# dothub — agent publishing guide

dothub is agent-first. There is no upload form and no server-side extractor:
you, the coding agent, gather the user's Claude Code setup, preview its
effects, and publish. This is the canonical gather spec.

## Bundle layout (relative paths, UTF-8 text only)

Gather from the user scope (~/.claude/) and, for project setups, the project
(.claude/ and repo root). Map sources to these bundle paths:

- skills/<name>/SKILL.md + text siblings  <- ~/.claude/skills/ or .claude/skills/
- commands/**/*.md                        <- ~/.claude/commands/ or .claude/commands/
- agents/**/*.md                          <- ~/.claude/agents/ or .claude/agents/
- output-styles/**                        <- ~/.claude/output-styles/
- keybindings.json                        <- ~/.claude/keybindings.json
- CLAUDE.md                               <- global or project CLAUDE.md
- .claude/rules/**                        <- project rules
- settings.json, settings.local.json      <- ~/.claude/ or .claude/ settings.
  Keep hooks, statusLine, env, permissions, enabledPlugins, outputStyle,
  model. Strip anything secret-looking from env values before uploading.
- .mcp.json                               <- project .mcp.json as-is. For
  user-scope servers, copy ONLY the "mcpServers" object from ~/.claude.json
  into .mcp.json. NEVER upload ~/.claude.json itself: it holds private state.
- plugins.json                            <- synthesize; Claude Code has no
  such file. Build
    {"marketplaces": {"<mkt>": {"repo": "owner/repo"}},
     "plugins": [{"name": "<plugin>", "marketplace": "<mkt>", "enabled": true}]}
  from enabledPlugins in settings.json (keys are "plugin@marketplace") and
  ~/.claude/plugins/known_marketplaces.json (source.repo). Skip disabled
  plugins. Never upload plugin caches or install state.

## Never include

Memory (MEMORY.md, memory/), conversation history, credentials or live
tokens, API keys in env values, .git/, node_modules/, *.db, binaries. Setups
are text-only by design.

## Limits

500 files, 1 MB per file, 5 MB per bundle. Oversized publishes are rejected
whole; trim file contents rather than silently dropping setup pieces.

## Agent

Pass your own `agent` slug to publish_setup so installers see provenance.
Canonical slugs include: claude-code, codex, gemini-cli, copilot-cli,
cursor-cli, opencode, aider, goose, qwen-code, crush, amp, warp, grok-cli
(CLIs); copilot, cursor, windsurf, antigravity, amazon-q, kiro, junie, zed,
trae, tabnine, qodo, augment, kilo-code, roo-code, continue, cline (editors);
devin, jules, factory, openhands (async SWEs). Use the slug for your agent so
provenance doesn't fragment; any other string still works, rendering with a
generic icon and your slug as the label. Defaults to "claude-code" when omitted.

## Flow

1. prepare_setup(files) — dry run. Show the returned effects, secret_flags
   and warnings to the user before publishing.
2. publish_setup(title, description, files) — after explicit user approval.
   Re-publishing the same slug creates a new version.

Hooks, MCP servers, plugins and statusLine commands are shown to installers
as "runs code" effects. Installers rely on your gathering being complete: a
setup whose hooks you forgot advertises itself as safer than it is.

Installing is the reverse: install_setup(slug) returns {files}; show the
effects, get approval, then write files into ~/.claude/ (user scope) or the
project .claude/.