Dev Workspace (dilberry) — Summary

Dev Workspace is a git-branch-scoped context management system for Claude Code that turns each feature branch into an isolated workspace with its own directory tree, conversation history, plans, research, tasks, and CLAUDE.md. The four pillars are: directory system (structured dev/workspace/ subtree), git branch management (each branch gets a fresh workspace, archives persist), slash commands (8 workspace commands for branch creation, discovery, merge preflight), and Ruby hooks (session_start injects tree + context, session_end captures and summarizes conversation transcripts, user_prompt_submit runs word-triggered behaviors). At 39 GitHub stars, it is a developer-centric workspace scaffold combining git isolation, context injection, and history capture into a coherent workflow.

differs_from_seeds: Closest to agent-os (Archetype 4 — markdown scaffold + commands that write the structure) but significantly more sophisticated: Dev Workspace adds git-branch isolation, Ruby-based hook system with the claude_hooks library, automated conversation capture with headless Claude summarization, and an archiving system that persists branch context after merge. Unlike ccmemory (global Neo4j graph), memory is file-based and branch-scoped, deliberately switching context when you switch branches.

Dev Workspace (dilberry) — Overview

Origin

Created by dilberryhoundog. No license specified. Shell-primary (Ruby hooks via claude_hooks gem). Active as of April 2026.

Philosophy

"Shape, store, build and reveal context generated by and used for Claude AI while working on a codebase."

"Changing branches, changes context. This makes our context very targeted and focused, which Claude responds to excellently."

The core insight: each git branch represents a distinct work context. When you switch branches, you should also switch to the relevant context — plans, research, conversation history, current goals. Dev Workspace implements this literally: each branch has its own dev/workspace/ directory, and switching branches means Claude sees a completely different workspace.

"This system doesn't touch: CLAUDE.md, agents, skills or mcp servers, you are free to use these tools as you wish."

The explicit positioning as non-invasive: it adds a structured workspace layer without overriding Claude Code primitives.

Conversation Capture Philosophy

"This is a premium context capture tool. Every conversation you have with claude can be parsed and recorded in your workspace."

The conversation history feature is distinctive: it creates rich, token-efficient history files by:

• Filtering out tool output content (keeps file paths but not file contents)
• Using headless Claude to rename and summarize each session file
• Capturing thinking levels, conversation branches, and tool usage patterns

Git Management Philosophy

"Once in the flow, branches can come and go freely."

Workspaces are ephemeral per branch but archived permanently in dev/branches/. This gives both the fresh-slate feeling of a new branch and the permanence of accumulated context.

Dev Workspace (dilberry) — Architecture

Distribution

Standalone repo — clone and configure. Ruby gem dependency (claude_hooks).

Directory Layout

.claude/
├── commands/
│   ├── workspace/
│   │   ├── new-workspace.md         # /workspace:new-workspace
│   │   ├── discover.md              # /workspace:discover
│   │   ├── health.md                # /workspace:health
│   │   ├── research-v2.md           # /workspace:research-v2
│   │   ├── research.md              # /workspace:research
│   │   ├── tree.md                  # /workspace:tree
│   │   └── workspace-PR.md          # /workspace:workspace-PR
│   └── project/                     # Project-level commands
├── rules/                           # Claude Code rules
├── settings.json                    # Permissions + hook registration
└── skills/                          # Claude Code skills

dev/
└── workspace/
    ├── context/          # Discovery output, tree views
    ├── filebox/          # Scratch files dump
    ├── history/          # Session transcripts (headless-summarized)
    ├── plans/            # Plan files with templates
    ├── prompts/          # Structured prompt templates
    ├── research/         # Research output
    ├── reviews/          # Code review output
    ├── tasks/            # Task lists
    ├── CLAUDE.md         # Branch-specific Claude context
    └── WORKSPACE.md      # Branch workspace configuration

Archive Structure

dev/branches/
└── {branch-name}/       # Archived workspace snapshots
    └── {timestamp}/     # Each archive is timestamped
        ├── context/
        ├── history/
        ├── plans/
        └── ...

Required Runtime

• Claude Code
• Ruby (for hooks via claude_hooks gem)
• Git

Hook System

Uses the claude_hooks Ruby library which:

• Provides single entry point per hook event
• Intelligently combines multiple hooks
• Outputs correct JSON for Claude Code
• Produces comprehensive hook logs in ~/.claude/ directory

Dev Workspace (dilberry) — Components

Commands (8, under workspace: namespace)

Note: merge-preflight and merge-branch commands mentioned in README may live in project/ namespace.

Hooks (3 events via Ruby claude_hooks library)

Templates

Key Files

Dev Workspace (dilberry) — Uniqueness

differs_from_seeds

Closest to agent-os (Archetype 4 — structured markdown scaffold that writes your context) but significantly more developed: Dev Workspace adds git-branch isolation (each branch = distinct workspace), Ruby claude_hooks library hooks (session_start/end/prompt), automated conversation capture with headless Claude summarization, and an archiving system that persists per-branch context permanently. Unlike ccmemory (global Neo4j for cross-session recall), this is explicitly branch-scoped — switching branches switches memory. Unlike planning-with-files (hook-enforced three-file discipline), this provides a full directory taxonomy (8 subdirs) and treats conversation history as a first-class artifact.

Key Innovations

1. Branch-scoped memory — git branch switch = workspace switch; each branch has distinct CLAUDE.md, context, history, tasks
2. Conversation capture with headless summarization — session_end saves transcript as token-efficient history file; headless Claude renames and summarizes
3. Archive persistence — workspace snapshots in dev/branches/ survive branch deletion; archives commit to main for team access
4. claude_hooks Ruby library — single entry point per hook event with logging, JSON handling, and error management built in
5. Tree injection — auto-builds project directory tree at session start; gives Claude instant structural awareness without having to search

Observable Failure Modes

1. Ruby dependency — requires Ruby and claude_hooks gem; barrier for non-Ruby developers
2. No Windows support — Ruby hook scripts are Unix-specific
3. Headless Claude cost — session_end spawns a headless Claude session for summarization; adds API cost per session
4. History file quality — headless summarization quality depends on model; cheap models may produce poor summaries
5. Branch proliferation — if developers never merge, archives grow indefinitely in dev/branches/
6. No content in history — history files capture file paths but not contents; good for tokens, bad for deep context retrieval

Comparison to planning-with-files