Claude Hook Experiment — Development skill for Claude Code

claude-hook-experiment

Reproducible Claude Code hook experiments.

Snapshot

Included snapshot:

• Claude Code 2.1.72
• OS Darwin 25.3.0

Observed in the included runs:

• settings-based PreToolUse, Stop, and SubagentStop worked in the tested cases
• plain stdout from PostToolUse did not appear as model-visible lint feedback
• hookSpecificOutput.additionalContext was observed in -p and in the checked true TTY interactive case
• the tested plugin_dir settings-hook cases worked
• frontmatter Stop was not observed on the tested plugin-backed subagent path

Raw snapshots are stored in `results/`.

Layout

• config/settings.local.json: shared settings-based hook fixture
• scripts/: experiment runners
• results/: saved JSONL snapshots from 2026-03-11
• cases/: minimal truth-case fixtures
• logs/: selected TTY logs

Requirements

• Claude Code CLI (claude)
• jq
• bash

Quick Start

For a manual smoke test:

./scripts/quick_test.sh

This prints a 3-case manual check and creates a small fixture under the system temp directory.

Re-run the automated matrix

Baseline matrix:

./scripts/run_auto_tests.sh

Default outputs:

• ${TMPDIR:-/tmp}/claude-hook-experiment/auto_cases/
• ${TMPDIR:-/tmp}/claude-hook-experiment/auto_results.generated.jsonl

Focused matrix:

CLAUDE_MODEL=sonnet ./scripts/run_focus_tests.sh

Default outputs:

• ${TMPDIR:-/tmp}/claude-hook-experiment/focus_cases/
• ${TMPDIR:-/tmp}/claude-hook-experiment/focus_results.generated.jsonl

Compare against the saved snapshots:

diff -u results/auto_results.jsonl "${TMPDIR:-/tmp}/claude-hook-experiment/auto_results.generated.jsonl"
diff -u results/focus_results.jsonl "${TMPDIR:-/tmp}/claude-hook-experiment/focus_results.generated.jsonl"

Manual truth cases

These fixtures were used to check stdout vs `additionalContext`, and local vs plugin-backed frontmatter `Stop`.

• cases/tty_truth_post_stdout_sonnet