v0.preview.93

v0.preview.93 - 2026-09-02

Changed

• Tool activity in the TUI shows project-relative paths (or ~ for other home-directory paths), labels completed shell commands as Ran, and keeps header punctuation visually neutral
• Compaction learns a provider's actual context window after an overflow, scales summaries and anchors to fit, and retries once instead of walking a shrinking retry ladder

Fixed

• Parallel tool results share one context cap, preventing oversized batches from entering a repeated clear-and-rerun loop; cleared-result markers now identify the original tool call
• Buffered provider requests stop at the configured idle timeout instead of potentially hanging until the connection fails
• A second Cerebras turn after reasoning and tool use no longer fails because replayed reasoning_content is unsupported
• Goal evaluation and compaction retry once without tool_choice when an OpenAI-compatible provider rejects that parameter
• OpenCode zen usage-window errors are classified as terminal quota errors instead of being retried as transient rate limits

Wire

• SessionInitialized adds skills and subagents, so SessionStart and SessionUpdated announce the capabilities equipped for the session; both fields default to empty when absent