UPDATES

A new sandboxed tool server (dieah-mcp) enforces uniform tool restrictions across all 6 CLI providers. Agents are classified into ToolSet tiers — Full, ReadOnly, ReadWrite, or None — controlling access to file reads, writes, shell commands, and web fetches. Every tool call is logged to telemetry. Bundled as a Tauri sidecar for zero-config deployment.

Peer-to-peer TCP mesh connects multiple dieahOS instances across machines. Pre-shared key authentication with SHA-256, persistent peer table, automatic keepalive, static peer connector with exponential backoff. New CLI commands: dieah peers, dieah mesh status, dieah mesh key.

Full multi-bot support for Telegram. Each bot gets its own config, polling loop, offset tracking, and cancellation token. Add, edit, start, stop, and delete bots from a redesigned Channels page. Daemon hot-reloads bot configs via telegram.reload IPC — close the app, your bots keep running.

Agents no longer discuss or speculate about application internals. A system-level instruction across all 5 prompt assembly points ensures agents stay in character and focus on the user's work.

New dieah-mcp crate joins the workspace alongside dieah-core, dieah_os, dieah-daemon, and dieah-cli. MCP proxy is auto-built on cargo tauri dev and bundled as a signed sidecar in release builds.

core/mod.rs split from 7,267 lines into 4 focused modules: mod.rs (4,760), commands.rs (973), streaming.rs (1,508), process.rs (61). Zero behavior changes — pure structural cleanup.

Agents were silently building inside the dieah-os directory when a user-specified project folder didn't exist yet. Now creates the directory first instead of falling back.

Plan and workspace confirmation dialogs now auto-proceed after 15 seconds (matching the existing pattern). Backend has a 20-second timeout safety net — no more infinite hangs if the user doesn't click.

All agent context loading (memory, journal, soul, household) was using global vault search — first match wins. Multi-team setups could cross-contaminate. Added 10 team-scoped _for(team_id) variants threaded through ~20 callsites.

HQ agent roster displayed the full personality paragraph instead of the functional role title. Fixed 4 spots. Also fixed team dropdown not updating the active agent — switching teams previously left the previous team's orchestrator in the chat.

Board showed all teams' boards, streaming indicators fired for wrong teams, chat threads shared by bare agent name causing collisions. Added composite teamId:agentName keys, per-team navigation state, streaming team tags, board filtering. Store migration v2.

Scheduler extracted to dieah-core (~800 lines) for daemon reuse. Daemon now runs both the scheduler loop and Telegram auto-start on startup. Close the app — scheduled jobs and Telegram bots keep running.

New bots added via the app survive app close. The app notifies the daemon via telegram.reload RPC whenever bot configs change. Daemon hot-reloads: stops removed/disabled bots, starts new/enabled ones.

New headless daemon binary runs your AI team 24/7 without the desktop app. Scheduled jobs fire, Telegram bot stays online, agents remain reachable. dieahd start/stop/status for lifecycle management. dieahd install creates a launchd (macOS) or systemd (Linux) service so your team survives reboots.

New dieah CLI binary connects to the daemon via IPC. Query team status, list agents, check errors, dump settings — all from the terminal. JSON-RPC 2.0 over Unix domain socket.

Talk to your agents from your phone. Configure a bot token, bind chats to specific agents via /agent command, and get Markdown-formatted responses. Chat ID whitelist for security. Auto-starts on app launch if previously enabled. Works in both daemon and embedded modes.

Periodic orchestrator check-ins that assess team health and flag issues. Configurable intervals with active hours, timezone support, and check rotation. Board-aware skipping so heartbeats never interrupt active delegations. Status bar indicator (HB OK / ALERT / OVERDUE / PAUSED) and HQ alert cards with Pulse Now button.

Visual graph of agent-to-agent delegation patterns with node sizing by activity and edge weights by interaction frequency. Filterable interaction ledger with timestamps, duration, and token usage. Aggregate stats dashboard.

Toggleable PATH shim system that intercepts verbose CLI tool output (git, cargo, cat) and compresses it before agents see it. One-click toggle in Settings. Transparent to agents — real binaries still accessible.

Desktop app auto-detects a running daemon on startup. If found, connects via IPC and skips starting its own scheduler and Telegram poller. If not found, falls back to embedded mode. Status bar shows DAEMON or EMBEDDED.

Backend split from a single Rust package into four crates: dieah-core (Tauri-free shared engine), dieah-app (desktop GUI), dieah-daemon (headless binary), dieah-cli (terminal client). Core library has zero GUI dependencies — runs on desktop, server, or container.

All 65 event emission sites in the 6,948-line orchestration engine converted from Tauri-specific to platform-agnostic. Three implementations: Tauri WebView, daemon IPC broadcast, and no-op for tests.

Fixed camelCase/snake_case deserialization mismatch that silently dropped all settings changes. Frontend sent camelCase but Rust expected snake_case. Added serde rename_all with alias support for backward-compatible disk reads.

Killed native OS select styling platform-wide. All dropdowns now render with flat tactical appearance, custom sharp SVG chevron, JetBrains Mono font, and left-accent focus state.

Agents now default to Online on app launch instead of showing Stopped. Auto-registers all agents from vault at startup. Renamed Start/Stop buttons to Online/Offline to reflect the stateless CLI nature of agents. Dashboard agent count now correct immediately.

Version bump to 0.2.2. Housekeeping release with no user-facing changes.

Skills marketplace Browse tab now shows an inline preview of skill content before installing. CI fix for pnpm workspace configuration.

Resolved 19 tsc build errors including unused imports, missing icon components, and incorrect marked library types.

Complete frontend rewrite from Dioxus WASM to React 19 + Vite + Zustand + Tailwind v4. Every page rebuilt with real IPC wiring, 60+ typed invoke wrappers, and Tauri event subscriptions. Dramatically improved responsiveness and developer velocity.

Full CSS variable system with brass (#d6b76e), seafoam (#669c94), charcoal (#16191a) palette. Zero border-radius everywhere, clip-cyber polygon utilities on buttons/cards/nav, dot grid background, film grain overlay, Rajdhani headings + JetBrains Mono body. Router redraw animations on page transitions.

First-run wizard auto-detects installed CLI providers (Claude, Codex, Gemini, Ollama) from PATH. Walks users through provider configuration with status indicators. Settings → Providers tab for ongoing management.

Mermaid diagrams render as interactive SVGs inline in chat and Board. Expand-to-browser opens full-viewport HTML with trackpad pinch-to-zoom and two-finger pan. Board auto-generates dependency diagrams. Timeline and report rendering for structured agent output.

Single source of truth at ~/Obsidian/. Auto-migration from legacy ~/Documents/Dieah/ path at startup. Each vault has daily/, context/, skills/, and MEMORY.md. Skills install writes .md notes to vault/skills/.

Expandable agent cards with model selector dropdown (available Claude CLI models), role selector, and personality textarea. Per-agent model persists in YAML frontmatter. Sub-agents inherit parent model if none set.

Dedicated scheduler UI with job CRUD, cron expression editor, run history browser, and enable/disable toggle per job. Visual next-run countdown and last-run status.

Fixed critical event name mismatches: backend emits agent:stream but frontend listened on stream (chat streaming completely broken), error:new vs error:log. Fixed React 19 Strict Mode race condition causing orphaned Tauri listeners and double HQ messages. Added message dedup by ID.

New "Errors" tab in the sidebar with real-time backend error capture. Errors from delegation, planning, sub-agents, and HQ are automatically logged with timestamps and source badges. Red badge on sidebar nav shows error count. "Copy All" for sharing with developers, "Clear" to reset.

New "Import" tab in the Skills Marketplace with three import methods: paste a GitHub URL pointing to a SKILL.md file, browse local .md files via file picker, or paste raw SKILL.md content directly. Imported skills appear alongside marketplace installs and can be toggled per-agent.

Model dropdown on every agent card. Choose from available Claude CLI models or use the CLI default. Selection persists in the agent's definition.md YAML frontmatter and survives restarts.

Reply to a specific agent in HQ without manually typing @AgentName. Hover any message to reveal a "Reply" button; clicking it auto-prepends the @mention with a reply preview above the input. Escape to dismiss.

Planning output is now validated before execution: agent names checked against the delegate roster, phase numbering verified, classification validated. If the LLM produces malformed JSON or invalid plans, errors are fed back for self-correction (up to 2 retries). Unknown agents are auto-fixed with fuzzy matching as a final fallback.

dieahOS now checks for updates on launch via the Tauri 2 updater plugin. A banner appears at the top of the app with the new version number and an "Update & Restart" button. One click to download, install, and relaunch.

Fixed 5 bugs preventing dieahOS from running on Windows: PATH splitting used : instead of ;, binary search didn't probe PATHEXT extensions, fallback used /bin/sh, expanded_path() used Unix-only separator, and .cmd/.bat files caused "not valid Win32 application" errors.

Orchestrator and HQ conversations now reuse Claude CLI sessions across messages via --resume instead of re-sending full history every turn. Session invalidation on agent delete and skill toggle.

CLI-style accumulated status log during delegation, replacing the old single-line spinner. Five status events with timing, token counts, and pulsing progress indicators.

Fixed critical bug where sub-agent CLI invocations used plain-text mode instead of stream-json, causing all output to be silently discarded and every sub-agent to report "failed."

Agent cards show elapsed time next to status labels and reflect current activity: "Starting...", "Thinking...", "Using tool...", "Processing..." — inferred from the agent's stream in real-time.

Skills with audits from skills.sh show colored risk badges (Safe/Low/Medium/High/Critical) with hover breakdown from Gen Agent Trust Hub, Socket, and Snyk.

Create user-defined agent roles with custom behavioral templates, archetypes, delegate descriptions, and write-capability flags. Role selector shows built-in + custom roles with inline creation.

Create teams instantly from pre-built or custom fleet templates without calling Claude CLI. Built-in "General Software Development" fleet with 9 roles. Custom fleets via 3-tab builder with import/export YAML support.

Skills tool installer now supports macOS (brew), Linux (apt-get with brew fallback), and Windows (winget with scoop fallback).

Agents can now use Gemini CLI, Codex CLI, and Ollama in addition to Claude. Per-agent or per-team provider override with auto-detection from PATH.

Browse, install, and per-agent toggle of 600+ community skills from the skills.sh ecosystem. Three-tab UI with search, security audit integration, and 8K per-skill injection cap.

Full cron implementation with preset/custom intervals. Sub-agents stream output in real-time. Planner uses high-effort mode with parallel execution by default. @build fast path, @think/@plan/@review chat commands.

Run multiple teams simultaneously, each with own orchestrator, vault, and agents. Personality profile and mission statement injected into all agent prompts.

Each team gets its own vault with full directory structure. Per-team HQ conversations with @mention autocomplete. Task classification, tiered briefings, dispensation system, git worktree isolation.

Chat UX with Shift+Enter multiline, sub-agent delegation with Board-first planning, CLI subprocess engine spawning claude/codex/gemini as child processes, and 3-step agent creation wizard.