* Fix hooks failing silently on Windows due to python3 not found
Add `|| python` fallback to all hook commands in hooks.json. On Windows,
`python3` either doesn't exist or resolves to a Microsoft Store stub that
fails silently, breaking all plugin hooks.
Fixes#20
* Bump version to 0.8.2
In default mode, after the memory-updater processes files and SubagentStop
clears dirty-files, a git commit via Bash would re-trigger PostToolUse,
extract committed files via git diff-tree, and re-add them to dirty-files.
This caused the Stop hook to fire again for already-processed files.
Add early-return guard so default mode skips git commit processing (files
are already tracked via Edit/Write hooks). Also fix existing test to run
commit context enrichment in gitmode where it belongs.
Merge stop.py into trigger.py to handle PreToolUse, Stop, and
SubagentStop hooks via a single entry point routed by hook_event_name.
Add PreToolUse hook for gitmode commit gating, SubagentStop hook for
automatic dirty-files cleanup, and update tests and docs accordingly.
- Update dirty-files path references to .claude/auto-memory/dirty-files
- Clarify commit context is stored inline, not in separate JSON file
- Fix Data Files table with correct locations
- Update test docstring with correct path
The /auto-memory:init command was generating CLAUDE.md files that didn't
follow the defined templates - using wrong marker format like
"BEGIN AUTO-MANAGED" instead of "AUTO-MANAGED:" and wrong section names.
Root cause: The skill referenced templates with @templates/... import syntax
which provides context but not enforcement. The algorithm lacked explicit
instructions about marker format and required sections.
Changes:
- Add Marker Format section to guidelines.md as shared reference
- Update codebase-analyzer description to third-person with trigger phrases
- Add explicit Marker Syntax section with format and anti-patterns
- Add Section Definitions tables for root and subtree CLAUDE.md
- Rewrite Algorithm Step 7 with prescriptive generation instructions
- Format code files with ruff
Document that CLAUDE.md should never include ephemeral data:
- Version numbers
- Test counts/coverage percentages
- Progress metrics
- Dates/timestamps
- Line counts/file sizes
These become stale after each commit and waste tokens.
- Consolidate commit-context.json into dirty-files with inline format:
/path/to/file [hash: commit message]
- Switch memory-updater agent from haiku to sonnet (fixes#7)
- Add /auto-memory:sync command for manual file changes (fixes#5)
- Add configurable trigger modes: default vs gitmode (fixes#6)
- Update tests for new paths and model expectation
- Stop hook now instructs Claude to read CLAUDE.md after memory-updater
completes, giving the main session fresh memory immediately
- Added `from __future__ import annotations` for Python 3.8 compatibility
(fixes TypeError: 'type' object is not subscriptable)
- Standardized "memory" terminology across commands (matches official
Claude Code docs and plugin name)
- Fixed tests for flat commands/ directory structure
- Removed name field requirement from command YAML validation
- Rename commands: memory-{init,calibrate,status}.md -> {init,calibrate,status}.md
- Update internal reference to use /auto-memory:calibrate namespace
- Update CLAUDE.md architecture section
- Bump version to 0.4.0 in pyproject.toml and plugin.json
The command namespace approach (commands/auto-memory/) didn't work as
expected. Reverted to flat command structure (commands/memory-*.md)
which Claude Code discovers automatically by convention.
- Rename commands to shorter names (init, calibrate, status)
- Move to commands/auto-memory/ for proper namespacing
- Commands now invoked as /auto-memory:init, etc.
- Update README.md and CLAUDE.md documentation
- Update tests for new paths
- Change "zero tokens" to "minimal tokens" for accuracy
- Emphasize that isolated agent doesn't consume main context window
- Update badge and comparison table
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
