[BUG] Opus 4.7 subagents enter generation loop on post-task finalization/return text (work completes, agent doesn't exit)

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?

Environment

• Claude Code: latest
• Model: claude-opus-4-7[1m] (Opus 4.7, 1M context)
• Invocation: Agent tool with subagent_type: "general-purpose", model: "opus", run_in_background: true

Reproduction

Dispatch a Opus 4.7 subagent with a moderately complex multi-step task (e.g., "apply 3 surgical fixes to file X, add ~6 regression tests, run pytest, commit"). Subagent completes the work successfully (verified via git log showing the expected commit landed) but then enters a generation loop emitting the same lines repeatedly. The completion notification never fires because the agent process never exits.

Observed pattern (3 incidents in one session)

┌──────────┬──────────────────────────────────────────────────┬───────────────────────────────────────────────────┐
│ Incident │ Work state │ Loop content │
├──────────┼──────────────────────────────────────────────────┼───────────────────────────────────────────────────┤
│ 1 │ All edits + tests + analysis done │ Repeated Update(test_file.py) calls + same prose │
├──────────┼──────────────────────────────────────────────────┼───────────────────────────────────────────────────┤
│ 2 │ Implementer correctly detected conflict + bailed │ (no loop — model self-recognized) │
├──────────┼──────────────────────────────────────────────────┼───────────────────────────────────────────────────┤
│ 3 │ Commit <SHA> already landed cleanly │ Repeated Update(test_file.py) calls in transcript │
└──────────┴──────────────────────────────────────────────────┴───────────────────────────────────────────────────┘

Mitigation we use

1. User watches live transcript for repetition
2. Controller kills via TaskStop
3. Always git status after kill — work is often 95-100% complete; commit may already be done
4. Add explicit "EXIT CLEANLY — no loops" to subagent prompts (works ~50% of time)

Expected

Agent exits with completion notification after final tool call (commit).

Actual

Agent process continues generating repetitive text indefinitely. No notification. Background dispatches wait forever.

Impact

• Wasted model compute on the looping generation
• Controller has to manually poll / be alerted by user
• Background dispatches with run_in_background: true give no feedback path
• The Agent tool has no timeout parameter to bound this

Suggested fix

Either:

• Add agent-side loop detection (same N consecutive identical tool calls or N tokens of identical text → auto-abort)
• Expose a timeout_seconds parameter on the Agent tool so controllers can bound runaway agents
• Surface partial-completion notifications so the controller can git status and salvage

PS: Bug description done and written by Claude Opus 4.7

What Should Happen?

Exit cleanly with no repetitive text

Error Messages/Logs

Repetitive text ..

Steps to Reproduce

It has happened about 4-5 times in my current coding. There is no specific way to reproduce. It just comes.

Claude Model

Opus

Is this a regression?

I don't know

Last Working Version

_No response_

Claude Code Version

2.1.150

Platform

Anthropic API

Operating System

Other Linux

Terminal/Shell

Xterm

Additional Information

_No response_