2.0.28

New Features Standalone inference and provider subscriptions The CLI can now run in standalone inference mode and authenticate directly with supported provider subscriptions. Subscription usage limits are visible in the interface, and subscription-backed models no longer consume Cosine credits. Terminal widgets Agents can now display structured terminal-native widgets for progress, results, tables, metrics, diagrams, and live monitoring. Pinned widgets remain visible across conversation compaction. Session goals You can now define and refine a goal for the current session using the  /goal  command. Goal suggestions and experiment configuration have also been improved. Themes and interface customization Added configurable themes, built-in light themes, and transparent-background support. The home animation and other interface elements now follow the selected theme consistently. New and updated models Added support for GPT-5.6 Codex subscription models, Claude Opus 5, Kimi K3, and DeepSeek V4 Flash. Model selection now shows provider information and uses more reliable defaults and context-window limits. Browser and computer use Browser automation and stateful computer-use workflows are now more capable and reliable. Web research can also search historical snapshots and pin archived sources for later reference. Sparklr experiments Sparklr now supports managed Tinker reinforcement-learning runs, configurable agent models, pooled provider subscriptions with failover, session goals, progress widgets, shareable sites, and inference cost tracking. Timeline and links URLs in the timeline are now clickable. Optional inline diffs make file changes easier to review without leaving the conversation. Easier bug reporting CLI diagnostics and bug reporting have been consolidated under  /bug-report , making it easier to collect debugging information and file actionable reports. Fixes • Expired ChatGPT or Codex authentication now starts a reauthentication flow instead of leaving the session unusable. • Temporary TLS and rate-limit failures are retried more reliably. • Resuming a session from inside the CLI now behaves consistently with  cos resume . • Interrupted agent runs are isolated correctly and no longer corrupt subsequent activity. • Plan mode now exits correctly when auto-approval is enabled, preserves legacy plans, and retains shell access where required. • ChatGPT sessions now compact at route-aware limits based on the measured context window, with safer fallback behavior when compaction fails. • Provider-specific Codex context windows are now calculated correctly. • Model switching is more reliable, with corrected defaults and automatic migration away from deprecated Fireworks models. • Headless Codex requests now stream responses correctly. • Configured subagent models are now honored, and the subagent dock adapts better to narrow terminal windows. • Autocommit operations are now isolated to the active worktree. • Concurrent checkpoints no longer overwrite one another. • Dropped file paths are preserved when entering slash commands. • Files outside the current drive root can now be read correctly after approval. • Browser and computer-use sessions recover more reliably across multi-step interactions. • LSP servers now start only when needed and shut down with the session, reducing unnecessary resource usage. • Large conversations render more efficiently, especially when inline timeline diffs are enabled. • Agent settings and commit preferences now persist correctly. • Plugin loading and CLI plugin integration are more reliable. • Sparklr experiments recover more consistently after reconnects, idle pauses, and transient runtime failures. • Long-running BYOK experiments now renew their provider leases and keep retrying when capacity is temporarily unavailable. • The footer no longer displays stale model information. • Disabled enterprise features no longer appear in menus. • Team MCP connections have been removed from the CLI.