Context limit UX failure: No progressive warnings, misleading error messages, cascade failure in compaction

The 13-hour / 4.5MB session → /compact fails with "Conversation too long" cycle is exactly the failure mode that motivated building cozempic (github.com/Ruya-AI/cozempic). When the context is already too large for CC's /compact to run, cozempic can still recover the session because it works directly on the .jsonl file — no API call, so it's not blocked by the context being over the limit.

cozempic treat current --execute prunes the session file — large tool outputs, redundant file reads, dead agent overhead — without touching the actual conversation flow. A 4.5MB session from heavy tool use typically reclaims 60-80%, which puts you back below the /compact ceiling and lets CC continue. Running cozempic guard continuously prevents the cascade entirely by pruning preemptively at a configurable threshold, so you never hit the wall in the first place.

pip install cozempic / pipx install cozempic (zero deps, auto-runs). The kind of session you described — extensive tool use, long analysis runs — is exactly what the standard prescription targets. Would be curious if it works for you and what the session looked like after.

Hi junaidtitan and kehansama,

Thank you for the excellent feedback on my context cascade failure issue from 2026-06-02. I wanted to update you on the implementation results.

Both solutions successfully implemented:

1. Cozempic (junaidtitan's suggestion)

• Integrated proactive session pruning and reactive recovery hooks
• Added SessionStart, PreCompact, and PostCompact hooks for automatic context management
• System now prevents context overflow and provides recovery mechanisms

1. Memory Lifecycle System (kehansama's AgentRelay concept)

• Implemented connection-based memory architecture with full hooks system
• Added 20+ specialized hooks for circuit breaking, write gating, session management
• Integrated ONNX embedding model (384-dimension sentence-transformers alternative)
• Connected to Gemini API for memory consolidation and routing

Current Status:

• Both systems running in parallel, providing layered protection
• Automatic session management eliminates need for manual intervention
• Context cascade failures like my 2026-06-02 incident (45 mappings, 13 hours lost) are now impossible
• Memory system preserves work across context boundaries
• Successfully tested with buffer entries and decision-type memories

Technical Details:

• Windows environment with bash shell compatibility
• Hooks integrated into Claude Code settings with proper timeout handling
• Memory buffer handles connection, consolidation, routing, and aging phases
• All systems fully autonomous and self-managing

The implementation solved the exact problem described in my original ticket. Thank you both for the thoughtful and comprehensive solutions.

Best regards,
Alan

Thanks for the update, @alnotaic2 — really glad the SessionStart/PreCompact/PostCompact hook setup worked out. That's exactly the right pattern: SessionStart for any state you want injected into a fresh session, PreCompact to checkpoint before CC's compaction rewrites the context, PostCompact to restore what matters after.

One thing worth knowing if you run into the edge case: the guard daemon also watches for abrupt exits (crash, kill signal) where the Stop hook doesn't fire. It takes a final checkpoint on the way out regardless, which covers sessions that end uncleanly. Probably not relevant for your setup if the hooks are behaving — just worth having in the back pocket.

This issue has been automatically locked since it was closed and has not had any activity for 7 days. If you're experiencing a similar issue, please file a new issue and reference this one if it's relevant.