v0.0.13: Memory Seed Bank + Compiler Speed

What's New

Memory Seed Bank

• 19 gold-standard seed IRs covering all 5 domains and 13 industries
• intentc memory seed — one command to populate the knowledge base
• Idempotent: safe to re-run, skips duplicates
• Each seed has quality_score=100 and passes schema validation

Compiler Speed (4 Optimizations)

1. Prefix Cache Reorder — fixed rules+schema+example placed first in prompt → DeepSeek auto prefix cache hits. Measured: input tokens drop 93% on repeat calls
2. Schema Minification — stripped all description fields, 4291→1708 chars (60% smaller, ~646 tokens saved)
3. Response Cache — LRU in-memory cache, 5-min TTL. Same input returns in 0ms (no API call)
4. Dynamic Token Budget — ≤4 sections→2048, ≤7→3072, else 4096

Benchmark Results

Cold start:       14975ms,  760 input tokens
Prefix cache hit:     0ms,   51 input tokens (93%↓)
Response cache:       0ms,    0 API call

Other

• Few-shot examples compressed (design+layout only)
• Prompt rules condensed 20→12 lines
• CLI shows ⚡ CACHE HIT indicator
• CONTRIBUTING.md cleaned up

Full Changelog: https://github.com/Deanbucai/intent-compiler/blob/master/CHANGELOG.md

v0.0.12

v0.0.12

Fixes

• Auto-create output directory (no more ENOENT crash)
• Compiler no longer auto-selects models — user controls via --fast and --model

Features

• intentc deploy — one-click Netlify deployment
• standalone.html — zero-install browser playground
• --fast flag for flash model
• Smart model selection removed from compiler (model choice = Agent concern, not compiler concern)

Install

npm install -g @deanbucai/intent-compiler
echo '你的描述' | intentc --render html --output page.html

v0.0.11: Bench Feedback Loop + Quality Memory + Skill

v0.0.11

Bench Feedback Loop

Errors detected by bench are stored in error_log with industry tags. Next compilation in same industry gets fix hints injected. Self-improving without model training.

Quality-Weighted Memory

Few-shot examples sorted by bench score. 100/100 IRs preferred over 75/100.

Claude Code Skill

intent-compiler skill registered. Agent auto-triggers on "做个网页".

MCP Memory Fix

compile_intent calls now auto-record to knowledge graph.

New Domains

ecommerce_content (5 types) + business_report (4 types). Total: 34 section types across 5 domains.

v0.0.10: Multi-page sites + intentc bench + experiment

v0.0.10

Added

• Multi-page sites: site.json shared design + nav + footer, intentc site init|build
• intentc bench: Automated quality benchmark (100-point scoring: sections, items, design, memory)
• Anti-error compiler rules: NO OMISSION / NO SHRINKAGE / NO DRIFT in compiler prompt
• Controlled experiment: IR proves 0% omission vs 17% without IR

Fixed

• README: removed outdated roadmap, old project structure, stale content
• .intent-compiler/ gitignored

Docs

• EXPERIMENT.md: controlled test methodology and results
• CLAUDE.md: experiment evidence + quality gate
• CHANGELOG updated

See full changelog for details.