CHANGELOG.md

Changelog

v0.1.17

• Hardened the telephony gateway runtime by splitting call lifecycle, ARI/event-hub, RTP bridging, playback, recording, inbound/outbound, cleanup, reporting, and MatrixClaw API code into smaller focused modules.
• Fixed telephony realtime playback and made call shutdown safer with root context cancellation, active-call pruning, setup-aware tool visibility, shared phone-number normalization, and stronger recording format validation.
• Improved call recording finalization by retrying stored-recording downloads with backoff, preserving ARI stop/download diagnostics, and logging nearby stored-recording candidates when Asterisk does not expose the expected file.
• Guarded managed browser setup from unsafe shell execution and made browser runtime startup use the managed Chromium executable consistently.
• Improved long-running session reliability by failing orphaned running runs, routing background workers through safego, recovering remaining background panics, refreshing Telegram typing indicators during active runs, and surfacing external runtime and subagent aftermath store errors.
• Split realtime voice and local voice control-plane flows into smaller option, setup, status, provider-selection, runtime, and action modules.
• Added architecture, browser, Telegram, testing, and refactoring documentation that records the current module boundaries and cleanup decisions.
• Reset the legacy broad test suite and added focused coverage for run recovery, subagent lifecycle, Telegram typing, managed browser guards, realtime/voice setup inputs, telephony phone normalization, and recording retry behavior.

v0.1.16

• Added provider-neutral realtime voice setup for Gemini Live and Grok Voice, including provider selection, API key validation, model/voice/language controls, and provider-specific status in the control plane.
• Added xAI Grok Voice Agent support for realtime speech-to-speech sessions, including language hints, manual audio turn commits, tool-call routing, and transcript handling tuned for cumulative Grok transcription events.
• Refactored control-plane navigation around consistent Back/Close behavior so menus, pickers, status views, and action dialogs return to their parent surface instead of leaking stale pickers or collapsing the menu stack.
• Added the optional matrixclaw-telephony-gateway binary for self-hosted Asterisk/SIP deployments, bridging ARI externalMedia RTP audio into MatrixClaw realtime voice sessions.
• Added approval-gated telephony_call tooling, outbound call objectives, inbound caller allowlists, phone-specific prompts, final call transcripts, post-call reports, and temporary MP3 call recording plumbing.
• Improved telephony runtime stability with faster inbound answering, a single long-lived ARI app listener, hangup-extension filtering, safer RTP/VAD turn handling, and cleanup for realtime close races.
• Updated release packaging, installer, uninstall script, local release build, and Homebrew template to include matrixclaw-telephony-gateway.
• Documented realtime voice providers, telephony gateway boundaries, SIP/PBX deployment assumptions, environment variables, and privacy considerations.

v0.1.15

• Reworked Telegram around normal private-chat sessions again: /new, /sessions, session use, and session deletion no longer depend on forum topics or message_thread_id routing.
• Added Telegram inline and guest delivery support so the bot can be invoked from other chats, edit the inline placeholder with the assistant answer, and answer guest-mode requests through the new Bot API flow.
• Improved Telegram live delivery with draft previews, a compact thinking placeholder, clearer running states, persistent inline request recovery, and fewer duplicate or stale delivery updates.
• Fixed Telegram voice/TTS routing so generated audio follows the originating chat or inline message, while the model only sends extra written text when it is actually useful.
• Added Telegram inline location handling so geolocation attached by Telegram is passed into the assistant request instead of being lost.
• Added browser module plumbing and managed browser MCP configuration paths, including daemon/client/control-plane APIs for browser provider state.
• Expanded storage APIs with base64 byte reads for durable and temporary files, and made storage not-found handling use typed errors instead of text matching.
• Made voice and session-LLM error handling rely on typed errors, removing more phrase-based diagnostics and locale-sensitive behavior.

v0.1.14

• Added smart web research with web_research, web_research_ask, and web_research_status, returning compact answers, facts, sources, warnings, next actions, and reusable research_id values.
• Added the shared internal/work storage layer with work_jobs, work_artifacts, and work_facts for heavy assistant jobs and large runtime artifacts.
• Moved new web research persistence onto the shared work tables, leaving raw page text, HTML, browser snapshots, and extraction artifacts out of the main provider prompt by default.
• Added deterministic web extraction for titles, snippets, page text chunks, schema-style facts, ratings, and review counts, with browser fallback support for dynamic or sparse pages.
• Updated web_fetch compatibility behavior: task mode now routes through web research extraction, while plain URL fetch returns compact diagnostics and artifact references instead of long raw excerpts.
• Kept web_search as a compact compatibility search tool and updated runtime guidance to prefer web_research for current, source-backed answers.
• Added MCP browser adapter wiring so web research can use a configured browser MCP server as its fallback renderer.
• Mirrored subagent lifecycle state into the shared work layer and made read_subagent_result return compact job summaries and refs instead of child transcript excerpts.
• Refactored web tool wiring around a single injected web service adapter, removing hidden global state between web_fetch, web_search, and web_research.

v0.1.13

• Added live Terminal subagent cards for delegate_task and spawn_subagent, with Matrix-style codenames, running/completed/failed/canceled states, expandable task text, and metadata previews.
• Added async subagent state merging so spawned background agents keep updating the original tool card after the spawn result is returned.
• Added queued busy input behavior: Enter queues while the assistant is busy, with /queue, /steer, /interrupt, and /busy commands for explicit control.
• Reworked the TUI status line to show the main model phase first and append active subagent or queued-input details.
• Added /context clear, clear markers, compact markers, context blocks, and corrected header token estimates based on effective post-clear context.
• Added persistent session input storage for queued/steered/interrupted user messages.
• Refined chat scrolling, viewport restoration, command pickers, permission rendering, and subagent/tool previews.
• Split local voice runtime management into Piper, Supertonic, and Whisper.cpp drivers with install/status coverage.
• Added README release highlights for the live-subagents/context release.

v0.1.12

• Added MatrixClaw subagents through delegate_task, with native child-session runs and external Codex/Claude Code runtime options.
• Added model-facing subagent guidance so assistants know when to delegate bounded work and which runtimes are available.
• Added subagent task persistence, parent/child session links, result delivery, and terminal rendering for delegated work.
• Added durable memory and assistant-facing memory tools, plus API, daemon-client, and controlplane support.
• Added session model/title improvements and external-agent runtime discovery updates.
• Added TUI self-restart support after daemon updates.

v0.1.11

• Refactored the Terminal UI stack by moving shared command-menu components into reusable surface components.
• Reworked command, picker, prompt, confirm, form, and info dialogs for more consistent rendering and navigation.
• Added dialog occlusion handling and simplified controlplane picker presentation.
• Tightened setup screen rendering, provider editing, storage/temp views, and module picker behavior.
• Updated CI lint configuration for Go 1.26 and limited golangci-lint to new issues.

v0.1.10

• Added daemon-first local voice modules: Text to Speech now supports Piper and Supertonic 3, while Speech to Text supports Whisper.cpp through the same shared module UI used by Terminal and Telegram.
• Added local voice runtime installation with install.sh --voice-runtime and scripts/install_voice_runtime.sh for Piper, Supertonic, Whisper.cpp, and ffmpeg.
• Added Run Per Task and Always Running modes for local voice providers. Run Per Task is the memory-saving default; Always Running keeps Piper, Supertonic, or Whisper.cpp warm for lower startup latency.
• Added online catalog-backed voice/model selection: Piper voices by language, Supertonic voice styles and language modes, and Whisper.cpp model tiers from tiny through large-v3.
• Added local Whisper.cpp speech-to-text execution through whisper-cli and whisper-server, with STT request limits sized for Telegram voice/audio uploads.
• Added voice status screens with installed storage, selected provider/model, runtime mode, and live RAM usage for managed local processes.
• Fixed local Piper text-to-speech so longer responses are generated without returning only the first chunk.
• Added Telegram voice delivery for TTS tool results and /tts, with generated audio saved into Matrixclaw Storage under telegram/audio/.
• Fixed Telegram TTS/STT daemon calls to use the long voice-runtime timeout instead of the short JSON timeout.
• Added storage/temp file documentation and kept Telegram-downloaded files in Matrixclaw storage with collision-safe names.
• Documented daemon-first architecture, local voice run modes, storage/temp files, Telegram voice/file flow, and open-source voice runtime installation.

v0.1.9

• Added OpenAI Codex subscription OAuth provider support and provider-login CLI plumbing.
• Added Telegram image/document upload handling backed by Matrixclaw storage, including temporary files and explicit save/delete controls.
• Added Telegram voice/audio transcription and text-to-speech delivery flows.
• Added daemon API and controlplane support for local voice modules.
• Added MCP, storage, automation, provider, and module command refinements.
• Added daemon stop controls, Piper runtime management, and process status helpers for local runtime processes.
• Improved Ubuntu install/runtime discovery, automation delivery fan-out to Telegram, and voice runtime activation guards.
• Documented Storage and Voice modules.

v0.1.8

• Added session capabilities so Matrixclaw and external-agent sessions expose only the controls that apply to their runtime.
• Marked Provider, Permission Mode, and Planning Mode as Matrixclaw-only for Codex sessions, with explicit explanations instead of silent no-op behavior.
• Refined the New Session picker copy for Matrixclaw and Codex runtime choices.
• Reworked Codex module options into editable Path and Enabled controls, with Enabled opening a standard picker and Path using the standard text prompt.
• Fixed external-agent path updates so changing the Codex binary path does not accidentally reset the enabled state.
• Renamed user-facing Goal/Plan labels to Planning Mode.

v0.1.7

• Added a persistent Planning Mode runner with SQLite-backed checkpoints for current item, last run, status, attempts, and errors.
• Changed plan execution from "model runs the whole plan" to one executable item at a time, with the daemon selecting the next leaf task.
• Added task/subtask execution semantics: parent items with open children are treated as sections and auto-close when all children are terminal.
• Made successful plan-run steps close the checkpointed item in core, reducing reliance on the model remembering to call plan_update_item.
• Kept blocked plan steps open and recorded blocked runner state instead of incorrectly marking work done.
• Improved TUI planning panel behavior, auto-run continuation, and plan summary display during multi-item execution.
• Documented Planning Mode architecture in docs/PLANNING.md.

v0.1.6

• Added TUI startup update checks against the latest GitHub Release, with a shared confirmation dialog and matrixclaw update CLI commands.
• Added update installation through the release installer and a follow-up TUI prompt to restart the daemon so the service binary is refreshed.
• Added /modules -> External Agents management with enable/disable controls, installed state, resolved binary path, mode, and version details.
• Moved external-agent daemon wiring into a built-in registry factory so future agents can be added without changing core session logic.
• Fixed Server Status back navigation by giving status info dialogs a return command and expanding the generic Info back key handling.
• Documented auto-update and external-agent management flows.

v0.1.5

• Improved Codex external-agent sessions: restored thread resume handling, normalized Codex tool activity into shared tool-call events, and preserved streamed text -> tool -> text ordering.
• Fixed TUI rendering so assistant text before and after tool calls is shown in the correct order without reimplementing Codex edits or diffs.
• Restored mouse-wheel scrolling in the terminal chat while keeping keyboard copy support for selected chat blocks.
• Cleaned up external-agent runtime plumbing and documentation so future runtimes can reuse the same event path.
• Refined TUI and Telegram command-menu parity.

v0.1.4

• Added Codex as an external agent runtime with app-server session attachment, CLI discovery/start commands, and daemon API support.
• Moved session architecture toward runtime-scoped settings: sessions now carry runtime, provider/model, and permission mode state.
• Added runtime-aware session creation for MatrixClaw and Codex in the shared controlplane used by TUI and Telegram.
• Moved Provider and Permission Mode out of the top-level menu and into session-scoped actions.
• Fixed Telegram provider switching callbacks and DeepSeek/OpenAI-compatible reasoning-content handling.

v0.1.3

• Replaced the repository-hosted README demo GIF with a GitHub attachment link and removed the large media file from git history.
• Changed empty-provider setup continuation to use the shared confirmation card.

v0.1.2

• Fixed macOS installer compatibility by removing a GNU-specific sed script from latest-release detection.
• Fixed installer cleanup after download failures so network errors do not trigger a secondary tmp: unbound variable failure.
• Added matrixclaw tui [WORKDIR] for opening a terminal session rooted at an explicit project directory, including external macOS volumes.
• Improved filesystem tool errors to show the active working directory when a requested path is outside the session root.

v0.1.1

• Improved provider setup and TUI provider editing: model pickers now open on the active model, tool-use pickers no longer show a misleading active marker, and provider edit dialogs keep consistent back/save navigation.
• Refreshed README positioning around matrixclaw as local personal AI infrastructure and moved README media assets under .github/assets.

v0.1.0

• Added daemon-backed terminal and Telegram clients.
• Added SQLite-backed sessions, runs, approvals, files, and client deliveries.
• Added setup flow for providers, daemon settings, Telegram, timezone, and assistant profile.
• Added automation jobs for reminders and scheduled AI tasks.
• Added release-readiness hardening for automation fires, SSE fan-out, Telegram monitoring, and daemon bind safety.
• Simplified setup/provider and storage API contracts to reduce daemon API/client drift.