v4.5.0 - Secure Computer Use, Agent Budgets & Public Documentation

This release builds on v4.4.0 with a native-computer-control, agent-safety,
local-retrieval, and documentation pass. It adds an opt-in Computer Use beta
for native Windows and macOS applications, introduces checkpoint-safe work
budgets and configurable child-agent capacity, makes local embedding recall
recover cleanly when a cached model is unavailable or still loading, hardens
browser and live-control isolation, and publishes a comprehensive searchable
user guide with reviewed screenshots and machine-readable references.

Native Computer Use (Beta)

• Native application control - adds a provider-neutral computer_use tool
  for launching allowlisted applications, selecting one target window, reading
  its accessibility tree or screenshot, clicking, typing, pressing keys,
  scrolling, and dragging in native desktop interfaces.
• Browser remains separate - keeps DOM-based browser automation as the
  preferred path for websites and routes native applications and operating
  system dialogs through the new Computer Use engine instead.
• Task-scoped exclusive sessions - gives one interactive local task an
  exclusive Computer Use lease covering discovery, capture, optional Vision
  fallback, and input, preventing overlapping agents from controlling the
  desktop at the same time.
• Generation-bound targeting - uses opaque target and element tokens that
  expire after observations, reconnects, approvals, target drift, and user
  takeover so stale coordinates or accessibility nodes cannot be replayed.
• Live control card - shows the active application, bounded ephemeral
  preview, current state, and direct Stop, Take over, and Resume controls in
  chat while shielding the preview during user or approval handoff.
• Human takeover - cancels queued mutations before handing control to the
  user, retains the paused lease, and requires a fresh target-window
  observation before automation can resume.
• Point-of-risk approvals - classifies routine and consequential desktop
  actions, always confirms foreground escalation and sensitive consequences,
  and hands credentials, OTPs, CAPTCHAs, biometrics, UAC/TCC, terminals,
  password managers, and secure desktops back to the user.
• Vision fallback - can send an ephemeral target-window screenshot to the
  configured Vision provider only when accessibility information is
  insufficient, with the active local or cloud disclosure shown in Settings.
• Interactive-only boundary - keeps Computer Use unavailable to schedules,
  channels, background workflows, child agents, headless/server callers, and
  plugin or general MCP exposure.

Computer Use Setup, Privacy & Recovery

• Off-by-default setup - adds a state-driven Settings flow for enablement,
  installation, health checks, a Calculator verification run, repair,
  reinstall, removal, and an explicitly configured reviewed system binary.
• Reviewed Cua runtime - pins Cua Driver Rust 0.7.1 for Windows x86-64,
  Windows ARM64, and macOS universal, downloads it only after an explicit
  Install action, verifies the platform SHA-256, and extracts it into Row-Bot's
  private data directory without invoking the upstream installer or updater.
• Mandatory telemetry disclosure - requires Continue or Cancel before any
  Cua executable invocation and documents the reviewed upstream telemetry
  fields and endpoint; Row-Bot still adds no first-party telemetry.
• Narrow driver allowlist - exposes only the reviewed application,
  target-window observation, input, health, permission, configuration, and
  session operations while blocking recording, desktop-wide capture, CDP,
  arbitrary configuration, updates, autostart, telemetry mutation, skills,
  process control, and maintenance surfaces.
• Ephemeral sensitive data - excludes typed values from logs, history,
  checkpoints, approval payloads, memory, and durable media, and does not
  persist target-window screenshots.
• macOS permission recovery - detects missing Accessibility and Screen
  Recording access, explains the recovery steps, links directly to the
  relevant Privacy & Security panes, and rechecks readiness after permissions
  are granted or the app is re-added.
• Runtime readiness diagnostics - distinguishes disabled, disclosure,
  install, permission, test, ready, repair, and unsupported states with
  actionable, non-technical recovery guidance.

Agent Runtime Budgets & Delegation

• Checkpoint-safe work budgets - gives every new logical agent turn an
  explicit model-iteration budget, charges one round per successful model
  response, persists progress through interrupts, and derives the LangGraph
  recursion ceiling from remaining capacity.
• Graceful budget completion - reserves an exactly-once, tool-free final
  response when the work-round budget is exhausted instead of surfacing a raw
  recursion error or leaving the run in an ambiguous state.
• Repeated-action protection - detects identical tool requests without
  storing their arguments, blocks the fourth repeat, and terminates continued
  no-progress loops on the fifth request.
• Configurable runtime limits - adds Settings controls for maximum work
  rounds, nested agent levels, active children per parent, active children
  across the app, and an optional child active-time limit.
• Reviewed defaults - starts new runs with 90 work rounds, one child level,
  three active children per parent, eight active children application-wide,
  and no child timeout.
• Queued child capacity - queues excess child agents in order until both
  parent and global capacity are available instead of rejecting them, while
  preserving Stop behavior for queued work.
• Run snapshots and progress - snapshots effective settings into durable
  Agent Run rows, records safe model-iteration counters and heartbeats, and
  keeps active runs on the limits they started with when Settings change.
• Optional active-time timeout - applies the configured child limit only
  while a child is executing, excludes queue time, and records a clear terminal
  timeout reason.

Local Embeddings & Memory Recall Reliability

• Strict cache-only runtime - resolves local embedding models from the
  existing Hugging Face cache and sets local-only loading so normal recall,
  indexing, status checks, and startup cannot trigger a surprise model
  download.
• Explicit download and repair - keeps network access behind the user-run
  local-model download or repair action and reports whether the selected model
  is cached, missing, loading, ready, or failed.
• Shared background loading - coalesces concurrent recall callers onto one
  local embedding load and gives the first recall a bounded grace period rather
  than starting duplicate model loads.
• Fast deterministic fallback - remembers missing, failed, or timed-out
  local-model state so later recall attempts fail fast and continue through
  bounded lexical and graph retrieval instead of stalling the response.
• Visible fallback notice - emits one safe per-generation notice explaining
  when semantic memory search fell back and points to the local model setup
  action without exposing memory content.
• Recall diagnostics - records semantic status, fallback code, wait time,
  retrieval-stage timings, selected counts, and bounded ranking details for
  troubleshooting.
• Workflow-safe retrieval - applies the same bounded fallback to agent
  turns and workflow knowledge lookups while preserving existing provenance,
  ranking, and no-mutation recall policy.

Browser & Interactive Control Hardening

• Thread and task isolation - scopes browser tabs and live-control state to
  their owning thread/task and cleans them up when work finishes or a thread is
  removed.
• Navigation policy - validates browser destinations and redirects before
  mutation so unsafe or out-of-scope navigation cannot silently broaden an
  agent's authority.
• Consequence-aware actions - applies shared approval and consequence
  classification to browser mutations while keeping observation and control
  state explicit.
• History redaction - prevents sensitive typed values and unsafe action
  details from leaking into durable browser history or tool traces.
• Cancellation and readiness - makes browser operations responsive to Stop,
  reports unavailable runtimes clearly, and keeps takeover/resume state
  consistent with Computer Use.
• MCP result isolation - normalizes private Computer Use MCP results without
  registering its driver as a general MCP server or exposing unreviewed tools.

Public Documentation & Website

• Comprehensive user guide - expands the Docusaurus site with concepts,
  request lifecycle, profiles/goals/agents, workflows, integrations, extension
  trust, operations, mobile/native guidance, knowledge provenance and repair,
  wiki vault, settings, and reference pages.
• Control-level Settings reference - generates a searchable inventory of
  visible Settings controls, defaults, declared ranges, dependencies, restart
  notes, security notes, and source locations.
• Reviewed UI screenshot library - refreshes desktop Settings and Home
  images and adds chat, approval, tool trace, profiles, workflows, Developer,
  Designer, Skills Hub, Plugin Marketplace, MCP, and mobile screenshots from an
  isolated neutral demonstration profile.
• Safe capture automation - refuses the normal Row-Bot data directory,
  disables background autostart and network status checks, seeds display-only
  provider/channel/plugin/MCP states, and validates screenshot review metadata.
• Searchable static publication - adds Pagefind search, committed GitHub
  ...

v4.4.0 - Mobile Companion, Channel Streaming & Runtime Reliability

This release builds on v4.3.0 with a mobile-access, messaging, provider, and
runtime-reliability pass. It adds a secure browser-first mobile companion,
brings live agent streaming and interactive approvals to external channels,
surfaces child-agent approvals and completion notices in parent conversations,
makes Stop cancel stalled provider streams and subprocess-backed tools, restores
automatic model discovery with last-known-good catalog protection, hardens
Gemini tool-schema compatibility and Google Calendar concurrency, and expands
the recommended MCP catalog with Xquik.

Mobile Web Companion

• Phone-native shell - adds a full-screen mobile interface with Chat,
  Activity, Workflows, Knowledge, and Settings navigation instead of shrinking
  the desktop drawers and studio chrome onto a phone display.
• Mobile chat workflow - adds conversation list/detail views, new-thread
  creation, attachments, skills, profile and model controls, generation status,
  and a persistent Stop control using the same durable threads as desktop.
• Activity and approvals - shows active chat generations, running workflows,
  recent workflow runs, tool approvals, and workflow approvals in one phone-safe
  surface with approve, deny, and stop actions.
• Mobile workflow editor - supports simple prompt workflows, schedules,
  profiles, model overrides, approval policy, persistent threads, channel
  delivery, enablement, and metadata while preserving advanced graph steps for
  desktop editing.
• Mobile-safe settings - adds stacked provider status and credential
  summaries, local skill enable/pin controls, and installed plugin enablement.
  Skills Hub and Plugin Marketplace installation/configuration remain desktop
  actions in Mobile V1.
• Installable web app - adds a PWA manifest, service worker, install
  metadata, and offline page while deliberately excluding authenticated and
  private Row-Bot surfaces from service-worker caching.
• Mobile Access settings - adds route discovery and phone pairing under
  Settings -> System for local-network access, optional Tailscale direct or
  Serve routes, an existing ngrok tunnel, and advanced custom origins.
• QR pairing and device control - creates short-lived single-use QR links,
  stores session tokens in HttpOnly cookies, lists paired phones and access
  events, and lets an authorized user revoke a phone immediately.
• Remote access gate - protects remote HTTP and WebSocket traffic while
  keeping direct loopback desktop access open, rejecting forwarded-header
  localhost bypasses, and redirecting unpaired remote clients to pairing.
• Local-first auth storage - stores hashed pairing and device secrets,
  device scopes, revocation state, failed-attempt lockouts, and display-safe
  audit events in a dedicated local SQLite database.

Channel Streaming, Approvals & Delivery

• Shared streaming engine - adds deterministic token and tool-event
  delivery with coalesced edits, typing keepalives, bounded previews,
  rate-limit-aware retries, platform-safe splitting, preview cleanup, and final
  send fallback when an edit cannot be completed.
• Discord streaming - streams through edited messages, keeps typing active,
  splits long final answers, retries with a fresh final message when necessary,
  and presents interactive approval buttons.
• Slack native streaming - uses Slack's native stream APIs when available
  and appropriate, falls back to edited messages when they are not, and adds
  Block Kit approval actions plus bounded retry-after handling.
• Telegram draft streaming - uses native message drafts for supported
  private chats, falls back to edit streaming elsewhere, respects UTF-16 message
  limits, and adds inline approve/deny controls.
• WhatsApp and SMS behavior - adds WhatsApp edit streaming, typing, split
  finals, and approval resume. SMS intentionally remains final-text-only with
  safe chunking and explicit YES/NO approval handling.
• Goal Mode and plugin channels - routes normal turns, Goal Mode turns,
  interrupted/resumed turns, and plugin-owned channel turns through the shared
  delivery path without duplicate final responses.
• Checkpoint repair - persists delivered assistant answers when a channel
  checkpoint contains only the human turn, without duplicating an assistant
  message already written by the agent graph.
• Durable terminal notices - delivers compact, once-only child-agent and
  Goal Mode completion/failure notices, retains failed notices for retry, and
  reconciles them when configured channels start again.

Child Agents, Approvals & Generation Control

• Parent-thread approval surfacing - appends durable, deduplicated child
  agent approval requests to the parent conversation so background work cannot
  wait invisibly.
• Clear approval reasons - prefers the model-supplied reason in approval
  cards, redacts and bounds display text, and retains the raw action payload for
  the actual safety decision.
• Shell and Developer rationale - adds an explicit approval-reason field to
  shell and Developer commands so requested command execution can explain why
  it is needed.
• Child-agent lifecycle messages - preserves direct and delegated run
  metadata, completion summaries, queued-turn ordering, checkpoint reloads, and
  once-only rendering across parent and child conversations.
• Shared cancellation scope - links an active generation to provider
  requests, subprocesses, tool waits, and its spawned child runs so Stop has one
  consistent cancellation path.
• Stalled provider cancellation - closes in-flight direct OpenAI,
  Anthropic, xAI, MiniMax, OpenRouter, OpenCode, subscription, and compatible
  transport responses so a blocked network stream no longer leaves a thread
  stuck in Thinking.
• Tool and process cancellation - extends Stop to shell commands, Developer
  processes, MCP probes and calls, browser operations, voice turns, and other
  subprocess-backed work, with a clear stopped result where appropriate.
• Scoped child cancellation - stops only child runs linked to the cancelled
  generation and wakes/detaches queued generation state without terminating
  unrelated agent work.

Providers, Models & Tool Schemas

• Automatic catalog discovery restored - returns Codex and other registered
  providers to targeted and scheduled model refreshes, including discovery of
  newly available subscription models.
• Last-known-good catalogs - preserves provider-specific cached rows when a
  refresh is empty, fails, or loses a later pagination page, and replaces only
  rows for a provider whose refresh completed successfully.
• Visible catalog provenance - reports live, cached, and fallback outcomes
  in provider settings so a retained catalog is distinguishable from a fresh
  provider response.
• Provider-scoped schema policy - adds a compatibility inspector for tool
  input schemas without rewriting the tools used by unaffected transports.
• Gemini array compatibility - checks the effective Google adapter output
  for typed array items, filters optional incompatible tools while preserving
  order, and fails clearly when an explicitly requested tool is incompatible.
• Built-in schema repairs - gives Gmail recipient arrays and Goal
  evidence/blocker arrays concrete item types while continuing to normalize
  legacy scalar or structured inputs safely.
• Provider matrix coverage - expands deterministic and opt-in live-provider
  policy checks for catalog ownership, pagination, routing, current-model
  discovery, media capability, and tool-schema conversion.

Google Calendar & MCP

• Request-scoped Calendar clients - creates an independent Google Calendar
  service for each operation so concurrent tool calls do not share an unsafe
  client instance.
• Safe concurrent OAuth refresh - makes Calendar token refresh single-flight
  and atomic when several operations discover an expired token together.
• Serialized mutations and bulk create - orders same-turn writes and adds a
  native multi-event create operation that preserves result order and reports
  partial failures clearly.
• Retry and reconciliation - retries transient SSL and backend failures with
  fresh services, returns structured permanent errors, and reconciles ambiguous
  timeouts so a backend-committed event is not created twice.
• Typed Calendar operations - preserves search, create, update, move, and
  delete while exposing typed attendee, conference, timezone, notification,
  calendar, and bulk-event inputs.
• Xquik MCP catalog entry - adds Xquik for X/Twitter data access through
  streamable HTTP with API-key header setup, capability metadata, and explicit
  high-risk approval guidance for its generic executor.
• MCP configuration integrity - preserves marketplace metadata when an
  installed server is edited, refreshes cached agent tools after MCP changes,
  normalizes interrupted probes, and avoids repeatedly restarting failed
  servers until an explicit refresh.
• Generated reference refresh - updates public generated references for the
  expanded channel, mobile, MCP, provider, approval, settings, storage, and tool
  inventories.

UI Fixes, Tests & Release Validation

• Active-thread spinner restored - returns the sidebar generation spinner
  to the active thread and limits it to genuinely streaming state while
  preserving pinned-thread and recency ordering.
• Mobile test coverage - adds deterministic subsystem and integration tests
  for pairing, cookies, access gating, remote routes, PWA privacy, Tailscale,
  chat, workflows, settings, and mobile-shell routing.
• **Channel test ...

v4.3.0 - Plugin System v2, Requesty, Prompt Cache & Release Hardening

This release builds on v4.2.0 with a broad extension, provider, workflow, and
release-readiness pass. It completes Plugin System v2, adds plugin-owned
channels and marketplace tooling, introduces Requesty as a first-class
provider, makes prompt context and Anthropic prompt-cache behavior explicit,
moves workflows to profile-first agent execution, adds Developer Studio
worktrees, replaces the macOS tray fallback with a native tray host, removes
the old Thoth rebrand migration from startup, and hardens CI, dependency, and
installer validation before the 4.3.0 release.

Plugin System v2 & Marketplace

• Manifest v2 contract - defines schema_version: 2 plugin manifests with
  supported extension surfaces for native tools, plugin-packaged MCP servers,
  bundled skills, and channels.
• Declarative plugin metadata - adds validation for plugin IDs, versions,
  minimum Row-Bot versions, permissions, settings, secrets, auth declarations,
  health checks, and supported provides entries.
• Plugin API expansion - extends plugins.api with public tool, channel,
  attachment, outbound callback, result, settings, secret, webhook, pairing,
  allowlist, and health-check helpers so plugins can integrate without
  importing Row-Bot internals.
• Native Plugin Center - expands the in-app plugin UI to render metadata,
  permissions, settings, secrets, auth status, health checks, tools, channels,
  bundled skills, logs, marketplace updates, enablement, reload, and uninstall
  controls from Row-Bot-owned UI.
• Marketplace install/update flow - adds cached marketplace indexes,
  checksum-aware local or remote installs, stale-cache fallback, installed
  version/update metadata, disabled-by-default installation, and immediate
  runtime refresh after plugin changes.
• Plugin developer tools - adds plugin linking, reload, doctor, validation,
  local marketplace index generation, devtool helpers, and examples for native
  tools, settings/secrets, MCP-backed tools, and fake channels.
• Plugin templates - adds native-tool, MCP-tool, and channel templates under
  src/row_bot/plugins/templates/ for repeatable plugin authoring.
• Plugin safety sandbox - tightens loader validation around dangerous
  constructs, unsupported extension surfaces, UI framework imports, Row-Bot
  internal imports, old Thoth manifests, stale code, plugin dependencies, and
  runtime refresh behavior.
• Plugin skill scoping - keeps plugin-provided skills tied to plugin
  enablement and Agent Profile tool boundaries, so selected-tool profiles do
  not receive unrelated plugin instructions.
• Plugin MCP bridge - lets plugin-packaged MCP server declarations follow
  plugin enablement and appear or disappear with the owning plugin's runtime
  inventory.

Plugin Channels, Messaging & Bot Auth

• Public plugin channel runtime - adds a core bridge that lets plugin-owned
  channels route inbound messages through Row-Bot's normal channel, agent,
  Goal Mode, approval, media, and generated-file delivery paths.
• Plugin channel attachment handling - reuses the shared channel media
  pipeline for plugin-channel audio transcription, image analysis, document
  extraction, inbox persistence, workspace copy, size limits, and generated
  image/video/file delivery.
• Plugin webhooks - adds namespaced plugin webhook registration under
  /plugin-webhooks/{plugin_id}/{name}, with webhook handlers disabled when
  the owning plugin is disabled, unloaded, uninstalled, or fails load.
• Bot Framework auth helpers - adds plugin-facing helpers for Bot Framework
  JWT validation, OpenID/JWKS discovery, issuer/audience checks, and
  display-safe failure reporting.
• Channel registry integration - teaches the channel registry and tool
  factory how to include plugin-owned channels while preserving native channel
  capability checks and generated send/photo/document tools.
• Thread-scoped channel commands - fixes Telegram model switching and
  related channel commands so model changes stay scoped to the active channel
  thread instead of leaking across conversations.
• Channel output assembly - centralizes channel answer/tool-report assembly
  so plugin and native channel responses have the same readable final shape.
• Channel UI polish - refreshes channel monitor and sidebar status surfaces
  so running native and plugin channels stay visible without crowding the
  conversation list.

Providers, Prompt Context & Workflow Agents

• Requesty provider - adds Requesty as a first-class OpenAI-compatible
  provider with provider definition, setup/settings support, auth mapping,
  catalog URL handling, provider-qualified refs, and routing through the shared
  OpenAI-compatible transport.
• Requesty catalog normalization - maps Requesty's context_window,
  supports_tool_calling, supports_reasoning, supports_vision,
  modality, task, and metadata fields into Row-Bot capability snapshots.
• Requesty surface filtering - filters embedding, audio, image, video,
  moderation, realtime, and other non-chat Requesty rows out of Brain/agent
  surfaces while preserving vision and tool-capable chat models.
• Provider credential dialogs - moves provider credential collection into
  row-specific dialogs so Settings can avoid exposing or confusing unrelated
  provider secret fields.
• Provider selection hardening - improves provider resolution, catalog
  cache behavior, readiness checks, and Quick Choice compatibility for
  provider-qualified refs, Requesty, xAI OAuth, and existing providers.
• Explicit prompt context contract - splits prompt assembly into named
  stable and ephemeral sections so identity, profile, platform, self-knowledge,
  tools, skills, plugins, background overrides, memory recall, date/time,
  Developer, Designer, channel, and history context have deterministic cache
  behavior.
• Anthropic prompt-cache markers - applies Anthropic cache_control
  markers only to eligible stable system content for the direct Anthropic API,
  while keeping conversation history and ephemeral turn data unmarked.
• Prompt cache metrics - normalizes provider prompt-cache read/write token
  counts across common metadata shapes for diagnostics and status reporting.
• Profile-first workflow agents - migrates workflows toward Agent Profile
  execution, maps compatible legacy workflow policies to built-in or generated
  profiles, preserves review/blocked status when policy cannot be mapped, and
  surfaces migration notes in workflow editing.
• Workflow profile defaults - adds a default workflow Agent Profile path so
  new workflows start from explicit profile policy rather than implicit legacy
  skill/tool snapshots.

Developer Studio, UI & Desktop Polish

• Developer worktrees - adds durable per-thread, child-agent, and workflow
  worktree allocation with owner records, branch naming, base branch/commit
  tracking, cleanup state, metadata, and failure preservation.
• Worktree seeding - can seed a Developer worktree from current staged,
  unstaged, and untracked changes or from the last commit, while requiring a
  real Git repository root and safe path handling.
• Developer workspace UI refresh - reorganizes Developer Studio around
  workspace identity, thread selection, profile/run controls, branch controls,
  inspector panels, responsive layout, and safer overflow behavior.
• Recent workspace visibility - fixes Developer Studio so all recent
  workspaces can be shown instead of being hidden by an overly narrow recent
  list.
• Sidebar pinning and thread actions - adds thread pinning, places the pin
  toggle before other row actions, and updates thread action helpers so pinned
  chat and Developer threads sort predictably.
• Natural progress updates - improves long-running interactive status text
  so agent runs can report progress in a more human, less repetitive way while
  preserving durable run state.
• Monochrome sidebar cleanup - refreshes sidebar iconography, status
  badges, action placement, and monitor styling for a cleaner desktop shell.
• Streaming/render reliability - hardens streaming reattach, grouped tool
  trace rendering, final content persistence, generated media handling, and
  disconnected-client timer behavior across chat, Developer, workflows, and
  channels.
• Native macOS tray host - adds a native Objective-C macOS tray helper and
  launcher integration so packaged macOS tray content stays visible and uses
  platform-native status item behavior.
• Startup speed and legacy cleanup - removes the old automatic
  Thoth-to-Row-Bot rebrand migration and post-migration notice from the hot
  startup path, keeping current startup focused on Row-Bot data only.

Docs, CI, Packaging & Supply Chain

• Public docs completion pass - completes the user-guide pass across
  Docusaurus docs, generated reference pages, settings/home/chat/integration
  pages, real UI screenshot metadata, Pagefind integration, and docs validation.
• Plugin documentation - adds the Plugin System v2 technical reference,
  authoring workflow, validation commands, marketplace fixture layout, and
  plugin test guidance.
• Prompt cache documentation - adds the prompt context/cache contract with
  stable/ephemeral section inventory, provider gating rules, and verification
  expectations.
• Installer verification plan - adds the installer and CI verification plan
  covering local pre-push checks, PR gates, manual Installer Verify workflow,
  release artifact builds, and deferred clean-machine smoke coverage.
• Test matrix source of truth - adds scripts/run_test_matrix.py with fast,
  contracts, subsystem, cover...

v4.2.0 - Agent Profiles, Goal Mode, xAI Grok & Public Docs

This release builds on v4.1.0 with a larger orchestration and provider pass. It
introduces durable Agent Profiles, Goal Mode, child-agent delegation, first-class
xAI Grok OAuth support, Grok Imagine media generation, a new public docs site,
real UI documentation capture, and several provider/settings hardening fixes
that make model selection, OAuth status, and headless secret handling more
predictable.

Agent Profiles, Goal Mode & Delegation

• Agent Profiles runtime - adds durable Agent Profiles with profile
  instructions, handoff contracts, usage guidance, tool policy, skill policy,
  context policy, workspace policy, approval policy, and enabled/disabled
  state.
• Thread profile injection - carries the active profile into normal agent
  runs and chat-only turns, including a structured profile prompt, policy
  summary, and warning path when a selected profile is missing or disabled.
• Built-in profile library - adds and refreshes built-in profiles for
  focused roles, including profile metadata, when-to-use guidance, tool
  constraints, and runtime snapshots used by child agents.
• Profile commands and UI - adds profile command handling, a Profile
  Library, profile picker, profile summaries, and profile-selection surfaces so
  users can choose or inspect an agent role without editing config files.
• Goal Mode v1 - introduces durable per-thread goals with objective,
  status, progress, evidence, blockers, next step, turn count, active run id,
  and formatted status output.
• Goal tool integration - adds goal_update and goal_status tools so
  agents can report real progress, blockers, evidence, and completion state
  through the same durable Goal Mode record shown in the UI.
• Child Agent runs - adds durable child-agent runs with queued/running/
  terminal states, parent thread linkage, profile snapshots, context summaries,
  status messages, event logs, stop requests, and wait handling.
• Agent delegation tools - adds delegate_work, agent_status,
  agent_wait, agent_stop, agent_profiles, agent_profile_save,
  agent_message, and agent_promote for controlled multi-agent workflows.
• Agent promotion paths - lets completed child runs be promoted into a new
  Agent Profile or into a disabled manual workflow that can be reviewed before
  reuse.
• Tool allowlists - carries profile/tool allowlists into agent graph
  construction, plugin tools, and MCP tool injection so delegated work can be
  narrower than the parent thread's full tool surface.
• Write-lock and queue safeguards - adds active-run queue handling,
  single-writer protections, and approval coverage so child agents do not
  silently collide with parent work.
• Activity and Goal UI - adds Goal UI, Agent drawer, channel monitor
  surfaces, Command Center grouping, live profile/run status, and streaming
  updates for long-running delegated work.
• Channel goal runtime - extends channel command/runtime paths across
  Discord, Slack, SMS, Telegram, and WhatsApp so channel-driven work can carry
  goal and agent context consistently.

xAI Grok OAuth Provider

• xAI Grok first-class provider - adds xai_oauth as a first-class,
  provider-qualified runtime for xAI Grok subscription access, separate from
  the existing xAI API-key provider.
• OAuth PKCE support - adds OAuth flow helpers, token storage, refresh
  handling, client-id status reporting, account/user/email hash metadata, and
  token health checks for the xAI Grok runtime.
• xAI Responses transport - adds a dedicated xAI OAuth Responses transport
  for chat/model runtime creation, prompt conversion, streaming behavior, and
  response normalization.
• Model catalog integration - wires xAI Grok into cloud model refresh,
  provider-qualified model refs, context-window lookup, model availability
  checks, cache entries, and status-only catalog reads.
• Vision capability probing - adds xAI OAuth vision probe support so
  vision-capable Grok models can be confirmed and reported instead of relying
  only on static assumptions.
• Provider readiness and status - adds runtime availability checks,
  provider status details, token-health timing, last runtime probe, last vision
  probe, model-count status, and OAuth client-id diagnostics.
• Model picker integration - adds xAI Grok provider labels, picker icons,
  inactive reasons, canonical provider refs, surface filtering, and short-lived
  provider-status caching for faster settings and picker rendering.
• Setup and settings integration - adds xAI Grok provider cards, connect
  state, setup-wizard copy, model settings rows, default handling, and
  provider-specific hardening in Settings.
• Default/model-settings hardening - fixes edge cases where xAI OAuth
  defaults, inactive provider state, or unsupported surfaces could leave model
  settings in a misleading state.

Grok Imagine Media Generation

• xAI OAuth media runtime - adds Grok Imagine image and video generation
  through the OAuth-backed xAI Grok provider, including availability checks that
  use OAuth runtime status rather than API-key presence.
• Curated Grok Imagine models - adds grok-imagine-image,
  grok-imagine-image-quality, and grok-imagine-video options for xAI and
  xAI Grok media surfaces.
• xAI media provider module - adds a dedicated xAI media implementation for
  image, video, and image-to-video request construction and response handling.
• Media picker filtering - updates image and video model option builders so
  provider-qualified cache rows and OAuth-backed models appear only on
  compatible media surfaces.
• Image/video tool integration - updates image generation, video
  generation, and Row-Bot Status media reporting so xAI Grok media models are
  discoverable and diagnosable.
• xAI API-key catalog improvements - merges /models and
  /language-models, hides unusable rows, preserves curated chat extras, and
  exposes Grok Imagine media rows without leaking them into chat surfaces.
• Image-to-video request fix - fixes the xAI image-to-video request body so
  media generation requests match the provider's expected payload shape.

Provider Secrets, Settings & Model Picker Reliability

• Session-only secret fallback - hardens provider secret handling when the
  OS keyring is unavailable, especially on WSL/headless Linux, by allowing
  newly entered secrets to work for the current process while keeping local
  metadata files secret-free.
• Headless Linux guidance - updates README and architecture docs to explain
  that persistent provider secrets need Secret Service, KWallet, or another
  secure Python keyring backend.
• Custom endpoint auth errors - improves custom OpenAI-compatible endpoint
  behavior when auth is required but no secret is available, returning a
  provider-specific message instead of silently falling through.
• Provider settings feedback - updates provider settings and setup wizard
  copy so unavailable keyring state, missing secrets, OAuth state, and provider
  readiness are surfaced more clearly.
• Read-only picker path - changes Quick Choice listing to avoid mutating
  provider config during ordinary picker reads, reducing settings churn and
  surprising writes.
• Surface-aware inactive reasons - annotates Quick Choices with clearer
  inactive reasons when a model is configured but unsupported on the requested
  chat, agent, vision, image, or video surface.
• Provider status cache - adds a short-lived picker cache for provider
  status checks so settings/model surfaces do not repeatedly refresh token
  health during one render pass.

Public Docs, Website & Automation

• 4.1.0 landing page refresh - updates the public landing page for the
  published v4.1.0 artifacts, refreshed provider positioning, social preview
  image, Linux install command, accessibility for video facades, and updated
  product demos.
• Public docs site scaffold - adds a Docusaurus-based docs site under
  docs-site with docs navigation, generated reference pages, custom styling,
  static architecture/contact pages, brand assets, favicon, CNAME, and package
  lockfile.
• Docs content metadata - adds structured metadata under docs-content for
  settings tabs, home tabs, dialogs, docs routes, UI surfaces, real UI
  surfaces, screenshots, how-to guides, and review status.
• Real UI screenshots - adds real app screenshot assets for app shell,
  chat, setup, home tabs, and settings tabs so docs show actual Row-Bot
  surfaces instead of placeholder illustrations.
• Docs generation pipeline - adds scripts for inventory collection, MDX
  generation, real UI docs generation, screenshot capture, demo data seeding,
  review report creation, llms.txt generation, schema helpers, and public
  docs validation.
• Docs capture hooks - adds app-side docs capture support so documentation
  scripts can drive and snapshot stable UI states.
• Search and navigation - adds Pagefind search integration, search page,
  generated reference index pages, settings docs, home docs, chat docs,
  integration docs, troubleshooting docs, and skill/plugin/MCP docs.
• Docs CI workflow - adds a GitHub Actions docs workflow and automated
  validation tests so generated docs, metadata, and screenshot references can
  be checked before publishing.

UI, Status & Runtime Polish

• Agent-aware Row-Bot Status - expands Row-Bot Status output to include
  agent/profile/run and media-provider reporting used by diagnostics and
  delegated work.
• Streaming and render updates - updates streaming, render, transcript,
  tool-trace, and chat component paths so profile context, child-a...

v4.1.0 - Providers, Controlled Self-Evolution, Skills & Diagnostics

This release builds on v4.0.1 with a broad provider and runtime reliability
pass. It adds first-class Atlas Cloud support, introduces a Claude Subscription
provider path, lands controlled self-evolution, improves skill activation and
pinning, hardens custom tool creation, and fixes several model-picker,
streaming, voice, vision, setup, and diagnostics regressions that surfaced
after the 4.0.0 rebrand.

Provider Runtime & Model Catalog

• Atlas Cloud first-class provider - adds Atlas Cloud as a native provider
  instead of treating it as a generic custom endpoint, with provider identity,
  setup copy, authentication wiring, runtime routing, and model references that
  behave like the existing first-class providers.
• Atlas Cloud model catalog fetching - adds live Atlas model discovery,
  cache integration, provider-qualified model refs, catalog refresh handling,
  and status/readiness checks so Atlas models appear through the same catalog
  path as other providers.
• Atlas Cloud agent capability classification - maps Atlas-hosted models
  into chat and agent-ready surfaces using the provider's API metadata plus
  curated fallbacks for known frontier/provider families, including OpenAI,
  Anthropic, Gemini, Qwen, Kimi, GLM, MiniMax, DeepSeek, and similar
  tool-capable chat models.
• Atlas Cloud vision capability support - classifies Atlas-hosted
  multimodal chat models as vision-capable where supported, including hosted
  OpenAI, Gemini, Anthropic, Qwen-VL, Kimi-VL, GLM vision, and related model
  families.
• Atlas Cloud media-model filtering - keeps Atlas image-generation and
  video-generation models out of chat, agent, and vision picker surfaces for
  this phase, preventing non-chat media models from leaking into incompatible
  workflows.
• Atlas Cloud streaming fixes - scopes the OpenAI-compatible buffered
  tool-call path so Atlas can stream assistant text after tool calls without
  disturbing existing OpenAI-compatible providers.
• Atlas Claude transport handling - adds Atlas-specific Claude behavior for
  streaming, tool-call replay, and native tool-history cleanup so
  Anthropic-hosted models behind Atlas can complete agent turns reliably.
• OpenAI-compatible transport regression coverage - expands tests around
  streaming, buffered tool output, Claude-shaped tool calls, and
  provider-specific transport behavior to protect OpenRouter and other existing
  compatible endpoints.
• Provider capability resolution - strengthens the shared
  capability-resolution path used by catalogs, readiness checks, vision
  routing, and agent eligibility so provider metadata, curated known-good
  families, and cached model data agree more consistently.
• Provider status and readiness improvements - updates provider readiness,
  runtime selection, status reporting, catalog cache behavior, and auth-store
  integration to support the new providers without changing the behavior of
  existing ones.

Claude Subscription Provider

• Claude Subscription provider support - adds a first-class provider path
  for Claude Subscription usage, with provider registration, auth-state
  detection, model references, runtime selection, and setup/status surfaces.
• Claude Subscription messages transport - adds a dedicated transport for
  Claude Subscription message exchange, including prompt conversion, tool-call
  handling, and response normalization.
• Claude Subscription auth and diagnostics - adds provider
  subscription-auth helpers, external credential handling, readiness checks,
  and tests so the app can report whether the subscription runtime is actually
  available.
• Provider selection integration - wires Claude Subscription into
  provider/model selection without taking over Anthropic API-key behavior or
  other Claude-compatible provider paths.

Controlled Self-Evolution

• Controlled self-evolution engine - introduces the first controlled
  self-evolution runtime, with structured change proposals, reviewable
  execution boundaries, persistence, and test coverage.
• Self-reflection skill updates - adds bundled self-reflection guidance so
  Row-Bot can reason about improvement opportunities through a constrained
  skill flow instead of ad-hoc code changes.
• Dream-cycle integration - connects controlled improvement work into the
  existing dream-cycle and memory-policy systems so reflection output can be
  captured and revisited safely.
• Prompt and agent integration - updates agent and prompt wiring so
  self-improvement behavior is explicit, bounded, and aligned with the rest of
  the assistant runtime.
• Command Center visibility - adds UI/status hooks for evolution state and
  related activity so controlled self-evolution is observable instead of hidden
  background behavior.

Skills, Developer Tools & Custom Tool Builder

• Skill pinning defaults - adds default pinning behavior and activation
  tests so important skills can remain discoverable and stable across sessions.
• Skill activation reliability - improves the skill activation path and
  channel command handling, with coverage for pinned skills, command routing,
  and activation edge cases.
• Custom tool builder hardening - strengthens Git and virtualenv handling
  in the custom tool builder so new tool projects are created more reliably
  across local environments.
• Developer Studio storage and capsules - improves developer storage, tool
  capsule handling, and Developer Studio UI behavior used by the custom tool
  flow.
• Tool-builder guidance updates - refreshes the custom tool builder guide
  to reflect the safer Git/venv workflow and the current implementation.

Chat, Voice & Model Picker Reliability

• Anthropic thinking-block normalization - fixes normalization of
  Anthropic thinking blocks so reasoning content does not corrupt downstream
  transcript handling.
• Local voice talk submission fix - repairs local voice talk submission so
  voice input can be sent through the normal chat path again.
• Ollama vision cache handling - fixes Ollama vision model detection to
  respect cached capability data instead of losing vision support after catalog
  refreshes.
• Migration wizard repair - fixes migration wizard UI and Ollama status
  behavior so setup and upgrade flows do not report misleading provider state.
• Model-picker regression coverage - adds tests around chat-only, vision,
  provider readiness, and model-picker behavior to prevent capability labels
  from drifting again.
• Streaming batcher coverage - expands streaming tests around batched output
  so incremental rendering remains responsive after provider and tool-call
  changes.
• Chat keybinding coverage - adds chat keybinding tests to protect composer
  behavior while provider and streaming internals continue to evolve.

Insights, Status & Diagnostics

• Insights status tray diagnostics - fixes the insights status tray
  diagnostic path so provider and runtime issues are surfaced with more useful
  state.
• macOS tray reliability - restores the packaged macOS launcher to the
  pystray green/grey status dots after the native menu-bar icon path proved
  unreliable on test installs.
• Row-Bot status tool updates - refreshes the Row-Bot status tool and
  guide, including provider/media reporting paths used during diagnostics.
• Provider settings and status UI - updates provider settings, status
  checks, status bar, setup wizard, and related UI state for the new
  provider/runtime readiness model.
• Home and performance stability - improves home-screen performance
  behavior and adds UI performance coverage for the post-rebrand shell.
• Application stability tests - adds broader app-stability hardening tests
  around setup, settings, provider state, streaming, and catalog interactions.

Documentation, Website & Architecture

• Architecture docs refresh - updates the architecture documentation and
  diagrams to match the Row-Bot rebrand and current runtime/provider structure.
• Website download links - updates the docs site download links for the
  v4.0.1 package line.
• Runtime and provider documentation alignment - updates README and docs
  surfaces touched by provider setup, installer guidance, and architecture
  diagrams.

Tests & Release Validation

• Atlas Cloud coverage - adds first-class Atlas tests for model catalog
  fetching, capability classification, auth/setup behavior, OpenAI-compatible
  transport behavior, streaming, tool calls, and vision refs.
• Claude Subscription coverage - adds Claude Subscription auth, transport,
  provider runtime, and subscription-readiness tests.
• Controlled self-evolution coverage - adds controlled self-evolution tests
  around proposal handling, persistence, guardrails, and integration points.
• Skill and command coverage - adds tests for skill pinning, skill
  activation, channel skill commands, and custom tool builder flows.
• Provider runtime coverage - expands provider catalog, runtime, selection,
  readiness, auth-store, API-key storage, and subscription-auth tests.
• UI and workflow coverage - expands tests for migration wizard behavior,
  insights provider status, status media, chat keybindings, streaming batching,
  home performance, and settings/provider contracts.

Breaking Changes And Caveats

• Atlas Cloud requires an Atlas Cloud API key and a successful catalog refresh
  before its live model list can be used.
• Atlas Cloud image-generation and video-generation models are intentionally
  hidden from chat, agent, and vision surfaces in this phase.
• Atlas capability labels depend on provi...

v4.0.1 - Ollama Model Picker Hotfix

This patch fixes a Settings -> Models regression in v4.0.0 where local Ollama
model selections could fail when Row-Bot saved a provider-qualified family ref
such as model:ollama:llama3 but the Ollama daemon exposed the installed model
as a tagged runtime name such as llama3:latest.

• Ollama model switching - local Ollama family aliases now resolve to the
  installed daemon tag when there is one unambiguous match, while explicit tags
  and ambiguous families remain unchanged.
• Provider runtime coverage - the same alias resolution now applies through
  both the legacy model helpers and the provider runtime constructor used by
  chat-only/provider-backed local model paths.
• Regression tests - added focused coverage for unique and ambiguous Ollama
  family aliases, context lookup, provider-qualified picker values, and provider
  runtime construction.

Fixes #178. Thanks to @lihouwenbin for the original PR and investigation.

schema: 1
files:
  Row-Bot-4.0.1-Linux-x86_64.tar.gz: sha256=290750d91934de621d1a18f9814c9ec329a8c39b2f700710c81ea24be5447659
  Row-Bot-4.0.1-Windows-x64.exe: sha256=b61980850d85915a4d8e86dc357302582e6bb4f55a2093ce8ea3ef5ce74cdc9b
  Row-Bot-4.0.1-macOS-arm64.dmg: sha256=27ce614b56bbeac6c08e132db579113152df7451d50d5e325d9e81d06745d6f3

v4.0.0 - Row-Bot Rebrand, Skills Hub, Voice, Providers & Installer Reliability

This is the public Row-Bot rebrand release. It moves the app from Thoth to
Row-Bot across product identity, repository metadata, installers, runtime paths,
release artifacts, documentation, launcher behavior, updater contracts, and
user-data locations. Existing 3.x data is preserved through a copy-first
migration path, so users can upgrade without losing rollback access to their old
Thoth data. Beyond the rebrand, v4.0.0 ships major upgrades to Smart Skills,
Skills Hub, realtime voice, provider discovery, approval modes, thread
organization, packaging, startup reliability, and release validation.

Row-Bot Rebrand

• Product identity - app copy, bundled skills, tool guides, docs, release
  workflows, installer scripts, updater metadata, icons, and public repository
  links now use Row-Bot naming.
• Repository rename support - canonical public repository references now
  target github.com/siddsachar/row-bot.
• Website contract - public site configuration now supports row-bot.ai
  and the Row-Bot repository identity.
• Brand assets - adds Row-Bot glyphs, favicon, installer icon, docs imagery,
  runtime brand helpers, and brand-constant tests.
• Release asset names - v4 artifacts use Row-Bot-X.Y.Z-Windows-x64.exe,
  Row-Bot-X.Y.Z-macOS-{arm64|x86_64}.dmg, and
  Row-Bot-X.Y.Z-Linux-ARCH.tar.gz.
• Linux command rename - Linux installs expose row-bot as the user command
  and use the Row-Bot XDG data tree.
• macOS app rename - macOS packaging now builds Row-Bot.app instead of the
  legacy Thoth app bundle.
• Windows launcher rename - Windows launcher scripts, installer scripts, and
  shortcut entry points now use Row-Bot names.
• Documentation refresh - README, release docs, architecture docs, installer
  docs, issue templates, contributing docs, and release workflows now reflect
  the Row-Bot identity.

Migration From Thoth 3.x

• Copy-first migration - Row-Bot reads legacy Thoth data, copies it into the
  new Row-Bot locations, and leaves the old data intact for rollback or manual
  recovery.
• One-shot migration guard - migration records completion state so normal
  launches do not repeatedly repair already-migrated data.
• Legacy data coverage - provider settings, channels, skills, MCP servers,
  plugins, Buddy assets, Designer workspaces, conversations, memories, tasks,
  media, updater state, and runtime config are covered by migration logic.
• Plugin manifest repair - legacy plugin manifests are repaired during
  migration so old minimum-version metadata does not block migrated plugins.
• Post-migration notice - adds UI and tests for showing users that migration
  completed and where their legacy data remains.
• Compatibility fixtures - migration tests cover copied data, guarded
  reruns, plugin metadata repair, runtime data paths, and Row-Bot brand/runtime
  assets.
• Manual recovery path - interrupted migrations can be retried after backing
  up both old and new data directories.

Source Layout & Runtime Packaging

• Package source layout - runtime code now lives under src/row_bot, with
  compatibility cleanup for the old root-level layout.
• Version module relocation - version metadata now lives under the Row-Bot
  package and release scripts read it from the source-layout path.
• Payload manifest - adds app payload manifest generation and packaging
  compatibility checks.
• Source layout docs - adds docs/SOURCE_LAYOUT.md to document the package
  layout and compatibility expectations.
• Smoke script updates - app smoke and release helper scripts now understand
  the source-layout package.
• Import compatibility coverage - tests cover runtime imports, package
  metadata, installer payloads, and compatibility shims.

Provider Runtime & Model Discovery

• OpenCode providers - adds first-class OpenCode provider support with
  runtime, selection, auth, catalog, and regression coverage.
• MiniMax live discovery - MiniMax models are discovered from the provider
  API instead of requiring hard-coded updates for every new model.
• MiniMax capability mapping - discovered MiniMax models are mapped into the
  model catalog with provider capabilities where available.
• MiniMax stale cleanup - stale MiniMax models can be removed automatically
  when they are no longer returned by the provider API.
• Custom endpoint cleanup - stale custom endpoint model references are
  cleaned up so old provider selections do not linger incorrectly.
• Custom reasoning fixes - custom OpenAI-compatible endpoints handle
  reasoning fields more reliably.
• Custom vision fixes - custom endpoint vision references and capability
  handling are repaired.
• OpenAI-compatible transport coverage - tests expand coverage around custom
  endpoint request shaping, provider refs, model catalog behavior, and
  live-provider discovery.
• Provider settings updates - provider UI and runtime settings now better
  reflect Row-Bot naming and newer provider catalog behavior.

Chat, Attachments & Channels

• Chat attachment bridge - fixes the filesystem bridge used by chat
  attachments so local file references survive the intended handoff path.
• Custom provider channel routing - channel workflows now route
  custom-provider turns through the correct provider/runtime path.
• Channel command support - channel command handling moves into the Row-Bot
  package layout and gains focused coverage.
• Workflow delivery defaults - workflow/channel routing preserves approval
  and provider context across resumed turns.
• Runtime status updates - Row-Bot status tools and channel/runtime
  diagnostics now report Row-Bot naming and effective runtime state more
  consistently.

Smart Skills & Skills Hub

• Smart Skills activation - adds skill activation logic for suggesting,
  enabling, disabling, and applying manual skills in chat context.
• Slash command support - adds slash-command infrastructure and tests for
  skill-aware chat commands.
• Command palette skills - command palette integration can surface skills and
  skill actions more directly.
• Composer skill parity - Designer and Developer chat composers gain access
  to shared skill and slash-command behavior.
• Skills Hub marketplace - adds a Skills Hub for browsing, detecting,
  importing, searching, and installing skills from supported sources.
• Marketplace sources - adds source adapters for GitHub, pasted Markdown, URL
  inputs, well-known skill indexes, and marketplace-style catalogs.
• Import detection - pasted or linked skill content can be detected and
  normalized before installation.
• Search index - Skills Hub includes local search/index helpers for browsing
  available skills.
• Bundled skill updates - bundled skills and tool guides are updated for
  Row-Bot naming and newer runtime behavior.
• Skills tests - adds broad tests for skills activation, Skills Hub sources,
  import detection, search, UI contracts, and slash commands.

Realtime Voice

• Realtime voice overhaul - adds a new realtime voice runtime with provider
  interfaces, coordinator, client contracts, presenter state, and lifecycle
  helpers.
• OpenAI realtime support - adds OpenAI realtime provider/client pieces and
  tests.
• Voice actions - adds structured voice action handling so realtime voice can
  interact with Row-Bot behavior more safely.
• Agent bridge - realtime voice can bridge into agent/runtime behavior
  through a dedicated layer.
• Cue policy - adds conversational cue policy, speech policy, output
  coordination, and realtime event handling.
• Local voice provider support - adds local provider scaffolding for voice
  runtime selection.
• Browser dispatch coverage - tests cover realtime browser dispatch and
  voice event surfaces.
• Voice UI lifecycle - adds UI helpers for voice lifecycle and realtime event
  presentation.

Approval Modes, Threads & Developer UX

• Unified approval modes - approval behavior is consolidated so chat,
  Developer, tools, and workflows can use clearer shared approval semantics.
• Approval gate tooling - adds approval-gate helpers for tool execution.
• Thread rename - conversations can be renamed and thread rename behavior is
  covered by tests.
• Thread actions - adds shared thread-action helpers and tests.
• Developer grouping - Developer/code threads are grouped and restored more
  cleanly from the sidebar.
• Developer workspace state - Developer UI, storage, thread context, and
  workspace contracts are updated for the new package layout and grouping
  behavior.
• Sidebar refinements - sidebar filtering and Developer grouping behavior are
  covered by focused tests.
• Buddy avatar behavior - default assistant avatar handling now respects
  selected Buddy identity more consistently.

Windows Launch, Update & Startup Reliability

• Launcher diagnostics - launcher events now write timing and failure details
  to launcher.log.
• Splash hardening - Tk splash failures are logged, and visible Windows
  console splash fallback is opt-in instead of appearing unexpectedly.
• Window picker hardening - first-run picker behavior is hardened to avoid
  blank-console launch paths.
• Packaged Windows Tk validation - installer build logic now validates
  bundled Tk support in embedded Python.
• Native dependency bundling - Windows Tk smoke checks account for required
  native DLLs, explicit DLL directories, and bundled zlib.
• Ollama startup gating - batch-level Ollama auto-start is gated behind an
  explicit environment var...

v3.23.1 - Custom Endpoint Tool-Calling Hotfix

This hotfix repairs custom OpenAI-compatible endpoint tool calling for local servers such as LM Studio. Streamed tool-call fragments are now assembled before execution, malformed empty-name fragments are dropped, custom endpoint tool turns fall back to non-stream unless streamed tool calling has been explicitly probed, and endpoint status labels now distinguish local custom endpoints from Ollama.

schema: 1
files:
  Thoth-3.23.1-Linux-x86_64.tar.gz: sha256=c78b303ca16767a84da7c8b98ec57b0b28f572c7988f9adf78b2d7c493b05207
  Thoth-3.23.1-macOS-arm64.dmg: sha256=c273cdcd44f5f10525275a5cb8d80ad0dcfa7d11a1f3aa0affe67bbe71485a03
  ThothSetup_3.23.1.exe: sha256=0b465838879c1dea6903848b520aa641c8af3ec76c074853995d76044eadad01

v3.23.0 - Provider Runtime, Memory Recall & UI Performance Hardening

This release hardens the runtime paths that were expanded in v3.22.0. The headline work is provider compatibility: Thoth now preserves provider-qualified model identity end to end, routes incompatible models into a safer chat-only path, probes custom OpenAI-compatible endpoints before trusting tool support, and normalizes tricky provider transcripts before replay. It also ships a major memory recall uplift, with deterministic bounded recall, lexical and graph-expanded candidates, audit metadata, review states, and provenance surfaces. Around that, v3.23.0 makes large transcripts and Settings screens lighter, adds task database recovery, improves local/self-hosted setup, and expands regression coverage around real provider behavior.

Provider Runtime & Custom Endpoints

• Provider-qualified model identity — model choices now keep their provider identity across Settings, catalog pinning, defaults, thread overrides, status displays, setup wizard choices, and runtime construction.
• No accidental OpenRouter fallback — unknown bare model IDs no longer silently route to OpenRouter when the original provider cannot be inferred.
• Runtime readiness routing — provider/runtime checks now distinguish full agent mode, chat-only mode, and blocked configurations before a broken run starts.
• Context-window guardrails — small context windows block agent mode with clearer guidance, while medium windows can use chat-only mode when tool schemas would not fit reliably.
• Unified context policy — local, cloud, and custom endpoint context caps now flow through one policy path with model maximums, user caps, and request-time context parameters where supported.
• Context cache invalidation — changing local or cloud context settings clears stale LLM clients so subsequent turns use the new limits.
• Custom endpoint profiles — OpenAI-compatible endpoints can use profile behavior for common local and proxy servers such as LM Studio, vLLM, llama.cpp, LocalAI, LiteLLM, SGLang, oMLX-style servers, and generic OpenAI-compatible backends.
• Custom endpoint probing — self-hosted endpoints can be probed for catalog availability, streaming support, tool-call behavior, and model compatibility, with probe results persisted for later readiness decisions.
• Native metadata discovery — LM Studio and llama.cpp metadata paths are used when available to discover context windows and native tool support more accurately.
• No-auth endpoint support — local endpoints that do not require API keys can refresh catalogs without unnecessary secret lookups.
• OpenAI-compatible transport — adds a dedicated transport for custom OpenAI-compatible chat, streaming, tool serialization, tool-call chunks, reasoning fields, runtime context overrides, and clearer HTTP error messages.
• Unsupported payload cleanup — custom endpoint profiles can drop unsupported parameters such as tools, tool choice, parallel tool calls, reasoning, response formats, or tool history when a backend cannot accept them.
• Tool-call recovery — local models that emit tool-call envelopes as text or reasoning can be recovered into structured tool calls when safe.
• Reasoning-only response handling — reasoning-only outputs after successful tool calls can be promoted into final visible content, while reasoning-only failures after tool errors produce actionable errors instead of silent empty replies.
• Custom tool validation repair — local/custom providers can receive a repair message when a tool call misses required fields such as query, reducing dead-end schema failures.
• Ollama tool probing — unknown or uncertain local Ollama models can be promoted to agent mode only after a real tool round-trip succeeds.
• Ollama launch cleanup — the launcher now starts Ollama only when saved Brain or Vision settings actually need local Ollama, and --no-ollama forces the skip.
• Ollama reasoning behavior — Ollama reasoning is enabled only for detected reasoning models instead of being forced globally.
• Vision provider refs — local Vision calls strip provider-qualified Ollama refs at the runtime edge while provider/cloud refs still route through the correct path.
• Designer runtime readiness — Designer text refinement and speaker-note generation now use the active model override and verify that the selected model is agent-ready.

Chat-Only Runtime & Transcript Compatibility

• Chat-only runtime path — non-tool or tool-incompatible models can answer normal chat without building the full tool graph.
• Compact chat-only prompt — chat-only mode uses a smaller prompt that avoids implying tools, workflows, or task actions are available.
• Tool-free history shaping — prior tool turns are summarized for chat-only context without replaying full tool bodies or invalid protocol shapes.
• Chat-only streaming persistence — chat-only responses stream and persist through the normal conversation paths.
• Runtime surface tagging — chat, channels, workflow approvals, Designer, and forced agent surfaces now tag their runtime mode so provider readiness can make the right routing decision.
• Provider transcript diagnostics — model-facing transcripts are inspected for invalid tool calls, duplicate tool IDs, orphan tool results, and reasoning-field hazards.
• Transcript normalization — provider-facing messages drop no-op assistant turns, strip invalid tool calls, rewrite duplicate tool-call IDs, drop orphan tool results, and remove unsafe reasoning fields for custom-tool artifacts.
• Thinking retention — non-empty thinking/reasoning text is preserved through streaming, reattach, persisted transcript rendering, and final message display.
• Reasoning-only final guard — reasoning-only chunks are no longer mistaken for final assistant content when there is no visible answer.
• Checkpoint transcript loading — transcript loading can read checkpoint messages and token usage without importing or constructing the agent graph.
• Legacy checkpoint repair — checkpoint version values are normalized when older integer versions are encountered.
• Detached stream finalization — detached clients can finalize with scoped transcript refreshes instead of rebuilding the full main UI.
• Optimistic message preservation — user messages remain visible during detached finalize and reconnect flows.

Memory Recall & Knowledge Audit

• Bounded auto-recall policy — Agent turns now use deterministic memory recall with query building, context-aware token budgeting, scoring, filtering, and trace output.
• Hybrid recall candidates — recall combines semantic search, FTS5 lexical search, keyword fallback, and graph-neighbor expansion.
• Graph-expanded recall — strong seed memories can pull in related graph nodes with relation confidence and hop metadata.
• Recall-safe candidate retrieval — candidate inspection no longer mutates recall timestamps until the final selected memories are injected.
• Recall reinforcement — selected memories are touched with recalled_at and recall-count metadata after they are actually used.
• Memory tier scoring — recall ranks core, semantic, episodic, and resource memories differently based on source, confidence, evidence, recency, and query fit.
• Status-aware filtering — archived, needs-review, superseded, stale, weak, greeting-only, runtime-status, and unanchored resource memories are filtered out of normal auto-recall.
• Recall traces — recent recall decisions are written to a compact trace file for debugging why memories were included or rejected.
• FTS5 memory index — knowledge graph entities now maintain a lexical search index for faster exact/keyword recall.
• Memory evolution helpers — new integrity helpers normalize status, tier, confidence, evidence, source context, manual edits, review state, superseding, archival, and journal entries.
• Memory review states — memories can now be active, needs review, superseded, or archived without losing the underlying entity.
• Audit metadata — extracted, document-derived, wiki-synced, and manually edited memories preserve stronger provenance, confidence, evidence, and source context.
• Conflict handling — extraction can mark conflicting memories for review instead of overwriting high-authority user facts.
• Low-confidence relation filtering — background extraction skips weak inferred relations instead of adding noisy graph edges.
• Extraction journal — memory extraction records run summaries, per-thread details, skipped relations, and extraction outcomes.
• Resource hub memories — document extraction creates or updates resource-style hub memories with provenance and audit fields.
• Wiki sync provenance — wiki vault sync preserves audit/status metadata and appends memory-evolution journal entries.
• Knowledge audit UI — Settings and entity editor surfaces now expose audit badges, filters, review queues, recall traces, and evolution journal entries.
• Entity review actions — individual memories/entities can be archived, marked for review, superseded, restored to active, or marked as user-modified from the editor.
• Memory tool output — memory search/list/save/update output now includes IDs, status, confidence, tier, and recall-aware results so agents can modify the right memory.

UI Performance & Transcript Loading

• UI performance utilities — adds generation tokens, timed UI sections, slow-section logging, and safe UI callback/task wrappers.
• Bounded transcript windows — large conversations render a bounded visible window with an explicit load-earlier path instead of rebuilding every message at once.
• Async model picker cache — model picker options are cached and refreshed asynchronously so chat inp...

v3.22.0 - Developer Studio, Custom Tools, Workflow Delivery & Stability Overhaul

This release turns Thoth into a broader workbench for chat, workflows, code, documents, and user-built tools. The headline feature is Developer Studio: a Codex-style coding workspace for connecting local Git repositories, reviewing code, planning and applying changes, running tests, preparing PRs, and working inside an optional Docker shadow sandbox. It also adds Custom Tools, letting users turn GitHub repos or local folders into reusable Thoth tools through a guided or conversational flow. Around that, v3.22.0 substantially improves workflow delivery defaults, Home status visibility, Settings organization, onboarding, model catalog performance, embedding provider choice, chat tool traces, and app stability diagnostics.

Developer Studio

• Developer workspace surface — adds a new Developer home tab for code workspaces, recent repos, explicit local-folder linking, explicit clone destinations, and code-thread restoration from the sidebar.
• Code threads — Developer conversations are marked as code threads, reopen directly into Developer Studio, keep workspace context, and preserve code-specific state separately from normal chat and Designer threads.
• Repository context injection — Developer turns receive compact, authoritative workspace context including repo path, branch, dirty state, remote URL, top-level files, approval mode, execution mode, and shell guidance, without showing that context in the user message.
• Codex-style approval modes — Developer Studio supports coding approval modes such as read-only, ask before changes, auto edit, and agent run. The mode is changeable at any time and reflected in the Developer Inspector safety policy.
• Developer-native tools — adds workspace-scoped tools for repo info, file listing, file reads, search, git status, branch create/switch, commit, push, fast-forward merge, diffs, todos, detected test commands, shell commands, patch preview/apply, file writes, sandbox imports, and agent-owned change reverts.
• Developer skills and tool guides — adds Developer-focused bundled skills for coding, review, PR prep, and Custom Tools, plus a concise Developer tool guide. These are wired for Developer context instead of bloating normal chat by default.
• Developer todo planning — adds persistent visible todos for coding threads, with status updates surfaced in the inspector so long coding jobs can keep a checkpointed plan.
• Developer Inspector — adds a right-side Developer Inspector with Overview, Safety Policy, Sandbox, Todos, Changes, Files, Agent Changes, Tests, and GitHub/PR sections.
• Live inspector snapshots — the inspector refreshes from debounced background snapshots instead of full UI rebuilds, preserving expanded sections and reducing disconnect/crash risk during long runs.
• Resizable inspector — the Developer Inspector can be widened for diffs, files, and test output without crowding the main chat.
• File tree view — the Files section renders a tree-style repo view instead of a flat label list, making larger repos easier to scan.
• Diff and change review — changed files show added/removed line counts, per-file diffs, and agent-owned change sets.
• Safe revert support — agent-owned edits are recorded and can be reverted when files have not drifted.
• GitHub CLI integration — Developer Studio detects gh from common Windows install paths and gates PR/push operations through Developer approval policy.
• Long coding turn budget — Developer Studio now gets its own recursion/step budget, separate from normal chat and workflows, with Developer-specific wind-down prompts that checkpoint progress instead of failing with a generic tool-loop message.

Docker Sandbox

• Optional Docker execution mode — Developer workspaces can run commands in a Docker shadow copy instead of the real repo folder.
• Persistent sandbox container — Docker Sandbox uses a persistent per-workspace container and shadow workspace, so repeated commands share the same sandbox state until cleaned or rebuilt.
• Import-gated edits — changes made in Docker Sandbox are recorded as pending patches and only affect the real repo after explicit import.
• Network policy — Docker Sandbox can run with network off, ask, or on. Network commands and package installs are blocked early when network is off.
• Sandbox image selection — users can choose the Docker image for a workspace; changing it cleans the current sandbox copy before the next Docker command.
• Sandbox process controls — long-running sandbox processes can be started and stopped through Developer tooling.
• Clear Docker startup errors — stopped Docker Desktop, missing images, and credential-helper failures now produce actionable messages instead of raw pipe/file-not-found errors.
• Local fallback remains available — users who do not want Docker can keep using local execution with the existing Developer approval policy.

Custom Tools

• Custom Tools product surface — Developer home now includes a Custom Tools area separate from code workspaces, with cards for created tools, commands, test output, enablement, promotion, and removal.
• Guided Custom Tool wizard — adds a Source -> Inspect -> Test -> Enable flow for turning a repo URL, local folder, or current workspace into a reusable Thoth tool.
• Conversational Custom Tool Builder — adds one agent-facing custom_tool_builder utility so users can ask Thoth to inspect a repo, draft commands, refine them, create the tool, and promote it without manually writing a manifest.
• LLM-assisted command proposals — Custom Tool creation can use a lightweight model pass to infer useful read-only commands from a repository, with deterministic fallback when AI analysis is unavailable.
• Safety validation — proposed Custom Tool commands are validated for dangerous shell patterns, unreviewed network use, write operations, and missing query placeholders.
• One-time command tests — Custom Tools can be tested before enablement; local/read-only commands can run directly, while network or riskier commands route through the normal approval policy.
• Promotion to normal chat — tested Custom Tools can be promoted into the plugin/tool surface and optionally made available in normal chat through the Utilities toggle.
• Plugin integration — promoted Custom Tools register as synthetic plugin tools, appear in plugin/tool management, and can be disabled or removed safely.
• Source transparency — Custom Tool cards show source URL, local install path, version, command count, availability, and enablement state.
• Terminology cleanup — user-facing UI uses “Custom Tool” instead of the earlier “capsule” wording.

Workflow Delivery & Workflow Console

• Workflow-level delivery defaults — adds a default delivery channel selector for background workflows so new workflows do not default to every channel.
• Multi-channel defaults — default delivery can target multiple configured channels while every workflow still always reports run status to the web app.
• Per-workflow overrides — workflows can inherit the global default or keep a specific override; changing the global default updates only workflows tied to default.
• No extra LLM delivery pass — delivery defaults reuse existing workflow outputs instead of adding an extra model call.
• Delivery UI polish — the workflow delivery control was moved and restyled so it no longer reads as part of the multi-select label.
• Collapsible workflow console — the right workflow console can collapse/expand, persists its state, and works in browser and pywebview.
• Approval attention state — collapsed workflow console shows an attention state when a workflow approval is waiting.
• Workflow console compact badges — collapsed state shows compact badges for running workflows, approvals, and insights while expanded state keeps the normal console layout.
• Recent and upcoming runs — workflow console surfaces running, approvals, upcoming scheduled runs, quick launch, and recent runs in a denser layout.
• Workflow Buddy sync — Buddy state now clears correctly after workflow approval/denial, timeout, stop, cancel, and successful completion.

Home Status & Buddy Reliability

• Expanded Home health bar — Home status now includes compact icon pills for Ollama, active model, cloud API, tunnel, Gmail OAuth, Calendar OAuth, X OAuth, workflows, knowledge, wiki vault, documents, search, skills, tracker, Buddy, MCP, plugins, network, tools, disk, threads DB, FAISS index, Dream Cycle, TTS, and logging.
• Accurate document/vector status — document status pills now use the same indexed-file/vector metadata path as Settings.
• MCP and plugin visibility — Home status now covers MCP and plugin health instead of only older core checks.
• Sleek icon-only pills — status pills use compact icons with hover tooltips, plus amber/red warning indicators for degraded states.
• Background progress inside status bar — document extraction and Buddy generation progress remain inside the Home status area while the icon row stays compact.
• Buddy state machine cleanup — Buddy state transitions are more deterministic around workflow approvals, denials, pending states, and workflow endings.
• Desktop overlay focus behavior — desktop Buddy more reliably appears when the app is minimized or unfocused and hides when the app returns to focus.

Chat, Streaming & Tool Traces

• Shift+Enter newline fix — Shift+Enter now inserts a newline in chat inputs instead of sending the message, matching normal chat app behavior.
• Input-level model picker — the main chat model selector moved into the chat input area to match Designer and reduce top-bar clutter.
• **Cloud/...