v3.4.0: Caveman-compressed renderer + one-line curl installer

Renderer: kill header sentence, wrapper tags, indentation, and redundant
src/path metadata. Short tag names (<r>, <perf>, <intent>). Session block
collapsed to pipe-separated one-liner. ~40% fewer structural tokens,
same information content, still valid XML.

Install: one-line curl script creates ~/.dhee with hidden venv, installs
package, configures Claude Code hooks automatically. No pip + settings.json
dance required. pip install and git clone remain for users who manage
their own environment.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Author: Ashish-dwi99

CHANGELOG.md

@@ -4,6 +4,31 @@ All notable changes to this project will be documented in this file.
 
 The format is based on [Keep a Changelog](https://keepachangelog.com/), and this project adheres to [Semantic Versioning](https://semver.org/).
 
+## [3.4.0] - 2026-04-16 — Caveman Injection + One-Line Install
+
+### Renderer compression (Caveman-inspired)
+
+Every token that isn't technical substance is now gone. The old injection wrapped content in a header sentence, `<dhee-context>` tag, `<docs>`/`<insights>`/`<memories>` wrappers, 2-space indentation, and verbose attributes like `src="CLAUDE.md" path="Repository Guidelines › Coding Style & Naming Conventions"`. All of it was scaffolding the LLM didn't need.
+
+- **Killed the header sentence.** ~25 tokens per injection, zero value.
+- **Killed wrapper tags.** Items are flat children of `<dhee>`, not nested inside `<docs>`/`<memories>`/etc.
+- **Killed indentation.** No leading spaces on any line.
+- **Short tag names.** `<rule>` → `<r>`, `<row>` → `<perf>`, `<i>` for insights.
+- **Dropped redundant metadata.** Doc chunks no longer ship `src="..."` or `path="breadcrumb"` — the content is authoritative on its own.
+- **Session block collapsed.** One-line pipe-separated format instead of nested `<decisions>`/`<files>`/`<todos>` tags.
+
+Net: ~40% fewer structural tokens per injection, same information content. Same XML parseability (still valid XML).
+
+### One-line install
+
+```bash
+curl -fsSL https://raw.githubusercontent.com/Sankhya-AI/Dhee/main/install.sh | sh
+```
+
+Creates `~/.dhee` with a hidden venv, installs the package, configures Claude Code hooks automatically. No `python -m venv`, no `pip install`, no editing `~/.claude/settings.json`. Re-running updates. `pip install dhee` and `git clone` remain available for users who prefer to manage their own environment.
+
+---
+
 ## [3.3.1] - 2026-04-16 — Honest Injection
 
 Dogfooding v3.3.0 revealed the Claude Code hook was a net-negative feature:

README.md

@@ -43,13 +43,34 @@ Your files stay exactly where they are. You maintain them the same way. Dhee jus
 
 ## Quick Start
 
+**One command. No venv. No config. No pasting into settings.json.**
+
+```bash
+curl -fsSL https://raw.githubusercontent.com/Sankhya-AI/Dhee/main/install.sh | sh
+```
+
+That's it. The installer creates `~/.dhee` with a hidden venv, installs the `dhee` package, and wires Claude Code hooks automatically. Next time you open Claude Code in any project, cognition is on.
+
+<details>
+<summary>Other install options</summary>
+
+**Via pip (if you manage your own venv):**
 ```bash
 pip install dhee
-dhee install    # hooks into Claude Code lifecycle
-dhee ingest     # chunks your CLAUDE.md, AGENTS.md into vector memory
+dhee install       # configure Claude Code hooks
+```
+
+**From source (contributors):**
+```bash
+git clone https://github.com/Sankhya-AI/Dhee.git
+cd Dhee
+./scripts/bootstrap_dev_env.sh
+source .venv-dhee/bin/activate
+dhee install
 ```
+</details>
 
-Done. Every Claude Code session now gets selective context injection.
+After install, Dhee auto-ingests project docs (CLAUDE.md, AGENTS.md, etc.) on the first session. Run `dhee ingest` manually any time to re-chunk.
 
 ---
 
@@ -93,10 +114,10 @@ Dhee manages information through a complete lifecycle — not just storage and r
                        ▼                     ▼
               ┌─────────────────────────────────────┐
               │  Token-budgeted XML injection        │
-              │  <docs><rule src="CLAUDE.md"         │
-              │    path="Testing" s="0.83">          │
-              │    Always run pytest before...       │
-              │  </rule></docs>                      │
+              │  <dhee>                              │
+              │    <r s="0.83">Always run pytest...│
+              │    </r>                              │
+              │  </dhee>                             │
               └─────────────────────────────────────┘
                                    │
                             LLM sees only
@@ -286,7 +307,7 @@ Zero LLM calls on hot path. Pure pattern matching + statistics.
 - **Chunker** — heading-scoped splits that respect code fences, paragraph boundaries, size limits
 - **Ingest** — SHA-tracked. Re-ingesting unchanged files is a no-op. Changed files get atomic chunk replacement.
 - **Assembler** — vector similarity search filtered by `kind=doc_chunk`, score threshold, token budget
-- **Renderer** — structured XML: `<docs><rule src="CLAUDE.md" path="Testing" s="0.83">...</rule></docs>`
+- **Renderer** — Caveman-compressed XML: `<dhee><r s="0.83">...</r></dhee>`. No header, no wrapper tags, no indentation — every byte earns its place. ~40% fewer structural tokens vs v3.3.
 
 ---
 

dhee/__init__.py

@@ -32,7 +32,7 @@
 # Default: CoreMemory (lightest, zero-config)
 Memory = CoreMemory
 
-__version__ = "3.3.1"
+__version__ = "3.4.0"
 __all__ = [
     # Memory classes
     "Engram",

pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "dhee"
-version = "3.3.0"
+version = "3.4.0"
 description = "Cognition layer for AI agents — persistent memory, performance tracking, and insight synthesis"
 readme = "README.md"
 requires-python = ">=3.9"