0.1.3

Added

• Provider presets and API key connection cards for Xiaomi MIMO, Atlas Cloud, and Astraflow, plus updated OpenAI, Codex, Claude, and Gemini model catalogs and reasoning/context options
• ReadContext tool to search and page through archived conversation content, tool results, edit arguments, and retained terminal output; large MCP tool catalogs load full schemas on demand
• Optional per-tab composer drafts preserve unsent text and attachments when switching chats, including a New Chat tab with an unfinished draft
• Structured question inputs for text, multiline text, numbers, and dates, with required-answer validation and persisted answers
• Live per-model cache usage details distinguish cached, non-cached, and unknown-cache input tokens, and show cache-write tokens when reported
• Automated regression tests and CI checks for provider protocols, agent workflows, runtime installation, packaged extension activation, and leaked secrets

Changed

• Minimum supported VS Code version lowered from 1.125 to 1.96 for compatibility with IDEs using older extension APIs
• Context compaction saves a separate working checkpoint and todo state while preserving the full chat transcript; archived content remains retrievable and repeated prompt/context payloads are reduced
• Default agent step limit increased from 50 to 200; background subagents deliver results as they finish, and the parent waits for outstanding work before completing
• Provider transports updated for OpenAI Responses and Codex reasoning continuity, Gemini thought signatures, and Anthropic thinking/tool history; transient provider failures retry with backoff
• OAuth sign-in exposes authorization links with Open browser and Copy link actions, clearer errors, and manual recovery when the callback port is unavailable
• Local feature runtimes install on demand from a pinned dependency manifest with integrity verification and recovery from interrupted or concurrent installs
• Large text reads use bounded memory and paginated output; line diffs bound processing work and display clearer truncation notices

Fixed

• Todo updates tolerate provider field variants, missing fields, and partial updates without wiping existing tasks or freezing the run; continuation handling avoids premature completion and repetitive task loops
• Approval rules remain enforced across mode changes, subagents, compound shell commands, and symlinked external paths; saved user approval preferences survive activation
• File edits and undo operations preserve unsaved editor content and newer changes, serialize concurrent writes, and retain original bytes for restoration; reverting an earlier message only targets edits owned by that chat and turn range
• Pending edit reviews clear when tracked changes become clean after a Git commit
• Queued messages retain their original conversation, model, and mode across tab switches; Stop and chat deletion wait for run cleanup, and restored live questions remain answerable
• Switching chats reliably scrolls to the latest content as long transcripts finish rendering
• Cancelled or replaced OAuth sign-ins cannot save stale accounts; token refresh and account persistence preserve settings and avoid restoring disconnected accounts
• Usage totals retain concurrent parent, subagent, summary, and title requests without double-counting streamed updates; quota refresh/reset controls recover from timeouts, ignore stale replies, and prevent duplicate credit resets
• Provider requests preserve tool and image history, report incomplete streams, and respect explicit provider selections; model discovery retains duplicate model IDs across providers and times out unresponsive catalogs
• Background shell output ordering, cancellation, timeout, and exit-status reporting are more reliable; Grep pagination/counts and web-search result parsing handle incomplete scans and changing result layouts
• Semantic indexes preserve existing data after failed embeddings or incomplete scans and refresh when ignore rules change; semantic and documentation searches reject incompatible embedding configurations. GGUF downloads and model startup handle filename collisions and concurrent loads
• MCP requests and lifecycle hooks honor cancellation and timeouts, surface failures, and apply hook permission decisions; MCP connections refresh after server configuration changes