plugins/glean/.cursor-plugin/plugin.json
index 232df50..e6b0d3c 100644
--- a/plugins/glean/.cursor-plugin/plugin.json
+++ b/plugins/glean/.cursor-plugin/plugin.json
@@ -1,7 +1,7 @@
 {
   "name": "glean-vnext",
   "displayName": "Glean vNext",
-  "version": "0.2.36",
+  "version": "0.2.37",
   "description": "Search and act across your company's apps — Jira, Slack, Salesforce, Google Workspace, and more — without leaving Cursor.",
   "author": {
     "name": "Glean"
diff --git a/plugins/glean/dist/index.js b/plugins/glean/dist/index.js
index b88e79b..27adc09 100644
--- a/plugins/glean/dist/index.js
+++ b/plugins/glean/dist/index.js
@@ -25662,6 +25662,7 @@ async function handleFindSkills(remoteClient, skillsBaseDir, args) {
 
 // src/tools/run-tool.ts
 import fs4 from "node:fs/promises";
+import os2 from "node:os";
 import path4 from "node:path";
 
 // src/tools/approval-args.ts
@@ -25907,6 +25908,20 @@ function primeElicitationCancellation(mcpServer) {
   void mcpServer.request({ method: "ping" }, EmptyResultSchema).catch(() => {
   });
 }
+function permissionModeMarkerPath() {
+  const base = process.env.CLAUDE_PLUGIN_DATA || path4.join(os2.homedir(), ".glean");
+  const sessionId = resolveSessionId().replace(/[^a-zA-Z0-9_-]/g, "-").slice(0, 64);
+  return path4.join(base, "glean-hitl-mode", `${sessionId}.json`);
+}
+async function currentPermissionMode() {
+  try {
+    const raw = await fs4.readFile(permissionModeMarkerPath(), "utf-8");
+    const parsed = JSON.parse(raw);
+    return typeof parsed.permission_mode === "string" ? parsed.permission_mode : null;
+  } catch {
+    return null;
+  }
+}
 async function handleRunTool(remoteClient, mcpServer, skillsBaseDir, args) {
   const serverId = args.server_id;
   const toolName = args.tool_name;
@@ -25937,8 +25952,9 @@ async function handleRunTool(remoteClient, mcpServer, skillsBaseDir, args) {
     throw err;
   }
   const hitlEnabled = process.env.ENABLE_HITL === "true";
-  if (hitlEnabled && mcpServer.getClientCapabilities()?.elicitation) {
-    if (toolMeta?.requires_approval) {
+  if (hitlEnabled && toolMeta?.requires_approval && mcpServer.getClientCapabilities()?.elicitation) {
+    const bypass = await currentPermissionMode() === "bypassPermissions";
+    if (!bypass) {
       const message = await buildApprovalMessage(
         mcpServer,
         toolName,
diff --git a/plugins/glean/hooks/auto-approve-run-tool.mjs b/plugins/glean/hooks/auto-approve-run-tool.mjs
index 1ef1585..7951f94 100644
--- a/plugins/glean/hooks/auto-approve-run-tool.mjs
+++ b/plugins/glean/hooks/auto-approve-run-tool.mjs
@@ -12,6 +12,7 @@
 // write. When ENABLE_HITL is not "true" the hook does nothing and the normal
 // permission flow runs.
 import fs from "node:fs";
+import os from "node:os";
 import path from "node:path";
 
 function readStdin() {
@@ -49,6 +50,34 @@ try {
 }
 
 if (env.ENABLE_HITL === "true") {
+  // Record Claude Code's live permission mode so the MCP server can skip its
+  // own elicitation gate when the user launched with
+  // --dangerously-skip-permissions (permission_mode "bypassPermissions").
+  // Written on every run_tool call and keyed by session id, so it is always
+  // fresh for the call that immediately follows and never leaks across
+  // sessions. The base dir and session-id sanitization MUST match
+  // run-tool.ts (permissionModeMarkerPath) and start.sh's PLUGIN_DATA_DIR:
+  // CLAUDE_PLUGIN_DATA when set, else ~/.glean. Best-effort — marker I/O must
+  // never break the approval decision below.
+  try {
+    const permissionMode = String(input.permission_mode ?? "");
+    const sessionId = String(input.session_id ?? "")
+      .replace(/[^a-zA-Z0-9_-]/g, "-")
+      .slice(0, 64);
+    if (permissionMode && sessionId) {
+      const base =
+        process.env.CLAUDE_PLUGIN_DATA || path.join(os.homedir(), ".glean");
+      const dir = path.join(base, "glean-hitl-mode");
+      fs.mkdirSync(dir, { recursive: true });
+      fs.writeFileSync(
+        path.join(dir, `${sessionId}.json`),
+        JSON.stringify({ permission_mode: permissionMode, ts: Date.now() }),
+      );
+    }
+  } catch {
+    // Ignore: a failed marker write just means the server keeps prompting.
+  }
+
   process.stdout.write(
     JSON.stringify({
       hookSpecificOutput: {
diff --git a/src/tools/run-tool.ts b/src/tools/run-tool.ts
index 27edacf..7b62228 100644
--- a/src/tools/run-tool.ts
+++ b/src/tools/run-tool.ts
@@ -3,9 +3,11 @@ import type { Server } from "@modelcontextprotocol/sdk/server/index.js";
 import type { CallToolResult, Tool } from "@modelcontextprotocol/sdk/types.js";
 import { EmptyResultSchema } from "@modelcontextprotocol/sdk/types.js";
 import fs from "node:fs/promises";
+import os from "node:os";
 import path from "node:path";
 import { callRemoteTool } from "../remote-client.js";
 import { buildCompactArgs, writeApprovalArgsFile } from "./approval-args.js";
+import { resolveSessionId } from "../session-id.js";
 
 const DEFAULT_FILE_ARG_MAX_BYTES = 1 * 1024 * 1024;
 
@@ -240,6 +242,45 @@ function primeElicitationCancellation(mcpServer: Server): void {
   });
 }
 
+// Path to the per-session permission-mode marker the PreToolUse hook writes
+// immediately before each run_tool call (see hooks/auto-approve-run-tool.mjs).
+// This resolution MUST match the hook's exactly. The hook cannot see the
+// server-only PLUGIN_DATA_DIR that start.sh derives, so both sides key off
+// CLAUDE_PLUGIN_DATA (falling back to ~/.glean) — the one anchor available to
+// both processes. Under start.sh, PLUGIN_DATA_DIR resolves to this same path.
+function permissionModeMarkerPath(): string {
+  const base =
+    process.env.CLAUDE_PLUGIN_DATA || path.join(os.homedir(), ".glean");
+  const sessionId = resolveSessionId()
+    .replace(/[^a-zA-Z0-9_-]/g, "-")
+    .slice(0, 64);
+  return path.join(base, "glean-hitl-mode", `${sessionId}.json`);
+}
+
+// Claude Code's live permission mode for THIS session, as captured by the hook
+// on the current call. Returns null when the marker is missing, unreadable, or
+// malformed — the caller treats null as "unknown" and keeps the approval gate,
+// so any failure fails toward prompting, never toward a silent bypass.
+//
+// Resume safety: the PreToolUse hook rewrites this marker with the CURRENT mode
+// on every run_tool call (see hooks/auto-approve-run-tool.mjs), and PreToolUse
+// always runs before the tool executes, so the value read here is the one
+// written for this exact call. A session first launched with
+// --dangerously-skip-permissions and later resumed WITHOUT it (same session id)
+// therefore has its stale bypass marker overwritten with the resumed mode on
+// the resumed session's first run_tool call, re-engaging the gate.
+async function currentPermissionMode(): Promise<string | null> {
+  try {
+    const raw = await fs.readFile(permissionModeMarkerPath(), "utf-8");
+    const parsed = JSON.parse(raw) as { permission_mode?: unknown };
+    return typeof parsed.permission_mode === "string"
+      ? parsed.permission_mode
+      : null;
+  } catch {
+    return null;
+  }
+}
+
 export async function handleRunTool(
   remoteClient: Client,
   mcpServer: Server,
@@ -288,8 +329,20 @@ export async function handleRunTool(
   }
 
   const hitlEnabled = process.env.ENABLE_HITL === "true";
-  if (hitlEnabled && mcpServer.getClientCapabilities()?.elicitation) {
-    if (toolMeta?.requires_approval) {
+  if (
+    hitlEnabled &&
+    toolMeta?.requires_approval &&
+    mcpServer.getClientCapabilities()?.elicitation
+  ) {
+    // In bypassPermissions mode (`claude --dangerously-skip-permissions`) the
+    // user has opted out of every approval prompt for the session, so our own
+    // elicitation gate is just a redundant popup — skip it and execute
+    // directly. The mode comes from the PreToolUse hook, which writes it keyed
+    // by session id immediately before this call, so it reflects the current
+    // call and never leaks across sessions. Any other or unknown mode keeps the
+    // gate. Only bypassPermissions is skipped (deliberately narrow).
+    const bypass = (await currentPermissionMode()) === "bypassPermissions";
+    if (!bypass) {
       const message = await buildApprovalMessage(
         mcpServer,
         toolName,
diff --git a/tests/auto-approve-hook.test.ts b/tests/auto-approve-hook.test.ts
index fbe7cf1..3b72cdc 100644
--- a/tests/auto-approve-hook.test.ts
+++ b/tests/auto-approve-hook.test.ts
@@ -11,27 +11,69 @@ const HOOK = path.resolve(
   "../plugins/glean/hooks/auto-approve-run-tool.mjs",
 );
 
+interface HookResult {
+  out: string;
+  // Parsed contents of the single permission-mode marker the hook wrote, or
+  // null when none was written. markerFiles lists the filenames present.
+  marker: { permission_mode?: string; ts?: number } | null;
+  markerFiles: string[];
+}
+
 async function runHook(
   toolName: string,
   env: Record<string, string>,
-): Promise<string> {
+  extraInput: Record<string, unknown> = {},
+  seed?: { sessionId: string; mode: string },
+): Promise<HookResult> {
   const root = await fs.mkdtemp(path.join(os.tmpdir(), "approve-hook-"));
   await fs.writeFile(
     path.join(root, ".mcp.json"),
     JSON.stringify({ mcpServers: { glean: { env } } }),
   );
+  // Isolate the marker under a throwaway CLAUDE_PLUGIN_DATA so the hook never
+  // touches the developer's real ~/.glean during tests.
+  const dataDir = path.join(root, "plugin-data");
+  // Optionally pre-seed a leftover marker (e.g. from a prior
+  // --dangerously-skip-permissions session) to prove the hook overwrites it.
+  if (seed) {
+    const dir = path.join(dataDir, "glean-hitl-mode");
+    await fs.mkdir(dir, { recursive: true });
+    await fs.writeFile(
+      path.join(dir, `${seed.sessionId}.json`),
+      JSON.stringify({ permission_mode: seed.mode, ts: 0 }),
+    );
+  }
   try {
-    return await new Promise<string>((resolve, reject) => {
+    const out = await new Promise<string>((resolve, reject) => {
       const child = spawn("node", [HOOK], {
-        env: { ...process.env, CLAUDE_PLUGIN_ROOT: root },
+        env: {
+          ...process.env,
+          CLAUDE_PLUGIN_ROOT: root,
+          CLAUDE_PLUGIN_DATA: dataDir,
+        },
       });
-      let out = "";
-      child.stdout.on("data", (d) => (out += d.toString()));
+      let o = "";
+      child.stdout.on("data", (d) => (o += d.toString()));
       child.on("error", reject);
-      child.on("close", () => resolve(out));
-      child.stdin.write(JSON.stringify({ tool_name: toolName }));
+      child.on("close", () => resolve(o));
+      child.stdin.write(JSON.stringify({ tool_name: toolName, ...extraInput }));
       child.stdin.end();
     });
+
+    let markerFiles: string[] = [];
+    let marker: HookResult["marker"] = null;
+    try {
+      const dir = path.join(dataDir, "glean-hitl-mode");
+      markerFiles = await fs.readdir(dir);
+      if (markerFiles.length) {
+        marker = JSON.parse(
+          await fs.readFile(path.join(dir, markerFiles[0]), "utf-8"),
+        );
+      }
+    } catch {
+      // No marker directory: nothing was written.
+    }
+    return { out, marker, markerFiles };
   } finally {
     await fs.rm(root, { recursive: true, force: true });
   }
@@ -40,25 +82,94 @@ async function runHook(
 const glean = (tool: string) => `mcp__plugin_glean-vnext_glean__${tool}`;
 const hitlOn = { ENABLE_HITL: "true" };
 const hitlOff = { ENABLE_HITL: "false" };
+const bypass = { permission_mode: "bypassPermissions", session_id: "sess-1" };
 
 describe("auto-approve-run-tool hook (Claude Code PreToolUse)", () => {
   it("allows glean run_tool when HITL is on", async () => {
-    const out = await runHook(glean("run_tool"), hitlOn);
+    const { out } = await runHook(glean("run_tool"), hitlOn);
     expect(JSON.parse(out).hookSpecificOutput.permissionDecision).toBe("allow");
   });
 
   it("never allows when HITL is off (safety)", async () => {
-    const out = await runHook(glean("run_tool"), hitlOff);
+    const { out } = await runHook(glean("run_tool"), hitlOff);
     expect(out.trim()).toBe("");
   });
 
   it("ignores a non-glean run_tool (scoped to this plugin)", async () => {
-    const out = await runHook("mcp__other-server__run_tool", hitlOn);
+    const { out } = await runHook("mcp__other-server__run_tool", hitlOn);
     expect(out.trim()).toBe("");
   });
 
   it("ignores glean tools other than run_tool (e.g. find_skills)", async () => {
-    const out = await runHook(glean("find_skills"), hitlOn);
+    const { out } = await runHook(glean("find_skills"), hitlOn);
+    expect(out.trim()).toBe("");
+  });
+});
+
+describe("auto-approve-run-tool hook (permission-mode marker)", () => {
+  it("records the permission_mode marker for run_tool when HITL is on", async () => {
+    const { marker, markerFiles } = await runHook(
+      glean("run_tool"),
+      hitlOn,
+      bypass,
+    );
+    expect(marker).toMatchObject({ permission_mode: "bypassPermissions" });
+    expect(typeof marker?.ts).toBe("number");
+    expect(markerFiles).toContain("sess-1.json");
+  });
+
+  it("keys the marker file by session id (parallel sessions don't collide)", async () => {
+    const { markerFiles } = await runHook(glean("run_tool"), hitlOn, {
+      permission_mode: "default",
+      session_id: "other-session",
+    });
+    expect(markerFiles).toEqual(["other-session.json"]);
+  });
+
+  it("does not write a marker when HITL is off", async () => {
+    const { out, marker } = await runHook(glean("run_tool"), hitlOff, bypass);
     expect(out.trim()).toBe("");
+    expect(marker).toBeNull();
+  });
+
+  it("does not write a marker for a non-glean run_tool", async () => {
+    const { marker } = await runHook(
+      "mcp__other-server__run_tool",
+      hitlOn,
+      bypass,
+    );
+    expect(marker).toBeNull();
+  });
+
+  it("writes no marker when permission_mode is absent from the payload", async () => {
+    const { out, marker } = await runHook(glean("run_tool"), hitlOn, {
+      session_id: "sess-1",
+    });
+    // Still auto-approves, just has no mode to record.
+    expect(JSON.parse(out).hookSpecificOutput.permissionDecision).toBe("allow");
+    expect(marker).toBeNull();
+  });
+
+  it("sanitizes the session id used for the marker filename", async () => {
+    const { markerFiles } = await runHook(glean("run_tool"), hitlOn, {
+      permission_mode: "default",
+      session_id: "weird/../id with spaces",
+    });
+    expect(markerFiles).toHaveLength(1);
+    expect(markerFiles[0]).toMatch(/^[a-zA-Z0-9_-]+\.json$/);
+  });
+
+  it("overwrites a leftover bypass marker when the session is resumed without the flag", async () => {
+    // Session was first launched with --dangerously-skip-permissions (leftover
+    // marker = bypassPermissions), then resumed WITHOUT the flag (current mode
+    // = default). The hook rewrites the same per-session marker, clearing the
+    // stale bypass so the server re-engages its gate on this call.
+    const { marker } = await runHook(
+      glean("run_tool"),
+      hitlOn,
+      { permission_mode: "default", session_id: "sess-1" },
+      { sessionId: "sess-1", mode: "bypassPermissions" },
+    );
+    expect(marker).toMatchObject({ permission_mode: "default" });
   });
 });
diff --git a/tests/run-tool.test.ts b/tests/run-tool.test.ts
index 468db8e..91ae09e 100644
--- a/tests/run-tool.test.ts
+++ b/tests/run-tool.test.ts
@@ -283,6 +283,22 @@ async function writeToolJson(
   );
 }
 
+// Mirrors the marker the PreToolUse hook writes: <dataDir>/glean-hitl-mode/
+// <sessionId>.json. The server reads it via CLAUDE_PLUGIN_DATA + GLEAN_SESSION_ID.
+async function writeModeMarker(
+  dataDir: string,
+  sessionId: string,
+  mode: string,
+) {
+  const dir = path.join(dataDir, "glean-hitl-mode");
+  await fs.mkdir(dir, { recursive: true });
+  await fs.writeFile(
+    path.join(dir, `${sessionId}.json`),
+    JSON.stringify({ permission_mode: mode, ts: Date.now() }),
+    "utf-8",
+  );
+}
+
 describe("handleRunTool (HITL)", () => {
   let tmpDir: string;
   const baseArgs = {
@@ -553,6 +569,69 @@ describe("handleRunTool (HITL)", () => {
     expect(elicit).not.toHaveBeenCalled(); // no prompt for unreadable input
     expect(remote.callTool).not.toHaveBeenCalled();
   });
+
+  it("skips the elicitation gate and executes directly in bypassPermissions mode", async () => {
+    vi.stubEnv("ENABLE_HITL", "true");
+    vi.stubEnv("CLAUDE_PLUGIN_DATA", tmpDir);
+    vi.stubEnv("GLEAN_SESSION_ID", "sess-bypass");
+    await writeToolJson(tmpDir, "jirasearch", { requires_approval: true });
+    await writeModeMarker(tmpDir, "sess-bypass", "bypassPermissions");
+    const remote = makeRemote();
+    const elicit = vi.fn().mockResolvedValue({ action: "accept" });
+    const server = makeServer({ elicitation: true, elicit });
+
+    await handleRunTool(remote, server, tmpDir, baseArgs);
+
+    expect(elicit).not.toHaveBeenCalled();
+    expect(remote.callTool).toHaveBeenCalledTimes(1);
+  });
+
+  it("still elicits when the session's permission mode is not bypass", async () => {
+    vi.stubEnv("ENABLE_HITL", "true");
+    vi.stubEnv("CLAUDE_PLUGIN_DATA", tmpDir);
+    vi.stubEnv("GLEAN_SESSION_ID", "sess-default");
+    await writeToolJson(tmpDir, "jirasearch", { requires_approval: true });
+    await writeModeMarker(tmpDir, "sess-default", "default");
+    const remote = makeRemote();
+    const elicit = vi.fn().mockResolvedValue({ action: "accept" });
+    const server = makeServer({ elicitation: true, elicit });
+
+    await handleRunTool(remote, server, tmpDir, baseArgs);
+
+    expect(elicit).toHaveBeenCalledTimes(1);
+    expect(remote.callTool).toHaveBeenCalledTimes(1);
+  });
+
+  it("still elicits when no permission-mode marker exists (fails toward the gate)", async () => {
+    vi.stubEnv("ENABLE_HITL", "true");
+    vi.stubEnv("CLAUDE_PLUGIN_DATA", tmpDir);
+    vi.stubEnv("GLEAN_SESSION_ID", "sess-none");
+    await writeToolJson(tmpDir, "jirasearch", { requires_approval: true });
+    // Deliberately write no marker.
+    const remote = makeRemote();
+    const elicit = vi.fn().mockResolvedValue({ action: "accept" });
+    const server = makeServer({ elicitation: true, elicit });
+
+    await handleRunTool(remote, server, tmpDir, baseArgs);
+
+    expect(elicit).toHaveBeenCalledTimes(1);
+  });
+
+  it("ignores a bypass marker written for a different session (no cross-session leak)", async () => {
+    vi.stubEnv("ENABLE_HITL", "true");
+    vi.stubEnv("CLAUDE_PLUGIN_DATA", tmpDir);
+    vi.stubEnv("GLEAN_SESSION_ID", "sess-A");
+    await writeToolJson(tmpDir, "jirasearch", { requires_approval: true });
+    // Another concurrent session opted into bypass; ours did not.
+    await writeModeMarker(tmpDir, "sess-B", "bypassPermissions");
+    const remote = makeRemote();
+    const elicit = vi.fn().mockResolvedValue({ action: "accept" });
+    const server = makeServer({ elicitation: true, elicit });
+
+    await handleRunTool(remote, server, tmpDir, baseArgs);
+
+    expect(elicit).toHaveBeenCalledTimes(1); // gate preserved for THIS session
+  });
 });
 
 describe("buildCompactArgs", () => {