### Preflight Checklist - [x] I have searched [existing requests](https://github.com/anthropics/claude-code/issues?q=is%3Aissue%20label%3Aenhancement) and this feature hasn't been requested yet - [x] This is a single feature request (not multiple features) ### Problem Statement ## Problem Statement Currently, Claude Code's auto-compact feature creates summaries that omit critical debugging information, including: - Error messages and stack traces - Failed attempts and why they failed - Prob

[FEATURE] Pre-Compact Auto-Save and Improved Summarization

Resolved 💬 5 comments Opened Dec 6, 2025 by EMarkODell Closed Feb 27, 2026

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

Currently, Claude Code's auto-compact feature creates summaries that omit critical debugging information, including:

Error messages and stack traces
Failed attempts and why they failed
Problem-solving iterations
Complete context of what was tried
Temporary variables and intermediate results

User Impact

When users search conversation history for past errors or debugging sessions, they find information has "vanished." This:

Breaks debugging workflows - Cannot review what went wrong previously
Causes repeated mistakes - AI may repeat failed approaches it doesn't remember
Blocks learning - Users cannot learn from documented failures
Erodes trust - Information disappears without warning or explanation
No user control - Cannot prevent, delay, or preview compaction

Current Behavior

User sees brief notification: "Conversation compacted · ctrl+o for history"
Compaction happens automatically at token limits (~180K-200K tokens)
Summary omits errors, debugging context, and failed attempts
No advance warning or ability to save before compaction
No way to recover lost information

---

Proposed Solution

Proposed Solutions

Priority 1: Pre-Compact Auto-Save (CRITICAL)

Implement automatic /save_claude before every auto-compact operation.

Implementation:

1. When token usage approaches compact threshold (e.g., 175K tokens)
2. Automatically trigger: /save_claude
3. Wait for save to complete
4. Then proceed with auto-compact
5. Notify user: "Conversation saved before compacting"

Benefits:

✅ Preserves complete transcript including all errors/debugging
✅ No user action required
✅ Backward compatible - existing saves continue to work
✅ Minimal performance impact
✅ Simple implementation

Alternative: Configurable Hook

{
  "hooks": {
    "beforeCompact": {
      "command": "~/AI_claude/hist/.save_conversation.sh",
      "timeout": 5000
    }
  }
}

This allows users to customize pre-compact behavior.

---

Priority 2: Richer Summaries (HIGH)

Preserve debugging context in auto-compact summaries.

Include in summaries:

✅ Error messages with full stack traces
✅ Failed attempts and why they failed
✅ Debugging iterations and solutions
✅ Warnings and edge cases discovered
✅ Performance issues identified
✅ Unresolved problems/blockers

Current summary format (simplified):


Summary:
User requested feature X.
Assistant implemented feature X.
Tests passed.

Improved summary format:

Summary:
User requested feature X.
Assistant implemented feature X.

Initial approach failed with error: [full error]
Reason: [explanation]

Second approach succeeded with changes:
- [specific changes]

Tests passed after fixing:
- [test failure details]
- [solution applied]

Unresolved issues:
- [open problems]

---

Priority 3: Compaction Preview (MEDIUM)

Show users what will be lost before compacting.

UI Flow:

Token limit approaching (180K/200K)

Options:
[1] Compact now (show preview of what will be summarized)
[2] Save full transcript first, then compact
[3] Continue without compacting (may affect performance)

Preview:
- 47 error messages will be summarized
- 12 failed attempts will be condensed
- 200+ lines of debugging output will be reduced

Benefits:

Users make informed decisions
Can save important sessions manually first
Transparency builds trust

---

Priority 4: Manual Compaction Control (LOW)

Allow users to trigger compaction manually.

Commands:

/compact              # Compact now
/compact preview      # Show what would be compacted
/compact disable      # Disable auto-compact (with warnings)
/compact force        # Force compact even if below threshold

Use cases:

User wants to compact at a logical stopping point
User knows next task will need full context
User wants to start fresh after debugging

---

Priority 5: Full Transcript Preservation (LOW)

Keep complete uncompacted history in background storage.

Implementation:

Auto-save full transcript to .claude/cache/full_history/
Summaries used for active context
Full history available via: /history full
Configurable retention (7 days, 30 days, forever)

Benefits:

No information ever lost
Users can retrieve any past detail
Debugging history always available

---

Alternative Solutions

_No response_

Priority

High - Significant impact on productivity

Feature Category

CLI commands and flags

Use Case Example

Current Behavior (Problematic)

Session 1: User debugging HTM encoding bug
- Tried approach A: Failed with error X
- Tried approach B: Failed with error Y
- Tried approach C: Success!

[Auto-compact happens]

Session 2: User wants to review why A and B failed
- Searches transcript: "error X" → NO RESULTS
- Searches transcript: "approach A" → NO RESULTS
- Information lost, user confused

Proposed Behavior (Fixed)

Session 1: User debugging HTM encoding bug
- Tried approach A: Failed with error X
- Tried approach B: Failed with error Y
- Tried approach C: Success!

[Token limit reached: 178K tokens]
→ Auto-save triggered: ~/AI_claude/hist/2025-12-06_14-30-00.txt
→ Saved 178K tokens, 2.1 MB
→ Creating summary (preserving errors and failures)
→ Auto-compact complete

Session 2: User wants to review why A and B failed
- Searches transcript: "error X" → FOUND in summary
- OR opens saved file: ~/AI_claude/hist/2025-12-06_14-30-00.txt
- Complete information preserved

Additional Context

Feature Request: Pre-Compact Auto-Save and Improved Summarization

Date: 2025-12-06
Product: Claude Code CLI
Priority: High - UX/Trust Issue
Category: Conversation Management

---

Summary

Request automatic conversation saving before auto-compact and improved summarization that preserves debugging context and error history.

---

Problem Statement

Currently, Claude Code's auto-compact feature creates summaries that omit critical debugging information, including:

Error messages and stack traces
Failed attempts and why they failed
Problem-solving iterations
Complete context of what was tried
Temporary variables and intermediate results

User Impact

When users search conversation history for past errors or debugging sessions, they find information has "vanished." This:

Breaks debugging workflows - Cannot review what went wrong previously
Causes repeated mistakes - AI may repeat failed approaches it doesn't remember
Blocks learning - Users cannot learn from documented failures
Erodes trust - Information disappears without warning or explanation
No user control - Cannot prevent, delay, or preview compaction

Current Behavior

User sees brief notification: "Conversation compacted · ctrl+o for history"
Compaction happens automatically at token limits (~180K-200K tokens)
Summary omits errors, debugging context, and failed attempts
No advance warning or ability to save before compaction
No way to recover lost information

---

Proposed Solutions

Priority 1: Pre-Compact Auto-Save (CRITICAL)

Implement automatic /save_claude before every auto-compact operation.

Implementation:

1. When token usage approaches compact threshold (e.g., 175K tokens)
2. Automatically trigger: /save_claude
3. Wait for save to complete
4. Then proceed with auto-compact
5. Notify user: "Conversation saved before compacting"

Benefits:

✅ Preserves complete transcript including all errors/debugging
✅ No user action required
✅ Backward compatible - existing saves continue to work
✅ Minimal performance impact
✅ Simple implementation

Alternative: Configurable Hook

{
  "hooks": {
    "beforeCompact": {
      "command": "~/AI_claude/hist/.save_conversation.sh",
      "timeout": 5000
    }
  }
}

This allows users to customize pre-compact behavior.

---

Priority 2: Richer Summaries (HIGH)

Preserve debugging context in auto-compact summaries.

Include in summaries:

✅ Error messages with full stack traces
✅ Failed attempts and why they failed
✅ Debugging iterations and solutions
✅ Warnings and edge cases discovered
✅ Performance issues identified
✅ Unresolved problems/blockers

Current summary format (simplified):

Summary:
User requested feature X.
Assistant implemented feature X.
Tests passed.

Improved summary format:

Summary:
User requested feature X.
Assistant implemented feature X.

Initial approach failed with error: [full error]
Reason: [explanation]

Second approach succeeded with changes:
- [specific changes]

Tests passed after fixing:
- [test failure details]
- [solution applied]

Unresolved issues:
- [open problems]

---

Priority 3: Compaction Preview (MEDIUM)

Show users what will be lost before compacting.

UI Flow:

Token limit approaching (180K/200K)

Options:
[1] Compact now (show preview of what will be summarized)
[2] Save full transcript first, then compact
[3] Continue without compacting (may affect performance)

Preview:
- 47 error messages will be summarized
- 12 failed attempts will be condensed
- 200+ lines of debugging output will be reduced

Benefits:

Users make informed decisions
Can save important sessions manually first
Transparency builds trust

---

Priority 4: Manual Compaction Control (LOW)

Allow users to trigger compaction manually.

Commands:

/compact              # Compact now
/compact preview      # Show what would be compacted
/compact disable      # Disable auto-compact (with warnings)
/compact force        # Force compact even if below threshold

Use cases:

User wants to compact at a logical stopping point
User knows next task will need full context
User wants to start fresh after debugging

---

Priority 5: Full Transcript Preservation (LOW)

Keep complete uncompacted history in background storage.

Implementation:

Auto-save full transcript to .claude/cache/full_history/
Summaries used for active context
Full history available via: /history full
Configurable retention (7 days, 30 days, forever)

Benefits:

No information ever lost
Users can retrieve any past detail
Debugging history always available

---

Use Case Example

Current Behavior (Problematic)

Session 1: User debugging HTM encoding bug
- Tried approach A: Failed with error X
- Tried approach B: Failed with error Y
- Tried approach C: Success!

[Auto-compact happens]

Session 2: User wants to review why A and B failed
- Searches transcript: "error X" → NO RESULTS
- Searches transcript: "approach A" → NO RESULTS
- Information lost, user confused

Proposed Behavior (Fixed)

Session 1: User debugging HTM encoding bug
- Tried approach A: Failed with error X
- Tried approach B: Failed with error Y
- Tried approach C: Success!

[Token limit reached: 178K tokens]
→ Auto-save triggered: ~/AI_claude/hist/2025-12-06_14-30-00.txt
→ Saved 178K tokens, 2.1 MB
→ Creating summary (preserving errors and failures)
→ Auto-compact complete

Session 2: User wants to review why A and B failed
- Searches transcript: "error X" → FOUND in summary
- OR opens saved file: ~/AI_claude/hist/2025-12-06_14-30-00.txt
- Complete information preserved

---

Technical Requirements

Pre-Compact Auto-Save

Required:

Hook into compaction trigger (before summarization)
Call existing /save_claude functionality
Handle save failures gracefully
Notify user of save location
Continue with compact after save

Estimated complexity: Low (1-2 days)
Estimated risk: Low (uses existing save mechanism)

Improved Summaries

Required:

Modify summarization prompt to preserve:
Error messages (full text)
Failed attempts (with reasons)
Debugging iterations
Unresolved issues
Increase summary budget for error context
Add structured sections (Errors, Warnings, Unresolved)

Estimated complexity: Medium (3-5 days)
Estimated risk: Medium (requires prompt tuning)

---

Success Metrics

Measure improvement by:

User Trust

% of users who report "lost information" issues
Target: Reduce by 90%

Debugging Efficiency

Time to find past error information
Target: < 30 seconds (vs. "not found" currently)

Repeated Mistakes

Instances where AI repeats previously failed approach
Target: Reduce by 80%

Manual Saves

Frequency of manual /save_claude commands
Target: Reduce by 50% (automation replaces manual)

---

Alternatives Considered

1. Larger Context Window

Pros: Delays compaction
Cons: Higher cost, doesn't solve fundamental issue
Verdict: Not sufficient alone

2. User-Only Manual Saves

Pros: Simple, already exists
Cons: Relies on user remembering, easy to forget
Verdict: Insufficient - users forget or don't know when to save

3. External Transcript Logging

Pros: Complete preservation
Cons: Users must configure separately, extra complexity
Verdict: Good supplement but not replacement for auto-save

4. No Compaction (Unlimited Context)

Pros: No information loss
Cons: Technically infeasible, extremely expensive
Verdict: Not realistic

---

Backward Compatibility

All proposed changes are backward compatible:

✅ Pre-compact save uses existing /save_claude infrastructure
✅ Improved summaries enhance (don't replace) existing format
✅ Manual controls are additive (new commands)
✅ Full preservation is opt-in configuration
✅ No breaking changes to existing workflows

---

Implementation Priority

Phase 1 (MVP): Pre-Compact Auto-Save

Impact: High - Solves 80% of the problem
Effort: Low - Uses existing save mechanism
Timeline: 1-2 weeks

Phase 2: Improved Summaries

Impact: Medium - Better immediate context
Effort: Medium - Requires prompt engineering
Timeline: 3-4 weeks

Phase 3: Preview & Control

Impact: Medium - Better UX
Effort: Medium - New UI components
Timeline: 4-6 weeks

Phase 4: Full Preservation

Impact: Low - Nice-to-have for power users
Effort: High - Storage, retrieval, UI
Timeline: 2-3 months

---

User Feedback

Actual user quote:

"this may explain why so often when I go looking back in the transcript, stuff seems to have vanished ... most notably every time 'you' make an 'error' or something does not work as expected ... which is a HUUUUUUUUUUUUUUUUUUUGE bug to me ... this is a MAJOR problem"

Pain points identified:

Information disappears without warning
Cannot review past errors
Cannot learn from documented failures
Erodes trust in the tool
No control or visibility into compaction

---

Request

Please implement Priority 1 (Pre-Compact Auto-Save) as soon as possible. This single change would dramatically improve user trust and debugging workflows.

Minimal viable implementation:

def auto_compact():
    if token_count > COMPACT_THRESHOLD:
        # NEW: Save before compacting
        save_result = execute_slash_command("/save_claude")
        notify_user(f"Conversation saved: {save_result.path}")

        # Existing: Proceed with compact
        summary = create_summary()
        replace_history_with_summary(summary)
        notify_user("Conversation compacted · ctrl+o for history")

Estimated effort: 1-2 days
Impact: Solves critical user trust issue
Risk: Low (uses existing tested save mechanism)

---

Contact

For questions or discussion about this feature request:

GitHub Issues: https://github.com/anthropics/claude-code/issues
Community Forum: [if available]

---

Thank you for considering this request. This feature would significantly improve the Claude Code debugging experience and user trust.

View original on GitHub ↗

This issue has 5 comments on GitHub. Read the full discussion on GitHub ↗