New scripts/build-mcpb.mjs packs a Claude Desktop extension bundle
(scripts/mcpb-wrapper.mjs + manifest + icon) into dist-mcpb/tiletopia.mcpb.
The wrapper reads the bearer token from %APPDATA% at launch and execs
`npx -y mcp-remote`, so no secrets are baked in and Regenerate keeps
working transparently. Run via `pnpm run build:mcpb`.
McpPanel gets a "Download .mcpb" button linking to the releases page; the
help-overlay tip and README MCP section both lead with the bundle install
path and keep the .mcp.json shim recipe as the Claude Code fallback.
Session-log entry in memory.md covers the design choices, especially why
the wrapper-script approach beat the alternatives (user_config prompt
would defeat one-click; baked-in token would be wrong for everyone else).
Three of the highest-power v2 tools, plus a defense-in-depth pass
on SSH-specific risk.
write_pane sends keystrokes (or any bytes) to a pane's PTY. The
policy engine matches against the text content directly so rules
like write_pane(npm test*) match by what would run, and the
compiled-in hard-deny catches rm -rf /, fork bombs, etc. regardless
of policy. Per-pane token-bucket rate limiter (30 calls / 10s,
3/sec refill) prevents a runaway loop from spamming the user with
confirm modals or burning audit-log capacity. The frontend handler
truncates the text in modal/audit summaries to ~60 chars + escapes
control characters so secrets pasted into write_pane don't echo
verbatim into the UI.
spawn_pane mirrors the existing SpawnSpec enum (WSL distro,
PowerShell, SSH) as the tool schema. New splitLeafWith helper
inserts a caller-built LeafNode (with a pre-generated id) so the
handler can await waitForPaneRegistration on that exact leaf before
replying with the resulting {leafId, paneId}. 15s spawn timeout
covers cold-start WSL distros; 30s for connect_host covers SSH
handshake + auth. Outer dispatch timeout bumped 30s → 60s. SSH
spawns without a saved hostId are refused — LeafNode only persists
sshHostId, no inline params, so use connect_host.
connect_host is a thin wrapper that looks up a saved SSH host by
id and routes through the same spawn machinery.
McpConfirm.tsx gains an optional ssh context — when the call
targets or spawns an SSH pane, a red warning banner renders
explaining that pattern matching is best-effort on the bytes we
send (remote shell expands aliases/subshells before executing).
buildConfirmSummary became buildConfirmInfo and returns the SSH
context alongside the summary string.
PR-3.5 — SSH safeguards. Two new switches in the Policy tab,
both off by default, both gated by mcp_policy::SshSafeguards:
allowOpenSsh: when off, connect_host and spawn_pane(kind=ssh)
refuse server-side with a clear "ssh-disabled" message pointing
at the Policy tab. User must open SSH manually via the titlebar
🔑 picker and toggle 🤖 on to grant Claude access.
autoAllowSpawnedSsh: when off, an SSH pane Claude spawns starts
with mcpAllow=false. User must explicitly toggle 🤖 before
Claude can read scrollback or send keystrokes. The second switch
is disabled in the UI when the first is off.
The safe-by-default design means a fresh install gives Claude no
ability to autonomously touch SSH — full safety with one click per
level to enable when consciously wanted. Both switches read fresh
per call so policy edits take effect without a server restart.
ErrorBoundary.tsx — last-resort guard against React render
exceptions. Wraps the App root + each MCP panel tab independently
so a bug in one tab doesn't blank the entire app. Shows a small
red error card with the exception message and a "Try again"
button. Caught a serde rename_all bug during PR-3.5 testing where
PolicyTab read policy.sshSafeguards but Rust serialized
ssh_safeguards (snake_case); without the boundary the whole window
went black.
newId() now exported from tree.ts for the splitLeafWith path.
McpPolicy struct gained #[serde(rename_all = "camelCase")] so
sshSafeguards survives the IPC round-trip cleanly; older policy
files without the field still load (serde defaults to safe).
App.tsx now listens on "mcp://request" and resolves each call:
needsConfirm=true queues a confirm modal (Accept/Reject, or
"Always allow <tool>" which appends the bare tool name to the
policy's allow bucket on the fly); needsConfirm=false runs straight
through. Replies via mcp_action_reply with externally-tagged
Result. The only wired-up tool for now is set_label, which delegates
to the existing ops.setLabel path.
McpConfirm.tsx (new) — themed amber-bordered modal sibling to the
existing overlays. Enter = accept, Esc = reject. Shows tool, the
policy reason that triggered the prompt, a human-readable summary
("Rename pane X → Y"), and an expandable raw-args section.
Audit log: subscription lifted from AuditTab up to App.tsx so events
fired while the panel is closed (or on Config/Policy tab) still land
in the ring. AuditTab becomes presentational; McpPanel forwards
entries + clearAudit + computes the unread badge from a baseline
seen-count.
StrictMode race fix: both new App-level listeners (mcp://audit and
mcp://request) use the cancelled-flag pattern so a late-resolving
listen() Promise after a strict-mode pretend-unmount tears itself
down instead of leaking a second subscription. Previously this
manifested as duplicate audit rows and a need-to-click-twice on
modal buttons.
Foundation for Claude-drives-the-workspace writes. Nothing wired
end-to-end yet (App.tsx dispatcher comes next); this lands the
machinery + UI.
mcp_policy.rs (new) — three-tier allow/ask/deny policy with
deny-first precedence and a compiled-in non-overridable hard-deny
list (10 patterns covering rm -rf /, fork bombs, mkfs on device, dd
to raw disk, /etc/passwd overwrite, curl|sh, chmod -R 777 /, etc.).
Shell-operator-aware glob matcher mirroring Claude Code's Bash(*)
syntax. Restrictive default — empty policy means every non-hard-
denied call falls to Ask. Persisted to mcp-policy.json in
app_config_dir. Includes a PolicyClassifier scaffold (no-op) for a
future v2.1 LLM-classifier hook. 1152 lines incl. ~100 unit + fuzz
tests covering the matchers and lookalike negatives.
mcp.rs — TileService now holds AppHandle + Arc<PendingActions>
(oneshot registry keyed by uuid). New async dispatch_action helper
runs the policy check, emits "mcp://request" for the frontend to
handle, awaits a oneshot reply (30s timeout), then emits "mcp://
audit" with the outcome regardless. set_label tool wired through
this path as the demo for PR-1b's dispatcher.
commands.rs / lib.rs — new Tauri commands mcp_action_reply,
mcp_policy_load, mcp_policy_save; PendingActions registered as
managed state.
McpPanel.tsx — refactored into Config / Audit / Policy tabs.
AuditTab listens on mcp://audit, keeps a 200-entry ring with
ok/denied/failed chips. PolicyTab edits the allow/ask/deny buckets
(stacked vertically — three columns overflowed the panel) and shows
the hard-deny rules read-only at the bottom with "Cannot be
disabled" badges. Themed scrollbar on mcp-body to match xterm panes.
Caveat: set_label calls from Claude will currently time out — the
App.tsx side that listens on mcp://request and replies via
mcp_action_reply lands in PR-1b.
Co-authored by Sonnet (policy engine, backend plumbing, panel UI)
and Haiku (hard-deny fuzz test suite); integration + bug fixes here.
