0 v0.28.10

Provider and runtime fixes

• Add Together AI as an explicitly selected API-key provider. Preserve account bindings across worker forks and discover token-priced chat models, including zero-priced entries. Public catalog rows do not verify account access; the documented US endpoint surcharge is not included in catalog estimates.
• Choose custom OpenAI-compatible models after connecting. Bind discovery and caches to the current account/endpoint, reject connection-switch races, display errors, and retain manual model IDs.
• Give Claude Code the native loop's tools through an authenticated local bridge. Keep tool execution in 0, refresh definitions without replay, close the owned client when its loop ends, and reconcile usage receipts. Regression coverage uses synthetic processes/MCP clients; no real model assessment was run for this qualification.
• Preserve the last in-flight user/tool history after confirmed force-stop teardown. Late snapshots cannot restore callbacks, runtime/scope authority, or checkpoints; unconfirmed cleanup remains quarantined.
• Correct picker and active-display context metadata for the four published OpenCode Go model IDs. Match actual Go endpoints, keep explicit limits first, and leave other gateways and similarly named free variants unchanged. Published context metadata is not a capacity probe.

Original contributor commits are preserved. Existing execution-profile and process-supervision eligibility remain unchanged; direct Claude-client cleanup is not a universal descendant-process containment guarantee.