diff --git a/README.md b/README.md
index 876662f..2f0f09d 100644
--- a/README.md
+++ b/README.md
@@ -133,6 +133,9 @@ Use it when you want OpenCode to:
 > [!NOTE]
 > Depending on the task and the model you choose these tasks might take a long time and it's generally recommended to force the task to be in the background or move the agent to the background.
 
+> [!IMPORTANT]
+> Rescue tasks default to a write-capable run using OpenCode's stock `build` agent under your own OpenCode permission configuration. Permission categories OpenCode gates behind an approval prompt (external-directory access, `.env` reads, doom-loop protection) are automatically **denied** in these headless runs — the plugin never approves a gated request on your behalf. Note that OpenCode permissions are approval-level controls, not an operating-system sandbox; run the OpenCode server in a container or as a restricted user if you need a hard filesystem boundary.
+
 It supports `--background`, `--wait`, `--resume`, and `--fresh`. If you omit `--resume` and `--fresh`, the plugin can offer to continue the latest rescue thread for this repo.
 
 Examples:
diff --git a/docs/opencode-adaptation-plan.md b/docs/opencode-adaptation-plan.md
index 46ebd18..0e4f03a 100644
--- a/docs/opencode-adaptation-plan.md
+++ b/docs/opencode-adaptation-plan.md
@@ -108,6 +108,7 @@ Claude Code
 ### Read-only vs write (the `sandbox` analog)
 - **Review / read-only task:** send the turn with `agent: "plan"` (OpenCode's read-only agent) and/or create the session with `permission` rules denying edits.
 - **`--write` task:** `agent: "build"` and create the session with `permission` rules set to `allow` for the edit/shell tools (equivalent to Codex `approvalPolicy: "never"` + `workspace-write`). In headless mode there's no human to approve, so the client must **auto-approve**: either pre-set `permission` on `POST /session`, or reply `allow` to `permission.asked` / `permission.v2.asked` events via `POST /session/{id}/permissions/{permID}`. Design the client to auto-approve in write mode and auto-deny-edits in read-only mode. **This is the single most important implementation detail to get right.**
+  - **Superseded (issue #26):** the auto-approve design above disables OpenCode's own safety guards. Session-level rules are merged *after* the agent ruleset with last-match-wins, so a broad session allow strips the stock `build` agent's `external_directory`/`.env`/`doom_loop` guards. The implementation now sends **no** session `permission` rules and **rejects** every headless `permission.asked` — under the stock agents an ask only fires when a guard trips.
 
 ### Model / effort mapping
 - `--model spark` → `openai/gpt-5.3-codex-spark`; `--model openai/gpt-5.4` → `{providerID:"openai", modelID:"gpt-5.4"}`; unset → OpenCode default. Add an alias map + a `provider/model` splitter.
diff --git a/plugins/opencode/scripts/lib/opencode.mjs b/plugins/opencode/scripts/lib/opencode.mjs
index 7af1f15..e80a337 100644
--- a/plugins/opencode/scripts/lib/opencode.mjs
+++ b/plugins/opencode/scripts/lib/opencode.mjs
@@ -103,14 +103,6 @@ function buildTaskSessionName(prompt) {
   return excerpt ? `${TASK_SESSION_PREFIX}: ${excerpt}` : TASK_SESSION_PREFIX;
 }
 
-function buildWritePermissionRules() {
-  // OpenCode's PermissionRule.permission is a free-form string but only real
-  // permission keys take effect, and the built-in `build` agent's own defaults
-  // still leave some categories on "ask". Use the same wildcard-allow rule the
-  // `build` agent ships with so headless write turns never stall on approval.
-  return [{ permission: "*", action: "allow", pattern: "*" }];
-}
-
 function buildCreateSessionParams(cwd, options = {}) {
   const write = Boolean(options.write);
   const agent = options.agent ?? (write ? WRITE_AGENT : READ_ONLY_AGENT);
@@ -119,14 +111,22 @@ function buildCreateSessionParams(cwd, options = {}) {
   // BadRequest — the model is selected per-message instead (buildMessageParams).
   // `directory` is not accepted either: the session inherits it from the
   // `opencode serve` working directory, which server-lifecycle spawns with
-  // `cwd`. `title` must be a string when present. Read-only turns rely on the
-  // read-only `plan` agent instead of a permission override.
+  // `cwd`. `title` must be a string when present.
+  //
+  // Deliberately no session-level `permission` rules for either mode. OpenCode
+  // appends session rules AFTER the agent's ruleset and resolves each request
+  // with the LAST matching rule, so any broad session rule silently overrides
+  // the stock agents' safety guards (`external_directory`, `.env` reads, and
+  // `doom_loop` stay on "ask") — and a session-level guard would likewise
+  // clobber the agent's allowances for OpenCode's own tool-output directories.
+  // Write turns rely on the stock `build` agent, read-only turns on the
+  // read-only `plan` agent; guard categories that reach "ask" are denied
+  // headlessly in respondToPermission (issue #26).
   const rawTitle = options.title ?? options.threadName ?? null;
   const title = typeof rawTitle === "string" && rawTitle.trim() ? rawTitle : null;
   return {
     agent,
-    ...(title ? { title } : {}),
-    ...(write ? { permission: buildWritePermissionRules() } : {})
+    ...(title ? { title } : {})
   };
 }
 
@@ -378,8 +378,7 @@ function createTurnCaptureState(sessionID, options = {}) {
     recoveryError: null,
     response: null,
     fallbackTimer: null,
-    onProgress: options.onProgress ?? null,
-    write: Boolean(options.write)
+    onProgress: options.onProgress ?? null
   };
 }
 
@@ -495,20 +494,38 @@ function applyMessageParts(state, parts, sessionID) {
   }
 }
 
+function describePermissionRequest(event) {
+  const category =
+    (typeof event?.permission === "string" ? event.permission : null) ??
+    (typeof event?.properties?.permission === "string" ? event.properties.permission : null);
+  const rawPatterns = event?.patterns ?? event?.properties?.patterns ?? null;
+  const patterns = Array.isArray(rawPatterns) ? rawPatterns.filter((value) => typeof value === "string") : [];
+  if (!category) {
+    return null;
+  }
+  return patterns.length > 0 ? `${category}: ${patterns.join(", ")}` : category;
+}
+
 async function respondToPermission(client, state, event, sessionID) {
   const permissionID = extractPermissionId(event);
   if (!permissionID || !sessionID) {
     return;
   }
 
-  const response = state.write ? "always" : "reject";
+  // Never self-approve, in write mode included. Headless turns have no human
+  // to ask, and under the stock agents a request only reaches "ask" when a
+  // safety guard trips (external-directory access, `.env` reads, doom-loop
+  // protection) or the user configured a category as interactive. Rejecting is
+  // the only answer that preserves those guards; the model sees the rejection
+  // and adapts (issue #26).
+  const described = describePermissionRequest(event);
   emitProgress(
     state.onProgress,
-    `${state.write ? "Allowing" : "Denying"} OpenCode permission request ${permissionID}.`,
-    state.write ? "running" : "investigating"
+    `Denying OpenCode permission request ${permissionID}${described ? ` (${described})` : ""}: headless runs never self-approve gated permissions.`,
+    "running"
   );
   try {
-    await client.respondPermission(sessionID, permissionID, response);
+    await client.respondPermission(sessionID, permissionID, "reject");
   } catch (error) {
     state.error = error;
     emitProgress(state.onProgress, `OpenCode permission response failed: ${error.message}`, "failed");
@@ -1118,7 +1135,6 @@ export async function runServerTurn(cwd, options = {}) {
           ),
         {
           onProgress: options.onProgress,
-          write,
           resumed: resumedSession,
           turnTimeoutMs: options.turnTimeoutMs
         }
diff --git a/tests/fake-opencode-fixture.mjs b/tests/fake-opencode-fixture.mjs
index 24cee8e..d4d8af6 100644
--- a/tests/fake-opencode-fixture.mjs
+++ b/tests/fake-opencode-fixture.mjs
@@ -159,11 +159,11 @@ function structuredOutputParts(body) {
 }
 
 async function waitForPermission(permissionID) {
-  await new Promise((resolve) => {
-    const timeout = setTimeout(resolve, 2000);
-    pendingPermissions.set(permissionID, () => {
+  return new Promise((resolve) => {
+    const timeout = setTimeout(() => resolve(null), 2000);
+    pendingPermissions.set(permissionID, (reply) => {
       clearTimeout(timeout);
-      resolve();
+      resolve(reply ?? null);
     });
   });
 }
@@ -185,10 +185,23 @@ async function handleMessage(req, res, sessionID) {
 
   emit({ type: "session.next.step.started", sessionID });
   if (session.agent === "build" || body.agent === "build") {
-    const permissionID = "perm_" + messageID;
-    emit({ type: "permission.asked", sessionID, permissionID, permission: { id: permissionID, tool: "edit" } });
-    await waitForPermission(permissionID);
+    // Workspace edits are covered by the stock build agent's wildcard allow
+    // and never produce a permission round-trip.
     emit({ type: "file.edited", sessionID, path: "generated.txt" });
+    // A guard category (e.g. external_directory) reaches "ask". The companion
+    // must deny it; only an (incorrect) approval lets the gated edit proceed.
+    const permissionID = "perm_" + messageID;
+    emit({
+      type: "permission.asked",
+      sessionID,
+      permissionID,
+      permission: { id: permissionID, tool: "edit" },
+      patterns: ["/outside/workspace/secret.txt"]
+    });
+    const reply = await waitForPermission(permissionID);
+    if (reply && reply.response !== "reject") {
+      emit({ type: "file.edited", sessionID, path: "/outside/workspace/secret.txt" });
+    }
   }
 
   const finalText = prompt.includes("follow up")
@@ -448,7 +461,7 @@ const server = http.createServer(async (req, res) => {
     saveState(state);
     const resolve = pendingPermissions.get(permissionID);
     pendingPermissions.delete(permissionID);
-    resolve?.();
+    resolve?.(body);
     sendJson(res, { ok: true });
     return;
   }
diff --git a/tests/runtime.test.mjs b/tests/runtime.test.mjs
index c98e185..12fc5ad 100644
--- a/tests/runtime.test.mjs
+++ b/tests/runtime.test.mjs
@@ -659,7 +659,7 @@ test("foreground task runs through opencode serve and stores a visible session",
   }
 });
 
-test("write task auto-allows headless permission prompts and records touched files", { skip: LOCAL_LISTEN_SKIP }, () => {
+test("write task denies gated permission asks and keeps stock agent guards (issue #26)", { skip: LOCAL_LISTEN_SKIP }, () => {
   const repo = makeTempDir();
   const binDir = makeTempDir();
   installFakeOpencode(binDir);
@@ -674,17 +674,18 @@ test("write task auto-allows headless permission prompts and records touched fil
 
     assert.equal(result.status, 0, result.stderr);
     const payload = JSON.parse(result.stdout);
+    // Only the ungated workspace edit lands; the gated out-of-workspace edit
+    // stays denied, so it never shows up in touched files.
     assert.deepEqual(payload.touchedFiles, ["generated.txt"]);
 
     const fakeState = readFakeState(binDir);
     assert.equal(fakeState.sessions[0].agent, "build");
+    // No session-level permission override: the stock build agent's ask-guards
+    // (external_directory, .env reads, doom_loop) must stay in effect.
+    assert.deepEqual(fakeState.sessions[0].permission, []);
     assert.equal(fakeState.permissions.length, 1);
-    assert.equal(fakeState.permissions[0].body.response, "always");
+    assert.equal(fakeState.permissions[0].body.response, "reject");
     assert.equal(fakeState.permissions[0].body.action, undefined);
-    assert.deepEqual(
-      fakeState.sessions[0].permission.filter((rule) => rule.action === "allow").map((rule) => rule.permission).sort(),
-      ["*"]
-    );
   } finally {
     cleanupServer(repo, env);
   }