v4.11.0

Added

• ACP server adapter (ccc_acp.py). Exposes CCC over the Agent Client Protocol so editors and ACP clients (VS Code, JetBrains, Zed, Hermes) can drive Claude Code sessions over JSON-RPC stdio. Optional: requires pip install agent-client-protocol and is launched separately (python3 ccc_acp.py); the core stdlib-only server is unaffected.
• Added a standalone group-chat live view with chat-style reading, posting, nudging, and a searchable participant picker.
• Kilo Code CLI engine. Spawn headless Kilo Code CLI runs as a first-class engine alongside Claude, Codex, Cursor, and Antigravity. New /api/sessions/spawn-kilo and /api/sessions/spawn-kilo/availability endpoints, an engine entry in the spawn dropdowns, and worktree support. Override the binary with CCC_KILO_BIN and the default model with CCC_KILO_MODEL.
• Kilo Code session ingestion. Externally-launched Kilo Code sessions now appear on the board alongside Claude/Codex/Cursor/Antigravity. CCC reads Kilo's SQLite store (~/.local/share/kilo/kilo.db) read-only via find_kilo_conversations, surfaces each session's title, cwd, model, and live status, and renders full transcripts (user / assistant / tool calls) through _parse_kilo_conversation. Detection (_is_kilo_session) now probes the DB so historical and terminal-launched sessions route to the Kilo transcript loader.
• Added a microphone button to both the main conversation input bar and the group chat input row, enabling users to dictate speech directly into the input textareas using the Web Speech API.
• Added support for new model releases (Gemini 3.5 Pro/Flash, Claude Sonnet/Haiku 4.8, OpenAI o4/o4-mini, and StepFun/GPT-6.0) in the user interface model selection dropdown.
• Added a premium step-by-step onboarding experience to guide first-time users through agent CLI detection (Claude Code, Antigravity, Cursor, Codex), account setup, and spawning their first session.
• /api/sessions/spawn accepts an optional report_to (return address) — the dispatching session's UUID. When set, the spawned session is instructed to POST one structured completion report (STATUS, summary, file paths, failure reason) back to that session via /api/inject-input when it finishes. Aliases return_to / reply_to accepted; value is validated as a session-id-shaped string.
• Public stats page at https://ccc.amirfish.ai/stats backed by a new read-only /v1/stats endpoint on the telemetry worker. Shows live aggregates only: distinct opt-in installs, active-today, anonymous boots per day, version + platform breakdown, and per-install sessions-in-last-24h (just an 8-char install_id prefix, never the full id). Worker caches the response 5 min at the edge so the page is hammer-safe.
• Added a dedicated Token Throughput Analyzer page (/throughput) and corresponding /api/throughput endpoint to measure active LLM processing speed (TPM/TPS) and turn durations.

Changed

• Draws a visual horizontal line indicator with a progress percentage label and a custom accent handle as the user drags horizontally to collapse/fold earlier conversation messages.

Fixed

• UX fixes queue worker rows now show the current queue item and latest queued item, updating when annotations add new fixes.