v0.2.0 — Claude Code-inspired guardrails + context management

What's New

Code Quality Guardrails (Claude Code–inspired)

• 18 system prompt guardrails: Code Discipline, Verification, Honesty, Security, File Creation, Self-Review, Failure Diagnosis, Output Discipline, Comment Preservation, Proactive Bug Discovery
• Project instructions override: WAVE.md/CLAUDE.md rules explicitly override AI's default behavior
• Warm context injection: technology-filtered conventions refreshed per file type during edits
• Sibling file reference: auto-attached when creating new files for style matching

Smart Context Management

• API-level tool result compaction: old results cleared in API request, full data kept in chatstore for UI
• Screenshot stripping: base64 removed from history after model sees it once
• Compressed screenshots: 512px/q30 JPEG (~60% smaller)
• Conversation compaction: auto-cleanup when context exceeds 30KB
• Model switch clears chat: prevents format mismatch between OpenAI Responses and Chat Completions

Edit Tool Improvements

• Marker mode (start_str + end_str): replace entire blocks between unique markers — perfect for HTML sections
• replace_all flag: global replacements without needing sed
• Trailing whitespace stripping: catches mismatches from AI-generated whitespace
• Flexible whitespace matching: ±2 line flex for normalized matching
• Line numbers in errors: shows exact lines when old_str matches multiple times
• Read limit increased: 2000 lines / 200KB default (was 500/100KB)

Image Generation

• generate_image tool: text-to-image via MiniMax image-01 API
• Inline rendering with Save button, image not stored in conversation history

UX Improvements

• Frontend watchdog: tracks actual activity, not total session time
• Stop button: visible during both "streaming" and "submitted" states
• Elapsed time counter: "AI is thinking... (15s)"
• Immediate thinking indicator: shows on submit, not just during streaming
• Context compaction indicator: yellow line when old results are cleared
• Single browser rule: AI uses one browser at a time

Bug Fixes

• Quick Add Model: fixed secret name mismatch (hyphens → underscores) for all 5 providers
• LLM parameter normalization: file tools accept file, path, file_path as aliases
• Last-read-file fallback: continuation reads work without re-specifying filename
• Retry on connection drop: 2s backoff for MiniMax API drops
• Shell integration timeout: increased to 20s, non-fatal on timeout
• Default web URL: changed from waveterm to mits-pl/wove
• Default AI timeout: increased from 90s to 180s

Documentation

• Improved README with step-by-step Quick Add Model setup
• Provider table with API key links
• Comprehensive MODIFICATIONS.md with all changes documented

Full Changelog

36 commits since v0.1.0

Built on Wave Terminal by Command Line Inc. — Apache 2.0