claude-code - 💡(How to fix) Fix [BUG] ralph-loop Stop hook silently fails on Windows after Claude Code 2.1.126 (worked on 2.1.119) [1 comments, 2 participants]

Preflight Checklist

• I have searched existing issues and this hasn't been reported yet
• This is a single bug report (please file separate reports for different bugs)
• I am using the latest version of Claude Code

What's Wrong?

The ralph-loop plugin Stop hook fires correctly on Claude Code 2.1.119 but silently fails to fire on Claude Code 2.1.126 (Windows + Git Bash). The hook script itself is valid — when invoked manually it returns the correct decision: block JSON. But Claude Code 2.1.126 doesn't appear to invoke it at end-of-turn anymore.

This effectively kills the plugin on the latest Windows builds — the loop runs exactly 1 iteration and goes idle waiting for a hook that never fires.

Evidence — same machine, same plugin, only Claude Code version differs

Counted hook execution events (type:"attachment" + attachment.type:"hook_system_message" + hookEvent:"Stop") across JSONL transcripts:

The session that broke shows the agent ending iter 1 normally, then nothing:

• 15:52:19.316 UTC — assistant text "Iter 1 done — commit f7066f9"
• 15:52:19.393 UTC — system/turn_duration event (turn ended)
• 15:55:22.817 UTC — system/away_summary event (3 min later, session went idle)
• No hook_system_message event between turn end and away
• No decision:"block" reinjection
• 18 minutes of idle until the user typed "continue manually"

Manual hook invocation works fine

Running stop-hook.sh directly with simulated stdin input returns the correct JSON:

{
  "decision": "block",
  "reason": "<original prompt>",
  "systemMessage": "🔄 Ralph iteration 2 | To stop: output <promise>DONE</promise>..."
}

The iteration: field in .claude/ralph-loop.local.md increments correctly. So the script is healthy — Claude Code 2.1.126 just isn't dispatching the Stop event to it on Windows builds.

What I ruled out

• Not the known bugs from issue #394 — no cd in agent commands; state file not deleted; relative path resolves correctly because cwd stayed at project root throughout the run
• Not a missing utility — bash, jq, perl, sed, awk, grep, cat all present in Git Bash 2.49
• Not away_summary itself — sessions on 2.1.119 also produced away_summary events alongside hook events (10 away_summary in one atlas session, coexisting with 175 hook events)
• Not lifecycle/permissions/settings — same ~/.claude/settings.json for both working (atlas, 2.1.119) and broken (guberman3, 2.1.126) sessions

About me

I'm a software engineer building agent systems on Claude Code. Happy to provide more JSONL artifacts, transcripts, or repro environment if useful for triage.

What Should Happen?

After the agent ends a turn while a ralph-loop.local.md state file is active in the project root:

1. Claude Code dispatches the Stop event to registered plugin hooks
2. The plugin's stop-hook.sh script executes with the standard stdin payload (session_id, transcript_path)
3. The script reads the state file, validates iteration < max_iterations, checks the last assistant message for <promise>COMPLETION_PROMISE</promise>, and either:
   • Allows exit if max_iterations reached or completion promise emitted, OR
   • Returns {"decision": "block", "reason": "<original prompt>", "systemMessage": "🔄 Ralph iteration N..."} to re-feed the original prompt and increment the iteration counter
4. Claude Code re-injects the reason field as a new user message, surfaces the systemMessage to the user, and the agent proceeds with iter N+1

This is the documented behavior of the plugin and matches what was observed on Claude Code 2.1.119, where a single overnight session produced 175 hook events (one per Stop), allowing the loop to run continuously for hours.

On Claude Code 2.1.126, none of the above happens — hook_system_message events are entirely absent from the JSONL transcript, even though the state file is present, the script is executable, and manual invocation produces the correct output.

Suggestions

1. If /loop dynamic mode + ScheduleWakeup is the intended successor (it's a great design — works without bash, context accumulates, Anthropic-maintained), consider deprecating ralph-loop officially in the plugin README with a migration note pointing at /loop. The plugin still installs and runs setup successfully, which misleads users into thinking the issue is their setup.

2. If ralph-loop is meant to keep working, investigate why the Stop event isn't dispatched to plugin hooks on Claude Code 2.1.126 Windows builds. Possible angles: change in event lifecycle around away_summary, change in plugin hook registration timing, change in how Stop fires when bypassPermissions mode is active.

3. Either way, surface a clear log/error when a hook is registered but not invoked — silent failure is the worst-case UX. A debug log entry like [hook] Stop event fired, dispatching to plugin X would have made this trivially diagnosable.

Error Messages/Logs

Steps to Reproduce

Environment

• OS: Windows 11 Home Single Language 10.0.26200
• Shell: PowerShell + Git Bash 2.49
• Claude Code (broken): 2.1.126 (native Windows installer, single binary at ~/.local/bin/claude.exe)
• Claude Code (works): 2.1.119 (same machine, same plugin, prior version)
• Plugin: [email protected] from claude-plugins-official
• Plan: Max
• ~/.claude/settings.json excerpt:

  {
    "model": "opus[1m]",
    "permissions": { "defaultMode": "bypassPermissions" },
    "alwaysThinkingEnabled": true,
    "skipDangerousModePermissionPrompt": true
  }

Reproduction steps

1. Install Claude Code 2.1.126 for Windows (native installer)

2. Install Git for Windows (provides Git Bash with bash/jq/perl/sed/awk/grep)

3. Install the plugin:

   claude plugin install ralph-loop@claude-plugins-official

4. Confirm install:

   claude plugin list

   Expect: ralph-loop@claude-plugins-official Version: 1.0.0 Status: ✔ enabled

5. Create a test git project:

   mkdir test-ralph && cd test-ralph
   git init
   echo "iter 0" > log.md
   git add log.md && git commit -m "init"

6. Start Claude Code in that directory and run:

   /ralph-loop:ralph-loop "Append 'iter N done' to log.md, commit, then end your turn." --max-iterations 5 --completion-promise "ALL_DONE"

7. Observed: agent appends iter 1, commits, ends turn. Session goes idle. After ~3 min, system/away_summary appears in the JSONL transcript. No further iterations occur. Loop is dead.

8. Verify the hook script itself works (proves the script is healthy, the issue is dispatch):

   # In Git Bash, from the project root:
   echo '{"session_id":"<paste from JSONL>","transcript_path":"<paste path to current session JSONL>"}' \
     | bash "$HOME/.claude/plugins/marketplaces/claude-plugins-official/plugins/ralph-loop/hooks/stop-hook.sh"

   Expected: returns valid JSON with "decision":"block" and the prompt re-fed in reason. Actual: returns valid JSON correctly. So the script works — Claude Code just isn't calling it.

9. Compare to working version (if you have access to 2.1.119):

   • Same project, same plugin, same .claude/ralph-loop.local.md
   • Run the same /ralph-loop:ralph-loop command
   • Hook fires after each turn; hook_system_message events appear in JSONL; loop iterates correctly

Files to inspect

• ~/.claude/projects/<encoded-cwd>/<session-id>.jsonl — transcript. Search for hook_system_message (zero matches on 2.1.126 broken sessions; many on 2.1.119 working sessions).
• ~/.claude/plugins/marketplaces/claude-plugins-official/plugins/ralph-loop/hooks/stop-hook.sh — the script that should be invoked but isn't.
• <project>/.claude/ralph-loop.local.md — state file. Verify iteration: stays at 1 (not incremented) and session_id: may be empty if CLAUDE_CODE_SESSION_ID env var isn't set during setup.

Claude Model

Opus

Is this a regression?

Yes, this worked in a previous version

Last Working Version

2.1.119

Claude Code Version

2.1.126

Platform

Anthropic API

Operating System

Windows

Terminal/Shell

PowerShell

Additional Information

No response