v1.3.2

SiliconFlow Providers, Streamdown Chat & Desktop Streaming Polish

• Add built-in SiliconFlow providers for the two independent platforms: siliconflow (https://api.siliconflow.com/v1) and siliconflow-cn (https://api.siliconflow.cn/v1).
• Custom providers saved without an API key now appear in the Desktop/TUI model list (keyless local servers such as Ollama, vLLM, and LM Studio).
• Stream assistant markdown with Vercel Streamdown (word-level blurIn) instead of the custom paced-text drip, so tokens fade in instead of arriving in chunks.
• Coalesce high-frequency Server SSE frames in the Desktop main process before IPC (sse-ipc-bridge), reducing main/renderer stalls during fast token streams.
• Generate the first-turn session title with the currently selected model and wait for it to finish (or fall back) before starting the chat request, so exclusive local agent backends are not killed by a parallel title stream.
• Title requests to OpenAI-compatible reverse proxies (including Cursor Local API) send max_tokens instead of max_completion_tokens, and skip developer / store / stream_options fields those gateways reject.
• Milestone-based progress narration nudges keep intermediate status text tied to real phase changes instead of one line per exploratory tool call.

v1.3.1

Packaged Desktop Fix & Tool Runtime Polish

• Fix a packaged Desktop launch crash: main.cjs required browser-host.cjs, which was missing from the asar, so opening the app failed with Cannot find module './browser-host.cjs'.
• Restore grep / ls in the default SDK/session tool set; named child agents no longer incorrectly enable performance_gate from the performance-run environment.
• update_plan accepts status aliases such as done/complete/in-progress; nested subagent bash commands now default to a timeout so unbounded commands cannot hang.
• Desktop Inspector shows a shine border while the model controls the browser; tool cards, command grouping, and session titles (including thinking text) are more stable.

v1.3.0

Reliable-headless Default, Desktop Browser & Tool UI Polish

• Remove the legacy execution profile: print/json, TUI, Desktop/server, RPC, and child agents all use reliable-headless; --execution-profile legacy is a compatibility warning and remaps to the new profile.
• Add host-side TaskExecutionController / runReliableTurn: conversational turns no longer fail solely because an oracle is missing; chat-aware Controller entry requires instruction-owned acceptance evidence.
• Restore root-session performance_gate so Build/Plan G0–G7 frontier advancement works; named children still report evidence via ChildResult.
• Desktop Inspector embeds a browser: browser_* tools, the metis-browser skill, and automatic reconnect between the local Server and METIS_BROWSER_HOST.
• Align tool cards with the OpenCode expanded layout (write/edit flush diffs, terminal transcripts, generic output panels); use short natural titles (open/snapshot/screenshot/plan task, and so on); delay-mount terminal expansion and truncate large output to reduce stutter.
• Desktop interaction fixes for session archive, task chips / subagent conversation views, and Review Diff runtime refresh stability.

v1.2.0

Desktop Beautiful UI Redesign, Workspace Review & Performance Admission

• Rebuild the Desktop UI (Beautiful UI) with a full design-token system, atomic components (shadcn/ui), refined typography, and Inter Variable.
• Add a workspace change review center (Workspace Review Panel) with sidebar file-tree filtering, precise diff highlighting, and one-click Git commits.
• Introduce Performance Admission and the performance_admit tool to harden autonomous planning decisions and execution gating.
• Improve streamed message pacing (PacedMarkdown), smart auto-scroll (useAutoScroll), and interaction motion.
• Refine the onboarding brand logo, thinking-chain placement, and system status indicator motion.

v1.1.15

Compaction Optimization & Desktop Interaction Refinement

• Slim compaction input: drop redundant thinking blocks by default and recursively truncate huge tool-call arguments to prevent context bloat.
• Recalibrate multilingual token estimates with CJK weighting, improving compaction triggers and context accounting.
• Fix summary-model thinking over-truncation and silent loss of Desktop compaction records.
• Polish Desktop interaction details (theme-responsive useSystemTheme hook, sidebar, tool cards, and settings dialog).

v1.1.14

Usage Quota Card, Subagents Indicator & Desktop UX Upgrades

• Added interactive UsageQuotaCard with rich 5-hour/7-day quota meters, daily token & cost breakdowns, and full OAuth integration.
• Added 120Hz continuous floating hover indicator to Inspector SubagentsList.
• Exposed lastMessage in SessionInfo to improve session preview freshness and conversation list fidelity.
• Refined desktop token usage bar, turn files summary, message auto-scroll, and settings dialog styling.

v1.1.13

Subagent Workspace Isolation, UI Polish & ALE Verifier Hardening

• Enhanced workspace isolation with snapshot copy for non-git root environments and hardened git branch worktree validation.
• Throttled subagent spawn progress emissions and decoupled subagent lifecycle from internal timeout kills.
• Hardened official ALE benchmark verifier output parsing, score classification, and rescore workflows.
• Polished Desktop PlanPreview auto-scroll/clearance, UserBubble prompt details, and Settings dialog styles.

v1.1.12

Performance Runtime Governance & Multi-Process Concurrency Hardening

• Hardened Performance Runtime gate governance with cross-process file locks and state reload verification.
• Enforced mission binding checks and lease validation to prevent stale worker overwrites and deadlocks.
• Enhanced subagent spawn timeout resilience and expanded performance persistence test coverage.

v1.1.11

Desktop Chat Auto-Scroll & Message Rendering Polish

• Refined Desktop chat auto-scroll behavior with unified ResizeObserver and layout effect synchronizations.
• Hardened scroll-to-bottom tracking during streaming, dynamic content expansions, and conversation switches.
• Added comprehensive desktop chat scroll test fixtures and regression coverage.

v1.1.10

Custom Provider Thinking & Model Registry Enhancements

• Enhanced custom provider thinking capability discovery and reasoning level mappings (GLM, Qwen, Zhipu, OpenRouter, Together).
• Added flexible thinking option parsing for custom OpenAI-compatible endpoints.
• Updated provider documentation and refined model capability tests.