.claude/ -- Claude Code Configuration

This directory configures Claude Code for the AIPass project.

Related: DPLAN-0184 (Hook Migration), DPLAN-0053 (original hook architecture research).

How Hooks Work (Post-Migration)

All AIPass hooks run through a three-layer pipeline:

~/.claude/settings.json      Provider settings (Claude Code reads these)
        |
        v
claude.py (bridge)           Thin entry point -- normalizes stdin, calls engine
        |
        v
engine.py (dispatcher)       Reads .aipass/hooks.json, imports + calls handlers
        |
        v
handlers/                    Native Python handlers (the actual hook logic)

Provider settings in ~/.claude/settings.json call the bridge with an event type:

{
  "type": "command",
  "command": "$AIPASS_HOME/.venv/bin/python3 $AIPASS_HOME/src/aipass/hooks/apps/handlers/bridges/claude.py PreToolUse"
}

The bridge supports two invocation forms:

claude.py EventType -- dispatch ALL enabled hooks for that event
claude.py EventType:hook_name -- dispatch ONLY one specific hook (used for UserPromptSubmit where each hook needs its own system-reminder block)

Per-project configuration lives in .aipass/hooks.json. Each hook entry specifies:

enabled -- whether the hook fires
handler -- dotted import path to the handler function
matcher -- tool name filter (empty string = match all)
timeout -- optional timeout in seconds

Quick Setup

Run setup.sh from the repo root. It creates the venv, installs the package, and wires bridge entries into ~/.claude/settings.json automatically.

./setup.sh

If hooks get out of sync, aipass doctor --fix can auto-wire missing hook entries.

No manual script copying is needed. No global_hooks directory. No git rev-parse tricks.

What's In This Directory

.claude/
├── settings.json              # Project settings (permissions, env vars)
├── hooks/                     # Legacy hook scripts (all disabled) + testing tools
│   ├── *.py(disabled)             # 18 disabled scripts (pre-migration)
│   ├── hook_log.py                # Shared logger -- hooks call run_and_log()
│   ├── hook_report.py             # Report tool -- reads JSONL log, shows table
│   ├── hook_test.py               # Test harness -- direct + integration tests
│   └── probes/                    # Opt-in per-event diagnostic probes
├── agents/                    # Agent definitions
│   └── builder.md
├── commands/                  # Slash commands
│   └── memo.md                    # /memo -- memory update workflow
├── sounds/                    # Audio files for sound hooks
└── README.md                  # This file

Hook logic has moved to src/aipass/hooks/apps/handlers/. See the handler README for the full layout.

Handler Layout

All 14 hooks are native Python handlers organized by domain:

src/aipass/hooks/apps/handlers/
├── bridges/
│   └── claude.py              # Provider bridge (called from settings.json)
├── config/
│   ├── loader.py              # Finds and reads .aipass/hooks.json
│   └── diagnostics.py         # JSONL logging for hook execution
├── prompt/
│   ├── global_loader.py       # UserPromptSubmit -- AIPass global prompt
│   ├── branch_loader.py       # UserPromptSubmit -- branch-specific prompt
│   └── identity.py            # UserPromptSubmit -- passport identity injection
├── notification/
│   ├── email.py               # UserPromptSubmit -- unread email count
│   ├── tool_sound.py          # PreToolUse -- key-press sound
│   ├── stop_sound.py          # Stop -- achievement bell
│   └── announce.py            # Notification -- notification sound
├── security/
│   ├── git_gate.py            # PreToolUse -- blocks raw git/gh commands
│   ├── edit_gate.py           # PreToolUse -- cross-branch write block
│   └── subagent_gate.py       # SubagentStop -- seedgo checklist gate
└── lifecycle/
    ├── auto_fix.py            # PostToolUse -- pyright + ruff after edits
    ├── auto_watchdog.py       # PostToolUse -- watchdog reminder after dispatch
    ├── compact.py             # PreCompact -- save context before compaction
    └── rollover.py            # PreCompact -- memory rollover on compaction

What Gets Injected Every Turn

Global Prompt -- system context, terminology, commands, rules (.aipass/aipass_global_prompt.md)
Branch Prompt -- branch-specific instructions based on CWD (.aipass/aipass_local_prompt.md)
Identity -- passport summary: role, traits, purpose (.trinity/passport.json)
Email -- notification only if unread mail exists (.ai_mail.local/inbox.json)

Each is dispatched as a separate UserPromptSubmit:hook_name call so it gets its own system-reminder block.

Project Settings

Defined in settings.json (this directory). These fire from subdirectories.

Environment:

CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1 -- makes PostToolUse hooks fire inside subagents
AIPASS_HOME -- repo root path, used by bridge commands

Permissions:

Denied: git reset, git rebase, git config, git push --force, EnterPlanMode
Default mode: acceptEdits

Adding a New Hook

Create a handler in src/aipass/hooks/apps/handlers/<domain>/your_hook.py with a handle(event_type, stdin_data, config) function
Add an entry to .aipass/hooks.json under the appropriate event type
If the hook needs its own system-reminder output (like prompt injectors), add a separate bridge entry in ~/.claude/settings.json using the EventType:hook_name form
Run setup.sh or aipass doctor --fix to sync provider settings

Architecture Notes

Why provider settings? Claude Code project settings (.claude/settings.json) do not fire UserPromptSubmit hooks from subdirectories. Since AIPass citizens launch from src/aipass/{name}/, prompt injection must live in provider settings (~/.claude/settings.json). The bridge pattern makes this clean -- one bridge binary, many handlers.

Why separate bridge calls for UserPromptSubmit? Each UserPromptSubmit hook entry gets its own system-reminder block in the conversation. Bundling them into one call would merge all prompt output into a single block, losing separation.

Why .aipass/hooks.json? Decouples hook configuration from provider settings. The engine reads this at dispatch time, so hooks can be enabled/disabled without editing ~/.claude/settings.json.

6.5 KiB Raw Blame History