Core Plugin

Core plugin: dotclaude

The core plugin dotclaude is the base of the marketplace. It uses only documented extension points of Claude Code: a forced output style, five agents, one skill, one hooks module, a few classic hooks, and a status line command. It sets no Claude Code setting, because a plugin cannot set permissions, environment variables, or models. The settings snippet gives those keys, and you copy them. This page is the reference. The other plugins work without it.

What it does

• Forces the output style dotclaude (force-for-plugin: true), which holds the role, the tone, and the format. The module replaces the long system prompt with the shared sections and a prompt variant for each model (System prompt).
• Gives five agents with a model, an effort, and a turn limit each (Agents).
• Asks before the first contact with a repository of another owner that has an AI policy file (Contributions).
• Checks work with hooks that use no model call: git asks, edit asks, code-span denies, and approval claims (Hooks).
• Tells a subagent to report before its context or turn bound (Context bound).
• Puts each sentence of Markdown and commit text on its own line (Semantic line breaks).
• Shortens attachments, agent lines, and tool descriptions by exact id (Prompt surface).
• Shows a status line and a subagent status line (Status line).
• Reports differences from the settings snippet at session start (Drift check).
• Sets up OpenSpec with the setup skill (OpenSpec).

Install

Add the marketplace, install the plugin, then copy the keys that you want from the settings snippet.

/plugin marketplace add xsyetopz/dotclaude
/plugin install dotclaude@dotclaude

Claude Code 2.1.296 or later and git are needed. Node.js 22.18 or later is needed for the classic hooks, the status line, and the setup script. gh is needed for the policy ask. See Install for the steps and the update path.

Skills

Entry	Invoked by	What it does
/dotclaude:setup	The user only (disable-model-invocation: true)	Sets up OpenSpec in the current project. It shows each step and asks before it runs one. It changes no Claude Code setting. The script is skills/setup/scripts/openspec.mjs.

A user-invoked skill has no listing line, so it costs no tokens in a request. The name and description of a model-invocable skill have 150 to 300 bytes together (SKILL_META_LOW_BYTES and SKILL_META_HIGH_BYTES in tools/limits.mjs). The 300 is a bound, and a skill above it fails the size check. The 150 is a target, and a skill below it gives only an advisory. In a measure on Claude Code 2.1.293, skillOverrides did not apply to a plugin skill, so a plugin skill keeps its full listing line until you turn off the whole plugin. This is why docs is a separate add-on (Library docs). dotclaude ships no skill from another author.

Agents

Agent	Work
investigator	Read-only answers from files, logs, git history, or web pages.
implementer	One scoped slice of work.
debugger	A failure with an unknown cause.
reviewer	Read-only review.
test-runner	Runs checks and returns the exact failures.

The frontmatter of agents/*.md is the only source of the model, effort, and maxTurns of each agent. See Agents for the table with the values and the pin tests.

Hooks

The module hooks/register.mjs handles nine events: tool.check, tool.call (on Write, Edit, and Bash), prompt.compose, agent.spawn, prompt.section, prompt.attachment, agent.offer, tool.describe, and session.start. Classic command hooks cover Bash prints, git asks, edit checks, and subagent bounds, and a prompt hook on the model claude-sonnet-5-5 covers Stop. The page Hooks lists each one with its matcher.

The checks in short:

Check	Result	Code
Policy ask	Ask, once for each session and repository	features/guard/rules.mjs
File print	Deny, with the Read call to use	features/reads/
Git move	Ask for git stash, git reset, and a git checkout or git restore of a path	features/adherence/git.mjs
Test or limits file	Ask before a changed or removed existing line	features/adherence/rules.mjs
Lint or type-check config, CI file, or hook file	Ask before an edit	features/adherence/kinds.mjs
Skipped or focused test, or inline lint suppression	Ask before an added one	features/adherence/kinds.mjs
Code item outside a code span	Deny, in a Markdown write or a commit message	features/adherence/spans.mjs
Approval claim	Deny new text that claims an approval that no prompt of the user gives	features/adherence/approval.mjs
Stop rules	Block once, on Stop, a message that breaks one of three rules: a done claim with no passing check, an opener of agreement, praise, thanks, or an apology in place of the state, or a weakened check, test, bound, config, or hook	hooks/Stop.prompt.txt

A recursive rm is not in this list. It is an ask rule of the settings snippet, and no hook parses it. The module has no catch-all, so a hook that throws is skipped and the verdict of the engine stands.

Settings and options

The plugin has no option, and plugins/dotclaude/settings.json is an empty object. The settings are keys that you copy:

• Settings snippet gives the JSON.
• Settings snippet keys gives the reason and default of each key.
• Off switches lists the built-in features that the snippet turns off.
• Settings schema tells how the schema is made.

The status line needs two keys with the absolute path of the plugin folder:

node <plugin folder>/features/statusline/cli.mjs
node <plugin folder>/features/statusline/subagent-cli.mjs

Drift check

The session.start handler of the module compares the active settings with the settings snippet. It appends one note to the conversation only when it finds a difference, and the note asks Claude to tell you. It changes no setting.

Difference	Line of the note
A built-in plugin of the snippet is not false in enabledPlugins	Names each plugin and the setting that turns it off.
A skill of the listing has the source syncedSkills	Gives the count and names syncClaudeAiSkills.
The Claude Code version is not 2.1.296	Gives the version and the tested version.
A section id of the system prompt is unknown and has no colon	Names the id. dotclaude keeps the section as Claude Code made it.
The policy source of the settings has a key	Reports that the guard is seated, and names the parts that it skips.
CLAUDE_CODE_EFFORT_LEVEL or CLAUDE_CODE_SUBAGENT_MODEL_FORCE is set	Reports that the model and effort of the agent files do not apply.

A fact that cannot be read gives no line, so a failed read cannot make a false report. The section ids, the plugin names, and the version are data that Claude Code does not document (Undocumented reads).

Limits

Each bound is in one file, and tests pin the copies.

Bound	Value	File
Context window	300,000 tokens	lib/budget.mjs (CONTEXT_WINDOW)
Status line colors	Yellow at 75% and red at 90% of the window	lib/budget.mjs (USAGE_LEVELS)
bashOutputMaxChars	4,000 characters	lib/budget.mjs
Policy gh call	5,000 ms	lib/budget.mjs (POLICY_TIMEOUT_MS)
Policy text in the ask	2,000 characters	lib/budget.mjs (POLICY_FILE_MAX_CHARS)
Subagent context	100,000 tokens, and 85,000 for Haiku 5.5	features/context/limits.mjs
Subagent turns left at the note	2	features/context/limits.mjs (TURNS_LEFT_AT)
Subagent report	400 words	features/context/limits.mjs (REPORT_MAX_WORDS)
adherence hook timeout	10 s	features/adherence/limits.mjs
File lists of the asks for configs, CI files, skipped tests, and suppressions	Data in features/adherence/limits.mjs	features/adherence/limits.mjs
File size that adherence reads	1,000,000 bytes	features/adherence/limits.mjs
Rows of the status line	3	features/statusline/limits.mjs

The size of the first request is bound in tools/limits.mjs (FIRST_REQUEST_LOW_BYTES to FIRST_REQUEST_HIGH_BYTES), and the Plugin layout page explains the tiers.

Files

Path under plugins/dotclaude/	Part
hooks/register.mjs, hooks/hooks.json	The hooks module and the classic hooks.
hooks/Stop.prompt.txt, hooks/description.txt	The text of the Stop prompt hook and the hook description. just hook-text (tools/hook-text.mjs) builds hooks.json from them.
output-styles/dotclaude.md	The forced style: role, tone, and format.
features/prompt/shared.mjs, features/prompt/variants.mjs	The shared rules and the model variants of the system prompt.
agents/*.md	The five agents.
skills/setup/	The setup skill and its script.
features/<name>/	One folder for each concern: adherence, context, drift, guard, prompt, reads, sembr, setup, statusline, and surface.
lib/	Pure shared code.
tests/*.test.ts	The hook lab, run with just lab.

Troubleshooting

See Troubleshooting.

Related pages

• Hooks
• Guards
• Settings snippet
• Undocumented reads
• Plugin layout