v0.2.0

What's changed

• Automatic translation is now off by default; use /talk-like-a-pro on to enable it, or /!! for a one-off forced translation.
• Hardened analyzer model routing, authentication handling, output validation, timeouts, configuration recovery, and pass-through behavior.
• Fixed Chinese fuzzy-request detection and reliable /! / /!! prefix handling.
• Added bounded concurrent repository context and prompt-injection safeguards.
• Reworked the English and Chinese README with clearer installation, controls, privacy, tuning, known limits, development, and MIT license guidance.

See CHANGELOG.md for the complete release notes.

v0.1.1 — 控制指令不再误触发翻译

Fixed

• 对话控制指令（确认执行/继续/开始/OK/continue 等）不再触发翻译——它们是确认或指挥，不是模糊需求
• 多段匹配器支持组合形式（好的，开始吧 / 确认，执行），带实际内容的输入不受影响

v0.1.0 — 模糊需求进，精确 prompt 出

v0.1.0- 2026-08-01

First release — minimal usable version: vague request in, precise prompt out.

Added

• Vague-request detection: heuristic prefilter (fuzzy verbs / short inputs) + force prefixes (/!! translate, /! pass through)
• Context probe: cwd, git status --short, AGENTS.md/CLAUDE.md head, last 6 session entries
• LLM intent analysis (per-model effort): judges clarity, produces a refined prompt or 1–3 clarifying questions
• Multi-turn Q&A: ctx.ui.select options + free-text input, then re-synthesis into the final prompt
• TUI settings (/talk-like-a-pro): enable switch, detail level (1 入门 / 2 标准), analyzer model picker, effort (filtered per model's supported thinking levels)
• Effort safety net: unsupported effort falls back by priority (high > max > … > off); empty output auto-retries one level higher
• Bilingual README (English + 简体中文 with switcher)

Fixed

• Model picker crash: SelectList theme now follows the official pi-tui contract (selectedText/description/scrollInfo/noMatch/selectedPrefix, onSelect/onCancel props)
• deepseek-v4-flash returns empty content at off effort — handled by the effort safety net
• Analyzer model could not be resolved when scopedModels is empty (no scoping configured) — now falls back to registry models with configured auth
• JSON parse reliability: maxTokens cap (1500), refinedPrompt ≤500 chars constraint, and parse-retry that feeds the broken output back to the LLM for repair