docs: fold Task 7 security findings into the plan

Caret and control characters added to the filename guard, file-keyed in-flight
registry replacing the row-scoped one, try/catch around spawn, single finish()
exit path, UTF-8 chunk decoding, bounded output buffer.
This commit is contained in:
meels
2026-07-28 14:39:52 +02:00
parent fa1e14c8d1
commit 84fe67632a

View File

@@ -1447,6 +1447,14 @@ test("isSafeFilename rejects shell metacharacters", () => {
}
});
test("isSafeFilename rejects the cmd.exe escape character and control characters", () => {
// `^` escapes the next character in cmd.exe, so it can defuse the closing
// quote. NUL additionally makes spawn() throw synchronously.
for (const bad of ["a^b.md", "a\u0000b.md", "a\u001bb.md", "a\u007fb.md"]) {
assert.equal(isSafeFilename(bad), false, `should reject: ${JSON.stringify(bad)}`);
}
});
test("isSafeFilename rejects path traversal", () => {
assert.equal(isSafeFilename("../secrets.md"), false);
assert.equal(isSafeFilename("a/../../b.md"), false);
@@ -1476,9 +1484,17 @@ Insert into `main.js` below `reconcileConcepts`:
```js
// `shell: true` is required on Windows to resolve `claude.cmd`, which puts the
// filename into a shell string. Reject anything cmd.exe or a POSIX shell would
// interpret. Apostrophes and cyrillic are safe inside double quotes and are
// present in real filenames, so they stay allowed.
const UNSAFE_CHARS = /["`$&|;<>%\r\n]/;
// interpret, plus every control character.
//
// `^` is cmd.exe's escape character and belongs to the same class Node escapes
// in its CVE-2024-27980 mitigation for exactly this spawn-through-.cmd shape.
// Control characters are rejected because a NUL byte makes spawn() throw
// synchronously, which would otherwise strand the row mid-run.
//
// Apostrophes, spaces, cyrillic, em dashes and `!` stay allowed — they are safe
// inside double quotes and appear in real filenames in this vault. (`!` would
// matter only under `setlocal enabledelayedexpansion`, which is not in play.)
const UNSAFE_CHARS = /["`$&|;<>%^\u0000-\u001f\u007f]/;
function isSafeFilename(name) {
if (typeof name !== "string" || name.length === 0) return false;
@@ -1496,7 +1512,7 @@ Add `isSafeFilename` to the `__test__` export object.
node --test test/safety.test.js
```
Expected: PASS, 4 tests.
Expected: PASS, 5 tests.
- [ ] **Step 5: Add ingest-state styles**
@@ -1519,6 +1535,19 @@ Append to `styles.css`:
`getBasePath()` exists on `FileSystemAdapter`; on mobile the adapter is a different class and `child_process` is unavailable, which is why the manifest sets `isDesktopOnly`.
```js
Add the in-flight registry as the first line of `onload()`, before the code-block
processor is registered. It lives on the plugin instance so it survives the row
re-renders that a DOM-scoped guard cannot:
```js
async onload() {
this.running = new Set();
// ... existing registerMarkdownCodeBlockProcessor call follows unchanged
```
Then add both methods to the class:
```js
vaultPath() {
const adapter = this.app.vault.adapter;
@@ -1527,11 +1556,19 @@ Append to `styles.css`:
}
runIngest(file, rowEl) {
if (rowEl.dataset.wdRunning === "1") return;
const Notice = OB ? OB.Notice : null;
const notify = (msg) => { if (Notice) new Notice(msg); };
// In-flight state is keyed on the file, not the row. Obsidian rebuilds the
// row on every re-render, so a row-scoped guard would let a re-render hand
// out a fresh row whose guard is unset — and a second click would then run
// a second unsupervised agent against the same file, concurrently writing
// the same wiki pages as the first.
if (this.running.has(file.path)) {
notify(`Already ingesting ${file.name}.`);
return;
}
if (!isSafeFilename(file.name)) {
rowEl.createDiv({
cls: "wd-output",
@@ -1554,54 +1591,76 @@ Append to `styles.css`:
return;
}
rowEl.dataset.wdRunning = "1";
// A retry reuses the same row. Clear the previous run's status and output
// so they are replaced rather than stacked on top of each other.
rowEl.querySelectorAll(".wd-status, .wd-output").forEach((el) => el.remove());
const button = rowEl.querySelector("button");
if (button) button.disabled = true;
this.running.add(file.path);
const status = rowEl.createSpan({ cls: "wd-status wd-status-running", text: "running 0s" });
const output = rowEl.createDiv({ cls: "wd-output", text: "" });
const started = Date.now();
const timer = window.setInterval(() => {
status.setText(`running ${Math.round((Date.now() - started) / 1000)}s`);
}, 1000);
this.registerInterval(timer);
const output = rowEl.createDiv({ cls: "wd-output", text: "" });
// Bounded as it accumulates, not only when displayed, so a long or noisy
// run cannot grow this string without limit.
let buffered = "";
const append = (chunk) => {
buffered += chunk.toString();
output.setText(buffered.slice(-4000));
const append = (text) => {
buffered = (buffered + text).slice(-8000);
output.setText(buffered);
output.scrollTop = output.scrollHeight;
};
const child = spawn("claude", ["-p", `ingest "${file.name}"`], {
// Single exit path: every way this run can end clears the timer, releases
// the file, and re-enables the button.
const finish = (cls, label) => {
window.clearInterval(timer);
this.running.delete(file.path);
status.className = `wd-status ${cls}`;
status.setText(label);
if (button) button.disabled = false;
};
let child;
try {
child = spawn("claude", ["-p", `ingest "${file.name}"`], {
cwd: base,
shell: true,
});
} catch (err) {
// spawn() throws synchronously for some argument shapes. Without this the
// timer would run forever and the row would stay disabled until reload.
append(`\nCould not start claude: ${err.message}`);
finish("wd-status-failed", "failed");
return;
}
// Decode as UTF-8 across chunk boundaries. Raw Buffer chunks split wherever
// the OS buffer ends, and this vault's output is full of Cyrillic and em
// dashes that would otherwise decode as replacement characters.
child.stdout.setEncoding("utf8");
child.stderr.setEncoding("utf8");
child.stdout.on("data", append);
child.stderr.on("data", append);
child.on("error", (err) => {
window.clearInterval(timer);
status.className = "wd-status wd-status-failed";
status.setText("failed");
append(`\nCould not start claude: ${err.message}\nIs it on PATH?`);
rowEl.dataset.wdRunning = "0";
if (button) button.disabled = false;
finish("wd-status-failed", "failed");
});
child.on("close", (code) => {
window.clearInterval(timer);
const secs = Math.round((Date.now() - started) / 1000);
if (code === 0) {
status.className = "wd-status wd-status-done";
status.setText(`done in ${secs}s`);
finish("wd-status-done", `done in ${secs}s`);
notify(`Ingested ${file.name}. Reopen the dashboard to refresh.`);
} else {
status.className = "wd-status wd-status-failed";
status.setText(`failed - exit ${code}`);
finish("wd-status-failed", `failed - exit ${code}`);
}
rowEl.dataset.wdRunning = "0";
if (button) button.disabled = false;
});
}
```
@@ -1650,7 +1709,7 @@ Bare `node --test` auto-discovers the test files. Do not pass `test/` as an
argument — Node 24 resolves a bare directory path as a module and fails with
`MODULE_NOT_FOUND` before running anything.
Expected: PASS, 26 tests across three files.
Expected: PASS, 27 tests across three files.
- [ ] **Step 11: Commit**
@@ -1683,4 +1742,4 @@ The spec's "v1 does not write the coverage file from the plugin" is honored —
**Type consistency.** `rawFiles` items are `{path, name, size}` in Tasks 2 and 3. `sourcePages` items are `{path, name, rawPath}` in both. Coverage rows are `{concept, status, stations, pinned, line}` in Tasks 4, 5, and 6. `onIngest(file, rowEl)` is declared in Task 3 and implemented with the same signature in Task 7. `addIcon(parent, name)` is defined in Task 3 and reused in Task 6. `STATIONS` is defined in Task 1 and consumed by `groupByStation` in Task 4.
**Test count.** Task 2 adds 9, Task 4 adds 11, Task 6 adds 2, Task 7 adds 4 — 26 total, matching Step 10 of Task 7.
**Test count.** Task 2 adds 9, Task 4 adds 11, Task 6 adds 2, Task 7 adds 5 — 27 total, matching Step 10 of Task 7.