Windows compatibility: 4 bugs prevent auto-save pipeline from working

[sixside]

Summary

The remember plugin's auto-save pipeline (PostToolUse hook -> save-session.sh -> Haiku summarization) silently fails on Windows. Four independent bugs combine to make the pipeline completely non-functional. The SessionStart hook and /remember skill work fine since they don't hit the same code paths.

Environment

• Windows 11 Pro
• Claude Code CLI (bash shell via Git Bash)
• Python 3.13
• Plugin version 0.1.0

Bugs

1. save-session.sh and run-consolidation.sh compute wrong PROJECT_DIR

Both scripts use:

PROJECT_DIR="$(cd "$(dirname "$0")/../../.." && pwd)"

This resolves relative to the script's location in the plugin cache (~/.claude/plugins/cache/claude-plugins-official), not the actual project directory. The session-start-hook.sh correctly uses CLAUDE_PROJECT_DIR, but these two scripts don't.

Fix: Use CLAUDE_PROJECT_DIR (and pass it via env vars from the calling hooks since nohup background processes need it explicitly):

PROJECT_DIR="${CLAUDE_PROJECT_DIR:-.}"
PIPELINE_DIR="${CLAUDE_PLUGIN_ROOT:-$(cd "$SCRIPT_DIR/.." && pwd)}"

2. Wrong session directory slug algorithm

All three shell scripts and extract.py use tr '/' '-' (or .replace("/", "-") in Python) to generate the project slug for finding JSONL session files. On Windows, paths use backslashes and colons (C:\Users\...), so this produces C:-Users-primemate-Solutions-_General.

Claude Code actually generates slugs like C--Users-primemate-Solutions--General -- it replaces all non-alphanumeric characters (not just /) with -.

Fix (shell):

SESSION_DIR="$HOME/.claude/projects/$(echo "$PROJECT" | sed 's/[^a-zA-Z0-9-]/-/g')"

Fix (Python, extract.py _session_dir):

import re
slug = re.sub(r'[^a-zA-Z0-9-]', '-', project_dir)
return os.path.expanduser("~/.claude/projects/" + slug)

3. PIPELINE_DIR points to wrong location

save-session.sh and run-consolidation.sh set:

PIPELINE_DIR="${PROJECT_DIR}/.claude/remember"

But the pipeline module is in the plugin cache, not in the project's .claude/remember/ directory. Should use CLAUDE_PLUGIN_ROOT.

4. Python open() calls lack UTF-8 encoding

Windows Python defaults to cp1252 encoding. The JSONL session files contain UTF-8 characters, causing UnicodeDecodeError in extract.py's count_lines(), extract_messages(), and get_last_save_line(). Multiple open() calls in shell.py and prompts.py have the same issue.

Fix: Either add encoding="utf-8" to all open() calls, or set export PYTHONUTF8=1 in the shell scripts before calling Python.

Affected files

• scripts/save-session.sh (bugs 1, 3)
• scripts/run-consolidation.sh (bugs 1, 3)
• scripts/post-tool-hook.sh (bug 2, also needs to pass env vars to nohup)
• scripts/session-start-hook.sh (bug 2, also needs to pass env vars to background calls)
• pipeline/extract.py (bugs 2, 4)
• pipeline/shell.py (bug 4)
• pipeline/prompts.py (bug 4)

Result

After patching all four bugs locally, the full pipeline works: extraction, Haiku summarization, now.md creation, NDC compression to today-*.md, and position tracking via last-save.json.

[fdaviddpt]

Fixed in v0.2.0 (commit 22fd949). All four bugs addressed:

1. ✅ PROJECT_DIR — was fixed in PR Fix PROJECT_DIR path traversal when installed via plugin cache #2 (March 23), but marketplace pins to install-time commit
2. ✅ Path slug — tr '/' '-' replaced with sed 's/[^a-zA-Z0-9]/-/g' across all 4 scripts
3. ✅ PIPELINE_DIR — was fixed in PR Fix PROJECT_DIR path traversal when installed via plugin cache #2 alongside PROJECT_DIR
4. ✅ Python encoding — encoding="utf-8" added to all open() calls (~20 sites)

To update: we've submitted the updated version to Anthropic's marketplace. Once they approve it, run /plugin in Claude Code, find "remember", and reinstall it.

In the meantime, you can install directly from the repo to get the fix now:

1. Uninstall the marketplace version via /plugin
2. Clone https://github.com/Digital-Process-Tools/claude-remember into your project's .claude/remember/ directory
3. Follow the manual setup in the README

Thanks for the thorough diagnosis — you basically wrote the fix for us.

[sixside]

The v0.2.0 commit (22fd949) fixed 3 of the 4 bugs listed here but missed one.

pipeline/haiku.py:72 still has cwd="/tmp" in call_haiku(). On Windows, Python's subprocess.run() doesn't resolve /tmp the way Git Bash does, so it throws NotADirectoryError: [WinError 267]. This kills the entire consolidation pipeline — recent.md, archive.md, and core-memories.md never get generated.

The shell scripts (save-session.sh, run-consolidation.sh) use cd /tmp && which works fine under Git Bash. It's only the Python cwd="/tmp" that breaks.

Fix is one line in haiku.py:

# line 72, change:
cwd="/tmp",
# to:
cwd=os.environ.get("TEMP", os.environ.get("TMPDIR", "/tmp")),

[fdaviddpt]

Fixed the remaining cwd="/tmp" bug in pipeline/haiku.py (commit 618fc72). Now uses tempfile.gettempdir() which returns the platform-appropriate temp directory. All Python /tmp references are resolved — the shell script ones work fine under Git Bash.

[fdaviddpt]

Thanks @sixside for the detailed report — great catch on the Python vs Git Bash distinction. Wouldn't have found this without your analysis.