ByteStream.Feed() in UTF-8 mode permanently stalls on raw C1
control bytes (0x80-0x9F). DecodeRune returns RuneError, the loop
breaks, and the bad byte stays in the buffer forever — blocking
all subsequent input and leaking memory.
Stream.handleGround() has correct C1 handlers but they are
unreachable through ByteStream. Includes reproduction case and
suggested fix (Latin-1 fallback for invalid UTF-8 bytes).
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Full Go implementation under go/ replacing Python pyte with go-te:
- HTTP server with WebSocket protocol, SSE, screenshot SVG rendering
- PTY terminal sessions and Docker exec sessions
- Docker watcher (label-based container discovery + event stream)
- CPU stats collection with sparkline SVG rendering
- Session manager with TwoWayMap routing and replay buffers
- C1 normalization, DA filtering, identity generation, theme palettes
Audit fixes for 9 concurrency/correctness issues:
- HTTP transport leak: shared client pool for Docker socket calls
- WebSocket concurrent writes: all writes routed through send channel
- Closed channel panic: atomic.Bool guard on enqueueWSData
- GetFirstRunningSession: use UnsafeForward under SessionManager lock
- NewSession TOCTOU: re-check routeKey after re-acquiring lock
- waitErr data race: protect with mutex in both session types
- Replay buffer fragmentation: copy to new slice on eviction
- go-te dirty tracking: check screen.Dirty before incrementing counter
- Identity modulo bias: rejection sampling for uniform distribution
All Go tests pass (including -race). Python baseline unchanged (397 tests).
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
pyte 0.8.2 does not handle CSI S (SU — Scroll Up) or CSI T (SD —
Scroll Down). When TERM=xterm-256color, tmux sends CSI n S for bulk
scrolling instead of DECSTBM+index. pyte silently ignored these
sequences, leaving old content in the screen buffer — visible as
ghost content in SVG screenshots.
Fix: monkeypatch pyte's CSI dispatch tables to map S→scroll_up and
T→scroll_down, and implement both methods on AltScreen with proper
scroll-region (DECSTBM) support.
Adds 6 tests for SU/SD functionality.
Ink (React CLI framework) clears its output using repeated EL2+CUU1
sequences, one per previously-drawn line. When /clear resets Ink's
internal line counter, the next frame only erases a few lines instead
of the full previous output. In a real terminal the old content is in
scrollback and invisible, but pyte's fixed-size screen retains it,
producing ghost content (e.g. duplicated prompts) in SVG screenshots.
Added AltScreen.expand_clear_sequences() which detects runs of 3+
EL2+CUU1 pairs that don't reach row 0 and extends them to erase all
lines up to the top of the screen. Both DockerExecSession and
TerminalSession call this before feeding data to pyte.
Also made on_session_end() idempotent (contextlib.suppress KeyError)
to prevent a race when close_session() and natural session exit both
call it.
Added docs/ink-clear-fix.md with root cause analysis, byte-level
explanation, and reproduction script.
- close_session() now calls on_session_end() to remove the session from
sessions dict and routes, preventing zombie entries that persist after
the container is gone
- _remove_container() now closes the active session before removing the
app from apps_by_slug/apps, so session cleanup can still reference the
app during teardown
- Updated and added tests to verify session tracking cleanup
- handle DECSET ?47 as an alternate screen mode so tmux clear redraws don't overlay stale content in screenshots
- keep AltScreen mode checks aligned with 47/1047/1048/1049 variants used by full-screen TUIs
- document the screenshot debugging workflow in .github/skills/screenshot-debugging/SKILL.md for repeatable escape-sequence analysis
Allows per-container customization of the auto command. For example:
WEBTERM_DOCKER_AUTO_COMMAND='tmux new-session -ADs {container}'
This creates a tmux session named after the container instead of using
a fixed session name for all containers.
When a container has no labels, Docker returns {"Labels": null} in the
inspect response. The code was using .get("Labels", {}) which only
returns the default when the key is missing, not when it's null.
This caused _has_webterm_label() to raise TypeError, which was silently
caught by the event watcher's exception handler, causing it to reconnect
and miss the container start event.
Fixed by using .get("Labels") or {} pattern in:
- _handle_event() when processing start events
- _get_container_command() when extracting command label
- _get_container_theme() when extracting theme label
Added test for null labels case.
GitHub Copilot