v0.0.12: Agentic Search & Wiki, standalone MCP server, live token streaming

A quick tour

Agentic Wiki — Overview	Agentic Search — Results	MCP Server — Claude Code
Agentic Wiki — Knowledge Graph	Agentic Search — Capability Graph	Console Settings — schema-driven

Highlights

• Agentic Search — graph-first, cited answers over saved multi-source workspaces, with shareable runs, live token streaming for true TTFT, and a sturdier, faster engine. Self-facing agentic_search tool lets task-spawned agents search too; every tier's "Auto" model defaults to GPT OSS 120B.
• Agentic Wiki — cited-source Q&A engine with an abstract-entity multiplex over the tree-sitter code graph, worktree create-from-base, and resilient checkpoint-aware indexing/recovery.
• Standalone MCP server — the wiki/search substrate extracted into the mewbo_graph capability library; structured-outputs endpoint and 7 new MCP tools.
• Console & engine — rich attachments, worktree-aware sessions, DeepWiki-backed Ask AI modal, Iconify icon system, and a schema-driven Settings page (faceted UI generated from the config schema, not a bespoke form).
• Build & CI hardening — pinned @playwright/cli in the base image (killing latest drift that broke the build), deterministic wiki-job recovery ordering, and a host-guarded agent-pickup workflow.

Container images (multi-arch: linux/amd64 and linux/arm64)

• ghcr.io/bearlike/mewbo-base:0.0.12
• ghcr.io/bearlike/mewbo-console:0.0.12
• ghcr.io/bearlike/mewbo-api:0.0.12

Full Changelog: v0.0.11...v0.0.12

v0.0.11: Widget builder, capability gating, & command palette

• Widget builder — built-in plugin lets agents generate interactive Streamlit widgets rendered inline in the chat timeline via stlite (browser Wasm). Includes a component library: PlantUML diagrams, GitHub repo cards, stock tickers, and search cards.
• Capability gating — sessions declare capabilities; agents, skills, and plugins are hidden and unavailable when requirements aren't met.
• Command palette — web console gains Cmd+K command palette and /commands slash-command system.
• Live sub-agent visibility — CheckAgentsCard and SpawnAgentCard render inline in the chat timeline.
• Tool search — fuzzy-search available tools by name or description.

What's Changed

• ✨ feat: widget-builder builtin plugin with stlite console rendering, capability gating, and SessionTool protocol by @bearlike

Full Changelog: v0.0.10...v0.0.11

v0.0.10: LSP, plugin system, web IDE, & conversation fork

• LSP tool — language server integration (Python/TS/JS/Go/Rust) with diagnostics, go-to-definition, references, and hover. Auto-discovered per session; passive diagnostics injected after every file edit.
• Plugin system — install/uninstall plugins from local manifests or a marketplace. Plugins contribute agent definitions, skills, hooks, and MCP tools. Managed via CLI (/plugins), API, and console.
• Conversation fork/edit — branch from any past message with edit-and-regenerate, plus per-message model override.
• Web IDE — opt-in per-session VS Code (code-server) containers accessible from the console. Requires agent.web_ide.enabled.
• Context compaction — accounts for tool schema and system prompt overhead before compacting; emits context_compacted events.
• Console — ReviewPane with last-turn / uncommitted / branch diff scopes; searchable ConfigMenu and updated session header layout.

What's Changed

• ✨ feat: LSP tool, plugin system, conversation fork/edit, and web IDE by @bearlike

Full Changelog: v0.0.9...v0.0.10

v0.0.9: Plan-mode approval, channel adapters, model fallback, & recovery pipeline

• Plan-mode approval — agents draft plans to scratch files with path-scoped edits and a shell allowlist, then propose for user approval. Budget nudges and revision tracking persist across runs.
• Model-cascade fallback — on retryable LLM failures the system walks an ordered fallback_models list, distinguishing transient errors from non-retryable ones (context overflow, auth).
• Retry/continue recovery — /retry replays the failed turn; /continue stitches a recovery prompt onto the existing transcript. Both available from CLI and per-message buttons.
• Natural-completion loop — replaces step-counted execution; the tool-use loop runs until the LLM emits a text response, with budget warnings injected as system messages instead of hard kills.
• Nextcloud Talk adapter — HMAC-SHA256 webhook auth, room-scoped sessions, @mention gating, threaded replies, and a slash command registry.

• Email adapter — IMAP polling, allowed_senders access control, Markdown-to-HTML rendering, and email thread tracking via References/In-Reply-To headers.
• Console UX overhaul — floating input bar, unified ConfigMenu, inline PlanCard with approve/reject, EditableTitle, ModelLabel with provider icons, and RetryFromHereButton on every message.

What's Changed

• ✨ feat: plan-mode overhaul, episodic approval, retry/recovery, channel adapters, model-cascade fallback by @bearlike

Full Changelog: v0.0.8...v0.0.9

v0.0.8: Active hypervisor, Agent Skills, Docker Compose, & console polish

• An active agent hypervisor that monitors running agents in real time — budget tracking, stall detection, and bidirectional steering without force-killing.
• Agent Skills support following the Agent Skills open standard. Place SKILL.md files in ~/.claude/skills/ or .claude/skills/ and invoke via slash commands or LLM auto-activation.
• Two-mode context compaction (full/partial) with structured summaries and post-compact file restoration. Auto-compacts near the context budget.
• Error-isolated hook system with session lifecycle events (on_session_start, on_session_end, on_compact) and external command hooks with fnmatch-based tool matching.
• Persistent MCP connection pool with automatic reconnection after consecutive errors and config change detection.
• Concurrent tool execution: tools are partitioned into parallel-safe and exclusive batches with per-tool timeouts.
• Multi-project support in the console with per-project tool and skill scoping.
• Console visual overhaul: typewriter greeting animation, color-coded trace activity feed, semantic CSS tokens aligned to Anthropic's design palette, and responsive layout improvements.

• Docker Compose deployment with pre-built GHCR images, nginx reverse proxy, and runtime config injection — no rebuild needed to change API URLs or keys.
• Hierarchical instruction discovery: CLAUDE.md files from user, project, rules, and local levels with priority ordering. Git context injected into the system prompt.
• Sub-agent spawning returns structured AgentResult with content, status, steps used, compressed summary, and artifacts instead of raw text.

What's Changed

• ✨ feat(core): active hypervisor, lifecycle-phased delegation, structured results, skills, compaction, hooks, MCP pool by @bearlike in #98
• 📝 docs: refresh screenshots, add banner GIF, and Docker Compose docs by @bearlike in #103
• ⬆️ Dependency updates: langchain-litellm, picomatch, yaml, ruff, mypy, pre-commit by @dependabot

Full Changelog: v0.0.7a...v0.0.8

v0.0.7a: Mewbo Console, unified tool-use loop, & multi-agent orchestration

• A new web console (apps/mewbo_console/) for task orchestration, replacing the Streamlit chat UI. Session management, real-time event polling, execution traces, diff viewing, and MCP tool selection — all backed by the REST API. Designed using Claude Design 🍀
• A unified async tool-use loop where the LLM drives tool selection via native bind_tools, with sub-agent spawning managed by the AgentHypervisor control plane.
• End-to-end MCP tool scoping: tool selections from the console flow through the API to filter the tool schema before LLM binding.
• User steering mid-session: send messages or interrupt the current step while a session is running.
• Uploaded file content is read from disk and injected into the LLM system prompt.
• Session status accurately reflects cancellation, errors, and step limits.
• Per-project Makefiles and Docker Compose stacks for consistent install, run, and deploy workflows across the monorepo.
• Frontend CI pipeline (ESLint, TypeScript, Vitest, Vite build) and mono-repo-aware pre-push checks.

What's Changed

• 📝 docs(readme): add console roadmap preview by @bearlike in #50
• ✨ feat(api): add session runtime + API session controls by @bearlike in #50
• ✨ feat(core): Add notification/share APIs & simplify orchestration planning schema by @bearlike in #56
• 🐛 fix(core): fix orchestration loops, over-correction, and redundant LLM calls by @bearlike
• ✨ feat(core): unified tool-use loop with agent hypervisor by @bearlike in #78
• ✨ feat(console): replace Streamlit chat with React console, align API & core contracts by @bearlike in #80
• Project packages renamed from Meeseeks to Mewbo.

Full Changelog: v0.0.6...v0.0.7a

v0.0.6: Local tooling, tighter orchestration, & safer approval

20260207-1157-59.4642545-3x.mp4

• A more reliable orchestration core that plans, executes, and recovers with tighter lifecycle control.
• Built-in local tooling (Aider adapters), plus config-driven behavior so you can safely edit files, browse directories, and run shell commands with clearer approvals.
• Cleaner setup docs and configs, so new users can get productive faster while trying the alpha.

What's Changed

• ♻️ refactor(orchestration): split & tighten orchestration lifecycle by @bearlike in #45
• 📝 docs(docs): document repo-managed git hooks and sync docs with README by @bearlike in #47
• ✨ feat(orchestration): add Aider tools and config-driven orchestration by @bearlike in #48

Full Changelog: v0.0.4...v0.0.6

v0.0.4: Auto‑replanning orchestration, MCP support, & uv-based modular installs.

• Orchestration now replans automatically on tool failures, so workflows complete reliably.
• Modular uv installs let you deploy only what you need (Core, API, Chat, CLI, HA).
• MCP discovery is more resilient, using cached manifests and clear disable reasons.
• Home Assistant integration is sturdier with explicit module checks, better errors, and cleaner entity caching.

Meeseeks.pre-alpha.demo.v2.mp4

What's Changed

• 📚 docs: add versioned MkDocs documentation site by @bearlike in #7
• 🐛 fix: configure git identity for mike deploys by @bearlike in #8
• 🐛 fix(docs): set default docs version on main and add MkDocs/Mike site scaffolding by @bearlike in #9
• ✨ feat(core): context-aware orchestration & MCP-ready CLI by @bearlike in #10
• build(deps): bump mcp from 1.12.4 to 1.23.0 in the pip group across 1 directory by @dependabot[bot] in #16
• build(deps): bump marshmallow from 3.21.2 to 3.26.2 in the pip group across 1 directory by @dependabot[bot] in #15
• build(deps): bump urllib3 from 2.2.1 to 2.6.3 in the pip group across 1 directory by @dependabot[bot] in #13
• build(deps): bump protobuf from 4.25.3 to 6.33.5 in /meeseeks-chat in the pip group across 1 directory by @dependabot[bot] in #11
• build(deps): bump urllib3 from 2.2.1 to 2.6.3 in /meeseeks-chat in the pip group across 1 directory by @dependabot[bot] in #12
• build(deps): bump protobuf from 4.25.3 to 6.33.5 in the pip group across 1 directory by @dependabot[bot] in #14
• ✨ feat(platform): add orchestration failure context, MCP discovery resilience, and CLI transparency by @bearlike in #18
• 🏗️ build(build): migrate monorepo to uv-based package layout by @bearlike in #40

New Contributors

• @dependabot[bot] made their first contribution in #16

Full Changelog: v0.0.3...v0.0.4

v0.0.3 - Added Home Assistant Conversation Plugin for Mewbo

What's Changed

• 🚀 RELEASE: Version bump to 0.0.3. Added Home Assistant Conversation Integration. by @bearlike in #3
• Now, you can use Home Assistant Assist or Conversations to talk with Mewbo.

Full Changelog: v0.0.2...v0.0.3

v0.0.2 - Added REST API support 🚀

Running Mewbo with Docker 🐳 (Recommended)

# Run the chat application 
docker run --env-file docker-meeseeks.env --add-host=host.docker.internal:host-gateway -p 8502:8502 ghcr.io/bearlike/mewbo-chat:1.1.0

# Run the REST API Server
docker run --env-file docker-meeseeks.env -p 5123:5123 --add-host=host.docker.internal:host-gateway ghcr.io/bearlike/mewbo-api:1.1.0

What's Changed

• 🚀 RELEASE: Version bump to 0.0.2. Added a basic REST-API wrapper. by @bearlike in #2
• 📦 NEW: Wrapped REST API endpoint around Meeseeks core.
• 📦 NEW: Docker image for API & Build action update

Full Changelog: v0.0.1...v0.0.2