Orka Code 0.3.2 — live model lists and the standing upgrade path

0.3.2 is the first release admitted under the standing upgrade path: exact pins stay, unknown wire shapes still fail closed, and the hand work leaves.

Models come from the harness, live

• The composer chip, Create, Profile → Models, Assistant, and Atlas pickers offer exactly the models your account is served by the harness at session start: Claude Fable 5.1 shows up the moment Anthropic serves it, and the Codex list follows OpenAI’s current catalog. Only a change in the wire shape fails closed; membership changes never do.
• A session whose model is no longer offered says so and lets you switch; nothing is substituted silently.
• The recorded catalogs stay as offline fallbacks and parser tests. The Global default-model picker still uses the recorded catalog for now.

The upgrade path

• npm run harness:admit -- <claude> <codex> performs the mechanical admission steps (currency, Claude recapture, catalog sync, verification) and prints the exact manual commands for the rest; humans read only surfaces whose recorded shape changed.
• One standing evidence policy replaces the per-release risk amendments and private-visibility flags.
• Every admission carries a generated changelog delta sorted into four buckets (integrate now, needs adaptation, scrap, settings exposure).

Diffs show the whole file

• Diffs for tracked modifications and renames open with the complete saved file as context by default, hydrated in bounded, snapshot-verified reads; stale, filtered, binary, oversized, untracked, and added/deleted cases fall back to a truthful compact or metadata-only view.
• Diff syntax colors now use the same Timbre roles as the code and Markdown editors, in light and dark.

Continue on the other harness

• Pick a Codex model while in a Claude session (or the reverse) and Orka opens a new tab on that harness, seeded with a concise transcript of the conversation so far, and continues from there. The original tab stays intact.
• A session parked at a usage limit shows Continue with Codex (or Claude) on the limit card for the same hand-off.
• Only your messages and the assistant’s replies cross over; tool output stays with the harness that produced it.

Completed turns, quieter

• The completed-turn footer is a status icon, a compact runtime, and a locale-aware completion time; secondary telemetry sits behind an accessible hover/focus card. Failures and interruptions stay explicit. Legacy history without a recorded completion says so rather than guessing.

Harnesses

• Managed Claude Code 2.1.259 and Codex 0.153.0, admitted in one pass under the standing path. Every wire change between the previous pins was additive.
• Codex 0.153.0 can attach structured questions to a message. Orka has no answer control for them yet, so it shows the message and an explicit “not supported here, answer in the composer” notice carrying every question verbatim instead of dropping them.
• The updater’s retained-runtime declaration now lists every managed runtime the app keeps, so sessions bound to the 0.3.1 runtimes no longer block “Restart & update”.

Composer

• Pasting a long text (≥1,500 characters or ≥30 lines) stages a pasted_text_….txt pill instead of flooding the composer; the model still receives the full content, and the sent bubble shows the pill.
• One-row control rail, thin scrollbars, Enter creates a workspace, compact model picker with digit shortcuts.

Fixes from the 0.3.1 dogfood

• The bundled orka MCP works again on Claude Code 2.1.258 (the sidecar’s grant now travels inside the private MCP config).
• The root error boundary no longer fires on the first live event of a session (unbound queueMicrotask).
• “Restart & update” shows why an install was refused instead of doing nothing; the live-catalog probe retries instead of sticking on built-ins.