Logsift Workflow Research¶
Error analysis, filtering methodology, and systematic fixing approach for command output.
Research Overview¶
Date: 2025-12-03 (initial), 2025-12-04 (updated)
Implementation: /logsift and /logsift-auto slash commands
Related: Commit Agent, Context Engineering
Problem Statement¶
Command Output Overflow¶
Running installation/test scripts produces massive output:
- Test script: 10,000+ lines
- Claude Code context limit: 200k tokens (~30k lines)
- Single test run: ~50% of context window
- Multiple runs: Context overflow
Result: Cannot debug iteratively, context fills with success messages
Logsift Solution¶
What it does: Filters command output to show only errors and warnings
Input: 10,000+ lines of command output
Output: ~200 lines of errors/warnings/key messages
Compression: ~50x reduction
How It Works¶
logsift monitor -- bash tests/install/test-install.sh
# 1. Runs command in background
# 2. Captures all output
# 3. Shows periodic status updates
# 4. Analyzes when done
# 5. Reports only issues
5-Phase Error Methodology¶
Phase 1: Initial Analysis¶
Wait for logsift to complete
Read the FULL error report
Identify ALL errors before acting
Look for patterns across failures
Anti-pattern: Jump to first error immediately
Phase 2: Root Cause Investigation¶
Determine relationships:
Related errors (shared root cause):
→ All point to missing foo package
Independent errors:
→ Three unrelated issues
Reality check: Don't force connections that don't exist
Phase 3: Solution Strategy¶
When related: Fix single root cause
When independent: Fix each individually (this is correct!)
Always read files before editing:
# ❌ Don't guess
Edit file.sh
# ✅ Do this
Read file.sh # Understand context
Edit file.sh # Make informed change
Phase 4: Iterative Fix-and-Rerun¶
- Re-run SAME logsift command
- Compare new errors to previous
- Verify fixes resolved issues
- Continue until all errors eliminated
Phase 5: Verification¶
- Confirm solution is robust
- Ensure no errors suppressed
- Verify fix aligns with codebase
Two Command Variants¶
/logsift - Explicit Command¶
Pros:
- Fast, no interpretation
- Explicit and unambiguous
- Claude gets straight to analysis
Cons:
- Need to know exact path/flags
- More typing
/logsift-auto - Natural Language¶
Pros:
- Natural language
- Claude figures out paths
- Less typing
Cons:
- Slight interpretation overhead
- May need clarification
Comparison: Track via metrics to see which works better
Integration with Commit Agent¶
Commit agent uses logs ift for pre-commit:
# Phase 4: Background (suppress auto-fixes)
pre-commit run > /dev/null 2>&1 || true
# Phase 5: Logsift (show errors only)
logsift monitor -- pre-commit run --files file1.py file2.sh
Token savings: ~950 tokens per pre-commit run
Guiding Principle¶
Prioritize correctness and root cause fixes over token savings
Logsift already saved massive context by filtering logs. Now use that savings to fix things properly. If thorough investigation requires reading files or exploring code, DO IT.
Related Research¶
- Context Engineering - Compression strategy
- Commit Agent - Uses logsift for pre-commit
- Prompt Engineering - Systematic methodology
References¶
- Logsift Tool
- URL: https://github.com/user/logsift (project-specific)
- Topics: Log filtering, error extraction
Research Date: 2025-12-03, updated 2025-12-04
Status: Production use in slash commands