Skip to content

v0.2.14

Choose a tag to compare

@github-actions github-actions released this 05 Sep 19:08

Added

  • New models show up on their own. LoomRouter now re-reads every enabled
    provider's catalog, and Codex's own native list, every fifteen minutes. A
    model your provider started serving since the last time you pressed Fetch
    models appears in the providers panel by itself, switched off, with its
    context window already filled in. Turning it on stays your call, and stays
    the moment LoomRouter checks which wire dialect the model actually answers
    on, so a first real turn is never routed on a guess.

  • OrcaRouter is a preset. Pick it from the provider list instead of typing
    the base URL by hand. It seeds no models: the 200 or so it serves arrive
    through discovery. Prompt caching is reported as unavailable, because the
    gateway accepts the cache header but never reports a cached token back.

Fixed

  • The desktop app no longer freezes while loading its config. The merged
    catalog was built from codex debug models, which returns a trimmed subset
    of the schema Codex Desktop expects. Without the missing fields the app hung
    on config load, while the LoomRouter panel cheerfully reported the
    integration as active. The catalog now carries Codex's full model schema.

  • Delegated work can write again. The subagent server was started without
    a sandbox profile, so Codex fell back to read-only and every worker, tester,
    debugger, migrator and refactorer task failed at its first write, silently.
    It now inherits the sandbox mode from your own Codex config. If that says
    read-only, it is respected rather than quietly widened.

  • Claude progress appears while the turn is still running. Progress
    reached Codex as one open item, so a long subagent run showed nothing but
    Thinking until the entire turn finished. Each step now closes as its own
    item and surfaces as it happens.

  • deepseek-v4-flash works through OpenCode Zen. Every request came back
    as HTTP 502. That tier is served only over chat completions, while the Go
    gateway keeps the Responses API for the same model id, so the two presets no
    longer share a single dialect.