Claude Code

Claude Code settings, hooks, plugins, slash commands, and skill links managed under ~/.claude.

Claude Code is Anthropic’s terminal coding agent. The repo manages its user-level settings, hook scripts, custom slash commands, and the skill links it loads. The claude binary itself is not installed by chezmoi.

Files

None of these are role-gated. Hook commands in settings.json use absolute /home/prad/... paths; on a machine with a different home directory, edit them after cloning.

settings.json

Model and UI

No statusLine is configured; Claude Code uses its default status line.

Environment

env sets two variables for every session:

Permissions

• defaultMode is auto, so sessions start in auto mode and the classifier approves routine actions.
• The allow list pre-approves seven read-only tools from the codegraph MCP server: codegraph_search, codegraph_context, codegraph_callers, codegraph_callees, codegraph_impact, codegraph_node, and codegraph_status.
• There is no deny or ask list.

autoMode.environment is a list of trust-boundary notes that auto mode reads when judging risk. It declares no org, cloud provider, secrets manager, or CI target. It names one private personal LeetCode repository as the trusted repo and treats routine work there as low risk. It marks .env* files as sensitive. Any host, namespace, or container with prod or production as a whole name segment counts as a sensitive remote target. IaC changes touching IAM, RBAC, networking, quota, and node pools count as protected.

Hooks

settings.json wires six hook commands:

Event wiring:

Plugins

enabledPlugins:

The LSP plugins give the agent language-server diagnostics for Go, Lua, Solidity, TypeScript/JavaScript (vtsls), and YAML.

Marketplaces

extraKnownMarketplaces registers eleven plugin sources:

mempalace and omc are registered but no plugin from them is enabled. context-mode is enabled but its marketplace is not declared here; it comes from Claude Code’s own marketplace state.

Hook scripts

cmem-to-obsidian.py

A stdlib-only Python script run on SessionEnd. It reads the hook JSON from stdin and exports that session’s claude-mem observations to an Obsidian note. It fails soft: any error is logged and the script exits 0, so it never blocks the session.

1. Opens the claude-mem SQLite database read-only (?mode=ro), so it never locks the live DB or its WAL.
2. Maps the Claude Code session_id to claude-mem’s memory session via sdk_sessions.content_session_id, then loads that session’s observations and latest summary.
3. Drops any observation containing <private>.
4. Writes <vault>/<project>/<date> <project> <session-prefix>.md. The file name is keyed to the session, so re-runs overwrite the same note.
5. Creates <vault>/<project>.md as a map-of-content note if it is missing, so the [[project]] wikilink resolves.

Note layout: frontmatter (date, project, session, source: claude-mem, tags claude-mem and project/<slug>), an abstract callout from the session summary, then sections by observation type. Each bullet links back to the claude-mem viewer at http://localhost:<port>/?observation=<id>.

Log file: ~/.claude/hooks/cmem-to-obsidian.log.

herdr-agent-state.sh

Generated and versioned by herdr (integration id claude, version 7). The header says reinstalling herdr’s integration overwrites it, so put custom hooks in separate files. It runs on SessionStart with the session action and exits immediately for any other action.

It only acts inside a herdr pane: HERDR_ENV=1, HERDR_SOCKET_PATH, and HERDR_PANE_ID must be set, and python3 must be on PATH. It then sends one pane.report_agent_session request over herdr’s Unix socket with the pane id, the Claude session id, the transcript path, and the session start source. herdr uses this to link the pane to the agent session. Subagent events and SubagentStop are ignored, so a finished subagent cannot mark an idle pane as working again. Socket errors are swallowed and the script has a 0.5 s timeout.

gitnexus/gitnexus-hook.cjs

A Node PreToolUse handler that adds GitNexus code-graph context to searches. settings.json does not reference it, so it is inactive until you add a hook entry.

• Runs only when .gitnexus exists in the working directory or one of its four nearest ancestors.
• Pulls a search pattern from Grep (the pattern), Glob (the first identifier-like path segment), or Bash commands that call rg or grep (the first non-flag argument). Patterns under three characters are skipped.
• Runs gitnexus augment <pattern>. It tries the bun global install at ~/.bun/install/global/node_modules/gitnexus first, then require.resolve, then npx -y gitnexus (8 s or 15 s timeout).
• Returns the CLI’s stderr output as hookSpecificOutput.additionalContext.

Slash commands

dot_claude/commands/ becomes ~/.claude/commands/, one user-level slash command per file.

Skills

~/.claude/skills/ holds two kinds of entries:

• Symlinks (symlink_<name> in the source). Each file contains a relative target such as ../../.agents/skills/blume, so ~/.claude/skills/blume resolves to ~/.agents/skills/blume. The shared library is documented on the Skills page.
• Real directories for Claude-only skills: absurd, agents-sdk, cloudflare, cloudflare-email-service, cloudflare-one, cloudflare-one-migrations, create-beads-orchestration, durable-objects, sandbox-sdk, todoist-cli, turnstile-spin, web-perf, workers-best-practices. See Claude-only skills.

Repo-local settings

This repo’s own .claude/settings.json applies only to Claude Code sessions opened inside the chezmoi source directory. chezmoi ignores dot-prefixed source entries, so it is never deployed. It runs bd prime on SessionStart and PreCompact, which reloads beads issue context at session start and before compaction.

Integrations

• kitty tab bar shows a robot glyph for tabs running claude.
• The fish ai picker launches claude --dangerously-skip-permissions; see fish.
• The claude-mem plugin feeds both cmem-to-obsidian.py and the /cmem-to-* commands.