cowork-os 0.3.92 → 0.3.94

package/CHANGELOG.md

@@ -7,6 +7,77 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ## [Unreleased]
 
+## [0.3.94] - 2026-02-26
+
+### Added
+- **DuckDuckGo free search fallback**: built-in web search provider that requires no API key. Works out of the box by scraping DuckDuckGo's HTML endpoint. Automatically used as a last-resort fallback when paid providers fail or are not configured. The `web_search` tool is now always available — users no longer need to configure a search provider to use web search. Supports web search with region and date range filters. Configured providers still take priority; DuckDuckGo is appended at the end of the fallback chain.
+- **Discord MCP Connector**: full REST API connector for Discord Bot management with 19 tools — guild listing, channel CRUD, message sending with rich embed support, thread creation, role CRUD, reactions, webhooks, and member listing. Includes 429 rate-limit retry (2 attempts, 10s cap), 2000-char message validation, embed schema enforcement (10 embeds max, typed fields), and privileged intent error hints for `list_members` and `get_messages`. Configurable in Settings > Connectors with bot token, application ID, and optional default guild ID.
+- **Mobile Development plugin pack**: new bundled pack with 4 skills (React Native setup, iOS development, Android development, build pipeline) and a mobile-developer agent role. Covers SwiftUI, Jetpack Compose, Fastlane, code signing, and simulator/emulator management.
+- **Game Development plugin pack**: new bundled pack with 4 skills (Unity, Unreal Engine, Godot, cross-engine performance) and a game-developer agent role. Covers C#/MonoBehaviour, C++/Blueprints, GDScript, draw call optimization, LOD, and GPU profiling.
+- **iOS Development skill**: deep standalone skill covering SwiftUI, UIKit, @Observable, SwiftData/Core Data, async/await, push notifications, `xcodebuild`, `xcrun simctl`, code signing, and App Store submission.
+- **Android Development skill**: deep standalone skill covering Jetpack Compose, ViewModel, Room, Retrofit, Hilt, Coroutines/Flow, Gradle builds, ADB, ProGuard/R8, and Play Store submission.
+- **Unity Development skill**: standalone skill for Unity/C# — MonoBehaviour lifecycle, ScriptableObjects, Addressables, object pooling, Shader Graph, UI Toolkit, editor scripting, and Unity CLI batch builds.
+- **Unreal Engine Development skill**: standalone skill for UE5/C++ — Gameplay Framework, Enhanced Input, UCLASS/UPROPERTY/UFUNCTION macros, Niagara, Lumen/Nanite, multiplayer replication, and UnrealBuildTool packaging.
+- **Game Performance Optimization skill**: cross-engine standalone skill covering draw call batching, LOD configuration, occlusion culling, texture optimization, object pooling, memory budgets, and platform-specific tuning (mobile/PC/console).
+- **Terraform Operations skill**: comprehensive IaC skill for `terraform init/plan/apply/destroy`, state management, module development, workspace management, backend configuration, drift detection, and provider patterns.
+- **Kubernetes Operations skill**: comprehensive skill for `kubectl` operations, manifest generation (Deployment, Service, Ingress, ConfigMap, Secret, HPA), Helm charts, RBAC, NetworkPolicy, kustomize overlays, and debugging workflows.
+- **Cloud Migration skill**: migration playbook covering the 6 Rs assessment framework, database migration strategies, network migration, cost estimation, cutover planning, rollback procedures, and multi-cloud patterns.
+- **Docker Compose Operations skill**: skill for Docker Compose v2 commands, Compose file authoring, multi-stage builds, health checks, override patterns, environment management, and production vs dev configurations.
+- **Enhanced DevOps plugin pack (v1.1.0)**: added 4 new inline IaC skills (Terraform plan review, Kubernetes manifest generation, cloud migration assessment, Docker Compose file generation) and expanded keywords for infrastructure-as-code discovery.
+- **Proactive session compaction with structured summaries**: context compaction now triggers proactively at 90% utilization (aligned with Codex CLI) instead of waiting for overflow. Generates comprehensive, Claude-Code-style structured summaries (up to 4096 tokens) covering user messages, work completed, errors/fixes, decisions, pending tasks, and recommended next steps. Uses role-aware transcript formatting (user messages get 3x more room than tool results), Codex-style "handoff to another LLM" framing, model-aware budget scaling (safe for small-context models), and an overflow guard to prevent the summary from re-inflating context. Timeline UI renders summaries as collapsible sections with token-freed stats.
+- **Polymarket prediction market skill**: query odds, trending markets, price momentum, orderbook depth, open interest, trade history, and resolution timelines across Gamma, CLOB, and Data APIs — no authentication required. Includes formatted output with percentages, volume breakdowns, and multi-outcome event support.
+- **Humanizer writing skill**: rewrite AI-generated text to sound natural and human-written. Detects and removes 50+ LLM tells across 7 layers — vocabulary, sentence mechanics, paragraph structure, emotional register, content depth, document architecture, and tone. Includes flagged word lists, 6 tone presets (casual, professional, academic, journalistic, technical, warm), and a systematic 7-step rewriting process.
+- **YouTube video intelligence skill**: fetch transcripts, metadata, chapters, and captions from YouTube videos via yt-dlp or youtube-transcript-api. Supports auto-generated and manual captions, multi-language, translation, audio extraction, playlists, and timestamp deep links. Includes 6 workflow recipes (summarize, search, compare, extract resources, blog conversion, key quotes).
+- **Stock analysis skill**: comprehensive market intelligence across 3 data sources (Yahoo Finance API, yfinance, Alpha Vantage). Covers real-time quotes, 60+ fundamental metrics, 50+ technical indicators, financial statements, earnings, options chains, analyst ratings, institutional holders, dividends, stock screening, and an 8-dimensional scoring framework. Supports stocks, ETFs, indices, crypto, and forex.
+- **Calendly scheduling skill**: manage Calendly via the v2 API with Personal Access Token auth. Covers event types, scheduled events (upcoming/past/cancelled), invitees with custom Q&A and UTM tracking, availability schedules and busy times, cancellation with reasons, no-show management, one-off booking links, webhooks, and organization members. Includes 8 workflow recipes and timezone-aware agenda formatting.
+- **Moltbook agent social network skill**: interact with Moltbook — the social network for AI agents. Post content, reply to discussions, browse feeds (hot/new/top/rising), upvote/downvote, join submolt communities, follow agents, search semantically, and track engagement via the home dashboard. Includes agent registration, verification challenge solving, rate limit awareness, and formatted feed output.
+- **Marketing Strategist skill**: comprehensive marketing strategy across 25 disciplines — positioning, messaging frameworks (StoryBrand, JTBD, Category Design), buyer psychology (Cialdini's 7 principles, 10 cognitive biases), 5 copywriting frameworks (PAS, AIDA, BAB, 4Ps, FAB), SEO (on-page checklist, technical audit, keyword research, link building), landing page CRO (above-the-fold formula, 10-item checklist, A/B testing methodology), paid ads (Google, Meta, LinkedIn with platform-specific guidance), funnel architecture with lead magnet ranking and nurture sequences, analytics and attribution models, pricing psychology, product launch playbook, growth loops and referral design, competitive intelligence templates, and marketing operations. Includes 8 workflow recipes and integrates with existing channel-specific skills.
+- **OpenAI-Compatible as built-in LLM provider**: promoted from custom provider to first-class built-in provider with dedicated settings panel, SSRF-safe base URL validation (blocks private/metadata IPs, DNS rebinding), model list fetching, and automatic migration from legacy custom provider config. Configurable in Settings > AI Model.
+- **Task continuation after turn-limit exhaustion**: failed tasks stopped by the global turn limit now show a "Continue" button in the timeline. Clicking it reconstructs the executor from persisted events, resets budgets, and resumes execution from the last plan checkpoint. Respects concurrency limits — overflow continuations are queued and resume automatically when a slot opens.
+- **Step-level user feedback**: users can now interact with in-progress plan steps via retry, skip, stop, or redirect ("drift") actions. Feedback controls appear in both the timeline view and the bubble view on actively running steps. Skipped steps are marked with a new `skipped` status and the executor advances to the next step.
+- **Tool allow-list for child tasks**: parent tasks can now specify an `allowedTools` list in `agentConfig`. When both parent and child define allow-lists, the child receives the intersection — preventing privilege escalation. An empty intersection throws a clear error at child task creation.
+- **LoreService (shared workspace history)**: new service that auto-records milestones from completed tasks into `.cowork/LORE.md`. Entries are debounced, deduplicated, capped at 40, and written into the `<!-- cowork:auto:lore:start -->` marked section. Rebuilds from recent task history on startup.
+- **VIBES.md workspace kit file**: new `.cowork/VIBES.md` for tracking current workspace energy and mode (e.g., crunch, deep-focus, balanced). Loaded before SOUL.md in agent context so current vibes influence personality interpretation. Auto-updated by agents based on cues.
+- **LORE.md workspace kit file**: new `.cowork/LORE.md` for shared history between user and agent. Milestones section is auto-populated by LoreService. Loaded after MISTAKES.md and before daily logs in agent context.
+- **Low-progress loop detection**: detects agents stuck repeatedly probing the same target with mixed tools (≥6 of last 8 calls on same base target across ≥2 tool categories). Injects a course-correction nudge, then escalates to a final warning if looping persists.
+- **Stop-reason nudge**: detects consecutive `tool_use` stops (≥6) or `max_tokens` stops (≥2) and injects a message telling the agent to wrap up and produce a direct answer.
+- **Turn-budget guard for max_tokens recovery**: max_tokens retry is now skipped when insufficient turns remain in the budget, preventing wasted turns on truncated responses that can't be completed.
+- **MCP server tools in context panel**: the active context data now includes tool names from connected MCP servers, giving the UI visibility into what tools each connector provides.
+- **Citation Engine**: per-task citation tracker that intercepts `web_search` and `web_fetch` results, deduplicates URLs, assigns sequential [N] indices, and injects formatted citation lists into the LLM system prompt. Citations appear inline in agent responses and are displayed in a dedicated Citation Panel UI with URL, title, domain, snippet, and access timestamp.
+- **Scratchpad Tools**: session-scoped note-taking system for agents during long-running tasks. `scratchpad_write` stores key-value notes (max 100-char keys, 10,000-char values); `scratchpad_read` retrieves all or specific notes. Persists to `.cowork/scratchpad-{taskId}.json` for crash recovery.
+- **Workflow Pipeline**: multi-phase task execution for complex workflows. The Workflow Decomposer detects multi-step prompts (connectives like "then", "after that", "next", "finally") and splits them into sequential phases (research, create, deliver, analyze, general). Each phase creates a child task with output piped to the next phase. Includes LLM-powered fallback decomposition for complex prompts.
+- **Deep Work Mode**: extended execution mode for complex tasks with longer timeouts, progress journaling visible in the task timeline, and automatic memory compression pause during active execution to avoid context disruption.
+- **Document Generation Tools**: three LLM-callable tools (`generate_document` → PDF, `generate_presentation` → PPTX, `generate_spreadsheet` → XLSX) registered as native agent tools with artifact registration and MIME type metadata.
+- **Event Trigger Service**: condition-based automation engine that fires actions (`create_task`, `send_message`, `wake_agent`) in response to channel gateway messages, cron events, or webhooks. Supports AND condition logic, configurable cooldowns (default 1 min), event variable substitution, and history tracking (last 50 fires per trigger). Configure in Settings > Event Triggers.
+- **File Hub Service**: unified file aggregation combining local workspace files, task artifacts, and cloud storage into a single searchable interface. Includes filename-based search, recent files tracking, and MIME type detection for 20+ formats.
+- **Web Access Server**: serves CoWork OS as a web application over HTTP/WebSocket with bearer token authentication (timing-safe comparison), CORS origin whitelisting, REST endpoints mapped to IPC channels, static file serving with SPA fallback, and a health check endpoint.
+- **OAuth connectors (Google, DocuSign, Outreach, Slack)**: enterprise OAuth 2.0 authentication flows for Google Workspace (Calendar, Drive, Gmail with PKCE and scope mapping), DocuSign (signature scope), Outreach (sales intelligence scopes), and Slack (team-domain support). Uses local callback server on port 18765 with CSRF state validation.
+- **Vision tools improvements**: SHA1-keyed result cache (128 entries) preventing redundant vision API calls, automatic image downscaling for images >2MB (1600×1200 at 80% quality), multi-provider fallback chain (OpenAI → Anthropic → Gemini → Bedrock), retry logic for transient errors (429, 5xx, timeouts), and PDF page conversion via `pdftoppm` at 72 DPI.
+- **5 Financial plugin packs**: Equity Research (earnings analysis, sector analysis, coverage initiation, price target, catalyst tracking), Financial Analysis (DCF modeling, ratio analysis, financial statement analysis, peer benchmarking, valuation summary), Investment Banking (deal screening, pitch book, M&A analysis, due diligence, comps analysis), Private Equity (deal sourcing, LBO modeling, portfolio monitoring, exit analysis, fund reporting), Wealth Management (portfolio construction, asset allocation, client reporting, risk assessment, tax optimization). Each pack includes a dedicated agent role and 5 "Try Asking" prompts.
+- **7 financial standalone skills**: DCF Valuation, Earnings Analyzer, ESG Scorer, Financial Modeling, Market Screener, Portfolio Optimizer, Risk Analyzer, Tax Optimizer.
+- **Symlink escape detection**: file tools now resolve symlinks via `realpath()` and verify the resolved path remains within the workspace boundary before any read/write operation. Symlinks pointing outside the workspace are rejected.
+- **Memory compression pause/resume**: the MemoryService pauses background compression during active task execution (especially deep work mode) to avoid context disruption, and resumes automatically when the task completes.
+- **Bedrock provider maxTokens pre-clamping**: the Bedrock LLM provider now pre-clamps `maxTokens` to model-specific limits before sending requests, preventing API rejections. Includes regex fix for error message parsing and `_callId` injection for log correlation.
+- **Browser navigation deduplication**: `browser_navigate` tool now normalizes URLs to detect and skip duplicate navigations to the same page.
+
+### Changed
+- **Task queue limits raised**: default concurrent tasks increased from 5 → 8, max configurable ceiling raised from 10 → 20, default timeout increased from 30 → 60 minutes. Existing users who never changed their settings are automatically upgraded to the new defaults.
+- **Sub-agent safety cap**: sub-agents that bypass the normal concurrency limit are now capped at 40 total running tasks (2× max configurable) to prevent runaway resource consumption.
+- **Resumed tasks respect concurrency**: tasks resuming after app restart now respect the concurrency limit instead of all starting simultaneously. Overflow tasks are re-queued at the front and start as slots open.
+- **Memory kit skill updated**: VIBES.md and LORE.md templates added to the memory-kit skill. Bootstrap guide updated with steps for reviewing vibes and lore files.
+- **HeartbeatService singleton**: HeartbeatService now exposes `getHeartbeatService()` / `setHeartbeatService()` for global access without passing references through the dependency chain.
+- **PlanStep status**: plan steps now support a `skipped` status in addition to pending, in_progress, completed, and failed.
+- **Codebase formatting**: applied oxfmt across ~40 renderer components, IPC handlers, extensions, tray modules, and utility files for consistent line wrapping and indentation.
+
+### Fixed
+- **Focused mode tabs**: Mission Control now correctly included in focused mode tab list.
+- **Bedrock inference profile fallback**: improved profile scoring to prefer same-family models and prevent silent downgrade to a different model family. MaxTokens is now clamped to model-specific limits on the retry path.
+- **Browser HTTP status errors**: browser tools returning HTTP 4xx/5xx no longer immediately circuit-break the tool. These URL-specific failures are now treated as input-dependent with a higher threshold before disabling.
+- **Skipped-tool-only turn detection**: executor now detects agents stuck emitting only policy-blocked tool calls with no text output; injects nudge after first turn, force-stops after 2 consecutive blocked-only turns.
+- **Follow-up tool call lock placement**: relocated `followUpToolCallsLocked` from `executeStep` to `executeGoalStep` where the loop detection state is maintained, fixing misplaced lock reference.
+
+## [0.3.91] - 2026-02-24
+
 ### Added
 - **Agent-initiated memory (`memory_save` tool)**: agents can now explicitly save observations, decisions, insights, and errors to the workspace memory database during task execution. Memories are persisted across sessions and recalled via hybrid search in future tasks. Respects workspace memory settings, privacy modes, and sensitive data filtering.
 - **Enhanced `search_memories`**: now searches both the memory database AND `.cowork/` workspace markdown files (MEMORY.md, daily logs, project contexts, etc.). Results are merged, deduplicated, and ranked by relevance. Response includes `source` ("db" or "markdown") and file `path` for markdown hits.
@@ -23,9 +94,84 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 - **Weekly Usage Insights dashboard**: new Settings > Usage Insights panel showing task metrics, cost/token tracking by model, activity heatmap by day-of-week and hour, and top skills usage. Supports 7/14/30-day period selection.
 - **Starter mission templates**: expanded from 3 onboarding suggestions to 10 one-click missions with categories. Displayed in onboarding and on the empty-state welcome screen.
 - **Competitive research skills**: new `competitive-research.json` and `idea-validation.json` skill files for market analysis, competitor scanning, and MVP scoping.
-- **Adaptive complexity (power user toggle)**: three-tier UI density (`focused | standard | power`) controlling which settings tabs and features are visible. Focused mode hides advanced settings; power mode shows everything.
 - **Plain-language settings labels**: renamed jargon in Settings sidebar (MCP Servers → Connected Tools, Guardrails → Safety Limits, LLM Provider → AI Model, Control Plane → Remote Access, SkillHub → Skill Store). Technical names preserved in tooltips.
 - **Native Infrastructure tools**: built-in cloud sandbox (E2B), domain registration (Namecheap), crypto wallet (USDC on Base), and x402 payment protocol support — all registered as native agent tools with no external MCP subprocess. Configurable in Settings > Infrastructure.
+- **Dispatched Agents Progress Panel**: when a task mentions agents (e.g. `@Security Analyst`), the parent task's main window now shows a Collab-mode-style progress panel with agent chips, phase indicator (Dispatched → Working → Complete), and a real-time event stream from all child agent tasks. Click any agent chip to navigate to its full task view.
+- **`userPrompt` field on Task**: child tasks dispatched to agents now store the original user prompt separately, so the UI displays the user's actual request instead of the internal agent-dispatch formatting.
+- **Digital Twin (Persona Templates)**: create role-specific AI digital twins from 10 pre-built templates across 5 categories (Engineering, Management, Product, Data & Analytics, Operations). Each template includes a tailored system prompt, capabilities, proactive heartbeat tasks, cognitive offload categories, and recommended skills. Templates: Software Engineer, Hardware Engineer, QA/Test Engineer, DevOps/SRE, Technical Writer, Engineering Manager, Technical Director, VP Engineering, Product Manager, and Data Scientist. Activation flow: browse gallery → customize name/heartbeat/tasks → create twin as a new AgentRole with background proactive tasks. Accessible via "Add Digital Twin" button in Mission Control agents panel.
+- **Knowledge graph system**: FTS5 search and graph traversal for structured knowledge representation.
+- **Twitter/X content writer skill**.
+- **Proactive suggestions service** with edge case handling.
+
+### Changed
+- **Execution strategy orchestration** in daemon lifecycle: strategy derivation at task creation, runtime strategy re-application for queued/legacy tasks, relationship-memory outcome recording on top-level task completion.
+- **Friendlier error messages, mode hints, and settings grouping** across the UI.
+
+### Fixed
+- **Task cancellation now cascades to child tasks**: stopping a parent task also cancels all dispatched agent sub-tasks instead of leaving them running.
+- **TypeScript errors in scraping tools** resolved.
+- **Wrap-up skips synthesis phase** due to premature phase transition.
+- **Task cancellation spinner** stays active after cancel (fixed).
+- **Collapsed sidebar header padding** adjusted for new session button.
+
+## [0.3.90] - 2026-02-23
+
+### Added
+- **Collaborative mode**: multi-agent collaborative thoughts and capability matching.
+- **Multi-LLM orchestration**: config validation and comparison service for running multiple LLM providers.
+- **Git worktree manager**: worktree isolation for parallel agent branches with comparison service.
+- **Git tools**: new agent tools for git operations.
+- **Task pinning**: pin important tasks for quick access.
+- **Anthropic streaming**: streaming support with progress callbacks for Anthropic provider.
+- **Bedrock as vision provider** for `analyze_image` tool.
+- **Crypto trading and email marketing skill definitions**.
+
+### Changed
+- **Action-first planning**: executor enhanced with resourcefulness and companion identity.
+- **Context summarization events** surfaced in task timeline.
+- **Image attachment support** for task creation and follow-ups.
+- **Bedrock display names**: improved formatting, model ID parsing, and AP region regex fix.
+
+## [0.3.89] - 2026-02-20
+
+### Added
+- **Declarative plugin system**: connectors, skills, and agent roles defined via JSON manifests.
+- **Built-in plugin packs**: pre-configured plugin bundles for common workflows.
+- **Slash command autocomplete** in input areas.
+- **`create_spreadsheet` and `create_presentation` tool support** for agents.
+- **Inline spreadsheet preview** in task view.
+
+### Changed
+- **Sonnet 4.6 model**: added as new model option and updated as Bedrock default.
+- **Oxfmt formatter** and promoted Oxlint rules to errors.
+- **Conway Terminal**: auto-provision API key and inject into MCP server env.
+- **Model selector redesign** and defensive electronAPI guards.
+
+### Fixed
+- **Bedrock transcript mismatches**: repaired and preserved tool-call pairs during compaction.
+- **Follow-up error handling**: prevent pinned message insertion between tool pairs.
+- **Electron sandbox**: disabled for preload and hardened attachment validation.
+
+## [0.3.88] - 2026-02-19
+
+### Added
+- **Vision/image support** across all LLM providers.
+- **LOOM email protocol** support alongside IMAP/SMTP.
+- **Conway Terminal integration** with payment safety caps.
+- **Adaptive complexity (power user toggle)**: three-tier UI density (`focused | standard | power`) controlling which settings tabs and features are visible. Focused mode hides advanced settings; power mode shows everything.
+- **Validation schemas and config sanitization** for security hardening.
+
+### Changed
+- **Executor enhanced** with image attachments and tool approval flow.
+- **Email settings, toast positioning, XLSX viewer, and styling** improvements.
+- **Contextual welcome cards** in the UI.
+
+### Fixed
+- **Electron env variable handling** and strict dev port configuration.
+
+## [0.3.87] - 2026-02-17
+
+### Added
 - **Intent-routed task strategy runtime**: tasks are now classified into `chat`, `advice`, `planning`, `execution`, or `mixed` and receive derived strategy defaults (conversation mode, turns, quality passes, answer-first flags).
 - **Relationship memory lifecycle APIs**: added relationship memory and commitment IPC/preload surfaces:
   - `memory:relationshipList`
@@ -37,23 +183,25 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 - **Dedicated relationship-agent architecture + UAT docs**:
   - `docs/relationship-agent-architecture.md`
   - `docs/relationship-agent-uat.md`
-
-- **Dispatched Agents Progress Panel**: when a task mentions agents (e.g. `@Security Analyst`), the parent task's main window now shows a Collab-mode-style progress panel with agent chips, phase indicator (Dispatched → Working → Complete), and a real-time event stream from all child agent tasks. Click any agent chip to navigate to its full task view.
-- **`userPrompt` field on Task**: child tasks dispatched to agents now store the original user prompt separately, so the UI displays the user's actual request instead of the internal agent-dispatch formatting.
-- **Digital Twin (Persona Templates)**: create role-specific AI digital twins from 10 pre-built templates across 5 categories (Engineering, Management, Product, Data & Analytics, Operations). Each template includes a tailored system prompt, capabilities, proactive heartbeat tasks, cognitive offload categories, and recommended skills. Templates: Software Engineer, Hardware Engineer, QA/Test Engineer, DevOps/SRE, Technical Writer, Engineering Manager, Technical Director, VP Engineering, Product Manager, and Data Scientist. Activation flow: browse gallery → customize name/heartbeat/tasks → create twin as a new AgentRole with background proactive tasks. Accessible via "Add Digital Twin" button in Mission Control agents panel.
+- **XLSX file extraction** and auto-generate missing hooks token.
+- **Attachment chips** in chat bubbles and collapsible toggle fix.
+- **Use-case skill templates** and Pi finder/librarian skills.
+- **Resend email connector** and expanded gateway router commands.
 
 ### Changed
-- **Execution strategy orchestration** in daemon lifecycle:
-  - strategy derivation at task creation
-  - runtime strategy re-application for queued/legacy tasks
-  - relationship-memory outcome recording on top-level task completion
+- **Execution strategy orchestration** in daemon lifecycle: strategy derivation at task creation, runtime strategy re-application for queued/legacy tasks, relationship-memory outcome recording on top-level task completion.
 - **Memory context composition** now combines `UserProfileService` facts with layered relationship memory context for prompt injection.
 - **Soft-deadline behavior** in executor now switches from deep step execution to best-effort finalization before hard timeout.
+- **Approval modal replaced** with centered notification toasts and session approve-all.
+- **MCP connectors** kept disabled until configured.
 
 ### Fixed
-- **Task cancellation now cascades to child tasks**: stopping a parent task also cancels all dispatched agent sub-tasks instead of leaving them running.
 - **Timeout-abort completion gap**: timeout-triggered cancellation paths now attempt best-effort finalization instead of exiting without a user-facing answer.
 - **Cancellation observability**: executor cancellation logs now include cancellation reason (`user`, `timeout`, `shutdown`, etc.) for clearer diagnosis.
+- **Stuck 'executing' status**: prevented with varied failure loop detection and improved artifact handling.
+- **Slack socket stability** and WhatsApp connection flap detection.
+- **Max iterations** increased for complex agent operations.
+- **User name rules** strengthened to prevent OS username leakage.
 
 ## [0.3.86] - 2026-02-14
 
@@ -655,6 +803,12 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 | Version | Date | Highlights |
 |---------|------|------------|
+| 0.3.91 | 2026-02-24 | Digital twins, web scraping, AI playbook, build mode, knowledge graph, usage insights, infrastructure tools |
+| 0.3.90 | 2026-02-23 | Collaborative mode, multi-LLM orchestration, git worktree isolation, Anthropic streaming |
+| 0.3.89 | 2026-02-20 | Declarative plugin system, built-in packs, Sonnet 4.6, slash command autocomplete |
+| 0.3.88 | 2026-02-19 | Vision support across all LLM providers, LOOM email, Conway Terminal, density mode |
+| 0.3.87 | 2026-02-17 | Intent-routed strategy, relationship memory, timeout recovery, Resend connector |
+| 0.3.86 | 2026-02-14 | ACP canvas endpoints, i18n, talk mode, skill registry |
 | 0.3.84 | 2026-02-14 | Fixes CI installability check module resolution so release validation passes and desktop packaging can continue |
 | 0.3.83 | 2026-02-14 | Fixes release workflow shell parsing so installability validation and desktop asset publishing complete successfully |
 | 0.3.82 | 2026-02-14 | Removes script-enabled recovery installs that triggered electron-winstaller SIGKILL and hardens runtime repair install flags |
@@ -676,7 +830,13 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 | 0.1.0 | 2025-01-24 | First public release with core features |
 | 0.0.1 | 2025-01-20 | Initial development setup |
 
-[Unreleased]: https://github.com/CoWork-OS/CoWork-OS/compare/v0.3.84...HEAD
+[Unreleased]: https://github.com/CoWork-OS/CoWork-OS/compare/v0.3.91...HEAD
+[0.3.91]: https://github.com/CoWork-OS/CoWork-OS/releases/tag/v0.3.91
+[0.3.90]: https://github.com/CoWork-OS/CoWork-OS/releases/tag/v0.3.90
+[0.3.89]: https://github.com/CoWork-OS/CoWork-OS/releases/tag/v0.3.89
+[0.3.88]: https://github.com/CoWork-OS/CoWork-OS/releases/tag/v0.3.88
+[0.3.87]: https://github.com/CoWork-OS/CoWork-OS/releases/tag/v0.3.87
+[0.3.86]: https://github.com/CoWork-OS/CoWork-OS/releases/tag/v0.3.86
 [0.3.84]: https://github.com/CoWork-OS/CoWork-OS/releases/tag/v0.3.84
 [0.3.83]: https://github.com/CoWork-OS/CoWork-OS/releases/tag/v0.3.83
 [0.3.82]: https://github.com/CoWork-OS/CoWork-OS/releases/tag/v0.3.82

package/README.md

@@ -7,8 +7,8 @@
 </p>
 
 <p align="center">
-  <strong>The most complete open-source AI assistant platform</strong><br>
-  Security-hardened, local-first runtime — 30+ LLM providers, 14 channels, 100+ skills
+  <strong>Looking for an OpenClaw alternative? CoWork OS is a local-first option for production workflows.</strong><br>
+  Security-hardened, local-first AI operating system — 30+ LLM providers, 14 channels, 139 skills
 </p>
 
 <p align="center">
@@ -20,6 +20,8 @@
 
 <p align="center">
   <a href="docs/getting-started.md">Getting Started</a> &middot;
+  <a href="docs/showcase.md">Use Cases</a> &middot;
+  <a href="docs/openclaw-comparison.md">OpenClaw Alternative Guide</a> &middot;
   <a href="docs/">Documentation</a> &middot;
   <a href="CHANGELOG.md">Changelog</a> &middot;
   <a href="SECURITY.md">Security</a> &middot;
@@ -34,17 +36,33 @@
 
 - **30+ LLM providers** — Anthropic, OpenAI, Google, Ollama, AWS Bedrock, OpenRouter, and more. Bring your own keys.
 - **14 messaging channels** — WhatsApp, Telegram, Discord, Slack, Teams, iMessage, Signal, and more. Chat with your AI from anywhere.
-- **100+ built-in skills** — Documents, code review, web search, image generation, cloud integrations, and more.
+- **139 built-in skills** — Documents, code review, web search, image generation, cloud integrations, game development, mobile development, financial analysis, infrastructure-as-code, and more.
 - **Digital Twin Personas** — Pre-built AI twins for every role (engineer, manager, PM, director). Each twin absorbs cognitively draining work — PR triage, meeting prep, status reports, dependency tracking — so the human stays in flow.
+- **Plugin Platform** — 17 bundled role-specific packs (including Mobile Development, Game Development, and 5 Financial packs) with 55+ skills, in-app Plugin Store for installing community packs, remote registry, and enterprise admin policies for organization-wide control.
+- **Active Context sidebar** — Always-visible panel showing connected MCP connectors with branded Lucide icons and enabled skills, auto-refreshed every 30 seconds.
 - **Agent teams** — Multi-agent collaboration with shared checklists, collaborative mode, multi-LLM synthesis, and persistent teams.
 - **Think With Me mode** — Socratic brainstorming that helps you clarify thinking without executing actions.
 - **Build Mode** — Go from idea to working prototype with a phased canvas workflow (Concept → Plan → Scaffold → Iterate) and named checkpoints.
 - **AI Playbook** — Auto-captures what worked from successful tasks and injects relevant patterns into future prompts.
-- **Usage Insights** — Dashboard showing task stats, cost/token tracking by model, activity heatmaps, and top skills.
+- **Usage Insights** — Dashboard showing task stats, cost/token tracking by model, activity heatmaps, top skills, and per-pack analytics.
 - **ChatGPT History Import** — Import your full ChatGPT conversation history. CoWork OS instantly knows your preferences, past projects, and context — no cold start. All data stays encrypted on your Mac and never leaves your machine.
 - **Security-first** — Approval workflows, sandboxed execution, configurable guardrails, encrypted storage, and 3200+ tests.
 - **Local-first & BYOK** — Your data and API keys stay on your machine. No telemetry. No middleman.
 
+### Looking for an OpenClaw Alternative?
+
+OpenClaw has a large and passionate community for agent experimentation. CoWork OS is a respectful alternative for teams and builders who want local-first operations with approvals, guardrails, and reliable multi-channel execution.
+
+| Positioning Lens | OpenClaw | CoWork OS |
+|---|---|---|
+| Category | Agent toolkit / experimentation layer | Security-hardened AI operating system |
+| Primary use case | Build and iterate on agent workflows | Run agent workflows in production |
+| Security posture | Varies by workflow and setup | Built-in approvals, guardrails, sandbox isolation, encrypted local storage |
+| Runtime surfaces | CLI-centric control paths | Desktop app + headless daemon + 14-channel gateway |
+| Data ownership model | BYOK/self-hosting workflows | Local-first, BYOK, no telemetry, optional offline Ollama |
+
+See the full guide: [OpenClaw alternative: CoWork OS vs OpenClaw](docs/openclaw-comparison.md)
+
 ## Quick Start
 
 ### Download the App
@@ -120,24 +138,45 @@ Advanced web scraping powered by [Scrapling](https://github.com/D4Vinci/Scraplin
 ### Integrations
 
 - **Cloud Storage**: Notion, Box, OneDrive, Google Workspace, Dropbox, SharePoint
-- **Enterprise Connectors**: Salesforce, Jira, HubSpot, Zendesk, ServiceNow, Linear, Asana, Okta, Resend
+- **Enterprise Connectors**: Salesforce, Jira, HubSpot, Zendesk, ServiceNow, Linear, Asana, Okta, Discord, Slack, Resend, Google Workspace, DocuSign, Outreach
 - **Developer Tools**: Claude Code-style `glob`/`grep`/`edit_file`, Playwright browser automation, MCP client/host/registry
 
 [Learn more](docs/features.md)
 
+### Active Context Sidebar
+
+Real-time overview of your active integrations, always visible in the right panel. Shows connected MCP connectors with branded Lucide icons (HubSpot, Salesforce, Slack, GitHub, Postgres, and 30+ more) and green status dots, plus enabled skills from active packs. Each section shows 4 items with internal scrolling for more. Auto-refreshes every 30 seconds. [Learn more](docs/plugin-packs.md#context-panel)
+
 ### Usage Insights
 
-Dashboard with task metrics, cost/token tracking by model, activity heatmaps (day-of-week and hourly), and top skills usage with 7/14/30-day period selection. Access from **Settings** > **Usage Insights**. [Learn more](docs/features.md#usage-insights)
+Dashboard with task metrics, cost/token tracking by model, activity heatmaps (day-of-week and hourly), top skills usage, and per-pack analytics (skill usage grouped by plugin pack) with 7/14/30-day period selection. Access from **Settings** > **Usage Insights**. [Learn more](docs/features.md#usage-insights)
 
 ### LLM Providers
 
 12 built-in providers + 20+ compatible/gateway providers. Use cloud APIs or run fully offline with Ollama. [Learn more](docs/providers.md)
 
+### Plugin Platform & Customize
+
+Unified plugin platform with 17 bundled role-specific packs (Engineering, DevOps, Product, Sales, QA, Finance, and more), each bundling skills, agent roles, connectors, and "Try asking" prompts. Packs can link to Digital Twin personas for proactive background work.
+
+- **Search & filter**: Real-time sidebar search across pack names, descriptions, categories, and skill names
+- **Per-skill control**: Enable or disable individual skills within a pack without toggling the whole pack
+- **Persistent toggles**: Pack and skill states survive app restarts
+- **Update detection**: Background version checks against the registry with visual indicators
+- **"Try asking" in chat**: Empty chat shows randomized prompt suggestions from active packs
+- **Plugin Store**: In-app marketplace for browsing, installing (Git/URL), and scaffolding custom packs
+- **Remote Registry**: Community pack catalog with search and category filtering
+- **Admin Policies**: Organization-level controls — allow/block/require packs, restrict installations, set agent limits, distribute org-managed packs from a shared directory
+- **Per-pack analytics**: Usage Insights dashboard groups skill usage by parent pack
+
+Access from **Settings** > **Customize**. [Learn more](docs/plugin-packs.md)
+
 ### Extensibility
 
-- **100+ built-in skills** across developer, productivity, communication, documents, and more
+- **139 built-in skills** across developer, productivity, communication, documents, game development, mobile development, financial analysis, infrastructure-as-code, and more
 - **Custom skills** in `~/Library/Application Support/cowork-os/skills/`
-- **Declarative plugins** with 5 built-in plugin packs
+- **17 bundled plugin packs** with 55+ role-specific skills and Digital Twin integration
+- **Plugin Store** — browse, install from Git/URL, or scaffold custom packs
 - **MCP support** — client, host, and registry
 
 ### Voice Mode
@@ -150,7 +189,7 @@ Built-in structured entity and relationship memory backed by SQLite. The agent b
 
 ### Memory & Context
 
-Persistent memory with privacy protection, FTS5 search, LLM compression, and workspace kit (`.cowork/`) for durable project context. **Import your ChatGPT history** to eliminate the cold-start problem — CoWork OS knows you from day one. All imported data is stored locally and encrypted on your Mac. [Learn more](docs/features.md#persistent-memory-system)
+Persistent memory with privacy protection, FTS5 search, LLM compression, and workspace kit (`.cowork/`) for durable project context. **Import your ChatGPT history** to eliminate the cold-start problem — CoWork OS knows you from day one. All imported data is stored locally and encrypted on your Mac. **Proactive session compaction** automatically generates comprehensive structured summaries when context reaches 90% capacity — preserving user messages, decisions, file changes, errors, and pending work so the agent continues seamlessly without losing critical context. [Learn more](docs/features.md#persistent-memory-system) | [Context Compaction](docs/context-compaction.md)
 
 <p align="center">
   <img src="screenshots/cowork-os-main3.png" alt="Collaborative Mode" width="700">
@@ -243,9 +282,12 @@ See [CHANGELOG.md](CHANGELOG.md) for the full history of completed features.
 | Guide | Description |
 |-------|-------------|
 | [Getting Started](docs/getting-started.md) | First-time setup and usage |
+| [Use Case Showcase](docs/showcase.md) | Comprehensive guide to what you can build and automate |
 | [Features](docs/features.md) | Complete feature reference |
+| [OpenClaw Alternative Guide](docs/openclaw-comparison.md) | Respectful category framing for users evaluating an OpenClaw alternative |
 | [Channels](docs/channels.md) | Messaging channel setup (14 channels) |
 | [Providers](docs/providers.md) | LLM provider configuration |
+| [Migration from OpenClaw](docs/migration.md#from-openclaw-to-cowork-os) | Practical checklist for teams moving from OpenClaw to CoWork OS |
 | [Development](docs/development.md) | Build from source, project structure |
 | [Architecture](docs/architecture.md) | Technical architecture deep-dive |
 | [Security Guide](docs/security-guide.md) | Security model and best practices |
@@ -254,7 +296,10 @@ See [CHANGELOG.md](CHANGELOG.md) for the full history of completed features.
 | [VPS/Linux](docs/vps-linux.md) | Headless server deployment |
 | [Remote Access](docs/remote-access.md) | Tailscale, SSH tunnels, WebSocket API |
 | [Knowledge Graph](docs/knowledge-graph.md) | Structured entity/relationship memory |
+| [Context Compaction](docs/context-compaction.md) | Proactive session compaction with structured summaries |
 | [Mission Control](docs/mission-control.md) | Agent orchestration dashboard |
+| [Plugin Packs](docs/plugin-packs.md) | Plugin platform, Customize panel, and Plugin Store |
+| [Admin Policies](docs/admin-policies.md) | Enterprise admin policies and organization pack management |
 | [Digital Twins](docs/digital-twins.md) | Role-based AI twin personas and cognitive offload |
 | [Digital Twins Guide](docs/digital-twin-personas-guide.md) | Comprehensive guide with scenarios and expanded job areas |
 | [Troubleshooting](docs/troubleshooting.md) | Common issues and fixes |

package/connectors/discord-mcp/README.md

@@ -0,0 +1,64 @@
+# Discord MCP Connector
+
+This connector exposes Discord Bot APIs to CoWork OS through MCP tools.
+
+## Requirements
+
+Provide credentials via environment variables:
+
+- `DISCORD_BOT_TOKEN` (required) — Bot token from the Discord Developer Portal
+- `DISCORD_APPLICATION_ID` (optional) — Application ID for the bot
+- `DISCORD_GUILD_ID` (optional) — Default guild ID; tools that require a guild will use this if `guild_id` is not passed per-request
+
+### Bot Setup
+
+1. Go to https://discord.com/developers/applications
+2. Create an application (or use an existing one)
+3. Navigate to **Bot** and click **Reset Token** to get `DISCORD_BOT_TOKEN`
+4. Under **Privileged Gateway Intents**, enable:
+   - **Server Members Intent** — required for `discord.list_members`
+   - **Message Content Intent** — required to read message content via `discord.get_messages`
+5. Navigate to **OAuth2 > URL Generator**, select scopes `bot` and `applications.commands`, then select the permissions your bot needs (Manage Channels, Manage Roles, Send Messages, etc.)
+6. Use the generated URL to invite the bot to your server
+
+## Build & Run
+
+```bash
+npm install
+npm run build
+npm start
+```
+
+## Add to CoWork MCP Settings
+
+- **Command**: `node`
+- **Args**: `/absolute/path/to/connectors/discord-mcp/dist/index.js`
+- **Env**: set the variables above
+
+## Tools
+
+- `discord.health` — Verify bot token and connection
+- `discord.list_guilds` — List guilds the bot is in
+- `discord.get_guild` — Get guild details (members, channels count)
+- `discord.list_channels` — List all channels in a guild
+- `discord.get_channel` — Get detailed information about a channel
+- `discord.create_channel` — Create text, voice, category, forum, or stage channels
+- `discord.edit_channel` — Update channel name, topic, etc.
+- `discord.delete_channel` — Delete a channel (irreversible)
+- `discord.send_message` — Send a message with optional embeds
+- `discord.get_messages` — Fetch recent messages from a channel
+- `discord.create_thread` — Create a thread (from a message or standalone)
+- `discord.list_roles` — List all roles in a guild
+- `discord.create_role` — Create a new role with color, hoist, mentionable
+- `discord.edit_role` — Edit an existing role (name, color, hoist, mentionable)
+- `discord.delete_role` — Delete a role (irreversible)
+- `discord.add_reaction` — React to a message with an emoji
+- `discord.create_webhook` — Create a webhook for a channel
+- `discord.list_webhooks` — List all webhooks for a channel
+- `discord.list_members` — List guild members (requires Server Members Intent)
+
+## Rate Limiting
+
+The connector automatically retries on HTTP 429 (rate limit) responses, waiting for the `retry_after` duration returned by Discord. Retries are capped at 2 attempts with a max delay of 10 seconds per retry.
+
+See `docs/enterprise-connectors.md` for the contract.