---

name: code-agent-insights-workflow description: Mandatory workflow and conventions for working on code-agent-insights project (project)

Code Agent Insights Workflow

Purpose: Enforce best practices and prevent common mistakes when working on code-agent-insights.

Scope: This project only - mandatory workflow for AI agents and developers.

---

⚠️ CRITICAL: Before ANY Operation

1. Use Recall Tool FIRST (Mandatory)

Before debugging, investigating issues, or running commands:

Copy# Check for past solutions
recall <issue-keywords>

# Example searches:
recall path normalization project_path    # Path issues
recall sync default no-global             # Sync conventions
recall indexing parser session            # Indexing bugs
recall database foreign key               # Database issues

Why: We've likely solved this before. Check past learnings to save 30+ minutes.

2. Check CLAUDE.md Conventions

Review the "Command Defaults & Conventions" section in CLAUDE.md:

• Command defaults (sync, index, clean)
• Flag behavior (--no-global, --force, --dry-run)
• Debugging protocol

3. Always Use Dry-Run First

For commands that modify data:

Copycai sync --dry-run          # Preview before syncing
cai clean --dry-run         # Preview before cleaning
cai index --force --verbose # Show what's happening

---

Command Defaults & Critical Flags

cai sync

• Default behavior: Project-only learnings (--no-global)
• NEVER run without: --dry-run first
• Rare cases only: --global (includes global learnings)

Copy# ✓ CORRECT workflow
cai sync --dry-run    # 1. Preview
# Review output carefully
cai sync              # 2. Apply (project-only by default)

# ❌ WRONG workflow
cai sync              # Skipped preview!
cai sync --global     # Includes irrelevant global learnings!

cai index

• Default: Incremental (new sessions only)
• Use --force: After parser fixes, path changes, schema updates
• Use --verbose: When debugging parse issues

Copycai index                    # Normal: index new sessions
cai index --force --verbose  # After fixes: reindex all

cai clean

• ALWAYS use --dry-run first: Deletions are permanent

---

Debugging Protocol (Mandatory Steps)

When a user reports an issue, follow this order:

Step 1: Use Recall (5 seconds)

Copyrecall <error-keywords> <component-name>

Check if we've seen this before. Review past learnings.

Step 2: Check CLAUDE.md (30 seconds)

• Read "Command Defaults & Conventions"
• Check "When to Use Recall" section
• Review recent learnings in the auto-generated section

Step 3: Check Git History (1 minute)

Copygit log --all --grep="<keyword>" --oneline
git show <commit-hash>

Look for recent fixes to the same component.

Step 4: Investigate Code (only after above)

Now you can dive into the code with context from past solutions.

---

Common Issues & Solutions

Issue: "cai correlate not finding sessions"

First action: recall path normalization project_path mismatch Likely cause: Path inference bug (dashes vs slashes) Solution documented in: Past learnings about inferProjectPath()

Issue: "Global learnings in project CLAUDE.md"

First action: recall sync default no-global Root cause: Forgot to use default behavior Solution: cai sync defaults to --no-global (project-only)

Issue: "Sessions not indexed"

First action: recall indexing discovery session jsonl Check: Session file locations, parser updates, path resolution Solution: Often needs cai index --force after parser changes

---

Saving New Learnings

When you discover something important:

Copy# Use remember tool
remember --type pattern --scope project "Your learning here"

# Add relevant tags
--tags debugging,path-issues,sync,convention

# Then sync to CLAUDE.md (if appropriate)
cai sync --dry-run
cai sync

---

What NOT to Do

❌ Don't Skip Recall

Copy# User: "cai correlate not working"
# ❌ WRONG: grep -r "correlate" packages/
# ✓ CORRECT: recall correlate sessions path

❌ Don't Skip Dry-Run

Copy# ❌ WRONG: cai sync
# ✓ CORRECT: cai sync --dry-run, review, then cai sync

❌ Don't Assume Defaults

Copy# ❌ WRONG: Assume sync includes global
# ✓ CORRECT: Check CLAUDE.md conventions

❌ Don't Repeat Past Mistakes

Copy# ❌ WRONG: Debug path issues from scratch
# ✓ CORRECT: recall path normalization (finds past solutions)

---

Workflow Checklist

Before executing commands that modify data:

• Used recall tool with relevant keywords
• Checked CLAUDE.md "Command Defaults & Conventions"
• Ran command with --dry-run flag first
• Reviewed output carefully
• Verified scope (global vs project-specific)
• Ready to document new learnings if discovered

---

Key Principles

1. Memory First: Check recall before debugging
2. Preview First: Use --dry-run before applying
3. Context First: Read CLAUDE.md conventions
4. Document: Save learnings for next time
5. Project-Only: Default to project-specific learnings

---

Success Metrics

• ✓ Zero repeated debugging of solved issues
• ✓ Zero accidental global learning pollution
• ✓ Zero data loss from skipped dry-runs
• ✓ 100% workflow checklist compliance

Remember: This project is about building memory and learning from the past. Practice what we preach!