Repository navigation
feat: v1.2.2.0 — Windows cookie import, hardening wave, caveman lock - #2
Merged
Merged
Conversation
…n, CDP fallback Ports gstack garrytan#892 + garrytan#1021 hardening. Windows-first CaveStack users can now import cookies from Chrome/Edge/Brave without falling back to mac/linux paths. - BrowserPlatform extended to win32; windowsDataDir for Chrome/Edge/Brave/Chromium - getBaseDir('win32') -> ~/AppData/Local - findBrowserMatch checks Network/Cookies first on Windows (Chrome 80+) - getWindowsAesKey() reads os_crypt.encrypted_key from Local State JSON - dpapiDecrypt() via PowerShell ProtectedData.Unprotect (stdin/stdout) - decryptCookieValue branches: AES-256-GCM (Windows) vs AES-128-CBC (mac/linux) - v20 App-Bound Encryption detected with specific error code - CDP-based extraction fallback (importCookiesViaCdp) for v20 cookies - launches Chrome headless with --remote-debugging-port on real profile - extracts via Network.getAllCookies over CDP WebSocket - requires Chrome closed (v20 keys path-bound to user-data-dir) - Inline security comment on CDP debug port (127.0.0.1 only, collision-avoid port) - One-time Chrome version log on CDP entry via /json/version - openDb() uses copy-then-read on Windows (exclusive locks) - Hardcoded /tmp -> TEMP_DIR from platform.ts in openDbFromCopy - cookie-import-browser.ts:494 error rethrow preserves underlying message Upstream: garrytan/gstack@1211b6b (community wave v0.18.1.0) Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Ports gstack garrytan#996 (msr-hickory) + the v20 CDP fallback wiring from garrytan#892. - Add hasActivePicker() export — returns true while any picker code or session is still live. Used by server watchdog to avoid tearing down the server mid-import and stranding the picker UI with "Failed to fetch". - Picker import path now auto-falls back to importCookiesViaCdp() when standard decrypt returns zero cookies AND hasV20Cookies() signals v20 App-Bound Encryption. Surface actionable error with close-chrome hint when CDP fallback itself fails. Rename kept: cavestack_picker cookie name (vs gstack's gstack_picker). Upstream: garrytan/gstack@1211b6b Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Ports gstack garrytan#994 (joelgreen) + garrytan#1020 gating + SIGINT wrapper fix (garrytan#1025). The browse server was dying between Claude Code Bash tool invocations because: 1. SIGTERM: Claude Code sandbox sends SIGTERM to child processes when a Bash command completes. Server called shutdown() and deleted state file. 2. Parent watchdog: 15s BROWSE_PARENT_PID poll shut down on parent death. Both made every new $B call start a fresh server with no cookies or state. Fix (scoped to normal headless mode): - SIGTERM handler: in headless mode, log and ignore. Idle timeout (30min) handles eventual cleanup. Headed/tunnel modes respect SIGTERM and shutdown cleanly (idle timeout early-returns in those modes, so ignoring would leak). - Parent watchdog: same three-way resolution. Active cookie picker always stays alive to avoid stranding picker UI with "Failed to fetch". - SIGINT (Ctrl+C) always shuts down — intentional stop. - Outer IS_HEADED_WATCHDOG gate skips watchdog entirely in headed mode. - shutdown() called via arrow wrapper so signal name arg doesn't become NaN exitCode (fixes legacy exit-code-1 bug). Imports hasActivePicker() from cookie-picker-routes (added in prior commit). /open-cavestack-browser path used in place of /open-gstack-browser in inline docs. Upstream: garrytan/gstack@1211b6b Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Ports gstack garrytan#993 (byliu-labs) + its follow-up extending coverage to design resolver + 5 more skill templates. Shell variable assignments that use a bare \`~/\` (rather than a quoted "\$HOME/...") trigger Claude Code's sandbox permission prompt on every run. Replace with explicit \$HOME expansion. Source-of-truth fixes: - scripts/resolvers/design.ts — 3 spots (D=, B=, _DESIGN_DIR=) - cavestack-upgrade/SKILL.md.tmpl — _SNOOZE_FILE= - design-consultation/SKILL.md.tmpl — _DESIGN_DIR= - design-review/SKILL.md.tmpl — REPORT_DIR= - design-shotgun/SKILL.md.tmpl — _DESIGN_DIR= - plan-design-review/SKILL.md.tmpl — _DESIGN_DIR= Uses the same .replace(/^~/, '') pattern from upstream for ctx.paths.* substitutions so rendered output becomes "\$HOME/..." across all hosts. Upstream: garrytan/gstack@1211b6b Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Regenerate claude-host SKILL.md files to absorb the tilde fix from scripts/resolvers/design.ts and the 5 edited .tmpl files. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Normalizes YAML frontmatter on the 4 hand-authored OpenClaw skills so stricter parsers like Codex can load them. - Drops non-standard version + metadata fields - Rewrites descriptions into simple "Use when..." form (no inline colons) - Adds regression test enforcing strict frontmatter (name + description only) Ported from garrytan#864 (cathrynlavery); adjusted for cavestack- prefix. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Cavestack divergence from upstream caveman. Users get a single compression level — classic caveman "full" — with no intensity toggle. Removes: - /caveman lite and /caveman ultra commands (mode tracker) - lite/ultra entries from VALID_MODES (config validator) - caveman-lite.json and caveman-ultra.json voice profiles - lite/ultra rows, examples, intensity tables in SKILL docs - lite/ultra mentions in README, web docs, setup hint, voices/README - lite/ultra tests in voice-density, caveman-voice-verify, skill-validation Wenyan variants (wenyan-lite, wenyan-full, wenyan-ultra) and sibling skills (caveman-commit/review/compress) stay untouched. "stop caveman" / "normal mode" continues to disable per session; "/caveman" re-enables to full. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Follow-up to 06e5f97. CLAUDE.md section described the tracker as handling "/caveman lite|full|ultra" toggles, which is no longer true after the lock. Updated to describe the locked-full behavior and the surviving wenyan + stop paths. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Test was checking for LF-only sentinels in file content, but Windows git checkouts produce CRLF line endings for tracked .md files. Normalize to LF before checking boundaries so the test works across platforms. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Ports 5 upstream PRs from garrytan#1028 (community wave v0.18.1.0): - garrytan#892 Windows cookie import (DPAPI + v20 CDP fallback) - garrytan#996 Cookie picker survives CLI exit - garrytan#994 + garrytan#1020 SIGTERM persistence + headless-mode gating - garrytan#993 Tilde-in-assignment fix (design resolver + 5 templates) - garrytan#864 OpenClaw native skill frontmatter normalization (codex-friendly) Plus local change: caveman mode locked to full (lite/ultra removed). Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
… pwsh fallback, key zeroing, temp cleanup) Adversarial review turned up 5 weaknesses in the ported Windows cookie import code. All fail visibly (no silent corruption) but every one is addressable. Tree-kill on Windows (killProcessTree helper) Bun.spawn.kill() uses TerminateProcess and does not reap child processes. Chrome spawns renderer/GPU/utility procs that survive as orphans and keep the profile locked. We now route every kill path (DPAPI PowerShell, macOS Keychain, secret-tool, Chrome headless) through taskkill /F /T /PID on Windows with a fallback to proc.kill elsewhere. Chrome debug port discovery via DevToolsActivePort Replaced the random [9222,9321] port picker with --remote-debugging-port=0. Chrome writes the chosen port to DevToolsActivePort in user-data-dir; we poll the file until it appears. Avoids the silent-timeout failure mode where two Chrome-based tools collide on the same random port. powershell.exe / pwsh.exe resolution Hardened Windows installs may ship without powershell.exe (deprecated on Win11) or without pwsh.exe (no PowerShell 7). Added findPowerShellExe() which probes PATH + canonical install locations for both, and raises a clear keychain_error if neither is found. AES key zeroing + encrypted-blob cache Refactored the key cache so Windows caches the DPAPI-encrypted blob (safe at rest — DPAPI binds to the current user) instead of the decrypted AES-256 key. getDerivedKeys now returns a SessionKeys wrapper with ephemeral buffers that zeroSessionKeys() scrubs in the finally block of importCookies. macOS / Linux still cache the derived key (Keychain prompt makes re-derivation expensive, and the threat model there is the same as the source DB). Temp DB sweep on crash Cookie DB copies land in os.tmpdir() and are normally unlinked when Database.close() runs. A process crash (SIGKILL, uncaught exception) bypassed cleanup and left .db / .db-wal / .db-shm behind. Added an exit + uncaughtException + unhandledRejection sweep driven by a tempDbPaths Set so crashes no longer leak temp files. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
On Windows, PIDs recycle aggressively. The parent-process watchdog was polling `process.kill(BROWSE_PARENT_PID, 0)` every 15s; if the OS reassigned the parent's PID to an unrelated new process within that window, the existence check kept succeeding and the server stayed alive forever thinking its parent was still around. Now capture the parent's creation time at server launch (WMIC on Windows, `ps -o lstart=` on mac/linux) and compare it on every poll. Mismatch means PID reuse, so we treat the parent as exited even though the PID still resolves. Logged with a "(PID reused)" note so the shutdown reason is explicit. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
The v20 CDP fallback must launch Chrome headless against the user's
real profile directory because v20 App-Bound Encryption keys are
path-bound to user-data-dir. If Chrome is force-killed mid-launch
(SIGKILL, machine sleep, /ship interrupt) the profile's Preferences
or Local State can be corrupted. Risk is low but non-zero.
Added a GET /cookie-picker/preflight endpoint that runs hasV20Cookies
and returns { v20Detected, warning }. The picker UI calls it after
loadDomains(). When v20 is detected, the banner is upgraded to a
warning style and import actions are gated behind a confirm dialog
(ensureV20Confirmed) so the user explicitly acknowledges the risk
before Chrome is launched against their real profile. Ack persists
per browser+profile; switching either resets it.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Covers the new behavior introduced by the Windows hardening pass:
- /cookie-picker/preflight requires browser param, returns the
{v20Detected, warning} shape, and gates on auth
- hasV20Cookies returns false on non-Windows and unknown browsers
without throwing
- importCookies key reuse survives the SessionKeys wrapper (second
sequential import on the same profile decrypts to the same value)
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
…throwing Maintainability pass after adversarial review: - extractCookiesViaCdp shared a 15s deadline across two sequential poll loops. If the first (DevToolsActivePort write) took 8s, the second (page target discovery) silently shrank to 7s and could time out unrelated to actual Chrome state. Give target-discovery its own fresh 15s budget. - ensureTempDbExitHandler re-threw from uncaughtException and unhandledRejection after sweeping, which masked any other listener's behavior and could double-report the error. Attach sweep-only listeners; let Node's default handler produce the diagnostic and terminate. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Watchdog ticks every 15s and spawned wmic / PowerShell / ps synchronously on each tick to read the parent's start time. On Windows this blocked the event loop ~50-200ms per tick, starving inbound HTTP requests and delaying the Playwright driver's own I/O. Switch getProcessStartTime to Bun.spawn + await proc.exited so the probes run concurrently with everything else the server is doing. Cache the launch-time start-time in a Promise so every tick resolves the same value without re-spawning a probe at startup. Guard against overlapping ticks with a watchdogBusy flag (belt-and-suspenders if a single probe ever exceeds 15s). Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
The heuristic "if every v10 decrypt failed AND the DB contains v20 bytes, try CDP" was duplicated between the picker route and the write-commands direct-import path. Drift risk — one site could change the trigger condition while the other kept the old behavior. Extract importCookiesWithV20Fallback in cookie-import-browser.ts as a single exported entry point. Picker route keeps its custom "v20_encryption" JSON error response by catching the CDP exception; write-commands loses a few lines and is easier to read. No behavior change on the happy path. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
hasActivePicker gates the SIGTERM handler and the parent-PID watchdog shutdown path. A regression that made it always return false would silently tear down the server mid-import (stranding the picker UI). A regression that made it always return true would keep the server alive forever. Add three direct tests: baseline boolean, TTL-live code, and session established via a route exchange. Previously this behavior was only tested indirectly through higher-level route + signal tests. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
The initial v1.2.2.0 entry covered the 5-PR upstream port wave. This branch also applied 7 adversarial-review hardening fixes (tree-kill, port discovery, PID reuse guard, temp sweep, pwsh fallback, key zeroing, preflight UI warning) plus 3 maintainability fixes (async watchdog, CDP poll deadline, v20 fallback DRY). Regenerate the entry to cover all 18 commits on the branch. No version bump — same branch, same shipping artifact. Date moved to 2026-04-18 since the hardening + regeneration landed today. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Ports 5 upstream PRs from garrytan/gstack's community wave v0.18.1 plus
applies 7 hardening fixes surfaced by adversarial review and 3
maintainability fixes from the specialist pass. 18 commits, one v1.2.2.0
release.
Windows cookie import + CDP v20 fallback (PRs garrytan#892, garrytan#996, garrytan#993, garrytan#994, garrytan#1020)
auto-discovery under
%LOCALAPPDATA%\Google\Chrome\User Data, Chrome80+ cookie path handling (
profile/Network/Cookies), AES-256-GCMWindows decrypt with AES-128-CBC mac/linux branch, v20 App-Bound
Encryption detection.
Chrome headless on the real profile and extract cookies via
Network.getAllCookiesover CDP. Real profile is unavoidable — v20keys are path-bound to user-data-dir.
hasActivePicker()gate keeps the picker UI alive whilecodes/sessions are pending, shutting down cleanly after all expire.
mode ignores SIGTERM so the server survives Claude Code's Bash sandbox
killing the parent shell between tool invocations. Headed + tunnel
modes still shut down cleanly on SIGTERM.
scripts/resolvers/design.tsand 5 skill templates so shell setup lines don't trigger Claude Code
permission prompts.
OpenClaw native skill frontmatter normalization (PR garrytan#864)
Codex CLI can parse them. Dropped non-standard
version+metadatafields; rewrote descriptions into "Use when..." formwithout inline colons.
Caveman voice locked to full (local)
/caveman liteand/caveman ultraremoved.caveman-lite.jsonandcaveman-ultra.jsonvoice profiles deleted. Wenyan variants stillavailable.
Hardening (adversarial review response)
taskkill /F /T /PIDon Windows so Chrome renderer/GPU/utility children don't orphan.
with
--remote-debugging-port=0+ file-polling the chosen port.(WMIC / PowerShell /
ps -o lstart=) and rejects recycled PIDs.handlers unlink temp cookie DB copies.
and zeros decrypted plaintext after each import.
risk disclosure; user confirms before Chrome opens).
Maintainability
Bun.spawnawait proc.exitedso the 15s tick doesn't block the event loop50-200ms on Windows.
no longer share one 15s budget.
importCookiesWithV20Fallback()helper consolidates the v10→CDPheuristic duplicated between the picker route and CLI import.
uncaughtException/unhandledRejectionsweep-only, let Node's default diagnostic fire.Test Coverage
openclaw frontmatter).
hasActivePicker()lifecycle,/cookie-picker/preflightendpoint shape + auth,
hasV20Cookiescross-platform safety, sessionkey reuse through the
SessionKeysrefactor.working directory — identical count on
main; documented inCLAUDE.md). Zero in-branch regressions.
Pre-Landing Review
(orphan Chrome processes, port collisions, PID reuse, temp file leak
on crash, hardened-Windows missing powershell.exe, AES key memory
lifetime, profile corruption risk). All 7 fixed with follow-on commits.
Testing: 1 MEDIUM auto-fixed (hasActivePicker tests); 3 informational
on harder-to-reach internal paths noted but not addressed.
Maintainability: 3 auto-fixed (cdp timeout, watchdog sync, v20
fallback DRY), 1 false positive (hasActivePicker race — Node is
single-threaded), rest skip.
Eval Results
Skipped. Diff touched voice profile infra and SKILL regens; user opted
to skip the paid eval suite since the changes are deletions +
OS-specific + frontmatter normalization rather than new prompt
semantics. Blast radius on prompt quality is low.
Scope Drift
CLEAN. Intent: port gstack v0.18.1 + caveman lock + harden surfaced
adversarial findings. Delivered: exactly that.
TODOS
None completed. No scope additions.
Test plan
bun testpasses on touched files (54/54 targeted)1.2.2.0 → 1.2.1.0 → 1.2.0.0 → …)install; deferred to post-merge — cannot reliably test on current
path-with-spaces working dir).
🤖 Generated with Claude Code