Publish jobs now only run when the commit message contains:
- "deploy: frontend" — deploys frontend only
- "deploy: backend" — deploys backend only
- "deploy: backend, frontend" or "deploy: frontend, backend" — both
Case-insensitive matching. Commits without a deploy directive
will only run CI (check/fmt/clippy), not publish.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- Overlay now reliably dismissed: use window.setTimeout and inline
overlay hide instead of separate function that could lose DOM ref
- Minimum 1s duration to prevent flash for short hotkey sequences
- Port dropdown preserves user's selection across page refreshes
by tracking selectedPort locally instead of always using device's
active_port
- Prevent duplicate change listener on port select across remounts
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
When switching ports via the console dropdown:
- Input (keyboard, mouse, wheel) is suspended immediately to prevent
stray events from interfering with the OSCAR hotkey sequence
- A semi-transparent overlay with spinner and countdown timer appears
over the console canvas
- Duration is calculated from the actual key_pause_duration setting
multiplied by the number of * pause tokens in the port's hotkey,
plus a 500ms buffer
- Input resumes and overlay disappears when the timer expires
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
On load, check sessionStorage for stored credentials (saved during
login). If present, auto-authenticate and skip the login form. Falls
back to showing login if the stored credentials fail.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Three fixes for KVM port management:
1. HTML unescape scraped values — the > character in hotkey strings
(e.g., PrintScreen>0>1) gets entity-encoded to > in the device's
HTML. Added html_unescape() to the scraper so hotkeys round-trip
correctly.
2. Send hotkeys over WebSocket — port switching via the Belkin web form
only changes the active port number, it doesn't send the hotkey
sequence to the downstream KVM. Now when switching ports from the
console dropdown, blekin parses the Belkin hotkey syntax and sends
the key press/release sequence over the existing WebSocket connection.
3. New hotkey.ts parser — converts Belkin hotkey syntax to scancode
sequences:
- > and -> = sequential (tap each key)
- + = simultaneous (hold all, release in reverse)
- Supports all key names: PrintScreen, Ctrl, Alt, Shift, F1-F12,
letters, digits, navigation keys, numpad
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
The Belkin firmware requires:
1. A hidden __templates__ field in all form POSTs
2. Image button actions submitted as name.x=0&name.y=0 (not name=value)
3. ECG_kvm_powerport_cnt hidden field
Without these, the device returns "Permission denied" on port config
saves. Also fixed the switch endpoint with the same pattern.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
When the port count dropdown changes, the table now re-renders with
the new number of rows, preserving existing values for ports that
were already visible. This allows configuring ports > 16 after
switching to a higher port count.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Adds a "Send Key" dropdown menu to the console toolbar for keys that
browsers intercept before JavaScript can capture them (Print Screen,
Scroll Lock, Pause/Break). Also includes Escape, Tab, Caps Lock,
Num Lock, and Ctrl+Alt+Del.
Each menu item sends a press+release scancode pair directly over the
WebSocket, bypassing browser key capture entirely. This enables
activating downstream KVM switch menus (e.g., Avocent OSCAR via
Print Screen) from the blekin interface.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Backend (crates/ericrfb-proxy):
- Session cookie now persisted in AppState for device API calls
- New kvm.rs with three REST endpoints:
GET /api/kvm/ports — scrapes kvm.asp, returns port config as JSON
PUT /api/kvm/ports — saves port names, hotkeys, visibility, count
POST /api/kvm/switch — switches active KVM port via home2.asp
- HTML scraping extracts form values from predictable firmware HTML
Frontend (crates/ericrfb-frontend):
- New shell.ts: sidebar navigation with page routing pattern
(Console, Ports — extensible for Virtual Media, Users, etc.)
- Console refactored into pages/console.ts with mount/unmount lifecycle
- Port switcher dropdown in toolbar (fetches port list, switches on change)
- WebSocket auto-reconnects after port switch
- New pages/ports.ts: editable port configuration table
- Port count, key pause duration, per-port name/hotkey/show-in-console
- Save, reload, and per-port switch buttons
- Active port highlighted
- Dark theme sidebar with active state indicators
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
The mapping was built assuming keynr followed physical keyboard order
with Escape=0. In reality, KeyTranslator.java maps Java VK_* codes to
keynr values with a different layout:
- keynr 0 = Backquote (not Escape)
- keynr 59 = Escape
- keynr 27 = Enter (not 40)
- keynr 40 = Backslash (not 42)
This caused the number row and QWERTY row to be off by 1 (Escape
was inserted at position 0, pushing everything). The home row
(CapsLock=28, A=29...L=37) happened to align by coincidence, which
is why 'g'(33) worked but 'r'(should be 18, was 19=T) didn't.
Also fixes: F1-F12 off by 1, PrintScreen/ScrollLock/Pause off by 1,
numpad operator keys swapped with numpad 7/8/9.
Corrected both Rust (input.rs) and TypeScript (input.ts) mappings
to match the authoritative KeyTranslator.java VK→keynr table.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
When pal_selector=0 (no predefined palette), the Java code reads two
separate full bytes for the 2-color palette (line 421-422 of
ByteColorRFBRenderer.java). We were reading one packed byte and
splitting into nibbles, consuming 1 byte too few and misaligning all
subsequent reads including the zlib compressed data.
This caused "deflate decompression error" on stream 1 because the
zlib header (78 da) was offset by 1 byte.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
codec/iip.rs:
- TileCache: (fb_width/16 × fb_height/16) tiles, 8 versions × 256 bytes
each, matching t.java's (8, 16*16) allocation
- TileEntry: versioned read/write at byte offsets within tile data
- decode_iip() handles all 4 modes from the control byte:
- Mode 0/12 (cache-read): tile control bytes select which cached
version of each 16x16 tile to display, no new pixel data on wire
- Mode 4 (write-only): Tight-decoded pixel data written to cache
- Mode 8 (update+read): conditionally writes new data to cache
(bit 7 of control byte = 0 means update), then reads from cache
- Tile control bytes compressed via zlib (varint length) when >= 12
- Sub-types 1-4/8 map to bit-depths 1/2/4/4/8 bpp
- Cache resized on framebuffer resize (ModeChange msg 128)
Wired into session dispatch for encoding 9. Not advertised in default
encoding list — only active if explicitly requested.
codec/tight.rs: made get_or_init() pub for IIP's zlib access.
39 tests passing.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Frontend reconnection:
- WebSocket auto-reconnects with exponential backoff (1s → 30s)
- Re-authenticates with OmniView to get fresh APPLET_ID on reconnect
- Credentials stored in sessionStorage for automatic re-login
- Status bar shows connection state and reconnect countdown
- Disconnect button returns to login screen
Encoding 10 (Raw with tile interleave):
- codec/raw_tile.rs: decodes encoding 10 per ByteColorRFBRenderer.for()
- Flag byte bit 0 selects plain raw vs 16x16 tile-interleaved data
- Deinterleave handles edge tiles smaller than 16x16
- Wired into session dispatch
- 2 unit tests
39 tests passing.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Simplified deploy steps — rsync directly to final paths using
sudo rsync on the remote instead of temp-file staging.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
publish.yml — triggered on push to main, two parallel jobs:
frontend:
- Builds Vite frontend (fnm + npm ci + npm run build)
- Rsyncs dist/ to gitea_ci@UI_HOST:UI_PATH/
- Rsyncs nginx config to UI_HOST, creates sites-enabled symlink,
runs nginx -t && systemctl reload nginx
backend:
- Builds release binary (cargo build --release -p ericrfb-proxy)
- Stops blekin.service on WS_HOST
- Rsyncs binary to WS_HOST:/usr/local/bin/ericrfb-proxy via sudo rsync
- Rsyncs systemd unit to WS_HOST:/etc/systemd/system/blekin.service
- Enables and starts the service
asset/nginx/blekin.kosherinata.internal.conf:
- Serves static frontend from UI_PATH
- Reverse proxies /api/ to frootmig:3000 with WebSocket upgrade
- 24h read/send timeouts for long-lived KVM sessions
asset/systemd/blekin.service:
- Runs ericrfb-proxy with BLEKIN_HOST=10.3.0.130
- Restart on failure with 5s backoff
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
codec/hextile.rs:
- Full Hextile (encoding 5) decoder per ByteColorRFBRenderer.int()
- Handles: Raw tiles, BackgroundSpecified, ForegroundSpecified,
AnySubrects, SubrectsColoured flags
- Background/foreground colors persist across tiles
- 4 unit tests covering all subencoding paths
framebuffer.rs:
- Added fill_rect() for Hextile background/subrect fills
session.rs:
- Wired Hextile encoding 5 into the rect dispatch
examples/record.rs:
- 30-second (configurable) recording session
- Saves 1 PNG per second to out/ directory
- Requests encodings [5, 1, 0] (Hextile, CopyRect, Raw)
- Tested against real OmniView: 10 frames in 10s, no errors
27 tests passing.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
The OmniView defaults to 16bpp when no SetPixelFormat is sent.
The Java applet sends SetPixelFormat (msg type 0) at
ByteColorRFBRenderer.new() line 76 to request 8bpp RGB332
(bpp=8, depth=8, true_color, red_max=7, green_max=7, blue_max=3,
shifts 0/3/6). Without this, pixel data arrives as 16bpp pairs
that produce green stripe artifacts when interpreted as 8bpp.
Verified: snapshot now matches the reference screenshot from
http://10.3.0.130/screenshot.jpg (dark screen with "Free" label).
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Fixes CI fmt check failure — rustfmt wants multi-line assert_eq! for
long struct literals and the varint roundtrip assertion.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Three parallel jobs on the rust runner:
- cargo check --workspace
- rustfmt --check on all .rs files
- cargo clippy --workspace -- -D warnings
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Cross-referenced every wire-format claim against rc-src and corrected
several errors that would have caused stream desync during implementation:
- rect header is 4×u16 + i32 (12 bytes), not varints + u8
- ping payload is i32 (4 bytes), not 1 byte
- PointerEvent is 8 bytes (includes trailing extra_u16)
- ServerInit is 19 bytes, not 16
- pixel-format struct is variable-length, not fixed 25 bytes
- string primitive uses modified-UTF-8, not ISO-8859-1
- bandwidth probe is read-only (no echo response)
- Phase 1 uses sync Read traits (async deferred to Phase 2)
- corrected all stale line-number references in ac.java/aw.java
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
