Muster 0.1.6

[0.1.6] - 2026-06-23

Muster 0.1.6 is the integration and CLI usability release. It moves Muster
closer to being a practical agent harness people can run directly from the
terminal or connect behind product/chat surfaces.

Added

• Interactive terminal chat overhaul: muster now opens a focused TUI chat
  surface with slash-command and @agent completion, command pickers, session
  naming/resume, provider/model selection flows, token/status commands, and a
  cleaner site-aligned palette.
• Capability-pack catalog expansion: all built-in plugin rows now resolve to
  real capability packs instead of policy-only placeholders. The catalog covers
  developer tools, browser QA, web frameworks, research, artifacts, data,
  security, daily ops, channels, providers, Codex, Claude Code, and MCP bridge
  workflows.
• Provider setup packs for OpenAI, Anthropic/Claude, vLLM, and
  Codex/Claude CLI runtimes with setup guidance, readiness checks, model-policy
  hints, and latency triage.
• Auth-heavy MCP setup workflows for GitHub, Notion, Linear, Google Drive,
  Firecrawl, Postgres, and n8n, including env readiness, OAuth setup/import
  paths, installability checks, and safe default tool policies.
• Built-in channel packs for Slack, Google Chat, Discord, WhatsApp, Teams,
  and Telegram. Telegram remains optional for regions where it is unavailable.
• Web framework pack for Frappe/ERPNext, React, Vue, Vite, Nuxt, and common
  web stacks, with local/dev/build/deploy/integration runbooks and production
  readiness checks.
• Indexed scoped memory retrieval using SQLite/FTS so long-running personal
  and project memory stays fast as conversations grow.
• Token ledger visibility in CLI flows so model spend, replay waste, and
  skill attribution are inspectable instead of hidden.
• Latency probe command: muster latency "prompt" --runs 3 separates
  provider time from Muster overhead, reports p50/p95, and flags whether a slow
  turn is provider-bound or caused by recall/prompt/persistence overhead.
• Local workspace-read fast path: trivial prompts that only ask to list the
  current folder are answered by Muster directly with an audited
  muster-local/workspace-read episode, avoiding unnecessary Codex round trips.
• Source-backed integration catalog expansion: muster plugins catalog now
  exposes 116 built-in entries after checking current Hermes and OpenClaw
  source trees. New setup-plan surfaces cover additional providers, channels,
  voice, document extraction, file transfer, webhooks, policy, token/cost
  governance, diagnostics, QA, migration, and memory backends without
  overclaiming execution before a pack/MCP is wired.

Changed

• muster plugins setup/check/enable now reports concrete prerequisites:
  missing env vars, setup URLs, MCP/channel availability, configured defaults,
  capability-pack readiness, and next commands.
• Runtime/provider UX is now picker-oriented: users can switch provider, model,
  runtime, MCPs, plugins, and skills from inside chat instead of memorizing
  exact IDs.
• Chat fast mode now means a warm native session with lighter Muster context
  rather than a cold one-shot Codex exec path, so repeat Codex turns can reuse
  app-server state while skipping recall, ambient skill scoring, and memory
  writes.
• Codex and Claude Code are modeled as first-class runtime capability packs,
  matching the way Hermes treats autonomous agent CLIs while preserving Muster's
  own session, memory, and token accounting.
• Browser and web-search workflows now prefer explicit MCP setup and cited
  source paths over hidden provider behavior.
• Default runtime guidance now uses Codex CLI instead of accidental localhost
  endpoints, keeping fast self-hosted providers opt-in rather than accidental.

Fixed

• Serialized warm Codex app-server turns per session and added injected-context
  hashes to native session handles, preventing concurrent turn event mixing and
  stale memory/skill/rule reuse across resumed provider sessions.
• Native session reuse now hashes stable instructions separately from volatile
  per-turn recall/skill context, preventing ordinary memory changes from
  cold-starting a fresh Codex app-server session.
• Native Codex and Claude Code attempts now stamp actual create/continue
  session mode and session id into the token ledger, so replay-waste warnings
  catch bloated continued Codex sessions instead of only Pi/API paths.
• Removed the “pack=no” gap from the built-in plugin catalog.
• Fixed CLI setup output so release-critical integrations show real setup links
  and readiness state rather than vague availability text.
• Hardened OAuth/MCP setup paths so credentials are checked and never printed in
  CLI output.