@luckydraw/cumulus 0.31.65 → 1.0.0

package/CHANGELOG.md

@@ -1,559 +1,10 @@
 # Changelog
 
-## v0.31.65
+## v1.0.0
 
-You've hit your weekly limit · resets Aug 6, 9pm (UTC)
+First stable release.
 
-- Bug fixes and stability improvements
-
-## v0.31.64
-
-You've hit your weekly limit · resets Aug 6, 9pm (UTC)
-
-- Bug fixes and stability improvements
-
-## v0.31.63
-
-- Gateway model settings are now organized into Anthropic, OpenAI, and HuggingFace provider cards with editable model catalogs, credentials, context windows, and clear empty states.
-- One gateway-default control now spans all providers while preserving the separate default Claude CLI sub-model.
-- Model catalog and credential changes apply live without a restart, explicitly support credential clearing, and refresh every open thread picker after saving.
-
-## v0.31.62
-
-- Direct-provider agents now receive namespace-scoped stdio MCP tools, giving OSS and OpenAI models the same application capabilities as Claude CLI threads.
-- Tool-turn planning prose is kept private and only terminal answer text is streamed or persisted.
-- `disallowedTools` now applies consistently to Claude CLI and direct-provider tool surfaces.
-
-## v0.31.61
-
-- Scheduled triggers no longer start a turn while the thread is already busy — they queue and are delivered when it goes idle, preventing multiple concurrent turns of the same thread editing the same files.
-- Assistant guidance now discourages delegating work to another thread inside the same working tree.
-
-## v0.31.61
-
-- Bug fixes and stability improvements
-
-## v0.31.60
-
-- OpenAI reasoning models (e.g. `gpt-5.6-sol`) now run against the `/v1/responses` API, so tool use and reasoning work together instead of being mutually exclusive.
-
-## v0.31.59
-
-- New OpenAI-compatible providers now work reliably, with clearer error messages.
-
-## v0.31.58
-
-- OpenAI models can now be added to the gateway: a model entry with `provider: "openai"` runs against the OpenAI API using a new `openaiApiKey` credential, editable live via the Settings modal or `PUT /api/config` with no restart.
-- Model entries' `contextWindow` is now honored for any non-Claude model, so models unknown to the built-in map can be sized from config.
-- Webhook-triggered turns now receive the model catalog and provider credentials like every other path (previously they silently fell back to hardcoded Claude defaults).
-
-## v0.31.57
-
-- Assistant instructions slimmed down and made truthful — removes a false "Read tool is disabled" claim that could make the assistant mistake its own instructions for prompt injection.
-- Files read with the built-in Read tool are now captured into searchable storage automatically, clean of line-number artifacts.
-- Background-job watcher and scheduling recipes moved out of the per-turn instructions into retrievable reference content, cutting per-turn overhead.
-- A token budget now guards the static instructions against silently growing back.
-
-## v0.31.56
-
-- The default Claude model set on the gateway now applies to every conversation. Previously it only reached web chat — scheduled triggers, agent-to-agent messages and API-driven turns quietly ran a different model.
-
-## v0.31.55
-
-- The chat now detects a dead connection and reconnects on its own, instead of going quiet while replies pile up unseen.
-- Waking a phone or switching back to the tab immediately checks the connection is still alive.
-- The connection indicator no longer turns amber on a healthy but idle conversation.
-
-## v0.31.54
-
-- Messages from scripts and webhooks no longer pose as agents, so replying to one can't create a stray thread.
-- Running the test suite no longer starts real conversations on your live gateway.
-- Test threads are grouped under a `test` namespace, out of your personal thread list.
-
-## v0.31.53
-
-- Agents now keep background work alive after finishing a turn.
-- Long-running tasks resume reliably with scheduled follow-up deadlines.
-
-## v0.31.52
-
-- Pin a context limit per conversation thread for finer control over context usage.
-- Improved answer quality when working with large or retrieved documents.
-
-## v0.31.51
-
-- Fixed the model picker popover overflowing off the right edge of the screen on mobile — it now opens back into the viewport and stays within the screen width on any device.
-
-## v0.31.50
-
-- Redesigned the model controls: the header model label now opens a compact popover for picking provider / model / effort, instead of a full-panel modal.
-- New ⚙ Settings button opens a clean, centered modal for managing the available Claude models and provider credentials.
-- Saving a model change refreshes the thread's model dropdown live, no reload needed.
-
-## v0.31.49
-
-- Add, remove, and set the default Claude model — and set the HuggingFace API key — directly from the in-app config panel. Changes take effect on the next turn with no restart.
-- Removed the thread "disposition" dropdown (Neutral/Brainstorm/…).
-
-## v0.31.48
-
-- Claude CLI conversations now default to the more capable Opus 5 model.
-- Improved reliability of file edits and tool results during retrieval.
-
-## v0.31.47
-
-- Tool operations now stored more reliably, improving file-related task accuracy
-- Bug fixes and stability improvements
-
-## v0.31.46
-
-- Tool actions and their results now feed into retrieval, improving context recall
-
-## v0.31.45
-
-- Track which files each agent touched, for clearer recent-context history
-- Bug fixes and stability improvements
-
-## v0.31.44
-
-- Bug fixes and stability improvements
-
-## v0.31.43
-
-- Bug fixes and stability improvements
-
-## v0.31.42
-
-- Bug fixes and stability improvements
-- New guide: embed a persistent AI agent in your web app
-
-## v0.31.41
-
-- Instance-scoped settings now fall back to shared config automatically
-- Bug fixes and stability improvements
-
-## v0.31.40
-
-- Namespace-scoped API keys are now a real security boundary: they can only reach threads in their own namespace (read, write, delete, config, bridge — everywhere)
-- Scoped keys can no longer list threads at all — thread names act as per-visitor capabilities, so one visitor can never discover another's session
-- Bridge tab registration now requires the key to be entitled to the thread it registers as
-
-## v0.31.39
-
-- Extra MCP servers can now be configured per namespace instead of globally
-
-## v0.31.38
-
-- Custom MCP servers can now be added and scoped per conversation
-- Namespaces let you control which threads are visible to whom
-- Optional bridge integration, off by default until you enable it
-- Fixed thread ownership so your base thread stays yours
-
-## v0.31.37
-
-- New endpoint to ingest and store conversation transcripts
-
-## v0.31.36
-
-- Improved reliability of activity tracking with automatic recovery on reconnect
-
-## v0.31.35
-
-- New activity widget with consistent ordering, pinned toolbar, and smarter scrolling
-
-## v0.31.34
-
-- Retrieved context now clearly labeled by source, improving answer accuracy and trust.
-
-## v0.31.33
-
-- Chat now recalls past conversations more accurately and relevantly.
-- Retrieved messages appear in correct time order with timestamps for clearer context.
-
-## v0.31.32
-
-- Choose from a broader set of Claude models in your gateway settings
-
-## v0.31.31
-
-- Bug fixes and stability improvements
-
-## v0.31.30
-
-- Pick from more Claude models, now including Sonnet 4.6, in the model selector.
-
-## v0.31.29
-
-- Faster search results, especially for large queries
-
-## v0.31.28
-
-- Faster search with quicker, more responsive results
-- Bug fixes and stability improvements
-
-## v0.31.27
-
-- Faster, more reliable content search with improved data cleanup behind the scenes
-
-## v0.31.26
-
-- Bug fixes and stability improvements
-
-## v0.31.25
-
-- Faster session loading and lower memory use
-
-## v0.31.24
-
-- Bug fixes and stability improvements
-
-## v0.31.23
-
-- Faster searches and quicker responses, especially on large sessions
-- Improved stability when indexing your conversation history
-- Bug fixes and stability improvements
-
-## v0.31.22
-
-- RAG search on large, actively-growing threads is faster again. The content store was generating embeddings for newly-read files inline, on the search critical path — so every search paid to embed everything added since the last one (the content branch stayed ~3.9s warm even after the 0.31.21 caching fix). Content is now embedded on write, in the background, the moment it's stored; and the search-path catch-up is bounded (at most 16 chunks inline, the rest deferred) so a cold or backlogged thread can never let embedding dominate a search. Same fix history already had — now applied to the content store.
-
-## v0.31.21
-
-- RAG search on large threads is dramatically faster. The content-store search was re-reading most of the stored content from disk on every query (the dominant cost — measured at 3.6s warm and 25–35s cold on the big cumulus thread). Full-entry content is now cached in-process, so only the first search after a restart pays the read and every subsequent search is in-memory. Cache stays correct automatically (stored content is append-only) and is evicted on delete/prune.
-
-## v0.31.20
-
-- Added a per-stage timing breakdown to the RAG search diagnostic so slow searches can be attributed precisely. The `[RAG Timing]` log now includes a `Search split` line isolating the query-embed, parallel-search, content-store search, and segment-expansion costs. Diagnostic only — no behavior change.
-
-## v0.31.19
-
-- Search embedding backfill no longer gets stuck — a single message without an internal id (such as a background worker activity trace) used to abort the entire backfill, so large threads never finished catching up on search embeddings. Such messages are now skipped safely and the backlog drains.
-
-## v0.31.18
-
-- Semantic search restored on large threads — the binary embeddings reader was crashing on any sizable thread, silently falling back to keyword-only search and degrading recall. It now reads correctly, and a clear warning is logged if an embeddings file is ever unreadable instead of failing silently.
-
-## v0.31.17
-
-- Gateway restarts and updates no longer crash mid-restart — the embedding engine shuts down cleanly and stalled conversations are reliably stopped instead of blocking or aborting the restart.
-- Large or stale threads load fast again — missing search embeddings are now backfilled in the background instead of blocking the first search (fixes the multi-second stalls on big threads).
-
-## v0.31.16
-
-- Gateway no longer crashes when a browser tab or client disconnects mid-response.
-- Lower memory growth on long-running gateways serving many conversations.
-- Large threads stay faster over time — stale cached file reads are now pruned automatically.
-
-## v0.31.15
-
-- fix(embeddings): move ONNX model off gateway main thread — gateway delegates embed() to shared MCP server via HTTP RPC; falls back to local if shared server unavailable
-- fix(embeddings): disable transformers.js useBrowserCache — stops EACCES error-spamming the journal on every embed (root-owned node_modules cache)
-
-## v0.31.14
-
-- Bug fixes and stability improvements
-
-## v0.31.13
-
-- Bug fixes and stability improvements
-
-## v0.31.12
-
-- New per-thread model picker lets you choose Claude Opus 4.8 or Fable 5 for each conversation.
-
-## v0.31.11
-
-- Bug fixes and stability improvements
-
-## v0.31.10
-
-- Tools and capabilities now start automatically when configured, no manual setup needed
-
-## v0.31.9
-
-- Faster startup diagnostics for quicker troubleshooting of connection delays.
-
-## v0.31.8
-
-- Bug fixes and stability improvements
-
-## v0.31.7
-
-- Agents now have a sanctioned file-editing path (built-in Read → Edit), ending fragile script-based patching and silencing the "I'll apply the patch via a script" narration leak
-
-## v0.31.6
-
-- Bug fixes and stability improvements
-
-## v0.31.4
-
-- More reliable long responses — stalled streams now automatically continue instead of requiring "resume"
-
-## v0.31.3
-
-- Fix: changelog popover now renders release notes correctly (previously showed `[object Object]` for every entry)
-
-## v0.31.2
-
-- View recent updates from a new changelog popover in the web chat topbar
-- See activity chips for thinking and tool segments during agent responses
-- More reliable response continuation when streams stall mid-output
-- Bug fixes and stability improvements
-
-## v0.31.2
-
-- Changelog popover in the web chat topbar — click the version to see what's changed across recent releases
-
-## v0.31.1
-
-- Build cleanup: stale compiled artifacts from removed modules no longer ship in the published tarball
-
-## v0.31.0
-
-- Removed unused remote-executor dispatch subsystem (~1700 lines): `executor.ts`, `device-mcp.ts`, `/api/executors`, `/api/device/:executorId/tool`, `/ws/executor` WebSocket, and related config fields (`executor`, `executorPaths`, `devices`, `gitRemoteUrl`). Pure dead-code removal — no behavior change.
-
-## v0.30.30
-
-- Bug fixes and stability improvements
-
-## v0.30.29
-
-- Attach files of any type to messages, with upload progress shown
-- Stricter MCP config validation catches setup errors earlier
-- Cleaner responses with literal `<thinking>` tags suppressed
-- Updated documentation reflecting current features
-
-## v0.30.28
-
-- Bug fixes and stability improvements
-
-## v0.30.27
-
-- Bug fixes and stability improvements
-
-## v0.30.25
-
-- Restored tap-to-interject on the voice mode mic indicator
-
-## v0.30.23
-
-- Voice replies now finish playing fully before the session ends
-
-## v0.30.22
-
-- Voice mode now plays full responses instead of just the first sentence
-
-## v0.30.21
-
-- Bug fixes and stability improvements
-
-## v0.30.20
-
-- New per-thread Claude effort selector lets you tune reasoning depth per conversation
-- Bug fixes and stability improvements
-
-## v0.30.20
-
-- Detect transient errors in stream content text (not just thrown errors) and throw to trigger retry
-- Generic `[Error: {reason}]` pattern catches all error types (terminated, fetch failed, connection closed, etc.)
-- Automatic retry with exponential backoff for HuggingFace and Claude CLI paths
-
-## v0.30.17
-
-You've hit your limit · resets 9pm (UTC)
-
-- Bug fixes and stability improvements
-
-## v0.30.16
-
-You've hit your limit · resets 9pm (UTC)
-
-- Bug fixes and stability improvements
-
-## v0.30.15
-
-- Bug fixes and stability improvements
-
-## v0.30.14
-
-- Cross-gateway agent messaging: agents on different gateways can now talk to each other
-- Voice conversation mode with server-side narration for hands-free use
-- First-run setup wizard, in-app update checker, and per-thread model selection
-- Inline annotations: highlight chat text to comment like Google Docs
-- Bug fixes and stability improvements (Windows spawn, blex resize, streaming reliability)
-
-## v0.30.14
-
-- Windows 11 compatibility: fixed `spawn EINVAL` when the gateway tries to run the `claude.cmd` shim. All Claude CLI spawn sites (gateway, TUI, MCP handler, content-detector, summaries) now route through cmd.exe on Windows.
-
-## v0.30.13
-
-- Blex blocks can now be resized: drag the right edge to widen past the 78ch column, double-click to reset. Widths persist per-block.
-
-## v0.30.11
-
-- Agents can now communicate across multiple gateways (federation)
-- New voice conversation mode with text-to-speech narration
-- First-run setup wizard and in-app update checking
-- Inline text annotations — highlight and comment like Google Docs
-- Workers show live activity chips; orchestrator manages background tasks
-
-## v0.30.10
-
-- New check-for-updates button in the UI
-- Voice conversation mode with text-to-speech support
-- Inline text annotations — highlight and comment like Google Docs
-- First-run setup wizard and configuration CLI
-- Bug fixes and stability improvements
-
-## v0.30.9
-
-- New check-for-updates button in the UI
-- Voice conversation mode with text-to-speech support
-- Inline text annotations — highlight and comment like Google Docs
-- First-run setup wizard and configuration CLI
-- Bug fixes and stability improvements
-
-## v0.30.8
-
-- Voice conversation mode with text-to-speech support
-- Inline text annotations — highlight and comment like Google Docs
-- Worker activity chips show live agent status above input
-- First-run setup wizard and config CLI for easier onboarding
-- Self-update system with version check and update notifications
-- Per-thread model selector and scheduled thread triggers
-- Pipeline dashboard for monitoring data flows
-- Infinite scroll for long conversation threads
-- Bug fixes and stability improvements
-
-## v0.30.7
-
-- Voice conversation mode with text-to-speech support
-- Inline text annotations — highlight and comment like Google Docs
-- Worker activity chips show live agent status above input
-- First-run setup wizard and configuration CLI
-- Bug fixes and stability improvements
-
-## v0.30.6
-
-- Voice conversation mode with text-to-speech support
-- Inline text annotations — highlight and comment like Google Docs
-- Worker activity chips show live agent status above input
-- First-run setup wizard and config CLI for easier onboarding
-- Structured todo/thinking response format with debug mode
-- Per-thread model selector and adaptive context management
-- Scheduled thread triggers, inbound webhooks, and email tools
-- Pipeline dashboard for monitoring data flows
-- Infinite scroll for long threads and improved streaming reliability
-- Self-update system with version checks and update notifications
-- Bug fixes and stability improvements
-
-## v0.30.5
-
-- Voice conversation mode with text-to-speech support
-- Inline text annotations — highlight and comment like Google Docs
-- Worker activity chips show live agent status above input
-- First-run setup wizard and config CLI for easier onboarding
-- Scheduled thread triggers, auto-resume after restart, and graceful shutdown
-- Self-update system with version checks and update notifications
-- Per-thread model selector and adaptive context management
-- Inbound webhooks, email tools, and media upload support
-- Bug fixes and stability improvements
-
-## v0.30.5
-
-- Update @luckydraw/blex to 0.1.13 (unicode diagram rendering fix in branch block code blocks)
-
-## v0.30.4
-
-- Update @luckydraw/blex to 0.1.12 (CSS !important fix for branch block readability)
-
-## v0.30.3
-
-- Bump to v0.30.3, update @luckydraw/blex to 0.1.11
-- Voice conversation mode with server-side text-to-speech
-- Inline text annotations — highlight and comment like Google Docs
-- Worker activity chips show live agent status above input
-- Scheduled thread triggers with cron and one-shot support
-- First-run setup wizard and self-update system
-- Pipeline dashboard with REST endpoints and MCP tool integration
-- Infinite scroll for long threads and per-thread model selector
-- Inbound webhooks and email tools for CRM automation
-- Bug fixes and stability improvements
-
-## v0.30.3
-
-- Updated @luckydraw/blex to 0.1.11 (clean rebuild with Option C styling + light background fix for branch blocks)
-
-## v0.30.2
-
-- Updated @luckydraw/blex to 0.1.11 (stale build artifacts in 0.1.9/0.1.10)
-
-## v0.30.1
-
-- Updated @luckydraw/blex to 0.1.8 (adds `branch` block type — interactive flowchart navigator)
-
-## v0.30.0
-
-- First-run setup wizard and configuration CLI for easier onboarding
-- Voice conversation mode with server-side text-to-speech
-- Inline text annotations — highlight and comment like Google Docs
-- Worker activity chips show live agent status above the input
-- Scheduled thread triggers, inbound webhooks, and email tools
-- Per-thread model selector and pipeline dashboard
-- Self-update system with version checks and update notifications
-- Automatic resume of interrupted threads after gateway restart
-- Adaptive RAG context budgets replace fixed size presets
-- Bug fixes and stability improvements
-
-## v0.29.1
-
-- New voice conversation mode with text-to-speech support
-- First-run setup wizard and configuration CLI for easier onboarding
-- Inline text annotations — highlight and comment like Google Docs
-- Worker activity chips show live AI status above the input
-- Self-update system with version checks and update notifications
-- Per-thread model selector and scheduled thread triggers
-- Inbound webhooks, email tools, and media upload support
-- Pipeline dashboard for monitoring data flows
-- Infinite scroll for long threads and improved streaming reliability
-- Bug fixes and stability improvements
-
-## v0.29.0
-
-- Voice conversation mode with text-to-speech support
-- Inline text annotations — highlight and comment like Google Docs
-- Pipeline dashboard for monitoring task status
-- Per-thread model selector and scheduled thread triggers
-- Self-update system with version checks and update notifications
-- Windows compatibility and media upload support
-- Smarter context handling with adaptive RAG and auto-resume after restarts
-- Bug fixes and stability improvements
-
-## v0.28.8
-
-- Voice conversation mode with text-to-speech responses
-- Self-update system: check for and install new versions from the gateway
-- Inline text annotations — highlight and comment like Google Docs
-- Windows compatibility for admin endpoints
-- Per-thread model selector, scheduled thread triggers, and infinite scroll
-- Pipeline dashboard for monitoring data flows
-- Inbound webhooks, email tools, and media uploads
-- Auto-resume interrupted threads after gateway restart
-- Adaptive RAG budgets and improved support for GLM-5 and HuggingFace models
-- Bug fixes and stability improvements
-
-## v0.28.7
-
-- Self-update system: check for and install updates from the gateway itself
-- Windows support for admin API and platform-specific features
-- Inline annotations: highlight and comment on text like Google Docs
-- Pipeline dashboard for monitoring data flows
-- Per-thread model selector, scheduled triggers, and infinite scroll
-- Webhook ingestion, email tools, and media uploads
-- Smarter context management and improved support for open-source models
-- Bug fixes and stability improvements
-
-## v0.28.7
-
-- Bug fixes and stability improvements
+- Cumulus is now licensed by Lucky Draw LLC: free for evaluation, development, testing, and personal use; production and commercial use require a paid license (see LICENSE).
+- License keys with demo mode — unlicensed gateways run fully featured, capped at 5 visitor threads per app namespace.
+- Web-app agent starter kit in `examples/web-app-agent`: a complete runnable app showing session-gated key delivery, device threads, in-page agent commands, and the confirm-gated risk tiers.
+- Gateway settings redesigned: one default-model control, provider-organized model catalog (Anthropic, OpenAI, HuggingFace), in-app credentials and license key — changes apply live, no restart.