Nest — 你的偏好，越用越合身

WASC June Challenge: 自成長 · 越用越懂你

A self-growing Claude Code Skill that learns your coding preferences from both dialog corrections and code-diff behavior, then silently applies them — reducing repetition with every session.

The more you use it, the more it fits. The more it learns, the quieter it gets.

Architecture

Python does the mechanical work (signal capture, storage, JIT injection). Claude Code does the intelligent work (semantic grouping, classification, confidence judgment).

flowchart LR
    User["User"] --> Host["Claude Code"]
    Host --> Skill["SKILL.md"]
    Skill --> Capture["Signal Capture"]
    Skill --> Agent["Agent Orchestrator"]

    Capture --> Dual["Dual-Track Signals"]
    Dual --> Regex["Regex Patterns"]
    Dual --> Claude["Claude Semantic"]

    Agent --> Classify["Classifier"]
    Classify --> Store["Memory Store"]
    Store --> JIT["JIT Injector"]

    JIT -->|"Top 20, scoped"| Host

Loading

Memory Lifecycle

flowchart LR
    Raw["raw signal"] -->|"classify"| Mature["mature (conf 40-79)"]
    Mature -->|"PS confirm"| Rule["rule (conf 80+)"]
    Rule -->|"silent apply"| Silent["silent application"]
    Mature -->|"stale / unused"| Decay["decay → deprecated"]
    Rule -->|"stale / unused"| Decay

Loading

Highlights

Dimension	Implementation
Signal Source	Dual-track: dialog text + code-diff behavior
Classification	Claude Code native (zero external API)
Memory Application	JIT context injection — Top 20 most relevant, filtered by project/directory/scope
Confidence Lifecycle	raw (1-39) → mature (40-79, PS confirmation) → rule (80+, silent) → decay
Scope Awareness	global / project / directory — same project, different folders, different rules
Validation	108 real Claude Code sessions, 100/100 WASC rubric score

v1 vs v2

	v1	v2
Signal Source	Dialog text only	Dialog + Diff behavior
Classification	Python + DeepSeek API	Claude Code native
Application	Full injection	JIT context injection (Top 20)
User Interaction	6 MCP tools	CLI scripts + lightweight PS confirmation
External Dependencies	anthropic SDK	Zero
Signal Capture	Regex only	Dual-track: Regex + Claude gap-filling

Repository Map

skill/                            — Host-agent skill entry (installed by Claude Code)
  SKILL.md                        — Skill instructions and trigger rules

src/                              — Core engine (zero pip dependencies)
  signal_capture.py               — Dual-track signal detection (regex + Claude)
  classifier.py                   — Memory classification and confidence assignment
  memory_store.py                 — Local JSON store with CRUD
  models.py                       — Memory and Signal data models
  agent.py                        — Orchestrator: JIT injection, pulse, summary

scripts/                          — CLI tools
  demo.py                         — 8-step WASC demo
  view_memory.py                  — List all memories
  edit_memory.py                  — Edit memory by ID
  delete_memory.py                — Delete memory by ID
  reset_memory.py                 — Clear all memories

tests/                            — Test suite
  test_harness.py                 — Rubric scoring (100/100)
  test_store.py / test_agent.py   — Unit tests
  test_signal_capture.py          — Signal detection tests
  test_classifier.py              — Classification tests

Quick Start

pip install -e .
python3 scripts/demo.py           # 8-step WASC demo
python3 tests/test_harness.py     # Rubric scoring (100/100)

CLI Commands

In Claude Code, invoke via /memory:

/memory view                      # List all active memories
/memory edit <id> '<json>'        # Edit a memory by ID
/memory delete <id>               # Delete a memory by ID
/memory reset                     # Clear all memories

Requirements

• Python 3.12+
• Claude Code
• Zero pip dependencies
• Zero external API keys

中文說明