Claude Code configuration and hooks

Know which layer owns the behavior

Edit the owner. Changes in ~/.claude/plugins/ disappear at the next reconciliation.

Concern	Source	Inspect
Shared instructions	agents-common.md, CLAUDE.md.tmpl	chezmoi diff
Settings/status line	settings.json, executable_statusline.sh	jq . ~/.claude/settings.json
Interactive model default	claude-defaults.sh, .zshrc, .bashrc	type claude
Plugin declarations	claude-plugins.txt	claude plugin list --json
Hook scopes	claude-hook-scope.py, hook-scope.sh	jq . ~/.claude/hook-scope-state.json

The shared voice arrives through Claude’s cade output style. Project instructions still own project build, test, and deployment work.

The user setting selects claude-opus-5-5 at xhigh effort, with 1M context enabled by default, with Sonnet 5.5 as the configured fallback. The effort is also stored in modelSettings.claude-opus-5-5.effortLevel: Opus 5.5 ignores the older top-level user effortLevel setting. On NVIDIA hosts, server-managed settings select Sonnet again at startup even though Opus remains allowed. The interactive shell launcher therefore supplies --model claude-opus-5-5 --effort xhigh for session commands. Explicit model or effort flags win independently; management commands such as claude mcp, claude auth, claude plugin, claude doctor, and claude update bypass the launcher defaults unchanged. Resumed sessions retain the model stored in their transcript.

After applying the managed files, open a new shell and run:

claude

Expected result: a new session starts on Opus 5.5 at xhigh; confirm with /model and /effort status. Higher effort spends more tokens on reasoning. See model configuration for model-specific effort settings and context behavior.

Plugin and hook lifecycle

install/claude.sh installs the CLI, reads the marketplace inventory, refreshes declared marketplaces, reconciles plugins, and regenerates narrow hook gates. An inventory failure stops before marketplace or plugin changes; resolve authentication or managed-settings failures before retrying. After an intentional direct plugin update:

bash ~/dotfiles/install/claude.sh sync-hooks
jq . ~/.claude/hook-scope-state.json

Expected result: generated scope and drift state are printed. The generator narrows reviewed Hookify and Lean gates; upstream implementation drift restores the full upstream check and records a warning. Hookify still follows rule globs; Lean prompt checks remain for /lean4: and Bash checks use physical ancestor project markers.

Observe overhead without inventing a speedup

The status line stores aggregate transcript statistics and offsets under ${XDG_CACHE_HOME:-~/.cache}/claude-statusline, not transcript text. It incrementally handles appends; replacement, truncation, equal-size timestamp changes, and boundary-fingerprint changes invalidate cache state. This is not an end-to-end latency claim.

The shared chezmoi guard uses fresh ownership observation and resolves lexical/physical aliases. Third-party adapters fail open if its executable is unavailable; inspect the deployed helper before treating it as enforcement. Preserve inputs, cache state, versions, and focused test evidence. See hook overhead.

Practical routes

• researcher: source-grounded, read-heavy work; no editor tools.
• reviewer: adversarial review after implementation; no edits.
• df-agent-doctor: post-bootstrap component check; not account, GUI, or running-session proof.

Before changing schemas, check current hooks and plugins documentation.