{Bug} Obsidian inbox processing error

Preflight Checklist

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

What's Wrong?

Title: SessionStart hooks fire mid-task, interrupting long-running skill loops

What happened
While running /pipeline (from the arscontexta/Obsidian-inbox plugin) sequentially across 26 inbox files, the session was reset by SessionStart hook injections at least twice mid-loop. Each reset:

• Re-injected the full using-superpowers skill body and the deferred-tools list
• Re-played the user's original prompt ("run arscontexta on my obsidian inbox") as if it were fresh
• Wiped the active TodoWrite state, forcing me to re-list the inbox and rebuild the todo list before resuming
• Coincided with huggingface and hunter MCP servers flapping disconnect → reconnect

Expected
Long-running skill loops should not be interrupted by SessionStart-style context refreshes, or those refreshes should be additive (preserve in-flight todo + tool state) rather than replaying the user's prompt.

Impact

• Wasted tokens re-establishing context 3× for the same task
• Required manual reconciliation of "which files already moved out of inbox" to avoid double-processing
• User had to re-issue "continue" several times

Environment

• Claude Code, Opus 4.7 (claude-opus-4-7)
• macOS Darwin 25.4.0, zsh
• Plugins active: arscontexta, superpowers, plus several MCP servers (huggingface, hunter, bio-research, zoom, greptile)
• Date: 2026-05-09

Repro hint
Trigger any skill that loops Skill tool invocations across N≥10 files in one turn, with multiple MCP servers connected. The flap + SessionStart replay seems correlated.

What Should Happen?

The processing for the obisidian inbox content to obidian markdown notes with proper tages and headers and deleting inbox notes once done

Error Messages/Logs

Steps to Reproduce

Steps to reproduce

1. Set up Claude Code with the arscontexta plugin (Obsidian vault at ~/Documents/Obsidian Vault/ with an inbox/ folder).
2. Drop ~25–30 markdown files into inbox/ (any size; company-knowledge notes were ~5–20 KB each).
3. Have several MCP servers configured in ~/.claude.json — at minimum huggingface and hunter, plus a few plugin MCPs (bio-research, zoom, greptile) so connection churn is likely.
4. In a fresh Claude Code session, prompt: run arscontexta on my obsidian inbox.
5. When the assistant offers options, choose /pipeline (full intake) and All inbox items.
6. Let it loop Skill(pipeline, <file>) calls one after another. Each pipeline invocation takes ~30–90s and forks subagents.
7. Within the first 2–3 files, observe: a SessionStart hook fires, MCP servers (huggingface, hunter) disconnect/reconnect, and the user's original prompt is replayed as if new. TodoWrite state is gone.

Frequency
Hit it 2 out of 2 long pipeline runs in the same session — appears deterministic when MCP flap occurs during a multi-minute Skill loop.

Workaround used
Re-ls the inbox each time, diff against the original 29-file list to figure out what was already processed, rebuild the todo, and resume. No data loss because /pipeline archives each source file as it completes.

Claude Model

Opus

Is this a regression?

Yes, this worked in a previous version

Last Working Version

_No response_

Claude Code Version

opus 4.7

Platform

Anthropic API

Operating System

macOS

Terminal/Shell

Terminal.app (macOS)

Additional Information

_No response_