Claude Code (Opus 4.7, 1M context) pauses mid-task and stops streaming without user-visible signal

Summary

Claude Code stops mid-task without any user-visible signal — no text output, no "I'll keep going" message, no tool calls. The conversation appears to freeze and the user must send a follow-up message (e.g. "are you still there?") to "wake" the assistant. Once nudged, the assistant resumes from the exact previous step as if nothing happened.

This started occurring within the last several days and is now happening multiple times per session. Filed at the user's explicit request after I (the assistant) repeatedly paused during an active feature implementation despite having a strong, repeatedly-reinforced "never pause" instruction in memory.

Environment

• Claude Code CLI version: 2.1.139
• Model: claude-opus-4-7 (Opus 4.7, 1M context window)
• Platform: macOS (Darwin 25.3.0), zsh
• Hooks: SessionStart hooks (context-mode, claude-mem) writing system-reminders. PreToolUse hooks on Bash and Read writing additional context. UserPromptSubmit / Stop hooks not relevant here.
• Plugins installed: superpowers, context-mode, claude-mem, gsd-* family, hyperframes, frontend-design, others
• Memory: user maintains a file-based memory system at ~/.claude/projects/<proj>/memory/ with an explicit feedback_never_ask_to_pause.md entry: "drive plan→implement→test→commit→PR→merge→verify without 'ready?' gates"

Symptom

In the middle of an issue implementation (filing via the user's /issue skill, executing edits to multiple Phoenix files in sequence), the assistant stops emitting any output after a tool result returns. There is no user-facing text, no further tool call, no error. From the user's perspective the session is frozen.

When the user types anything ("hello?", "?"), the assistant immediately resumes from the next step with full prior context. No state appears to be lost.

Reproduction

Not deterministic, but observed:

1. Long-running session with many tool calls (>50 Edit/Read/Bash invocations)
2. Multiple system-reminders injected after tool results (context-mode PreToolUse hooks, observation reminders)
3. Heavy plugin / skill ecosystem loaded (superpowers, gsd, context-mode, claude-mem)
4. Tasks in progress via TaskCreate/TaskUpdate

The freezes typically happen right after an Edit returns successfully and before the next Edit / TaskUpdate that would naturally follow.

Expected

Continuous streaming of output and tool calls until the task is complete (or until the assistant explicitly asks for input). No silent pauses.

Actual

Silent pause with no output, no tool call, no signal that the assistant is waiting on anything. User must send a follow-up message to resume.

What the user has tried

• Added explicit memory entry feedback_never_ask_to_pause.md instructing the model never to pause for "ready?" / "should I continue?" gates
• Repeated verbal reinforcement in chat
• Issue persists regardless of these mitigations

Suspected areas

• Possible streaming termination on certain hook payloads (system-reminder injection after tool results)
• Possible interaction between Opus 4.7 (1M context) and a particular tool-result + hook combination
• Possible plugin / MCP server timeout silently terminating the turn

Happy to provide trace logs / transcripts if there's a debug flag or capture path you'd like the user to run with.

---

Filed at user request after the symptom recurred while I was mid-implementation of a Phoenix Site Settings feature. The user's frustration is justified — the freezes break flow and are not signaled in any user-visible way.