peerd-preview-v0.7.1

Changed

• The extension runtime is substantially lighter to start. The MV3 service
  worker now loads exact background and offscreen module surfaces instead of
  whole-module barrels, and Acorn-backed toolbox validation no longer sits in
  the Chrome service worker's cold graph. That static graph drops from about
  4.63 MB to 2.02 MB, and the combined service worker and offscreen cold graph
  drops by about half, so the worker wakes with far less to parse.
• Release staging compacts authored cold-path JavaScript in the disposable
  packaging tree only, preserving modules, names, lazy imports, and vendored
  bytes. Development is unchanged and still runs the source directly with no
  build step.

Fixed

• Privileged UI, keepalive, local-model, and toolbox-parser messaging is now
  pinned to the exact document that owns it, so a message cannot be answered by
  a surface that was never meant to handle it.

Added

• Apps and Notebooks now have browser-native Git history, branches,
  checkpoints, restore, diff inspection, and consent-bound Smart HTTP remotes.
  Repository operations quiesce live editors before touching the working tree.
• Pods add an ephemeral, isolated WASI shell for agent-driven command and file
  workflows. Cancellation, timeout, workspace ownership, and teardown are
  explicit terminal states rather than best-effort background cleanup.
• The preview dweb gains the cryptographic and runtime foundation for
  same-person device enrollment and private state restoration: rootless device
  grants, device certificates, private rendezvous, authenticated snapshot
  transfer, live custody wiring, and bounded partial-result reporting. The
  rendered enrollment controller and canonical hosted ceremony remain explicit
  release gates, so this is not presented as an end-user enrollment feature yet.
• Provider onboarding now verifies key shape and readiness before activation,
  with a first-run flow that does not reappear on established installations.
• On-device WebGPU inference now supports more than one model. The engine,
  residency tracking, chat picker, and Settings cards are driven by a model
  registry instead of a single hard-coded model, so a new on-device model is a
  registry entry rather than engine surgery. Weights already downloaded on an
  existing install continue to be recognised without re-downloading.
• Muse Glimmer 30B runs on-device via a second local engine: the Muse Glimmer
  WebGPU GGUF runtime (custom WGSL kernels, from the webml-community Space,
  vendored with provenance and hash-locked). The engine streams the ~12.4 GB
  Q2_K_XL GGUF from Hugging Face, caches it in IndexedDB, splits the model's
  reasoning channel out of the visible stream, and reports the effective
  context window it enforces to the trim layer. Which runtime a model needs is
  a registry fact (engine on the model spec), and each model's support
  verdict comes from its own engine's check, before any download.

Changed

• The composer, confirmation card, Stop state, onboarding, and narrow-sidebar
  layouts share the refreshed monochrome interaction system, with restored
  focus, clearer recovery actions, and expanded screen-reader coverage.
• App history and remote operations use the same mutation lanes and recovery
  contracts as the underlying OPFS workspace, including immutable approval
  targets and cleanup of repository data when an App is deleted.
• Self-device state transfer uses strict versioned schemas, authenticated
  unavailable-surface rows, deterministic byte identities, bounded caches and
  retries, and fail-closed completeness accounting.
• Release automation, dependency observation, package integrity, Firefox
  validation, and update-feed monitoring use tighter provenance and lifecycle
  gates.
• A local model download is refused up front when the vendored runtime cannot
  load its architecture, when another model is already downloading, or when the
  model id is unknown, each with the reason, rather than a failure mid-transfer.
• One on-device model is held in GPU memory at a time; requesting a different
  one unloads the previous model first and loads from cache, never silently
  answering from whichever model happened to be resident.

Fixed

• Closed portable-identity substitution, root-custody, revocation rollback,
  extension-origin, roster propagation, stale-offer, false-success, and
  post-cancellation mutation gaps found during adversarial review.
• Git remote approval can no longer authorize one origin and operate on a
  concurrently replaced origin; recent Notebook edits are flushed before
  version operations.
• Edited Apps receive a new byte-derived transfer identity, so a later restore
  cannot mistake changed content for an already-installed retry.
• Browser and actor lifecycle recovery now fences cold-start egress, private
  navigation, sign-in transitions, host loss, and stale async completions more
  consistently across Chrome and Firefox.
• Provider readiness survives worker restarts, preview updates avoid disrupting
  active work, and established vaults no longer regress into first-run setup.
• Passkey signup could stall right after the biometric prompt: the first
  message from the UI could reach the service worker before its route
  dispatcher had registered, and the browser could then retire the worker with
  that message still unclaimed. Early messages are now held until the
  dispatcher is ready.
• Settings no longer hides a configured Ollama provider during a temporary
  outage, an available local WebGPU host can be chosen as the default, a
  locked vault offers a direct unlock instead of misleading fallback copy, and
  the Ollama recovery commands are safe to copy and paste.
• An invalid Pod URL now reaches its visible failure state instead of leaving
  the tab in an indeterminate one.
• A queued focus restore in the Library could land long after the interaction
  that scheduled it and steal focus from whatever had claimed it since. Only
  the newest restore now acts, and it moves focus only when focus is orphaned
  or still parked where that restore left it.

---

Install (preview): Firefox: click the .xpi on this page. Chrome: drag
the .crx into chrome://extensions with Developer mode on. Most users
want the store packages (see the README). The auto-update feeds are
attached here and served at peerd.ai/updates/.