v0.9.47 — ↻ Model lists that stay current · agents that don't stall waiting for you

Added

• ↻ Live model lists — Claude included. The pickers used to show whatever was
  hardcoded on release day, so a model that shipped today stayed invisible until
  the next office release. Every provider's list is now pulled from its own
  /models 20 s after boot and every 12 h, plus on demand: ⚙ → CONNECT →
  ↻ Refresh model list (all providers at once, with counts + when it last ran)
  and the ↻ beside the Model field (that provider only). Claude was the one
  provider with no live fetch at all — Anthropic's list needs your own
  credentials rather than a provider key, so the office uses ANTHROPIC_API_KEY
  when you've set one and otherwise the Claude Code CLI's existing login on this
  machine (read-only; sent nowhere but api.anthropic.com, macOS Keychain
  included). The newest model becomes the pre-selected default — and no existing
  agent's brain is ever rewritten for you. New POST /registry/models/refresh.
• 🔓 Auto-approve tool permissions (opt-in, ⚙ → TOOLS). Hand out work, walk
  away: a permission prompt with nobody there to click Allow used to park the job
  for 50 s and then be denied. With the switch on, requests are allowed
  automatically. Off by default — it is a real trade-off, so it's your call.
• Every running task is visible now, including the ones that used to run
  silently: 👻 sub-agents (ghosts) get their own live rows, and meetings /
  coffee breaks report as tasks like everything else.

Changed

• Delegated work is owned end-to-end. Teammates now carry a <work-autonomy>
  mandate: decide what's within your remit and keep going — come back only when
  genuinely blocked (a missing credential, a truly ambiguous requirement, an
  irreversible outward action that needs the owner's call). Work no longer stalls
  on questions the teammate could have answered itself.
• One agent hitting a limit no longer stops the floor. A sustained 429
  (rate/usage ceiling) now triggers the same failover as a sustained 5xx — the task
  moves to the fallback brain if you've set one — and the job pool runs 3 lanes
  (reg.maxJobs; 🌱 eco mode still keeps it to one), so the rest of the team keeps
  working while one agent waits out its limit and comes back by itself.
• Unified markdown rendering for all agent text — chat, feed, workflow analysis
  and results, proposals and notes now go through one vendored, XSS-safe md.js
  (marked + DOMPurify) instead of showing raw markdown and literal <b> tags.
  Thanks @bmdy5
  (#36).

Fixed

• The macOS installer could die silently right after downloading the shell —
  printing a success line while never installing Godot, the Claude hooks or the
  bagidea CLI. macOS ships bash 3.2, where ./source is a POSIX special
  builtin: sourcing a missing file exits the shell before || true can catch
  it — and on the prebuilt-shell path Rust is never installed, so ~/.cargo/env
  never exists. All four such lines now test the file first. Root-caused, with the
  exact version-independent fix, by
  @lyfer-bob
  (#38).
• The team stopped gathering in the CEO's room. The exec-room fence was only
  enforced on named waypoints, so social behaviours (chat, high-five, chase)
  walked agents straight in on raw coordinates. Every destination now passes
  through the fence: non-residents are rerouted to the lobby and only the CEO +
  Director stand there. Walking past is still fine.
• A missing security light crashed the patrol tween (rp_target on a freed
  object) on world builds that don't have one — now guarded, so the flash is
  skipped instead of erroring.
• The npm bootstrapper forces HTTPS + TLS 1.2 when fetching the install script,
  so it can't be pulled over a downgraded channel. Thanks
  @anupamme
  (#37).

No shell change in this release — bagidea update picks it all up.