v0.2.15

Changed

• Delegation runs on Codex's own subagents now. Ask for it in plain
  language -- "use multi agents with deepseek to investigate this project" --
  and the work goes to a real child agent thread you can open and watch, on any
  model you have enabled. LoomRouter used to run those workers itself, behind a
  single blocking call that showed nothing until it finished and gave up
  entirely at the ten minute mark. Codex's own spawn_agent accepts LoomRouter
  slugs, so there was never a reason to.

Removed

• The Agents page is gone. Saving a persona there never affected anything:
  every delegated worker was built from the request at the moment it was
  spawned, and the saved profiles were never read. The one thing an empty
  roster did do was switch delegation off, which is the opposite of what the
  page appeared to offer. Nothing you have to do; profiles already written to
  ~/.codex/agents stay on disk, and Codex still reads them as its own.

  The onboarding step went with it. Multi-agent stays a switch under Codex
  Integration, and delegation through LoomRouter models works whether it is on
  or off.

Fixed

• Models from a new Codex release finally show up. Once the integration was
  applied, LoomRouter asked Codex for its catalog and Codex answered with the
  file LoomRouter had written -- so every refresh re-read its own output and the
  native list could never grow. It now also reads the catalog compiled into the
  Codex binary, which no round trip can echo. On a current install that is
  gpt-6-astra and the two Daybreak models appearing for the first time.

• Claude Code shows every model your subscription serves. The list was
  five entries fixed at build time, so Sonnet 5 and Fable 5.1 could not appear
  however often you pressed Fetch models. It now comes from the same public
  catalog LoomRouter already downloads for context windows. New models arrive
  switched off, as everywhere else. If that catalog is unreachable you keep the
  list you had.