End-to-end reverse-engineering of the command-code npm package (v0.33.0). Continues from the v0.28.1 decompilation with 5 new models, expanded repair coverage, the new deslop design tool, taste system, and context compaction.

The core insight: model errors at tool boundaries are a small, finite, compositional set. ~170 lines of repair code handle 36K+ production repairs. The architecture:

safeParse(input)
  ├─ success → ship it (valid inputs never touched)
  └─ failure → walk issue list
       ├─ for each issue path: try 5 rules in order
       │   1. renameAliasedField (60 aliases across 7 tools)
       │   2. dropNullOrUndefinedField
       │   3. dropEmptyObjectPlaceholder
       │   4. parseJsonStringifiedArray
       │   5. wrapBareStringAsArray
       └─ re-parse → success: log tool_input_repaired
                    → failure: return model-readable error

v0.33.0 additions: +14 field aliases, schema-level numeric coercions, extractRepairNotes for round-tripping repair hints.

The /design skill system uses work-pattern-first composition (7 patterns: Monitor, Operate, Compare, Configure, Learn, Decide, Explore) and a 10-smell catalog to detect AI-generated design tells.

New in v0.33.0: deslop tool — requires all 3 diagnostic reports, then applies 10 antidotes with strict fix ordering (composition → color → type → depth → motion → decoration).

The 21x cost reduction comes from: cheaper model ($0.95/$3.15 vs $5/$25) × fewer retry spirals (repairs eliminate failed tool calls) × lower token waste (schema coercions prevent failures entirely). DeepSeek V4 Pro beats Opus 4.7 6/10 times on internal evals.

• Taste System: taste-1 neuro-symbolic model with 3-tier storage (project/global/remote), uses Kimi K2.5 for onboarding
• Context Compaction: 3-tier auto-compaction (50%/80%/90% thresholds), CompactAgent using DeepSeek V4 Pro
• Retry System: Two layers (5 attempts auxiliary, 10 main), error classification, per-attempt telemetry

source/
├── tool-repair/
│   ├── repair-system.js    # Core validate-then-repair pipeline (5 rules, 60 aliases)
│   ├── pre-parse.js        # Pre-parse strategies (JSON repair, control char escaping)
│   └── read-file-defaults.js # Relational defaults (offset/limit)
├── models/
│   ├── catalog.js          # 28 models + provider routing + legacy aliases
│   └── pricing.js          # Per-provider pricing + subscription plans
├── agents/
│   └── builtin-agents.js   # Explore + Plan agents
└── tools/
    ├── self-knowledge.js   # Product knowledge tool
    └── skill-loader.js     # Skill resolution logic