Agent hook overhead

The common hooks preserve their existing responsibilities while avoiding work that does not apply to the current event. RTK remains enabled; model defaults are independent of these optimizations.

Runtime and shell paths

df-task and df-google-mcp use the runtime-only path of install/agent-runtime.sh. It shares the layout resolver with installers but never loads installer traps, unrelated credential files, or compiler setup. Flat mode needs no CPU detection; PLAT mode selects the current machine on each invocation, including GUI launches and shared homes. Relocating ~/.local continues to resolve its current physical target.

Bash and zsh login profiles still establish the full environment. Their shared partials use shell builtins to select platform specs and Node aliases/versions, and pass the shell explicitly to brew shellenv. They do not cache a machine’s PATH or disable login initialization.

Relevant Claude hooks

install/claude-hook-scope.py generates narrow negative gates in the installed Hookify and Lean plugin hook declarations. Plugin skills, commands, bootstrap, and applicable validation remain available. The source of truth is the repo’s patcher and home/dot_claude/executable_hook-scope.sh, not hand-edited cache files.

• Hookify checks its upstream rule-file glob every invocation. Creating a rule during a session activates it immediately.
• Lean prompt validation runs for explicit /lean4: invocations anywhere.
• Lean Bash validation retains force overrides and physical ancestor detection, including symlinked project directories and bounded stdin handling.
• Lean session bootstrap still runs so later Lean use has its runtime.

Bootstrap reconciles these gates after installing/updating plugins. After a direct claude plugin update, run:

bash ~/dotfiles/install/claude.sh sync-hooks

~/.claude/hook-scope-state.json records whether each installed plugin is scoped. A changed upstream gate implementation restores full upstream checks, emits a warning, and records the reason. Review the changed source and update its known hash only after testing applicability; do not remove the check to force a pass. Claude sessions that already loaded plugin declarations may need restarting to use the regenerated commands.

Status line, edit guard, and Cursor

Claude’s status line parses the incoming payload once. Its stdlib-only helper stores aggregate transcript statistics and offsets under ${XDG_CACHE_HOME:-~/.cache}/claude-statusline, never transcript text. Appends are processed incrementally. Replacement, truncation, timestamp changes at unchanged size, and boundary fingerprints invalidate old state. Arbitrary edits in the middle of an otherwise append-only transcript are outside that cache contract; remove that transcript’s derived cache to force a rescan. An unavailable cache only affects performance.

The shared chezmoi guard uses one fresh ownership query for a multi-file patch. There is no persistent ownership cache that could miss newly managed files. Lexical and physical destination paths, symlink aliases, add/delete/move targets, and filenames with spaces/newlines remain protected. Root normalization adds a small cost to a single-file check while larger batches are much faster.

Cursor compares hashes of live settings and their source on each event. Unchanged prompts do no import or extension inventory. Valid changed settings import immediately; invalid/racing writes are rolled back and retried. Private cache files contain hashes only. Extension inventory runs at session end or manually, under a separate lock, so prompts do not wait for it. Direct manual sync:

bash ~/dotfiles/install/cursor.sh sync-extensions

Validation and measurements

Run ./tests/run.sh for the canonical Docker suite. Focused tests cover applicable/inapplicable rules, upstream drift, symlink aliases, legacy jq syntax, cache append/reset/concurrency, Cursor imports/inventory, GUI runtime paths, flat/PLAT layouts, and fresh shell environment changes. The execution case in tests/agent_behavior.py exercises generated zsh code in the native selected model; its verifier checks counts, spaces, sorting, alternate cwd, and unchanged inputs.

For a repeatable Cursor comparison against a saved pre-change git revision:

uv run tests/cursor-hook-benchmark.py --baseline PRE_CHANGE_REF

Benchmark medians are local subprocess measurements, not additive end-to-end model latency. Keep original commands, iteration counts, cache warmth, input sizes, and runtime state with any reported comparisons. A successful synthetic hook probe does not establish that an already-open app reloaded its hook config. See troubleshooting for missing PyYAML, shell variable collisions, Cargo test arguments, and bounded output recipes.