ultimate-pi 0.1.7 → 0.2.2

package/vault/wiki/sources/wozcode.md

@@ -1,79 +0,0 @@
----
-type: source
-source_type: product-documentation
-author: WithWoz Inc.
-date_published: 2026
-url: https://www.wozcode.com/how-it-works
-confidence: medium
-key_claims:
-  - "25-55% token reduction vs vanilla Claude Code"
-  - "Three levers: smarter search, batched edits, quality loop"
-  - "AST truncation stubs function bodies, returns only signatures"
-  - "Haiku subagents handle ~40% of coding work (exploration)"
-  - "100% local architecture, zero cloud code transmission"
-  - "Post-edit syntax validation (TS, JSON/YAML, SQL) catches errors before model retries"
-  - "Fuzzy edit matching tolerates whitespace drift and visually-identical characters"
-  - "Claims measured from live Anthropic API usage fields, not simulations"
-tags:
-  - wozcode
-  - token-reduction
-  - claude-code
-  - agent-architecture
-created: 2026-04-30
-updated: 2026-04-30
-status: ingested
-title: "WOZCODE"
-
----# WOZCODE
-
-WOZCODE is a Claude Code plugin by WithWoz Inc. that reduces token spend through three compounding levers. Patent-pending. Runs 100% locally — code never leaves the developer's machine.
-
-## Three Core Levers
-
-### 1. Smarter Search (Input Token Reduction)
-- Ranked snippets instead of full-file grep dumps
-- AST truncation: stubs function bodies, returns only signatures
-- Returns only what the model needs for the current task
-
-### 2. Batched & Fuzzy Edits (Output Token Reduction)
-- Multiple file edits merged into a single model call
-- Fuzzy edit matching: tolerates whitespace drift, indentation changes, curly vs straight quotes, em-dashes
-- Near-misses land without retry round-trips
-
-### 3. Quality Loop (Rework Reduction)
-- Post-edit syntax validation: TS compiler, JSON/YAML/HTML parsers, SQL linter
-- SQL dialect auto-fix: rewrites backtick identifiers, unquoted reserved aliases, `COUNT(DISTINCT a, b)`, `date_trunc("month", col)`
-- Better error context: dialect-specific hints, real file content diffs instead of "string not found"
-
-## Agent Architecture
-
-| Agent | Model | Role |
-|-------|-------|------|
-| `woz:code` | User's frontier model (Opus/Sonnet) | Main thread, writes/edits code with full tool access |
-| `woz:explore` | Haiku | Read-only exploration, returns summaries to parent. ~15× cheaper. |
-
-~40% of coding work is exploration — automatically routed to Haiku.
-
-## Privacy Architecture
-
-- **Never sent**: source code, file contents, file paths, grep output, tool inputs/outputs, prompts, model responses, Anthropic API key
-- **In the loop for tool execution, not API transport**: API requests go direct from machine to Anthropic
-- **What WOZCODE servers do**: stats dashboard, auth endpoint, billing (via Stripe)
-- **Aggregated anonymous stats**: session-level token/cost/time metrics only
-
-## Competitor Comparison (from WOZCODE)
-
-| Tool | Approach | WOZCODE's Critique |
-|------|----------|-------------------|
-| SDL-MCP | Graph-based code exploration | Covers only exploration, not editing/validation |
-| Caveman | System-prompt prose compression | Most tokens are in tool calls, not prose |
-| WOZCODE | Full-session coverage | Covers exploration, editing, and validation |
-
-## Claims to Verify
-
-Claims are **medium confidence** (single source, commercial product, patent-pending means details obscured):
-- [ ] 25-55% token savings — benchmarkable on our codebase
-- [ ] 5-10× faster on DB tasks — domain-specific claim
-- [ ] AST truncation effectiveness on dynamic languages
-- [ ] Fuzzy edit matching false-positive rate
-- [ ] Post-edit validation latency overhead