CHANGELOG.md

Changelog

All notable changes to OpenCrab will be documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

[0.2.50] - 2026-03-04

Changed

• Config hot-reload via watch channel — Replaced per-channel Mutex copies of config (allowlists, voice, respond_to, idle_timeout) with a single tokio::sync::watch<Config> channel. All channels now read the latest config per-message from the watch receiver. Removed allowed_users/allowed_phones HashSet fields from all channel states and 4 separate allowlist callbacks in ui.rs
  • src/channels/factory.rs, src/channels/{telegram,discord,slack,whatsapp}/{mod,agent,handler}.rs, src/cli/ui.rs, src/brain/tools/{telegram,discord,slack,whatsapp}_connect.rs, src/brain/tools/whatsapp_send.rs
• TTS voice/model read from [providers.tts] — Added voice and model fields to TtsProviders so voice = "echo" under [providers.tts] is actually picked up. Previously serde silently ignored the field
  • src/config/types.rs, src/channels/factory.rs, src/channels/{telegram,discord,slack,whatsapp}/handler.rs
• Default TTS voice changed from "ash" to "echo"; both stt_enabled and tts_enabled now default to false (user must opt in)
  • src/config/types.rs

Fixed

• Telegram "Session not found" after TUI quit — The retry logic checked for "SessionNotFound" (camelCase) but the error Display produces "Session not found" (lowercase), so recovery never triggered. Now correctly matches and creates a fresh session (closes #24)
  • src/channels/telegram/handler.rs
• Duplicate data delivery on Telegram — LLM sent data both as streaming text response AND via telegram_send, resulting in the same content appearing twice. Added channel context prefix to all handlers telling the LLM its text response is auto-delivered (closes #23)
  • src/channels/{telegram,discord,slack,whatsapp}/handler.rs
• Groq API key in test fixture — Replaced real-format Groq key in redact_secrets test with obviously fake placeholder (closes #25)
  • src/utils/sanitize.rs

0.2.49 - 2026-03-04

Added

• Channel commands (/help, /usage, /models, /stop) — All four commands now work on Telegram, Discord, Slack, and WhatsApp. Shared commands.rs module handles parsing; each channel renders platform-native responses (inline keyboards, action rows, Block Kit buttons)
  • src/channels/commands.rs (new), src/channels/mod.rs, src/channels/telegram/handler.rs, src/channels/discord/handler.rs, src/channels/slack/handler.rs, src/channels/whatsapp/handler.rs
• /stop cancels running agent on channels — CancellationToken per session, equivalent to double-Escape in TUI. Immediately aborts streaming/tool loop mid-run
  • src/channels/telegram/mod.rs, src/channels/discord/mod.rs, src/channels/slack/mod.rs, src/channels/whatsapp/mod.rs, all handler files
• /models interactive model switching on channels — Platform-native buttons (Telegram InlineKeyboardMarkup, Discord ActionRow, Slack Block Kit) with model: callback handlers
  • src/channels/telegram/agent.rs, src/channels/discord/agent.rs, src/channels/slack/handler.rs
• Agent slash_command tool returns real data — /models, /usage, /help, /doctor, /sessions now execute and return actual context instead of "TUI-only" errors, enabling the agent to read config, check health, and switch models via config_manager
  • src/brain/tools/slash_command.rs, src/brain/tools/trait.rs, src/brain/agent/service/tool_loop.rs
• service_context on ToolExecutionContext — Tools can now access ServiceContext for DB queries (used by /usage and /sessions)

Fixed

• Image API key stored under wrong path — Onboarding wrote to flat [image] section in keys.toml instead of [providers.image.gemini], inconsistent with all other provider keys. Added ImageProviders struct, merge logic, and legacy fallback
  • src/config/types.rs, src/tui/onboarding/config.rs
• Channel commands section in README — Documented /help, /usage, /models, /stop for all channels including WhatsApp
• keys.toml parse errors now surface visibly — Invalid TOML (e.g. unquoted emails) previously caused silent key merge failure, breaking provider startup with no error. Now prints warning to stderr and logs error. /doctor validates keys.toml syntax
  • src/config/types.rs, src/brain/provider/factory.rs, src/brain/tools/slash_command.rs

Improved

• Telegram tool calls as individual messages — Each tool call now gets its own message (context + result) instead of all tools stacked in the response. Response streams cleanly at the bottom
  • src/channels/telegram/handler.rs
• Intermediate agent texts visible on Telegram — Agent commentary between tool rounds (e.g. "Found one! Let me reply to this:") now appears as individual messages, matching TUI behavior
  • src/channels/telegram/handler.rs

0.2.48 - 2026-03-04

Added

• Telegram thinking/reasoning stream — Live 💭 reasoning content streams during inference, vanishes on tool calls and response chunks, keeping the conversation clean
  • src/channels/telegram/handler.rs
• quick_jump mode for /onboard:<step> deep-links — Any /onboard:step (except ModeSelect) opens locked to that single step: no progress dots, centered title, Enter confirms, Esc exits to chat. Step-change detection reverts navigation attempts
  • src/tui/onboarding/wizard.rs, src/tui/onboarding/input.rs, src/tui/app/messaging.rs, src/tui/onboarding_render.rs
• Deferred health re-check on Enter — In /doctor quick_jump mode, Enter resets all checks to Pending (visible flash), tick resolves them next frame. Reloads config from disk so external changes are picked up
  • src/tui/onboarding/config.rs, src/tui/onboarding/fetch.rs, src/tui/app/state.rs
• YOLO (permanent) approval button on all channels — Telegram, Discord, Slack, and WhatsApp now offer a 🔥 YOLO button alongside Always (session), persisting auto-always to config.toml so approval survives restarts
  • src/channels/telegram/handler.rs, src/channels/telegram/agent.rs, src/channels/discord/handler.rs, src/channels/discord/agent.rs, src/channels/slack/handler.rs, src/channels/whatsapp/handler.rs, src/channels/whatsapp/mod.rs, src/utils/approval.rs, src/utils/mod.rs

Fixed

• Redundant check_approval_policy() in tool loop — Removed config-level short-circuit that was bypassing per-tool approval logic, fixing 3 approval policy test failures on CI
  • src/brain/agent/service/tool_loop.rs
• CI and Release workflows running redundantly on tag push — Added tags-ignore: v* to CI, added test gate (needs: test) to Release workflow
  • .github/workflows/ci.yml, .github/workflows/release.yml
• /doctor standalone mode (closes #21) — No onboarding chrome, Enter/Esc exit, removed redundant /onboard:health command
  • src/tui/onboarding_render.rs, src/tui/app/state.rs, src/tui/render/help.rs
• Trello API Token not loaded in from_config() — Health check falsely reported "No API Token provided" even when configured
  • src/tui/onboarding/wizard.rs
• Model selector filter not working (closes #20) — Filter text was typed but never applied to the displayed model list
  • src/tui/render/dialogs.rs
• UTF-8 crash on multi-byte text in all channels — 12 unsafe byte-index string slices replaced with truncate_str(), fixing panic on accented/emoji characters (e.g. Portuguese õ)
  • src/channels/telegram/handler.rs, src/channels/discord/handler.rs, src/channels/slack/handler.rs, src/channels/whatsapp/handler.rs
• quick_jump blocking all step navigation — Guard was catching internal step changes (field switching, channel sub-steps), not just step completion. Moved guard into next_step() so only step completion exits in deep-link mode
  • src/tui/onboarding/input.rs, src/tui/onboarding/navigation.rs, src/tui/onboarding/wizard.rs
• Approval policy not persisting from channels — Channels only offered "Always (session)" which wrote auto-session, downgrading the default YOLO policy. Now properly offers both session and permanent options
• Updated README and commands.toml.example with all /onboard:* sub-commands, /doctor, /whisper
  • README.md, commands.toml.example

0.2.47 - 2026-03-03

Changed

• Centralized tool approval into shared utils::approval module — Replaced per-channel auto_approve_session: Mutex<bool> fields in Discord, Slack, Telegram, and WhatsApp with a single config-driven source of truth. Two new functions (check_approval_policy, persist_auto_session_policy) read/write config.toml directly, and the core tool_loop.rs checks policy first before delegating to any channel callback. Approval callbacks moved from mod.rs to handler.rs as free functions across all channels
  • src/utils/approval.rs (new), src/utils/mod.rs, src/brain/agent/service/tool_loop.rs, src/channels/discord/mod.rs, src/channels/discord/handler.rs, src/channels/slack/mod.rs, src/channels/slack/handler.rs, src/channels/telegram/mod.rs, src/channels/telegram/handler.rs, src/channels/whatsapp/mod.rs, src/channels/whatsapp/handler.rs, src/channels/trello/handler.rs

Fixed

• Telegram streaming message stuck at top between tool calls — Streaming now uses separate tools and response fields with a recreate flag that deletes the old message and creates a fresh one below the approval buttons after each tool completion, so the conversation flows naturally downward instead of getting stuck above approval messages. Thanks @opryshok for reporting #17 and #16 — your bug reports directly drove this fix and the v0.2.46 improvements
  • src/channels/telegram/handler.rs, src/channels/telegram/agent.rs
• Race condition in approval registration across all channels — Pending approval is now registered BEFORE sending the approval message (not after), preventing a window where the user could click before the handler was ready
  • src/channels/discord/handler.rs, src/channels/slack/handler.rs, src/channels/telegram/handler.rs
• TUI "Always" approval choice not persisting — Clicking "AllowAlways" in the TUI now writes approval_policy = "auto-session" to config.toml so the choice survives restarts and is respected by all channels
  • src/tui/app/input.rs, src/tui/app/messaging.rs, src/tui/app/state.rs

Added

• Tracing/logging across all channel approval flows — Every approval request, response, and edge case now logs via tracing::info! / tracing::warn! for easier debugging
  • src/channels/discord/agent.rs, src/channels/discord/handler.rs, src/channels/slack/handler.rs, src/channels/telegram/agent.rs, src/channels/telegram/handler.rs, src/channels/whatsapp/handler.rs
• Cross-channel approval awareness — TUI reads approval policy from config.toml on session create/load, so a policy set via Telegram or any other channel is picked up everywhere
  • src/tui/app/messaging.rs, src/tui/app/state.rs

0.2.46 - 2026-03-03

Fixed

• Telegram tool approval stuck after clicking Yes/Always (3716bf9) — Three root causes fixed: (1) ApprovalCallback now returns (bool, bool) where the second bool propagates "Always" back into tool_context.auto_approve, so it persists across the entire tool loop instead of resetting after a few steps. (2) Race condition: pending approval is registered BEFORE sending the message, not after. (3) Tool input truncated to 3500 chars to prevent silent Telegram API rejection on long inputs. Closes #17
  • src/brain/agent/service/types.rs, src/brain/agent/service/tool_loop.rs, src/channels/telegram/mod.rs, src/channels/discord/mod.rs, src/channels/slack/mod.rs, src/channels/whatsapp/handler.rs, src/tui/app/state.rs, src/cli/ui.rs
• Telegram missing tool call context and formatting (3716bf9) — ToolCompleted events were being dropped by the progress callback; tool indicators used unsupported markdown. Now shows tool start/completion with proper labels. Improved markdown_to_telegram_html with headers, links, lists, underscore italic, and strikethrough. Closes #16
  • src/channels/telegram/handler.rs
• Multiline Up/Down arrow keys never navigated lines (e27f59b) — The multiline branch consumed all Up/Down events when the buffer contained newlines, even at boundaries (cursor at position 0 or end), blocking fall-through to history navigation. Now yields at boundaries: Up at position 0 falls through to history, Down at end of buffer does nothing
  • src/tui/app/input.rs
• Light mode unreadable — user messages and UI text invisible on light terminals (009e8e3) — Removed hardcoded dark user message background Rgb(30,30,38). Replaced Color::White (invisible on light backgrounds) with Color::Reset (terminal's default foreground) across all render files. Diff backgrounds changed from dark RGB to ANSI named colors (Green/Red/DarkGray) that adapt to both themes
  • src/tui/render/chat.rs, src/tui/render/input.rs, src/tui/render/help.rs, src/tui/render/dialogs.rs, src/tui/render/sessions.rs, src/tui/render/tools.rs
• Streaming token count duplicated ctx counter in input bar (2f8bc09) — The per-chunk tiktoken counter was adding ctx + output together and feeding it back into last_input_tokens, making the tool group display show the same "28K ctx" already in the input bar, plus a duplicate timer. Now output tokens are tracked separately and displayed as a per-response count next to the timer: (7s · 42 tok). The duplicate ctx+timer block below tool groups is removed
  • src/tui/events.rs, src/tui/app/state.rs, src/tui/app/messaging.rs, src/tui/render/chat.rs, src/cli/ui.rs

Changed

• Removed auto-backup logic (e142698) — Git handles versioning; the custom backup mechanism was redundant

0.2.45 - 2026-03-03

Added

• Real-time token count during streaming (65a0278) — The context usage display in the input box now increments live as the model responds: each streaming chunk is counted via tiktoken (cl100k_base) and fires a TokenCountUpdated event, so the counter ticks up token by token (e.g. 45K → 45.1K → 45.3K) instead of jumping at the end of each API round-trip. The API-reported real count resets the baseline after each response, keeping the display accurate across multi-tool loops
  • src/cli/ui.rs
• Elapsed time + ctx in thinking indicator (65a0278) — The "OpenCrabs is thinking..." spinner now shows elapsed seconds and current context size: ⠙ OpenCrabs is thinking... 3s · 45K ctx
  • src/tui/render/mod.rs
• Running token count below active tool groups (65a0278) — While tool calls execute, a subtle 45K ctx · 3s line is rendered below the live tool group so you can see context growth during multi-tool sequences
  • src/tui/render/chat.rs
• opencrabs daemon command (be61993) — New headless subcommand: same full channel setup (Telegram, Discord, Slack, WhatsApp) as the TUI, but no terminal UI. Blocks on Ctrl-C. Designed for use by the systemd/LaunchAgent service installed during onboarding. Fixes the daemon not working after opencrabs init (issue #12)
  • src/cli/mod.rs, src/cli/ui.rs
• 28 CLI parsing unit tests (be61993) — Full test coverage for all CLI subcommands including the new daemon command. Wired into lib.rs under #[cfg(test)]
  • src/tests/cli_test.rs, src/tests/mod.rs, src/lib.rs
• Hot-reload for all three config files (1675fd2) — config_watcher now watches config.toml, keys.toml, and commands.toml. Changing any of them is picked up within ~300ms without restart. Provider is swapped live when keys change (via AgentService::swap_provider). TUI refreshes approval policy and slash commands on reload
  • src/utils/config_watcher.rs, src/cli/ui.rs, src/tui/app/state.rs
• config.toml and commands.toml annotated examples (4fdc1a6) — Full annotated config.toml example added to the README Configuration section. New commands.toml section with complete syntax and action types reference. New commands.toml.example file in the project root matching the style of keys.toml.example. Two new Table of Contents entries added
  • README.md, commands.toml.example (new)

Fixed

• Daemon service not starting after install (be61993) — systemd ExecStart was missing the daemon subcommand arg and systemctl --user start was never called after enable. macOS LaunchAgent plist was also missing the daemon arg in ProgramArguments. Both fixed. Closes #12
  • src/tui/onboarding/config.rs
• config_watcher test hanging the test runner (be61993) — Blocking rx.recv() loop inside spawn_blocking kept the tokio runtime from shutting down after tests. Fixed with a 200ms-poll loop and hard 3s deadline so the blocking thread exits cleanly
  • src/utils/config_watcher.rs
• Nightly rustfmt CI failures (3208ac7) — telegram/mod.rs and whatsapp/handler.rs had formatting differences between local stable rustfmt and the nightly toolchain used by CI. Fixed by running cargo fmt through the pinned nightly toolchain from rust-toolchain.toml
  • src/channels/telegram/mod.rs, src/channels/whatsapp/handler.rs
• Redundant .max(0) on usize after saturating_sub (00fc64d) — Clippy unnecessary_min_or_max lint: usize::saturating_sub(1) already clamps at 0, .max(0) was always a no-op. Removed from three fields in onboarding channels
  • src/tui/onboarding/channels.rs
• llama-cpp-2 Metal segfault on macOS 26 arm64 (118ea65) — Bumped llama-cpp-2 from 0.1.134 to 0.1.137 which includes the upstream Metal fix. Thanks @Pibomeister (PR #13)
  • Cargo.toml, Cargo.lock

Changed

• Default approval policy changed to auto-always for new users (3ed02ef) — New installations no longer prompt before every tool call. The agent works autonomously out of the box. Existing users with approval_policy set in config.toml are unaffected (serde default only applies when the field is absent). To opt back into per-call prompts: run /approve → "Approve-only (always ask)"
  • src/config/types.rs, README.md
• Telegram allowlist hot-reload extended to Discord and Slack (2b9b8c6, bd95b52) — allowed_users lists for all three text channels now update at runtime when config.toml changes, without restart. Builds on the allowlist hot-reload foundation contributed by @Pibomeister (PR #14)
  • src/channels/telegram/mod.rs, src/channels/discord/handler.rs, src/channels/slack/handler.rs, src/utils/config_watcher.rs

0.2.44 - 2026-03-02

Added

• Google Gemini provider (e715536) — Full Provider trait implementation against the Gemini REST API (generativelanguage.googleapis.com/v1beta). Streaming via SSE, tool use with functionDeclarations/functionCall/functionResponse, vision (multimodal inlineData), 1M–2M token context window. Live model list fetched from the Gemini API during onboarding and /models. Auth via ?key= query param
  • src/brain/provider/gemini.rs (new), src/brain/provider/factory.rs, src/brain/provider/mod.rs
• Image generation & vision tools (e715536) — Two new agent tools powered by gemini-3.1-flash-image-preview ("Nano Banana"), independent of the main chat provider:
  • generate_image — Generate an image from a text prompt; saves PNG to ~/.opencrabs/images/; returns file path for channel delivery
  • analyze_image — Analyze an image file path or URL via Gemini vision; works even when the main model doesn't support vision
  • src/brain/tools/generate_image.rs (new), src/brain/tools/analyze_image.rs (new), src/brain/tools/mod.rs
• ImageSetup onboarding step (e715536, 1336b89, f534b24) — Step 7 in Advanced mode (after VoiceSetup, before Daemon). Toggle Vision Analysis and Image Generation independently; API key input with mask/replace mode; existing key detection. Model labeled as gemini-3.1-flash-image-preview (🍌 Nano Banana). Persistent "get a free key at aistudio.google.com" hint shown when no key is set. Navigation: Space/↑↓ to toggle, Tab/Enter to continue, BackTab/Esc to go back
  • src/tui/onboarding/types.rs, src/tui/onboarding/wizard.rs, src/tui/onboarding/navigation.rs, src/tui/onboarding/fetch.rs, src/tui/onboarding/config.rs, src/tui/onboarding_render.rs
• /onboard:image deep-link (e715536) — Jump directly to the ImageSetup step from chat at any time
  • src/tui/app/messaging.rs
• On-demand brain file loading (3224048) — build_core_brain() replaces build_system_brain() at startup — injects only SOUL.md + IDENTITY.md (~1-2k tokens). All other brain files listed in a memory index; loaded by the agent via load_brain_file(name) tool on demand. name="all" loads everything. Dramatically reduces baseline token overhead for every message
  • src/brain/prompt_builder.rs, src/brain/tools/load_brain_file.rs (new), src/cli/ui.rs
• write_opencrabs_file tool (8f3d648) — Writes any file inside ~/.opencrabs/ (brain files, config, keys). Replaces the broken agent pattern of using edit_file/write_file which are locked to the working directory by validate_path_safety()
  • src/brain/tools/write_opencrabs_file.rs (new), src/brain/tools/mod.rs
• respond_to selector in Telegram/Discord/Slack onboarding (9ecc8f0) — New field in each channel's setup step; choose all / dm_only / mention mode during onboarding instead of editing config.toml manually
  • src/tui/onboarding/types.rs, src/tui/onboarding/fetch.rs, src/tui/onboarding_render.rs, src/tui/onboarding/config.rs
• Google Image API Key in health check (6923174) — When image features are enabled, the health check step verifies the Google AI key is present
  • src/tui/onboarding/config.rs
• send_file action — discord_send and slack_send (905e9ef) — New action uploads a local file as a native attachment. Discord: file attachment in channel. Slack: file upload via API. Both tools now at 17 actions
  • src/brain/tools/discord_send.rs, src/brain/tools/slack_send.rs
• add_attachment action — trello_send (ac44fc3) — New action uploads a local image or file as a Trello card attachment via multipart upload; returns the hosted Trello URL. Tool now at 22 actions
  • src/brain/tools/trello_send.rs, src/channels/trello/client.rs
• Full file/image/audio input pipeline across all channels (9aed2ea, 5bc33f5) — Unified classify_file(bytes, mime, filename) → FileContent utility routes incoming files across every channel: images → vision pipeline (<<IMG:path>>), text/code/data files → extracted inline (up to 8 000 chars), audio → STT, PDFs → note to paste or use analyze_image. Trello: card attachments are fetched and processed on every incoming comment. Slack: voice/STT support added (was missing). All channels now handle images, text files, documents, and audio with consistent behavior
  • src/utils/file_extract.rs (new), src/utils/mod.rs, src/channels/telegram/handler.rs, src/channels/discord/handler.rs, src/channels/whatsapp/handler.rs, src/channels/slack/handler.rs, src/channels/slack/agent.rs, src/brain/tools/slack_connect.rs, src/channels/trello/client.rs, src/channels/trello/handler.rs, src/channels/trello/models.rs
• TUI text file input (3e4460e) — Paste or type any text file path in the TUI input field — the file is read and inlined automatically as [File: name]\n```\ncontent\n``` . Works at paste time and submit time. Supports .txt, .md, .json, .yaml, .toml, .rs, .py, .go, .sql, and 20+ other formats
  • src/tui/app/state.rs, src/tui/app/messaging.rs

Fixed

• Generated images delivered as native media across all channels (60584ff) — <<IMG:path>> markers in agent replies are now unwrapped and delivered natively on every channel: Telegram send_photo, WhatsApp image message, Discord file attachment, Slack file upload, Trello card attachment + ![filename](url) embed in comment. Previously the raw marker string was sent as plain text
  • src/channels/telegram/handler.rs, src/channels/discord/handler.rs, src/channels/whatsapp/handler.rs, src/channels/slack/handler.rs, src/channels/trello/handler.rs
• Trello outgoing images — upload attachment + embed inline (3e4460e) — Agent replies containing <<IMG:path>> on Trello are now uploaded as card attachments via add_attachment_to_card and embedded in the comment as ![filename](url). Previously the marker was silently dropped
  • src/channels/trello/handler.rs
• Channel tool approval + TUI real-time updates follow-up (248b719) — Follow-up fixes to tool approval flows and TUI live-update reliability across all remote channels after the v0.2.43 multi-channel expansion
  • src/channels/*/mod.rs, src/brain/agent/service/tool_loop.rs, src/tui/app/state.rs
• Clippy: collapse nested if blocks (2595550) — Fixed two collapsible_if lint errors in messaging.rs (TUI text file detection) and whatsapp/handler.rs (document attachment handling)
  • src/tui/app/messaging.rs, src/channels/whatsapp/handler.rs
• TUI silent message queue after errors (dc815ce) — After any agent error, processing_sessions was never cleared for the current session, causing all subsequent send_message calls to be silently queued with no agent running. Fixed by unconditionally removing the session from processing_sessions and session_cancel_tokens in the TuiEvent::Error handler before branching on current vs background session
  • src/tui/app/state.rs
• TUI real-time updates during channel tool loops (b44f1ff) — Remote channel tool loops (Telegram, WhatsApp, etc.) were not firing session_updated_tx on each chunk, causing the TUI to only refresh at the end of a long tool sequence. Now fires after every tool call completion
  • src/brain/agent/service/tool_loop.rs
• Attachment input shows "Image #N" instead of full path (2609583) — Attachment display in the TUI input bar was showing the full file path; now shows Image #N, Document #N placeholders matching the <<IMG:...>> / <<DOC:...>> injection format
  • src/tui/render/input.rs
• WhatsApp TTS — upload media before sending audio message (135b4d6) — TTS audio was being sent via send_audio before uploading to WhatsApp media servers, causing delivery failures. Now uploads first, then sends with the returned media ID
  • src/channels/whatsapp/handler.rs
• WhatsApp handler regression — empty allowed_phones + connect tool (5a32d49) — Empty allowed_phones in config was incorrectly blocking all messages including the owner. whatsapp_connect tool now correctly writes the config entry. Owner bypass re-validated
  • src/channels/whatsapp/handler.rs, src/brain/tools/whatsapp_connect.rs
• WhatsApp security — block owner→contact processing (0fc8b2e) — Messages sent by the owner to a contact were being processed as if the contact sent them, exposing the agent to arbitrary tool execution from outgoing messages
  • src/channels/whatsapp/handler.rs
• WhatsApp outgoing allowed_users enforcement (1707e0f) — Outgoing messages to contacts not in allowed_users were being processed; now strictly gated
  • src/channels/whatsapp/handler.rs
• Context display reset immediately after compaction (2c4ca8e) — After /compact, the context percentage in the TUI header was not resetting to the new value until the next message; now resets immediately
  • src/tui/app/state.rs

Changed

• Per-channel config structs (f28e229) — Replaced the single flat ChannelConfig with 8 dedicated structs (TelegramConfig, DiscordConfig, SlackConfig, WhatsAppConfig, TrelloConfig, etc.) for cleaner config parsing, better type safety, and simpler channel-specific fields. Trello board_ids replaces the previous allowed_channels field
  • src/config/types.rs, all channel modules, src/tui/onboarding/

0.2.43 - 2026-03-02

Added

• Telegram full control — 16 actions + live streaming + approval buttons (c1ba37c) — telegram_send tool expanded from send to 16 actions: send, reply, edit, delete, pin, unpin, forward, send_photo, send_document, send_location, send_poll, send_buttons, get_chat, ban_user, unban_user, set_reaction. LLM response streams live into a Telegram message with ▋ cursor (edits every 1.5 s). Session resilience: re-fetches bot from DB if lost across restarts. Idle session timeout per-user
  • src/brain/tools/telegram_send.rs, src/channels/telegram/handler.rs, src/channels/telegram/mod.rs
• Discord full control — 16 actions + session idle timeout (3459d0b) — discord_send tool expanded to 16 actions mirroring Telegram: send, reply, edit, delete, pin, unpin, forward, send_photo, send_document, send_location, send_poll, send_buttons, get_guild, kick_user, ban_user, set_reaction. Idle session timeout for per-user sessions
  • src/brain/tools/discord_send.rs, src/channels/discord/
• Slack full control — 16 actions + sender context injection (89c9e71) — slack_send tool expanded to 16 actions: send, reply, react, unreact, edit, delete, pin, unpin, get_messages, get_channel, list_channels, get_user, list_members, kick_user, set_topic, send_blocks. Non-owner messages now prepend sender identity [Slack message from {uid} in channel {ch}]
  • src/brain/tools/slack_send.rs, src/channels/slack/handler.rs
• WhatsApp typing indicator (9f3b1fa) — Sends composing chat state on message receipt, paused on completion so the user sees a native typing indicator while the agent processes
  • src/channels/whatsapp/handler.rs
• Tool approval — 3-button UI across all channels (f6b8523, 586cccd, 816147c) — All four remote channels now show ✅ Yes / 🔁 Always (session) / ❌ No approval prompts matching the TUI, powered by channel-native interactive elements (WhatsApp ButtonsMessage, Telegram inline keyboard, Discord CreateButton, Slack SlackBlockButtonElement). "Always" sets session-level auto_approve_session flag — no further prompts for that session
  • src/channels/whatsapp/mod.rs, src/channels/telegram/mod.rs, src/channels/discord/mod.rs, src/channels/slack/mod.rs
• Tool input context in Telegram streaming indicator (3da472a, af4b96b) — Streaming status line now shows a brief hint of what the tool is doing (e.g. ⚙ bash: git status) so the user has context while waiting
  • src/channels/telegram/handler.rs
• TUI auto-refresh when remote channels process messages (7b95209) — After every run_tool_loop completion, AgentService fires a session_updated_tx notification. The TUI listens, calling load_session if the updated session is the current one (and not already being processed by the TUI), or marking it as unread otherwise. Real-time TUI updates when Telegram/WhatsApp/Discord/Slack messages are processed — no manual session switch required
  • src/brain/agent/service/builder.rs, src/brain/agent/service/tool_loop.rs, src/tui/events.rs, src/tui/app/state.rs, src/cli/ui.rs

Fixed

• SQLite WAL mode + larger pool (1ec5c3b) — Enables write-ahead logging so concurrent reads (TUI) and writes (channel agents) don't block each other; pool size increased from 5 to 20 connections. Eliminates channel concurrency timeouts
  • src/services/ (DB setup)
• WhatsApp sender identity (00cc01b) — Strips device suffix from JID (:N@s.whatsapp.net → @s.whatsapp.net) before phone-number comparison; injects [WhatsApp message from {name} ({phone})] for non-owner messages; fetches contact display name when available
  • src/channels/whatsapp/handler.rs
• WhatsApp reply to chat JID instead of device JID (24c1e5d) — Was replying to the device-scoped JID (:0@s.whatsapp.net) causing delivery failures in group chats and multi-device setups; now replies to the canonical chat JID
  • src/channels/whatsapp/handler.rs
• Inject sender context for non-owner Discord and Telegram messages (e00374a) — Non-owner messages now prepend [Discord/Telegram message from {name} (ID {uid}) in channel {ch}] so the agent knows who it's talking to instead of assuming the owner
  • src/channels/discord/handler.rs, src/channels/telegram/handler.rs
• Secret sanitization — redact API keys from all display surfaces (436808e, d3a2380) — New utils::redact_tool_input() function recursively walks tool input JSON, redacting values for sensitive keys (authorization, api_key, token, secret, password, bearer, etc.) and inline bash command patterns (Bearer xxx, api_key=xxx, URL passwords). Applied to TUI tool history, TUI approval dialogs, and all four remote channel approval messages
  • src/utils/sanitize.rs (new), src/tui/render/tools.rs, src/channels/*/mod.rs
• WhatsApp upstream log noise suppressed (f6b8523) — Added whatsapp_rust::client=error and whatsapp_rust=warn directives to filter upstream TODO stub log lines
  • src/logging/logger.rs

Changed

• Context budget enforcement refactored (d8ab8f0) — Extracted repeated 80%/90% compaction logic into enforce_context_budget() helper on AgentService. 80 %: triggers LLM compaction. 90 %: hard-truncates to 80 % first, then compacts. Up to 3 retries on LLM compaction failure, then warns user to run /compact
  • src/brain/agent/service/tool_loop.rs
• send_message_with_tools_and_callback — Per-call approval and progress callback overrides; remote channels pass their own callbacks without touching service-level defaults
  • src/brain/agent/service/messaging.rs, src/brain/agent/service/tool_loop.rs

0.2.42 - 2026-03-01

Added

• Native Trello channel (80c7b05) — TrelloAgent authenticates and makes credentials available for tool use. Default mode is tool-only — the AI acts on Trello only when explicitly asked via trello_send. Opt-in polling available via poll_interval_secs in config; when enabled, only responds to explicit @bot_username mentions from allowed users. Board names resolved automatically — mix human-readable names and 24-char IDs freely
  • src/channels/trello/ (agent, client, handler, models, mod)
• trello_connect tool (80c7b05) — Verify credentials, resolve boards by name, persist to config, spawn agent, confirm with open card count. Accepts comma-separated board names or IDs
  • src/brain/tools/trello_connect.rs
• trello_send tool — 21 actions (80c7b05) — Full Trello control without exposing credentials in URLs: add_comment, create_card, move_card, find_cards, list_boards, get_card, get_card_comments, update_card, archive_card, add_member_to_card, remove_member_from_card, add_label_to_card, remove_label_from_card, add_checklist, add_checklist_item, complete_checklist_item, list_lists, get_board_members, search, get_notifications, mark_notifications_read
  • src/brain/tools/trello_send.rs
• /onboard:<step> deep-links (e4975e4) — Jump directly to any onboarding step: /onboard:provider, /onboard:channels, /onboard:voice, /onboard:health, etc. /doctor alias for /onboard:health
  • src/tui/app/messaging.rs, src/tui/app/state.rs, src/tui/render/help.rs

Fixed

• WhatsApp voice notes silently dropped (8e29655) — Handler was skipping all non-text messages including voice notes (ptt). Now only skips if no text AND no audio AND no image
  • src/channels/whatsapp/handler.rs
• STT key missing from channel factory (8e29655, d0a7651) — ChannelFactory was built with config.voice which has stt_provider=None. All channel agents (WhatsApp, Discord, dynamic Telegram) now receive the fully resolved VoiceConfig with stt_provider/tts_provider populated
  • src/cli/ui.rs
• Channel allowed_users unified (e4975e4) — Removed allowed_ids from ChannelConfig, unified into allowed_users: Vec<String> with backward-compat deserializer accepting legacy TOML integer arrays. Fixed health check false failures: Discord and Slack channel IDs were read from wrong field
  • src/config/types.rs, channel agents
• Channel config not passed to agents (406503b) — telegram_connect, discord_connect, slack_connect now pass respond_to and allowed_channels from persisted config to agent constructors (previously hardcoded to defaults)
  • src/brain/tools/telegram_connect.rs, discord_connect.rs, slack_connect.rs
• Tool expand (Ctrl+O) shows full params (0aba196) — Expanded tool view now shows complete untruncated input params line by line. In-flight calls show a "running..." spinner. DB-reconstructed entries degrade gracefully
  • src/tui/render/tools.rs, src/tui/app/state.rs
• Error/warning messages auto-dismiss after 2.5 s (4408a69, d0a7651) — Timer resets correctly on user action; covers all clear-sites in input, messaging, and dialogs
  • src/tui/app/dialogs.rs, input.rs, messaging.rs
• Thinking indicator sticky above input (406503b) — Moved out of scrollable chat into a dedicated layout chunk — never scrolls away
  • src/tui/render/mod.rs, chat.rs
• /onboard resets to first screen (d0a7651) — Pre-loads existing config values while resetting to ModeSelect so health check shows correct state
  • src/tui/app/messaging.rs, src/tui/onboarding/wizard.rs
• CI Windows build (001ed00) — Replaced removed aws-bedrock/openai features with telegram,discord,slack in Windows CI workflow
  • .github/workflows/ci.yml
• Trello agent tool-only by default (7ca6b6b) — Removed automatic polling and auto-replies. Agent starts in tool-only mode (credentials stored, no polling). poll_interval_secs in [channels.trello] config opts in to polling; even then only @mentions from allowed users trigger a response. Adds poll_interval_secs: Option<u64> to ChannelConfig
  • src/channels/trello/agent.rs, src/config/types.rs, src/cli/ui.rs, src/brain/tools/trello_connect.rs, config.toml.example

0.2.41 - 2026-03-01

Fixed

• WhatsApp onboarding — always test connection on Enter (676ab29) — Pressing Enter on the phone field now always triggers a test message, matching Telegram/Discord/Slack behavior. Previously the test was gated on whatsapp_connected, so re-opening the app with an existing session silently skipped the test and just advanced
  • src/tui/onboarding/channels.rs
• WhatsApp onboarding — reconnect from existing session for test (676ab29) — When no live client is in memory (app reopened after prior pairing), test_whatsapp_connection now calls reconnect_whatsapp() which reuses the stored session.db without wiping it — no new QR scan required
  • src/brain/tools/whatsapp_connect.rs, src/tui/app/dialogs.rs
• WhatsApp test message includes brand header (676ab29) — Test message now prepends 🦀 *OpenCrabs*\n\n (the MSG_HEADER constant) so it reads consistently with all other WhatsApp messages sent by the agent
  • src/tui/app/dialogs.rs
• WhatsApp onboarding — post-QR UX overhaul (676ab29) — After scanning the QR code the popup dismisses, the wizard advances to the phone allowlist field, shows any previously configured number (sentinel pattern), and allows confirm-or-replace before testing. Navigation keys (Tab/BackTab/S) always work regardless of test state; only Enter is blocked while a test is in-flight
  • src/tui/onboarding/channels.rs, src/tui/onboarding_render.rs, src/tui/app/dialogs.rs, src/tui/app/state.rs, src/brain/tools/whatsapp_connect.rs
• Clippy collapsible_match errors (ff66828) — Collapsed nested if-in-match arms into match guards across input.rs (WhatsApp paste handler) and markdown.rs (Tag::BlockQuote, TagEnd::Heading, TagEnd::Item, Event::HardBreak|SoftBreak)
  • src/tui/onboarding/input.rs, src/tui/markdown.rs
• CI nightly clippy/rustfmt (a65c0ab) — Added rustfmt and clippy components to rust-toolchain.toml so nightly CI jobs resolve the tools without network fallback; pinned workflow to main branch trigger
  • rust-toolchain.toml, .github/workflows/

0.2.40 - 2026-02-28

Added

• Live plan checklist widget (7e1b4db) — A real-time task panel appears above the input box whenever the agent is executing a plan. Shows plan title, progress bar (N/M ████░░ X%), and per-task status rows (✓ completed, ▶ in-progress, · pending, ✗ failed) with per-status colors. Height is min(task_count + 2, 8) rows; zero height when no plan is active. Panel is session-isolated — each session tracks its own plan file (~/.opencrabs/agents/session/.opencrabs_plan_{uuid}.json) and reloads on session switch
  • src/tui/render/plan_widget.rs (new), src/tui/render/mod.rs, src/tui/app/state.rs, src/tui/app/messaging.rs

Fixed

• Live ctx counter during agent tool loops (1cb46a9) — TokenCountUpdated events now sync last_input_tokens so the ctx: N/M display in the status bar ticks up live during streaming and tool execution instead of freezing until ResponseComplete
  • src/tui/app/state.rs
• Ctx shows base context on session load and new session (1cb46a9) — Status bar no longer starts at – or 0 on a fresh session. It immediately reflects system prompt + tool definition token cost via base_context_tokens() (system prompt tokens + tool count × 60)
  • src/brain/agent/service/builder.rs, src/tui/app/messaging.rs
• Plan tool auto-approves on finalize (9fca3ec) — finalize now sets PlanStatus::Approved directly and instructs the agent to begin executing tasks immediately. Previously the tool returned PendingApproval and printed "STOP — wait for user response", causing a double-approval (tool dialog + follow-up message) and blocking task execution
  • src/brain/tools/plan_tool.rs, src/brain/prompt_builder.rs
• read_only_mode dead code removed (9fca3ec) — Remnant field and all callers from the deleted Plan Mode feature purged from ToolExecutionContext, tool implementations, send_message_with_tools_and_mode, A2A handlers, and tests
  • src/tui/app/messaging.rs, src/tui/app/state.rs, src/brain/tools/bash.rs, src/brain/tools/edit_file.rs, src/brain/tools/write_file.rs, src/brain/tools/code_exec.rs, src/brain/tools/notebook.rs
• MiniMax </think> block stripping (9b0b8d0) — MiniMax sometimes closes reasoning blocks with </think> instead of <!-- /reasoning -->. Extended the think-tag filter to handle this closing variant
  • src/brain/provider/custom_openai_compatible.rs

Changed

• Complete TUI color overhaul — gray, orange, and cyan palette (2796889, 3d88f11, a33fddc) — All three legacy accent colors replaced for a cohesive warm-neutral scheme:
  • Color::Blue / Rgb(70,130,180) → Color::Gray / Rgb(120,120,120) — borders, titles, section headers
  • Color::Yellow / Rgb(184,134,11) → Color::Rgb(215,100,20) muted orange — active/pending states, ctx warning, approval badge
  • Color::Green / green-dominant Rgb values → Color::Cyan / Rgb(60–80,165–190,165–190) — success states, completed tasks, diff additions, ctx-ok indicator
  • src/tui/render/chat.rs, src/tui/render/dialogs.rs, src/tui/render/help.rs, src/tui/render/input.rs, src/tui/render/plan_widget.rs, src/tui/render/sessions.rs, src/tui/render/tools.rs, src/tui/onboarding_render.rs

0.2.39 - 2026-02-28

Added

• Status bar below input (02220e7, 9dd4cab) — Persistent one-line status bar replaces the old sticky overlay. Displays session name (orange), provider / model, working directory, and approval policy badge. Session and directory were moved from the header into the status bar; the full-width header bar was removed entirely
  • src/tui/render/mod.rs, src/tui/render/input.rs
• Immediate thinking spinner in chat (57ffc40) — A spinner and "OpenCrabs is thinking..." line appears in the chat area as soon as a request is submitted, before any streaming content arrives. Eliminates the blank gap while the provider is warming up
  • src/tui/render/chat.rs
• Per-session context token cache (57ffc40) — When switching between sessions or reloading, the last known input token count is restored from an in-memory cache instead of showing –. Accurate token counts are re-confirmed on the next API response
  • src/tui/app/state.rs, src/tui/app/messaging.rs

Fixed

• ctx shows accurate token count for providers that report zero usage (033043f) — Providers like MiniMax always return usage: {total_tokens: 0} in streaming responses. The provider now uses its pre-computed message_tokens + tool_schema_tokens (serialised OpenAI JSON) as the fallback, so the ctx display (e.g. 29K/200K) matches the debug log exactly instead of showing the lower raw-text estimate (~14K)
  • src/brain/provider/custom_openai_compatible.rs, src/brain/agent/service/tool_loop.rs
• Compact app title in sessions/help screens (bc80a0f) — Removed blank lines and border from the app title block in non-chat screens. Title now occupies exactly one row, reclaiming vertical space
  • src/tui/render/mod.rs
• Extra blank space below chat history (d469f01) — Scroll calculation used reserved = 3 left over from removed borders/overlay. Changed to reserved = 1 (top padding only), eliminating the gap at the bottom of the chat area
  • src/tui/render/chat.rs
• Duplicate thinking indicators removed (aa08d68, 57ffc40) — "OpenCrabs is thinking" was appearing twice: once as an inline tool-group hint and once in the status bar. Removed both; the single spinner in the chat area is the sole indicator
  • src/tui/render/chat.rs, src/tui/render/input.rs
• Muted orange replaces bright yellow (02220e7) — Color::Yellow replaced with Color::Rgb(215, 100, 20) for ctx percentage, sessions spinner, and pending-approval badge. Intentional dark-golden Rgb(184, 134, 11) unchanged
  • src/tui/render/input.rs, src/tui/render/sessions.rs

0.2.38 - 2026-02-27

Fixed

• Splash screen shows actual custom provider name — resolve_provider_from_config() was returning the hardcoded string "Custom" instead of the actual provider name (e.g. "nvidia", "moonshot"). Now correctly returns the name key from providers.active_custom()
  • src/config/types.rs
• Full request payload in debug logs — Removed .take(1000) truncation from OpenAI-compatible request debug log. The API request itself was never truncated; only the log display was. Now logs the full payload for accurate debugging
  • src/brain/provider/custom_openai_compatible.rs
• Standalone reasoning render during thinking-only phase — Providers that emit reasoning before any response text (e.g. Kimi K2.5, DeepSeek) now render a visible 🦀 OpenCrabs is thinking... block with live reasoning content while streaming_response is still empty. Previously the screen was blank until the first response chunk
  • src/tui/render/chat.rs
• Streaming redraws per chunk — Drain loop in runner now breaks immediately on ResponseChunk events, triggering a redraw after each text chunk. Previously ReasoningChunk events also broke the loop, preventing response text from rendering in real-time on some providers
  • src/tui/runner.rs
• Approval dialog shows full tool parameters — Tool approval dialog previously truncated parameter values at 60 characters. Now renders all parameters line-by-line without truncation so the full context is visible when deciding whether to approve
  • src/tui/render/tools.rs
• Tool approval waits indefinitely — Removed 120-second timeout on tool approval callbacks. The dialog now waits as long as needed for the user to approve or deny
  • src/tui/app/state.rs
• Green dot pulse slowed — Animated ● dot in tool call groups now pulses on a ~1.6s cycle (animation_frame / 8) instead of the previous fast flicker (animation_frame / 3)
  • src/tui/render/tools.rs

Removed

• Plan Mode completely removed (~1400 lines deleted) — All plan execution code, UI, keyboard shortcuts, and state removed. Includes plan_exec.rs module, AppMode::Plan variant, PlanApprovalState/PlanApprovalData structs, Ctrl+P/Ctrl+A/Ctrl+R/Ctrl+I shortcuts, plan approval intercept in input handler, plan help screen section, and plan re-exports. Plan Mode section removed from README
  • src/tui/app/plan_exec.rs (deleted), src/tui/app/input.rs, src/tui/app/messaging.rs, src/tui/app/mod.rs, src/tui/app/state.rs, src/tui/events.rs, src/tui/mod.rs, src/tui/render/chat.rs, src/tui/render/help.rs, src/tui/render/mod.rs, src/tui/render/tools.rs, README.md

0.2.37 - 2026-02-26

Added

• Per-session provider selection (5689cd9) — Each session can now have its own LLM provider. Configure per-session via /models or in config.toml under [session.*.provider]. Parallel execution of multiple sessions with different providers supported
  • src/brain/agent/service.rs, src/brain/mod.rs, src/tui/app/state.rs, src/tui/render.rs, config.toml.example
• Arrow key navigation in multiline input (9b544f9) — Arrow Up/Down now navigate between lines in the multiline input field, not just recall history. Cursor moves within the multiline content as expected
  • src/tui/app/input.rs, src/tui/render/input.rs
• Test units for multi-session and multi-model (cf7ff0d) — Added unit tests covering session-aware approval policies, model switching within sessions, and provider key isolation
  • src/brain/agent/service/tests/approval_policies.rs, src/brain/agent/service/tests/basic.rs

Fixed

• Session-aware tool approvals (846f228) — Tool approval policies now correctly apply per-session. Approval state is stored with session ID, not globally. Async model fetching improved with better error handling
  • src/brain/agent/service.rs, src/brain/mod.rs, src/tui/app/state.rs
• Custom provider name field (c22a05a) — Onboarding now pre-fills the custom provider name field. Model fetching uses existing key if available instead of requiring re-entry. Provider name displays correctly in /models dialog
  • src/tui/onboarding.rs, src/tui/app/dialogs.rs, src/brain/provider/custom_openai_compatible.rs

Refactored

• Split agent/service.rs (8f9c160) — Extracted into module directory: service/builder.rs, service/context.rs, service/helpers.rs, service/messaging.rs, service/mod.rs. Improved code organization and testability
• Split render.rs (6247666) — Extracted 3312-line file into render/ module directory with render/mod.rs, render/input.rs, render/dialogs.rs, render/components.rs
• Cargo fmt pass (d02fcf7) — Full codebase formatting enforcement

0.2.36 - 2026-02-26

Fixed

• Custom provider /models dialog (fc0626c) — Model name is now a free-text input instead of a hardcoded list. Labels show the actual provider name (e.g. "Moonshot") instead of generic "Custom". Onboarding flow updated to match
  • src/tui/app/dialogs.rs, src/tui/render.rs, src/tui/onboarding.rs, src/config/types.rs, src/tui/app/state.rs, src/brain/provider/anthropic.rs, src/brain/provider/custom_openai_compatible.rs, README.md
• Input UX improvements (7804ab3) — Esc scrolls viewport to bottom of conversation. Arrow Up recalls previously cleared/stashed input text. Cursor renders as a block highlighting the current character instead of a thin line. Escape timer resets when processing completes so next Esc behaves correctly
  • src/tui/app/input.rs, src/tui/app/messaging.rs, src/tui/app/state.rs, src/tui/render.rs
• Strip Kimi HTML comment markup (47b1d58) — Kimi K2.5 embeds reasoning and hallucinated tool calls as HTML comments (<!-- reasoning -->, <!-- tools-v2: -->) in the content field. Extended filter_think_tags and strip_think_blocks to strip these alongside <think>. Fixed extract_reasoning to handle multiple reasoning blocks per message. Added Moonshot/Kimi pricing (K2.5, K2 Turbo, K2) to compiled-in defaults and usage_pricing.toml.example
  • src/brain/provider/custom_openai_compatible.rs, src/pricing.rs, src/tui/app/messaging.rs, usage_pricing.toml.example
• /models provider switch: never overwrite user API keys (5120bf5) — Killed sentinel string "__EXISTING_KEY__" from /models dialog entirely. Replaced with boolean flag model_selector_has_existing_key. Only writes to keys.toml when user actually types a new key. Disables all other providers on disk before enabling selected one. Added is_real_key guard in merge_provider_keys for all providers
  • src/config/types.rs, src/tui/app/dialogs.rs, src/tui/app/state.rs, src/tui/render.rs
• Model change context hint for agent (ce8e422) — When user switches model via /models, a [Model changed to X (provider: Y)] hint is prepended to the next user message via pending_context (same mechanism as /cd), so the LLM is aware of the switch. TUI status message also shown in chat. Custom provider uses user-configured name (e.g. "nvidia") instead of generic label. Fallback provider key changed from providers.custom.default to providers.custom to avoid stale config entries
  • src/tui/app/dialogs.rs, src/tui/app/messaging.rs

0.2.35 - 2026-02-26

Added

• Animated tool call dots — Green ● dot pulses (●/○) while tools are actively processing, stays solid when finished. Visually distinguishes active tool execution from completed groups
• Inline thinking indicator during tool execution — "OpenCrabs is thinking..." now renders inline above the active tool group instead of as a sticky overlay, preventing overlap with tool call content
• .github/CODEOWNERS — Auto-assigns @adolfousier as reviewer on all PRs

Fixed

• TUI spacing improvements — Removed double blank lines between messages and tool groups. Added proper spacing before thinking sections and between thinking hint and expanded content
• Inline code background removed — bg(Color::Black) on backtick code spans in markdown renderer removed for cleaner look. Thinking hints use subtle Rgb(90,90,90) text with no background
• Sudo prompt bleeding into TUI — Added -p "" flag to sudo -S to suppress sudo's native "Password:" prompt from writing directly to the terminal
• cargo fmt full codebase pass — Enforced official Rust style guide across 92 files
• Test fixes — stream_complete() tests updated to destructure (LLMResponse, Option<String>) tuple return with reasoning assertions. write_secret_key doctest fixed (missing import + Result return type)

0.2.34 - 2026-02-26

Added

• Reasoning/thinking persistence — MiniMax (and other providers that emit reasoning_content) now accumulate thinking content during streaming, persist it to DB with <!-- reasoning --> markers, and reconstruct it on session reload. Reasoning is rendered as a collapsible "Thinking" section on assistant messages
• Real-time message persistence per step — Assistant text is written to DB after each tool iteration, not just at the end. Crash or disconnect mid-task no longer loses intermediate text
• Collapsible reasoning UI — Ctrl+O now toggles both tool groups and reasoning sections. Collapsed by default, expandable inline with dimmed italic style matching the streaming "Thinking..." indicator

Fixed

• MiniMax intermediate text lost on reload — Tool call indices from OpenAI-compatible providers collided with the text content block at index 0 in stream_complete(), overwriting accumulated text. Tool indices now offset by +1. Fixes #10
• TUI unresponsive after onboarding — rebuild_agent_service() only attached the approval callback, dropping progress_callback, message_queue_callback, sudo_callback, and working_directory. All callbacks are now preserved from the existing agent service. Fixes #10
• Tool loop false positives eliminated — Replaced 115-line per-tool signature matching with 7-line universal input hash. Different arguments = different hash = no false detection. Same args repeated 8 times = real loop
• Chat history lost on mid-task exit — Exiting while the agent was between tool iterations discarded the conversation. Now persists accumulated text before exit
• Clippy warnings — Collapsed nested if statements in service.rs and input.rs

0.2.33 - 2026-02-25

Added

• Streaming /rebuild — Live compiler output streamed to chat during build. On success, binary is exec()-replaced automatically (no prompt, no restart). Auto-clones repo for binary-only users if no source tree found
• Centralized usage_pricing.toml — Runtime-editable pricing table for all providers (Anthropic, OpenAI, MiniMax, Google, DeepSeek, Meta). Edit live, changes take effect on next /usage open without restart. Written automatically on first run during onboarding
• All-time /usage breakdown — Shows cost grouped by model across all sessions. Historical sessions with stored tokens but zero cost get estimated costs (yellow ~$X.XX prefix). Unknown models shown as $0.00 instead of silently ignored
• /cd context injection — When user changes working directory via /cd, a context hint is queued and prepended to the next message so the LLM knows about the directory change without the user having to explain. Uses new pending_context vec on App state
• Tool approval policy preservation across compaction — Compaction summary prompt now includes ## Tool Approval Policy section. All 4 continuation messages (pre-loop, mid-tool-loop, emergency, mid-loop) inject CRITICAL: Tool approval is REQUIRED when auto-approve is off. Agent can no longer "forget" approval policy after context resets
• Dropped stream detection + retry — Detects when provider streams end without [DONE]/MessageStop (stop_reason is None). Retries up to 2 times transparently, discarding partial responses. After 2 failures, proceeds gracefully with partial response

Fixed

• Context compaction streamed, not frozen — compact_context uses stream_complete so the TUI event loop stays alive during compaction. Previously froze the UI for 2-5 minutes on large contexts
• Compaction summary visible in chat — Summary fires via CompactionSummary progress event after streaming, rendered in chat so user can see what was preserved
• TUI state reset post-compaction — Resets streaming_response + active_tool_group on compaction so the UI is clean for continuation
• Compaction request budget cap — Capped at 75% of context window with 16k token overhead (was 8k). Prevents the compaction request itself from exceeding the provider limit (was sending 359k tokens)
• Real-time context counter — Live token count updates in header during streaming
• /models paste support — API keys can be pasted into the model selection dialog
• Pricing: $0 cost for all sessions — PricingConfig struct used HashMap<String, Vec<PricingEntry>> but TOML has entries = [...] wrapper. Added ProviderBlock to match schema correctly
• Pricing: MiniMax $0 — Stream chunks don't include model name. Falls back to request model
• Pricing: legacy format migration — Auto-migrates [[usage.pricing.X]] on-disk format to current schema
• Clippy: collapsible_if — Fixed in rebuild.rs and pricing.rs

0.2.32 - 2026-02-24

Added

• A2A Bearer token authentication -- JSON-RPC endpoint (/a2a/v1) now supports Authorization: Bearer <key> when api_key is configured. Agent card and health endpoints remain public for discovery. Key can be set in config.toml or keys.toml under [a2a]
• A2A task persistence -- Tasks are persisted to SQLite (a2a_tasks table, auto-migration) on create, complete, fail, and cancel. Active tasks are restored from DB on server startup so in-flight work survives restarts
• A2A SSE streaming (message/stream) -- Real-time task updates via Server-Sent Events per A2A spec. Each SSE data: line is a JSON-RPC 2.0 response containing a Task, TaskStatusUpdateEvent (with final: true on completion), or TaskArtifactUpdateEvent. Agent card now advertises streaming: true

0.2.31 - 2026-02-24

Fixed

• Tool calls stacking into one giant group on reload — Removed cross-iteration merge logic that collapsed all consecutive tool groups into a single "N tool calls" block, eating intermediate text between iterations. Each iteration's <!-- tools-v2: --> marker now produces its own collapsible group, matching live session behavior
• Tool group ordering during live streaming — IntermediateText handler flushed the previous iteration's tool group after pushing the new step's text, causing tools to appear below the wrong text. Now flushes tools first, matching DB order
• Ctrl+O blocked during approval — All non-approval keys were eaten when an approval dialog was pending, preventing users from collapsing expanded tool groups to see the approval. Ctrl+O now works during approval
• Auto-collapse tool groups on approval — When an approval request arrives, all tool groups are automatically collapsed so the approval dialog is immediately visible without manual intervention
• EXA MCP fallback on empty API key — Empty string API key ("") caused EXA to attempt direct API mode instead of free MCP. Now treats empty keys as absent, correctly falling back to MCP (aaefd3d)
• Brave search registered without enabled flag — brave_search tool registered whenever an API key existed, ignoring enabled = false in config.toml. Now requires both enabled = true and a valid API key

0.2.30 - 2026-02-24

Added

• Agent-to-Agent (A2A) Protocol — HTTP gateway implementing A2A Protocol RC v1.0 for peer-to-peer agent communication via JSON-RPC 2.0. Supports message/send, tasks/get, tasks/cancel. Contributed by @koatora20 in #9
• Bee Colony Debate — Multi-agent structured debate protocol based on ReConcile (ACL 2024) confidence-weighted voting. Configurable rounds with knowledge-enriched context from QMD memory search
• Dynamic Agent Card — /.well-known/agent.json endpoint with skills generated from the live tool registry
• A2A Documentation — Config example, README section with curl examples, TOOLS.md/SECURITY.md/BOOTSTRAP.md reference templates updated

Fixed

• Tool calls vanishing from TUI — Tool call context (the collapsible bullet with tool names and output) disappeared from the chat after the agent responded. Tool group was being attached to a previous assistant message instead of rendered inline before the current response. Now matches the DB reload layout: tool calls appear above the response text, visible in both live and reloaded sessions
• Tool loop false positives — web_search and http_request calls with different arguments were treated as identical by the loop detector, killing legitimate multi-search flows. Signatures now include query/URL arguments. Thresholds raised (8 default, 4 for modification tools) with a 50-call history window
• Tool call groups splitting on session reload — Each tool-loop iteration wrote a separate DB marker, so "2 tool calls" became two "1 tool call" entries on reload. Fixed in v0.2.31
• Brave search registered without enabled flag — brave_search tool was available to the agent even when enabled = false in config.toml. Now requires both enabled = true and API key
• EXA MCP fallback on empty API key — Empty string API key ("") in keys.toml caused EXA to use direct API mode instead of free MCP mode. Now treats empty keys as absent, correctly falling back to MCP
• A2A: Removed unused rusqlite dependency — A2A handler no longer pulls in rusqlite; uses existing SQLite infrastructure
• A2A: UTF-8 slicing safety — Fixed potential panic on multi-byte characters in message truncation
• A2A: Restrictive CORS by default — No cross-origin requests allowed unless allowed_origins is explicitly configured
• A2A: Handler module split — Monolithic handler.rs split into handler/mod.rs, handler/service.rs, handler/processing.rs for maintainability

Changed

• A2A: Agent card uses tool registry — Skills reflect actual available tools instead of hardcoded list
• A2A: Server wiring — Proper integration with AppState, config, and tool registry
• Web search defaults in README — Updated to reflect DuckDuckGo + EXA as default (no key needed), Brave as optional

0.2.29 - 2026-02-24

Added

• Tool Parameter Normalization — Centralized alias map in tool registry corrects common LLM parameter name mistakes (query→pattern, cmd→command, file→path) before validation. Works across all tools
• Brain Tool Reference — System prompt lists exact required parameter names for each tool
• TOOLS.md Parameter Table — New user template includes tool parameter quick-reference table

Fixed

• Token Counting for OpenAI-Compatible Providers — stream_complete now reads input_tokens from MessageDelta events. Previously always 0 for MiniMax and other OpenAI-compatible providers, causing incorrect session token totals and context percentage
• Session Search UTF-8 Crash — Fixed panic on multi-byte characters when truncating message content (floor_char_boundary instead of raw byte slice)
• Session Search Deadlock — Search uses try_lock() on embedding engine mutex with FTS-only fallback when backfill is running
• Embedding Backfill Lock Contention — Processes one document at a time, releasing engine lock between each
• Tool Loop False Positive — session_search loop detector signature includes operation:query to distinguish calls
• Grep Traversal Performance — Skips target/, node_modules/, .git/ and other heavy directories; default limit of 200 matches
• Thinking Indicator Overlap — "OpenCrabs is thinking..." no longer overlaps chat content
• App Exit Hang — process::exit() prevents tokio runtime hanging on spawn_blocking threads
• Ctrl+C Force Exit — Cancel token + 1-second timeout fallback when tools are stuck

Changed

• App Module Split — app.rs (4,960 lines) split into state.rs, input.rs, messaging.rs, plan_exec.rs, dialogs.rs with mod.rs declarations only
• Doc Comments — Converted // to /// doc comments across codebase
• 7 Test Fixes — Fixed test_create_provider_no_credentials (PlaceholderProvider) and 6 onboarding tests (config pollution, channel routing)

0.2.28 - 2026-02-23

Added

• Brain Setup Persistence — BrainSetup step loads existing USER.md/IDENTITY.md from workspace as truncated preview on re-run. No extra files — brain files are the source of truth
• Brain Setup Skip — Esc to skip, unchanged inputs skip regeneration, empty inputs skip gracefully
• Brain Regeneration Context — On re-run, LLM receives current workspace brain files (not static templates), preserving manual edits as context. Generated content overwrites existing files
• Splash Auto-Close — Splash screen auto-closes after 3 seconds
• Slack Debug Logging — Added debug tracing for Slack message routing (user, channel, bot_id)

Fixed

• Model List Isolation — Minimax and Custom provider model lists no longer mix. Each provider loads only its own models from config.toml.example. Previously load_default_models() dumped all providers into one shared list
• Workspace Path Trim — Workspace path is trimmed on confirm, preventing ghost directories from trailing spaces
• HealthCheck Skipping BrainSetup — HealthCheck step returned WizardAction::Complete immediately, skipping BrainSetup. Now returns WizardAction::None to advance to BrainSetup
• Brain File Overwrite on Regeneration — apply_config() skipped writing brain files if they already existed, even after regeneration. Now overwrites when AI-generated content is available

Changed

• Renamed about_agent → about_opencrabs — Field and label renamed from "Your Agent" to "Your OpenCrabs" for clarity

0.2.27 - 2026-02-23

Added

• Named Custom Providers — Define multiple named OpenAI-compatible providers via [providers.custom.<name>] (e.g. lm_studio, ollama). First enabled one is used. Legacy flat [providers.custom] format still supported

Fixed

• Stream Deduplication — Fixed duplicated agent messages in chat when using LM Studio and other custom providers. Some providers send the full response in the final chunk's message field — falling back to message after receiving delta content duplicated everything
• Database Path Tilde Expansion — ~ in database path config was treated literally, creating a ~/ directory inside the repo. Added expand_tilde() to resolve to actual home directory
• WhatsApp Onboarding — Fixed WhatsApp channel setup to include QR code pairing step with auto-advance, skip and retry
• Channel Onboarding Allowed Lists — Fixed missing allowed users/channels/phones input fields on Telegram, Discord, WhatsApp and Slack setup screens

Changed

• README — Provider examples updated to named custom provider format ([providers.custom.lm_studio])
• config.toml.example — Database path uses smart default, custom providers use named format

0.2.26 - 2026-02-22

Added

• Streaming Tool Call Accumulation — OpenRouter and Custom providers now correctly handle streaming tool calls. Added StreamingToolCall/StreamingFunctionCall structs with optional fields for incremental SSE deserialization, plus ToolCallAccum state machine that accumulates id, name, and arguments across chunks and emits on finish_reason: "tool_calls" or [DONE]
• Input Sanitization — Paste handler strips \r\n, takes first line only, trims whitespace. Storage layer (write_secret_key, write_key) also sanitizes before writing to TOML files
• Auto-append /chat/completions — Custom provider factory auto-appends /chat/completions to base URLs that don't include it, preventing silent 404s
• Provider + Model in Completion — Onboarding completion message now shows which provider and model were selected

Fixed

• Streaming Tool Calls Failing on OpenRouter/Custom — Root cause: StreamingToolCall struct required id and type fields but SSE continuation chunks only send index + function.arguments. Made all fields optional except index. Removed unused type field
• API Key Header Panic — headers() used .expect() which panicked on invalid key characters (e.g. \r from paste). Now returns Result<HeaderMap, ProviderError> with descriptive error
• Log Directory Path — Logs were stored in cwd/.opencrabs/logs/ (inside the repo) instead of ~/.opencrabs/logs/ (user workspace). Fixed LogConfig, get_log_path(), and cleanup_old_logs() to use home directory
• Config/Keys Overwrite — Config::save() was called in app.rs and onboarding.rs, destructively overwriting the entire TOML file. Replaced all instances with individual write_key()/write_secret_key() calls that read-modify-write without losing unrelated sections
• Custom Provider Using Wrong Field — Custom provider used custom_api_key while all other providers used api_key_input. Unified to api_key_input across all providers
• Sentinel Prepended to Key — __EXISTING_KEY__ sentinel was prepended to actual API key on paste. Fixed CustomApiKey handlers to clear sentinel before appending new input
• URL Appended to Key — Pasting from clipboard could include \r and trailing URL text in API key field. Added paste sanitization at input handler and storage layer

Changed

• Renamed openai.rs → custom_openai_compatible.rs — Reflects that this module handles all OpenAI-compatible APIs (OpenRouter, Minimax, Custom, LM Studio, Ollama), not just official OpenAI
• Onboarding Simplified — Removed ~300 lines of dead in-memory config construction from apply_config(); all config writes now use individual write_key()/write_secret_key() calls
• keys.toml is Single Secret Source — All API keys, bot tokens, and search keys are stored in ~/.opencrabs/keys.toml. No more env vars or OS keyring for secrets. config.toml holds non-sensitive settings only

0.2.25 - 2026-02-21

Added

• Token Usage for MiniMax/OpenRouter — Added stream_options: {include_usage: true} to streaming requests; extracts and logs token usage from final chunk
• Shutdown Logo — Shows ASCII logo with rolling goodbye message on terminal when exiting

Fixed

• Duplicate Messages — Fixed duplicate assistant messages appearing when IntermediateText already added content
• Tool Call Flow — Tool calls now appear as separate messages after assistant text, flowing naturally between steps
• Empty Content Rendering — Fixed assistant messages showing empty during session (was showing correctly after restart)
• Thinking Indicator — Moved "OpenCrabs is thinking..." indicator to sticky position at bottom of chat (above input field), always visible to users

Changed

• Message Ordering — Queued messages now appear at very bottom of conversation (after all assistant/tool messages), above input field
• README — Added GitHub stars call-to-action

0.2.24 - 2026-02-21

Added

• MiniMax Provider Support — Added MiniMax as new LLM provider (OpenAI-compatible). Does not have /models endpoint, uses config_models for model list
• Onboarding Wizard — Full onboarding flow for first-time setup with provider selection
• Model Selector — Slash command /models to change provider and model with live fetching, search filter
• Tool Call Expanded View — Ctrl+O expands tool context with gray background; diff coloring (+ green, - red)
• API Keys in keys.toml — API keys now stored in separate ~/.opencrabs/keys.toml (chmod 600)
• STT/TTS Provider Config — Added providers.stt.groq and providers.tts.openai config sections

Fixed

• MiniMax Tool Calls — Fixed tool call parsing for MiniMax (empty arguments issue)
• Context Compaction Crash — Fixed orphaned tool_result crash after compaction
• Onboarding Persistence — Provider selection and settings now persist correctly
• Model Selector Flow — Multiple fixes for persistence, search, scrolling, Enter key behavior
• Compaction Crash (400 — Orphaned tool_result) — After any trim or compaction, a user(tool_result) message could be left at the front of history without its preceding assistant(tool_use). The Anthropic API rejects this with a 400 error, crashing the next compaction attempt. Fixed at three layers: trim_to_fit and trim_to_target now call drop_leading_orphan_tool_results() after each removal; compact_with_summary advances keep_start past any leading orphaned tool_result messages; compact_context skips them before sending to the API as a safety net. Conversation continues normally after compaction with no tool call drops
• Compaction Summary as Assistant Message — Compaction summary was stored in a details field and hidden behind Ctrl+O. Now rendered as a real assistant chat message in the conversation flow. Tool calls that follow appear below it as normal tool groups with Ctrl+O expand/collapse
• config.toml Model Priority over .env — ANTHROPIC_MAX_MODEL env var was overwriting the model set in config.toml, reversing the intended priority. Now config.toml wins; .env is only a fallback when no model is configured in TOML
• Stale Terminal on exec() Restart — /rebuild hot-restart left stale rendered content from the previous process visible briefly. Terminal is now fully cleared immediately after the new process takes over

Changed

• Remove Qwen and Azure — These providers are no longer supported
• README Updated — Added MiniMax documentation, keys.toml instructions

0.2.23 - 2026-02-20

Added

• session_search Tool — Hybrid FTS5+vector search across all chat sessions (list/search operations)
• History Paging — Cap initial display at 200k tokens, Ctrl+O loads 100k more from DB
• Onboarding Model Filter — Type to search models, Esc clears filter

Fixed

• Onboard Centering — Header/footer center independently, content block centers as uniform group
• Onboard Scroll — ProviderAuth tracks focused_line for proper scroll anchoring
• Content Clipping — Content no longer clips top border on overflow screens

Changed

• Compaction Display — Now clears TUI display fully, shows summary as fresh start
• Render history_marker — Rendered as dim italic in chat view

0.2.22 - 2026-02-19

Added

• /cd Command — Change working directory at runtime via slash command or agent NLP. Opens a directory picker (same UI as @ file picker). Persists to config.toml. Agent can also call config_manager with set_working_directory
• slash_command Tool — Agent-callable tool to invoke any slash command programmatically: /cd, /compact, /rebuild, /approve, and all user-defined commands from commands.toml. Makes the agent aware of and able to trigger any slash command
• Edit Diff Context — Edit tool now includes a compact unified diff in its output. Renderer colors + lines green, - lines red, @@ lines cyan — giving both user and agent clear visual context of changes

Fixed

• Stderr Bleeding into TUI — Replaced all unsafe libc dup2//dev/null hacks with llama-cpp-2's proper send_logs_to_tracing(LogOptions::default().with_logs_enabled(false)) API. Called once at engine init — kills all llama.cpp C-level stderr output permanently. Removed libc dependency entirely
• Compaction Summary Never Visible — System messages were rendered as a single Span on one Line — Ratatui clips at terminal width, so multi-paragraph summaries were silently swallowed. Fixed: newline-aware rendering with ⚡ yellow label. Compaction summary now goes into expandable details (Ctrl+O to read)
• Tool Approval Disappearing — Removed 4 messages.retain() calls that deleted approval messages immediately after denial, before the user could see or interact with them

Changed

• Install Instructions — README now includes "Make It Available System-Wide" section with symlink/copy instructions
• Brain Templates — BOOT.md, TOOLS.md, AGENTS.md updated to document /cd and config_manager working directory control

0.2.21 - 2026-02-19

Changed

• Module Restructure — Merged src/llm/ (agent, provider, tools, tokenizer) into src/brain/. Brain is now the single intelligence layer — no split across two top-level modules
• Channel Consolidation — Moved src/slack/, src/telegram/, src/whatsapp/, src/discord/, and src/voice/ into src/channels/. All messaging integrations + voice (STT/TTS) live under one module with feature-gated submodules
• Ctrl+O Expands All — Ctrl+O now toggles expand/collapse on ALL tool call groups in the session, not just the most recent one

Fixed

• Tool Approval Not Rendering — Fixed approval prompts not appearing in long-context sessions when user had scrolled up. auto_scroll is now reset to true when an approval arrives, ensuring the viewport scrolls to show it
• Tool Call Details Move — Fixed use of moved value for tool call details field in ToolCallCompleted handler

0.2.20 - 2026-02-19

Added

• /whisper Command — One-command setup for system-wide voice-to-text. Auto-downloads WhisperCrabs binary, launches floating mic button. Speak from any app, transcription auto-copies to clipboard
• SystemMessage Event — New TUI event variant for async tasks to push messages into chat

Fixed

• Embedding Stderr Bleed — Suppressed llama.cpp C-level stderr during embed_document() and embed_batch_with_progress(), not just model load. Fixes garbled TUI output during memory indexing
• Slash Autocomplete Dedup — User-defined commands that shadow built-in names no longer show twice in autocomplete dropdown
• Slash Autocomplete Width — Dropdown auto-sizes to fit content instead of hardcoded 40 chars. Added inner padding on all sides
• Help Screen — Added missing /rebuild and /whisper to /help slash commands list
• Cleartext Logging (CodeQL) — Removed all println! calls from provider factory that wrote to stdout (corrupts TUI). Kept tracing::info! for structured logging
• Stray Print Statements — Removed debug println! from wacore encoder, replaced eprintln! in onboarding tests with silent returns

Changed

• Docker Files Relocated — Moved docker/ from project root to src/docker/, updated all references in README and compose.yml
• Clippy Clean — Fixed collapsible_if warnings in onboarding and app, map_or → is_some_and

0.2.19 - 2026-02-18

Changed

• Cleaner Chat UI — Replaced role labels with visual indicators: ❯ for user messages, ● for assistant messages. User messages get subtle dark background for visual separation. Removed horizontal dividers and input box title for a cleaner look
• Alt+Arrow Word Navigation — Added Alt+Left / Alt+Right as alternatives to Ctrl+Left / Ctrl+Right for word jumping (macOS compatibility)
• Branding — Thinking/streaming indicators now show 🦀 OpenCrabs instead of model name

0.2.18 - 2026-02-18

Added

• OpenRouter Provider -- First-class OpenRouter support in onboarding wizard. One API key, 400+ models including free and stealth models (DeepSeek, Llama, Mistral, Qwen, Gemma, and more). Live model list fetched from openrouter.ai/api/v1/models
• Live Model Fetching -- /models command and onboarding wizard now fetch available models live from provider APIs (Anthropic, OpenAI, OpenRouter). When a new model drops, it shows up immediately — no binary update needed. Falls back to hardcoded list if offline
• Provider::fetch_models() Trait Method -- All providers implement async model fetching with graceful fallback to static lists

Changed

• Onboarding Wizard -- Provider step 2 now shows live model list fetched from API after entering key. Shows "(fetching...)" while loading. OpenRouter added as 5th provider option
• Removed cargo publish from CI -- Release workflow no longer attempts crates.io publish (was never configured, caused false failures)

0.2.17 - 2026-02-18

Changed

• QMD Vector Search + RRF -- qmd's EmbeddingEngine (embeddinggemma-300M, 768-dim GGUF) wired up alongside FTS5 with Reciprocal Rank Fusion. Local model, no API key, zero cost, works offline. Auto-downloads ~300MB on first use, falls back to FTS-only when unavailable
• Batch Embedding Backfill -- On startup reindex, documents missing embeddings are batch-embedded via qmd. Single-file indexes (post-compaction) embed immediately when engine is warm
• Discord Voice (STT + TTS) -- Discord bot now transcribes audio attachments via Groq Whisper and replies with synthesized voice (OpenAI TTS) when enabled
• WhatsApp Voice (STT) -- WhatsApp bot now transcribes voice notes via Groq Whisper. Text replies only (media upload for TTS pending)
• CI Release Workflow -- Fixed nightly toolchain for all build targets, added ARM64 cross-linker config
• AVX CPU Guard -- Embedding engine checks for AVX support at init; gracefully falls back to FTS-only on older CPUs
• Stderr Suppression -- llama.cpp C-level stderr output redirected to /dev/null during model load to prevent TUI corruption

0.2.16 - 2026-02-18

Changed

• QMD Crate for Memory Search -- Replaced homebrew FTS5 implementation with the qmd crate (BM25 search, SHA-256 content hashing, collection management). Upgraded sqlx to 0.9 (git main) to resolve libsqlite3-sys linking conflict
• Brain Files Indexed -- Memory search now indexes workspace brain files (SOUL.md, IDENTITY.md, MEMORY.md, etc.) alongside daily compaction logs for richer search context
• Dynamic Welcome Messages -- All channel connect tools (Telegram, Discord, Slack, WhatsApp) now instruct the agent to craft a creative, personality-driven welcome message on successful connection instead of hardcoded greetings
• WhatsApp Welcome Removed -- Replaced hardcoded WhatsApp welcome spawn with agent-generated message via whatsapp_send tool
• Patches Relocated -- Moved wacore-binary patch from patches/ to src/patches/, stripped benchmarks and registry metadata

Added

• Discord channel_id Parameter -- Optional channel_id input on discord_connect so the bot can send welcome messages immediately after connection
• Slack channel_id Parameter -- Optional channel_id input on slack_connect for the same purpose
• Telegram Owner Chat ID -- telegram_connect now sets the owner chat ID from the first allowed user at connection time
• QMD Memory Benchmarks -- Criterion benchmarks for qmd store operations: index file (203µs), hash skip (18µs), FTS5 search (381µs–2.4ms), bulk reindex 50 files (11.3ms), store open (1.7ms)

0.2.15 - 2026-02-17

Changed

• Built-in FTS5 Memory Search -- Replaced external QMD CLI dependency with native SQLite FTS5 full-text search. Zero new dependencies (uses existing sqlx), always-on memory search with no separate binary to install. BM25-ranked results with porter stemming and snippet extraction
• Memory Search Always Available -- Sidebar now shows "Memory search" with a permanent green dot instead of conditional "QMD search" that required an external binary
• Targeted Index After Compaction -- After context compaction, only the updated daily memory file is indexed (via index_file) instead of triggering a full qmd update subprocess
• Startup Background Reindex -- On launch, existing memory files are indexed in the background so memory_search is immediately useful for returning users

Added

• FTS5 Memory Module -- New async API: get_pool() (lazy singleton), search() (BM25 MATCH), index_file() (single file, hash-skip), reindex() (full walk + prune deleted). Schema: memory_docs content table + memory_fts FTS5 virtual table with sync triggers
• Memory Search Tests -- Unit tests for FTS5 init, index, search, hash-based skip, and content update re-indexing
• Performance Benchmarks in README -- Real release-build numbers: ~0.4ms/query, ~0.3ms/file index, 15ms full reindex of 50 files
• Resource Footprint Table in README -- Branded stats table with binary size, RAM, storage, and FTS5 search latency

Removed

• QMD CLI Dependency -- Removed all Command::new("qmd") subprocess calls: is_qmd_available(), ensure_collection(), search() (sync), reindex_background()

0.2.14 - 2026-02-17

Added

• Discord Integration -- Full Discord bot with message forwarding, per-user session routing, image attachment support, proactive messaging via discord_send tool, and dynamic connection via discord_connect tool
• Slack Integration -- Full Slack bot via Socket Mode (no public endpoint needed) with message forwarding, session sharing, proactive messaging via slack_send tool, and dynamic connection via slack_connect tool
• Secure Bot Messaging: respond_to Mode -- New respond_to config field for all platforms: "mention" (default, most secure), "all" (old behavior), or "dm_only". DMs always get a response regardless of mode
• Channel Allowlists -- New allowed_channels config field restricts which group channels bots are active in. Empty = all channels. DMs always pass
• Bot @Mention Detection -- Discord checks msg.mentions for bot user ID, Telegram checks @bot_username or reply-to-bot, Slack checks <@BOT_USER_ID> in text. Bot mention text is stripped before sending to agent
• Bot Identity Caching -- Discord stores bot user ID from ready event, Telegram fetches @username via get_me() at startup, Slack fetches bot user ID via auth.test at startup
• Troubleshooting Section in README -- Documents the known session corruption issue where agent hallucinates tool calls, with workaround (start new session)

Fixed

• Pending Tool Approvals Hanging Agent -- Approval callbacks were never resolved on cancel, error, supersede, or agent completion, causing the agent to hang indefinitely. All code paths now properly deny pending approvals with response_tx.send()
• Stale Approval Cleanup -- Cancel (Escape), error handler, new request, and agent completion all now send deny responses before marking approvals as denied
• Rustls Crypto Provider for Slack -- Install ring crypto provider at startup before any TLS connections, fixing Slack Socket Mode panics

Changed

• Proactive Message Branding Removed -- discord_send, slack_send, telegram_send tools no longer prepend MSG_HEADER to outgoing messages
• Agent Logging -- Improved iteration logging: shows "completed after N tool iterations" or "responded with text only"
• Auto-Approve Feedback -- Selecting "Allow Always" now shows a system message confirming auto-approve is enabled for the session

0.2.13 - 2026-02-17

Added

• Proactive WhatsApp Messaging -- New whatsapp_send agent tool lets the agent send messages to the user (or any allowed phone) at any time, not just in reply to incoming messages
• WhatsApp Welcome Message -- On successful QR pairing, the agent sends a fun random crab greeting to the owner's WhatsApp automatically
• WhatsApp Message Branding -- All outgoing WhatsApp messages are prefixed with 🦀 *OpenCrabs* header so users can distinguish agent replies from their own messages
• WhatsApp device_sent_message Unwrapping -- Recursive unwrap_message() handles WhatsApp's nested message wrappers (device_sent_message, ephemeral_message, view_once_message, document_with_caption_message) to extract actual text content from linked-device messages
• Fun Startup/Shutdown Messages -- Random crab-themed greetings on launch and farewell messages on exit (10 variants each)

Fixed

• WhatsApp Self-Chat Messages Ignored -- Messages from the user's own phone were dropped because is_from_me: true; now only skips messages with the agent's MSG_HEADER prefix to prevent echo loops while accepting user messages from linked devices
• WhatsApp Phone Format Mismatch -- Allowlist comparison failed because config stored +351... but JID user part was 351...; sender_phone() now strips @s.whatsapp.net suffix, allowlist check strips + prefix
• Model Name Missing from Thinking Spinner -- "is thinking" showed without model name because session.model could be Some(""); added .filter(|m| !m.is_empty()) fallback to default_model_name
• WhatsApp SQLx Store Device Serialization -- Device state now serialized via rmp-serde (MessagePack) instead of broken bincode; added rmp-serde dependency under whatsapp feature

Changed

• wacore-binary Direct Dependency -- Added as direct optional dependency for Jid type access (needed by whatsapp_send and whatsapp_connect tools for JID parsing)

Removed

• /model Slash Command -- Removed redundant /model command; /models already provides model switching with selected-model display

0.2.12 - 2026-02-17

Added

• WhatsApp Integration -- Chat with your agent via WhatsApp Web. Connect dynamically at runtime ("connect my WhatsApp") or from the onboarding wizard. QR code pairing displayed in terminal using Unicode block characters, session persists across restarts via SQLite
• WhatsApp Image Support -- Send images to the agent via WhatsApp; they're downloaded, base64-encoded, and forwarded to the AI backend for multimodal analysis
• WhatsApp Connect Tool -- New whatsapp_connect agent tool: generates QR code, waits for scan (2 min timeout), spawns persistent listener, updates config automatically
• Onboarding: Messaging Setup -- New step in both QuickStart and Advanced onboarding modes to enable Telegram and/or WhatsApp channels right after provider auth
• Channel Factory -- Shared ChannelFactory for creating channel agent services at runtime, used by both static startup and dynamic connection tools
• Custom SQLx WhatsApp Store -- wacore::store::Backend implementation using the project's existing sqlx SQLite driver, avoiding the libsqlite3-sys version conflict with whatsapp-rust-sqlite-storage (Diesel-based). 15 tables, 33 trait methods, full test coverage
• Nightly Rust Requirement -- wacore-binary requires #![feature(portable_simd)]; added rust-toolchain.toml pinning to nightly. Local patch for wacore-binary fixes std::simd::Select API breakage on latest nightly

Changed

• Version Numbering -- Corrected from 0.2.2 to 0.2.11 (following 0.2.1), this release is 0.2.12

0.2.11 - 2026-02-16

Fixed

• Context Token Display -- TUI context indicator showed inflated values (e.g. 640K/200K) because input_tokens was accumulated across all tool-loop iterations instead of using the last API call's actual context size; now AgentResponse.context_tokens tracks the last iteration's input_tokens for accurate display while usage still accumulates for correct billing
• Per-Message Token Count -- DisplayMessage.token_count now shows only output tokens (the actual generated content) instead of the inflated input + output sum which double-counted shared context
• Clippy Warning -- Fixed redundant_closure warning in trim_messages_to_budget

Changed

• Compaction Threshold -- Lowered auto-compaction trigger from 80% to 70% of context window for earlier, safer compaction with more headroom
• Token Counting -- trim_messages_to_budget now uses tiktoken (cl100k_base) instead of chars/3 heuristic; history budget targets 60% of context window (was 70%) to leave more room for tool results

Added

• 2 New Tests -- test_context_tokens_is_last_iteration_not_accumulated and test_context_tokens_equals_input_tokens_without_tools verifying correct context vs billing token separation (450 total)

Removed

• Dead Code -- Removed unused format_token_count function and its 5 tests from render.rs

0.2.1 - 2026-02-16

Added

• Config Management Tool -- New config_manager agent tool with 6 operations: read_config, write_config, read_commands, add_command, remove_command, reload; the agent can now read/write config.toml and commands.toml at runtime
• Commands TOML Migration -- User-defined slash commands now stored in commands.toml ([[commands]] array) instead of commands.json; existing commands.json files auto-migrate on first load
• Settings TUI Screen -- Press S for a real Settings screen showing: current provider/model, approval policy, user commands summary, QMD memory search status, and file paths (config, brain, working directory)
• Approval Policy Persistence -- /approve command now saves the selected policy to [agent].approval_policy in config.toml; policy is restored on startup instead of always defaulting to "ask"
• AgentConfig Section -- New [agent] config section with approval_policy ("ask" / "auto-session" / "auto-always") and max_concurrent (default: 4) fields
• Live Config Reload -- Config::reload() method and TuiEvent::ConfigReloaded event for refreshing cached config values after tool writes
• Config Write Helper -- Config::write_key(section, key, value) safely merges key-value pairs into config.toml without overwriting unrelated sections
• Command Management Helpers -- CommandLoader::add_command() and CommandLoader::remove_command() for atomic command CRUD
• 20 New Tests -- 14 onboarding tests (key handlers, mode select, provider navigation, API key input, field flow, validation, model selection, workspace/health/brain defaults) + 6 config tests (AgentConfig defaults, TOML parsing, write_key merge, save round-trip) -- 443 total

Changed

• config.toml.example -- Added [agent] and [voice] example sections with documentation
• Commands Auto-Reload -- After ConfigReloaded event, user commands are refreshed from commands.toml

0.2.0 - 2026-02-15

Added

• 3-Tier Memory System -- OpenCrabs now has a layered memory architecture: (1) Brain MEMORY.md -- user-curated durable memory loaded into system brain every turn, (2) Daily Memory Logs -- auto-compaction summaries saved to ~/.opencrabs/memory/YYYY-MM-DD.md with multiple compactions per day stacking in the same file, (3) Memory Search -- memory_search tool backed by QMD for semantic search across all past daily logs
• Memory Search Tool -- New memory_search agent tool searches past conversation logs via QMD (qmd query --json); gracefully degrades if QMD is not installed, returning a hint to use read_file on daily logs directly
• Compaction Summary Display -- Auto-compaction at 80% context now shows the full summary in chat as a system message instead of running silently; users see exactly what the agent remembered
• Scroll While Streaming -- Users can scroll up during streaming without being yanked back to the bottom; auto_scroll flag disables on user scroll, re-enables when scrolled back to bottom or on message send
• QMD Auto-Index -- After each compaction, qmd update is triggered in the background to keep the memory search index current
• Memory Module -- New src/memory/mod.rs module with QMD wrapper: availability check, collection management, search, and background re-indexing
• Path Consolidation -- All data now lives under ~/.opencrabs/ (config, database, brain, memory, history, logs)
• Context Budget Awareness -- Tool definition overhead (~500 tokens per tool) now factored into context usage calculation, preventing "prompt too long" errors

Changed

• Compaction Target -- Compaction summaries now write to daily logs (~/.opencrabs/memory/YYYY-MM-DD.md) instead of appending to brain workspace MEMORY.md; brain MEMORY.md remains user-curated and untouched by auto-compaction
• Local Timestamps -- Daily memory logs use chrono::Local instead of UTC for human-readable timestamps

0.1.9 - 2026-02-15

Added

• Cursor Navigation -- Full cursor movement in input: Left/Right arrows, Ctrl+Left/Right word jump, Home/End, Delete key, Backspace at cursor position, word delete (Alt/Ctrl+Backspace), character and paste insertion at cursor position, cursor renders at correct position
• Input History Persistence -- Command history saved to ~/.config/opencrabs/history.txt (one line per entry), loaded on startup, appended on each send, capped at 500 entries, survives restarts
• Real-time Streaming -- Added stream_complete() method that streams text chunks from the provider via StreamingChunk progress events, replacing the old blocking provider.complete() call
• Streaming Spinner -- Animated spinner shows "claude-opus is responding..." with streamed text below; "thinking..." spinner shows only before streaming begins
• Inline Plan Approval -- Plan approval now renders as an interactive inline selector with arrow keys (Approve / Reject / Request Changes / View Plan) instead of plain text Ctrl key instructions
• Telegram Photo Support -- Incoming photos download at largest resolution, saved to temp file, forwarded as <<IMG:path>> caption; image documents detected via image/* MIME type; temp files cleaned up after 30 seconds
• Error Message Rendering -- app.error_message is now rendered in the chat UI (was previously set but never displayed)
• Default Model Name -- New sessions show the actual provider model name (e.g. claude-opus-4-6) as placeholder instead of generic "AI"
• Debug Logging -- DEBUG_LOGS_LOCATION env var sets custom log directory; --debug CLI flag enables debug mode
• 8 New Tests -- stream_complete_text_only, stream_complete_with_tool_use, streaming_chunks_emitted, markdown_to_telegram_html_*, escape_html, img_marker_format (412 total)

Fixed

• SSE Parser Cross-Chunk Buffering -- TCP chunks splitting JSON events mid-string caused EOF while parsing a string errors and silent response drops; parser now buffers partial lines across chunks with Arc<Mutex<String>>, only parsing complete newline-terminated lines
• Stale Approval Cleanup -- Old Pending approval messages permanently hid streaming responses; now cleared on new message send, new approval request, and response completion
• Approval Dialog Reset -- approval_auto_always reset on session create/load; inline "Always" now sets approval_auto_session (resets on session change) instead of approval_auto_always
• Brain File Path -- Brain prompt builder used wrong path for workspace files
• Abort During Streaming -- Cancel token properly wired through streaming flow for Escape×2 abort

Changed

• README -- Expanded self-sustaining section with /rebuild command, SelfUpdater module, session persistence, brain live-editing documentation

0.1.8 - 2026-02-15

Added

• Image Input Support -- Paste image paths or URLs into the input; auto-detected and attached as vision content blocks for multimodal models (handles paths with spaces)
• Attachment Indicator -- Attached images show as [IMG1:filename.png] in the input box title bar; user messages display [IMG: filename.png]
• Tool Context Persistence -- Tool call groups are now saved to the database and reconstructed on session reload; no more vanishing tool history
• Intermediate Text Display -- Agent text between tool call batches now appears interleaved in the chat, matching Claude Code's behavior

Fixed

• Tool Descriptions Showing "?" -- Approval dialog showed "Edit ?" instead of file paths; fixed parameter key mismatches (path not file_path, operation not action)
• Raw Tool JSON in Chat -- [Tool: read_file]{json} was dumped into assistant messages; now only text blocks are displayed, tool calls shown via the tool group UI
• Loop Detection Wrong Keys -- Tool loop detection used file_path for read/write/edit; fixed to path
• Telegram Text+Voice Order -- Text reply now always sent first, voice note follows (was skipping text on TTS success)

Changed

• base64 dependency -- Re-added base64 = "0.22.1" for image encoding (was removed in dep cleanup but now needed)

0.1.7 - 2026-02-14

Added

• Voice Integration (STT) -- Incoming Telegram voice notes are transcribed via Groq Whisper (whisper-large-v3-turbo) and processed as text by the agent
• Voice Integration (TTS) -- Agent replies to voice notes with audio via OpenAI TTS (gpt-4o-mini-tts, ash voice); falls back to text if TTS is disabled or fails
• Onboarding: Telegram Setup -- New wizard step with BotFather instructions, bot token input (masked), and user ID guidance; auto-detects existing env/keyring values
• Onboarding: Voice Setup -- New wizard step for Groq API key (STT) and TTS toggle with ash voice label; auto-detects GROQ_API_KEY from environment
• Sessions Dialog: Context Info -- /sessions now shows token count per session (12.5K tok, 2.1M tok) and live context window percentage for the current session with color coding (green/yellow/red)
• Tool Descriptions in Approval -- Approval dialog now shows actual file paths and parameters (e.g. "Edit /src/tui/render.rs") instead of raw tool names ("edit_file")
• Shared Telegram Session -- Owner's Telegram messages now use the same session as the TUI terminal; no more separate sessions that could pick the wrong model

Changed

• Provider Priority -- Factory order changed to Qwen → Anthropic → OpenAI; Anthropic is now always preferred over OpenAI for text generation
• OPENAI_API_KEY Isolation -- OPENAI_API_KEY no longer auto-creates an OpenAI text provider; it is only used for TTS (gpt-4o-mini-tts), never for text generation unless explicitly configured
• Async Terminal Events -- Replaced blocking crossterm::event::poll() with async EventStream + tokio::select! to prevent TUI freezes during I/O-heavy operations

Fixed

• Model Contamination -- OPENAI_API_KEY in .env was causing GPT-4 to be used for text instead of Anthropic Claude; multi-layered fix across factory, env overrides, and TTS key sourcing
• Navigation Slowdown -- TUI became sluggish after losing terminal focus due to synchronous 100ms blocking poll in async context
• Context Showing 0% -- Loading an existing session showed 0% context; now estimates tokens from message content until real API usage arrives
• Approval Spam -- "edit_file -- approved" messages no longer clutter the chat; approved tool calls are silently removed since the tool group already shows execution progress
• 6 Clippy Warnings -- Fixed collapsible_if (5) and manual_find (1) across onboarding and telegram modules

0.1.6 - 2026-02-14

Added

• Telegram Bot Integration -- Chat with OpenCrabs via Telegram alongside the TUI; bot runs as a background task with full tool access (file ops, search, bash, etc.)
• Telegram Allowlist -- Only allowlisted Telegram user IDs can interact; /start command shows your ID for easy setup
• Telegram Markdown→HTML -- Agent responses are formatted as Telegram-safe HTML with code blocks, inline code, bold, and italic support
• Telegram Message Splitting -- Long responses automatically split at 4096-char Telegram limit, breaking at newlines
• Grouped Tool Calls -- Multiple tool calls in a single agent turn now display as a collapsible group with tree lines (├─ └─) instead of individual messages
• Claude Code-Style Approval -- Tool approval dialog rewritten as vertical selector with ❯ Yes / Always / No matching Claude Code's UX
• Emergency Compaction Retry -- If the LLM provider returns "prompt too long", automatically compact context and retry instead of failing

Changed

• Token Estimation -- Changed from chars/4 to chars/3 for more conservative estimation, preventing context overflows that the old estimate missed
• Compaction Accounts for Tools -- Auto-compaction threshold now reserves ~500 tokens per registered tool for schema overhead, preventing "prompt too long" errors
• Telegram Feature Default -- telegram feature now included in default features (no need for --features telegram)

Fixed

• Context % Showing 2369% -- context_usage_percent() was summing all historical token counts; now uses only the latest response's input_tokens
• TUI Lag After First Request -- active_tool_group wasn't cleaned up on error/abort paths, causing UI to hang
• Telegram Bot No Response -- Bot was calling send_message (no tools) instead of send_message_with_tools; also needed auto_approve_tools: true since there's no TUI for approval

0.1.5 - 2026-02-14

Added

• Context Usage Indicator -- Input box shows live Context: X% with color coding: green (<60%), yellow (60-80%), red (>80%) so you always know how close you are to the context limit
• Auto-Compaction -- When context usage exceeds 80%, automatically sends conversation to the LLM for a structured breakdown summary (Current Task, Key Decisions, Files Modified, Current State, Important Context, Errors & Solutions), saves to MEMORY.md, and trims context keeping the last 8 messages + summary for seamless continuation
• /compact Command -- Manually trigger context compaction at any time via slash command
• Brave Search Tool -- Real-time web search via Brave Search API (set BRAVE_API_KEY); great if you already have a Brave API key or want a free-tier option
• EXA Search Tool -- Neural-powered web search via EXA AI; works out of the box via free hosted MCP endpoint (no API key needed). Set EXA_API_KEY for direct API access with higher rate limits

Changed

• EXA Always Available -- EXA search registers unconditionally via free MCP endpoint; Brave still requires BRAVE_API_KEY

0.1.4 - 2026-02-14

Added

• Inline Tool Progress -- Tool executions now show inline in chat with human-readable descriptions (e.g. "Read src/main.rs", "bash: cargo check", "Edited src/app.rs") instead of invisible spinner
• Expand/Collapse Tool Details -- Press Ctrl+O to expand or collapse tool output details on completion messages, inspired by Claude Code's UX
• Abort Processing -- Press Escape twice within 3 seconds to cancel an in-progress agent request via CancellationToken
• Active Input During Processing -- Input box stays active with cursor visible while agent is processing; border remains steel blue
• Processing Guard -- Prevents sending a second message while one is already processing; shows "Please wait or press Esc x2 to abort"
• Progress Callback System -- New ProgressCallback / ProgressEvent architecture emitting Thinking, ToolStarted, and ToolCompleted events from agent service to TUI
• LLM-Controlled Bash Timeout -- Bash tool now accepts timeout_secs from the LLM (capped at 600s), default raised from 30s to 120s

Changed

• Silent Auto-Approved Tools -- Auto-approved tool calls no longer spam the chat; only completion descriptions shown
• Approval Never Times Out -- Tool approval requests wait indefinitely until the user acts (no more 5-minute timeout)
• Approval UI De-Emojified -- All emojis removed from approval rendering; clean text-only UI
• Yolo Mode Always Visible -- All three approval tiers (Allow once, Allow all session, Yolo mode) always visible with color-coding (green/yellow/red) in inline approval

Fixed

• Race Condition on Double Send -- Added is_processing guard in send_message() preventing overlapping agent requests

0.1.3 - 2026-02-14

Added

• Inline Tool Approval — Tool permission requests now render inline in chat instead of a blocking overlay dialog, with three options: Allow once, Allow all for this task, Allow all moving forward
• /approve Command — Resets tool approval policy back to "always ask"
• Word Deletion — Ctrl+Backspace and Alt+Backspace delete the last word in input
• Scroll Support — Arrow keys and Page Up/Down now scroll Help, Sessions, and Settings screens
• Tool Approval Docs — README section documenting inline approval keybindings and options

Changed

• Ctrl+C Behavior — First press clears input, second press within 3 seconds quits (was immediate quit)
• Help Screen — Redesigned as 2-column layout filling full terminal width instead of narrow single column
• Status Bar Removed — Bottom status bar eliminated for cleaner UI; mode info shown in header only
• Ctrl+H Removed — Help shortcut removed (use /help instead); fixes Ctrl+Backspace conflict where terminals send Ctrl+H for Ctrl+Backspace

Removed

• MCP Module — Deleted empty placeholder src/mcp/ directory (unused stubs, zero functionality)
• Overlay Approval Dialog — Replaced by inline approval in chat
• Bottom Status Bar — Removed entirely for more screen space

0.1.2 - 2026-02-14

Added

• Onboarding Wizard — 8-step wizard with QuickStart/Advanced modes for first-time setup
• AI Brain Personalization — Generates all 6 workspace brain files (SOUL, IDENTITY, USER, AGENTS, TOOLS, MEMORY) from user input during onboarding
• Session Management — /sessions command, rename sessions (R), delete sessions (D) from session list
• Mouse Scroll — Mouse wheel scrolls chat history
• Dynamic Input Height — Input area grows with content, 1-line default
• Screenshots — Added UI screenshots to README (splash, onboarding, chat)

Changed

• Unified Anthropic Provider — Auto-detects OAuth tokens vs API keys from env/keyring
• Pre-wrapped Chat Lines — Consistent left padding for all chat messages
• Updated Model List — Added claude-opus-4-6, gpt-5.1-codex-mini, gemini-3-flash-preview, qwen3-coder-next
• Cleaner UI — Removed emojis, reordered status bar
• README — Added screenshots, updated structure

0.1.1 - 2026-02-14

Added

• Dynamic Brain System — Replace hardcoded system prompt with brain loader that reads workspace MD files (SOUL, IDENTITY, USER, AGENTS, TOOLS, MEMORY) per-turn from ~/opencrab/brain/workspace/
• CommandLoader — User-defined slash commands via commands.json, auto-reloaded after each agent response
• SelfUpdater — Build/test/restart via Unix exec() for hot self-update (/rebuild command)
• RestartPending Mode — Confirmation dialog in TUI after successful rebuild
• Onboarding Docs — Scaffolding for onboarding documentation

Changed

• system_prompt → system_brain — Renamed across entire codebase to reflect dynamic brain architecture
• /help Fixed — Opens Help dialog instead of pushing text message into chat

0.1.0 - 2026-02-14

Added

• Anthropic OAuth Support — Claude Max / setup-token authentication via ANTHROPIC_MAX_SETUP_TOKEN with automatic sk-ant-oat prefix detection, Authorization: Bearer header, and anthropic-beta: oauth-2025-04-20 header
• Claude 4.x Models — Support for claude-opus-4-6, claude-sonnet-4-5-20250929, claude-haiku-4-5-20251001 with updated pricing and context windows
• .env Auto-Loading — dotenvy integration loads .env at startup automatically
• CHANGELOG.md — Project changelog following Keep a Changelog format
• New Branding — OpenCrab ASCII art, "Shell Yeah! AI Orchestration at Rust Speed." tagline, crab icon throughout

Changed

• Rust Edition 2024 — Upgraded from edition 2021 to 2024
• All Dependencies Updated — Every crate bumped to latest stable (ratatui 0.30, crossterm 0.29, pulldown-cmark 0.13, rand 0.9, dashmap 6.1, notify 8.2, git2 0.20, zip 6.0, tree-sitter 0.25, thiserror 2.0, and more)
• Rebranded — "OpenCrab AI Assistant" renamed to "OpenCrab AI Orchestration Agent" across all source files, splash screen, TUI header, system prompt, and documentation
• Enter to Send — Changed message submission from Ctrl+Enter (broken in many terminals) to plain Enter; Alt+Enter / Shift+Enter inserts newline for multi-line input
• Escape Double-Press — Escape now requires double-press within 3 seconds to clear input, preventing accidental loss of typed messages
• TUI Header Model Display — Header now shows the provider's default model immediately instead of "unknown" until first response
• Splash Screen — Updated with OpenCrab ASCII art, new tagline, and author attribution
• Default Max Tokens — Increased from 4096 to 16384 for modern Claude models
• Default Model — Changed from claude-3-5-sonnet-20240620 to claude-sonnet-4-5-20250929
• README.md — Complete rewrite: badges, table of contents, OAuth documentation, updated providers/models, concise structure (764 lines vs 3,497)
• Project Structure — Moved tests/, migrations/, benches/, docs/ inside src/ and updated all references

Fixed

• pulldown-cmark 0.13 API — Tag::Heading tuple to struct variant, Event::End wraps TagEnd, Tag::BlockQuote takes argument
• ratatui 0.29+ — f.size() replaced with f.area(), Backend::Error bounds added (Send + Sync + 'static)
• rand 0.9 — thread_rng() replaced with rng(), gen_range() replaced with random_range()
• Edition 2024 Safety — Removed unsafe std::env::set_var/remove_var from tests, replaced with TOML config parsing

Removed

• Outdated "Claude Max OAuth is NOT supported" disclaimer (it now is)
• Sprint history and "coming soon" filler from README
• Old "Crusty" branding and attribution