[Feature Request] Enhance /compact with structured context preservation sections

Preflight Checklist

• [x] I have searched existing requests and this feature hasn't been requested yet
• [x] This is a single feature request (not multiple features)

Problem Statement

The default /compact summary produces a chronological narrative that captures what happened but loses critical actionable context:

| Missing Element | Impact |
|----------------|--------|
| Compact counter | Can't track multi-compact sessions |
| Immediate next action | Claude doesn't know what to do next (buried in narrative) |
| Settled decisions | Decisions get re-debated, wasting time |
| Dead ends | Failed approaches get retried |
| Trust anchors | Working components get questioned |
| User preferences | Preferences get forgotten, user has to repeat |
| Task queue | Remaining work unclear |
| Breakthroughs | Key insights lost in prose |

Users have built elaborate infrastructure (PostgreSQL databases, Gemini API extraction, external scripts) to capture this context - all of which breaks because PreCompact hooks don't fire on manual /compact (#13572).

The Discovery

The /compact command already accepts custom instructions! Running /compact [text] guides the summarizer to include specified sections.

I've been using this successfully across multiple compact cycles:

/compact In addition to the default summary, explicitly include these sections at the END:

0) COMPACT NUMBER - This is compact #[N].
1) IMMEDIATE NEXT ACTION - [Specific imperative].
2) SETTLED DECISIONS - [Decisions with rationale]
3) DEAD ENDS - [What failed and WHY]
4) TRUST ANCHORS - [What's verified working]
5) USER PREFERENCES - [How user wants things done]
6) TASK QUEUE - [Remaining work with dependencies]
7) BREAKTHROUGHS - [Key insights and why they matter]

Result: All 8 sections appear accurately in post-compact context. The next Claude instance has explicit, structured access to critical context.

Proposed Solution

Make these structured sections part of the default /compact behavior, or provide a flag like /compact --structured or /compact --preserve-context.

The summarizer should automatically extract and explicitly list:

1. Session counter - For multi-compact tracking
2. Continue by - Clear next action (imperative, not narrative)
3. Settled decisions - With brief rationale (don't revisit)
4. Dead ends - What failed and why (don't retry)
5. Trust anchors - Verified working (don't question)
6. User preferences - PERMANENT markers for lasting preferences
7. Task queue - Remaining work with dependencies
8. Breakthroughs - "Aha!" moments, paradigm shifts, key insights

Why This Matters

1. Eliminates infrastructure burden - No PostgreSQL, no Gemini extraction, no external scripts
2. Works with manual /compact - Doesn't depend on broken PreCompact hooks
3. Claude IS the extraction engine - AI analyzing its own session is more accurate than external parsing
4. Prevents common post-compact failures:

• Re-debating settled decisions
• Retrying failed approaches
• Losing user preferences
• Missing immediate next action
• Forgetting key breakthroughs

Alternative Solutions

1. Fix PreCompact/PostCompact hooks for manual /compact (#13572, #3537) - Still valuable but more complex
2. Document the custom instructions approach - Users can manually add instructions (current workaround)
3. Configurable default instructions - Let users set default /compact instructions in settings.json

Priority

High - Significant impact on productivity

Feature Category

Core functionality - Context preservation

Use Case Example

Before (default /compact):

"The user asked about X, then we discussed Y, and decided to use Z approach..."

After (structured /compact):

[Standard summary] COMPACT #3 IMMEDIATE NEXT ACTION: Run test suite in src/tests/ SETTLED DECISIONS: - Use PostgreSQL over SQLite (rationale: concurrent access needed) - Claude-analysis over Gemini extraction (rationale: simpler, no external deps) DEAD ENDS: - WebSocket approach failed - browser security restrictions TRUST ANCHORS: - Auth flow verified working (tested 3x) USER PREFERENCES: - VERBATIM means NO TRUNCATION (PERMANENT) - Open files in TextMate for review TASK QUEUE: - Implement caching layer - Write integration tests BREAKTHROUGHS: - "Custom /compact instructions add explicit sections!" - Eliminates need for external extraction

Additional Context

I've been using this approach successfully and it completely eliminated my need for the PostgreSQL/Gemini extraction pipeline I had built. The workaround works, but it would be much better as a first-class feature.

Related issues:

• #13572 - PreCompact hook not triggered on manual /compact
• #3537 - No PostCompact hook
• #3349 - Custom pre/post compact commands
• #7359 - /compact options (different focus - selective deletion vs structured preservation)