mirror of https://github.com/garrytan/gstack.git
4 Commits
| Author | SHA1 | Message | Date |
|---|---|---|---|
Garry Tan
|
4328748136
|
fix(security): delete the dead ML layers — transcript classifier and DeBERTa ensemble
The L4b Haiku transcript classifier and the opt-in DeBERTa ensemble
(GSTACK_SECURITY_ENSEMBLE=deberta, a documented 721MB download) had ZERO
production callers since the chat-path agent that invoked them was ripped.
The only live ML path is scanPageContent (testsavant) inside the security
sidecar subprocess. Deleted by import graph:
- security-classifier.ts 614 -> 265 lines: HAIKU_MODEL, checkTranscript,
shouldRunTranscriptCheck, loadDeberta, scanPageContentDeberta, ToolCallInput,
all DEBERTA_* consts + load state. Header now states the live truth
(imported only by security-sidecar-entry.ts). downloadFile kept, name
intact — it is an enumerated egress sink (HF model download).
- security-bunnative.ts + test: a research skeleton self-described as 'NOT a
production replacement', shipped into src/ with zero importers.
- security-bench-ensemble{,-live}.test.ts + the Haiku response fixture: a
paid live-model benchmark for a layer that could not fire. The
security-classifier-tdz test's only case exercised checkTranscript — gone.
- security.ts: layer-model header rewritten to the live architecture;
StatusDetail.layers -> {testsavant, canary}; getStatus() no longer requires
the impossible transcript==='ok' for 'protected' (old on-disk session state
with a transcript key is tolerated on read, never re-emitted).
- security-sidecar-entry.ts needed zero changes: it serializes
getClassifierStatus() verbatim and no consumer read .transcript (verified
in sidecar-client + server.ts).
- BROWSER.md security section matches reality (ensemble knob gone, 112MB not
22MB, sidecar hosting documented). combineVerdict/THRESHOLDS retained as
the pure, tested combiner of record — comments now flag transcript/deberta
votes as producer-less.
Net: 26 pass in security.test.ts incl. a NEW regression test for stale-
transcript disk tolerance; egress-receipt tripwire green.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
|
|
|
Garry Tan
|
ed1e4be2f6
|
feat: gstack browser sidebar = interactive Claude Code REPL with live tab awareness (v1.14.0.0) (#1216)
* build: vendor xterm@5 for the Terminal sidebar tab
Adds xterm@5 + xterm-addon-fit as devDependencies and a `vendor:xterm`
build step that copies the assets into `extension/lib/` at build time.
The vendored files are .gitignored so the npm version stays the source
of truth. xterm@5 is eval-free, so no MV3 CSP changes needed.
No runtime callers yet — this just stages the assets.
* feat(server): add pty-session-cookie module for the Terminal tab
Mirrors `sse-session-cookie.ts` exactly. Mints short-lived 30-min HttpOnly
cookies for authenticating the Terminal-tab WebSocket upgrade against
the terminal-agent. Same TTL, same opportunistic-pruning shape, same
"scoped tokens never valid as root" invariant. Two registries instead of
one because the cookie names are different (`gstack_sse` vs `gstack_pty`)
and the token spaces must not overlap.
No callers yet — wired up in the next commit.
* feat(server): add terminal-agent.ts (PTY for the Terminal sidebar tab)
Translates phoenix gbrowser's Go PTY (cmd/gbd/terminal.go) into a Bun
non-compiled process. Lives separately from `sidebar-agent.ts` so a
WS-framing or PTY-cleanup bug can't take down the chat path (codex
outside-voice review caught the coupling risk).
Architecture:
- Bun.serve on 127.0.0.1:0 (never tunneled).
- POST /internal/grant accepts cookie tokens from the parent server over
loopback, authenticated with a per-boot internal token.
- GET /ws upgrades require BOTH (a) Origin: chrome-extension://<id> and
(b) the gstack_pty cookie minted by /pty-session. Either gate alone is
insufficient (CSWSH defense + auth defense).
- Lazy spawn: claude PTY is not started until the WS receives its first
data frame. Idle sidebar opens cost nothing.
- Bun PTY API: `terminal: { rows, cols, data(t, chunk) }` — verified at
impl time on Bun 1.3.10. proc.terminal.write() for input,
proc.terminal.resize() for resize, proc.kill() + 3s SIGKILL fallback
on close.
- process.on('uncaughtException'|'unhandledRejection') handlers so a
framing bug logs but doesn't kill the listener loop.
Test-only `BROWSE_TERMINAL_BINARY` env override lets the integration
tests spawn /bin/bash instead of requiring claude on every CI runner.
Not yet spawned by anything — wired in the next commit.
* feat(server): wire /pty-session route + spawn terminal-agent
Server-side glue connecting the Terminal sidebar tab to the new
terminal-agent process.
server.ts:
- New POST /pty-session route. Validates AUTH_TOKEN, mints a gstack_pty
HttpOnly cookie via pty-session-cookie.ts, posts the cookie value to
the agent's loopback /internal/grant. Returns the terminalPort + Set-Cookie
to the extension.
- /health response gains `terminalPort` (just the port number — never a
shell token). Tokens flow via the cookie path, never /health, because
/health already surfaces AUTH_TOKEN to localhost callers in headed mode
(that's a separate v1.1+ TODO).
- /pty-session and /terminal/* are deliberately NOT added to TUNNEL_PATHS,
so the dual-listener tunnel surface 404s by default-deny.
- Shutdown path now also pkills terminal-agent and unlinks its state files
(terminal-port + terminal-internal-token) so a reconnect doesn't try to
hit a dead port.
cli.ts:
- After spawning sidebar-agent.ts, also spawn terminal-agent.ts. Same
pattern: pkill old instances, Bun.spawn(['bun', 'run', script]) with
BROWSE_STATE_FILE + BROWSE_SERVER_PORT env. Non-fatal if the spawn
fails — chat still works without the terminal agent.
* feat(extension): Terminal as default sidebar tab
Adds a primary tab bar (Terminal | Chat) above the existing tab-content
panes. Terminal is the default-active tab; clicking Chat returns to the
existing claude -p one-shot flow which is preserved verbatim.
manifest.json: adds ws://127.0.0.1:*/ to host_permissions so MV3 doesn't
block the WebSocket upgrade.
sidepanel.html: new primary-tabs nav, new #tab-terminal pane with a
"Press any key to start Claude Code" bootstrap card, claude-not-found
install card, xterm mount point, and "session ended" restart UI. Loads
xterm.js + xterm-addon-fit + sidepanel-terminal.js. tab-chat is no
longer the .active default.
sidepanel.js: new activePrimaryPaneId() helper that reads which primary
tab is selected. Debug-close paths now route back to whichever primary
pane is active (was hardcoded to tab-chat). Primary-tab click handler
toggles .active classes and aria-selected. window.gstackServerPort and
window.gstackAuthToken exposed so sidepanel-terminal.js can build the
/pty-session POST and the WS URL.
sidepanel-terminal.js (new): xterm.js lifecycle. Lazy-spawn — first
keystroke fires POST /pty-session, then opens
ws://127.0.0.1:<terminalPort>/ws. Origin + cookie are set automatically
by the browser. Resize observer sends {type:"resize"} text frames.
ResizeObserver, tab-switch hooks, restart button, install-card retry.
On WS close shows "Session ended, click to restart" — no auto-reconnect
(codex outside-voice flagged that as session-burning).
sidepanel.css: primary-tabs bar + Terminal pane styling (full-height
xterm container, install card, ended state).
* test: terminal-agent + cookie module + sidebar default-tab regression
Three new test files:
terminal-agent.test.ts (16 tests): pty-session-cookie mint/validate/
revoke, Set-Cookie shape (HttpOnly + SameSite=Strict + Path=/, NO Secure
since 127.0.0.1 over HTTP), source-level guards that /pty-session and
/terminal/* are NOT in TUNNEL_PATHS, /health does NOT surface ptyToken
or gstack_pty, terminal-agent binds 127.0.0.1, /ws upgrade enforces
chrome-extension:// Origin AND gstack_pty cookie, lazy-spawn invariant
(spawnClaude is called from message handler, not upgrade), uncaughtException/
unhandledRejection handlers exist, SIGINT-then-SIGKILL cleanup.
terminal-agent-integration.test.ts (7 tests): spawns the agent as a real
subprocess in a tmp state dir. Verifies /internal/grant accepts/rejects
the loopback token, /ws gates (no Origin → 403, bad Origin → 403, no
cookie → 401), real WebSocket round-trip with /bin/bash via the
BROWSE_TERMINAL_BINARY override (write 'echo hello-pty-world\n', read it
back), and resize message acceptance.
sidebar-tabs.test.ts (13 tests): structural regression suite locking the
load-bearing invariants of the default-tab change — Terminal is .active,
Chat is not, xterm assets are loaded, debug-close path no longer hardcodes
tab-chat (uses activePrimaryPaneId), primary-tab click handler exists,
chat surface is not accidentally deleted, terminal JS does NOT auto-
reconnect on close, manifest declares ws:// + http:// localhost host
permissions, no unsafe-eval.
Plan called for Playwright + extension regression; the codebase doesn't
ship Playwright extension launcher infra, so we follow the existing
extension-test pattern (source-level structural assertions). Same
load-bearing intent — locks the invariants before they regress.
* docs: Terminal flow + threat model + v1.1 follow-ups
SIDEBAR_MESSAGE_FLOW.md: new "Terminal flow" section. Documents the WS
upgrade path (/pty-session cookie mint → /ws Origin + cookie gate →
lazy claude spawn), the dual-token model (AUTH_TOKEN for /pty-session,
gstack_pty cookie for /ws, INTERNAL_TOKEN for server↔agent loopback),
and the threat-model boundary — the Terminal tab bypasses the entire
prompt-injection security stack on purpose; user keystrokes are the
trust source. That trust assumption is load-bearing on three transport
guarantees: local-only listener, Origin gate, cookie auth. Drop any
one of those three and the tab becomes unsafe.
CLAUDE.md: extends the "Sidebar architecture" note to include
terminal-agent.ts in the read-this-first list. Adds a "Terminal tab is
its own process" note so a future contributor doesn't bolt PTY logic
onto sidebar-agent.ts.
TODOS.md: three new follow-ups under a new "Sidebar Terminal" section:
- v1.1: PTY session survives sidebar reload (Issue 1C deferred).
- v1.1+: audit /health AUTH_TOKEN distribution (codex finding #2 —
a pre-existing soft leak that cc-pty-import sidesteps but doesn't
fix).
- v1.1+: apply terminal-agent's process.on exception handlers to
sidebar-agent.ts (codex finding #4 — chat path has no fatal
handlers).
* feat(extension): Terminal-only sidebar — auth fix, UX polish, chat rip
The chat queue path is gone. The Chrome side panel is now just an
interactive claude PTY in xterm.js. Activity / Refs / Inspector still
exist behind the `debug` toggle in the footer.
Three threads of change, all from dogfood iteration on top of
cc-pty-import:
1. fix(server): cross-port WS auth via Sec-WebSocket-Protocol
- Browsers can't set Authorization on a WebSocket upgrade. We had
been minting an HttpOnly gstack_pty cookie via /pty-session, but
SameSite=Strict cookies don't survive the cross-port jump from
server.ts:34567 to the agent's random port from a chrome-extension
origin. The WS opened then immediately closed → "Session ended."
- /pty-session now also returns ptySessionToken in the JSON body.
- Extension calls `new WebSocket(url, [`gstack-pty.<token>`])`.
Browser sends Sec-WebSocket-Protocol on the upgrade.
- Agent reads the protocol header, validates against validTokens,
and MUST echo the protocol back (Chromium closes the connection
immediately if a server doesn't pick one of the offered protocols).
- Cookie path is kept as a fallback for non-browser callers (curl,
integration tests).
- New integration test exercises the full protocol-auth round-trip
via raw fetch+Upgrade so a future regression of this exact class
fails in CI.
2. fix(extension): UX polish on the Terminal pane
- Eager auto-connect when the sidebar opens — no "Press any key to
start" friction every reload.
- Always-visible ↻ Restart button in the terminal toolbar (not
gated on the ENDED state) so the user can force a fresh claude
mid-session.
- MutationObserver on #tab-terminal's class attribute drives a
fitAddon.fit() + term.refresh() when the pane becomes visible
again — xterm doesn't auto-redraw after display:none → display:flex.
3. feat(extension): rip the chat tab + sidebar-agent.ts
- Sidebar is Terminal-only. No more Terminal | Chat primary nav.
- sidebar-agent.ts deleted. /sidebar-command, /sidebar-chat,
/sidebar-agent/event, /sidebar-tabs* and friends all deleted.
- The pickSidebarModel router (sonnet vs opus) is gone — the live
PTY uses whatever model the user's `claude` CLI is configured with.
- Quick-actions (🧹 Cleanup / 📸 Screenshot / 🍪 Cookies) survive
in the Terminal toolbar. Cleanup now injects its prompt into the
live PTY via window.gstackInjectToTerminal — no more
/sidebar-command POST. The Inspector "Send to Code" action uses
the same injection path.
- clear-chat button removed from the footer.
- sidepanel.js shed ~900 lines of chat polling, optimistic UI,
stop-agent, etc.
Net diff: -3.4k lines across 16 files. CLAUDE.md, TODOS.md, and
docs/designs/SIDEBAR_MESSAGE_FLOW.md rewritten to match. The sidebar
regression test (browse/test/sidebar-tabs.test.ts) is rewritten as 27
structural assertions locking the new layout — Terminal sole pane,
no chat input, quick-actions in toolbar, eager-connect, MutationObserver
repaint, restart helper.
* feat: live tab awareness for the Terminal pane
claude in the PTY now has continuous tab-aware context. Three pieces:
1. Live state files. background.js listens to chrome.tabs.onActivated /
onCreated / onRemoved / onUpdated (throttled to URL/title/status==
complete so loading spinners don't spam) and pushes a snapshot. The
sidepanel relays it as a custom event; sidepanel-terminal.js sends
{type:"tabState"} text frames over the live PTY WebSocket.
terminal-agent.ts writes:
<stateDir>/tabs.json all open tabs (id, url, title, active,
pinned, audible, windowId)
<stateDir>/active-tab.json current active tab (skips chrome:// and
chrome-extension:// internal pages)
Atomic write via tmp + rename so claude never reads a half-written
document. A fresh snapshot is pushed on WS open so the files exist by
the time claude finishes booting.
2. New $B tab-each <command> [args...] meta-command. Fans out a single
command across every open tab, returns
{command, args, total, results: [{tabId, url, title, status, output}]}.
Skips chrome:// pages; restores the originally active tab in a finally
block (so a mid-batch error doesn't leave the user looking at a
different tab); uses bringToFront: false so the OS window doesn't
jump on every fanout. Scope-checks the inner command BEFORE the loop.
3. --append-system-prompt hint at spawn time. Claude is told about both
the state files and the $B tab-each command up front, so it doesn't
have to discover the surface by trial. Passed via the --append-system-
prompt CLI flag, NOT as a leading PTY write — the hint stays out of
the visible transcript.
Tests:
- browse/test/tab-each.test.ts (new) — registration + source-level
invariants (scope check before loop, finally-restore, bringToFront:false,
chrome:// skip) + behavior tests with a mock BrowserManager that verify
iteration order, JSON shape, error handling, and active-tab restore.
- browse/test/terminal-agent.test.ts — three new assertions for
tabState handler shape, atomic-write pattern, and the
--append-system-prompt wiring at spawn.
Verified live: opened 5 tabs, ran $B tab-each url against the live
server, got per-tab JSON results back, original active tab restored
without OS focus stealing.
* chore: drop sidebar-agent test refs after chat rip
Five test files / describe blocks targeted the deleted chat path:
- browse/test/security-e2e-fullstack.test.ts (full-stack chat-pipeline E2E
with mock claude — whole file gone)
- browse/test/security-review-fullstack.test.ts (review-flow E2E with real
classifier — whole file gone)
- browse/test/security-review-sidepanel-e2e.test.ts (Playwright E2E for
the security event banner that was ripped from sidepanel.html)
- browse/test/security-audit-r2.test.ts (5 describe blocks: agent queue
permissions, isValidQueueEntry stateFile traversal, loadSession session-ID
validation, switchChatTab DocumentFragment, pollChat reentrancy guard,
/sidebar-tabs URL sanitization, sidebar-agent SIGTERM→SIGKILL escalation,
AGENT_SRC top-level read converted to graceful fallback)
- browse/test/security-adversarial-fixes.test.ts (canary stream-chunk split
detection on detectCanaryLeak; one tool-output test on sidebar-agent)
- test/skill-validation.test.ts (sidebar agent #584 describe block)
These all assumed sidebar-agent.ts existed and tested chat-queue plumbing,
chat-tab DOM round-trip, chat-polling reentrancy, or per-message classifier
canary detection. With the live PTY there is no chat queue, no chat tab,
no LLM stream to canary-scan, and no per-message subprocess. The Terminal
pane's invariants are covered by the new browse/test/sidebar-tabs.test.ts
(27 structural assertions), browse/test/terminal-agent.test.ts, and
browse/test/terminal-agent-integration.test.ts.
bun test → exit 0, 0 failures.
* chore: bump version and changelog (v1.14.0.0)
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* fix(extension): xterm fills the full Terminal panel height
The Terminal pane only rendered into the top portion of the panel — most
of the panel below the prompt was an empty black gap. Three layered
issues, all about xterm.js measuring dimensions during a layout state
that wasn't ready yet:
1. order-of-operations in connect(): ensureXterm() ran BEFORE
setState(LIVE), so term.open() measured els.mount while it was still
display:none. xterm caches a 0-size viewport synchronously inside
open() and never auto-recovers when the container goes visible.
Flipped: setState(LIVE) → ensureXterm.
2. first fit() ran synchronously before the browser had applied the
.active class transition. Wrapped in requestAnimationFrame so layout
has settled before fit() reads clientHeight.
3. CSS flex-overflow trap: .terminal-mount has flex:1 inside the
flex-column #tab-terminal, but .tab-content's `overflow-y: auto` and
the lack of `min-height: 0` on .terminal-mount meant the item
couldn't shrink below content size. flex:1 then refused to expand
into available space and xterm rendered into whatever its initial
2x2 measurement happened to be.
Fixes:
- extension/sidepanel-terminal.js: reorder + RAF fit
- extension/sidepanel.css: .terminal-mount gets `flex: 1 1 0` +
`min-height: 0` + `position: relative`. #tab-terminal overrides
.tab-content's `overflow-y: auto` to `overflow: hidden` (xterm has
its own viewport scroll; the parent shouldn't compete) and explicitly
re-declares `display: flex; flex-direction: column` for #tab-terminal.active.
bun test browse/test/sidebar-tabs.test.ts → 27/27 pass.
Manually verified: side panel opens → Terminal fills full panel height,
xterm scrollback works, debug-tab toggle still repaints correctly.
---------
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
|
|
|
Garry Tan
|
d75402bbd2
|
v1.6.4.0: cut Haiku classifier FP from 44% to 23%, gate now enforced (#1135)
* feat(security): v2 ensemble tuning — label-first voting + SOLO_CONTENT_BLOCK Cuts Haiku classifier false-positive rate from 44.1% → 22.9% on BrowseSafe-Bench smoke. Detection trades from 67.3% → 56.2%; the lost TPs are all cases Haiku correctly labeled verdict=warn (phishing targeting users, not agent hijack) — they still surface in the WARN banner meta but no longer kill the session. Key changes: - combineVerdict: label-first voting for transcript_classifier. Only meta.verdict==='block' block-votes; verdict==='warn' is a soft signal. Missing meta.verdict never block-votes (backward-compat). - Hallucination guard: verdict='block' at confidence < LOG_ONLY (0.40) drops to warn-vote — prevents malformed low-conf blocks from going authoritative. - New THRESHOLDS.SOLO_CONTENT_BLOCK = 0.92 decoupled from BLOCK (0.85). Label-less content classifiers (testsavant, deberta) need a higher solo-BLOCK bar because they can't distinguish injection from phishing-targeting-user. Transcript keeps label-gated solo path (verdict=block AND conf >= BLOCK). - THRESHOLDS.WARN bumped 0.60 → 0.75 — borderline fires drop out of the 2-of-N ensemble pool. - Haiku model pinned (claude-haiku-4-5-20251001). `claude -p` spawns from os.tmpdir() so project CLAUDE.md doesn't poison the classifier context (measured 44k cache_creation tokens per call before the fix, and Haiku refusing to classify because it read "security system" from CLAUDE.md and went meta). - Haiku timeout 15s → 45s. Measured real latency is 17-33s end-to-end (Claude Code session startup + Haiku); v1's 15s caused 100% timeout when re-measured — v1's ensemble was effectively L4-only in prod. - Haiku prompt rewritten: explicit block/warn/safe criteria, 8 few-shot exemplars (instruction-override → block; social engineering → warn; discussion-of-injection → safe). Test updates: - 5 existing combineVerdict tests adapted for label-first semantics (transcript signals now need meta.verdict to block-vote). - 6 new tests: warn-soft-signal, three-way-block-with-warn-transcript, hallucination-guard-below-floor, above-floor-label-first, backward-compat-missing-meta. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * test(security): live + fixture-replay bench harness with 500-case capture Adds two new benches that permanently guard the v2 tuning: - security-bench-ensemble-live.test.ts (opt-in via GSTACK_BENCH_ENSEMBLE=1). Runs full ensemble on BrowseSafe-Bench smoke with real Haiku calls. Worker-pool concurrency (default 8, tunable via GSTACK_BENCH_ENSEMBLE_CONCURRENCY) cuts wall clock from ~2hr to ~25min on 500 cases. Captures Haiku responses to fixture for replay. Subsampling via GSTACK_BENCH_ENSEMBLE_CASES for faster iteration. Stop-loss iterations write to ~/.gstack-dev/evals/stop-loss-iter-N-* WITHOUT overwriting canonical fixture. - security-bench-ensemble.test.ts (CI gate, deterministic replay). Replays captured fixture through combineVerdict, asserts detection >= 55% AND FP <= 25%. Fail-closed when fixture is missing AND security-layer files changed in branch diff. Uses `git diff --name-only base` (two-dot) to catch both committed and working-tree changes — `git diff base...HEAD` would silently skip in CI after fixture lands. - browse/test/fixtures/security-bench-haiku-responses.json — 500 cases × 3 classifier signals each. Header includes schema_version, pinned model, component hashes (prompt, exemplars, thresholds, combiner, dataset version). Any change invalidates the fixture and forces fresh live capture. - docs/evals/security-bench-ensemble-v2.json — durable PR artifact with measured TP/FN/FP/TN, 95% CIs, knob state, v1 baseline delta. Checked in so reviewers can see the numbers that justified the ship. Measured baseline on the new harness: TP=146 FN=114 FP=55 TN=185 → 56.2% / 22.9% → GATE PASS Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * chore(release): v1.5.1.0 — cut Haiku FP 44% → 23% - VERSION: 1.5.0.0 → 1.5.1.0 (TUNING bump) - CHANGELOG: [1.5.1.0] entry with measured numbers, knob list, and stop-loss rule spec - TODOS: mark "Cut Haiku FP 44% → ~15%" P0 as SHIPPED with pointer to CHANGELOG and v1 plan Measured: 56.2% detection (CI 50.1-62.1) / 22.9% FP (CI 18.1-28.6) on 500-case BrowseSafe-Bench smoke. Gate passes (floor 55%, ceiling 25%). Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * docs(changelog): add v1.6.4.0 placeholder entry at top Per CLAUDE.md branch-scoped discipline, our VERSION 1.6.4.0 needs a CHANGELOG entry at the top so readers can tell what's on this branch vs main. Honest placeholder: no user-facing runtime changes yet, two merges bringing branch up to main's v1.6.3.0, and the approved injection-tuning plan is queued but unimplemented. Gets replaced by the real release-summary at /ship time after Phases -1 through 10 land. * docs(changelog): strip process minutiae from entries; rewrite v1.6.4.0 CLAUDE.md — new CHANGELOG rule: only document what shipped between main and this change. Keep out branch resyncs, merge commits, plan approvals, review outcomes, scope negotiations, "work queued" or "in-progress" framing. When no user-facing change actually landed, one sentence is the entry: "Version bump for branch-ahead discipline. No user-facing changes yet." CHANGELOG.md — v1.6.4.0 entry rewritten to match. Previous version narrated the branch history, the approved injection-tuning plan, and what we expect to ship later — all of which are process minutiae readers do not care about. * docs(changelog): rewrite v1.6.4.0; strip process minutiae Rewrote v1.6.4.0 entry to follow the new CLAUDE.md rule: only document what shipped between main and this change. Previous entry narrated the branch history, the approved injection-tuning plan, and what we expect to ship later, all process minutiae readers do not care about. v1.6.4.0 now reads: what the detection tuning did for users, the before/after numbers, the stop-loss rule, and the itemized changes for contributors. CLAUDE.md — new rule: only document what shipped between main and this change. Keep out branch resyncs, merge commits, plan approvals, review outcomes, scope negotiations, "work queued" / "in-progress" framing. If nothing user-facing landed, one sentence: "Version bump for branch-ahead discipline. No user-facing changes yet." --------- Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com> |
|
|
Garry Tan
|
97584f9a59
|
feat(security): ML prompt injection defense for sidebar (v1.4.0.0) (#1089)
* chore(deps): add @huggingface/transformers for prompt injection classifier
Dependency needed for the ML prompt injection defense layer coming in the
follow-up commits. @huggingface/transformers will host the TestSavantAI
BERT-small classifier that scans tool outputs for indirect prompt injection.
Note: this dep only runs in non-compiled bun contexts (sidebar-agent.ts).
The compiled browse binary cannot load it because transformers.js v4 requires
onnxruntime-node (native module, fails to dlopen from bun compile's temp
extract dir). See docs/designs/ML_PROMPT_INJECTION_KILLER.md for the full
architectural decision.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* feat(security): add security.ts foundation for prompt injection defense
Establishes the module structure for the L5 canary and L6 verdict aggregation
layers. Pure-string operations only — safe to import from the compiled browse
binary.
Includes:
* THRESHOLDS constants (BLOCK 0.85 / WARN 0.60 / LOG_ONLY 0.40), calibrated
against BrowseSafe-Bench smoke + developer content benign corpus.
* combineVerdict() implementing the ensemble rule: BLOCK only when the ML
content classifier AND the transcript classifier both score >= WARN.
Single-layer high confidence degrades to WARN to prevent any one
classifier's false-positives from killing sessions (Stack Overflow
instruction-writing-style FPs at 0.99 on TestSavantAI alone).
* generateCanary / injectCanary / checkCanaryInStructure — session-scoped
secret token, recursively scans tool arguments, URLs, file writes, and
nested objects per the plan's all-channel coverage decision.
* logAttempt with 10MB rotation (keeps 5 generations). Salted SHA-256 hash,
per-device salt at ~/.gstack/security/device-salt (0600).
* Cross-process session state at ~/.gstack/security/session-state.json
(atomic temp+rename). Required because server.ts (compiled) and
sidebar-agent.ts (non-compiled) are separate processes.
* getStatus() for shield icon rendering via /health.
ML classifier code will live in a separate module (security-classifier.ts)
loaded only by sidebar-agent.ts — compiled browse binary cannot load the
native ONNX runtime.
Plan: ~/.gstack/projects/garrytan-gstack/ceo-plans/2026-04-19-prompt-injection-guard.md
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* feat(security): wire canary injection into sidebar spawnClaude
Every sidebar message now gets a fresh CANARY-XXXXXXXXXXXX token embedded
in the system prompt with an instruction for Claude to never output it on
any channel. The token flows through the queue entry so sidebar-agent.ts
can check every outbound operation for leaks.
If Claude echoes the canary into any outbound channel (text stream, tool
arguments, URLs, file write paths), the sidebar-agent terminates the
session and the user sees the approved canary leak banner.
This operation is pure string manipulation — safe in the compiled browse
binary. The actual output-stream check (which also has to be safe in
compiled contexts) lives in sidebar-agent.ts (next commit).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* test(security): make sidebar-agent destructure check regex-tolerant
The test asserted the exact string `const { prompt, args, stateFile, cwd, tabId } = queueEntry`
which breaks whenever security or other extensions add fields (canary, pageUrl,
etc.). Switch to a regex that requires the core fields in order but tolerates
additional fields in between. Preserves the test's intent (args come from the
queue entry, not rebuilt) while allowing the destructure to grow.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* feat(security): canary leak check across all outbound channels
The sidebar-agent now scans every Claude stream event for the session's
canary token before relaying any data to the sidepanel. Channels covered
(per CEO review cross-model tension #2):
* Assistant text blocks
* Assistant text_delta streaming
* tool_use arguments (recursively, via checkCanaryInStructure — catches
URLs, commands, file paths nested at any depth)
* tool_use content_block_start
* tool_input_delta partial JSON
* Final result payload
If the canary leaks on any channel, onCanaryLeaked() fires once per session:
1. logAttempt() writes the event to ~/.gstack/security/attempts.jsonl
with the canary's salted hash (never the payload content).
2. sends a `security_event` to the sidepanel so it can render the approved
canary-leak banner (variant A mockup — ceo-plan 2026-04-19).
3. sends an `agent_error` for backward-compat with existing error surfaces.
4. SIGTERM's the claude subprocess (SIGKILL after 2s if still alive).
The leaked content itself is never relayed to the sidepanel — the event is
dropped at the boundary. Canary detection is pure-string substring match,
so this all runs safely in the sidebar-agent (non-compiled bun) context.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* feat(security): add security-classifier.ts with TestSavantAI + Haiku
This module holds the ML classifier code that the compiled browse binary
cannot link (onnxruntime-node native dylib doesn't load from Bun compile's
temp extract dir — see CEO plan §"Pre-Impl Gate 1 Outcome"). It's imported
ONLY by sidebar-agent.ts, which runs as a non-compiled bun script.
Two layers:
L4 testsavant_content — TestSavantAI BERT-small ONNX classifier. First call
triggers a one-time 112MB model download to ~/.gstack/models/testsavant-small/
(files staged into the onnx/ layout transformers.js v4 expects). Classifies
page snapshots and tool outputs for indirect prompt injection + jailbreak
attempts. On benign-corpus dry-run: Wikipedia/HN/Reddit/tech-blog all score
SAFE 0.98+, attack text scores INJECTION 0.99+, Stack Overflow
instruction-writing now scores SAFE 0.98 on the shorter form (was 0.99
INJECTION on the longer form — instruction-density threshold). Ensemble
combiner downgrades single-layer high to WARN to cover this case.
L4b transcript_classifier — Claude Haiku reasoning-blind pre-tool-call scan.
Sees only {user_message, last 3 tool_calls}, never Claude's chain-of-thought
or tool results (those are how self-persuasion attacks leak). 2000ms hard
timeout. Fail-open on any subprocess failure so sidebar stays functional.
Gated by shouldRunTranscriptCheck() — only runs when another layer already
fired at >= LOG_ONLY, saving ~70% of Haiku spend.
Both layers degrade gracefully: load/spawn failures set status to 'degraded'
and return confidence=0. Shield icon reflects this via getClassifierStatus()
which security.ts's getStatus() composes.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* feat(security): wire TestSavantAI + ensemble into sidebar-agent pre-spawn scan
The sidebar-agent now runs a ML security check on the user message BEFORE
spawning claude. If the content classifier and (gated) transcript classifier
ensemble returns BLOCK, the session is refused with a security_event +
agent_error — the sidepanel renders the approved banner.
Two pieces:
1. On agent startup, loadTestsavant() warms the classifier in the background.
First run triggers a 112MB model download from HuggingFace (~30s on
average broadband). Non-blocking — sidebar stays functional during
cold-start, shield just reports 'off' until warmed.
2. preSpawnSecurityCheck() runs the ensemble against the user message:
- L4 (testsavant_content) always runs
- L4b (transcript_classifier via Haiku) runs only if L4 flagged at
>= LOG_ONLY — plan §E1 gating optimization, saves ~70% of Haiku spend
combineVerdict() applies the BLOCK-requires-both-layers rule, which
downgrades any single-layer high confidence to WARN. Stack Overflow-style
instruction-heavy writing false-positives on TestSavantAI alone are
caught by this degrade — Haiku corrects them when called.
Fail-open everywhere: any subprocess/load/inference error returns confidence=0
so the sidebar keeps working on architectural controls alone. Shield icon
reflects degraded state via getClassifierStatus().
BLOCK path emits both:
- security_event {verdict, reason, layer, confidence, domain} (for the
approved canary-leak banner UX mockup — variant A)
- agent_error "Session blocked — prompt injection detected..."
(backward-compat with existing error surface)
Regression test suite still passes (12/12 sidebar-security tests).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* test(security): add security.ts unit tests (25 tests, 62 assertions)
Covers the pure-string operations that must behave deterministically in both
compiled and source-mode bun contexts:
* THRESHOLDS ordering invariant (BLOCK > WARN > LOG_ONLY > 0)
* combineVerdict ensemble rule — THE critical path:
- Empty signals → safe
- Canary leak always blocks (regardless of ML signals)
- Both ML layers >= WARN → BLOCK (ensemble_agreement)
- Single layer >= BLOCK → WARN (single_layer_high) — the Stack Overflow
FP mitigation that prevents one classifier killing sessions alone
- Max-across-duplicates when multiple signals reference the same layer
* Canary generation + injection + recursive checking:
- Unique CANARY-XXXXXXXXXXXX tokens (>= 48 bits entropy)
- Recursive structure scan for tool_use inputs, nested URLs, commands
- Null / primitive handling doesn't throw
* Payload hashing (salted sha256) — deterministic per-device, differs across
payloads, 64-char hex shape
* logAttempt writes to ~/.gstack/security/attempts.jsonl
* writeSessionState + readSessionState round-trip (cross-process)
* getStatus returns valid SecurityStatus shape
* extractDomain returns hostname only, empty string on bad input
All 25 tests pass in 18ms — no ML, no network, no subprocess spawning.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* feat(security): expose security status on /health for shield icon
The /health endpoint now returns a `security` field with the classifier
status, suitable for driving the sidepanel shield icon:
{
status: 'protected' | 'degraded' | 'inactive',
layers: { testsavant, transcript, canary },
lastUpdated: ISO8601
}
Backend plumbing:
* server.ts imports getStatus from security.ts (pure-string, safe in
compiled binary) and includes it in the /health response.
* sidebar-agent.ts writes ~/.gstack/security/session-state.json when the
classifier warmup completes (success OR failure). This is the cross-
process handoff — server.ts reads the state file via getStatus() to
surface the result to the sidepanel.
The sidepanel rendering (SVG shield icon + color states + tooltip) is a
follow-up commit in the extension/ code.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* docs(security): document the sidebar security stack in CLAUDE.md
Adds a security section to the Browser interaction block. Covers:
* Layered defense table showing which modules live where (content-security.ts
in both contexts vs security-classifier.ts only in sidebar-agent) and why
the split exists (onnxruntime-node incompatibility with compiled Bun)
* Threshold constants (0.85 / 0.60 / 0.40) and the ensemble rule that
prevents single-classifier false-positives (the Stack Overflow FP story)
* Env knobs — GSTACK_SECURITY_OFF kill switch, cache paths, salt file,
attack log rotation, session state file
This is the "before you modify the security stack, read this" doc. It lives
next to the existing Sidebar architecture note that points at
SIDEBAR_MESSAGE_FLOW.md.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* docs(todos): mark ML classifier v1 in-progress + file v2 follow-ups
Reframes the P0 item to reflect v1 scope (branch 2 architecture, TestSavantAI
pivot, what shipped) and splits v2 work into discrete TODOs:
* Shield icon + canary leak banner UI (P0, blocks v1 user-facing completion)
* Attack telemetry via gstack-telemetry-log (P1)
* Full BrowseSafe-Bench at gate tier (P2)
* Cross-user aggregate attack dashboard (P2)
* DeBERTa-v3 as third signal in ensemble (P2)
* Read/Glob/Grep ingress coverage (P2, flagged by Codex review)
* Adversarial + integration + smoke-bench test suites (P1)
* Bun-native 5ms inference (P3 research)
Each TODO carries What / Why / Context / Effort / Priority / Depends-on so
it's actionable by someone picking it up cold.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* feat(telemetry): add attack_attempt event type to gstack-telemetry-log
Extends the existing telemetry pipe with 5 new flags needed for prompt
injection attack reporting:
--url-domain hostname only (never path, never query)
--payload-hash salted sha256 hex (opaque — no payload content ever)
--confidence 0-1 (awk-validated + clamped; malformed → null)
--layer testsavant_content | transcript_classifier | aria_regex | canary
--verdict block | warn | log_only
Backward compatibility:
* Existing skill_run events still work — all new fields default to null
* Event schema is a superset of the old one; downstream edge function can
filter by event_type
No new auth, no new SDK, no new Supabase migration. The same tier gating
(community → upload, anonymous → local only, off → no-op) and the same
sync daemon carry the attack events. This is the "E6 RESOLVED" path from
the CEO plan — riding the existing pipe instead of spinning up parallel infra.
Verified end-to-end:
* attack_attempt event with all fields emits correctly to skill-usage.jsonl
* skill_run event with no security flags still works (backward compat)
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* feat(security): wire logAttempt to gstack-telemetry-log (fire-and-forget)
Every local attempt.jsonl write now also triggers a subprocess call to
gstack-telemetry-log with the attack_attempt event type. The binary handles
tier gating internally (community → Supabase upload, anonymous → local
JSONL only, off → no-op), so security.ts doesn't need to re-check.
Binary resolution follows the skill preamble pattern — never relies on PATH,
which breaks in compiled-binary contexts:
1. ~/.claude/skills/gstack/bin/gstack-telemetry-log (global install)
2. .claude/skills/gstack/bin/gstack-telemetry-log (symlinked dev)
3. bin/gstack-telemetry-log (in-repo dev)
Fire-and-forget:
* spawn with stdio: 'ignore', detached: true, unref()
* .on('error') swallows failures
* Missing binary is non-fatal — local attempts.jsonl still gives audit trail
Never throws. Never blocks. Existing 37 security tests pass unchanged.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* feat(ui): add security banner markup + styles (approved variant A)
HTML + CSS for the canary leak / ML block banner. Structure matches the
approved mockup from /plan-design-review 2026-04-19 (variant A — centered
alert-heavy):
* Red alert-circle SVG icon (no stock shield, intentional — matches the
"serious but not scary" tone the review chose)
* "Session terminated" Satoshi Bold 18px red headline
* "— prompt injection detected from {domain}" DM Sans zinc subtitle
* Expandable "What happened" chevron button (aria-expanded/aria-controls)
* Layer list rendered in JetBrains Mono with amber tabular-nums scores
* Close X in top-right, 28px hit area, focus-visible amber outline
Enter animation: slide-down 8px + fade, 250ms, cubic-bezier(0.16,1,0.3,1) —
matches DESIGN.md motion spec. Respects `role="alert"` + `aria-live="assertive"`
so screen readers announce on appearance. Escape-to-dismiss hook is in the
JS follow-up commit.
Design tokens all via CSS variables (--error, --amber-400, --amber-500,
--zinc-*, --font-display, --font-mono, --radius-*) — already established in
the stylesheet. No new color constants introduced.
JS wiring lands in the next commit so this diff stays focused on
presentation layer only.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* feat(ui): wire security banner to security_event + interactivity
Adds showSecurityBanner() and hideSecurityBanner() plus the addChatEntry
routing for entry.type === 'security_event'. When the sidebar-agent emits
a security_event (canary leak or ML BLOCK), the banner renders with:
* Title ("Session terminated")
* Subtitle with {domain} if present, otherwise generic
* Expandable layer list — each row: SECURITY_LAYER_LABELS[layer] +
confidence.toFixed(2) in mono. Readable + auditable — user can see
which layer fired at what score
Interactivity, wired once on DOMContentLoaded:
* Close X → hideSecurityBanner()
* Expand/collapse "What happened" → toggles details + aria-expanded +
chevron rotation (200ms css transition already in place)
* Escape key dismisses while banner is visible (a11y)
No shield icon yet — that's a separate commit that will consume the
`security` field now returned by /health.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* feat(ui): add security shield icon in sidepanel header (3 states)
Small "SEC" badge in the top-right of the sidepanel that reflects the
security module's current state. Three states drive color:
protected green — all layers ok (TestSavantAI + transcript + canary)
degraded amber — one+ ML layer offline but canary + arch controls active
inactive red — security module crashed, arch controls only
Consumes /health.security (surfaced in commit
|