/compact and auto-compaction both broken — fails across sessions/models/subscriptions; auto-compact sometimes fires at 30% context; /feedback also fails

Bug: /compact and auto-compaction both broken — manual and auto fail across many sessions/models/subscriptions. Auto-compact sometimes fires at ~30% context used.

Summary

Both manual /compact and automatic compaction have been non-functional across a full day of heavy use. Symptoms persist across many sessions, multiple Claude models (Opus 4.6, and others), and multiple subscription tiers. Additionally, the in-app /feedback bug report submission flow is also failing with a generic "Could not submit feedback. Please try again later." error, so users cannot report this issue through the normal channel.

Symptoms

1. Manual /compact fails. Running /compact produces an error and the conversation continues uncompacted. See attached screenshot of a failed /feedback submission (same underlying issue?).

1. Automatic compaction fires prematurely. In some sessions auto-compact triggers at approximately 30% context used (70% remaining) — far below any reasonable threshold. In other sessions auto-compact never fires at all until context is exhausted, at which point the session effectively becomes unusable.

1. Cross-session, cross-model, cross-subscription. The user reports this has been broken ALL DAY across:

• Multiple independent Claude Code sessions (different project dirs, different CWDs)
• Multiple Claude models (Opus 4.6 with 1M context was confirmed affected; other models also reported affected)
• Multiple subscription accounts

1. /feedback submission ALSO fails. Attempting to file this bug through the in-app feedback flow returns:

> Could not submit feedback. Please try again later.

See attached screenshot. This means users who hit the compaction bug cannot report it through the intended channel, which is presumably why this is being filed as a regular GitHub issue instead.

Impact

High severity. For long-running sessions (e.g., an orchestrator coordinating multiple instances over a full workday, as in this user's case), broken compaction means:

• Context bloats until the session becomes unresponsive
• The user must manually /clear and lose all continuity
• Workarounds like persisting state to external files (runbook.md, etc.) are already in use but cannot fully compensate for the lost ability to compress conversation history

Reproduction

The user cannot reliably reproduce on demand (since auto-compact timing is non-deterministic), but the bug is highly reliable across a full working day:

1. Start a long-running Claude Code session (Opus 4.6 recommended, 1M context)
2. Use the session continuously with many tool calls (bash, read, edit, etc.)
3. Observe either:

• (a) Auto-compact fires at ~30% context (too early), OR
• (b) Auto-compact never fires until the session is effectively dead

1. Try to manually run /compact — fails.
2. Try to file feedback via /feedback — also fails with the "try again later" error.

Environment

• Platform: win32 (Windows 11 Pro 10.0.26200)
• Shell: bash (Git Bash)
• Model confirmed affected: claude-opus-4-6[1m] (Opus 4.6, 1M context)
• Claude Code CLI (interactive mode)
• User reports: "COMPACTING AUTO AND MANUAL HAS NOT WORKED ALL DAY, ACROSS MANY SESSIONS, MANY MODELS, EVEN DIFFERENT SUBSCRIPTIONS. SOMETIMES IT EVEN TRIES TO AUTO COMPACT WITH LIKE 30% CONTEXT USED (70% REMAINING)."

Expected behavior

• Manual /compact should compress the conversation history and return a summary
• Automatic compaction should fire at the configured threshold (presumably ~90-95% context used), not at 30%
• /feedback submission should succeed and provide a confirmation message

Actual behavior

• Manual /compact: fails
• Auto-compact: fires at wrong threshold (e.g., 30%) or never fires
• /feedback: returns "Could not submit feedback. Please try again later."

Attached

Screenshot of the failed /feedback submission showing the verbatim error and the user's written description, from this evening's session.

---

Filed from a working Claude Code session via gh issue create since the in-app feedback flow is broken.