Context quality preservation — 4 improvements across the prompt→compact→recovery chain

[Jianping-Xiong]

Part of an architecture improvement initiative based on Claude Code source analysis.

Related:

• #21 Execution safety improvements
• #17 Experience extraction MVP

---

Four changes that work together to prevent information loss through the full context lifecycle.

1. Add reactive compact fallback

If a single tool output is huge, context can exceed the limit in one turn — before autocompact triggers. API error, session dies.

Claude Code has three layers: autocompact → blocking limit → reactive compact. The third one fires after API rejection, does emergency compression, retries. We only have the first layer.

2. Add 4 required sections to compaction summary

Our compaction already does well: code-extracted file lists, tool failures, identifier preservation, post-compaction refresh. But the summary body itself has no structure — LLM writes whatever it wants.

Suggest requiring:

• Pending Tasks — not done yet (most dangerous to lose)
• Key Decisions — constraints already established
• Current Work — where things stand
• User Instructions — what the user explicitly asked

3. Prompt assembly: static first, dynamic last

Prompt cache = prefix matching. First byte that differs → everything after = cache miss.

Our modular prompt is better than Claude Code's monolith for this, but only if order is right:

Good: [SOUL.md][TOOLS.md][AGENTS.md][USER.md] → [MEMORY.md] → [conversation]
      ← rarely changes ──────────────→        ← sometimes →  ← every turn →

Need to verify current assembly order and fix if needed.

4. Add session-end as memory flush trigger

Memory flush currently only fires before compaction. Misses:

• Short conversations that never trigger compaction — zero extraction
• Content after the last compaction — the tail is never captured

The existing session-memory hook already fires on /new and /reset. Add extraction there.