Bug #16: Edit Risk Model — Only CRITICAL Blocks

Status: RESOLVED in v3.15.1 Category: Risk Model / Agentic Workflow Severity: High (wasted token round-trips on every moderate edit) Resolution Date: 2026-03-02

Problem

When an edit changed more than 30% of a file, the server blocked the operation and asked the AI to retry with force: true. In agentic workflows (Claude Code, Claude Desktop), this caused a wasteful extra round-trip for every moderate edit — even though the change was perfectly safe.

What Happened

1. Claude called edit_file(path, old_text, new_text) to replace a block (~35% of file)
2. Server returned: OPERATION BLOCKED - MEDIUM risk ... add "force": true
3. Claude retried with force: true — the exact same edit, wasting tokens
4. The edit succeeded on retry, but the backup was never created for the blocked attempt

Root Cause

Two issues in core/edit_operations.go:

1. Wrong blocking condition: EditFile() used impact.IsRisky && !force to decide whether to block. Since IsRisky is true for MEDIUM (>=30%), HIGH (>=50%), and CRITICAL (>=90%), any edit changing more than 30% of a file was blocked. The method ShouldBlockOperation() in impact_analyzer.go correctly only blocked HIGH/CRITICAL, but was never called from EditFile().

2. Backup created too late: The backup creation (line 107) happened AFTER the blocking check (line 75). When the operation was blocked, no backup existed — so if the user had already made destructive changes, there was no safety net.

Why It Matters

• Token waste: Every MEDIUM risk edit (very common in refactoring) cost 2 tool calls instead of 1
• No backup on block: The safety net was missing exactly when it was most needed
• Agentic friction: AI agents follow instructions literally — the “add force: true” retry is pure overhead when the operation is safe

Solution

New Risk Model

Risk Level	% Change	Behavior
LOW	< 20%	Proceed silently
MEDIUM	20–74%	Auto-backup + proceed + risk notice in response
HIGH	75–89%	Auto-backup + proceed + prominent warning in response
CRITICAL	>= 90%	Auto-backup + block (require force: true)

Key Changes

1. Backup BEFORE block: Backup is now created before any blocking decision, so even CRITICAL-blocked operations have a safety net

2. ShouldBlockOperation() now used: EditFile() calls impact.ShouldBlockOperation(force) instead of the raw impact.IsRisky && !force check. Only "critical" risk blocks.

3. Non-blocking risk notice: MEDIUM and HIGH risk edits succeed immediately. The response includes the backup ID and a risk notice so the user can restore if needed:

   Risk notice [medium]
     40.0% of file changed (1 occurrence(s), ~200 chars)
     Auto-backup: 20260302-143022-abc123
     Restore with: restore_backup(backup_id)

4. Updated thresholds: MEDIUM raised from 30% to 20%, HIGH raised from 50% to 75% to better match real-world editing patterns

5. MultiEdit gets force parameter: Previously missing, now supports force: true for CRITICAL batch edits

Updated Tool Descriptions

All edit tools (edit_file, mcp_edit, intelligent_edit, smart_edit_file, multi_edit) now document:

• “Only blocks CRITICAL risk (>=90% file rewrite)”
• force parameter: “Force even if CRITICAL risk”

Files Changed

File	Change
core/impact_analyzer.go	Thresholds 20%/75%, ShouldBlockOperation() CRITICAL-only, new FormatRiskNotice()
core/edit_operations.go	EditResult.RiskWarning, backup before block, ShouldBlockOperation() call, MultiEdit force param + persistent backup
core/streaming_operations.go	SmartEditFile + streamingEditLargeFile get force param + risk assessment
core/claude_optimizer.go	Pass force through to SmartEditFile
core/engine.go	Fallback defaults 20.0/75.0
core/pipeline.go	Pass force to MultiEdit
main.go	CLI defaults, tool descriptions, handler responses, force params for all edit tools
tests/bug16_test.go	10 regression tests covering all risk levels
CHANGELOG.md	v3.15.1 entry

Testing

• 10 new regression tests in tests/bug16_test.go:
  • TestBug16_UpdatedThresholds — verify new default values
  • TestBug16_ShouldBlockOnlyCritical — table-driven test for all 4 risk levels
  • TestBug16_MediumRiskAutoProceeds — 40% change succeeds without force
  • TestBug16_HighRiskAutoProceeds — 80% change succeeds without force
  • TestBug16_CriticalRiskBlocked — full rewrite blocked without force
  • TestBug16_CriticalRiskForceProceeds — force=true overrides CRITICAL
  • TestBug16_LowRiskNoWarning — small edit has empty RiskWarning
  • TestBug16_BackupCreatedBeforeBlock — blocked error contains backup ID
  • TestBug16_MultiEditWithForce — MultiEdit accepts force, returns BackupID
  • TestBug16_FormatRiskNotice — notice contains backup ID and restore instructions
• Build: go build passes
• Full suite: go test ./tests/... ./core/... all pass
• Race detector: go test -race ./... clean