Refactor

• Swap env.CLAUDE_CODE_NO_FLICKER for top-level "tui": "fullscreen" — the tui setting is the schema-native way to select the flicker-free alt-screen renderer. Per the official Claude Code settings schema, tui: "fullscreen" is "equivalent to CLAUDE_CODE_NO_FLICKER=1". Using the schema field is more discoverable, validates against settings.json's JSON Schema, and keeps env reserved for variables without a native counterpart.

Notes

• Behaviour is identical — same fullscreen renderer, same virtualized scrollback.
• env still carries CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS and CLAUDE_CODE_DISABLE_ADAPTIVE_THINKING (no schema-native equivalents; alwaysThinkingEnabled: false has different semantics — it disables thinking entirely).

Full Changelog: v2.5.1...v2.5.2

Bug Fixes

• effortLevel default changed from max to xhigh — max is only accepted as a session-scoped override (/effort max). As a persistent default in settings.json (or via CLAUDE_CODE_EFFORT_LEVEL), Claude Code only allows low / medium / high / xhigh. The previous max default was silently ignored.
• Removed betas: ["extended-cache-ttl-2025-04-11"] — the 1-hour prompt cache TTL is now generally available, so the beta header is dead configuration.

Notes

• If you need max reasoning effort, invoke /effort max per session — this is Anthropic's deliberate design for the top tier.
• 1h cache TTL continues to work natively after removing the beta header.

Full Changelog: v2.5.0...v2.5.1

Highlights

• New plugin andrej-karpathy-skills (Karpathy coding guidelines: Think-First, Simplicity, Surgical, Goal-Driven), default on.
• everything-claude-code default off — moved into a new PLUGINS_OPTIONAL group; only installed via explicit --all or menu opt-in.
• Selection-aware enabledPlugins merge — what you pick in the menu is what's enabled. Unselected catalogue plugins become false; user-added plugins outside our catalogue are preserved verbatim.
• Menu regrouped into use-case categories: Workflow, Integrations, Design & Content, Memory & Lifestyle, Academic Research. The old Skills group was dissolved; its items live alongside related plugins now (e.g. humanizer in Design & Content, paper-reading in Academic Research, update-config in Workflow).
• Windows menu parity — install.ps1 now supports → (open sub-menu) and ← (back), matching install.sh.

Bug Fixes (post-review)

• Catalogue now unions base keys with $selected, so plugins picked in the menu but not in shipped settings.json (codex@openai-codex, health@claude-health, pua@pua-skills) land as true instead of being silently dropped.
• Fallback union merge order fixed: existing values now win on conflict; v2.4.x everything-claude-code: true survives a plain upgrade.
• install_jq hoisted in install_settings — fresh installs on jq-less machines with statusline+lessons both on no longer silently skip the plugin filter.
• --dry-run banner now branches on $INSTALL_PLUGINS so it accurately describes the real run.

Docs

• README and README.zh-CN trimmed from 349 → ~195 lines; per-plugin duplication collapsed to plugins/README.md while the full interactive-menu catalogue (9 groups, 41 items with links and defaults) is kept inline.

Install

One-line remote install:

bash <(curl -fsSL https://raw.githubusercontent.com/Mizoreww/awesome-claude-code-config/main/install.sh)

irm https://raw.githubusercontent.com/Mizoreww/awesome-claude-code-config/main/install.ps1 | iex

Full changelog: CHANGELOG.md

Features

• permissions.defaultMode = "auto" — Claude autonomously approves safe actions and blocks risky ones. Installer downgrades to bypassPermissions on Claude Code < 2.1.80.
• effortLevel: "max" — Pins /effort to the highest reasoning tier by default. Older CLIs that reject max fall back to the highest tier they support.
• betas: ["extended-cache-ttl-2025-04-11"] — Enables the 1-hour prompt-cache TTL beta instead of the default 5-minute TTL.
• CLAUDE_CODE_NO_FLICKER=1 — Switches Claude Code to fullscreen rendering.
• CLAUDE_CODE_DISABLE_ADAPTIVE_THINKING=1 — Pins thinking budget to MAX_THINKING_TOKENS. No effect on Opus 4.7 (always adaptive).

Design Rationale

• Unknown keys (effortLevel, betas) are silently ignored by older Claude Code versions → no version gating needed in the installer except for auto mode.
• Only auto mode genuinely needs fallback; existing _supports_auto_mode check in install.sh handles that.

Notes

• CLAUDE_CODE_DISABLE_ADAPTIVE_THINKING and CLAUDE_CODE_NO_FLICKER need Claude Code 2.1.104+. Harmless no-ops on earlier versions.
• Adaptive thinking is always on for Opus 4.7 regardless of the env var — switch model if you need strict fixed-budget behaviour.

Install

bash <(curl -fsSL https://raw.githubusercontent.com/Mizoreww/awesome-claude-code-config/main/install.sh)

irm https://raw.githubusercontent.com/Mizoreww/awesome-claude-code-config/main/install.ps1 | iex

Bug Fixes

• Windows remote install crash: Fixed ParameterBindingException when running irm URL | iex on PowerShell 5.x.

Root Cause

Join-Path $SCRIPT_DIR "skills" $skill passes 3 positional arguments — PS7+ supports this via -AdditionalChildPath, but PS5.1 only accepts 2 (-Path, -ChildPath). When $skill = "adversarial-review", the third argument becomes unbound, triggering the error.

Fix

• Join-Path calls wrapped in nested form: Join-Path (Join-Path $SCRIPT_DIR "skills") $skill
• Added $_safeArgs filter to prevent $args token leakage in irm | iex piped mode

What's New

Adaptive Statusline Wrapping

The statusline now dynamically wraps segments to the next line based on terminal width, instead of being truncated. Progress bars shrink adaptively (min 8 chars) when space is tight.

Smart Terminal Width Detection

Walks ancestor process file descriptors via /proc/$PID/fd/ to find the real terminal in pipe contexts where $COLUMNS and tput cols are unreliable.

DeepXiv SDK Integration

New Academic Research group with deepxiv_sdk plugin for academic paper search and analysis.

Bug Fixes

• macOS wc -L fallback for visible_len()
• Negative bar sizing clamp (previously silently used full 20-char width)
• Width cache reduces subshell forks from ~13 to ~7 per render
• COLUMNS=0 pipe context detection

Full Changelog: https://github.com/Mizoreww/awesome-claude-code-config/blob/main/CHANGELOG.md

What's New

Academic Research Group

• Merged Plugins — AI Research (6 plugins) and DeepXiv (3 skills) into a unified Academic Research menu group (9 items, all off by default)
• DeepXiv skills (deepxiv-cli, deepxiv-trending-digest, deepxiv-baseline-table) are fetched via git clone from DeepXiv/deepxiv_sdk at install time
• Menu groups reduced from 9 to 8

Installer Improvements

• Pre-flight git availability check before DeepXiv clone
• git clone stderr preserved for diagnostics (no longer suppressed)
• mktemp -d uses portable template for macOS/BSD compatibility
• --all mode bounded to known skill list (no unbounded upstream installs)
• INSTALL_WARNINGS incremented on all DeepXiv failure paths
• Uninstall uses deepxiv-* glob pattern instead of hardcoded names
• PowerShell: Invoke-Retry pattern aligned with all other call sites
• PowerShell: temp dir uses GetRandomFileName() for uniqueness

Documentation

• Both READMEs updated with unified Academic Research group
• plugins/README.md updated with DeepXiv skills section
• Version bumped to 2.2.1

What's New

Two-Level Interactive Menu

• Groups with [selected/total] counts → Enter/→ opens sub-menu, ←/Esc returns
• Review group: code-review + adversarial-review + Codex CLI (mutually exclusive)
• Per-plugin granularity: all 23 plugins individually selectable

Bug Fixes (bash 5.x / Linux)

• (( var++ )) from 0 crashes under set -e → fixed with ++var
• [[ test ]] && cmd without || true → all instances guarded
• local _menu_active invisible to trap handlers → made global
• _read_key EOF maps to QUIT (not ENTER) — prevents phantom actions
• local key initialized to "" — prevents set -u crash
• Removed fd probe (was consuming first input byte)
• install.ps1: removed stale $groups override breaking Windows menu

Install

bash <(curl -fsSL https://raw.githubusercontent.com/Mizoreww/awesome-claude-code-config/main/install.sh)

What's Changed

• Codex adversarial-review plugin: Replaced built-in adversarial-review skill with official Codex plugin (codex@openai-codex). Falls back to Claude's code-reviewer agent when Codex is unavailable.
• Skill rename: /update → /update-config to avoid global namespace collision. Directory renamed to skills/update-config/.
• Smart-merge enabledPlugins: Changed from "existing wins" to "union" — upgrade users automatically get new plugins.
• Installer migration: Automatically cleans up legacy skills/update and skills/adversarial-review directories.
• Plugin count: 22 → 23 plugins, 7 → 8 marketplaces (added codex@openai-codex).

Full Changelog: https://github.com/Mizoreww/awesome-claude-code-config/blob/main/CHANGELOG.md

What's Changed (since v1.9.3)

🚀 Auto Mode Default (v2.0.0)

settings.json now ships with defaultMode: "auto" instead of bypassPermissions.

Auto mode (announced 2026-03-24) lets Claude approve safe actions autonomously while blocking risky ones — a safer middle ground between full bypass and manual approval.

• A classifier reviews each tool call before execution: safe → proceed, risky → blocked
• Installer auto-detects Claude Code version (_supports_auto_mode in install.sh)
• Falls back to bypassPermissions for Claude Code < 2.1.80 or if not installed
• Distinguishes "not installed" vs "version too old" in warning messages
• Portable sed fallback (no sed -i for macOS compatibility)

Requirements: Claude Sonnet 4.6 or Opus 4.6 model. Not available on Haiku, claude-3, or third-party providers.

📄 Paper Reading Skill Overhaul (v1.9.4)

Replaced the unreliable ar5iv HTML + Playwright screenshot pipeline with pure PDF + pymupdf4llm automatic extraction.

Before: arXiv → ar5iv HTML → Playwright element screenshots → manual fallback to pymupdf bbox clipping. Failed when ar5iv had no HTML version, and manual clipping frequently captured surrounding text.

After: All papers → PDF → pymupdf4llm.to_markdown(write_images=True) → automatic noise filtering → visual review & rename.

• Handles raster images, vector graphics (plots, diagrams), and tables in one call
• Auto-filters noise (logos < 100x100px, decorative lines with ratio > 10:1)
• Graceful degradation: theoretical papers with no figures produce text-only summaries
• Fallback to manual pymupdf clip extraction for missed figures
• OCR disabled by default (use_ocr=False) to avoid tesseract dependency
• Template image placeholders changed from hardcoded figure_X.png to HTML comment guides
• Net reduction of ~50 lines in SKILL.md (643 → 592 lines)

📝 Files Changed

⚠️ Breaking Changes

• Default permission mode changed from bypassPermissions to auto — users on older Claude Code versions will be auto-downgraded to bypassPermissions by the installer
• Paper reading skill no longer uses ar5iv or Playwright — all figure extraction is now via pymupdf4llm

Full Changelog: v1.9.3...v2.0.0