Reveal mode and highlights were replacing text in sidebars, navigation,
headers, repo names, and other non-chat UI elements. Now skips:
- <nav>, <aside>, <header>, <footer> elements
- Elements with role="navigation", role="banner", role="complementary"
- Elements with class names containing sidebar, nav, menu, header
Added isInNonChatArea() check to all five code paths:
- revealInElement, unrevealInElement, highlightMatches
- MutationObserver addedNodes and characterData handlers
Also:
- Removed footer Options link (redundant with 5th Options tab)
- Updated README: ChatGPT marked as Tested
https://claude.ai/code/session_01SWSwDfMVij53bCTNSCLMwn
Accept (+) button:
- Was reading raw storage (getStorageData) which returns encrypted
blobs when encryption is enabled, silently failing to add mappings
- Fixed: adds directly to the local mappings array and persists via
setStorageData (storage bridge handles encryption transparently)
Auto-detect false positives:
- Was only checking identity values, not explicit mappings — values
already in the mappings table still got flagged as unconfigured PPI
- Fixed: now adds all mapping real/substitute values to the skip set
Ignore button:
- Changed from plain text link to grey pill button for better UX
- Still persists permanently via ss_ignored_ppi in storage
https://claude.ai/code/session_01SWSwDfMVij53bCTNSCLMwn
When at-rest encryption is enabled, storage.local contains encrypted
blobs. The injector reads raw storage (content script world, no
access to IndexedDB CryptoKey) and sees { _ssLocalEncrypted: true }.
It passed empty config to content.js → no mappings → no substitution.
This is why Firefox stopped working after encryption was enabled.
Chrome/Brave worked because the user hadn't set up encryption there.
Fixed: injector now detects encrypted data and asks the background
script for decrypted config via 'get:decrypted-config' message.
The background uses the Storage module (which has IndexedDB access)
to decrypt and return the data. Falls back to empty config if the
vault is actually locked.
https://claude.ai/code/session_01SWSwDfMVij53bCTNSCLMwn
The early hook approach (inline script, then external early-hook.js)
kept breaking Firefox due to CSP restrictions and timing issues with
the async external script load. Each fix for ChatGPT introduced a
new regression for Firefox/Brave.
Reverted to the original simple approach:
- Injector reads storage, injects content.js via <script src="...">
- content.js captures window.fetch at load time and patches it
- No inline scripts, no early hooks, no __ssOriginalFetch globals
This is what worked on Claude.ai across all browsers before the
ChatGPT fix attempts. ChatGPT support may need a different approach
later (possibly using declarativeNetRequest for header-only changes,
or a ChatGPT-specific content script), but it should not break the
core functionality on Claude.ai.
Kept the Request object handling in the fetch interceptor (needed for
some frameworks) but removed all early hook dependencies.
https://claude.ai/code/session_01SWSwDfMVij53bCTNSCLMwn
The inline <script> with textContent was blocked by claude.ai's
Content Security Policy on Firefox (Chrome is more permissive).
No fetch interception = no substitution = completely broken on FF.
Fixed by moving the early fetch hook to its own file (early-hook.js)
loaded via <script src="..."> which is CSP-compliant. Added to
web_accessible_resources in both Chrome and Firefox manifests.
Also fixed duplicate 'const api' declaration in injector.js that
would have crashed the content script.
https://claude.ai/code/session_01SWSwDfMVij53bCTNSCLMwn
Bug 1 - Reveal mode not working (all browsers):
sessionSubstitutions stored "ademo demo" (full name) as key, but
buildRevealPairs looked up "ademo" and "demo" individually. No match,
no reveal pairs, reveal did nothing. Fixed by also storing individual
words from multi-word replacements so both "ademo demo" AND "ademo"
AND "demo" are in the map.
Bug 2 - Missing await on _handleDecryptedMeta (sync.js):
Two call sites returned the Promise instead of the resolved value.
Downstream code checking decResult.data got undefined. Added await.
Bug 3 - Profile selector broken by safeHTML (popup.js):
DOMParser.parseFromString wraps content in <html><body> which
mangles <option> elements when moved to a <select>. Replaced with
new Option() DOM API which creates proper option elements.
https://claude.ai/code/session_01SWSwDfMVij53bCTNSCLMwn
Root cause: when importing an encrypted sync code from another device,
_decryptFromSync was replacing the local encryption config (salt) with
the source device's salt. This caused:
1. Local key cache derived from wrong salt
2. Local data (encrypted with local salt) became unreadable
3. _applyData tried to write with the wrong key
Fixed with a complete refactor of cross-device decryption:
- authenticateForSync() derives a TEMPORARY key using the source salt
- Temporary key stored separately as 'tempSyncKey' in IndexedDB
- Local encryption config and cached key are NEVER modified
- After decryption, _applyData writes via _writeSecure using the
LOCAL key (which uses the local salt)
- _handleDecryptedMeta extracted for code reuse
Also:
- Options.js auth handler detects pending sync import and routes to
authenticateForSync instead of regular authenticate
- After auth success, automatically retries the import
- README updated: imported passwords are protected (dots in UI,
vault password to reveal, AES-256 encrypted at rest)
- Bumped to v2.0.9
https://claude.ai/code/session_01SWSwDfMVij53bCTNSCLMwn
When importing an encrypted sync code, the flow was:
1. Click Apply → importSyncCode returns needsAuth
2. Auth prompt shown → user enters password → authentication succeeds
3. Green checkmark shown... but nobody re-runs the import
The user had to manually click Apply a second time. Now the import
automatically retries after successful authentication via a
__ssPendingSyncImport callback.
https://claude.ai/code/session_01SWSwDfMVij53bCTNSCLMwn
The early fetch hook was still inside the async init() function,
running AFTER await storage.local.get(). By the time storage
responded, ChatGPT's JS had already loaded and captured the
original fetch().
Moved the inline script injection to the TOP of the IIFE, before
any async operations. The sequence is now:
1. [synchronous] Inject inline <script> that captures fetch/XHR
2. [synchronous] Define mergeProfiles and other helpers
3. [async] Read storage for config
4. [async] Inject content.js with full substitution engine
This guarantees the fetch proxy is installed before any page
JavaScript runs, regardless of how long storage reads take.
https://claude.ai/code/session_01SWSwDfMVij53bCTNSCLMwn
Replace tab-load-only injection with scripting.registerContentScripts()
so custom domains inject at document_start (like built-in sites) and
persist across service worker restarts. Scripts re-register automatically
when domains change. Removed domains now revoke browser permissions.
https://claude.ai/code/session_01KF4i7Ra7zCEDskxDBaNtcT
Sync code import between browsers was failing with "Wrong key or
corrupted data" because each browser derives a different AES key
from the same password (different salt). The decrypt function was
using the local device's salt instead of the source device's salt
embedded in the sync envelope.
Fixed _decryptFromSync to always use the sync envelope's salt for
decryption. If the cached key's salt doesn't match, forces re-auth
so the password is re-entered and a new key is derived with the
correct salt.
Also fixed WebAuthn "device can't be used" error on extension pages.
chrome-extension:// and moz-extension:// origins are not valid for
WebAuthn. isWebAuthnAvailable() now returns false on extension pages.
https://claude.ai/code/session_01SWSwDfMVij53bCTNSCLMwn
The fetch interceptor was loaded via <script src="content.js"> which
is asynchronous. ChatGPT's Next.js framework stores a reference to
the original fetch() during module initialization — before our script
finishes downloading. By the time content.js patches window.fetch,
ChatGPT is already using its stored copy of the original.
Fixed with a two-step injection:
1. Injector injects a tiny INLINE <script> (synchronous, instant)
that stores the real fetch/XHR references and installs a thin
proxy. This runs before ANY page JavaScript.
2. Content.js loads normally, uses the stored __ssOriginalFetch
reference, and registers __ssInterceptFetch so the proxy can
route future calls through the full substitution engine.
This ensures the fetch hook is in place before frameworks like
Next.js, React, or any SPA framework can save a reference to
the original fetch().
https://claude.ai/code/session_01SWSwDfMVij53bCTNSCLMwn
Sync operations (browser sync, Gist, custom URL, sync code) now refuse to
operate without encryption enabled. Disabling encryption also turns off all
active sync channels. Activity log cap reduced from 200 to 100 entries for
both storage and display.
https://claude.ai/code/session_01KF4i7Ra7zCEDskxDBaNtcT
ChatGPT (and potentially other AI services) calls fetch() with a
Request object as the first argument: fetch(new Request(url, opts))
instead of fetch(url, opts). The interceptor only handled the second
form, so ChatGPT's conversation requests passed through unmodified.
Fixed by handling both fetch signatures:
- fetch(url, options) — existing path
- fetch(Request) — new: extracts URL, method, headers, reads body
via request.text() for JSON/text content types
Also handles non-string body types:
- Blob → text via blob.text()
- ArrayBuffer → text via TextDecoder
- URLSearchParams → string via toString()
These cover the various ways modern frameworks call fetch().
https://claude.ai/code/session_01SWSwDfMVij53bCTNSCLMwn
Auto-sync now:
- Persists Gist token in the auto-sync config so it survives
page reloads (previously only read from the input field)
- Saves token when the Gist token field changes (not just on toggle)
- Validates that credentials exist before enabling
- Shows current status: method, interval, last push/pull times
- Updated UI description to clarify it works in all browsers
https://claude.ai/code/session_01SWSwDfMVij53bCTNSCLMwn
Shows which sync methods work in which browsers:
- Sync Code, GitHub Gist, Custom URL: all browsers
- Browser account sync: Chrome and Firefox only
- Auto-Sync Folder: Chrome only (File System Access API)
https://claude.ai/code/session_01SWSwDfMVij53bCTNSCLMwn
_getAllData() was reading raw storage which contains encrypted blobs
when at-rest encryption is enabled. The sync code export would send
{ _ssLocalEncrypted: true, data: <blob> } instead of actual identity
data. The importing browser couldn't use these blobs.
Fixed _getAllData() to use Storage._readSecure() which decrypts
transparently. Fixed _applyData() to use Storage._writeSecure()
so imported data gets encrypted on the receiving end.
Also: hide the Auto-Sync Folder section entirely in browsers that
don't support File System Access API (Firefox, Brave, Safari)
instead of showing a broken-looking error message.
https://claude.ai/code/session_01SWSwDfMVij53bCTNSCLMwn
Proper noun detection (flagging capitalized phrases like "Getting
Started", "Generate Design", "Introduction Getting Started") now
disabled by default. Enable via popup → Options tab → Detect proper
nouns.
The pattern-based detection (IPs, emails, API keys, addresses, paths,
etc.) remains always-on and reliable. The capitalized phrase heuristic
was producing too many false positives on normal UI phrases, headings,
and instructions — adding words to the filter was a losing game.
Added detectProperNouns toggle to popup Options tab.
Updated README to note it's opt-in.
https://claude.ai/code/session_01SWSwDfMVij53bCTNSCLMwn
Updated "How to verify it's working" with step-by-step instructions
for each browser showing exactly where to click:
- Firefox: F12 → Network → POST events → Request tab → expand JSON
- Chrome: F12 → Network → POST chat/events → Payload tab
- Safari: Cmd+Opt+I → Network → POST → Request
Includes what to look for (replaced data should be there, real data
should not) and what to do if verification fails.
https://claude.ai/code/session_01SWSwDfMVij53bCTNSCLMwn
README changes:
- Added step-by-step "How to verify it's working" section with 6
methods, including Network tab inspection to prove real data never
reaches the AI
- Removed App Store distribution steps (internal, not user-facing)
- Rewrote "What Silent Send can't catch" — updated file uploads line
(we now scan them), clearer language
- Replaced "Legal" section with concise user-facing disclaimer —
removed internal litigation notes that read like developer-to-
developer instead of developer-to-user
- Shortened license section
https://claude.ai/code/session_01SWSwDfMVij53bCTNSCLMwn
Added:
- Safari installation section (full guide: build, test without
Apple Developer account via Allow Unsigned Extensions, App Store
distribution steps)
- Document scanning section with format support table
- Concatenated name patterns in smart pattern examples
- Options tab mention in "how to verify" section
- Browser support note (Chrome, Firefox, Safari)
- document-scanner.js and build-safari.sh in architecture section
- Updated service-worker.js and content.js descriptions
https://claude.ai/code/session_01SWSwDfMVij53bCTNSCLMwn
Popup:
- Added 5th "Options" tab with key settings accessible without
opening the full options page: secret scanning, auto-detect PPI,
auto-redact, show highlights, document scan preview
- "Open Full Options Page" button for sync/encryption/org/import
- Settings changes in popup immediately broadcast to content scripts
Safari:
- Added build-safari.sh that uses Apple's safari-web-extension-converter
to generate an Xcode project from the Firefox build
- browser-polyfill.js already handles Safari (uses browser.* like Firefox)
https://claude.ai/code/session_01SWSwDfMVij53bCTNSCLMwn
Fixed safeHTML in all three files — template.content.childNodes and
doc.body.childNodes are live NodeLists that shrink as nodes are moved.
Using Array.from() to create a static copy before spreading into
replaceChildren().
Expanded proper noun common words filter with ~500 additional verbs,
nouns, and adjectives (generate, design, manage, process, account,
button, dashboard, etc.) to prevent false PPI flags on titles,
headings, and UI button labels.
https://claude.ai/code/session_01SWSwDfMVij53bCTNSCLMwn
Massively expanded the common words filter — added ~500 common verbs,
nouns, and adjectives that frequently appear capitalized in titles,
headings, UI buttons, and instructions. 'Generate Design', 'Create
Account', 'Search Filter', etc. are no longer flagged as PPI.
Updated both content.js and auto-detect.js with the same word list.
https://claude.ai/code/session_01SWSwDfMVij53bCTNSCLMwn
sessionSubstitutions stored keys as toLowerCase() for lookup, but the
lowercase key was also used as the substitute value when building
reveal pairs. This caused unrevealText to replace "John Smith" with
"ademo demo" instead of "Ademo Demo".
Fixed by storing both the original-case substitute and the original
real value in sessionSubstitutions. The lowercase key is still used
for lookup, but the preserved-case value is used for reveal pairs.
https://claude.ai/code/session_01SWSwDfMVij53bCTNSCLMwn
When reveal mode was turned OFF, text stayed showing real data
(e.g. "John Smith") instead of reverting to the AI's actual text
("Ademo Demo"). The unreveal function relied on a WeakMap of
original text nodes which lost entries when DOM nodes were replaced
during streaming responses.
Fixed by replacing WeakMap-based restore with active reverse
replacement: unrevealText() replaces real→substitute (the opposite
direction of revealText). This is reliable regardless of DOM changes.
Removed originalTexts WeakMap entirely.
https://claude.ai/code/session_01SWSwDfMVij53bCTNSCLMwn
1. Removed data_collection_permissions from Firefox manifest — it
requires Firefox 140+ but we target 128+. Only produces a warning
when missing, not a blocking error.
2. Fixed JS syntax error in content.js line 450 — two extra closing
braces from a bad merge in the proper noun detection function
caused a parse error that blocked Mozilla validation.
https://claude.ai/code/session_01SWSwDfMVij53bCTNSCLMwn
Replaced all 31 innerHTML assignments across popup.js (12),
options.js (17), and content.js (3) to pass Mozilla AMO linter.
Each file gets a safeHTML(el, html) helper:
- popup.js/options.js: DOMParser-based (extension page context)
- content.js: <template> element pattern (page world context)
Empty innerHTML clears replaced with el.replaceChildren().
innerHTML += replaced with createElement + safeHTML + appendChild.
All event listener bindings after HTML rebuilds remain functional
since safeHTML uses replaceChildren() which populates DOM
synchronously before querySelectorAll + addEventListener calls.
https://claude.ai/code/session_01SWSwDfMVij53bCTNSCLMwn
Required for Firefox Add-ons (AMO) public listing submission.
Documents that all processing is 100% local, no data collection,
no servers, no analytics. Explains each permission and all
optional network requests (sync, org policy).
https://claude.ai/code/session_01SWSwDfMVij53bCTNSCLMwn
Outis