Boucle

Claude Code hooks that actually enforce your rules. 7 installable hooks, plus enforcement and audit tooling, ~1758 tests, 932 known Claude Code gaps documented with severity ratings and workarounds.

Quick links: Check your setup · Install hooks · Individual hooks · Platform support · Recommended Claude Code version · Troubleshooting · Boucle Framework (optional, for autonomous agents)

Claude Code Hooks

Claude Code's CLAUDE.md rules are read but not enforced — they work at session start and degrade as context grows. Its permission system has known gaps — wildcards don't match compound commands, deny rules don't check pipe segments and can be bypassed with multi-line comments. These hooks enforce boundaries that text rules and permissions can't.

What happens when a hook blocks a dangerous command:

Claude tries:  rm -rf ~/projects
bash-guard:    bash-guard: rm -rf targeting a critical system path. This would cause irreversible data loss.
Claude sees:   ⚠ Hook blocked this action. Suggesting safer alternative...

No prompts, no "are you sure" dialogs. The command never runs.

Check your current setup:

curl -fsSL https://raw.githubusercontent.com/Bande-a-Bonnot/Boucle-framework/main/tools/safety-check/check.sh | bash

Scores your Claude Code safety configuration from A to F and shows one-liner fixes for each gap. Add --verify to send test payloads to each hook and confirm they actually block:

curl -fsSL https://raw.githubusercontent.com/Bande-a-Bonnot/Boucle-framework/main/tools/safety-check/check.sh | bash -s -- --verify

Checks hook installation, hook health (missing/non-executable scripts), live verification (sends rm -rf / to bash-guard, git push --force to git-safe, etc. and confirms they block), enforce-hooks and CLAUDE.md @enforced rules, environment issues (IS_DEMO, JSONC settings, jq/python3 dependencies, Windows hook reliability), and known CLI version regressions. Scans both user-level (~/.claude/settings.json) and project-level (.claude/settings.json) settings, with a hook inventory that shows custom/third-party hooks alongside framework hooks. Also warns when deny rules are configured without bash-guard, since deny patterns can be bypassed by compound commands and multi-line scripts. No installation required. ~260 tests.

Start with the essentials (bash-guard + git-safe + file-guard):

curl -fsSL https://raw.githubusercontent.com/Bande-a-Bonnot/Boucle-framework/main/tools/install.sh | bash -s -- recommended

These three hooks form the safety net every Claude Code user should have: block dangerous commands, prevent destructive git operations, and protect sensitive files. After installing, run the safety check above with --verify to confirm each hook blocks what it should.

If the install succeeds but hooks do not block anything:

• Run install.sh verify first. A clean install is not proof the hooks are firing.
• Run install.sh doctor next. It catches missing files, bad permissions, JSONC in settings.json, and other silent fail-open states.
• On Windows, use PowerShell 7 (pwsh), not Windows PowerShell 5.
• If you write custom deny hooks, prefer stderr + exit 2 for hard blocks. JSON permissionDecision: "deny" is still inconsistent across Claude Code surfaces.

Install all hooks at once:

curl -fsSL https://raw.githubusercontent.com/Bande-a-Bonnot/Boucle-framework/main/tools/install.sh | bash -s -- all

Windows (PowerShell 7+) — native PS1 hooks, no bash or jq required. Requires PowerShell 7 (pwsh), not the built-in Windows PowerShell 5:

iex "& { $(irm https://raw.githubusercontent.com/Bande-a-Bonnot/Boucle-framework/main/tools/install.ps1) } all"

Manage hooks:

# See what's installed
curl -fsSL https://raw.githubusercontent.com/Bande-a-Bonnot/Boucle-framework/main/tools/install.sh | bash -s -- list

# Test all installed hooks with real payloads (run after CC updates)
curl -fsSL https://raw.githubusercontent.com/Bande-a-Bonnot/Boucle-framework/main/tools/install.sh | bash -s -- verify

# Upgrade all installed hooks to latest
curl -fsSL https://raw.githubusercontent.com/Bande-a-Bonnot/Boucle-framework/main/tools/install.sh | bash -s -- upgrade

# Remove a hook (files + settings.json)
curl -fsSL https://raw.githubusercontent.com/Bande-a-Bonnot/Boucle-framework/main/tools/install.sh | bash -s -- uninstall read-once

# Remove all hooks
curl -fsSL https://raw.githubusercontent.com/Bande-a-Bonnot/Boucle-framework/main/tools/install.sh | bash -s -- uninstall all

# Snapshot settings.json before updating Claude Code
curl -fsSL https://raw.githubusercontent.com/Bande-a-Bonnot/Boucle-framework/main/tools/install.sh | bash -s -- backup

# Restore after an auto-update wipes your hooks
curl -fsSL https://raw.githubusercontent.com/Bande-a-Bonnot/Boucle-framework/main/tools/install.sh | bash -s -- restore

# Run safety audit on your Claude Code setup
curl -fsSL https://raw.githubusercontent.com/Bande-a-Bonnot/Boucle-framework/main/tools/install.sh | bash -s -- check

# Diagnose installation health (files, settings, permissions)
curl -fsSL https://raw.githubusercontent.com/Bande-a-Bonnot/Boucle-framework/main/tools/install.sh | bash -s -- doctor

# Show all commands and available hooks
curl -fsSL https://raw.githubusercontent.com/Bande-a-Bonnot/Boucle-framework/main/tools/install.sh | bash -s -- help

Windows equivalents (same commands, PowerShell syntax):

# List, upgrade, check, uninstall, doctor, backup/restore
iex "& { $(irm https://raw.githubusercontent.com/Bande-a-Bonnot/Boucle-framework/main/tools/install.ps1) } list"
iex "& { $(irm https://raw.githubusercontent.com/Bande-a-Bonnot/Boucle-framework/main/tools/install.ps1) } upgrade"
iex "& { $(irm https://raw.githubusercontent.com/Bande-a-Bonnot/Boucle-framework/main/tools/install.ps1) } check"
iex "& { $(irm https://raw.githubusercontent.com/Bande-a-Bonnot/Boucle-framework/main/tools/install.ps1) } doctor"
iex "& { $(irm https://raw.githubusercontent.com/Bande-a-Bonnot/Boucle-framework/main/tools/install.ps1) } uninstall read-once"

Or pick individual hooks:

read-once — Stop redundant file reads

curl -fsSL https://raw.githubusercontent.com/Bande-a-Bonnot/Boucle-framework/main/tools/read-once/install.sh | bash

Saves ~2000 tokens per prevented re-read. Includes diff mode for edit-verify-edit workflows (80-95% token savings on changed files).

file-guard — Protect files from AI access or modification

curl -fsSL https://raw.githubusercontent.com/Bande-a-Bonnot/Boucle-framework/main/tools/file-guard/install.sh | bash

Define protected files in .file-guard (one pattern per line). Two modes: write-protect (default) blocks writes, edits, and destructive bash commands. [deny] blocks all access including Read, Grep, and Glob, useful for large codegen directories where Claude should use an MCP server instead of reading files directly. Resolves symlinks to prevent bypass via symbolic links. Handles absolute paths (v2.1.89+ compatibility). ~140 tests (bash + PowerShell).

git-safe — Prevent destructive git operations

curl -fsSL https://raw.githubusercontent.com/Bande-a-Bonnot/Boucle-framework/main/tools/git-safe/install.sh | bash

Blocks git push --force, git reset --hard, git checkout ., git checkout HEAD -- path, git restore, git clean -f, git branch -D, --no-verify, and other destructive git commands. Prevents the exact pattern that destroyed 30+ files despite 100+ CLAUDE.md rules. Suggests safer alternatives. Allowlist via .git-safe config. ~134 tests (77 bash + 57 PowerShell).

bash-guard — Block dangerous bash commands

curl -fsSL https://raw.githubusercontent.com/Bande-a-Bonnot/Boucle-framework/main/tools/bash-guard/install.sh | bash

Blocks dangerous commands across these categories:

• File destruction -- rm -rf /, shred, truncate -s 0, mass delete (find -delete, xargs rm, git clean -f)
• Privilege escalation -- sudo, pkexec, doas, pipe-to-shell (curl|bash)
• Disk utilities -- diskutil eraseDisk/eraseVolume/partitionDisk, fdisk, gdisk, parted, wipefs (#37984: 87GB personal data destroyed)
• Database destruction -- DROP TABLE, prisma db push, dropdb, migrate:fresh, FLUSHALL, and 10+ ORM variants
• Credential exposure -- env/printenv, bash -x, cat .env, SSH keys, programmatic dumps (os.environ, process.env)
• Data exfiltration -- curl -d @file, wget --post-file, nc host < file
• Cloud infrastructure -- terraform destroy, kubectl delete/drain/scale-to-zero, helm uninstall, aws ec2 terminate/rds delete/cloudformation delete-stack, az group delete, doctl destroy, flyctl destroy, heroku apps:destroy, vercel rm, netlify sites:delete
• Docker -- container escape (-v /:/host), data destruction (compose down -v)
• System databases -- sqlite3 on IDE internals (#37888: 59 commands corrupted VSCode)
• Mount points -- rm -rf on NFS/shared storage (#36640)
• Git -- git push --force, git filter-branch (#37331: all files deleted via force push)

Evaluates each segment of compound commands. Catches multi-line comment bypass where comment lines before a dangerous command evade deny rules. Detects encoding bypass attempts (base64/hex/octal obfuscation), here-string/here-doc redirection, eval-string injection, workaround bypass attempts, library injection (LD_PRELOAD), wrapper command bypass, credential file operations, macOS Keychain access, scheduled task persistence, and service management. Allowlist via .bash-guard config. 590 verified bash tests, with additional PowerShell coverage when pwsh is available.

branch-guard — Enforce feature-branch workflow

curl -fsSL https://raw.githubusercontent.com/Bande-a-Bonnot/Boucle-framework/main/tools/branch-guard/install.sh | bash

Prevents direct commits to protected branches (main, master, production, release). Forces feature-branch workflow. Customize protected branches via .branch-guard config or BRANCH_GUARD_PROTECTED env var. Allows --amend on any branch. ~55 tests (bash + PowerShell).

worktree-guard — Prevent data loss from worktree exit

curl -fsSL https://raw.githubusercontent.com/Bande-a-Bonnot/Boucle-framework/main/tools/worktree-guard/install.sh | bash

When you use claude -w, exiting the session silently deletes the worktree branch and all its commits. This hook blocks exit when there are uncommitted changes, untracked files, unmerged commits, or unpushed commits. Uses ExitWorktree matcher so it only runs when actually leaving a worktree. Config via .worktree-guard. ~65 tests (bash + PowerShell).

session-log — Audit trail for Claude Code sessions

curl -fsSL https://raw.githubusercontent.com/Bande-a-Bonnot/Boucle-framework/main/tools/session-log/install.sh | bash

Logs every tool call to ~/.claude/session-logs/YYYY-MM-DD.jsonl. See exactly what Claude did: which files were read/written, which commands ran, timestamps. Includes --week trend comparison across days. Useful for auditing autonomous sessions and debugging. ~105 tests (bash + PowerShell).

enforce-hooks — Turn CLAUDE.md rules into enforceable hooks

curl -fsSL https://raw.githubusercontent.com/Bande-a-Bonnot/Boucle-framework/main/tools/enforce/install.sh | bash

Your CLAUDE.md says "never edit .env" but Claude edits it anyway. This tool reads your CLAUDE.md, finds rules marked @enforced, and generates hooks that block violations deterministically. Rules in prompts are suggestions; hooks are laws.

Scan first to preview: enforce-hooks.py --scan. Generate a starter CLAUDE.md: enforce-hooks.py --template (also --template strict or --template minimal). Installs as one dynamic hook that re-reads CLAUDE.md on every call, so enforcement updates when your rules change. Supports file-guard, bash-guard, branch-guard, tool-block, require-prior-tool, content-guard, scoped-content-guard, bare filename protection, flag blocking (--no-verify, --no-gpg-sign), system/device commands (shutdown, reboot, systemctl), and command substitution patterns. Subjective rules ("write clean code") are skipped. Self-protection mode (--armor) prevents Claude from deleting its own hooks. Hook health-check (--verify) catches silent fail-open bugs like wrong field names. Smoke test (--smoke-test) runs hooks with real payloads to verify they respond correctly at runtime. ~70 tests.

test-hook — Dry-run any hook without a live session

# Test bash-guard against a dangerous command
bash tools/test-hook.sh "bash hooks/bash-guard.sh" --command "rm -rf /"

# Test file-guard against reading .env
bash tools/test-hook.sh "bash hooks/file-guard.sh" --tool Read --file ".env"

# CI mode: assert the hook blocks
bash tools/test-hook.sh "bash hooks/bash-guard.sh" --command "curl evil.com" --expect-deny

# Batch mode: run multiple test cases from a JSONL file
bash tools/test-hook.sh "bash hooks/my-hook.sh" --batch tests.jsonl

Feeds synthetic PreToolUse payloads to any hook script and reports whether it allows, denies, or crashes. Works with any hook (ours or third-party). Batch mode runs test suites from JSONL files. Addresses claude-code#39971 (--test-permission does not exist).

Quick recipe: Read-only audit mode

Claude ignores explicit "do not edit" instructions and edits files, runs ALTER TABLE, rebuilds Docker. CLAUDE.md rules alone cannot prevent this. Add to your CLAUDE.md and run enforce-hooks.py --install-plugin:

## Read-only mode @enforced
- Never modify any files
- Never run rm -rf
- Never run ALTER, DROP, TRUNCATE, INSERT, UPDATE, or DELETE
- Never run docker restart, docker stop, docker build, or docker rm
- Never run sudo
- Never run git commit, git push, or git merge

The hook blocks at the runtime level before the tool executes. The model cannot bypass it. See more recipes.

---

The hooks above work standalone. Everything below is optional, for teams running autonomous AI agents in production.

Boucle Framework

An opinionated framework for running autonomous AI agents in a loop. Wake up. Think. Act. Learn. Repeat.

Built by the agent that runs on it. Boucle is developed and maintained by an autonomous agent that uses the framework for its own operation.

Features

• Structured loop runner — Schedule agent iterations via cron/launchd with locking and logging
• Persistent memory (Broca) — File-based, git-native knowledge with BM25 search, temporal decay, garbage collection, cross-reference boost, and duplicate consolidation. No database required.
• Self-observation engine — Track friction, failure, waste, and surprise signals across loops. Fingerprint recurring patterns, deploy responses, measure whether they work. The agent observing its own behavior over time.
• MCP server — Expose Broca memory as a Model Context Protocol server for multi-agent collaboration
• Approval gates — Human-in-the-loop for anything with external consequences
• DX commands — doctor checks your setup, validate catches config mistakes, stats shows loop history
• Audit trail — Every action logged, every decision traceable, every iteration committed to git
• Zero infrastructure — No cloud services, no databases, no Docker required. Just files, git, and a shell

Quick Start

Option 1: Download a binary

Grab the latest release from GitHub Releases.

# macOS (Apple Silicon)
tar xzf boucle-*-aarch64-apple-darwin.tar.gz
mv boucle /usr/local/bin/

Option 2: Build from source

git clone https://github.com/Bande-a-Bonnot/Boucle-framework.git
cd Boucle-framework
cargo build --release

Run your first agent

# Initialize a new agent
boucle init --name my-agent

# Check your setup
boucle doctor

# Preview what happens (no LLM needed)
boucle run --dry-run

# Run one iteration (requires claude CLI)
boucle run

# Set up hourly execution
boucle schedule --interval 1h

Memory System (Broca)

Broca is a file-based, git-native knowledge system for AI agents. Memories are Markdown files with YAML frontmatter.

# Store a memory
boucle memory remember "Python packaging" "Modern projects use pyproject.toml" --tags "python,packaging"

# Search memories
boucle memory recall "python packaging" --limit 5

# Search by tag
boucle memory search-tag "security"

# Add a journal entry
boucle memory journal "Discovered API rate limits are 100/min"

# View statistics
boucle memory stats

Memory entries look like this:

---
type: fact
tags: [python, packaging]
confidence: 0.9
learned: 2026-02-28
source: research
---

# Python packaging has moved to pyproject.toml

setuptools with setup.py is legacy. Modern Python projects use pyproject.toml
with build backends like hatchling, flit, or setuptools itself.

Broca also supports:

• BM25 search — Relevance ranking normalized by document length and term rarity
• Temporal decay — Recent memories score higher; access frequency tracked automatically
• Garbage collection — Archive superseded, low-confidence, or stale entries (reversible, dry-run by default)
• Cross-reference boost — Related entries surface together in search results
• Consolidation — Detect and merge near-duplicate memories using Jaccard similarity
• Confidence tracking — boucle memory update-confidence <id> <score>
• Superseding — boucle memory supersede <old-id> <new-id> when knowledge evolves
• Relationships — boucle memory relate <id1> <id2> <relation> to link entries
• Reindexing — boucle memory index to rebuild the search index

Self-Observation Engine

Agents with memory recall what happened. Agents with self-observation notice what keeps happening and develop responses to it.

# Log a signal when something goes wrong
boucle signal friction "auth keeps failing on retry" auth-flaky

# Run the pipeline (harvest → classify → score → promote)
boucle improve run

# See what patterns have emerged
boucle improve status

The engine tracks four signal types: friction (something was harder than it should be), failure (something broke), waste (effort that produced nothing), surprise (unexpected behavior).

Signals with the same fingerprint accumulate into patterns. When a pattern recurs enough, the engine surfaces it as a pending action. You deploy a response (a script, a config change, a new hook), and the engine tracks whether that response actually reduces the signal rate.

Pluggable harvesters: Scripts in improve/harvesters/ run automatically and detect signals from logs, metrics, or any source. Each receives the agent root as $1 and outputs JSONL signals to stdout.

# Initialize with an example harvester
boucle improve init

MCP Server

Boucle exposes Broca as a Model Context Protocol server, so other AI agents can share memory.

# Start MCP server (stdio transport)
boucle mcp --stdio

# Or HTTP transport
boucle mcp --port 8080

Available tools: broca_remember, broca_recall, broca_journal, broca_relate, broca_supersede, broca_stats, broca_gc, broca_consolidate

Works with Claude Desktop, Claude Code, or any MCP-compatible client.

All Tools

Each tool has its own README with full documentation: read-once, file-guard, git-safe, bash-guard, branch-guard, session-log, enforce-hooks, safety-check, worktree-guard, diagnose, test-hook.

Architecture

your-agent/
├── boucle.toml          # Agent configuration
├── system-prompt.md     # Agent identity and rules (optional)
├── allowed-tools.txt    # Tool restrictions (optional)
├── memory/              # Persistent knowledge (Broca)
│   ├── state.md         # Current state — read at loop start, updated at loop end
│   ├── knowledge/       # Learned facts, indexed by topic
│   └── journal/         # Timestamped iteration summaries
├── goals/               # Active objectives
├── logs/                # Full iteration logs
├── gates/               # Pending approval requests
├── context.d/           # Scripts that add context sections (optional)
└── hooks/               # Lifecycle hooks (optional)
    ├── pre-run          # Before each iteration
    ├── post-context     # After context assembly (stdin: context, stdout: modified)
    ├── post-llm         # After LLM completes ($1: exit code)
    └── post-commit      # After git commit ($1: timestamp)

How It Works

Each loop iteration:

1. Wake — Lock acquired, context assembled from memory + goals + pending actions
2. Think — Agent reads its full state and decides what to do
3. Act — Agent executes: writes code, does research, creates plans, requests approvals
4. Learn — Agent updates its memory with what it learned
5. Sleep — Changes committed to git, lock released, agent waits for next iteration

Configuration

# boucle.toml
[agent]
name = "my-agent"
description = "A helpful autonomous agent"

[schedule]
interval = "1h"

[boundaries]
autonomous = ["read", "write", "research", "plan"]
requires_approval = ["spend_money", "post_publicly", "contact_people"]

[memory]
backend = "broca"
path = "memory/"

[llm]
provider = "claude"
model = "claude-sonnet-4-20250514"

Extension Points

Context Plugins (context.d/)

Executable scripts that inject context into each iteration. Each receives the agent directory as $1 and outputs Markdown to stdout.

#!/bin/bash
# context.d/weather — Add weather to context
echo "## Weather"
curl -s wttr.in/?format=3

Lifecycle Hooks (hooks/)

Hook	When	Arguments	Use case
pre-run	Before iteration	$1: timestamp	Setup, health checks
post-context	After context assembly	stdin: context	Modify/filter context
post-llm	After LLM completes	$1: exit code	Notifications, cleanup
post-commit	After git commit	$1: timestamp	Push to remote, deploy

Tool Restrictions (allowed-tools.txt)

Read
Write
Edit
Glob
Grep
WebSearch
Bash(git:*)
Bash(python3:*)

If this file doesn't exist, all tools are available.

CLI Reference

# Agent management
boucle init [--name <name>]      # Initialize new agent (default: my-agent)
boucle run                        # Run one iteration
boucle run --dry-run              # Preview context without calling LLM
boucle doctor                     # Check prerequisites and agent health
boucle validate                   # Validate config (catches typos, bad values, path issues)
boucle stats                      # Show aggregate loop statistics
boucle status                     # Show agent status
boucle log [--count <n>]          # Show loop history (default: 10 entries)
boucle schedule --interval <dur>  # Set up scheduled execution (e.g., 1h, 30m, 5m)
boucle plugins                    # List available plugins

# Self-observation
boucle signal <type> <summary> <fingerprint>  # Log a signal (friction/failure/waste/surprise)
boucle improve run [--budget <secs>]          # Run the improvement pipeline
boucle improve status                         # Show patterns, scores, pending actions
boucle improve init                           # Set up improve/ with example harvester

# Memory (Broca)
boucle memory remember <title> <content> [--tags <tags>] [--entry-type <type>]
boucle memory recall <query> [--limit <n>]
boucle memory show <id>
boucle memory search-tag <tag>
boucle memory journal <content>
boucle memory update-confidence <id> <score>
boucle memory supersede <old-id> <new-id>
boucle memory relate <id1> <id2> <relation>
boucle memory stats
boucle memory index
boucle memory gc [--apply]            # Archive stale/superseded entries
boucle memory consolidate [--apply]   # Merge near-duplicate entries

# MCP server
boucle mcp --stdio               # stdio transport
boucle mcp --port <port>         # HTTP transport

# Global options
boucle --root <path>             # Use specific agent directory
boucle --help                    # Show help
boucle --version                 # Show version

Design Principles

1. Files over databases. Memory is Markdown. Config is TOML. Logs are plain text. Everything is human-readable and git-diffable.

2. Boundaries are features. Approval gates make autonomous agents trustworthy. An agent that can spend your money without asking isn't autonomous, it's dangerous.

3. Compound knowledge. Every iteration should leave the agent smarter. Memory isn't a cache — it's an investment.

4. Transparency by default. If you can't see what the agent did and why, something is wrong.

Platform Support

	macOS	Linux	Windows (WSL)	Windows (native PS7)
bash-guard	Yes	Yes	Yes	Yes (.ps1)
git-safe	Yes	Yes	Yes	Yes (.ps1)
file-guard	Yes	Yes	Yes	Yes (.ps1)
read-once	Yes	Yes	Yes	Yes (.ps1)
branch-guard	Yes	Yes	Yes	Yes (.ps1)
worktree-guard	Yes	Yes	Yes	Yes (.ps1)
session-log	Yes	Yes	Yes	Yes (.ps1)
enforce-hooks	Yes	Yes	Yes (bash)	WSL or Git Bash
safety-check	Yes	Yes	Yes	Partial (bash required)
Installer	install.sh	install.sh	install.sh	install.ps1
Hook reliability	Full	Full	Full	~18%

Best experience: macOS or Linux. Windows: Use WSL for full reliability. Native PowerShell hooks work but Claude Code fires them inconsistently (#37988).

Recommended Claude Code Version

Use v2.1.90 or later. Earlier versions have hook-related regressions:

Version	Issue
v2.1.90	Current recommended. Improves exit-2 + JSON blocking behavior, fixes PostToolUse format-on-save, and fixes 4 PowerShell permission bypasses
v2.1.89	Pulled from npm — custom commands broken, systemMessage display broken, source map leak
v2.1.81-84	Permission bypass resets mid-session when PreToolUse hooks are installed
< v2.1.50	No hookSpecificOutput format support (deprecated decision: "block" still works but should be migrated)

Run claude --version to check. Run safety-check with --verify to confirm hooks fire correctly on your version.

Troubleshooting

JSONC comments in settings.json: If your ~/.claude/settings.json contains // or /* */ comments, hooks may silently stop working (claude-code#37540). Our installers detect JSONC and automatically strip comments (creating a .bak backup). If hooks aren't firing, check for comments in your settings file.

Hooks not blocking: Claude Code only fires hooks on tool calls, not on prompt assembly. Features like @-autocomplete inject file content before hooks can intercept. See claude-code#32928.

Permission bypass resets with hooks installed: If you use --dangerously-skip-permissions (common in autonomous setups), PreToolUse hooks can cause the permission state to reset mid-session, reverting all tools to manual approval. This is a platform bug, not a hooks bug. If tools suddenly require approval 30-120 minutes into a session, this is why.

IS_DEMO environment variable disables all hooks: If IS_DEMO=1 is set in your environment (sometimes via IDE or cloud workspace settings), Claude Code silently skips all hook execution by suppressing workspace trust without granting it. Run echo $IS_DEMO to check. Our safety-check tool detects this automatically.

CLAUDE_CODE_SIMPLE disables all hooks: When the CLAUDE_CODE_SIMPLE environment variable is set (to true or 1), Claude Code disables hooks, MCP tools, attachments, and CLAUDE.md file loading entirely (introduced in v2.1.50). No enforcement rules will fire. Run echo $CLAUDE_CODE_SIMPLE to check. Our safety-check tool detects this automatically.

--bare flag skips all hooks: The --bare CLI flag disables hooks, LSP, plugin sync, and skill directory walks for scripted -p calls. If your autonomous pipeline uses claude --bare -p, no hooks fire. Use OS-level controls (file permissions, containerization) for enforcement in bare mode.

Hook deny handling is still inconsistent across tools and versions: hookSpecificOutput.permissionDecision: "deny" has improved, but it is not a universal guarantee across Claude Code surfaces. Several upstream issues still document cases where deny handling is ignored or changes by tool/event type. That is why framework hooks that must hard-block dangerous actions use the most conservative path Claude Code currently respects most reliably: a human-readable reason on stderr plus exit 2, then we tell users to run safety-check --verify after install and after Claude Code updates. If you write custom hooks, do not assume a JSON deny response alone is enough just because it works in one local test.

Subagents may skip hook settings: Agents spawned via the Agent tool don't consistently inherit permission settings. Hooks in .claude/settings.json should still fire (shared config), but verify hook behavior when using subagent workflows.

Hook stderr may leak your filesystem paths: Claude Code's hook runner prefixes stderr output with the raw command path, exposing details like /Users/yourname/.claude/hooks/my-hook.sh in the conversation. This comes from the platform's execution layer, not from the hooks. Our hooks use clean prefixes ([bash-guard], [file-guard], etc.) for debug messages and never expose filesystem paths in either stdout or stderr. Debug logging is opt-in per hook (e.g., BASH_GUARD_LOG=1).

Internal git operations bypass all hooks: Claude Code runs background git operations (fetch + reset) programmatically every ~10 minutes without spawning an external git binary or making a tool call. Since hooks only fire on tool calls, git-safe and all other hooks are blind to these operations. This can silently destroy uncommitted changes to tracked files. Workaround: use git worktrees (immune) or commit frequently.

Permissions desync after editing settings.local.json: If Claude's Edit tool modifies .claude/settings.local.json during a session, the in-memory permission state desyncs from the file on disk. Allow rules stop working and the user is repeatedly prompted for commands that are already permitted. The file on disk is correct; the problem is the in-memory cache. Workaround: let Claude Code manage permission files through its own prompt mechanism, or restart the session after manual edits.

New in v2.1.89: PermissionDenied hook event: A new hook event fires after auto mode classifier denials. Hooks can return {"retry": true} to tell the model it can retry the denied operation. This is not yet in the official docs. Also in v2.1.89: hooks if conditions now match compound Bash commands (ls && git push matches Bash(git *)) and commands with env-var prefixes (FOO=bar git push).

SessionStart systemMessage not displayed (v2.1.89): The systemMessage field returned by SessionStart hooks is no longer rendered in the terminal. The hook runs and additionalContext is still injected into model context, but the visual output that previously appeared (e.g. "SessionStart:startup says: ...") is silently missing. If you rely on systemMessage for operator notifications or session identification, the output will not be visible. Related: #9090, #15344.

Hooks fail on first session in a new project: On the very first session in a project directory, SessionStart and UserPromptSubmit hooks fire before the project directory exists (~/.claude/projects/<encoded-path>/). Any hook that derives file paths from transcript_path and tries to write there will fail. Workaround: add mkdir -p for transcript_path-derived paths before writing.

Model self-execution in long sessions: In long unattended sessions, the model can hallucinate Human: text after task-notification delivery and then execute it as if it were a real user request, triggering unauthorized git operations and file modifications. Hooks cannot detect this because the resulting tool calls are genuine — only the trigger is hallucinated. Mitigation: use session time limits and avoid very long unattended sessions.

Worktree GIT_INDEX_FILE leak: Agents spawned via EnterWorktree can have their git index corrupted by marketplace plugin entries due to GIT_INDEX_FILE environment variable leaking across process boundaries. If worktree operations show unexpected files in git status, this may be the cause.

Background agents cannot be stopped: Agents spawned via the Agent tool with run_in_background cannot be reliably terminated by the user. In one reported case, 14 parallel agents wrote to the same file and consumed ~1.4M tokens ($55-106). There is no built-in kill mechanism. Mitigation: avoid spawning many background agents, and monitor token usage if you do.

cleanupPeriodDays setting may be ignored: The cleanupPeriodDays setting in settings.json can be silently bypassed, deleting session files even when set to very high values. One user lost 490 sessions despite setting it to 99999. If you rely on session persistence, back up ~/.claude/projects/ independently.

Symlinked .claude/ directories not discovered (Linux): Slash commands from symlinked .claude/commands/ are not loaded on Linux (regression). This is a common team pattern (store shared config in a central directory and symlink). Hooks and skills may also fail if .claude/ itself is a symlink. Workaround: copy files instead of symlinking.

Bundled ripgrep missing execute permission (Linux): The bundled rg binary can lose its execute permission on Linux, silently breaking all user-defined slash commands in ~/.claude/commands/. Fix: chmod +x the bundled binary.

v2.1.89 pulled from npm: This version was pulled due to multiple regressions: custom commands in .claude/commands/ are not discovered, SessionStart systemMessage display is broken, and the cli.js.map file was accidentally shipped. If you are on v2.1.89, downgrade to v2.1.87 or wait for the next release.

Non-interactive sessions hang on usage limit: In headless, --print, or remote-control mode, hitting a usage limit shows a confirmation prompt that cannot be answered because there is no stdin. The session hangs permanently. There is no programmatic workaround (#41503). If you run Claude Code in CI, cron, or autonomous loops, set session time limits and monitor for stuck processes.

Deny rules bypassed by pipes and compound commands: Built-in deny rules only match the full command string. Bash(rm *) blocks rm -rf / but not find /foo | xargs rm or something && rm -rf /. The docs say allow rules parse shell operators, but deny rules do not. Note: hooks if conditions were fixed upstream (late March 2026) to correctly match compound commands and env-var prefixes, so hooks fire correctly for these patterns. The gap is specifically in deny rules, not hooks. bash-guard parses each pipe segment and compound chain independently, catching these bypass patterns. See also #37662, #16180.

"Confirm each change individually" silently skipped: When exiting plan mode and selecting "confirm each change individually," changes apply without any prompt if the tools (Edit, Write, Bash) are in permissions.allow. The persistent allow rules override the user's explicit per-session choice. Workaround: remove broad tool allows and use hooks for enforcement instead.

SessionEnd hooks killed before completion: SessionEnd hooks that perform async work (API calls, LLM summarization, network requests) are killed mid-execution when Claude Code exits, regardless of configured timeout. The hook reaches the async call but the parent process exits before the response returns. Workaround: detach heavy work into a background process with nohup ... & disown, then exit 0 immediately.

"Always allow" directory access not persisted: Clicking "Yes, and always allow access to [folder]" does not reliably save. Claude re-prompts for the same directory in subsequent sessions. Adding to additionalDirectories in settings.json is also flaky. Related to #40606 (additionalDirectories leak across projects).

~/.claude/ writes stall automated sessions: Writes to paths under ~/.claude/ trigger a hardcoded sensitive-file prompt that cannot be suppressed by permissions.allow, PreToolUse hooks returning "allow", bypassPermissions mode, or skipDangerousModePermissionPrompt. Automated sessions (tmux, CI, autonomous loops) that need to modify Claude Code config files will hang on the interactive prompt. Workaround: use Bash tool commands (echo, cat, jq) to write files directly instead of the Edit/Write tools.

bash -c wrapping bypasses .claude/ directory write protection: The permission system protects .claude/ files from modification (edit, write, direct bash commands all trigger a confirmation modal). But wrapping the command in bash -c 'echo "..." >> .claude/file' bypasses the check entirely: no modal, write succeeds silently. The pattern matching inspects the top-level command string but not nested subshells. This means a model (or prompt injection via tool results) could modify hooks, settings.json, skills, or agent definitions without user confirmation. file-guard catches this if .claude/ is added to the .file-guard config, because it scans the full Bash command string for protected paths (including inside quoted arguments). Users relying solely on built-in permission checks are exposed.

WorktreeCreate hooks hang claude -w: Any WorktreeCreate hook causes claude -w to hang indefinitely, regardless of hook content. Even echo ok < /dev/null freezes the session. The hook executes and returns, but Claude Code never proceeds. Remove all WorktreeCreate hooks if you need claude -w to function.

Model can manipulate hook state files: The model has filesystem access and can overwrite files that hooks depend on — checkpoint files, lock files, counters. In one documented case, Claude computed the SHA256 hash of a checkpoint filename and wrote a fresh timestamp to bypass a content-read-gate. Hooks that rely on external state files for enforcement should assume the model can read and modify those files. Cryptographic signatures or out-of-process validation can mitigate this.

bypassPermissions not restored on session resume (VS Code): When bypassPermissions is configured via initialPermissionMode in VS Code settings, resumed conversations revert to default permission mode and prompt for every edit. New sessions may pick it up, but resumed sessions consistently fail. Hooks that depend on the session running in bypass mode cannot rely on it persisting across resume.

Worktree isolation breaks in git submodules: Using isolation: "worktree" on the Agent tool inside a git submodule creates the worktree in .git/modules/<path>/.claude/worktrees/ instead of the project's own .claude/worktrees/. This places the agent outside the project's permission scope, causing bypassPermissions to be silently downgraded and triggering unexpected permission prompts.

Skill approval not tied to content hash: When a user approves a skill, the approval is not anchored to the file's content hash. If the skill file is modified after approval (even mid-session), the modified version executes without re-prompting. Additionally, approving a skill can bypass tool-level deny rules in settings.json. This is a supply chain risk: anything with write access to ~/.claude/skills/ can escalate capabilities post-approval.

Stdio MCP servers never auto-reconnect: When a stdio-type MCP server process dies or disconnects, Claude Code marks it as failed and never retries. HTTP/SSE/WebSocket servers get automatic reconnection with exponential backoff (5 attempts), but stdio servers are explicitly excluded. Users must manually run /mcp to reconnect. This affects any MCP integration using stdio transport (the most common local pattern).

Plan mode bypass after first cycle: After completing one plan-approve-implement cycle, entering plan mode again does not reliably enforce read-only restrictions. Claude carries over the "approved" mental state and begins editing files before the user approves the new plan. Hooks that rely on plan mode as a safety boundary cannot trust it across multiple cycles in the same session.

Windows: All seven hooks have native PowerShell 7+ equivalents (hook.ps1) that require no external dependencies. Requires PowerShell 7 (pwsh), not the built-in Windows PowerShell 5. Install them with:

iex "& { $(irm https://raw.githubusercontent.com/Bande-a-Bonnot/Boucle-framework/main/tools/install.ps1) } all"

Or configure manually in .claude/settings.json with "command": "pwsh -File /path/to/hook.ps1". The enforce-hooks tool is a bash script that works from a WSL terminal or with Git for Windows (which provides /usr/bin/bash). Note: Claude Code has a known bug where hooks fire only ~18% of the time on Windows, so hook reliability is limited on native Windows regardless of shell. WSL remains the most reliable option. See #3.

Development

cargo test           # Framework tests (198 passing)
cargo fmt            # Format code
cargo clippy         # Run linter

# Hook tests (run individually)
bash tools/read-once/test.sh
bash tools/file-guard/test.sh
bash tools/git-safe/test.sh
bash tools/bash-guard/test.sh
bash tools/branch-guard/test.sh
bash tools/session-log/test.sh
bash tools/enforce/test.sh
bash tools/safety-check/test.sh
bash tools/worktree-guard/test.sh

Status

v0.12.0 — 198 Rust tests + ~1758 hook tests (bash + PowerShell). Zero clippy warnings. CI on Ubuntu + macOS + Windows. Docker support.

New in v0.12.0: Searchable Known Limitations page (now 932 entries). test-hook.sh dry-run tester. file-guard v2.1.89 absolute path fix. safety-check 3 silent failure fixes. Installer UX improvements. README navigation banner + platform matrix. See CHANGELOG for details.

64 stars, 4 external contributors, 4 forks.

Contributing

Contributions welcome. Please open an issue first to discuss what you'd like to change.

License

MIT