Strategic Investigation: Plan Mode Recovery

Strategic Investigation: Plan Mode Recovery

CORE PRINCIPLE: After 1-2 failed attempts, STOP trying variations. Enter plan mode and investigate properly using parallel subagents.

When to Use This Skill

Use this skill when:

• You've tried 1-2 approaches and they didn't work
• You don't understand WHY something isn't working
• You're tempted to try "just one more thing" or add a hacky workaround
• The error message doesn't make sense
• Multiple approaches seem valid and you're unsure which is idiomatic
• You find yourself guessing instead of knowing

Warning signs you should have used this earlier:

• You've been trying variations of the same approach for 5+ minutes
• You're adding workarounds or "temporary" hacks
• The code is getting messier instead of cleaner
• You're saying "this should work but doesn't"

The Process

Step 1: STOP and Acknowledge

When you recognize you're stuck:

1. Stop immediately - Don't try "just one more thing"
2. Tell the user:

   "I've tried [X approaches] and they're not working. Instead of continuing
   to guess, I'm going to step back and investigate this systematically using
   parallel subagents. This should find the idiomatic solution."
   

Step 2: Enter Plan Mode

Request plan mode to structure your investigation. Plan mode signals you're stepping back from implementation to investigate properly.

Step 3: Launch Parallel Subagents

Spawn 2-4 focused Explore agents in parallel to gather information:

Subagent Design Principles:

• Each agent gets ONE focused question
• Agents explore DIFFERENT angles (not variations of same thing)
• Agents are read-only (Explore type)
• Each returns concrete findings with file paths and code snippets

Common Explore Agent Types:

Explore Agent 1: Pattern Search
- Search labs/​packages/​patterns/ for similar patterns
- Search patterns/​examples/ for curated examples
- Look for patterns that solve similar problems
- Return: file paths, relevant code snippets, patterns used

Explore Agent 2: Documentation Deep-Dive
- Read relevant sections of ~/​Code/​labs/​docs/​common/
- Focus on DEBUGGING.md, PATTERNS.md, CELLS_AND_REACTIVITY.md
- Look for exact guidance on this situation
- Return: relevant doc sections, warnings, recommended approaches

Explore Agent 3: Community Knowledge
- Search community-docs/​superstitions/ for similar issues
- Check if others have hit this problem
- Return: relevant superstitions, known workarounds

Explore Agent 4: Framework Conventions
- Study how working patterns structure similar code
- What's the "frameworky" way to do this?
- Return: conventions observed, idioms to follow

Step 4: Synthesize Findings

After agents complete:

1. Review gathered information:

   • What patterns did similar code use?
   • What exactly do the docs say?
   • What community knowledge exists?

2. Identify the idiomatic solution:

   • Which approach follows framework conventions?
   • What do working examples do differently?
   • What would a framework expert do?

3. Understand WHY this is the right solution:

   • Don't just copy-paste from examples
   • Be able to explain why this works
   • Understand the underlying principle

Step 5: Execute with Confidence

Now implement the chosen approach:

• You understand WHY this approach is right
• You've verified it matches framework conventions
• You're not guessing anymore

Step 6: Handle Continued Failure

If the solution doesn't work after investigation:

1. Return to Step 3 with NEW context: "Approach X didn't work because Y"
2. Investigate why the expected solution failed
3. This often reveals the real underlying issue

After 2 plan-mode investigation cycles without resolution:

• Escalate to user
• Summarize what you investigated
• Show approaches you tried
• Explain what you learned
• This gives user actionable context to help

Example Investigation

Scenario: "Cell not updating reactively"

Failed attempts: Changed dependencies, added .get() calls, restructured code

Investigation plan (launch in parallel):

Explore Agent 1: Pattern Search
"Search labs/​packages/​patterns/ for computed() usage patterns.
Find 3+ examples of computed cells that update correctly.
Note what they do differently from my failing code."

Explore Agent 2: Documentation
"Read ~/​Code/​labs/​docs/​common/​CELLS_AND_REACTIVITY.md thoroughly.
Focus on computed() section and any gotchas.
Look for 'don't do X' warnings."

Explore Agent 3: Community Knowledge
"Search community-docs/​superstitions/ for 'computed', 'reactive', 'update'.
Check if this is a known edge case with solutions."

Explore Agent 4: Framework Behavior
"Study how other patterns structure reactive computations.
What cell access patterns do working examples use?"

Likely synthesis:

• Reactivity requires direct cell access, not derived values
• Creating new cells inside computed() breaks tracking
• Need to maintain stable cell references
• Idiomatic approach: lift cell creation to pattern body level

Good vs Bad Uses

GOOD - Use Strategic Investigation

• "I've tried computed() and derive() and neither updates. Let me investigate."
• "The type error doesn't make sense. Let me understand what the framework expects."
• "This feels hacky. Let me find the idiomatic way."
• "I'm not sure which approach is right. Let me explore the options properly."

BAD - Don't Use Strategic Investigation

• Simple typos - just fix them
• Missing imports - just add them
• Error message tells you exactly what's wrong - just apply the fix
• Problem you've solved before - apply known pattern

Key Principle: Idiomatic Over Working

The goal is not just to find SOMETHING that works, but to find the IDIOMATIC solution:

• Hacky: "I'll just cast to any and suppress the error"
• Idiomatic: "The framework expects X pattern, let me use that"

If investigation reveals the "right" way is significantly different from your approach, that's valuable - even if your hack might have worked.

Integration with Other Skills

Remember

Investigation time is NOT wasted time.

A 5-minute proper investigation often saves 30 minutes of trial-and-error on wrong approaches. Don't skip plan mode thinking "this should be simple" - that's often when you need it most.

When in doubt, investigate. It's better to investigate a simple problem than to spin wheels on a complex one.