v4.1.8 — Model Config with per-model effort, MCP key safety, 5h limit ring, Codex switch

Added

• Model Config panel with per-model effort — the model picker has a new Config button next to ✕. It opens a panel beside the picker (or in place of it when the chat panel is narrower than 880px) with one row per model: a show/hide toggle and effort buttons Auto · Low · Med · High · xHigh · Max. Levels are saved per model and shared by all workspaces; the plugin passes --effort every turn, and Auto lets the CLI decide. The model button shows the level, e.g. Opus 5.5 · High. Unsupported combinations are disabled: Haiku 4.5, Opus 4.5 and Sonnet 4.5 have no effort support, Opus 4.6 / Sonnet 4.6 have no xHigh, and Default follows the CLI's own /effort. The model in use and Default cannot be hidden
• Effort actually applies even with CLAUDE_CODE_EFFORT_LEVEL set — the CLI lets this variable (including one in a settings.json env block) override --effort, so every level silently ran at the variable's value. For models with an explicit level the plugin now also passes --settings with a file that clears it (verified by capturing the API request: --effort low went out as high before, low after). Models on Auto still follow your variable, and the Config panel shows a notice when one is detected
• 5-hour limit ring (5h ◔, left of Compact) for subscription logins, from the CLI's rate_limit_event. The tooltip shows usage and when the window resets, e.g. 5-hour limit: 19% used / Resets at 19:10 (in 25 min). It appears after the first reply and is remembered across restarts
• Opus 5 (claude-opus-5, $5/$25, 1M context) and Opus 4.7 are back in the picker; Opus 4.5 and Sonnet 4.5 are available but hidden by default (turn them on in Config)
• Codex switch in Settings → AI Assistant. When on and the Codex CLI is found on PATH, the system prompt teaches Claude to delegate via Bash codex exec (read-only by default, resume --last for follow-ups). Off by default
• Restore checkpoint upgrade — the confirmation shows how many files will be written back or deleted, a snapshot is taken before restoring so the restore can be undone (↩ Undo on the result card), and Claude is told the files were restored. Every message now has a restore point, even when nothing changed

Security

• No plaintext keys in MCP config files — Grok / Gemini keys are written as ${CHATUI_SECRET_*} placeholders and the real values are passed only in the CLI process environment of that turn
• Temporary MCP config folders are cleaned up — each turn deletes its own ~/.claude/mcp-* folder, and folders older than 24 hours are removed on start (previously they were never deleted)
• Debug log masks secrets, and restore snapshots no longer include debug_log.*
• Grok / Gemini key fields show a checkable fragment (first 7 characters + *** + last 4) instead of all dots

Changed

• Thinking Mode drops the xHigh level (effort now lives in Model Config); a saved xhigh is migrated to Ultrathink
• The generic Opus and Sonnet aliases were removed from the picker; workspaces that used them switch to Default
• Statistics fallback pricing uses per-model cache rates (Fable 5.1 reads at 2.5%, Opus 5.5 at 5%, others 10%) and counts 1-hour cache writes at 2× input. The statistics cache is rebuilt once

Removed

• Codex MCP template — Codex CLI 0.154.0 removed codex mcp-server, so the template (and its system prompt) could only produce a failing server. If you added codex-official earlier, delete or disable it in MCP settings to stop the "connection closed" error, then use the new Codex switch instead. The plugin no longer writes a Codex section into CLAUDE.md

Fixed

• [SYSTEM INFO] was added twice to the first message in Thinking mode
• Switching to Sonnet 5 showed Claude-sonnet-5 in the notification
• Debug log reported PATH: (missing) on Windows (the variable is named Path there)

Upgrade notes

• Codex users: if you added codex-official from the old template, delete or disable it in MCP settings (Codex 0.154.0+ no longer has codex mcp-server), then turn on Settings → AI Assistant → Codex
• If your ~/.claude/settings.json sets CLAUDE_CODE_EFFORT_LEVEL, models on Auto keep using it; levels you pick in Model Config take priority
• The statistics cache is rebuilt once after upgrading

Install

• VS Code Marketplace: search "Claude Code ChatUI" or update in place
• VSIX: download claude-code-chatui-4.1.8.vsix below → Ctrl+Shift+P → "Extensions: Install from VSIX..."