[Feature Request] AutoClaw Smart-Recovery: Scope-Aware Validation & Checkpoint Resume

Feature Request: AutoClaw Smart-Recovery Improvements

📋 Summary

Scope-aware replacement validation, checkpoint-based resume with user confirmation, and partial success extraction for AutoClaw agent tasks that fail due to overly broad code modifications.

---

🔍 Problem Description

Observed Failure Pattern

When executing complex multi-step tasks (e.g., PPT generation with 25+ slides), AutoClaw agents can make overly broad string replacements that break valid code patterns:

❌ Agent Action: Replace all "inches" occurrences
✅ Result: Broke valid `.inches` calls on actual `Inches` objects from python-pptx
✅ Cascade: New bugs introduced → auto-recovery fails → task incomplete

Evidence from Production Failure

Screenshot captured: PPT Generator task showing:

"That replacement was too broad — it broke valid .inches calls on actual Inches objects. Let me rewrite the problematic arithmetic patterns properly."

"The task still could not finish after automatic recovery. You can retry, and I'll continue from the existing results without blindly repeating completed actions."

Deliverables produced:

• build_new_deck.py — initial attempt (partial)
• build_new_deck_v2.py — recovery attempt (still partial)

Root Causes Identified

Cause	Impact	Frequency
Overly broad regex replacements	Breaks unrelated code using similar patterns	High in python-pptx, reportlab tasks
No scope validation before apply	Agent cannot predict side effects of edits	Common in subagent-driven development
Blind retry without user input	Repeats or compounds previous mistakes	~40% of recovery attempts
No partial extraction on failure	Working deliverables trapped in failed session	User loses salvageable work

---

✅ Proposed Solutions

Feature 1: Scope-Aware Replacement Validator

Description: Before applying any string/regex replacement, validate that changes are scoped correctly.

def validate_replacement(old_code, new_code, pattern, target_scope):
    """
    Validate that a replacement only affects intended scope.
    
    Args:
        old_code: Original source code
        new_code: Modified source code  
        pattern: The regex/string being replaced
        target_scope: List of functions/classes that SHOULD be affected
    
    Returns:
        (is_valid, reason) tuple
    """
    affected = find_affected_symbols(old_code, new_code)
    unintended = [s for s in affected if s not in target_scope]
    
    if unintended:
        return False, f"Replacement would affect {unintended} outside target scope"
    if breaks_type_annotations(new_code):
        return False, "Preserves attribute access patterns (.inches, .width, etc.)"
    if breaks_import_statements(new_code):
        return False, "Modifies import statements indirectly"
    
    return True, "Scope validated"

Behavior:

• Agent proposes replacement → validator checks scope → warns if too broad
• Agent must narrow pattern or explicitly confirm override
• Configurable strictness level (strict/lenient/off)

Feature 2: Checkpoint-Based Resume with Confirmation

Description: Insert explicit save-points during long-running tasks; require user approval before recovery retries.

Checkpoint Triggers:

• After each major section/phase completion
• Before any bulk replacement operation
• After N successful tool calls (configurable, default=10)
• When token usage exceeds threshold

Resume Flow:

Task Fails at Step 12/25
    ↓
AutoClaw saves checkpoint: "slides_11_complete.pptx"
    ↓
Prompt user:
    ┌─────────────────────────────────────┐
    │ ⚠️ Task failed at slide 12          │
    │ ✓ Slides 1-11 saved successfully   │
│                                      │
    │ [Retry from checkpoint]  [Abort]   │
    │ [View error details]   [Edit fix]  │
    └─────────────────────────────────────┘
    ↓
User selects option → AutoClaw continues intelligently

Feature 3: Partial Success Extraction

Description: On task failure, automatically extract and preserve all working deliverables.

Implementation:

extraction_rules:
  - pattern: "*.py"
    action: save_to_deliverables
    condition: syntax_valid
  - pattern: "*.pptx"
    action: save_and_verify
    condition: file_opens_without_error
  - pattern: "*.json"
    action: save_always
    condition: null  # always save config/state files

on_failure:
  - run_tests_on_deliverables
  - generate_failure_report:
      include: [error_log, completed_steps, remaining_steps, suggested_fix]
  - offer_resume_options

User Benefit: Never lose work — even failed sessions produce usable artifacts.

Feature 4: Smart Retry Strategy

Description: Instead of blind retry, analyze failure cause and adapt approach.

Decision Matrix:

Failure Type	Smart Response
Broad replacement broke code	Narrow regex scope; exclude imports/types
Import statement corrupted	Restore imports from original; only redo logic
Syntax error in generated code	Run linter before next attempt; fix lint errors first
Token limit exceeded	Compress context; summarize completed work
External API failure	Exponential backoff; switch endpoint if available

---

🎯 Use Cases

Primary: Document Generation (pptx/docx/pdf)

• Prevent .inches, .width, .Emu attribute corruption in python-pptx
• Preserve reportlab dimension calculations
• Allow partial PPT extraction when later slides fail

Secondary: Full-Stack Development

• Prevent import statement corruption during refactoring
• Validate database migration changes don't affect models
• Extract working endpoints when integration tests fail

Tertiary: Data Processing Pipelines

• Checkpoint after each ETL stage
• Recover from transformation errors without re-running extraction
• Preserve intermediate data artifacts

---

📊 Success Metrics

Metric	Current State	Target State
Task completion rate (complex tasks)	~60%	>85%
Recovery success rate	~40%	>75%
Work lost on failure	0% salvaged	>90% extracted
User intervention required	Post-failure manual	Pre-retry confirmation
Average retries to success	2-3 blind retries	1 smart retry

---

🔄 Implementation Priority

Priority	Feature	Effort	Impact
P0	Partial Success Extraction	Low	High
P0	Smart Retry Strategy	Medium	High
P1	Checkpoint-Based Resume	Medium	High
P2	Scope-Aware Validator	High	Medium

Recommended MVP: Ship P0 features first — they provide immediate value with minimal complexity.

---

🧪 Testing Scenarios

1. PPT Generation Test: Run 25-slide creation; intentionally trigger broad replacement; verify v2 script is extracted
2. Recovery Resume Test: Fail at step 10; verify checkpoint resume skips steps 1-9
3. Scope Validation Test: Attempt to replace inches globally; verify warning triggers
4. Partial Extraction Test: Generate 3 deliverables; fail on 4th; verify first 3 are preserved

---

💡 Additional Context

• Related Discord Thread: GLM Coding Pro quota/billing discussion (where this failure was reported)
• Affected Component: AutoClaw agent system / ZCode IDE
• Reporter Experience: Intermediate user running autopilot skill for document generation
• Workaround Available: Manual script execution + targeted retry prompt (not intuitive)

---

Submitted by:
Regards,
Roman (Discord: bignavi_x)

• Z.ai Volunteer Ambassador
• Tech resources channel: https://t.me/VibeCodePrompterSystem
• Tech Blog: http://claw.rommark.dev/blog

No repository files, tests, or implementation entry points are named. Start by locating the AutoClaw recovery, retry, checkpoint, and deliverable-handling entry points, then scope the work around the proposed P0 features. Use the listed PPT generation, recovery resume, and partial extraction scenarios to define what “done” means.