install.sh fills Claude Code sandbox tmpfs, causing session crashes (ENOSPC)

[dgdev25]

Summary

After running the ruflo-machine-ref install, Claude Code sessions start crashing repeatedly — the process stops completely and requires a full restart. Investigation traced three root causes introduced by the install.

Root Cause 1: Daemon processes fill the Claude Code sandbox tmpfs

Claude Code runs subprocess outputs through a size-limited sandbox tmpfs at /private/tmp/claude-501/<project>/tasks. The install starts ruflo daemon processes (via ruflo-setup-project → ruflo daemon start) that spawn background workers (audit, optimize, ultralearn, deepdive, consolidate, etc.) on 30–60 minute schedules. Combined with the existing per-prompt and per-Bash hooks in .claude/settings.json, the volume of subprocess I/O fills the sandbox to 0 MB free.

Exact error seen in Claude Code:

the temp filesystem at /private/tmp/claude-501/-Users-lyle-dev-SynOrca/<uuid>/tasks
is full (0MB free). The child process's stdout/stderr writes failed with ENOSPC.

This is what causes the session crash — Claude Code cannot write hook/tool output and terminates.

Root Cause 2: statusline.cjs injection runs sqlite3 queries on every render (no timeout)

install.sh injects a rufloActivationSegments() block into the project's .claude/helpers/statusline.cjs (between /* ruflo-seg:BEGIN */ and /* ruflo-seg:END */ markers) and appends a call to it on the statusline output line:

console.log(generateStatusline() + rufloActivationSegments(process.cwd()));

This function runs up to 5 sqlite3 subprocess calls (each with a 1500ms timeout) on every statusline render. The project's settings.json configures the statusline with no timeout:

"statusLine": { "type": "command", "command": "node .claude/helpers/statusline.cjs" }

If sqlite3 is slow or the database is locked, this hangs indefinitely.

Root Cause 3: daemon.autoStart: true in settings.json

The project settings.json has claudeFlow.daemon.autoStart: true, meaning the daemon restarts automatically on every Claude Code session open. This ensures Root Cause 1 re-occurs after every restart.

Steps to Reproduce

1. Clone a project that uses claude-flow V3 hooks (.claude/settings.json with UserPromptSubmit, PreToolUse, PostToolUse hooks)
2. Run ruflo-machine-ref install
3. Open Claude Code in that project
4. Use it for a while — sessions will crash within minutes

Fix Applied

# Kill daemon processes
kill <daemon-pids>

# Remove ruflo-seg injection from statusline.cjs
python3 -c "
import re
with open('.claude/helpers/statusline.cjs') as f: c = f.read()
c = re.sub(r'/\* ruflo-seg:BEGIN \*/.+?/\* ruflo-seg:END \*/\n', '', c, flags=re.DOTALL)
c = c.replace('generateStatusline() + rufloActivationSegments(process.cwd())', 'generateStatusline()')
with open('.claude/helpers/statusline.cjs', 'w') as f: f.write(c)
"

# Disable daemon autoStart
python3 -c "
import json
with open('.claude/settings.json') as f: s = json.load(f)
s['claudeFlow']['daemon']['autoStart'] = False
with open('.claude/settings.json', 'w') as f: json.dump(s, f, indent=2)
"

# Set CLAUDE_CODE_TMPDIR to main filesystem in ~/.claude/settings.json
# \"env\": { \"CLAUDE_CODE_TMPDIR\": \"/Users/<you>/tmp/claude-code\" }

Suggestions

• install.sh should not inject into statusline.cjs without a configurable timeout guard, or should skip the injection if the statusline has no timeout configured.
• The daemon should default to autoStart: false — it should be opt-in.
• The install should document the CLAUDE_CODE_TMPDIR workaround for users with many hooks.
• Consider whether starting a daemon with 10 background workers is appropriate as a side effect of ruflo-setup-project.

Environment

• macOS Darwin 25.5.0
• Claude Code (claude-sonnet-4-6)
• ruflo-machine-ref installed 2026-05-29